CN115130514A - Construction method and system for health index of engineering equipment - Google Patents

Abstract

The invention relates to a method and a system for constructing a health index of engineering equipment, and belongs to the field of reliability engineering. The method includes: acquiring test data of engineering equipment and calculating its time-frequency domain information; inputting the time-frequency domain information into a continuous depth belief network, and training it by means of a contrastive divergence algorithm; using the trained continuous depth belief network to extract the deep level Fault features; take the deep fault features as the input of the SOM neural network for unsupervised learning; by calculating the Euclidean distance between the state vector of the input layer of the trained SOM neural network and the weight vector of all units, the engineering equipment is constructed. health indicators. The method and system of the present invention use a continuous deep confidence network to extract the deep-level fault features contained in the time-frequency domain, and use the SOM neural network to construct the health index of the equipment on the premise of ensuring the original feature topology structure, thereby improving the accuracy of the construction of the health index.

Description

A method and system for constructing health indicators of engineering equipment

technical field

The invention relates to the technical field of reliability engineering, in particular to a method and system for constructing health indicators of engineering equipment.

Background technique

With the development of sensor monitoring technology and the improvement of electronic device technology, more and more test data can be collected under the condition that the equipment is less affected. will proliferate. In particular, in today's Internet of Things and Industry 4.0 era, the test data acquired by engineering equipment will present the characteristics of mass, nonlinearity and dynamics. When using the traditional method to determine the health index of the equipment, it is difficult to extract the deep-level fault features of the equipment, so the accuracy of the health index will be affected to a certain extent. In order to fully tap the hidden features contained in massive data, Hinton proposed an effective tool for analyzing massive data—deep learning, which has been highly praised by scholars and engineers since then. Deep learning has been successfully applied in many engineering fields, such as image processing, speech recognition, machine translation, etc. At the same time, the extraction of fault features and the construction of health indicators based on this technology have received a lot of attention, and a series of results have been achieved.

Due to the excellent characteristics and special structure of Restricted Boltzmann Machine (RBM), it can be used to solve various classification, dimensionality reduction, regression and other problems. As a bidirectional probabilistic graphical model, the restricted Boltzmann machine is a neural network with a two-layer structure, which is mainly composed of a visible layer and a hidden layer, which meets the requirements of no connection within the layer and full connection between layers. The advantage of restricted Boltzmann machine is that it can quantitatively represent the distribution of a set of inputs and outputs in the form of probability. It has certain advantages in processing binary data, but when faced with continuous data modeling, the model training effect is not impressive Satisfied, it presents a challenge to the fault feature extraction of engineering equipment with continuous test data. In addition, topology is another important factor that needs to be considered in health indicators, and it is difficult to ensure the topology of original features using deep learning alone. Once the topology is changed, more valuable information will be lost in the process of dimensionality reduction. Therefore, it is urgent to introduce other dimensionality reduction methods to facilitate the determination of health indicators on the premise of ensuring the topology structure.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method and system for constructing a health index of engineering equipment, adopting a continuous deep confidence network to extract the deep-level fault features contained in the time-frequency domain, and using the SOM neural network to construct the equipment's Health indicators to improve the accuracy of health indicator construction.

For achieving the above object, the present invention provides the following scheme:

A method for constructing health indicators of engineering equipment, comprising:

Obtain the test data of the engineering equipment and calculate the time-frequency domain information of the test data;

The time-frequency domain information is input into the continuous depth belief network, and the continuous depth belief network is trained by means of a contrastive divergence algorithm to generate a trained continuous depth belief network;

Use the trained continuous deep belief network to extract the deep fault features of the engineering equipment;

The deep-level fault feature is used as the input of the SOM neural network for unsupervised learning to generate a trained SOM neural network;

By calculating the Euclidean distance between the state vector of the input layer of the trained SOM neural network and the weight vectors of all units, the health index of the engineering equipment is constructed; the health index directly reflects the actual situation of the engineering equipment health status.

Optionally, obtaining the test data of the engineering equipment and calculating the time-frequency domain information of the test data specifically includes:

Obtain test data of engineering equipment;

Extract the time domain features of the test data; the time domain features include mean, root mean square, root square amplitude, mean absolute amplitude, skewness, kurtosis, variance, maximum value, minimum value, peak value, shape factor, Crest factor, impulse factor, margin factor, skewness factor, kurtosis factor;

Extract the frequency domain features of the test data; the frequency domain features include center frequency, root mean square frequency, and frequency variance;

The time-domain features and the frequency-domain features extracted form the time-frequency domain information of the engineering equipment.

Optionally, the time-frequency domain information is input into a continuous depth belief network, and the continuous depth belief network is trained by means of a contrastive divergence algorithm to generate a trained continuous depth belief network, specifically including:

constructing a continuous deep belief network formed by stacking a plurality of continuous restricted Boltzmann machines; the continuous restricted Boltzmann machine includes a visible layer, a hidden layer and a Gaussian noise unit;

Input the time-frequency domain information into the continuous depth belief network, update the connection weight and sigmoid function slope control parameters of the continuous depth belief network with the help of the contrast divergence algorithm, complete the continuous depth belief network parameter training, and generate a trained A continuous deep belief network for .

