CN120408530A

CN120408530A - Inspection method and system based on multimodal sensor fusion

Info

Publication number: CN120408530A
Application number: CN202510770669.7A
Authority: CN
Inventors: 崔佩仪; 赵伟; 刘奕清; 郭芷君; 林丽攀; 宋玲丽; 李依宁; 马敦豪; 方创明
Original assignee: Guangdong Junhua Energy Technology Co ltd
Current assignee: Guangdong Junhua Energy Technology Co ltd
Priority date: 2025-06-10
Filing date: 2025-06-10
Publication date: 2025-08-01

Abstract

The invention relates to the technical field of multi-mode data processing and discloses a multi-mode sensor fusion inspection method and system, wherein the method comprises the steps of collecting multi-mode original data of power equipment through a multi-mode sensor in an inspection robot and constructing a feature vector set; the method comprises the steps of carrying out self-adaptive weight calculation on the multi-mode sensors according to a feature vector set to obtain a sensor weight set, carrying out conflict recognition and solution on multi-mode original data to obtain a fusion data set, carrying out abnormal feature extraction on the power equipment based on the fusion data set to obtain an abnormal feature set, carrying out inspection track optimization based on the abnormal feature set to obtain a target inspection path sequence, and carrying out equipment state joint prediction by combining historical equipment inspection data to obtain an equipment failure prediction result.

Description

Inspection method and system for multi-mode sensor fusion

Technical Field

The invention relates to the technical field of multi-mode data processing, in particular to a multi-mode sensor fusion inspection method and system.

Background

The traditional power equipment inspection mainly relies on manual periodic inspection and single sensor monitoring, and the mode has the problems of long inspection period, limited coverage range, large interference of human factors and the like, and is difficult to discover potential faults and abnormal states of equipment in time. Along with the development of robot technology and sensor technology, intelligent inspection robots based on multi-mode sensors are gradually applied to the field of power equipment monitoring, and can be provided with various sensors such as thermal imaging, infrared rays, ultrasonic waves, electromagnetic fields, vibration and the like, so that the comprehensive and multidimensional monitoring of power equipment is realized.

However, multi-modal sensors face technical challenges of data collision and information fusion in practical applications. Different types of sensors may have inconsistent or even contradictory monitoring results of the same device at the same time due to differences in working principles, measurement accuracy, environmental adaptability and the like. The traditional data fusion method generally adopts a simple weighted average or majority voting mechanism, so that the problem of dynamic conflict among sensors can not be effectively solved, and fusion result distortion or misjudgment is easy to cause. In addition, the conventional routing inspection path planning mostly adopts a fixed route or an experience-based path selection, lacks comprehensive consideration of real-time states and fault risks of equipment, and is difficult to realize optimal configuration of inspection resources and accurate prediction of faults.

Disclosure of Invention

The invention provides a multi-mode sensor fusion inspection method and system, which can accurately identify data conflict among multi-mode sensors, realize intelligent screening and fusion of conflict data, and further realize more accurate equipment state joint prediction.

In a first aspect, the present invention provides a multi-mode sensor fusion inspection method, where the multi-mode sensor fusion inspection method includes:

Acquiring multi-mode original data of the power equipment through a multi-mode sensor in the inspection robot, and constructing a feature vector set;

performing self-adaptive weight calculation on the multi-modal sensor according to the feature vector set to obtain a sensor weight set;

Based on the sensor weight set, carrying out conflict recognition and solution on the multi-mode original data to obtain a fusion data set;

Performing abnormal feature extraction on the power equipment based on the fusion data set to obtain an abnormal feature set;

And carrying out inspection track optimization based on the abnormal feature set to obtain a target inspection path sequence, and carrying out equipment state joint prediction by combining historical equipment inspection data to obtain an equipment failure prediction result.

With reference to the first aspect, in a first implementation manner of the first aspect of the present invention, the collecting, by a multi-mode sensor in the inspection robot, multi-mode raw data of the power device, and constructing a feature vector set, includes:

acquiring multi-mode raw data of the power equipment through a multi-mode sensor of the inspection robot, wherein the multi-mode sensor comprises a thermal imaging sensor, an electromagnetic field sensor and a vibration sensor;

Noise filtering, data normalization and time synchronization processing are carried out on the multi-mode original data to obtain a preprocessed standardized data set, wherein the preprocessed standardized data set comprises thermal imaging sensor data, electromagnetic field sensor data and vibration sensor data;

extracting temperature distribution characteristics of the thermal imaging sensor data to obtain a thermal characteristic parameter set containing a temperature peak value, a temperature mean value and a temperature variance, and generating a thermal characteristic vector based on the thermal characteristic parameter set;

Performing electromagnetic field intensity spectrum analysis on the electromagnetic field sensor data to obtain an electromagnetic characteristic parameter set, and generating an electromagnetic characteristic vector based on the electromagnetic characteristic parameter set;

Performing vibration signal time-frequency domain transformation on the vibration sensor data to obtain a vibration characteristic parameter set, and generating a vibration characteristic vector according to the vibration characteristic parameter set;

and taking the thermal characteristic vector, the electromagnetic characteristic vector and the vibration characteristic vector as characteristic vector sets.

With reference to the first aspect, in a second implementation manner of the first aspect of the present invention, the performing adaptive weight calculation on the multi-modal sensor according to the feature vector set to obtain a sensor weight set includes:

Calculating output fluctuation of the multi-mode sensor under different temperature conditions based on the feature vector set to obtain a thermal stability index, analyzing the historical measurement accuracy of the multi-mode sensor to obtain an accuracy historical index, and evaluating the signal-to-noise ratio of the multi-mode sensor to obtain a signal strength index;

Respectively setting a differentiated thermal stability coefficient, an accuracy history coefficient and a signal intensity coefficient according to the type of the power equipment;

Obtaining a weight value of each sensor in the multi-mode sensor by calculating the sum of the product of the thermal stability index and the thermal stability coefficient, the product of the precision history index and the precision history coefficient and the product of the signal strength index and the signal strength coefficient;

The weight values of the sensors are arranged according to the sequence of the sensor identifiers and form a sensor weight set.

With reference to the first aspect, in a third implementation manner of the first aspect of the present invention, the performing conflict recognition and resolution on the multi-mode raw data based on the sensor weight set to obtain a fused data set includes:

calculating the data consistency of measurement values among different sensors based on the sensor weight set and the multi-mode raw data, and judging that data conflict exists when the data consistency is lower than a consistency threshold value to obtain a sensor conflict identification result;

Performing characteristic correlation analysis on the collision sensor according to the sensor collision recognition result to obtain a characteristic correlation function value;

Inputting SyncDrop-E conflict resolution algorithm into the sensor weight set, the characteristic correlation function value and the power equipment deviation threshold value to perform data screening to obtain an initial conflict resolution result;

And constructing a sensor trust network diagram according to the initial conflict resolution result, identifying the most reliable sensor set by applying a maximum weight spanning tree algorithm, and simultaneously executing time window analysis on periodic conflicts to obtain a fusion data set.

With reference to the first aspect, in a fourth implementation manner of the first aspect of the present invention, the inputting SyncDrop-E the sensor weight set, the feature correlation function value, and the power device deviation threshold value into the conflict resolution algorithm performs data screening to obtain an initial conflict resolution result, including:

inputting SyncDrop-E conflict resolution algorithm to the sensor weight set, the characteristic correlation function value and the power equipment deviation threshold value, identifying the sensor data pair with conflict and extracting the corresponding weight value to obtain a conflict data input set containing conflict sensor identification, measurement data and weight value;

based on the conflict data input set, weighting the measurement data of each conflict sensor to obtain weighted sensor data values;

Calculating the ratio of the absolute difference value between the data and the deviation threshold value of the power equipment according to the weighted sensor data value, and subtracting the ratio from 1 to obtain a deviation correction factor;

and performing product operation on the maximum value in the weighted sensor data value and the deviation correction factor to obtain the most reliable sensor data subjected to deviation correction, and generating a corresponding initial conflict resolution result based on the most reliable sensor data subjected to deviation correction.

With reference to the first aspect, in a fifth implementation manner of the first aspect of the present invention, the extracting abnormal features of the electrical device based on the fused data set to obtain an abnormal feature set includes:

performing feature mapping on the fusion data set to obtain a feature space vector of the power equipment;

setting a normal lower limit value and a normal upper limit value of an electrical characteristic, a thermal characteristic and a mechanical characteristic according to the type of the electrical equipment to obtain an equipment state boundary set containing a minimum boundary vector and a maximum boundary vector;

Calculating the minimum distance from the feature space vector to the equipment state boundary set and combining the equipment type sensitivity coefficient to obtain an abnormality index value;

and carrying out abnormality judgment and feature classification processing according to the abnormality index value to obtain an abnormal feature set.

With reference to the first aspect, in a sixth implementation manner of the first aspect of the present invention, the performing, based on the abnormal feature set, patrol track optimization to obtain a target patrol path sequence, and performing device state joint prediction with reference to historical device patrol data to obtain a device fault prediction result, where the method includes:

Performing fault probability density calculation based on the abnormal feature set to obtain a fault probability density value of the power equipment;

carrying out importance scoring calculation on the power equipment according to the fault probability density value to obtain an equipment evaluation parameter set;

inputting EIPO the fault probability density value, the equipment evaluation parameter set and the physical distance between the equipment into a patrol optimization algorithm to execute multi-objective constraint patrol track optimization to obtain a target patrol path sequence;

And carrying out equipment state joint prediction on the target inspection path sequence and the historical equipment inspection data to obtain an equipment failure prediction result.

