CN116956089A - Training method and detection method for temperature anomaly detection model of electrical equipment - Google Patents

Abstract

The utility model provides an electrical equipment temperature anomaly detection model training method, which comprises the following steps: training a depth self-encoder using the power and temperature data set; performing kernel density estimation on a reconstruction error obtained by the depth self-encoder according to power and temperature data to obtain probability density distribution of the reconstruction error; and executing a k-means clustering algorithm on the probability density distribution of the reconstruction error to obtain clustering of abnormal values.

Description

Electrical equipment temperature anomaly detection model training method and detection method

Technical field

The invention relates to the field of temperature detection of electrical equipment, and in particular to a method for training an abnormal temperature detection model of electrical equipment based on a multi-quantification mechanism and a detection method.

Background technique

Anomaly detection is a key task in predictive maintenance. A properly designed anomaly detection strategy can improve the stability of the operation of electrical equipment, such as uninterruptible power supplies (UPS). By closely monitoring the output power of electrical equipment, radiator temperature and other indicators (KPI), and detecting abnormal data of components such as IGBTs/diodes in operating equipment, we can predict the operating trend of electrical equipment, infer the cause of faults and trigger troubleshooting or troubleshooting in a timely manner. ease.

The goal of outlier detection is to reveal unusual patterns in data. Typical scenarios in predictive maintenance are identifying equipment failures, sensor failures or system intrusions. This is a challenging task, especially when there are interactions between data and outliers become hidden and difficult to judge in a single dimension.

There are multiple anomaly detection methods in the existing technology. Anomaly detection based on statistical methods assumes that the data obeys a specified probability distribution, and if the probability of generating a data point from the model is below a certain threshold, it is defined as an anomaly. The advantage of this model is that it provides probability as a decision rule for judging anomalies, but it lacks automatic judgment of the threshold and often requires manual determination. Proximity-based anomaly detection assumes that anomalous data is isolated from the majority of the data. Anomaly detection algorithms, including distance-based and density-based, evaluate the relationship of each outlier point by identifying dense areas or clusters that exist in the data, and form the relationship between each outlier point. An anomaly is defined if the proximity to the cluster centroid is above the threshold or the size of the nearest cluster is below the threshold, or the number of data points within a local region of the data points is below the threshold. The model also lacks automatic judgment of thresholds and requires manual determination. Supervised learning-based methods require manually marking the anomalies of each KPI.

However, in reality, most data are unlabeled, and in most cases knowledge needs to be obtained through unsupervised learning. Therefore, when faced with a large number of KPIs, neither traditional statistical methods nor methods based on supervised learning can work automatically and effectively.

Contents of the invention

Based on the above problems of the prior art, the present invention proposes a method for training an electrical equipment temperature anomaly detection model, which includes:

Train deep autoencoders using power and temperature datasets;

Perform kernel density estimation on the reconstruction error obtained by the deep autoencoder based on power and temperature data to obtain the probability density distribution of the reconstruction error;

A k-means clustering algorithm is performed on the probability density distribution of the reconstruction error to obtain clustering of outliers.

Preferably, before using the power and temperature data set to train the deep autoencoder, data shaping and data normalization are performed on the power and temperature data set.

Preferably, the clustered data of the outliers are marked as severe anomalies, general anomalies and normal.

Preferably, the electrical device is an uninterruptible power supply.

The invention also provides a method for detecting temperature anomalies in electrical equipment, including:

Collect the temperature and corresponding output power of the electrical equipment;

The collected temperature and output power data are input into the above-mentioned electrical equipment temperature abnormality detection model to determine whether the temperature of the electrical equipment is abnormal.

Preferably, if the temperature and output power data are marked as abnormal, it is determined whether the temperature and output power data are abnormal in trend change or abnormal in temporary mutation.

Preferably, when the power trend changes abnormally with the temperature, the accuracy of the temperature sensor of the electrical device is determined.

Preferably, when the temperature trend changes abnormally with the power, it is determined whether the temperature control system of the electrical equipment fails.

Preferably, both temperature and output power data marked as general anomalies and severe anomalies are alerted.