Optionally, the use of a trained continuous deep belief network to extract deep-level fault features of the engineering equipment specifically includes:

The time-frequency domain information X=[x ₁ , x ₂ , . . . , x _n ] of the test data is used as the input of the trained continuous deep belief network, and the output Y of the trained continuous deep belief network =[y ₁ , y ₂ ,...,y _m ] as the deep fault feature of the engineering equipment; where x _i (0≤i≤n) represents the feature sequence of the ith input neuron of the continuous deep belief network, n is the total number of input features; y _i (0≤i≤m) represents the feature sequence of the ith output neuron of the continuous deep belief network, and m is the total number of output neurons of the continuous deep belief network.

Optionally, the described deep-level fault feature is used as the input of the SOM neural network to carry out unsupervised learning, and the trained SOM neural network is generated, specifically including:

Taking a row of the deep fault feature Y=[y ₁ , y ₂ ,..., y _m ] as the input vector D of the SOM neural network, and updating the best matching unit and topological neighborhood according to the input vector D The weight vector of the SOM neural network is completed, and the unsupervised learning of the SOM neural network is completed, and the trained SOM neural network is generated.

A system for constructing health indicators of engineering equipment, comprising:

a time-frequency domain information extraction module, used for acquiring the test data of the engineering equipment and calculating the time-frequency domain information of the test data;

The continuous depth belief network training module is used to input the time-frequency domain information into the continuous depth belief network, train the continuous depth belief network by means of a contrastive divergence algorithm, and generate a trained continuous depth belief network;

The deep-level fault feature extraction module is used for extracting the deep-level fault features of the engineering equipment by using the trained continuous deep belief network;

The SOM neural network training module is used to perform unsupervised learning with the deep-level fault features as the input of the SOM neural network to generate a trained SOM neural network;

The engineering equipment health index building module is used to construct the health index of the engineering equipment by calculating the Euclidean distance between the state vector of the input layer of the trained SOM neural network and the weight vectors of all units; The indicators directly reflect the real health status of the engineering equipment.

Optionally, the time-frequency domain information extraction module specifically includes:

The test data acquisition unit is used to acquire the test data of the engineering equipment;

A time-domain feature extraction unit for extracting time-domain features of the test data; the time-domain features include mean, root mean square, root square amplitude, mean absolute amplitude, skewness, kurtosis, variance, maximum value, Minimum value, peak value, shape factor, crest factor, pulse factor, margin factor, skewness factor, kurtosis factor;

a frequency domain feature extraction unit, used for extracting the frequency domain feature of the test data; the frequency domain feature includes a center frequency, a root mean square frequency, and a frequency variance;

The time-frequency domain information construction unit is configured to form the time-frequency domain information of the engineering equipment from the extracted time-domain features and the frequency-domain features.

Optionally, the continuous deep belief network training module specifically includes:

A continuous deep belief network building unit for constructing a continuous deep belief network composed of multiple continuous restricted Boltzmann machines stacked; the continuous restricted Boltzmann machine includes a visible layer, a hidden layer and a Gaussian noise unit;

The continuous depth belief network training unit is used to input the time-frequency domain information into the continuous depth belief network, and update the connection weight and the sigmoid function slope control parameter of the continuous depth belief network by means of a contrastive divergence algorithm to complete the continuous depth belief network. Deep belief network parameter training to generate a trained continuous deep belief network.

Optionally, the deep-level fault feature extraction module specifically includes:

A deep-level fault feature extraction unit, configured to use the time-frequency domain information X=[x ₁ , x ₂ , . . . , x _n ] of the test data as the input of the trained continuous deep belief network, The output of a good continuous deep belief network Y=[y ₁ , y ₂ ,...,y _m ] is used as the deep fault feature of the engineering equipment; wherein, x _i (0≤i≤n) represents the first continuous deep belief network The feature sequence of i input neurons, n is the total number of input features; y _i (0≤i≤m) represents the feature sequence of the ith output neuron of the continuous deep belief network, m is the total number of output neurons of the continuous deep belief network .

Optionally, the SOM neural network training module specifically includes:

The SOM neural network training unit is configured to use a row of the deep fault features Y=[y ₁ , y ₂ ,..., y _m ] as the input vector D of the SOM neural network, and update according to the input vector D The weight vector of the best matching unit and topological neighborhood completes the unsupervised learning of the SOM neural network, and generates a trained SOM neural network.