With reference to the first aspect, in a seventh implementation manner of the first aspect of the present invention, the performing, by using the fault probability density value, the equipment evaluation parameter set, and the inter-equipment physical distance input EIPO, a multi-objective constrained inspection trajectory optimization algorithm to obtain a objective inspection path sequence includes:

Inputting EIPO the fault probability density value, the equipment evaluation parameter set and the physical distance between the equipment into a patrol optimization algorithm to construct a comprehensive optimization objective function comprising distance cost, fault risk and equipment importance;

Performing self-adaptive parameter setting on a distance weight coefficient, a fault probability weight coefficient and an equipment importance weight coefficient in the comprehensive optimization objective function according to the type of the inspection scene to obtain an optimization parameter configuration set;

Performing iterative solution based on the comprehensive optimization objective function and the optimization parameter configuration set to obtain a candidate patrol path solution set;

and performing convergence judgment and optimal solution selection on the candidate patrol path solution set to obtain a target patrol path sequence.

With reference to the first aspect, in an eighth implementation manner of the first aspect of the present invention, performing device state joint prediction on the target routing sequence and the historical device routing data to obtain a device failure prediction result, where the method includes:

Constructing an equipment association graph structure comprising equipment nodes, association edges and association strength weights based on the target routing inspection path sequence and the historical equipment routing inspection data;

carrying out power equipment state joint prediction on the equipment association diagram structure and the current equipment state sequence to obtain a predicted state data set;

Simulating the propagation process of the fault in the equipment network through a heat conduction equation based on the prediction state data set and the equipment association diagram structure, and calculating the equipment specific diffusion coefficient to obtain fault probability distribution data;

and performing fault risk assessment and visualization processing according to the fault probability distribution data to obtain a device fault prediction result comprising potential fault propagation paths and device risk levels.

In a second aspect, the present invention provides a multi-modal sensor-fused inspection system, the multi-modal sensor-fused inspection system comprising:

the acquisition module is used for acquiring multi-mode original data of the power equipment through a multi-mode sensor in the inspection robot and constructing a feature vector set;

the computing module is used for carrying out self-adaptive weight computation on the multi-modal sensor according to the characteristic vector set to obtain a sensor weight set;

the conflict recognition module is used for carrying out conflict recognition and resolution on the multi-mode original data based on the sensor weight set to obtain a fusion data set;

the feature extraction module is used for extracting abnormal features of the power equipment based on the fusion data set to obtain an abnormal feature set;

and the joint prediction module is used for optimizing the inspection track based on the abnormal feature set to obtain a target inspection path sequence, and combining historical equipment inspection data to perform equipment state joint prediction to obtain an equipment failure prediction result.

According to the technical scheme provided by the invention, the reliability index of each sensor can be dynamically quantized by constructing the sensor self-adaptive weight evaluation model based on the thermal stability, the precision history and the signal strength, the problem that the traditional fixed weight method cannot cope with the aging of the sensor, the environmental interference and the change of the equipment state is effectively solved, and the accuracy and the robustness of multi-mode data fusion are ensured. By combining the dynamic confidence coefficient matrix and the characteristic correlation function, the data conflict among the multi-mode sensors can be accurately identified, intelligent screening and fusion of conflict data are realized, and compared with a traditional simple weighted average or majority voting mechanism, the method has stronger conflict processing capability and data fusion precision. By establishing the three-dimensional feature space and the equipment state boundary, millisecond-level conflict data real-time processing is realized, the computational bottleneck of the traditional deep learning method on the embedded platform is avoided, and the real-time requirement of power equipment inspection is met. By constructing the power equipment fault probability density function as a core constraint, multi-objective optimized routing inspection path planning is realized, and equipment states, importance and spatial distribution are comprehensively considered. The method can effectively capture the spatial association relation between the power equipment, solves the limitation that the traditional time sequence model cannot process the spatial dependence by establishing the equipment association diagram and analyzing the physical connection, the electrical influence and the fault propagation relation, and realizes more accurate equipment state joint prediction. The propagation process of faults in the equipment network is simulated through a heat conduction equation, potential fault propagation paths and chain reaction risks can be identified, algorithm parameters and optimization strategies can be adaptively adjusted for different power equipment types and inspection scenes by the system, the algorithm parameters comprise sensor weight coefficients, deviation thresholds, inspection weight parameters and the like, and the universality and the effectiveness under different application environments are ensured.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings may be obtained based on these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram illustrating steps of a multi-mode sensor fusion inspection method according to an embodiment of the present invention;

Fig. 2 is a schematic structural diagram of a multi-mode sensor fusion inspection system according to an embodiment of the present invention.

Detailed Description

The embodiment of the invention provides a multi-mode sensor fusion inspection method and system. The terms "first," "second," "third," "fourth" and the like in the description and in the claims and in the above drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments described herein may be implemented in other sequences than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed or inherent to such process, method, article, or apparatus.

For ease of understanding, a specific flow of an embodiment of the present invention is described below, referring to fig. 1, and an embodiment of a multi-mode sensor fusion inspection method in an embodiment of the present invention includes:

Step S1, acquiring multi-mode original data of power equipment through a multi-mode sensor in the inspection robot, and constructing a feature vector set;

it can be understood that the execution body of the present invention may be a multi-mode sensor fusion inspection system, and may also be a terminal or a server, which is not limited herein. The embodiment of the invention is described by taking a server as an execution main body as an example.

Specifically, through the multimode sensing unit installed at the front end, the top or the rotatable platform of the inspection robot, corresponding thermal, electromagnetic and mechanical response information is synchronously acquired at the operation site of the power equipment, and the original multimode data stream with continuous time sequence is acquired through a high-frequency sampling mechanism. The method comprises the steps of sequentially executing three standardized processing steps of noise filtering, data normalization and time synchronization on original data, wherein a band-pass filtering algorithm is adopted for noise filtering, environmental vibration and thermal drift interference are removed while effective signals in a range from 10Hz to 500Hz are reserved, different physical quantities are uniformly mapped to a [ -1,1] range through a maximum and minimum value standardization strategy in normalization processing, the problem that the dimensions of heat, field and vibration signals are inconsistent is solved, and three data streams are aligned to a 1ms level through a high-precision timestamp correction algorithm in time synchronization, and a standardized multi-mode data set with uniform structure and consistent time sequence is constructed. And respectively carrying out targeted feature extraction operation on the three types of data, wherein the thermal imaging sensor data extracts a temperature peak value, a full-image temperature mean value and a local temperature variance of a key part in a thermal image through thermal image decoding and regional statistical analysis, constructs a thermal feature parameter set reflecting the abnormal trend of thermal distribution, and generates a thermal feature vector based on the thermal feature parameter set so as to describe the tiny change trend of the thermal state of the surface of equipment. The electromagnetic field sensor data takes frequency spectrum analysis as a core, adopts fast Fourier transform or wavelet packet transform to carry out frequency domain analysis on electromagnetic field intensity data, captures amplitude variation of the electromagnetic field intensity data in high frequency and harmonic intervals, constructs an electromagnetic characteristic parameter set containing main frequency intensity, frequency band energy distribution and spectrum balance factors by identifying signals such as power frequency fundamental wave offset, harmonic energy abnormality and the like, and generates electromagnetic characteristic vectors according to the electromagnetic characteristic parameter set, thereby describing the symptoms of electromagnetic leakage or insulation degradation of equipment. And for vibration sensor data, the vibration waveform is subjected to multi-scale feature extraction by combining a time-frequency domain analysis means and combining short-time Fourier transform and Hilbert-Huang transform to form a vibration feature parameter set comprising a vibration amplitude peak value, a main frequency component, an energy concentration degree, an impact response index and the like, and a vibration feature vector capable of distinguishing fault modes such as mechanical looseness, fatigue cracks and the like is generated. And combining the thermal feature vector, the electromagnetic feature vector and the vibration feature vector according to a unified structure to form a feature vector set.

S2, performing self-adaptive weight calculation on the multi-modal sensor according to the feature vector set to obtain a sensor weight set;

specifically, based on the feature vector set, three performance indexes of thermal stability, historical accuracy and signal intensity, which are exhibited by each sensor in the current inspection period, are analyzed and quantitatively modeled. The evaluation of the thermal stability index is to quantify the response consistency under the thermal disturbance condition by means of the output change data of each sensor under different environmental temperature conditions and by counting the output fluctuation of each sensor in a set temperature interval, namely, the standard deviation of thermal characteristic parameters such as temperature peaks or average values in the process of sampling multiple times of the same equipment area is measured and calculated, so that the capability of the thermal imaging or electromagnetic induction sensor for environmental adaptation is evaluated. Meanwhile, in order to reflect the measurement history reliability of each sensor, the coincidence degree between the sensor and the state of the known equipment in the previous multi-round inspection task is analyzed, the deviation between the characteristic value and the reference value generated in each acquisition period is calculated, and the deviation is summarized into a history precision index for describing the long-term stability and reliability trend of the sensor. The signal strength index is realized through signal-to-noise ratio evaluation, the system measures the power of the collected signals of each type of sensor in the current working period, and calculates the corresponding signal-to-noise power ratio by combining the background noise level, so that the effective data capturing capability of the sensor in the complex electromagnetic or vibration environment is obtained. Coefficient factors required for weight calculation, including a thermal stability coefficient, a precision history coefficient and a signal intensity coefficient, are set for equipment such as a transformer, a high-voltage switch, a power transmission line and the like. For example, when the transformer is inspected, the heat stability is more concerned, so that the heat stability weight is higher in the corresponding coefficient setting, and when the transmission line data is processed, the proportion of the signal strength coefficient is obviously improved due to stronger electromagnetic interference. Based on the differential configuration, weighting calculation is carried out on each sensor, the current thermal stability index of the sensor is multiplied by the thermal stability coefficient under the corresponding equipment type, the precision history index is multiplied by the corresponding precision coefficient, the signal strength index is multiplied by the signal strength coefficient, and the three weighted results are summed to obtain the total weight value of the current sensor in the current inspection task. And sequencing the calculation results according to the unique identifier of each sensor, and sequentially arranging the weight values of all the sensors according to a fixed sequence to construct a sensor weight set. The weight set is used as a confidence coefficient in the following data conflict recognition and fusion stage to participate in the decision process and is dynamically updated along with the change of the task environment, so that the real-time adaptation to the factors such as sensor aging, environmental disturbance, equipment state change and the like is realized.