Preferably, if the temperature and output power data are marked as serious abnormalities, it is determined that the electrical equipment is faulty.

Preferably, for temperature and output power data with abnormal trend changes in severe anomalies, a severe alarm is issued.

Preferably, before inputting the collected temperature and output power data into the electrical equipment temperature anomaly detection model, data shaping and data normalization are performed on the temperature and output power data.

According to business needs, the present invention designs a temperature anomaly detection method for electrical equipment based on a multi-quantification mechanism in a data-driven manner. It uses temperature-power time series data as model input and uses an unsupervised deep autoencoder to construct a time series correlation. The relationship model between KPIs, then perform kernel density estimation on the reconstruction error, calculate the error distribution curve, and use the number of maximum values of the kernel density curve as the number of k-means mean clustering, with larger k-nearest neighbor distance and variance Clustered data are defined as anomalies.

The electrical equipment temperature anomaly detection method based on a multi-quantification mechanism of the present invention uses unsupervised deep learning and automatic clustering to solve the outlier discrimination of mutually influencing periodic temperature-power data, and realizes temporary mutation anomalies and Detection of trend change anomalies and achieved good test results. Through the detection of abnormal power changes and abnormal temperature changes, it provides powerful data explanations for analyzing the power operating status of key components and monitoring the operating status of the temperature control system.

The data-driven temperature anomaly detection method is also suitable for outlier detection between other related KPIs on the IoT big data platform, and has a wide range of uses in predictive maintenance scenarios.

Description of the drawings

Figure 1 shows a block diagram of a UPS temperature anomaly detection method based on a multi-quantization mechanism according to an embodiment of the present invention.

Figure 2 shows the network structure of the deep autoencoder.

Figure 3 shows the temperature time series curve of the data set D1 according to one embodiment of the present invention.

Figure 4 shows the output power time series curve of the data set D1 according to one embodiment of the present invention.

Figure 5 shows a temperature time series curve of data set D2 according to an embodiment of the present invention.

Figure 6 shows the output power time series curve of the data set D2 according to one embodiment of the present invention.

Figure 7 shows the reconstruction result of the data set D1 according to one embodiment of the present invention.

Figure 8 shows the reconstruction error of the data set D1 according to one embodiment of the present invention.

Figure 9 shows a schematic slice diagram of a time period of the reconstruction result of the data set D1 according to an embodiment of the present invention.

Figure 10 shows a schematic slice diagram of another time period of the reconstruction result of the data set D1 according to an embodiment of the present invention.

Figure 11 shows the reconstruction result of data set D2 according to one embodiment of the present invention.

Figure 12 shows the reconstruction error of data set D2 according to one embodiment of the present invention.

Figure 13 shows a schematic slice diagram of a time period of the reconstruction result of the data set D2 according to an embodiment of the present invention.

Figure 14 shows a schematic slice diagram of another time period of the reconstruction result of the data set D2 according to an embodiment of the present invention.

Figure 15 shows the reconstruction error density distribution of the data set D1 according to one embodiment of the present invention.

Figure 16 shows automatic clustering of reconstruction errors of data set D1 according to one embodiment of the present invention.

Figure 17 shows the reconstruction error density distribution of data set D2 according to one embodiment of the present invention.

Figure 18 shows automatic clustering of reconstruction errors of data set D2 according to one embodiment of the present invention.

Figure 19 shows a schematic diagram of a temporary sudden change in power relative to temperature at a certain moment.

Figure 20 shows a schematic diagram of a temporary sudden change in temperature that is abnormally larger than the power at a certain moment.

Figure 21 shows a schematic diagram of the power trend changing abnormally with temperature at a certain moment.

Figure 22 shows a schematic diagram of the temperature trend changing abnormally with the power at a certain moment.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail through specific embodiments in conjunction with the accompanying drawings. It should be noted that the embodiments given in the present invention are for illustration only and do not limit the scope of the present invention.