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

The present invention provides a method and system for constructing health indicators of engineering equipment. The method includes: acquiring test data of engineering equipment and calculating time-frequency domain information of the test data; inputting the time-frequency domain information into continuous depth confidence network, train the continuous depth belief network by means of the contrast divergence algorithm to generate a trained continuous depth belief network; use the trained continuous depth belief network to extract the deep fault features of the engineering equipment; The hierarchical fault feature is used as the input of the SOM neural network for unsupervised learning to generate a trained SOM neural network; by calculating the Euclidean distance between the state vector of the input layer of the trained SOM neural network and the weight vector of all units, The health index of the engineering equipment is constructed; the health index directly reflects the real health state of the engineering equipment. The method and system of the present invention use a continuous deep confidence network to extract the deep fault features contained in the time-frequency domain, and use the SOM neural network to construct the health index of the equipment on the premise of ensuring the original feature topology structure, thereby improving the accuracy of the construction of the health index.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

Fig. 1 is the flow chart of a kind of construction method of engineering equipment health index of the present invention;

2 is a schematic diagram of a bearing experimental platform provided by an embodiment of the present invention;

3 is a schematic diagram of a typical horizontal vibration signal of a bearing provided in an embodiment of the present invention in a full life cycle;

FIG. 4 is a schematic diagram of the health index of the bearing under the first operating condition among the three operating conditions provided by the embodiment of the present invention;

FIG. 5 is a schematic diagram of the health index of the bearing under the second operating condition among the three operating conditions provided by the embodiment of the present invention;

FIG. 6 is a schematic diagram of the health index of the bearing under the third operating condition among the three operating conditions provided by the embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Aiming at engineering equipment under the background of big data, the present invention proposes a universal fault feature extraction and health index construction method and system. The continuous deep confidence network is used to extract the deep fault features contained in the time-frequency domain, and the SOM is used. The neural network constructs the health index of the device under the premise of ensuring the original feature topology structure, so as to improve the characteristics of the health index and improve the accuracy of the construction of the health index.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

FIG. 1 is a flowchart of a method for constructing a health index of engineering equipment according to the present invention. Referring to Fig. 1, a construction method of an engineering equipment health index of the present invention includes:

Step 1: Acquire test data of engineering equipment and calculate time-frequency domain information of the test data.

The present invention uses signal processing technology to analyze the collected mass test data, and determines the corresponding time domain and frequency domain characteristics. Typical time domain characteristics mainly include mean value, root mean square, root square amplitude, mean absolute amplitude and skewness. , kurtosis, variance, maximum value, minimum value, peak value, shape factor, crest factor, impulse factor, margin factor, skewness factor, kurtosis factor; frequency domain features mainly include center frequency, root mean square frequency, frequency variance Wait.

Therefore, the step 1 obtains the test data of the engineering equipment and calculates the time-frequency domain information of the test data, which specifically includes:

Step 1.1: Obtain test data of engineering equipment.

The engineering equipment includes mechanical equipment such as bearings, and electronic equipment such as chips, triodes, and MOS tubes.

Step 1.2: Extract the time domain features of the test data; the time domain features include mean, root mean square, root square amplitude, mean absolute amplitude, skewness, kurtosis, variance, maximum value, minimum value, peak value, Shape factor, crest factor, impulse factor, margin factor, skewness factor, kurtosis factor.

The calculation formula of each time domain feature is as follows:

max x _max = maxx(t) (8)

Minimum x _min = minx(t) (9)

Peak x _peak =max|x(t)| (10)

Among them, x _i represents the ith test data of the engineering equipment in a single sampling interval; N is the number of test data in a single sampling interval, and x(t) represents all the data sets of the engineering equipment in a single sampling interval.

Step 1.3: Extract the frequency domain features of the test data; the frequency domain features mainly include center frequency, root mean square frequency, frequency variance, and the like. In order to facilitate the understanding of the frequency domain characteristics, the calculation formulas are described in detail by taking the center frequency, root mean square frequency and frequency variance as examples:

Performing Fourier transform on the collected engineering equipment test data can extract the spectral information contained in the time domain signal, that is, the frequency value and the corresponding peak/amplitude value, and the frequency domain feature can be calculated based on the spectral information. Among them, f _i is the frequency value corresponding to the spectrum at the _i -th time, pi is the amplitude corresponding to the frequency domain at the i-th time, and f _c is the average frequency. Other frequency domain features can be determined according to the existing literature and in combination with the above frequency domain features.

Step 1.4: The time-frequency domain information of the engineering equipment is constituted by the extracted time-domain features and the frequency-domain features.

The time-frequency domain features of each monitoring moment are formed into a eigenvector, and for multiple monitoring moments, different eigenvectors form a feature matrix in time order, that is, the time-frequency domain information of the engineering equipment, as the continuous depth confidence network. enter.

Step 2: Input the time-frequency domain information into a continuous deep belief network, train the continuous deep belief network by means of a contrastive divergence algorithm, and generate a trained continuous deep belief network.

In the present invention, the extracted time-frequency domain information is input into a continuous deep belief network (Continuous Deep Belief Network, CDBN), the network model is trained by means of a contrastive divergence algorithm, and the deep-level fault features of the engineering equipment can be extracted after the training is completed.

In the step 2, the time-frequency domain information is input into the continuous depth belief network, and the continuous depth belief network is trained by means of the contrast divergence algorithm to generate a trained continuous depth belief network, which specifically includes:

Step 2.1: Construct a continuous deep belief network formed by stacking multiple continuous restricted Boltzmann machines; the continuous restricted Boltzmann machine includes a visible layer, a hidden layer and a Gaussian noise unit.