S3, carrying out conflict recognition and solution on the multi-mode original data based on the sensor weight set to obtain a fusion data set;

Specifically, the sensor weight set and the multi-mode raw data collected by the inspection robot are input into a data consistency calculation model in a combined mode, the model is based on time sequence, measurement values of a plurality of sensors (such as thermal imaging, electromagnetic induction and vibration sensors) in a same target area are paired and compared in a same time window, and the consistency measurement between the values is calculated according to the measurement values. The consistency index is realized by normalizing a relative error function according to the relative difference between measured values, and the measured value difference of each pair of sensors is normalized with the sum of the measured total value and a minimum value in the comparison process so as to avoid the problem of unstable numerical value caused by an extreme value. When the consistency index between the measured values of two or more sensors is lower than a preset threshold value (such as 0.85), the system judges that the data conflict occurs, and records the conflict moment, the conflict sensor number and the conflict type to form a preliminary sensor conflict identification result. Based on the identified conflicting sensors, a feature correlation analysis is performed to construct a correlation function that quantifies the similarity of measured trends between conflicting sensors. The function comprehensively examines response trends of the conflict sensors to key characteristic parameters (such as temperature peak values, main frequency of electric field intensity or vibration energy density) in a plurality of time steps in the past, and obtains a correlation value between each pair of conflict sensors by carrying out mean and variance normalization on the historical characteristic vectors of the conflict sensors and further determining whether data conflict is caused by short-time interference or caused by sensor precision decline or signal drift. The conflict recognition result, the sensor weight set, the feature correlation function value, and a deviation threshold set for a specific power device (e.g., the allowable temperature deviation of the transformer is 15 ℃, the surface of the insulator is 25 ℃, etc.) are input together to SyncDrop-E conflict resolution algorithm. The algorithm is based on a weighted voting mechanism and combines a maximum reliability optimization principle, data which is more stable and is close to a threshold range in a measured value multiplied by a weight is selected from a value pair with conflicts to serve as a reserved value, and a relative offset is taken as a compensation correction coefficient to generate a preliminary conflict resolution result. The maximum weighted effective value is directly adopted in a single conflict pair, and in a complex scene of multi-sensor conflict, the fusion optimization is further needed to be carried out by depending on a structured graph model. In order to enhance the processing capacity of the system to complex scenes, a sensor trust network diagram is constructed on the basis of the preliminary result, all relevant sensors are regarded as nodes in the diagram, and the edge weights among the nodes are data consistency indexes at corresponding moments. And automatically screening out a sensor set with the strongest global consistency and the optimal data quality through a maximum weight spanning tree algorithm in graph theory, so as to ensure that a final decision is based on a stable and reliable data subset. Meanwhile, in consideration of the phenomenon of repetitive collision caused by external periodic interference or structural vibration at certain parts of the power equipment, the system introduces a time window analysis mechanism in parallel, carries out sliding window analysis on the output trend of each suspected abnormal sensor in a plurality of continuous sampling periods, and eliminates misjudgment data which shows periodic fluctuation but has no actual fault characteristics. A fused dataset is obtained.

The set of sensor weights, the feature correlation function value, and the power device bias threshold are input SyncDrop-E into a conflict resolution algorithm. The algorithm traverses the data pairs marked as abnormal in the conflict identification result, retrieves the identifiers of the corresponding sensors, the original measured values of the current sampling period and the weight values at the moment one by one, and forms a structured conflict data input set together, wherein the input set encapsulates the numbers of the conflict sensors, the measured values and the confidence coefficients of the conflict sensors in a tuple mode. Based on the set of conflicting data inputs, weighting processing is performed on the measurement data of each conflicting sensor. And multiplying the measured values of two or more sensors in the same group of conflicts by corresponding weight coefficients respectively, and fusing the measured values into weighted sensor data values by a weighted average strategy. And carrying out deviation checking and correction processing on the weighted data, calculating the absolute difference of weighted values among different sensors, carrying out ratio operation on the difference and a deviation threshold set by the power equipment, and subtracting the ratio from 1 to obtain a deviation correction factor. The correction factor is understood to be a proportional attenuation of confidence correction in numerical logic, the closer the value is to 1, the smaller the deviation between data is, and the higher the credibility is, otherwise, if the deviation value is close to or even exceeds a set threshold value, the correction factor is close to 0, which indicates that the group of measurement results have larger deviation. And performing product operation on the maximum value in the weighted sensor data value and the deviation correction factor to obtain the most reliable sensor data subjected to deviation correction. And (3) the most reliable sensor data subjected to deviation correction is structured and output as an initial conflict resolution result, and flows into a subsequent trust network construction and multi-sensor fusion analysis flow as high-confidence data.

S4, carrying out abnormal feature extraction on the power equipment based on the fusion data set to obtain an abnormal feature set;

Specifically, the mapping and modeling of the feature space are performed on the current running state of the power equipment on the basis of the high-confidence fusion data set obtained after the multi-mode sensor conflict is resolved. And carrying out multidimensional feature mapping processing on the fusion data set by adopting MFEB model, and constructing a unified multi-modal feature vector through the mapping relation among three types of features of heat, electricity and force, wherein the vector not only reflects the physical properties of equipment surface temperature, internal electromagnetic behaviors, shell vibration and the like, but also realizes the same-space expression among cross-modal features, thereby converting the original physical parameters into standardized feature space vectors which are convenient for discriminant analysis. Each dimension in the vector structure corresponds to a predefined key index, such as a temperature peak value, electromagnetic harmonic intensity, vibration energy density and the like, so that the running state of the equipment is represented in a characteristic space as a point. And searching a preset equipment state boundary set according to the type (such as a transformer, a high-voltage circuit breaker, a lightning arrester or a disconnecting switch) of the current power equipment under inspection, wherein the set consists of a minimum boundary vector and a maximum boundary vector, and the minimum boundary vector and the maximum boundary vector respectively represent the lower limit and the upper limit of all characteristic dimensions of the equipment under the normal running state. The boundary parameters are obtained by historical operation data and expert experience statistical modeling, so that the boundary parameters can cover all state changes of the equipment in a normal fluctuation range. The system takes the boundary set as a reference frame, performs dimension-by-dimension comparison on the feature space vector and the boundary, and adopts a minimum distance calculation method to evaluate the degree of the current state deviating from the normal operation interval. The distance calculation considers the Euclidean or Manhattan distance between the vector and the boundary, and introduces a device type sensitivity coefficient as an adjusting factor to endow different types of devices with differentiated response sensitivity, so as to adapt to the differential requirements that a transformer is high in thermal stability but sensitive to electromagnetic disturbance and a breaker is frequent in vibration response but tolerant to temperature change. After sensitivity adjustment is performed on the minimum distance between the current feature space point and the boundary, an anomaly index value is calculated, wherein the value of the index value is closer to 1, so that the current state is more likely to deviate significantly, and if the index value is closer to 0, the equipment is within the safety boundary. And comparing the abnormality index value with a preset threshold value, triggering an abnormality judgment mechanism if the abnormality index value exceeds the threshold value, and entering a feature classification processing flow. In the process, classification labeling is carried out according to specific feature dimensions exceeding the boundary in the abnormal vector, if the thermal feature exceeds, the thermal abnormality is marked, if the electric field strength is deviated, the electric abnormality is marked, if the vibration parameter is suddenly changed, the mechanical abnormality is classified, and if a plurality of indexes are simultaneously exceeded, the compound abnormality is further classified. All device features determined to be abnormal are archived as an abnormal feature set that structures information such as device number, type of abnormality, value of abnormality, and specific overrun dimension.

And S5, carrying out inspection track optimization based on the abnormal feature set to obtain a target inspection path sequence, and carrying out equipment state joint prediction by combining historical equipment inspection data to obtain an equipment failure prediction result.

Specifically, the failure probability density calculation process is performed based on the state shift data and the abnormality determination result of each type of the electric power equipment included in the abnormality feature set. The process uses the distance between the fused feature vector and the equipment state boundary as a core parameter, combines the reliability factors mapped by the sensor weights in CWMM-Fusion algorithm, and carries out nonlinear normalization on the current abnormality degree through a sigmoid function to form a fault probability density value so as to quantify the possibility of fault occurrence of each equipment at the current moment. And calculating importance scores of the power equipment based on the fault probability density values, wherein the scores consider the current abnormality degree of the equipment, and introducing structural indexes and running environment factors for comprehensive evaluation. And integrating the factors such as position criticality, load importance level, standby redundancy rate and the like of the computing equipment in the power grid topological structure by adopting an analytic hierarchy process, and constructing an equipment evaluation parameter set by weighting and superposition, so that an integrated score reflecting the priority of each equipment in an operation and maintenance strategy is given to each equipment. The fault probability density value, the equipment evaluation parameter set and the actual physical distance data between the equipment are input EIPO into a routing optimization algorithm, the algorithm is a path optimization model under multi-objective constraint, the total length of the routing path, the equipment risk level and the importance score are comprehensively considered by an objective function, and the shortest path is ensured while the risk coverage and the key node access frequency are maximized. The optimization model adopts an improved genetic algorithm to realize path search, the path search is coded into a chromosome through equipment access sequence to carry out iterative evolution, sequence maintenance crossover and 2-opt local mutation operation are carried out in each generation, diversity and convergence speed of search space are ensured, optimization is stopped when the maximum iterative algebra or continuous multiple generations are not improved, a target inspection path sequence is output, and the sequence is generated according to logic of high risk and high weight equipment priority access and has dynamic self-adaptive characteristics. And introducing historical equipment inspection data and a state vector related to equipment in the current path to jointly construct an equipment state joint prediction model. The model builds a physical, electrical and historical fault propagation relation diagram among devices based on a graph rolling network, takes each device in an optimized path as a graph node, fuses a historical state change sequence as a graph node characteristic input model, captures high-order association information among devices through graph rolling operation, strengthens key device characteristic expression by using an attention mechanism, outputs a state value and a state change trend of the device in the next prediction period, and predicts a potential fault propagation path by means of a diffusion simulation method. And the equipment fault prediction result generated by the model is presented in a risk thermodynamic diagram form, and the expected risk level and the propagation range of each key node are marked.