Periodic KPIs are very common in practical applications and are business-related. The time series of temperature, power, and load rate in electrical equipment interact with each other and the data itself has a certain periodicity. It is difficult to accurately analyze the business using a single detection model. Starting from the need for unsupervised and automated anomaly detection, the present invention proposes a temperature anomaly detection method for electrical equipment based on a multi-quantification mechanism, a robust and unsupervised KPI anomaly detection method. The temperature anomaly detection method of electrical equipment based on multi-quantification mechanism is first based on Deep Auto Encoder (DAE), using temperature-power time series data as model input, and building a relationship model between time series related KPIs through deep learning. , and reconstruct the time series data. Then the reconstruction error is estimated by kernel density, and the error distribution curve is calculated. The number of maximum values of the kernel density curve is used as the number of k-means mean clusters. Clustered data with larger k-nearest neighbor distance and variance are defined as anomalies. , thereby realizing automatic detection of temperature anomalies in electrical equipment.

In the following, a UPS is taken as an example to describe in detail the electrical equipment temperature anomaly detection model training method and detection method of the present invention. However, those skilled in the art should understand that this method can also be applied to any other electrical equipment.

Figure 1 shows a block diagram of a UPS temperature anomaly detection method based on a multi-quantization mechanism according to an embodiment of the present invention. The UPS temperature anomaly detection method based on the multi-quantification mechanism includes four parts: data collection 101, data processing 102, anomaly detection model 103 and status identification 104. Among them, data collection refers to collecting UPS equipment operating information at different geographical locations through the Internet of Things (IOT) data pipeline inside the UPS, such as collecting the temperature and corresponding output power at a certain location. UPS usually includes multiple temperature sensors to detect the temperature of IGBTs, diodes and other components. Data processing 102 refers to data shaping and data normalization, that is, interpolating the missing data and normalizing the collected data. The anomaly detection model 103 includes deep autoencoding training, kernel density estimation, and mean clustering, and is used to analyze the processed data and obtain outliers. State identification 104 is used to classify the obtained abnormal values, for example, into healthy states, temporary mutations, and trend changes. Among them, the temporary sudden change abnormality refers to the temporary sudden change of the power that is abnormal to the temperature or the temporary sudden change of the temperature to the power. The abnormal trend change refers to the abnormal change of the power trend to the temperature or the abnormal change of the temperature trend to the power. The anomaly detection model 103 is described in detail below.

Deep autoencoder:

Autoencoder is an unsupervised learning model. It uses neural networks to generate low-dimensional representations of high-dimensional inputs, and can use nonlinear activation functions to meet the needs of nonlinear fitting. Autoencoders contain two main parts, encoder and decoder. The encoder is used to discover a compressed representation of the input data, and the decoder is used to reconstruct the original input data. During training, the decoder forces the autoencoder to select the most informative features and save them in the compressed representation.

The encoding process corresponds to formula (1), the decoding process corresponds to formula (2), and formula (3) is used to calculate the error caused by the encoding and decoding processes corresponding to the original data, which is called reconstruction error. The purpose of training is to minimize to reduce the reconstruction error.

h=σ(W _xh x+b _xh ) (1)

z＝σ(W _hx h+b _hx ) (2)

‖x-z‖ (3)

Among them, h is the hidden layer, x is the input data, and z is the output data. The dimensionality of the hidden layer h is lower than that of the input layer, and the decoding process needs to reproduce the input data x according to the hidden layer. This makes the autoencoder robust to data with white noise and only captures meaningful patterns of the data.

Deep learning uses the characteristics of artificial neural networks to build models with a hierarchical structure. Deep autoencoders (encoders containing at least one additional hidden layer) can approximate any mapping from input to encoding with arbitrary accuracy given a sufficient number of hidden units. In order to achieve this kind of reproduction, deep autoencoders must capture the most important factors that can represent the input data and find the main features that can represent the original information. Given unlabeled data, learn features in an unsupervised manner.

Figure 2 shows the network structure of the deep autoencoder. Among them, the encoder includes multiple hidden layers L1, L2...Ln, and the decoder includes multiple hidden layers L1', L2'...Ln'. The input layer inputs an n-dimensional input data x. The encoder f(x)=q converts the input data x into an m-dimensional data form q. The decoder will try to restore the data q to its original dimension. The entire network can be seen as trying to keep g(f(x)) and x as close as possible. Measuring the fidelity of reconstruction is the value of the loss function L(x,g(f(x))). The smaller the value,/> The closer to x. Because it is unlabeled data, when abnormal events occur in the model, the reconstruction error will increase significantly due to different characteristics of the data.

Kernel density estimation:

Kernel density estimation is a non-parametric method for estimating the probability density function. It completely uses the information of the data itself and does not attach any assumptions to the data distribution, so that it can approximate the sample data to the greatest extent. {X ₁ , X ₂ , _{X 3} ,...,

Among them, K(x) is a kernel function (non-negative, integral is 1, conforms to probability density properties and has a mean value of 0), which can be, for example, a Gaussian kernel function, a trigonometric kernel function, or an Epanechnikov kernel function. wait. Among them, h>0 is the smoothing parameter, called bandwidth. K _h (x) = 1/h·K (x/h), which is the scaled kernel function.

Given a given sample set, the choice of h can be made by minimizing the L2 risk function (mean integrated squared error).

Under the weak assumption, MISE(h)=AMISE(h)+o(1/(nh)+h/4), where AMISE is asymptotic MISE. And AMISE has:

in:

R(K)＝∫K(x) ² dx,m ₂ (K)＝∫x ² K(x)dx (7)

In order to minimize MISE(h), it is transformed into a pole finding problem:

get:

When the kernel function is determined, R, m, and f in formula (9) can all be determined, and h will have an analytical solution. If a Gaussian kernel function is used for kernel density estimation, the optimal choice of h (that is, the bandwidth that minimizes the mean integrated squared error) is:

here is the standard deviation of the sample. This approximation is called the normal distribution approximation, Gaussian approximation, or Silverman's (1986) rule of thumb.

Kernel density estimation uses a smooth peak function ("kernel") to fit the observed data points, thereby fitting a true probability distribution curve. Under the condition of understanding the probability distribution, if a certain number appears in observation, it can be considered that the probability density of this number is very large, and the probability density of numbers that are closer to this number will also be relatively large, and those that are far away from this number will also have a relatively large probability density. The probability density of numbers will be relatively small. Kernel density estimation will have boundary effects when estimating boundary areas. Therefore, the clustering algorithm based on probability density can be determined by the tightness of the sample distribution. Based on the density of the data set in the spatial distribution, the clustering can be divided according to the density boundary. kind.

k-means clustering algorithm:

k-means clustering algorithm is also known as k-means clustering algorithm. For a given data set of n samples (unlabeled), given the number of clusters k (k<n), initialize the category to which each sample belongs, and then classify each sample according to the distance. Assign it to the cluster with the nearest center point, and then iteratively update the cluster center point position until the termination condition is reached.

The main three points to implement the k-means algorithm are:

(1) Selection of the number of clusters k

The selection of k is generally determined according to actual needs. In the UPS temperature anomaly detection method based on the multi-quantization mechanism of the present invention, the k value is given based on the number of maximum values estimated by the kernel density.

(2) The distance between each sample point and the "cluster center"

A given data sample X contains n objects {X ₁ , X ₂ , X ₃ ,..., X _n }, where each object has attributes of m dimensions. The goal of the K-means algorithm is to cluster n objects into specified k clusters based on the similarity between objects. Each object belongs to and only belongs to one cluster with the smallest distance to the cluster center. For the k-means algorithm, you first need to initialize k cluster centers {C ₁ , C ₂ , C ₃ ,..., C _k }, 1<k≤n, and then calculate the distance from each object to each cluster center Euclidean distance.

In the above formula, X _i represents the i-th object (1≤i≤n), Cj represents the j-th cluster center (1≤j≤k), _and t≤m), C _jt represents the t-th attribute of the j-th cluster center.

(3) Update the "cluster center" according to the newly divided clusters

For each divided cluster, calculate the mean value of the sample points in each cluster, and use its mean value as the new cluster center. The K-means algorithm uses the center to define the prototype of the cluster. The cluster center is the mean value of all objects in the cluster in each dimension. Its calculation formula is as follows:

In the formula, C _t represents the center of the l-th cluster (1≤l≤k), ∣S _l∣ represents the number of objects in the l-th cluster, and X _i represents the i-th object in the l-th cluster. (1≤i≤∣S _l∣ ). Compare the distance between each object and each cluster center in turn, and assign the object to the cluster with the closest cluster center, obtaining k clusters {S ₁ , S ₂ , S ₃ ,..., S _k }.

Based on the above content, the present invention provides a method for training an uninterruptible power supply temperature anomaly detection model based on a multi-quantization mechanism. The method includes the following steps:

Step 101: Train a deep autoencoder using the power and temperature datasets.

Usually, UPS equipment operating information at different geographical locations can be collected through internal IOT data pipelines, such as collecting the temperature of a certain component and the output power of the corresponding component. UPS usually includes multiple temperature sensors to detect the temperature of IGBTs, diodes and other components. In this model training method, IOT provides 2 well-maintained UPS data sets, data sets D1 and D2. Each data set contains equipment operation time series at 5-minute intervals for 2 months, including KPIs such as temperature and corresponding output power. Temperature and power data have no tag information to indicate normal or abnormal. Since all UPS devices are running normally, the proportion of healthy data is much larger than the abnormal data, so the data is unbalanced. Figures 3 and 4 respectively show the temperature and output power time series curves of the data set D1; and Figures 5 and 6 respectively show the temperature and output power time series curves of the data set D2.

Use all samples of data sets D1 and D2 as training sets, and set the following parameters to build a deep autoencoder. The input and output layer dimensions are 2, each represented by 2 neurons; the hidden layer uses a 3-layer structure to amplify and compress the input data, with the first layer using 3 neurons, the second layer using 2 neurons, and the hidden layer using 3 neurons. The third layer uses 3 neurons. ReLU is used as the activation function in the encoder hidden layer, Sigmoid is used as the activation function in the decoder hidden layer, the neural network uses the adadelta gradient optimizer and the binary_crossentropy loss function, and the Euclidean distance is used as the reconstruction error of the time series data. The neural network automatically learns the patterns of most of the data in data sets D1 and D2 and minimizes the reconstruction error. When abnormal events occur in the model, the reconstruction error will increase significantly. Figures 7 to 14 show the reconstruction results and reconstruction errors of the temperature-power correlation relationship learned by the deep autoencoder.

Step 102: Perform kernel density estimation on the reconstruction error obtained by the deep autoencoder based on the power and temperature data to obtain the probability density distribution of the reconstruction error. The reconstruction error probability density distributions of data sets D1 and D2 are shown in Figure 15 and Figure 17 respectively.

After the training of the deep autoencoder is completed, the above power and temperature data sets can be re-entered into the deep autoencoder, and the resulting reconstruction error can be estimated by kernel density; other power and temperature data sets can also be input. into the deep autoencoder, and perform kernel density estimation on the resulting reconstruction error.

Step 103: Execute k-means clustering algorithm on the obtained probability density distribution to obtain clustering of outliers.

Using the number of maximum values of the reconstruction error probability density distribution of data sets D1 and D2 as the basis for division, four clusters were divided respectively through the k-means clustering algorithm. The k-nearest neighbor distance and variance size are used to divide the clustering results into abnormal levels, marked as serious anomalies and general anomalies. The abnormality level of the most remote clustering data is serious anomaly, and the anomaly level of more remote clustering data is general anomaly. Figures 16 and 18 show the automatic clustering results of reconstruction errors for data sets D1 and D2 respectively. Divided into four clusters, cn1-cn4. Cluster cn4 in Figure 16 and Figure 18 is the most remote cluster data and is marked as a serious anomaly. Cluster cn2 in Figures 16 and 18 is more remote cluster data and is marked as a general anomaly. Cluster cn1 and cluster cn3 in Figure 16 and Figure 18 are close to the center and marked as normal.

Anomaly detection involves distinguishing abnormal states from healthy states, or determining the source of the anomaly when an abnormal state occurs. In many cases, only measurements from healthy machines are available. The method for training an uninterruptible power supply temperature anomaly detection model based on a multi-quantification mechanism of the present invention uses output power indicators and ambient temperature indicators extracted from system data to train a decision-making model. The anomaly detection model can analyze indicators extracted from test data. to determine the current system status.

The present invention also provides an uninterruptible power supply temperature anomaly detection method based on a multi-quantification mechanism, which uses the above-mentioned uninterruptible power supply temperature anomaly detection model to perform temperature anomaly detection. The method includes the following steps:

Step 201: Collect the temperature and corresponding output power of the uninterruptible power supply;

Step 202: Input the collected temperature and output power into the above-mentioned temperature abnormality detection model to determine whether the temperature of the uninterruptible power supply is abnormal;

Step 203: If the temperature and output power data are marked as serious abnormalities, determine that the uninterruptible power supply has failed.

Anomaly detection in time series expressed as outliers relative to a standard signal. UPS only focuses on the most important types of business. Among outliers (such as clusters cn4 (serious anomalies) and cn2 (general anomalies)), UPS temperature-power anomalies are mainly divided into the following two types:

Temporary mutation anomaly: The abnormal value has a certain initial effect when T interference occurs at a certain moment, and then decays over time and tends to normal. As shown in FIG. 19 , the temporary sudden change of power is abnormal to the temperature; and as shown in FIG. 20 , the temporary sudden change of temperature is abnormal to the power.

Abnormal trend changes: The interference of outliers is that at a certain time T, the structure of the system changes, and continues to affect all behaviors after time T. The sequence often shows a horizontal shift in the mean value of the sequence before and after time T. As shown in Figure 21, the power trend changes abnormally with temperature, and the accuracy of the temperature sensor can be further analyzed. As shown in Figure 22, the temperature trend changes abnormally with the power, especially the temperature trend increases, which is particularly important for equipment health detection, and can further analyze whether the temperature control system has failed.

If the temperature and output power data are marked as serious abnormalities, the uninterruptible power supply is judged to have failed. Both general abnormality and severe abnormality data may contain temporary mutation anomalies and trend change anomalies. The trend change anomalies in severe anomalies are usually more serious. Therefore, a serious alarm is issued for data with abnormal trend changes in severe anomalies.

Develop sensitivity strategies and accuracy strategies based on the classification of abnormal detection results. For the sensitivity strategy, both general anomalies and severe anomalies are treated as abnormal data for alarming; for the accuracy strategy, only severely abnormal data is focused on and the possibility of an imminent failure is analyzed. The anomaly detection strategy marks each time point as healthy/abnormal and marks the degree of abnormality.

The UPS anomaly detection data set is an unbalanced data set, and the proportion of normal data is much larger than the abnormal data. According to business needs, accuracy is used as the evaluation index of the classification model. Abnormal detection results can be divided into two categories: true positive (TP) and false positive (FP). Among them, TP represents the actual anomalies classified as anomalies, while FP represents the classification of normal instances into anomaly types. Precision P (Precision) represents the percentage of correct predictions:

P＝TP/(TP+FP) (13)

The data sets provided by the internal IOT data pipeline are unlabeled, and the health status of the original data is not marked. The UPS temperature anomaly detection algorithm focuses on the most important temporary mutation anomalies and trend change anomalies in the business. By backtracking the data, it is consistent with the above Exception type data are marked as exceptions. The temperature anomaly detection method of the present invention performs classification and evaluation on the data sets D1 and D2, and the clusters of general anomalies G and severe anomalies S respectively. The results are shown in Table 1 below.

Table 1. Anomaly detection results of temperature anomaly detection methods for different data sets

clustering abnormal Accuracy D1_cn4 S 1 D2_cn4 S 1 D1_cn2 G 0.31 D2_cn2 G 0.28

As can be seen from Table 1, the UPS temperature anomaly detection method based on the multi-quantification mechanism has achieved a high accuracy rate in clustering serious abnormal S. The detection results of general anomaly G show that there are a certain number of abnormal samples mixed in the clustered samples. After comparing the original data, it is found that there are model errors, abnormal, normal and other data points in the clustered samples at the same time, and the sample may be in a critical state.

Equipment temperature data when electrical equipment is running has daily cycle characteristics, and is affected by power consumption behavior. When the load increases and the power increases, the temperature rises rapidly. When the power decreases, the temperature drops significantly due to the influence of the temperature control system. The anomaly detection algorithm of the present invention first constructs a temperature-power time series model through a deep autoencoding network, and unsupervised learning of the temperature-power correlation relationship. The non-parametric kernel density estimation method is used to calculate the distribution density of the temperature and power time series reconstruction errors, and the number of maximum values of the probability curve is used as the number of k-means mean clusters. The mean clustering algorithm defines clustered data with larger k-nearest neighbor distance and variance as abnormal states. An electrical equipment temperature anomaly detection algorithm based on a multi-quantification mechanism uses unsupervised learning to build correlations and uses automated means to identify abnormal data.

Outlier detection is a key task in predictive maintenance, and a well-designed anomaly detection strategy can improve the reliability of equipment operation. The electrical equipment temperature anomaly detection algorithm based on the multi-quantification mechanism of the present invention does not need to obtain normal and abnormal point labels in advance, and can directly infer the degree of abnormality relative to other points from the data points. This is convenient in anomaly detection since the characteristics of anomalies/outliers are often not known in advance. Starting from business needs and in a data-driven manner, the present invention designs a temperature anomaly detection algorithm for electrical equipment based on a multi-quantification mechanism, using an unsupervised deep learning model and an automated anomaly clustering method to achieve temporary mutation anomalies and trend change anomalies. The detection solves the problem of outlier discrimination of mutually influencing periodic temperature-power, and achieves good classification results, laying a data foundation for further analysis of equipment failure causes and prediction of failure trends.

Although the present invention has been described through preferred embodiments, the present invention is not limited to the embodiments described here and includes various modifications and variations without departing from the scope of the present invention.

Claims (12)

1. A method for training an electrical equipment temperature anomaly detection model, including: Train deep autoencoders using power and temperature datasets; Perform kernel density estimation on the reconstruction error obtained by the deep autoencoder based on power and temperature data to obtain the probability density distribution of the reconstruction error; A k-means clustering algorithm is performed on the probability density distribution of the reconstruction error to obtain clustering of outliers. 2. The electrical equipment temperature anomaly detection model training method according to claim 1, wherein before using the power and temperature data set to train the deep autoencoder, data shaping and data normalization are performed on the power and temperature data set. one. 3. The electrical equipment temperature anomaly detection model training method according to claim 1, wherein the clustered data of the abnormal values is marked as serious abnormality, general abnormality and normal. 4. The electrical equipment temperature anomaly detection model training method according to any one of claims 1 to 3, wherein the electrical equipment is an uninterruptible power supply. 5. A method for detecting temperature anomalies in electrical equipment, including: Collect the temperature and corresponding output power of the electrical equipment; The collected temperature and output power data are input into the electrical equipment temperature abnormality detection model obtained in any one of claims 1-4 to determine whether the temperature of the electrical equipment is abnormal. 6. The electrical equipment temperature abnormality detection method according to claim 5, wherein if the temperature and output power data are marked as abnormal, it is determined whether the temperature and output power data are abnormal trend changes or temporary mutation abnormalities. 7. The temperature anomaly detection method of electrical equipment according to claim 6, wherein when the power trend changes abnormally with the temperature, the accuracy of the temperature sensor of the electrical equipment is determined. 8. The temperature anomaly detection method of electrical equipment according to claim 6, wherein when the temperature trend changes abnormally with the power, it is determined whether the temperature control system of the electrical equipment fails. 9. The method for detecting temperature anomalies in electrical equipment according to claim 5, wherein the temperature and output power data marked as general anomalies and severe anomalies are both alarmed. 10. The electrical equipment temperature abnormality detection method according to claim 9, wherein if the temperature and output power data are marked as serious abnormalities, it is determined that the electrical equipment is faulty. 11. The electrical equipment temperature anomaly detection method according to claim 9, wherein a serious alarm is issued for temperature and output power data with abnormal trend changes in severe anomalies. 12. The electrical equipment temperature anomaly detection method according to any one of claims 5-11, wherein before inputting the collected temperature and output power data into the electrical equipment temperature anomaly detection model, the temperature and The output power data is subjected to data shaping and data normalization.