The continuous depth belief network constructed by the present invention is a neural network formed by stacking a plurality of Continuous Restricted Boltzmann Machines (CRBM). On the basis of the Mann machine, it is formed by adding a zero-mean Gaussian noise to the sampling Sigmoid unit, which mainly includes the visible layer, the hidden layer and the Gaussian noise unit. Let s _j denote the output state of neuron j with several neuron state sets {s _i } as input, then s _j can be expressed as:

Among them, W _ij is the connection weight, and N _j (0,1) represents Gaussian white noise with mean 0 and variance 1. σ is a constant term, which together with N _j (0,1) forms the noise input part, that is, n _j =σ·N _j (0,1), the corresponding probability density p(n _j ) is:

x_j为sigmoid函数作用之前的状态，满足

θ_L与θ_H分别表示φ_j(x_j)函数的下渐近线与上渐近线，c_j为一个控制sigmoid函数斜率的参数，称为sigmoid函数斜率控制参数。φ _j (x _j ) is a more general sigmoid function that satisfies

x _j is the state before the sigmoid function acts, satisfying

θ _L and θ _H represent the lower asymptote and the upper asymptote of the φ _j (x _j ) function, respectively, and c _j is a parameter that controls the slope of the sigmoid function, which is called the sigmoid function slope control parameter.

Step 2.2: Input the time-frequency domain information into the continuous depth belief network, update the connection weight of the continuous depth belief network and the sigmoid function slope control parameter by means of the contrast divergence algorithm, and complete the continuous depth belief network parameter training, Generate a trained continuous deep belief network.

{s _i } in step 2.1 represents each neuron in the input layer of the continuous restricted Boltzmann machine. For the convenience of describing the operation mechanism of the continuous restricted Boltzmann machine, it is uniformly represented by {s _i } , after the time-frequency domain features (that is, the time-frequency domain information) are input to the continuous deep confidence network, that is, the first continuous restricted Boltzmann machine input layer directly receives the time-frequency domain features, so here { s _i } needs to be replaced by time domain and frequency domain features, but for the following continuous restricted Boltzmann machine, {s _i } needs to be replaced by the output of the previous continuous restricted Boltzmann machine.

When the neurons in the visual layer input a specific state, the state of the neurons in the hidden layer can be obtained according to formula (20), and the state of the visual layer can be reconstructed based on the state of the neurons in the hidden layer. The state of the neurons in the hidden layer after the structure is constructed, and the analogy is carried out in this way. After infinite reconstructions, the training of the model parameters can be realized. However, this method also requires multiple Gibbs sampling (Gibbs sampling), which takes a long time and is not conducive to the real-time solution of the model. The invention adopts a minimization contrast divergence algorithm for model parameter training, which significantly reduces the amount of calculation and provides a convenient way for efficient training of model parameters. The continuous depth belief network weight and sigmoid function slope control parameters can be updated according to the following formula:

分别神经元i与神经元j一步采样状态值。对于单个连续受限玻尔兹曼机而言，当隐含层神经元总数小于可视层神经元总数时，能够有效降低输入状态的维数。ΔW_ij表示经过一步采样后W_ij的变化量，Δc_j表示经过一步采样后c_j的变化量，经过多步采样后，可实现参数更新。具体地，通过将ΔW_ij和Δc_j分别与上一次训练的参数W_ij和c_j相加来进行参数更新，从而完成连续深度置信网络参数训练，生成训练好的连续深度置信网络。In the formula, εw and _εc represent the network connection weight _{W ij} and the learning rate of the sigmoid function slope control parameter c _j , respectively. s _i s _j is the product of the current state of neuron i and neuron j.

The state values of neuron i and neuron j are sampled in one step, respectively. For a single continuous restricted Boltzmann machine, when the total number of neurons in the hidden layer is less than the total number of neurons in the visual layer, the dimension of the input state can be effectively reduced. ΔW _ij represents the change of W _ij after one-step sampling, Δc _j represents the change of c _j after one-step sampling, and parameter update can be realized after multi-step sampling. Specifically, the parameters are updated by adding ΔW _ij and Δc _j to the parameters W _ij and c _j of the previous training respectively, so as to complete the continuous deep belief network parameter training and generate a trained continuous deep belief network.

Step 3: Using the trained continuous deep belief network to extract the deep fault features of the engineering equipment.

Assuming the original characteristics of the test signal of the degraded device, that is, the extracted time-frequency domain information of the test data can be expressed as X=[x ₁ , x ₂ , . . . , x _n ], where x _i (0≤i≤n) Represents the feature sequence of the ith input neuron of the continuous deep belief network, n is the total number of original features, these original features are regarded as the input of the continuous deep belief network, and the above contrastive divergence algorithm is used to determine each continuous restricted glass layer by layer. The model parameters of the Boltzmann machine can be calculated based on the formula (20) to calculate the state of each layer of neurons in the continuous deep belief network, that is, the deep fault features reflecting the health status of the equipment can be extracted through multiple continuous restricted Boltzmann machines. . Therefore, the output of the continuous deep belief network can be regarded as the deep fault feature of the input data, which can be expressed as Y=[y ₁ , y ₂ ,...,y _m ], where y _i (0≤i≤m) represents the continuous The feature sequence of the ith output neuron of the deep belief network, m is the total number of output neurons of the continuous deep belief network.

The time-frequency domain information X=[x ₁ , x ₂ ,..., x _n ] of the test data contains the time domain and frequency domain features of all the data, i represents the feature sequence number, x _i (0≤i≤n)th The sequence values of i features are a sequence because they are sampled at multiple time points, and each feature directly corresponds to a neuron input to the continuous deep belief network. Each row in X represents a vector composed of time-domain and frequency-domain features at a certain time, and each column represents the value of a feature at different times, so after X is input into the continuous deep belief network, each row of X corresponds to The input neuron of the first continuous restricted Boltzmann machine of the continuous deep belief network, that is, corresponding to {s _i }, and for the following continuous restricted Boltzmann machine, {s _i } only represents The output of the continuous restricted Boltzmann machine of the previous layer.

Step 4: Use the deep-level fault features as the input of the SOM neural network to perform unsupervised learning to generate a trained SOM neural network.

In the present invention, the fault feature extracted by the continuous deep confidence network is used as the input of the SOM (Self-organizing map, self-organizing map) neural network to realize its unsupervised learning, and the Euclidean distance between the state vector and the weight vector of all units is calculated by calculating the Euclidean distance. , quantitatively characterize the health status of the equipment.

The SOM neural network constructed by the present invention mainly includes two parts: an input layer and an output layer. The input layer mainly receives actual medium and high-dimensional nonlinear data, also known as the competition layer; the output layer is composed of regular node grids. The weight vector between the lth neuron of the output layer and the input neuron can be expressed as W _l =[W _l1 ,W _l2 ,...,W _lm ], where m is the dimension of the input vector of the SOM neural network. For a row of the deep fault feature Y extracted in step 3, that is, the deep feature from the continuous deep belief network at a specific moment can be regarded as the input vector of the SOM neural network, which can be denoted as D. The weight vector of the best matching unit and the topological neighborhood is updated according to the input vector D, and the unsupervised learning of the SOM neural network is completed, thereby generating a trained SOM neural network.

The learning process is mainly to calculate the Euclidean distance between the input vector and the weight vector to determine the neuron closest to the input vector in each step of training. Generally, the closest neuron is called the Best Matching Unit (BMU). Based on the obtained BMU, the weight vector of the BMU and the topological neighborhood can be adjusted and updated. The update result can be expressed as:

表示以BMU为中心的拓扑邻域核函数，W_BMU为BMU的权重向量。In the formula, p is the index of the iteration step, W _l (p) and W _l (p+1) are the weight vectors of p steps and p+1 steps respectively, D is the input vector, ε(p) is the learning rate,

Represents the topological neighborhood kernel function centered on the BMU, and W _BMU is the weight vector of the BMU.

Step 5: Construct the health index of the engineering equipment by calculating the Euclidean distance between the state vector of the input layer of the trained SOM neural network and the weight vector of all units.

In the process of health assessment, the SOM structure is trained by using the health state data. After the training is completed, the state vector of the input layer is compared with the weight vector of all units to evaluate the health state of the equipment. Specifically, the present invention selects the minimum quantization error as the health index, that is, by calculating the Euclidean distance between the state vector and the weight vectors of all units, quantitatively characterizes the health state of the device, that is:

Z=min||DW _l || (25)

Since the input vectors at different times correspond to different health indicators, a series of health indicator sequences Z=[Z(t ₁ ), Z(t ₂ ), ..., Z(t _N )] can be obtained by using formula (25), where Z(t N )] (t _i ) (1≤i≤N) represents the health index at time t _i . It should be noted that the SOM neural network can convert multi-dimensional data into one-dimensional health indicators, which directly reflect the real health status of the equipment. Based on this health index, fault diagnosis and life prediction can be performed.

The health index sequence Z=[Z(t ₁ ), Z(t ₂ ),..., Z(t _N )] extracted by the present invention is a quantity that changes with time. The equipment is showing a trend of degradation, indicating that the health status is getting worse and worse. The health index is a comprehensive index extracted by integrating all the time-domain and frequency-domain features of the device with the continuous deep belief network and the SOM neural network, so the original features can be abstractly expressed. The following specific examples of bearings may reflect this process.

The effectiveness and superiority of the method are verified by using the published bearing experimental data below. The bearing experimental data set comes from the joint laboratory of mechanical equipment health monitoring jointly established by the Institute of Design Science and Basic Components of Xi'an Jiaotong University and Changxing Shengyang Technology Co., Ltd. The bearing experimental platform is mainly composed of an AC motor, a motor speed controller, a support shaft, two support bearings and a hydraulic loading device, as shown in Figure 2. The sensing devices of the experimental bench are two accelerometers vertically installed on the outer ring of the experimental bearing. The main function is to collect vibration signals in the horizontal and vertical directions. The single sampling frequency is 25.6kHz, the sampling duration is 1.28s, and the interval between two adjacent samplings is 1min.

The accelerated degradation test of the bearing was carried out under three experimental conditions, and 5 sets of bearings were tested under each experimental condition. Therefore, the data set mainly contains the whole life cycle vibration test data of 15 sets of bearings. Table 1 lists the operating conditions of the experimental bearings and Lifetime specific information.

Table 1 Specific information on the operating conditions and life of the experimental bearing

Since the load of the experiment is applied in the horizontal direction, the accelerometer installed in the horizontal direction can obtain more bearing degradation information, so the horizontal vibration signal is selected here to adaptively predict the RUL of the tested bearing. Figure 3 depicts a typical horizontal vibration signal over the life of a bearing. It can be seen intuitively from Figure 3 that the health state of the bearing can usually be divided into two stages: the stable operation stage and the degradation stage. In the stable operation stage, the vibration signal of the bearing fluctuates up and down in a small range around the 0 value, and there is no degradation trend, so the degradation of the bearing at this stage can be ignored; however, in the degradation stage, the amplitude of the vibration signal changes with time. There is a clear upward trend, which means that extremely rich bearing degradation information can be obtained at this stage.

In order to make full use of the feature information in the vibration signal, the 16 time-domain features and 13 frequency-domain features described in step 1.2 (including the center frequency, root mean square frequency, frequency variance and the remaining 10 parameters described in step 1.3) are selected here. Other frequency domain features well-known in the field), and input the value to the CDBN containing the 2-layer CRBM structure to extract the deep fault features contained in the vibration signal. The specific structures and initial parameters of CDBN and SOM neural networks are determined by cross-validation, as listed in Table 2 and Table 3.

Table 2 The specific structure and initial parameters of CDBN

Table 3 Specific structure and initial parameters of SOM

structure topological form number of initial neighborhoods set value 23*23 Hextop 3

After extracting the hidden deep fault features reflecting the bearing health status based on CDBN, the SOM neural network can be used to determine the bearing health indicators under three operating conditions, as shown in Figure 4, Figure 5, and Figure 6, respectively. Thus, the 29 features will be reduced to one-dimensional features. Figure 4 shows that the trend of the health index of the bearing is consistent with the amplitude trend of the vibration signal, which confirms the two-stage health state of the bearing to a certain extent. Therefore, the method described in the present invention is reasonable and effective in extracting deep fault features of bearings and determining health indicators.

In order to further reflect the superiority of the method of integrating the continuous depth belief network and the SOM neural network proposed by the present invention, the traditional dimensionality reduction method is used as a comparison method to determine the health indicators, such as PCA (Principle Component Analysis, principal component analysis), SOM neural network. Networks and Continuous Deep Belief Networks (CDBN). At the same time, the health indicator trend factor is selected to evaluate the quality of the health indicator. In general, the closer the trend factor is to 1, the stronger the trend. Table 4 lists the trend factor results of each bearing health index under the proposed method and the three comparison methods.

Table 4 Comparison of health index performance under four methods

From the data in Table 4, it can be found that the trend factor of the health index calculated by the method of the present invention is greater than the trend factor of the health index under the other three methods, which means that the health index determined by the method of the present invention is better than the three traditional dimensionality reduction methods. In addition, the trend factors of most bearing health indicators determined based on the method of the present invention are all greater than 0.5, which provides technical support for the degradation modeling of health indicators and RUL (Remaining Useful Life, remaining useful life) prediction.

The method for constructing a new type of engineering equipment health index proposed by the present invention can extract the deep-level fault features contained in the time-frequency domain for engineering equipment under the background of big data, and construct the equipment's health indicators on the premise of ensuring the original feature topology. health indicators. Compared with the prior art, the method of the invention fully extracts the deep-level fault features contained in the time-frequency domain information, and at the same time ensures the topology structure of the original features in the process of dimensionality reduction, further develops the fault feature extraction method, and enriches the equipment. Health indicators build a theoretical system, which effectively improves the performance characteristics of health indicators.

Based on the construction method of engineering equipment health indicators provided by the present invention, the present invention also provides a construction system for engineering equipment health indicators, the system comprising:

a time-frequency domain information extraction module, used for acquiring the test data of the engineering equipment and calculating the time-frequency domain information of the test data;

The continuous depth belief network training module is used to input the time-frequency domain information into the continuous depth belief network, train the continuous depth belief network by means of a contrastive divergence algorithm, and generate a trained continuous depth belief network;

The deep-level fault feature extraction module is used for extracting the deep-level fault features of the engineering equipment by using the trained continuous deep belief network;

The SOM neural network training module is used to perform unsupervised learning with the deep-level fault features as the input of the SOM neural network to generate a trained SOM neural network;

The engineering equipment health index building module is used to construct the health index of the engineering equipment by calculating the Euclidean distance between the state vector of the input layer of the trained SOM neural network and the weight vectors of all units; The indicators directly reflect the real health status of the engineering equipment.

The time-frequency domain information extraction module specifically includes:

The test data acquisition unit is used to acquire the test data of the engineering equipment;

A time-domain feature extraction unit for extracting time-domain features of the test data; the time-domain features include mean, root mean square, root square amplitude, mean absolute amplitude, skewness, kurtosis, variance, maximum value, Minimum value, peak value, shape factor, crest factor, pulse factor, margin factor, skewness factor, kurtosis factor;

a frequency domain feature extraction unit, used for extracting the frequency domain feature of the test data; the frequency domain feature includes a center frequency, a root mean square frequency, and a frequency variance;

The time-frequency domain information construction unit is configured to form the time-frequency domain information of the engineering equipment from the extracted time-domain features and the frequency-domain features.

The continuous deep belief network training module specifically includes:

A continuous deep belief network building unit for constructing a continuous deep belief network composed of multiple continuous restricted Boltzmann machines stacked; the continuous restricted Boltzmann machine includes a visible layer, a hidden layer and a Gaussian noise unit;

The continuous depth belief network training unit is used to input the time-frequency domain information into the continuous depth belief network, and update the connection weight and the sigmoid function slope control parameter of the continuous depth belief network by means of a contrastive divergence algorithm to complete the continuous depth belief network. Deep belief network parameter training to generate a trained continuous deep belief network.

The deep-level fault feature extraction module specifically includes:

A deep-level fault feature extraction unit, configured to use the time-frequency domain information X=[x ₁ , x ₂ , . . . , x _n ] of the test data as the input of the trained continuous deep belief network, The output of a good continuous deep belief network Y=[y ₁ , y ₂ ,...,y _m ] is used as the deep fault feature of the engineering equipment; wherein, x _i (0≤i≤n) represents the first continuous deep belief network The feature sequence of i input neurons, n is the total number of input features; y _i (0≤i≤m) represents the feature sequence of the ith output neuron of the continuous deep belief network, m is the total number of output neurons of the continuous deep belief network .

The SOM neural network training module specifically includes:

The SOM neural network training unit is configured to use a row of the deep fault features Y=[y ₁ , y ₂ ,..., y _m ] as the input vector D of the SOM neural network, and update according to the input vector D The weight vector of the best matching unit and topological neighborhood completes the unsupervised learning of the SOM neural network, and generates a trained SOM neural network.

The invention provides a new method and system for constructing health indicators of engineering equipment, which fully extracts the deep-level fault features contained in the time-frequency domain, and at the same time ensures the topology structure of the original features in the process of dimensionality reduction. First, the test data of the engineering equipment is collected and the corresponding time-frequency domain information is calculated; then, the time-frequency domain information is input into the continuous deep confidence network, and the network model is trained with the help of the contrastive divergence algorithm. The deep fault feature of engineering equipment is used as the input of SOM neural network to realize its unsupervised learning. By calculating the Euclidean distance between the state vector and the weight vector of all units, the health state of the equipment is quantitatively described. The method of the invention further develops the fault feature extraction method, enriches the theoretical system of equipment health index construction, effectively improves the performance characteristics of the health index, helps subsequent fault diagnosis and life prediction, and has broad application space.

For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

In this paper, specific examples are used to illustrate the principles and implementations of the present invention. The descriptions of the above embodiments are only used to help understand the methods and core ideas of the present invention; meanwhile, for those skilled in the art, according to the present invention There will be changes in the specific implementation and application scope. In conclusion, the contents of this specification should not be construed as limiting the present invention.

Claims (10)

1. A construction method of engineering equipment health indexes is characterized by comprising the following steps:

acquiring test data of engineering equipment and calculating time-frequency domain information of the test data;

inputting the time-frequency domain information into a continuous depth confidence network, training the continuous depth confidence network by means of a contrast divergence algorithm, and generating a trained continuous depth confidence network;

extracting deep fault features of the engineering equipment by adopting a trained continuous deep belief network;

the deep fault features are used as the input of the SOM neural network for unsupervised learning to generate a trained SOM neural network;

constructing a health index of the engineering equipment by calculating Euclidean distances between state vectors of an input layer of the trained SOM neural network and weight vectors of all units; the health index directly reflects the real health state of the engineering equipment.

2. The method according to claim 1, wherein the obtaining of the test data of the engineering equipment and the calculating of the time-frequency domain information of the test data specifically include:

acquiring test data of engineering equipment;

extracting time domain characteristics of the test data; the time domain features comprise a mean value, a root mean square, a square root amplitude, an average absolute amplitude, skewness, kurtosis, variance, a maximum value, a minimum value, a peak value, a form factor, a peak factor, a pulse factor, a margin factor, a skewness factor and a kurtosis factor;

extracting frequency domain characteristics of the test data; the frequency domain features comprise center frequency, root mean square frequency and frequency variance;

and forming the time-frequency domain information of the engineering equipment by using the extracted time domain features and the extracted frequency domain features.

3. The method according to claim 2, wherein the inputting the time-frequency domain information into a continuous deep confidence network, training the continuous deep confidence network by means of a contrast divergence algorithm, and generating a trained continuous deep confidence network specifically comprises:

constructing a continuous depth confidence network formed by stacking a plurality of continuous limited Boltzmann machines; the continuous limited Boltzmann machine comprises a visible layer, a hidden layer and a Gaussian noise unit;

and inputting the time-frequency domain information into the continuous deep belief network, updating the connection weight and the sigmoid function slope control parameter of the continuous deep belief network by means of a contrast divergence algorithm, completing parameter training of the continuous deep belief network, and generating a trained continuous deep belief network.