And constructing EIPO a multi-objective optimization model of the inspection optimization algorithm by taking the fault probability density value output by the fusion analysis module, the equipment evaluation parameter set obtained based on the hierarchical analysis and the actual physical layout distance between the power equipment as core input variables. The core of the model is to construct an adjustable comprehensive optimization objective function which jointly considers the length cost of the inspection path, the fault risk degree of the current state of the power equipment and the importance level of the equipment in the operation of the power system, thereby forming a path planning strategy with global constraint and task driving dual characteristics for the inspection sequence and the dispatching mode. The objective function consists of three items in a mathematical structure, namely a physical distance measurement item among devices for reflecting the inspection efficiency, a fault probability density weighting item for emphasizing the prior detection of high-risk devices, and a device importance scoring weighting item for ensuring that key nodes of the power system obtain more inspection resources. After the objective function is constructed, three kinds of weight parameters in the objective function, namely a distance weight coefficient, a fault probability weight coefficient and an equipment importance weight coefficient, are adaptively set in order to adapt to different inspection task types and operation scene characteristics. The setting is dynamically adjusted according to the situation classification of the inspection task, for example, a higher distance weight value is set when the path efficiency is emphasized in the conventional periodic inspection, a fault risk coefficient is lifted in an emergency fault inspection scene to preferentially cover high-risk equipment, and the equipment importance coefficient weight is placed at the first position in the key equipment maintenance or seasonal electricity-protecting task. The system presets various typical scene parameter configuration schemes, and supports dynamic generation of an optimized parameter configuration set through expert strategy or historical task feedback, so that adaptability of an optimized model to a complex running environment is enhanced. On the basis of the constructed comprehensive optimization objective function and the corresponding optimization parameter configuration set, the EIPO algorithm executes an iterative solving process, and the solving strategy is realized based on an improved genetic algorithm or a mixed ant colony algorithm. The equipment list is encoded into a group of access sequences to be ordered at this stage, each routing inspection path is a candidate solution individual, and the algorithm gradually develops a solution set through genetic operations such as crossing, mutation, selection and the like, and continuously approaches to a better solution. In the evolution process, the algorithm scores each generation of candidate paths according to an optimized objective function, performs sequencing selection through a fitness function, and performs path fine adjustment by combining a local search strategy (such as 2-opt or 3-opt), so as to ensure the balance between global convergence and local optimization, and form a solution set containing a plurality of high-quality inspection paths. And executing convergence judgment and optimal solution screening on the candidate routing inspection path solution set, triggering a convergence termination mechanism when the optimal solution of the continuous multi-generation solution set does not have obvious improvement or reaches the maximum iteration number, and selecting a path scheme with the lowest comprehensive cost, highest risk coverage and optimal hit rate of important equipment from the final solution set as a target routing inspection path sequence.

And constructing an equipment association graph structure with a topological structure and time sequence information by combining the target inspection path sequence output by the path optimization module and historical equipment inspection data accumulated in the power system for a long time. In the structure, each device to be inspected corresponds to a graph node, the node characteristics consist of a historical running state sequence, an abnormal mark, a sensor characteristic vector and the like, and the physical connection, the electrical interaction and the historical fault linkage record among the devices form edges in the graph. And (3) assigning an association strength weight to each edge, wherein the weight is obtained by combining the evaluation of the electrical dependency, the physical distance influence coefficient and the fault propagation probability among the devices, and a heterogeneous graph data structure reflecting the structural and functional dependency relationship among the devices is constructed. The device association graph structure and the current device state sequence are input into a graph rolling network model (GCN) to perform joint prediction of the power device state. The GCN takes the equipment nodes as a basic operation unit, the higher-order dependency relationship among the nodes is extracted through multi-layer graph convolution, state information is propagated layer by layer, and the model executes feature aggregation operation on each layer according to the adjacency matrix and the node features, so that the update of the states of all the nodes is not only influenced by the historical states of the nodes, but also the linkage trend of the states of peripheral equipment is reflected. Meanwhile, an attention mechanism is integrated in the model, feature enhancement processing is carried out on key nodes in the graph, high-risk equipment on a fault propagation path is paid priority attention to, and the accuracy and the discrimination of state prediction are improved. The model output is a predicted state dataset containing state estimates for each device at a next time step and providing an incremental index of state change trends for identifying exacerbating fault risk. And constructing a heat conduction type diffusion model based on the prediction state data set and the equipment association diagram structure together, wherein the heat conduction type diffusion model is used for simulating the propagation path and the influence range of the potential fault in the equipment network. By introducing a heat-like conduction equation, a diffusion relation between the state change rate and the state gradient of adjacent nodes is established between the nodes, and a device-specific diffusion coefficient of each node is calculated, wherein the coefficient is obtained according to the performance of the device serving as a fault propagation source or a propagation receptor in history, and the capability of the device serving as a fault conduction channel in a network is embodied. The model simulates the propagation process of fault signals in the graph through multi-step time evolution, dynamically deduces the probability of state degradation of each node in a plurality of time steps in the future, and forms fault probability distribution data. And inputting the fault probability distribution data into a fault risk assessment and visualization processing module. The risk assessment logic comprehensively considers the fault probability value, the diffusion rate, the path coverage and the key equipment hit rate, classifies the risk of all the nodes, and identifies the main propagation path of the fault. The path appears as a sequence of nodes with propagation probability and time delay estimates on each side, forming a complete potential fault diffusion chain. And the visualization module displays the prediction result in a graph thermodynamic diagram mode, different nodes label the fault risk level of the prediction result in a color intensity mode, and the propagation path is highlighted in a thickened and colored side mode to obtain the equipment fault prediction result.

In the embodiment of the invention, the reliability index of each sensor can be dynamically quantized by constructing the sensor self-adaptive weight evaluation model based on the thermal stability, the precision history and the signal strength, the problem that the traditional fixed weight method cannot cope with the aging of the sensor, the environmental interference and the change of the equipment state is effectively solved, and the accuracy and the robustness of multi-mode data fusion are ensured. By combining the dynamic confidence coefficient matrix and the characteristic correlation function, the data conflict among the multi-mode sensors can be accurately identified, intelligent screening and fusion of conflict data are realized, and compared with a traditional simple weighted average or majority voting mechanism, the method has stronger conflict processing capability and data fusion precision. By establishing the three-dimensional feature space and the equipment state boundary, millisecond-level conflict data real-time processing is realized, the computational bottleneck of the traditional deep learning method on the embedded platform is avoided, and the real-time requirement of power equipment inspection is met. By constructing the power equipment fault probability density function as a core constraint, multi-objective optimized routing inspection path planning is realized, and equipment states, importance and spatial distribution are comprehensively considered. The method can effectively capture the spatial association relation between the power equipment, solves the limitation that the traditional time sequence model cannot process the spatial dependence by establishing the equipment association diagram and analyzing the physical connection, the electrical influence and the fault propagation relation, and realizes more accurate equipment state joint prediction. The propagation process of faults in the equipment network is simulated through a heat conduction equation, potential fault propagation paths and chain reaction risks can be identified, algorithm parameters and optimization strategies can be adaptively adjusted for different power equipment types and inspection scenes by the system, the algorithm parameters comprise sensor weight coefficients, deviation thresholds, inspection weight parameters and the like, and the universality and the effectiveness under different application environments are ensured.

In a specific embodiment, the process of executing step S1 may specifically include the following steps:

extracting temperature distribution characteristics of thermal imaging sensor data to obtain a thermal characteristic parameter set containing a temperature peak value, a temperature mean value and a temperature variance, and generating a thermal characteristic vector based on the thermal characteristic parameter set;

The thermal eigenvector, the electromagnetic eigenvector, and the vibration eigenvector are set as eigenvectors.

Specifically, the inspection robot is provided with the multi-mode sensor array through the fixed or rotary cradle head structure in the process of executing operation, so that the inspection robot can acquire real-time data of the power equipment from multiple directions, and the inspection robot covers multiple physical dimensions including equipment shell surface heat distribution, electromagnetic leakage behavior in operation, mechanical structure operation stability and the like, so that the diversification of sensing dimensions and the comprehensiveness of data acquisition are ensured. In order to realize high-frequency and low-delay data capture, a distributed sampling structure is introduced in the design, and sampling frequency upper limits of different equipment types are set, for example, 100Hz, 75Hz and 50Hz sampling frequency configurations are adopted for a transformer, a high-voltage switch, a power transmission line and the like. And executing a unified preprocessing flow on the original multi-mode data, wherein the flow mainly comprises three key steps of noise filtering, data normalization and time synchronization. Because the acquisition environment of the original data is complex and a large amount of background interference noise exists, a band-pass filtering model is respectively constructed for each type of sensor signal, specific filtering parameter configuration is carried out according to the working frequency band of the band-pass filtering model, for example, effective signals are reserved in the range of 10Hz to 500Hz, and non-equipment signal interference caused by external vibration, environmental thermal fluctuation or electromagnetic disturbance is filtered. The output units and dimensions of different types of sensors are different, and direct fusion analysis has the problem of unbalanced numerical level, so that normalization operation is uniformly performed on all data, and all data are uniformly projected to the [ -1,1] numerical interval in a maximum-minimum mapping mode, so that structural alignment and statistical scale uniformity of multi-source data are realized. In order to ensure that the multi-mode data have strict time corresponding relation, a time synchronization mechanism based on high-precision time stamps is introduced, microsecond level alignment processing is carried out on various data according to the acquisition time of the data, and therefore a time basis with consistency among the multi-channel data is ensured. Through the above processing, a standardized data set is formed in which thermal imaging data, electromagnetic field data, and vibration data are uniformly stored and packaged into a structured time series data block. The method comprises the steps of respectively executing targeted feature extraction operation on three types of sensor data, namely, thermal imaging sensor data, wherein the data are presented in the form of infrared images or thermal matrixes, and extracting a temperature peak value, a full-frame temperature mean value and a temperature stability index constructed based on local area variance through local and global statistics on area thermal distribution of the images so as to form a thermal feature parameter set, wherein the parameter set reflects whether local overheating phenomenon exists on the surface of equipment or not, and describes fluctuation amplitude of long-term thermal distribution trend, and then the parameter set is encoded into a group of thermal feature vectors for subsequent model processing. For electromagnetic field sensor data, a frequency domain analysis method is adopted for unfolding processing, a time domain electric field signal is converted into frequency spectrum distribution by utilizing fast Fourier transform, then characteristic indexes such as main frequency component amplitude, harmonic energy density, frequency spectrum balance coefficient and the like are extracted, the indexes form an electromagnetic characteristic parameter set, and the electromagnetic characteristic parameter set is uniformly encoded into an electromagnetic characteristic vector by a system to be used for describing whether electromagnetic anomalies such as partial discharge, winding short circuit or grounding faults exist in the operation of an electric component. And carrying out multi-level processing on the vibration sensor data in the time-frequency domain joint transformation, extracting indexes such as instantaneous energy distribution, main frequency change, impact response index and the like of a vibration signal of the equipment by adopting a mode of combining short-time Fourier transformation and Hilbert-Huang transformation, forming a vibration characteristic parameter set, and converting the vibration characteristic parameter set into a vibration characteristic vector so as to describe potential problems such as bearing damage, structural looseness or external impact and the like in the operation of the equipment. and integrating the thermal feature vector, the electromagnetic feature vector and the vibration feature vector according to a unified format to construct a multi-dimensional feature vector set.

In a specific embodiment, the process of executing step S2 may specifically include the following steps:

Obtaining the weight value of each sensor in the multi-mode sensor by calculating the sum of the product of the thermal stability index and the thermal stability coefficient, the product of the precision history index and the precision history coefficient and the product of the signal strength index and the signal strength coefficient;

Specifically, based on the feature vector set, stability analysis is performed on output performances of various sensors at different environmental temperatures, so that a thermal stability index is calculated. The method comprises the steps of analyzing the fluctuation degree of output data of each sensor in different temperature sections by constructing a temperature interval sequence based on temperature mean value and temperature variance data in a thermal characteristic vector, namely extracting standard deviation of corresponding measured values of the sensor in each set temperature section, reflecting the value of the fluctuation in a normalization mode, and using the value as a thermal stability index to characterize the capability of the sensor for maintaining the measurement stability under Wen Biangan disturbance, wherein the lower the value is, the higher the stability is. And analyzing the measurement precision of each sensor based on the historical inspection data to construct a precision historical index. The process backtracks the output record of the sensor to the specific equipment characteristic value (such as temperature peak value, electromagnetic main frequency or vibration amplitude) in a plurality of past inspection cycles, compares the output record with the corresponding equipment operation actual state or expert evaluation result, calculates the deviation ratio of each historical measurement, calculates the historical measurement accuracy of the sensor by counting the proportion of the deviation ratios which accords with the set accuracy threshold value, and introduces the proportion as the accuracy historical index into a model, thereby reflecting the reliability trend of the sensor in the long-term use process, and the higher the numerical value is, the more reliable the long-term performance of the sensor is indicated. The signal intensity performance of the sensor in the current working environment is evaluated, namely, the signal-to-noise ratio is calculated to obtain the signal intensity index. And (3) carrying out energy analysis on the original data output by each sensor in the current period, respectively calculating the average power values of the signal component and the background noise component, and constructing a signal-to-noise ratio expression according to the average power values. In order to ensure the physical authenticity of the result, a main signal channel is extracted by utilizing band-pass filtering, the filtering result is regarded as noise channel input, the signal-to-noise value is obtained by comparing the power ratio of the main signal channel and the noise channel, and logarithmic scaling processing is carried out on the signal-to-noise value to form a signal strength index under a unified scale. The higher the index value is, the clearer and more distinguishable signal data can be provided by the sensor in the current environment, and the accuracy of fusion decision is improved. After the three indexes are calculated, background information of the equipment type is introduced, and a thermal stability coefficient, an accuracy history coefficient and a signal intensity coefficient are respectively set according to the type of the current equipment to be inspected, such as a transformer, a high-voltage circuit breaker, a lightning arrester or a cable branch box, and the like, and the coefficients reflect the weight of different equipment in different fault sensitivity dimensions. For example, the transformer is particularly sensitive to thermal drift, the thermal stability coefficient is set higher, the importance of electromagnetic measurement accuracy is higher and the precision coefficient is improved when the high-voltage circuit breaker is accompanied by electromagnetic surge, and the definition of a vibration signal is more important when the cable support system with a complex structure is adopted, so that the signal-to-noise ratio is higher. And automatically loading coefficient combinations adapting to the device types according to a preset rule table, so as to ensure that the evaluation model has structural adaptability to the device environment. After the corresponding matching of the indexes and the coefficients is completed, multiplying the three indexes of each sensor by the corresponding coefficients respectively, and then adding the three products to obtain the comprehensive weight value of the sensor under the current environment and task conditions. And sequentially arranging the weight values of all the sensors according to the corresponding unique identifiers thereof, and uniformly forming a structured sensor weight set. The weight set is cached in the system in the form of a vector and dynamically updated at a fixed period (e.g., 200 ms).

In a specific embodiment, the process of executing step S3 may specifically include the following steps:

Calculating the data consistency of measurement values among different sensors based on the sensor weight set and the multi-mode raw data, and judging that data conflict exists when the data consistency is lower than a consistency threshold value to obtain a sensor conflict recognition result;

And constructing a sensor trust network diagram according to the initial conflict resolution result, identifying the most reliable sensor set by applying a maximum weight spanning tree algorithm, and simultaneously executing time window analysis on the periodic conflict to obtain a fusion data set.

Specifically, data consistency of measurement values among different sensors is calculated based on the sensor weight set and the multi-modal raw data. The consistency calculation is based on the current measured value of each sensor and the confidence coefficient thereof in the weight set, a weighted normalization difference function is adopted for processing, namely, after difference ratio normalization is carried out on the same physical dimension data collected by any two sensors at the same time, sensor weights are introduced to carry out smooth regulation on the difference values, a consistency score between 0 and 1 is obtained, the closer the value is to 1, the more consistent the data is represented, if the value is lower than a set consistency threshold (such as 0.85), the obvious deviation exists between the data of the two sensors, and the system judges potential data conflict according to the obvious deviation. When the system identifies the sensors with data collision, namely, a sensor collision identification result is formed, the system starts a characteristic correlation analysis module to execute time sequence characteristic comparison on each pair of collision sensors. In the analysis process, historical feature vectors of each sensor in near K sampling periods are traced back, time sequence data are subjected to standardization processing in a mean value removal and variance normalization mode, and then the relevance degree of the time sequence data in feature trend dimensions is calculated by adopting pearson correlation coefficients or mutual information indexes to generate feature relevance function values reflecting the consistency of time sequence changes. The function value reflects numerically whether the observations of the two sensors have trend consistency over long periods of time despite short-term differences, if the correlation is strong, the system will give it a higher retention weight in subsequent conflict resolution. And inputting SyncDrop-E conflict solution algorithm into the sensor weight set, the characteristic correlation function value and the power equipment deviation threshold value to perform data screening. The algorithm takes the collision sensor measured value as an input, carries out confidence weighting calculation by combining the weight of the collision sensor measured value with the historical characteristic correlation of the paired sensor, then evaluates the deviation degree between each sensor data and the deviation threshold value, calculates a correction factor and applies the correction factor to the original data to carry out deviation compensation. The algorithm selects the maximum credible value obtained by multiplying the weight by the correction factor as the effective data output of the conflict pair, and marks the source sensor as the primary credible source. For the case where more than two sensors collide at the same time SyncDrop-E retains the trust factor for each trusted value. in order to improve the overall robustness of the multi-sensor system in a high-dimensional data fusion scene, a sensor trust network diagram is constructed based on the preliminary result, all sensors are used as diagram nodes, edge connection is constructed among the nodes through data consistency, and consistency scores are used as edge weights. And optimizing the trust graph by adopting a maximum weight spanning tree algorithm, and constructing a sub-graph structure with the maximum weight, wherein the sub-graph structure corresponds to the most reliable sensor subset. The non-trunk nodes are cut through the tree structure, so that automatic rejection of the low-confidence sensors is realized, high-reliability sensor groups with wide consistency connection relations are reserved, and the comprehensiveness and the credibility of the fusion data sources are ensured. In addition, the system synchronously executes a time window analysis mechanism for identifying periodic conflict problems. The mechanism builds a sliding window statistical model aiming at the sensor with abnormal fluctuation repeatedly occurring in a plurality of continuous sampling periods, analyzes the output abnormal value trend of the sensor, and judges the sensor as continuous pseudo-abnormality caused by internal device offset or environmental interference if the sensor is found to show non-single-point burst but continuous periodic abnormality in a plurality of windows, so that suppression weight is given during fusion, and the integral fusion result of the periodic distortion interference is avoided. After trust network screening and time window abnormal suppression are completed, all the reserved trusted sensor data are subjected to weighted integration, and a fused data set with high consistency and high confidence is output.