4. The method according to claim 3, wherein the extracting deep fault features of the engineering equipment by using the trained continuous deep belief network specifically comprises:

setting the time-frequency domain information X of the test data as [ X ═ X% ₁ ,x ₂ ,…,x _n ]As the input of the trained continuous deep belief network, the output Y of the trained continuous deep belief network is [ Y ═ Y ₁ ,y ₂ ,…,y _m ]As a deep fault feature of the engineering equipment; wherein x is _i (i is more than or equal to 0 and less than or equal to n) represents the characteristic sequence of the ith input neuron of the continuous deep belief network, and n is the total number of the input characteristics; y is _i And (i is more than or equal to 0 and less than or equal to m) represents the characteristic sequence of the ith output neuron of the continuous deep belief network, and m is the total number of the output neurons of the continuous deep belief network.

5. The method according to claim 4, wherein the unsupervised learning is performed by taking the deep fault feature as an input of an SOM neural network, and a trained SOM neural network is generated, specifically comprising:

the deep fault feature Y is [ Y ═ Y% ₁ ,y ₂ ,…,y _m ]And one row in the SOM neural network is used as an input vector D of the SOM neural network, the optimal matching unit and the weight vector of the topological neighborhood are updated according to the input vector D, the unsupervised learning of the SOM neural network is completed, and the trained SOM neural network is generated.

6. A construction system for health indexes of engineering equipment is characterized by comprising the following components:

the time-frequency domain information extraction module is used for acquiring test data of the engineering equipment and calculating the time-frequency domain information of the test data;

the continuous deep belief network training module is used for inputting the time-frequency domain information into a continuous deep belief network, training the continuous deep belief network by means of a contrast divergence algorithm and generating a trained continuous deep belief network;

the deep fault feature extraction module is used for extracting the deep fault features of the engineering equipment by adopting a trained continuous deep belief network;

the SOM neural network training module is used for performing unsupervised learning by taking the deep fault characteristics as the input of the SOM neural network to generate a trained SOM neural network;

the engineering equipment health index construction module is used for constructing the health index of the engineering equipment by calculating Euclidean distances between the state vector of the input layer of the trained SOM neural network and the weight vectors of all units; the health index directly reflects the real health state of the engineering equipment.

7. The system according to claim 6, wherein the time-frequency domain information extracting module specifically includes:

the test data acquisition unit is used for acquiring test data of the engineering equipment;

the time domain feature extraction unit is used for extracting the time domain features of the test data; the time domain features comprise a mean value, a root mean square, a square root amplitude, an average absolute amplitude, skewness, kurtosis, variance, a maximum value, a minimum value, a peak value, a form factor, a peak factor, a pulse factor, a margin factor, a skewness factor and a kurtosis factor;

the frequency domain characteristic extraction unit is used for extracting the frequency domain characteristics of the test data; the frequency domain features comprise center frequency, root mean square frequency and frequency variance;

and the time-frequency domain information construction unit is used for constructing the time-frequency domain information of the engineering equipment by using the extracted time domain characteristics and the extracted frequency domain characteristics.

8. The system of claim 7, wherein the continuous deep belief network training module specifically comprises:

the continuous depth confidence network construction unit is used for constructing a continuous depth confidence network formed by stacking a plurality of continuous limited Boltzmann machines; the continuous limited Boltzmann machine comprises a visible layer, a hidden layer and a Gaussian noise unit;

and the continuous deep belief network training unit is used for inputting the time-frequency domain information into the continuous deep belief network, updating the connection weight and the sigmoid function slope control parameter of the continuous deep belief network by means of a contrast divergence algorithm, completing continuous deep belief network parameter training and generating a trained continuous deep belief network.

9. The system according to claim 8, wherein the deep fault feature extraction module specifically includes:

a deep fault feature extraction unit, configured to extract [ X ] from the time-frequency domain information X of the test data ₁ ,x ₂ ,…,x _n ]As the input of the trained continuous deep belief network, the output Y of the trained continuous deep belief network is ═ Y ₁ ,y ₂ ,…,y _m ]As a deep fault feature of the engineering equipment; wherein x is _i (0 ≦ i ≦ n) representing the characteristic sequence of the ith input neuron of the continuous deep belief network,n is the total number of input features; y is _i And (i is more than or equal to 0 and less than or equal to m) represents the characteristic sequence of the ith output neuron of the continuous deep belief network, and m is the total number of the output neurons of the continuous deep belief network.

10. The system of claim 9, wherein the SOM neural network training module specifically comprises:

an SOM neural network training unit for training the deep fault feature Y ═ Y ₁ ,y ₂ ,…,y _m ]And one row in the SOM neural network is used as an input vector D of the SOM neural network, the optimal matching unit and the weight vector of the topological neighborhood are updated according to the input vector D, the unsupervised learning of the SOM neural network is completed, and the trained SOM neural network is generated.

Images