In a specific embodiment, the performing step inputs SyncDrop-E the sensor weight set, the feature correlation function value and the power device deviation threshold value to the conflict resolution algorithm for data filtering, and the process of obtaining the initial conflict resolution result may specifically include the following steps:

Inputting a sensor weight set, a characteristic correlation function value and a power equipment deviation threshold value into SyncDrop-E conflict resolution algorithm, identifying conflicting sensor data pairs, extracting corresponding weight values, and obtaining a conflict data input set containing conflict sensor identifications, measurement data and weight values;

Specifically, the dynamic weight value of the corresponding sensor is extracted from the sensor weight set, then a conflict data input set is constructed by combining the respective real-time measurement data and the unique sensor identification code recorded by the system, the input set is composed of a plurality of triplets, each triplet comprises the sensor identification, the current measurement value and the corresponding confidence weight thereof, and the triplets are basic information structures for judging the credibility and outputting the correction result in SyncDrop-E algorithm. When the conflicting data input set is constructed, syncDrop-E algorithm performs a weighting operation on each pair of conflicting sensors in the input set, calculating their weighted measurements. The weighting method is based on the measured value of each sensor, multiplies the measured value by a corresponding weight coefficient, and normalizes the product to ensure that the final value reflects the relative contribution specific gravity that the different sensors should occupy in the fusion decision. The weighted results form a new set of data points that physically represent a "confidence estimate" of the system for the current conflicting measurements. And judging the degree of deviation of the current data from the allowable operation fluctuation range of the equipment after the weighted sensor data values are obtained, and thus introducing a deviation threshold corresponding to the equipment type. Searching a preset maximum deviation tolerance value according to the type of equipment (such as a transformer, a circuit breaker, a bus connector and the like), then calculating absolute difference values of each pair of weighted data, and then carrying out ratio processing on the difference values and deviation threshold values to obtain a standardized deviation ratio, wherein the ratio reflects the relative degree of the current conflict data exceeding the tolerance range in value. To construct the correction factor, a first order inversion is performed on the ratio, i.e., 1 minus the ratio to construct a bias correction factor. The closer the factor is to 1 in value, the smaller the difference between weighted measurements, the more the data is concentrated within normal operating ranges with higher confidence, and when the factor approaches 0, the conflicting measurements have deviated significantly from tolerance boundaries and must be carefully handled or eliminated. The SyncDrop-E algorithm enters a final correction stage, extracts the maximum value from the weighted data value, and carries out product operation on the maximum value and the deviation correction factor obtained by calculation to obtain the 'most reliable sensor data' subjected to dynamic deviation correction. The product operation logic keeps strong signal contribution in data, and tolerance regulation is introduced, so that the final output value has steady-state inhibition capability when being close to the dominant measured value, and double control of weight driving and offset correction is realized in conflict decision. The process dynamically adjusts the contribution ratio of the maximum weight data to the final output through the deviation correction factor to form a preliminary conflict resolution result with high response precision and robustness. The most reliable correction data is marked as an initial conflict resolution result and saved into the fusion data set through a structured mapping and time sequence archiving mechanism. The system binds the correction data with the source sensor number thereof in the process to form a fusion identification vector so as to track the source path and the trust level thereof in the subsequent trust network diagram construction and maximum weight spanning tree algorithm.

It should be noted that SyncDrop-E algorithm is dedicated to handling data collision problem between multi-modal sensors. The working mechanism of the algorithm comprises the following key links that SyncDrop-E algorithm selects and corrects the deviation based on the double mechanism of weighted data, and the most reliable sensor data is screened by comprehensively considering the reliability weight and the data deviation degree of the sensor. The core calculation logic of the algorithm multiplies the measurement data of the conflict sensor with the corresponding weight value to obtain weighted data, and then performs final screening by combining with the deviation correction factor. Firstly, identifying the sensor data pairs with conflicts, extracting the weight value corresponding to each conflict sensor, and forming a conflict data input set containing sensor identification, measurement data and the weight value. Then, weighting processing is performed on the measurement data of each collision sensor, and the measurement data of each sensor is multiplied by its corresponding weight value to obtain a weighted sensor data value capable of reflecting the reliability of the sensor. The algorithm calculates the ratio of the absolute difference between conflicting sensor data to the power device deviation threshold and then subtracts the ratio from 1 to obtain the deviation correction factor. The design purpose of this mechanism is to reduce the confidence of the sensor data when it is highly diverse, and to maintain a high confidence when the data is less diverse. And carrying out product operation on the maximum value in the weighted sensor data value and the deviation correction factor, and taking the weight advantage of the sensor and the consistency degree between the data into consideration in the mode. And finally, the most reliable sensor data subjected to deviation correction is selected as a conflict resolution result, and the accuracy and the reliability of the fusion data are ensured. For complex conflict scenarios involving more than three sensors, the SyncDrop-E algorithm can be combined with graph theory methods to build a sensor trust network and identify the most trusted set of sensors through a maximum weight spanning tree algorithm. For periodic collisions, the algorithm also supports time window analysis to determine abnormal sensors by data trend for consecutive multiple sampling periods.

In a specific embodiment, the process of performing step S4 may specifically include the following steps:

And carrying out abnormality judgment and feature classification processing according to the abnormality index value to obtain an abnormality feature set.

Specifically, the thermal characteristics, the electromagnetic characteristics and the vibration characteristics in the multi-mode fusion data set are respectively subjected to numerical coding and vector splicing according to specific mapping rules, multiple physical signal types are uniformly converted into a standardized three-dimensional characteristic space through a preset characteristic mapping matrix, and characteristic components of an electric subspace, a thermal subspace and a mechanical subspace are respectively formed, and the three components are combined to form the multi-mode characteristic space vector. The vector is a data structure with definite dimension interpretation and physical meaning in mathematical form, and each dimension represents the measurement parameter value of a certain operation dimension of the equipment, such as voltage harmonic ratio, surface maximum temperature, vibration dominant frequency energy density and the like, and has traceability and real-time updating capability. The corresponding operation state boundary model is determined according to the type of the current equipment, and the process is matched depending on equipment asset tags and technical parameter files thereof. And according to equipment classification, such as a transformer, a high-voltage circuit breaker, a bus connector or a cable branch box, extracting the upper and lower limits of key physical quantities in the operation process from a standard parameter database, and constructing a minimum boundary vector and a maximum boundary vector. The two vectors respectively represent the minimum value and the maximum value interval allowed by the equipment in a normal state and are reference boundaries for abnormality judgment. For example, for a transformer, the surface temperature peak value should be lower than 90 ℃ and higher than 40 ℃, the electromagnetic leakage should be within a safety threshold, and the vibration spectrum should not have structural excitation frequency, while for a circuit breaker, the electromagnetic peak value is more sensitive to short-time arc, and the boundary setting standard is correspondingly transferred to dimensions such as high-frequency voltage distortion index, vibration recovery time after breakdown and the like. The system enters an abnormal calculation core stage after acquiring the feature space vector and the boundary set, namely, the deviation degree is quantified by calculating the minimum distance between the current feature vector and the normal boundary set. The distance function used is a weighted euclidean distance or manhattan distance, and mahalanobis distance is employed in some scenarios to enhance the adaptability to inter-dimensional covariance. The system performs item-by-item boundary comparison on the parameter value of each dimension, if the current value is between the maximum boundary and the minimum boundary, the dimension is regarded as normal, the offset is set to zero, if the current value exceeds the upper limit and the lower limit, the excess is used as an offset index, and the total offset distance is built by weighting and accumulating all the offset values. And carrying out fusion processing on the distance and the equipment type sensitivity coefficient to form a final abnormality index value. The sensitivity coefficient here acts as a regulator, and the tolerance of different devices to the same anomaly type is different, for example, the cable support system can tolerate a certain structural vibration and the transformer is more sensitive to continuous temperature rise, so that different characteristic dimensions are endowed with differential sensitivity according to the risk level and structural vulnerability of the devices, the numerical expression of the global anomaly degree A is formed, the numerical expression is limited in the [0,1] interval, and the closer the numerical value is to 1, the more serious the deviation is, and the more unstable the state is. After obtaining the abnormality index value, carrying out logic judgment on the abnormality index value and a threshold value, and carrying out abnormality judgment and feature classification processing by combining a rule base. setting multi-level exception response logic, e.g., an anomaly degree a below 0.3 is considered normal, 0.3 to 0.6 is considered mild exception, 0.6 to 0.8 is considered significant deviation, and greater than 0.8 triggers a fault alert mechanism. Meanwhile, the dimension of the feature space, in which the boundary crossing specifically occurs, is marked as 'thermal anomaly' if the thermal parameters such as the temperature peak value and the average value are out of range, as 'electric anomaly' if the frequency spectrum index is out of range, as 'mechanical anomaly' if the high-amplitude impact signal protrudes in the vibrator space, as 'composite anomaly' if the same-time multidimensional parameters are out of range, so that an anomaly feature set is formed.

In a specific embodiment, the process of performing step S5 may specifically include the following steps:

Specifically, an abnormal feature set is taken as input, and the set contains multiple information such as types, abnormal constant values, severity levels, time stamps, source sensors and the like of abnormal dimensions in the thermal, electric and mechanical multi-modal features of each device. On the basis, the probability that the equipment is likely to be in fault is quantitatively predicted by constructing a fault probability density function model. The core calculation method of the fault probability density is based on sigmoid activation function as a mathematical basis, various abnormal characteristic indexes are standardized and then are input as characteristic functions, sensitivity weights of the characteristics in historical fault cases are combined, multi-factor fault expressions are constructed in a proportional superposition mode, and a fault probability reasoning model with strong adaptability and dynamic feedback capacity is formed. The model not only considers the single-dimensional abnormal strength, but also integrates the structural combination influence of the abnormality, so that the model can output the comprehensive probability density value of the equipment failure under the nonlinear condition, and the numerical value is limited between 0 and 1, wherein the closer to 1 is the closer to the failure threshold value the current state of the equipment is. In order to realize efficient scheduling and resource priority allocation of the inspection tasks, importance scores are calculated on the power equipment based on the fault probability density values so as to form equipment evaluation parameter sets. In the process, a structural evaluation model of the power system is introduced, the model carries out numerical quantification on the network topology position (such as a trunk node or an edge node), the load level (such as a high-load circuit breaker or a standby feeder) and the substitution redundancy rate (i.e. the substitution after the equipment fails) of the equipment in the power grid through a hierarchical analysis method or a weighted index model, and the weighted equipment score is calculated by combining the failure probability density value. in the scoring function, the network position index is given the maximum weight to ensure the stability of the core transmission path, the load index reflects the operation pressure, and the redundancy rate is reversely weighted to improve the overall toughness of the system. The system-generated device evaluation parameter set structurally identifies the current failure risk level, importance weight, scope of influence, and priority being patrol for each device. And inputting EIPO the fault probability density value, the equipment evaluation parameter set and the physical distance between the equipment into a routing inspection path optimization algorithm model, and executing optimal path calculation under the multi-objective constraint condition. The EIPO algorithm takes three objective functions as cores, namely minimizing the total routing inspection path length, maximizing the coverage density of high-fault probability equipment and maximizing the access priority of key equipment, wherein an optimization objective function consists of a distance cost function, a fault risk function and an equipment importance function, and adopts a weight coefficient to flexibly regulate and control so as to adapt to different application scenes, such as taking path efficiency as a main part in routine routing inspection, taking risk response as a main part in emergency inspection and taking equipment grade as a main part in a key guarantee scene. The system adopts an improved genetic algorithm to realize path search, combines a mechanism of maintaining cross, 2-opt local mutation and dynamically adjusting fitness function by encoding equipment access sequences into chromosomes, iteratively evolves candidate path sets in each generation, and terminates searching when the maximum algebra or continuous multi-generation optimal path is not improved, and outputs a global optimal solution. The finally formed target inspection path sequence can meet the dual requirements of coverage rate and priority, and efficiency optimization under the condition of limited resources is realized by adjusting path density. And introducing historical inspection data and equipment states related in the current inspection path after path optimization is completed, and constructing an equipment state joint prediction model. The model is based on the fusion design of a graph convolution network and a time sequence prediction model, the system takes equipment in a target routing inspection path as graph nodes, takes physical connection, electrical interconnection and a historical fault propagation path between the equipment as an edge structure of a graph, and takes a multi-mode feature sequence in a historical routing inspection record as an attribute feature of the graph nodes. In the graph neural network, the structure dependency relationship and the state co-transformation mode among the excavating devices are operated through multi-layer graph convolution, and the higher-order influence of the key devices is highlighted by combining an attention mechanism, so that the state predicted value and the state change trend at the next moment are output on each node. The predicted state vector is input into an extended heat conduction type diffusion model, the process of latent faults propagating from high-risk nodes to adjacent equipment is simulated, and the risk space distribution in a plurality of time steps in the future is deduced. the result output by the joint prediction module is visually presented in the form of a fault risk distribution diagram and a propagation path diagram, wherein each device is marked with the predicted state level, the risk change trend and the fault diffusion probability.

In a specific embodiment, the performing step inputs EIPO the fault probability density value, the equipment evaluation parameter set and the physical distance between the equipment into an inspection optimization algorithm to perform multi-objective constrained inspection track optimization, and the process of obtaining the objective inspection path sequence may specifically include the following steps:

inputting EIPO a fault probability density value, a device evaluation parameter set and a physical distance between devices into a patrol optimization algorithm to construct a comprehensive optimization objective function comprising distance cost, fault risk and device importance;

Carrying out iterative solution based on the comprehensive optimization objective function and the optimization parameter configuration set to obtain a candidate patrol path solution set;

Specifically, the device fault probability density value is used as a main index for measuring the current potential failure risk of each device, the structural importance score in the device evaluation parameter set is used as a key basis for measuring the network influence and the routing inspection priority of the device, the physical distance matrix is used as a space basis for measuring the path cost, the three types of data together form an input variable set of EIPO algorithm, and the system constructs a multi-objective optimization function comprising three core constraint targets of distance cost, fault risk and device importance in an algorithm model according to the main index. The objective function structurally consists of three parts, wherein the first part is a path cost item which represents the physical total length of the inspection path and is an accumulated value of actual moving distances among all equipment nodes, and the item is modeled by Euclidean distance or terrain correction distance so as to minimize energy consumption and inspection time; the second part is a fault risk item, the covering capacity of the system to high-risk equipment on the inspection path is expressed by carrying out weighted summation on fault probability density values of all equipment in the path, the weight is dynamically regulated according to the importance of the equipment and the fault level so as to ensure the priority of risk response, and the third part is an equipment importance item, and the core of the third part is the equipment importance item for ensuring the priority of risk response to key equipment (such as main transformer, Bus node, high load switch) to give higher weight according to the importance factors marked in the equipment evaluation parameter set, so that the system can preferentially ensure the coverage frequency and timeliness of the key nodes in path optimization. The objective function is defined as a weighted combination form of the three sub-functions, the comprehensive objective is to minimize the path cost and maximize the risk coverage and importance response capability, and the system adjusts the proportion of each item according to different application scenes. In order to ensure that the optimization model has good adaptability under different inspection task scenes, an adaptive parameter configuration mechanism is introduced, namely three weight coefficients in an objective function are dynamically set according to the type, the emergency degree and the power grid running state of the inspection task. In the conventional daily inspection, the system sets a distance weight coefficient higher (such as alpha=0.6) so as to ensure path efficiency and resource saving, and resets the fault risk weight and importance weight as secondary, if the system is in a fault tracking state, the system improves the fault probability coefficient (such as beta=0.6) to preferentially cover high-risk equipment, and in the critical guarantee period such as peak-welcome summer, the importance weight coefficient (such as gamma=0.6) is improved to focus on the inspection task of the high-criticality equipment. And a plurality of predefined scene weight templates are built in the system, and meanwhile, weight fine adjustment is allowed to be carried out through expert parameters or historical task data, so that a complete set of optimized parameter configuration sets is generated and used as a scheduling basis of the path solving module. After the objective function construction and the weight parameter loading are completed, the system enters an iterative path solving stage, a EIPO algorithm adopts an improved genetic optimization mechanism to encode the equipment access sequence into an individual chromosome, an initial population consists of a plurality of random or semi-heuristic paths, the path individual is subjected to multi-round evolution through sequence maintenance intersection, inverted sequence variation (such as 2-opt or 3-opt) and a local climbing strategy in each generation, the fitness function is calculated according to the comprehensive objective function, and the total path cost, the risk coverage quality and the important node hit rate are subjected to weighted evaluation. In the population evolution process, a multipoint selection and optimal retention strategy is introduced to avoid sinking into local minima, and the crossover and variation strength is controlled through a dynamic self-adaptive operator, so that the algorithm maintains solution stability while ensuring exploration capability. The new path set generated by each generation forms a candidate patrol path solution set, and the system continuously monitors the path cost change trend and the risk coverage level in the solution set so as to evaluate the search progress and the solution space diversity. And when the candidate solution sets are accumulated to the appointed number or the algorithm is continuous for a plurality of generations, and no better path appears, the system enters a convergence judgment and optimal solution selection stage. The system performs scoring sorting on all paths in the solution set based on the comprehensive objective function, checks coverage breadth, risk concentration and path efficiency stability of the paths on key performance indexes, screens out a path with the optimal score as a target inspection path sequence of a current task, and attaches a task execution time window and a device state mark of each path to the sequence to provide a data interface for a follow-up path execution and task scheduling module. meanwhile, the path sequence is synchronously recorded into a history track library so as to be used for subsequent task prediction and model self-adaptive training.

It should be noted that EIPO inspection optimization algorithm is a multi-objective optimization algorithm specifically designed for the power equipment inspection track optimization, and can comprehensively consider a plurality of factors such as equipment state, importance, spatial distribution and the like to realize intelligent planning of the inspection path. The EIPO algorithm constructs a comprehensive optimization objective function that contains distance cost, fault risk, and device importance. The algorithm forms a unified optimization target by carrying out weighted summation on the distance function, the fault probability density function and the equipment importance scoring function. And automatically adjusting the optimization parameters according to different inspection scene types. For a conventional inspection scene, the algorithm pays more attention to path efficiency, sets higher distance weight, for a fault tracking scene, the algorithm gives priority to fault probability and improves fault risk weight, and for a key equipment inspection scene, the algorithm highlights equipment importance weight. The self-adaptive mechanism ensures the optimization effect of the algorithm under different application scenes. And (3) grading the importance of the power equipment by adopting an analytic hierarchy process, and comprehensively considering factors such as network position coefficient, load importance, standby rate and the like of the equipment in the power grid. And EIPO, an improved genetic algorithm solving mechanism is adopted to carry out optimization solving by adopting the improved genetic algorithm, and the equipment access sequence is used as a chromosome coding mode, so that the patrol path can be represented in a visual way. The algorithm adopts sequence maintenance crossover operation to ensure the validity of the chromosome, and uses 2-opt local search as mutation operation to improve the quality of the solution. A double termination condition is set, including no improvement in the continuous 50-generation optimal solution or maximum algebra 300. The design ensures the convergence of the algorithm and avoids the waste of calculation resources caused by excessive iteration. The algorithm continuously monitors the improvement condition of population fitness in the iterative process, and when the quality of the solution is stable, the search is terminated in time. The method can dynamically adjust the inspection track according to the real-time state of the power equipment, and when the equipment abnormality or the fault probability change is detected, the algorithm can recalculate the optimization objective function and update the inspection path. The real-time optimization capability enables the inspection robot to quickly respond to the state change of the equipment, and improves the pertinence and the efficiency of inspection. And (3) simultaneously processing various constraint conditions in the optimization process, including equipment physical position constraint, inspection time window constraint, robot energy consumption constraint and the like. By introducing punishment items into the objective function or adopting constraint processing technology, the generated inspection track is ensured to reach an optimal effect on the premise of meeting various practical constraints.

In a specific embodiment, the performing step performs device state joint prediction on the target inspection path sequence and the historical device inspection data, and the process of obtaining the device failure prediction result may specifically include the following steps:

Constructing an equipment association graph structure comprising equipment nodes, association edges and association strength weights based on the target routing inspection path sequence and historical equipment routing inspection data;

Specifically, a graph structure model is built based on a target routing inspection path sequence and a historical equipment operation record, each inspected equipment is used as a node entity in the graph, node attributes comprise the latest state characteristic vector (such as temperature, electric field intensity, vibration spectrum characteristics and the like) of the current equipment, abnormal indexes, historical fault frequency, state change tracks of past routing inspection periods and the like, connection relations among nodes, namely edge structures of the graph, establish edge weights according to factors such as space physical connection, electric communication relations, historical resonance or co-fault records and the like among the equipment, the numerical value of the edge weights is the 'associated intensity weight', the intensity degree of state transfer or influence between the two equipment is reflected, and the stronger the association is, the state of one equipment is more likely to influence the other equipment. And inputting the equipment association graph structure and the equipment state sequence acquired in the current period into a state joint prediction model, wherein the model is based on graph convolution network design, and characteristic aggregation and node embedding updating are performed on the equipment graph through multi-layer graph convolution operation, so that joint modeling of the equipment state in two dimensions of space and time is realized. In the calculation process of each layer of graph convolution, the state of a node not only depends on the historical state of the node, but also converges the state characteristics of adjacent nodes, and the higher-order dependence on a space structure is fused, meanwhile, the time evolution process of the state of the node is modeled by introducing a time sequence convolution or gating circulating unit (such as GRU or LSTM) structure of the historical equipment state, so that the state prediction result finally output by each node not only comprises the time trend of the node, but also is influenced by the state change of topological upstream and downstream equipment, and the next cycle prediction state under the current state drive, namely a prediction state data set, is formed. The data set records the state index values of all the devices at the next moment and the growing trend of the state index values in a vector form. Based on the predicted state dataset and the device association graph structure, a fault propagation modeling stage is entered. The heat conduction equation is introduced at this stage as a mathematical basis for fault diffusion simulation, the equipment state is regarded as an initial condition of a heat field, and a time evolution diffusion model is constructed on a graph structure. The system regards the equipment with high risk or obvious abnormal indexes in the prediction state as a heat source node, initializes the state of the equipment to be high value, and the other nodes to be low value or normal value, and then performs the numerical solution of the heat conduction equation on the whole graph network. The model gradually outputs the evolution trend of the state values of the equipment groups and the propagation paths of fault signals in the graph structure in a plurality of future periods through multi-time step deduction to form fault probability distribution data, wherein the data takes nodes as indexes, and the probability value of faults of each equipment at the time of t+1, t+2 and t+3 in the future is recorded. Based on the fault probability distribution data, fault risk assessment and visual processing operation are executed, a risk assessment module establishes a comprehensive risk grade grading model by setting risk grade grading standards, such as that the probability <0.3 is normal, the probability 0.3-0.6 is mild, the probability 0.6-0.8 is moderate and the probability >0.8 is high, and combining key node density in a fault propagation path, importance grading of equipment and path connectivity, and carries out final risk grade labeling on each equipment, and a visual module converts an equipment association diagram structure into an interactive thermodynamic diagram or a geographic information superposition diagram, different colors represent different risk grades, edge colors or thickness represent propagation probability intensity, and shows how fault heat spreads from a core fault point to peripheral nodes in an animation mode, so that operation and maintenance personnel can intuitively understand a potential fault evolution chain in space.

The above describes a method for multi-mode sensor fusion inspection in the embodiment of the present invention, and the following describes a multi-mode sensor fusion inspection system in the embodiment of the present invention, referring to fig. 2, an embodiment of the multi-mode sensor fusion inspection system in the embodiment of the present invention includes:

Through the cooperative cooperation of the components, the reliability index of each sensor can be dynamically quantized by constructing a sensor self-adaptive weight evaluation model based on thermal stability, precision history and signal strength, the problem that the traditional fixed weight method cannot cope with sensor aging, environmental interference and equipment state change is effectively solved, and the accuracy and the robustness of multi-mode data fusion are ensured. By combining the dynamic confidence coefficient matrix and the characteristic correlation function, the data conflict among the multi-mode sensors can be accurately identified, intelligent screening and fusion of conflict data are realized, and compared with a traditional simple weighted average or majority voting mechanism, the method has stronger conflict processing capability and data fusion precision. By establishing the three-dimensional feature space and the equipment state boundary, millisecond-level conflict data real-time processing is realized, the computational bottleneck of the traditional deep learning method on the embedded platform is avoided, and the real-time requirement of power equipment inspection is met. By constructing the power equipment fault probability density function as a core constraint, multi-objective optimized routing inspection path planning is realized, and equipment states, importance and spatial distribution are comprehensively considered. The method can effectively capture the spatial association relation between the power equipment, solves the limitation that the traditional time sequence model cannot process the spatial dependence by establishing the equipment association diagram and analyzing the physical connection, the electrical influence and the fault propagation relation, and realizes more accurate equipment state joint prediction. The propagation process of faults in the equipment network is simulated through a heat conduction equation, potential fault propagation paths and chain reaction risks can be identified, algorithm parameters and optimization strategies can be adaptively adjusted for different power equipment types and inspection scenes by the system, the algorithm parameters comprise sensor weight coefficients, deviation thresholds, inspection weight parameters and the like, and the universality and the effectiveness under different application environments are ensured.

It will be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described systems, systems and units may refer to the corresponding processes in the foregoing method embodiments, which are not repeated herein.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied essentially or in part or all of the technical solution or in part in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. The storage medium includes a U disk, a removable hard disk, a read-only memory (ROM), a random access memory (random access memory, RAM), a magnetic disk, an optical disk, or other various media capable of storing program codes.

While the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that the foregoing embodiments may be modified or equivalents may be substituted for some of the features thereof, and that the modifications or substitutions do not depart from the spirit and scope of the embodiments of the invention.

Claims

1. The multi-mode sensor fusion inspection method is characterized by comprising the following steps of:

2. The method for inspecting a multi-modal sensor fusion according to claim 1, wherein the acquiring multi-modal raw data of the power device by the multi-modal sensor in the inspection robot and constructing the feature vector set includes:

3. The method for inspecting multi-modal sensor fusion according to claim 1, wherein the performing adaptive weight calculation on the multi-modal sensor according to the feature vector set to obtain a sensor weight set includes:

4. The method for multi-modal sensor fusion inspection according to claim 1, wherein the performing conflict recognition and resolution on the multi-modal raw data based on the sensor weight set to obtain a fusion data set includes:

5. The method for multi-modal sensor fusion inspection according to claim 4, wherein the inputting SyncDrop-E of the sensor weight set, the feature correlation function value, and the power device bias threshold value into the conflict resolution algorithm performs data screening to obtain an initial conflict resolution result, comprising:

6. The multi-modal sensor fusion inspection method according to claim 1, wherein the performing, based on the fusion data set, abnormal feature extraction on the power device to obtain an abnormal feature set includes:

7. The multi-mode sensor fusion inspection method according to claim 1, wherein the performing inspection track optimization based on the abnormal feature set to obtain a target inspection path sequence, and performing device state joint prediction in combination with historical device inspection data to obtain a device failure prediction result comprises:

8. The method for multi-modal sensor fusion inspection according to claim 7, wherein the performing multi-objective constrained inspection trace optimization on the fault probability density value, the equipment evaluation parameter set, and the inter-equipment physical distance input EIPO inspection optimization algorithm to obtain a target inspection path sequence includes:

9. The method for multi-mode sensor fusion inspection according to claim 7, wherein performing device state joint prediction on the target inspection path sequence and historical device inspection data to obtain a device failure prediction result comprises:

10. A multi-modal sensor fused inspection system for performing the multi-modal sensor fused inspection method of any one of claims 1-9, the multi-modal sensor fused inspection system comprising: