Fault detection method and device for electrical equipment, equipment and storage medium

Abstract

The invention discloses a fault detection method and device for electrical equipment, equipment and a storage medium. The invention relates to the technical field of electrical equipment. The method comprises the following steps: acquiring time sequence data, image data and single-point data acquired by each sensor deployed by the electrical equipment at the current moment; extracting a first feature of the time sequence data, and extracting a second feature of the image data; fusing the first feature, the second feature and the single-point data to obtain a first feature vector; and determining a real-time fault state of the electrical equipment based on a pre-trained depth self-encoding Gaussian mixture model and the first feature vector. Through the above technical means, the deep self-encoding Gaussian mixture model can identify abnormal samples based on data distribution in the absence of negative sample data, the performance is excellent in a fault detection task, the problem that the model performance of a fault detection model is affected due to insufficient negative sample data in the prior art is solved, and the fault detection efficiency is improved. And the accuracy and reliability of fault detection are improved.

Description

Technical Field

[0001] The present application relates to the technical field of electrical equipment, and in particular to a fault detection method, device, equipment and storage medium for electrical equipment. Background Art

[0002] Electrical equipment plays an indispensable role in modern society. They support the normal operation of various fields from daily life to industrial production. Therefore, the stable and reliable operation of electrical equipment is directly related to the efficiency and quality of social and economic activities and people's daily life experience. In order to ensure the safe and reliable operation of electrical equipment, manual inspection was initially used, but the inspection quality was not high and it wasted manpower and material resources. Later, with the continuous updating and iteration of information automation technology, technologies such as big data and artificial intelligence were applied to the power system, converting traditional manual inspections into intelligent inspections, realizing the transformation from "planned maintenance" to "condition-based maintenance" and "predictive maintenance".

[0003] Traditional manual inspections rely on the senses of seeing, hearing, smelling, touching, and accumulated professional knowledge of operation and maintenance personnel to diagnose the health of on-site equipment. If intelligent inspections are to replace manual inspections, they also need to rely on information such as "seeing", "hearing", "smelling", and "touching" as well as expert knowledge to better replace humans to achieve the same functions. To collect this information, sensors of electrical equipment can be deployed, and various operating data of electrical equipment can be collected through sensors. For example, voiceprint collectors and vibration sensors realize the function of "listening", and temperature sensors realize the function of "touching". In addition, in the actual production process, as the commissioning time of electrical equipment continues to increase, its reliability, stability, and anti-interference ability are attenuated, that is, the health status of electrical equipment shows a trend of gradual attenuation. Therefore, it is difficult to accurately analyze the status of electrical equipment based on a single data.

[0004] In summary, in order to accurately analyze the status of electrical equipment, multi-source data fusion analysis can be performed. Multi-source data refers to a collection of data collected by homogeneous or heterogeneous sensors. Homogeneous sensors are sensors of the same type, but deployed in different locations. This type of data is often fused through correlation analysis methods. This type of data has good time synchronization and a relatively simple fusion method. Heterogeneous sensors are sensors of different categories, with different dimensions of collected signals and different hardware principles. Time synchronization is more difficult, but heterogeneous sensors have a wider "field of view" to monitor electrical equipment.

[0005] In the prior art, the method for detecting faults of electrical equipment based on multi-source data is to extract features of multi-source data, fuse the features and input them into a pre-trained fault detection model, and the fault detection model determines the fault diagnosis result of the electrical equipment. The training of the fault detection model depends on a large number of positive and negative sample data sets, but due to the reliability of the electrical equipment itself, it is difficult to collect negative sample data sets, resulting in insufficient negative samples for training the fault detection model, affecting the model performance of the fault detection model, thereby reducing the accuracy and reliability of fault detection. Summary of the invention

[0006] The present application provides a fault detection method, apparatus, device and storage medium for electrical equipment, which utilizes a deep autoencoder Gaussian mixture model as a fault detection model. In the absence of negative sample data, the deep autoencoder Gaussian mixture model can also recognize abnormal samples based on data distribution, and performs well in fault detection tasks, reducing dependence on negative sample data, solving the problem in the prior art that the model performance of the fault detection model is affected by insufficient negative sample data, and improving the accuracy and reliability of fault detection.

[0007] In a first aspect, the present application provides a fault detection method for an electrical device, comprising:

[0008] Obtain the time series data, image data and single point data collected by each sensor deployed on the electrical equipment at the current moment;

[0009] Extracting a first feature of the time series data and extracting a second feature of the image data;

[0010] Fusing the first feature, the second feature and the single point data to obtain a first feature vector;

[0011] Based on a pre-trained deep autoencoder Gaussian mixture model and the first feature vector, the real-time fault status of the electrical equipment is determined.

[0012] In a second aspect, the present application provides a fault detection device for an electrical device, comprising:

[0013] A multi-source data acquisition module is configured to acquire time series data, image data and single point data collected by various sensors deployed on the electrical equipment at the current moment;

[0014] a feature extraction module, configured to extract a first feature of the time series data and a second feature of the image data;

[0015] a feature fusion module, configured to fuse the first feature, the second feature and the single point data to obtain a first feature vector;

[0016] The fault detection module is configured to determine the real-time fault status of the electrical equipment based on a pre-trained deep autoencoder Gaussian mixture model and the first feature vector.

[0017] In a third aspect, the present application provides a fault detection device for an electrical device, comprising:

[0018] One or more processors; a storage device storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the fault detection method for electrical equipment as described in the first aspect.

[0019] In a fourth aspect, the present application provides a storage medium comprising computer executable instructions, which, when executed by a computer processor, are used to execute the fault detection method for an electrical device as described in the first aspect.

[0020] In the present application, the time series data, image data and single point data collected by each sensor deployed by the electrical equipment at the current moment are obtained; the first feature of the time series data is extracted, and the second feature of the image data is extracted; the first feature, the second feature and the single point data are fused to obtain the first feature vector; based on the pre-trained deep self-encoding Gaussian mixture model and the first feature vector, the real-time fault state of the electrical equipment is determined. Through the above technical means, the time series data, image data and single point data collected by various types of sensors are obtained as multi-source data for detecting the fault state of the electrical equipment, so as to comprehensively detect the fault state of the electrical equipment and improve the detection accuracy of the fault state. The features of the multi-source data are fused, and the abnormal features identified by the deep self-encoding Gaussian mixture model in the fused features are used to determine the real-time fault state of the electrical equipment based on the abnormal features. In the absence of negative sample data, the deep self-encoding Gaussian mixture model can also accurately detect the fault state of the electrical equipment, and performs well in the fault detection task, reducing the dependence on negative sample data, solving the problem that the model performance of the fault detection model is affected by insufficient negative sample data in the prior art, and improving the accuracy and reliability of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of a fault detection method for electrical equipment provided in an embodiment of the present application;

[0022] Figure 2 is a flowchart of extracting the first feature of time series data provided by an embodiment of the present application;

[0023] Figure 3 It is a schematic diagram of the structure of the Inception network provided in the embodiment of the present application;

[0024] Figure 4 It is a schematic diagram of the structure of the SeNet network provided in the embodiment of the present application;

[0025] Figure 5 It is a schematic diagram of the structure of the deep autoencoder Gaussian mixture model provided in an embodiment of the present application;

[0026] Figure 6 is a flow chart of determining the real-time fault status of electrical equipment based on a deep autoencoder Gaussian mixture model provided by an embodiment of the present application;

[0027] Figure 7 is a flow chart of a training feature extraction model and a fault detection model provided in an embodiment of the present application;

[0028] Figure 8 is a flow chart for predicting future fault states of electrical equipment provided by an embodiment of the present application;

[0029] Fig. 9 is a flow chart of a training feature prediction model provided in an embodiment of the present application;

[0030] Fig.10 It is a structural diagram of a state analysis model provided in an embodiment of the present application;

[0031] Fig.11 It is a structural schematic diagram of a fault detection device for electrical equipment provided in an embodiment of the present application;

[0032] Fig.12 It is a structural schematic diagram of a fault detection device for electrical equipment provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages of the present application clearer, the specific embodiments of the present application are further described in detail below in conjunction with the accompanying drawings. It is understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for the convenience of description, only the part related to the present application but not all the contents are shown in the accompanying drawings. Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow chart describes each operation (or step) as a sequential process, many of the operations therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of each operation can be rearranged. The process can be terminated when its operation is completed, but it can also have additional steps not included in the accompanying drawings. The process can correspond to a method, a function, a procedure, a subroutine, a subprogram, etc.

[0034] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0035] In the relevant implementation methods, the fault state analysis method of electrical equipment based on multi-source data is mainly divided into two steps: feature extraction and fault analysis. Among them, feature extraction is divided into two implementation methods: subjective identification and model extraction. Subjective identification is to extract features from multi-source data manually. For example, statistical time domain features such as effective value, peak value and kurtosis coefficient or some special frequency features are often extracted for vibration signals. Model extraction is to hand over feature extraction to the network model for implementation, and to mine the excellent features of multi-source data in the form of a black box. For example, DNN network and CNN network framework are trained separately to extract features for oil chromatogram and spectrum, and the features are fused to realize fault diagnosis, or CNN network and RNN network are trained separately to extract image features and voiceprint features, and then fused to realize health analysis of substation. However, the existing network models rely on a large number of positive and negative sample data sets for training when extracting features and fault analysis, otherwise the generalization ability of the model is poor and the performance is poor when predicting new data, and the model performance is poor. Due to the reliability of electrical equipment itself, it is difficult to collect negative sample data sets, resulting in insufficient negative samples for training network models, affecting the model performance of network models, thereby reducing the accuracy and reliability of fault detection.

[0036] To solve the above problems, this embodiment provides a fault detection method for electrical equipment, which utilizes a deep autoencoder Gaussian mixture model as a fault detection model. In the absence of negative sample data, the deep autoencoder Gaussian mixture model can also identify abnormal samples based on data distribution, and performs well in fault detection tasks, reducing dependence on negative sample data and improving the accuracy and reliability of fault detection.

[0037] The fault detection method for electrical equipment provided in this embodiment can be executed by a fault detection device for electrical equipment, which can be implemented by software and / or hardware, and can be composed of two or more physical entities, or can be composed of one physical entity. For example, the fault detection device for electrical equipment can be an intelligent terminal with strong processing capabilities such as a computer and a server. Among them, the server can be implemented by an independent server or a server cluster composed of multiple servers.

[0038] The fault detection device of the electrical device is installed with at least one type of operating system, and the fault detection device of the electrical device can install at least one application based on the operating system, and the application can be an application that comes with the operating system, or an application downloaded from a third-party device or server. In this embodiment, the fault detection device of the electrical device is installed with at least an application that can execute the fault detection method of the electrical device.

[0039] For ease of understanding, this embodiment is described by taking a server as an example of a main body that executes the fault detection method for electrical equipment.

[0040] Figure 1 A flow chart of a fault detection method for an electrical device provided in an embodiment of the present application is given.

[0041] refer to Figure 1 , the fault detection method of the electrical equipment specifically includes:

[0042] S110, obtaining time series data, image data, and single point data collected by various sensors deployed on the electrical equipment at the current moment.

[0043] Exemplarily, the electrical equipment is deployed with multiple heterogeneous sensors and multiple homogeneous sensors. The heterogeneous sensors are used to collect sensor data of different dimensions and formats, and the homogeneous sensors are used to collect sensor data of the same dimensions and formats. The sensors deployed on the electrical equipment are connected to the data management device via serial communication or network communication. The sensors transmit the sensor data collected in real time to the data management device via serial communication or network communication. The data management device establishes an association relationship between the sensor data and the device identification of the electrical equipment corresponding to the sensor, and reports the sensor data and the corresponding device identification to the server via network communication. The server analyzes the fault status of the electrical equipment corresponding to the device identification at the current moment based on the sensor data.

[0044] There are many types of electrical equipment, and the sensors deployed for different types of electrical equipment are different, and the sensor data collected is also different. For example, for motors, it is necessary to collect their temperature, vibration, soundprint, working current and noise, etc. For transformers, it is necessary to collect grounding current, partial discharge, hot spot temperature, vibration and noise, etc. For cables, it is necessary to collect temperature, voltage current, ambient temperature, etc. But in general, the data formats of sensor data commonly obtained for electrical equipment are mainly divided into time series form, single point form and image form, which correspond to time series data, single point data and image data respectively. For example, vibration, soundprint and high-frequency current are time series data, temperature, oil chromatogram and grounding current are single point data, infrared detection spectrum and ultrasonic detection spectrum are image data.

[0045] S120, extracting the first feature of the time series data and extracting the second feature of the image data.

[0046] For example, the first feature of the time series data and the second feature of the image data can be extracted by a pre-trained convolutional neural network. However, the single-point sensor data can only collect one value at the corresponding sampling point each time, and the analysis of the single-point value is relatively simple and can be directly used for subsequent fault status analysis.

[0047] It should be noted that the collection of time series data and image data depends on the properties of the corresponding sensors. When there are large differences in the properties of the sensors, such as acquisition frequency and data format, using the same feature extraction network will lead to problems such as unequal extracted information. Therefore, different feature extraction methods can be used for different types of sensor data.

[0048] Optionally, for image data, there are currently many deep learning networks that can achieve better image feature extraction. For this, the second feature of the image data can be extracted through a convolutional neural network.

[0049] For time series data, it involves the conversion of multiple dimensions, such as time domain, frequency domain and time-frequency domain, etc. The data extraction operation is relatively complicated. For this, the first feature of the time series data can be extracted by a combined network composed of the Inception network and the SeNet network. For example, Figure 2 1 is a flowchart of extracting the first feature of time series data provided by the embodiment of the present application. Figure 2 As shown, the step of extracting the first feature of the time series data specifically includes S1201-S1204:

[0050] S1201. Perform Fourier transform on the time series data to obtain frequency domain distribution data, and convert the time series data and the frequency domain distribution data into corresponding two-dimensional matrices.

[0051] For example, the frequency domain is an important feature domain for time series data, so before extracting features from the time series data, the time series data must first be Fourier transformed to obtain frequency domain distribution data. Since both the time series data and the frequency domain distribution data are one-dimensional arrays, in order for the subsequent Inception network to extract features from the time series data and the frequency domain distribution data, the time series data and the frequency domain distribution data are converted into two-dimensional matrices respectively.

[0052] For example, the time series data x(t) is subjected to a fast Fourier transform to obtain the frequency domain distribution data g(f). The length of the time series data is N, and the length of the frequency domain distribution is L. Then the relationship between N and L satisfies N=ωL, where ω is the Nyquist parameter. The specific value will be defined according to the Nyquist theorem and the sensor's own properties, and is generally 2.56. After that, the time series data x(t) and the frequency domain distribution data g(f) are subjected to a two-dimensional transformation. Since the two-dimensional matrices corresponding to the two data transformations need to be concatenated into a two-dimensional matrix, the greatest common divisor of N and L is determined to be the number of rows of the two-dimensional matrix. The time series data is transformed from a one-dimensional array to a two-dimensional matrix according to the number of rows, and the frequency domain distribution data is transformed from a one-dimensional array to a two-dimensional matrix according to the number of rows. For example, if the time series data x(t) is a one-dimensional array with a length of 12, and the frequency domain distribution data g(f) is a one-dimensional array with a length of 16, then the time series data x(t) is transformed into a 4*3 two-dimensional matrix, and the frequency domain distribution data g(f) is transformed into a 4*4 two-dimensional matrix.

[0053] S1202. Concatenate the two-dimensional matrix corresponding to the time series data and the two-dimensional matrix corresponding to the frequency domain distribution data to obtain a two-dimensional matrix in the time-frequency domain.

[0054] Exemplarily, the two-dimensional matrix corresponding to the time series data and the two-dimensional matrix corresponding to the frequency domain distribution data are spliced ​​row by row to obtain a two-dimensional matrix in the time-frequency domain. For example, after the time series data x(t) is converted into a 4*3 two-dimensional matrix and the frequency domain distribution data g(f) is converted into a 4*4 two-dimensional matrix, the two two-dimensional matrices are spliced ​​row by row to obtain a 4*7 two-dimensional matrix in the time-frequency domain.

[0055] S1203. Extract features of the two-dimensional matrix in the time-frequency domain through the Inception network to obtain a multi-dimensional feature matrix.

[0056] Exemplarily, the two-dimensional matrix in the time-frequency domain contains both the time domain information and the frequency domain information of the time series data. The two-dimensional matrix in the time-frequency domain can be input into the Inception network as input data of the Inception network. The Inception network convolves the two-dimensional matrix in the time-frequency domain with each convolution layer respectively, and uses the convolution result corresponding to the two-dimensional matrix in the time-frequency domain and each convolution layer as the feature matrix of a channel. The feature matrices of multiple channels corresponding to multiple convolution layers are summarized to obtain the multi-dimensional feature matrix extracted by the Inception network from the two-dimensional matrix in the time-frequency domain.

[0057] For example, Figure 3 Schematic diagram of the structure of the Inception network provided in the embodiment of the present application. Figure 3 As shown, the Inception network includes five convolutional layers, the first convolutional layer consists of a 1×1 convolutional kernel, the second convolutional layer consists of a 3×3 convolutional kernel, the third convolutional layer consists of two 3×3 convolutional kernels, the fourth convolutional layer consists of three 3×3 convolutional kernels, and the fifth convolutional layer consists of a pooling layer and a 1×1 convolutional kernel. The Inception network convolves the input time-frequency domain two-dimensional matrix with the convolutional kernel of each convolutional layer to obtain a two-dimensional matrix of five channels, and superimposes the two-dimensional matrices of the five channels on the channels to form a multi-dimensional feature matrix. This embodiment can capture the feature information of time-frequency domain two-dimensional matrices of different scales through the Inception network, thereby improving the efficiency and accuracy of feature extraction.

[0058] S1204. Perform feature weighted fusion processing on the multi-dimensional feature matrix through the SeNet network to obtain the first feature of the time series data.

[0059] Exemplarily, the multidimensional feature matrix output by the Inception network is used as the input data of the SeNet network. The SeNet network calculates the weight of the two-dimensional matrix under each channel in the multidimensional feature matrix, multiplies the weight by the corresponding two-dimensional matrix, and then performs channel superposition to obtain the first feature of the time series data, that is, the first feature is a multidimensional matrix.

[0060] For example, Figure 4 Schematic diagram of the structure of the SeNet network provided in the embodiment of the present application. Figure 4As shown, the SeNet network includes a compression module (Squeeze) and an excitation module (Excitation). The compression module compresses the two-dimensional matrix under each channel of the multi-dimensional feature matrix into a one-dimensional channel weight through global average pooling, thereby capturing the global information between channels. The excitation module weights the two-dimensional matrix under each channel according to the channel weight output by the compression module through a fully connected layer and an activation function, outputs the weighted two-dimensional matrix under each channel, and combines the weighted two-dimensional matrices under multiple channels to obtain the first feature. This embodiment can better understand and utilize the correlation information between feature channels through the SeNet network, improve the feature expression ability of the network, and thus improve the accuracy of feature extraction.

[0061] This embodiment uses a combination network of the Inception network and the SeNet network to extract features from time series data, which expands the perception range of the combined network for data, and can improve the performance of subsequent models for fault detection while maintaining computational efficiency, thereby reducing the risk of overfitting.

[0062] S130, fusing the first feature, the second feature and the single point data to obtain a first feature vector.

[0063] Exemplarily, the first feature is a multi-dimensional feature matrix, and the second feature is a three-dimensional feature matrix. The number of channels and matrix dimensions are different, and they cannot be directly spliced ​​with the single-point data. Therefore, the first feature and the second feature can be converted into one-dimensional feature vectors respectively through the maximum pooling layer and the fully connected layer, and the one-dimensional feature vector corresponding to the first feature, the one-dimensional feature vector corresponding to the second feature, and the numerical value corresponding to the single-point data are spliced ​​to obtain the first feature vector.

[0064] S140, determining the real-time fault status of the electrical equipment based on the pre-trained deep autoencoder Gaussian mixture model and the first eigenvector.

[0065] Exemplarily, a deep autoencoder Gaussian mixture model (DAGMM) is used as a fault detection model to analyze the first feature vector obtained by feature fusion of multi-source data to determine the real-time fault state of the electrical equipment. The real-time fault state is the state of the electrical equipment being faulty or operating normally at the current moment.

[0066] It can be understood that DAGMM is an unsupervised deep learning model. When negative samples are scarce, it can also learn the intrinsic representation of data in positive samples and identify abnormal points that are different from normal patterns. Secondly, DAGMM uses deep autoencoders for feature extraction. Deep autoencoders compress data into low-dimensional space through a series of hidden layers and try to reconstruct the original data from the low-dimensional representation. In this process, the deep autoencoder can learn the deep features of the data, which can be used to effectively distinguish normal data from abnormal data. Therefore, DAGMM does not need to explicitly mark abnormal data to train the model. Moreover, DAGMM combines the idea of ​​Gaussian mixture model (GMM). It simulates the behavior of GMM through the estimation network, that is, it estimates the probability of samples belonging to each Gaussian distribution model. This simulation method allows DAGMM to utilize the powerful modeling ability of GMM while avoiding the singular matrix problem that may occur in the actual solution process of GMM. During the model training process, the parameters of the estimation network will be continuously updated, thereby improving the anomaly detection performance of the model. Therefore, the unsupervised learning method, deep autoencoder feature extraction, and Gaussian mixture model simulation characteristics of DAGMM enable it to ensure high fault detection accuracy even in the absence of negative sample data.

[0067] For example, Figure 5 Schematic diagram of the structure of the deep autoencoder Gaussian mixture model provided in the embodiment of the present application. Figure 5 As shown, the deep autoencoder Gaussian mixture model includes a compression network and an estimation network. The compression network includes an encoder and a decoder, and the estimation network includes a perceptron and an activation function. The compression network compresses and reduces the dimension of the input first feature vector and reconstructs the information through the encoder and the decoder, and prevents the loss caused by dimensionality reduction by calculating the information loss. The estimation network further extracts features from the compressed features and the vector constructed by the reconstruction loss through the perceptron, and then calculates the probability value through the activation function. The probability value can replace the membership of the current data calculated in the E step of the EM algorithm to each Gaussian distribution model, so the fault state of the electrical equipment at the current moment can be determined based on the probability value.

[0068] Accordingly, Figure 6 1 is a flow chart of determining the real-time fault status of electrical equipment based on a deep self-encoding Gaussian mixture model provided by an embodiment of the present application. Figure 6 As shown, the step of determining the real-time fault state of the electrical equipment based on the deep self-encoding Gaussian mixture model specifically includes S1401-S1403:

[0069] S1401. Input the first feature vector into an encoder to obtain a compressed feature output by the encoder, and input the compressed feature into a decoder to obtain a reconstructed feature output by the decoder.

[0070] refer to Figure 5 , the first feature vector x is input into the encoder, and the encoder compresses and reduces the dimension of the first feature vector x to obtain the compressed feature z c , compression feature z c The calculation expression is z c =h(x,θ e ), where h() represents the encoder, θ e are the model parameters of the encoder.

[0071] Afterwards, the compressed feature z c Input decoder, the decoder is based on the compressed feature z c Perform information reconstruction to obtain the reconstructed feature x ′ , reconstruct feature x ′ The calculation expression is x ′ =g(z c ,θ d ), where g() represents the decoder, θ d are the model parameters of the decoder.

[0072] S1402: Determine a reconstruction error based on the reconstruction feature and the first feature vector, concatenate the reconstruction error and the compression feature, and input the concatenated features into a perceptron to obtain a second feature vector output by the perceptron.

[0073] refer to Figure 5 , the compression network calculates the reconstructed feature x based on the error calculation function ′ The error between the first eigenvector x and the reconstruction error z r , reconstruction error z r The calculation expression is z r =f(x,x ′ ), where f() is the error calculation function. Then, the reconstruction error z r and the compression feature z c The splicing feature z = [z c ,z r ]. The concatenated feature z is input into the perceptron, which further extracts features and outputs the second feature vector p. The calculation expression of the second feature vector p is p = MLN (z, θ m ), MLN() represents the perceptron, γ m are the model parameters of the perceptron.

[0074] S1403. Calculate the fault probability value through the activation function and the second eigenvector, and determine the real-time fault state of the electrical equipment according to the fault probability value.

[0075] refer to Figure 5 , substitute the second eigenvector p into the activation function to calculate the fault probability value Failure probability value The calculation expression is Among them, softmax() is the activation function.

[0076] After calculating the fault probability value, the fault probability value is compared with the preset probability threshold. When the fault probability value is higher than the preset probability threshold, it is determined that the electrical equipment is in a fault state; when the fault probability value is lower than or equal to the preset probability threshold, it is determined that the electrical equipment is in a normal operating state. The preset probability threshold can be set according to actual needs.

[0077] This embodiment can simulate the behavior of the GMM model through the DAGMM model, thereby realizing the identification of abnormal data based on data distribution, effectively fitting complex distribution shapes, and identifying abnormal data by estimating the membership of the data. Therefore, the advantage of the DAGMM model in abnormal data identification is that it can effectively capture the multi-peak characteristics and complex structure of the data, making it perform well in processing multi-modal data and fault detection tasks under complex backgrounds, and can achieve accurate detection of the fault state of electrical equipment even in the absence of negative sample data.

[0078] It should be noted that the GMM model can also be used as a fault detection model, but the GMM model is a machine learning model. Its direct use requires an existing feature data set and then trains the GMM model, that is, first train a feature extraction model for extracting features of time series data and image data, and then train the GMM model. This two-stage training method is prone to produce local optimal solutions, resulting in deviations in detection results and affecting the accuracy of fault state detection. Therefore, this embodiment uses an end-to-end DAGMM model as a fault detection model, and trains the feature extraction model and the fault detection model at the same time to prevent the model from falling into a local optimal solution and improve the accuracy of the model.

[0079] Optional, Figure 7 1 is a flow chart of a training feature extraction model and a fault detection model provided in an embodiment of the present application. Figure 7 As shown, the steps of training the feature extraction model and the fault detection model specifically include S210-S240:

[0080] S210 , obtaining positive sample data, where the positive sample data includes time series samples, image samples, and single-point samples collected at the same time by various sensors deployed on the electrical equipment.

[0081] Exemplarily, positive sample data can be understood as the time series data, image data, and single point data collected at the same time by the corresponding sensors deployed when the electrical equipment is in normal operation. Among them, the time series samples, image samples, and single point samples are sample data in the form of time series, image, and single point, respectively. The electrical equipment remains in normal operation most of the time during operation. The time series data, image data, and single point data collected by the corresponding sensors deployed at multiple time points of the electrical equipment can be integrated to obtain multiple positive sample data.

[0082] S220, extracting a first sample feature of the time series sample through a combined network, extracting a second sample feature of the image sample through a convolutional neural network, and concatenating the first sample feature, the second sample feature, and the single-point sample to obtain a first sample feature vector.

[0083] Among them, the combined network is a feature extraction model composed of the Inception network and the SeNet network for extracting features from time series data, and the convolutional neural network is a feature extraction model for extracting features from image data.

[0084] Exemplarily, the time series samples are input into the combination network to obtain the first sample features output by the combination network, and the image samples are input into the convolutional neural network to obtain the second sample features output by the convolutional neural network. Afterwards, the first sample features and the second sample features are compressed into one-dimensional vectors through the maximum pooling layer and the fully connected layer, and the one-dimensional vector corresponding to the first sample feature and the one-dimensional vector corresponding to the second sample feature and the numerical value of the single point sample are concatenated to obtain the first sample feature vector.

[0085] S230: Input the first sample feature vector into a deep autoencoder Gaussian mixture model to obtain a sample detection result output by the deep autoencoder Gaussian mixture model.

[0086] Exemplarily, the first sample feature vector is input into a deep autoencoder Gaussian mixture model, the deep autoencoder Gaussian mixture model compresses the first sample feature vector through an encoder to obtain a corresponding sample compression feature, the sample compression feature is reconstructed through a decoder to obtain a corresponding sample reconstruction feature, the sample reconstruction error between the sample reconstruction feature and the first sample feature vector is calculated based on an error calculation function, and the sample reconstruction error and the sample compression feature are concatenated to obtain a sample concatenation feature. The sample concatenation feature is input into a perceptron, the perceptron extracts features from the sample concatenation feature to obtain a fourth sample feature vector, the fourth sample feature vector is substituted into an activation function to obtain a fault probability value, and the fault probability value is used as a sample detection result.

[0087] S240. Calculate a first loss value of the sample detection result through a preset first loss function, and optimize the model parameters of the combined network, the convolutional neural network, and the deep autoencoder Gaussian mixture model through the first loss value. The first loss function includes reconstruction error, likelihood function loss, and regularization term.

[0088] Exemplarily, the expression of the first loss function is as follows:

[0089]

[0090] Among them, J(θ e ,θ d ,θ m ) is the first loss function, is the reconstruction error in the first loss function, is the likelihood function loss in the first loss function, θ is the regularization term in the first loss function. The regularization term is to prevent the values ​​on the diagonal of the covariance matrix in the GMM from becoming 0 and causing the matrix to be irreversible. e ,θ d ,θ m are the model parameters of the encoder, decoder and perceptron in the deep autoencoder Gaussian mixture model, respectively, i is the first sample feature vector, x′ i Reconstruct features for samples, z i is the sample splicing feature, E(z i ) is the opposite of the maximum likelihood function value, is the covariance of the Gaussian model, is the probability value calculated by the activation function, N is the number of samples, K is the number of Gaussian models, and d is the covariance matrix dimension of the Gaussian model. and are the mixed probability value, mean and variance of the kth Gaussian model respectively, and λ1 and λ2 are penalty parameters.

[0091] After calculating the first loss value of the sample detection result through the first loss function, the model parameters of the encoder, decoder and perceptron in the deep autoencoder Gaussian mixture model can be reversely optimized by the first loss value, as well as the model parameters of the combined network and the convolutional neural network, and the model parameters of the maximum pooling layer and the fully connected layer set between the feature extraction model and the fault detection model. When the model parameters of the combined network, the convolutional neural network and the deep autoencoder Gaussian mixture model reach the convergence condition, or the number of training times reaches the upper limit, it is confirmed that the combined network, the convolutional neural network and the deep autoencoder Gaussian mixture model have completed training.

[0092] Furthermore, after the feature extraction model and the fault detection model are trained using positive sample data, the model performance of the feature extraction model and the fault detection model can be tested using negative sample data in the test phase. When the model performance is poor, the model can be retrained using positive sample data to ensure that the model used to actually detect equipment faults can accurately identify the equipment fault problem. Optionally, the specific implementation process of testing the feature extraction model and the fault detection model based on negative sample data is as follows: obtaining negative sample data, the negative sample data includes time series samples, image samples, and single point samples collected by various sensors deployed by the electrical equipment at the same time; testing the combined network, convolutional neural network, and deep autoencoder Gaussian mixture model based on the negative sample data to obtain the test accuracy of the negative sample data; when the test accuracy is lower than the preset accuracy threshold, retraining the combined network, convolutional neural network, and deep autoencoder Gaussian mixture model based on the positive sample data. Wherein, the negative sample data is the time series data, image data, and single point data collected by various sensors deployed at the same time when the electrical equipment is in a fault state.

[0093] For example, the fifth sample feature of the time series sample in the negative sample data can be extracted through the combined network, the sixth sample feature of the image sample in the negative sample data can be extracted through the convolutional neural network, and the fifth sample feature, the sixth sample feature and the single point sample in the negative sample data can be spliced ​​to obtain the fifth sample feature vector. The fifth sample feature vector is input into the deep self-encoding Gaussian mixture model, the deep self-encoding Gaussian mixture model compresses the fifth sample feature vector through the encoder to obtain the corresponding sample compression feature, the decoder reconstructs the sample compression feature to obtain the corresponding sample reconstruction feature, and the sample reconstruction error between the sample reconstruction feature and the fifth sample feature vector is calculated based on the error calculation function, and the sample reconstruction error and the sample compression feature are spliced ​​to obtain the sample splicing feature. The sample splicing feature is input into the perceptron, the perceptron extracts the sample splicing feature to obtain the sixth sample feature vector, and the sixth sample feature vector is substituted into the activation function to obtain the fault probability value of the negative sample data. If it is determined that the electrical equipment is in a normal state based on the fault probability value, it is determined that the current model is inaccurate in testing the negative sample data. If it is determined that the electrical equipment is in a faulty state based on the fault probability value, it is determined that the current model is accurate in testing the negative sample data. The proportion of the number of negative sample data with accurate tests to the total number of all negative sample data is determined as the test accuracy of the negative sample data. When the test accuracy is higher than or equal to the preset accuracy threshold, it indicates that the current model can accurately detect whether the electrical equipment is in a faulty state, so the current model can be put into use in the actual fault detection process; when the test accuracy is lower than the preset accuracy threshold, it indicates that the current model cannot accurately detect whether the electrical equipment is in a faulty state, so the model should be retrained with positive sample data to ensure that the model used for actual detection of equipment faults can accurately identify the equipment fault problem.

[0094] In one embodiment, the feature prediction model can also be used to predict the first feature vector at the next moment based on the first feature vector at the previous moment and the first feature vector at the current moment, and the deep self-encoding Gaussian mixture model can be used to analyze the first feature vector at the next moment to determine whether the electrical equipment is in a fault state or a normal operating state at the next moment, thereby realizing the prediction of the future fault state of the electrical equipment. Figure 8 1 is a flowchart for predicting the future fault state of electrical equipment provided by an embodiment of the present application. Figure 8 As shown, the step of predicting the future fault state of the electrical equipment specifically includes S1501-S1503:

[0095] S1501. Store the first feature vector in a feature library, where the feature library stores the first feature vector at the previous moment and the first feature vector at the current moment.

[0096] Exemplarily, extract the first feature of the time series data collected at the current moment, extract the second feature of the image data collected at the current moment, and fuse the first feature, the second feature, and the single-point data collected at the current moment to obtain the first feature vector at the current moment. Input the first feature vector at the current moment into the feature library. Since the feature library only stores two first feature vectors, the earliest first feature vector stored will be automatically deleted. For example, at the current moment t i Before the first feature vector of is stored in the feature library, the feature library stores t i-1 The first eigenvector and time t i-2 The first eigenvector at the current time t i When the first eigenvector of is stored in the feature library, t i-2 The first feature vector at time t is deleted from the feature library so that the feature library stores the current time t i The first eigenvector and t i-1 The first eigenvector at time.

[0097] S1502: Input the first feature vector combination stored in the feature library into a pre-trained gated recurrent network to obtain the first feature vector output by the gated recurrent network at the next moment.

[0098] Among them, the first feature vector combination includes the current time t i The first eigenvector and t i-1 The first feature vector at the moment t is used as the first feature vector at the moment t. i The first eigenvector and t i-1 The first feature vector at a certain moment is input into the gated recurrent network, and the gated recurrent network is used to predict the first feature vector at the next moment.

[0099] S1503. Input the first feature vector at the next moment into a pre-trained deep autoencoder Gaussian mixture model to obtain the future fault state of the electrical equipment output by the deep autoencoder Gaussian mixture model.

[0100] Exemplarily, the first feature vector at the next moment is input into the deep self-encoding Gaussian mixture model, and the deep self-encoding Gaussian mixture model analyzes the first feature vector at the next moment to predict the fault state of the electrical equipment at the next moment. The future fault state is the state of the electrical equipment being faulty or operating normally at the next moment.

[0101] This embodiment uses a gated recurrent network to predict the first eigenvector at the next moment based on the first eigenvectors at the previous moment and the next moment, thereby predicting the future fault state of the electrical equipment based on the first eigenvector at the next moment through a deep autoencoder Gaussian mixture model, thereby achieving accurate prediction of the future fault state of the electrical equipment.

[0102] It should be noted that although the LSTM network can also predict the first eigenvector at a future moment, compared with the LSTM network, the gated recurrent network has a simpler structure and is easier to train, and its accuracy is similar to that of the LSTM network. Therefore, the gated recurrent network is selected as the feature prediction model to effectively improve the prediction efficiency of the future fault status of electrical equipment.

[0103] Since the gated recurrent network is a type of RNN network, it also needs to be trained in advance to have prediction capabilities. Optionally, the feature prediction model can be trained after the feature extraction model and the fault detection model are trained. The feature extraction model and the feature prediction model can be combined into one model, and the features output by the feature extraction model can be used to optimize the model parameters of the feature prediction model. Fig. 9 Flowchart of the training feature prediction model provided in the embodiment of the present application. Fig. 9 As shown, the steps of training the feature prediction model specifically include S310-S350:

[0104] S310 , obtaining a plurality of positive sample data, where the plurality of positive sample data are time series samples, image samples and single point samples collected by various sensors of the same electrical device at a plurality of consecutive moments.

[0105] Exemplarily, the time series data, image data and single point data collected by various sensors deployed corresponding to electrical equipment A within a certain time period can be used as multiple positive sample data, and one positive sample data includes the time series data, image data and single point data at the same time.

[0106] S320, extracting the third sample feature of the time series sample through the trained combined network, extracting the fourth sample feature of the image sample through the trained convolutional neural network, and concatenating the third sample feature, the fourth sample feature and the single point sample to obtain a second sample feature vector.

[0107] Exemplarily, step S320 may refer to step S220.

[0108] S330: Store the second sample feature vector in a feature library, input the second sample feature vector stored in the feature library into a gated recurrent network, and obtain a third sample feature vector output by the gated recurrent network.

[0109] For example, T k After the second sample feature vector at time is stored in the feature library, the feature library stores T k Moment and T k-1 The second sample feature vector at time T k Moment and T k-1The second sample feature vector at time T is input into the gated recurrent network to obtain the third sample feature vector output by the gated recurrent network, and the third sample feature vector can be regarded as T predicted by the gated recurrent network. k+1 The second sample feature vector at time instant.

[0110] S340, calculating a second loss value by using a preset second loss function and a third sample feature vector, and optimizing model parameters of the gated recurrent network by using the second loss value.

[0111] For example, the sensors deployed in the electrical equipment are obtained in the positive sample data at T k+1 The time series samples, image samples and single point samples collected at all times are extracted through the combined network k+1 The third sample feature of the time series sample collected at the moment is extracted by convolutional neural network T k+1 The fourth sample feature of the image sample collected at time T k+1 The third sample feature, the fourth sample feature and the single point sample at time T are fused to obtain k+1 The second sample feature vector at time T k+1 The second sample feature vector and the third sample feature vector at time (T predicted by the gated recurrent network) k+1 The second sample feature vector at the time (the second sample feature vector at the time) is substituted into the second loss function to calculate the second loss value, and the model parameters of the gated recurrent network are reversely optimized by the second loss value. The second loss function can be Smooth L1 Loss.

[0112] When the model parameters of the gated recurrent network reach the convergence condition or the number of training times reaches the upper limit, it is confirmed that the gated recurrent network has completed training.

[0113] In one embodiment, the feature extraction model, the fault detection model and the feature prediction model are combined into an end-to-end state analysis model, and the state analysis model is used to detect the real-time fault state of the electrical equipment and predict the future fault state of the electrical equipment. Fig.10 Schematic diagram of the state analysis model provided in the embodiment of the present application. Fig.10As shown, after acquiring image data, time series data and single-point data, the image data is input into CNN (convolutional neural network) to obtain the first feature, the time series data is converted into frequency domain distribution data, the time series data and frequency domain distribution data are converted into a two-dimensional matrix in the time-frequency domain, the two-dimensional matrix in the time-frequency domain is input into a combined network composed of an Inception network and a SeNet network to obtain the second feature, the first feature and the second feature are respectively input into the maximum pooling layer (Max Pool) and then into the fully connected layer (FC) to obtain the corresponding one-dimensional vector, the one-dimensional vector corresponding to the first feature, the one-dimensional vector corresponding to the second feature and the single-point value of the single-point data are concatenated to obtain the first feature vector, the first feature vector is input into a deep autoencoder Gaussian mixture model (DAGMM) and stored in a feature library, and the deep autoencoder Gaussian mixture model determines the real-time fault status of the electrical equipment based on the first feature vector. The gated recurrent network (GRU) obtains the first feature vectors of the current moment and the previous moment from the feature library, predicts the first feature vector of the next moment based on the first feature vectors of the current moment and the previous moment, and inputs the first feature vector of the next moment into the deep autoencoder Gaussian mixture model. The deep autoencoder Gaussian mixture model predicts the future fault state of the electrical equipment based on the first feature vector of the next moment.

[0114] In summary, the fault detection method of electrical equipment provided by the embodiment of the present application obtains the time series data, image data and single point data collected by each sensor deployed by the electrical equipment at the current moment; extracts the first feature of the time series data, extracts the second feature of the image data; fuses the first feature, the second feature and the single point data to obtain the first feature vector; and determines the real-time fault state of the electrical equipment based on the pre-trained deep self-encoding Gaussian mixture model and the first feature vector. Through the above technical means, the time series data, image data and single point data collected by various types of sensors are obtained as multi-source data for detecting the fault state of the electrical equipment, so as to comprehensively detect the fault state of the electrical equipment and improve the detection accuracy of the fault state. The features of the multi-source data are fused, and the abnormal features identified by the deep self-encoding Gaussian mixture model in the fused features are used to determine the real-time fault state of the electrical equipment based on the abnormal features. In the absence of negative sample data, the deep self-encoding Gaussian mixture model can also accurately detect the fault state of the electrical equipment, and performs well in the fault detection task, reducing the dependence on negative sample data, solving the problem that the model performance of the fault detection model is affected by insufficient negative sample data in the prior art, and improving the accuracy and reliability of fault detection.

[0115] Based on the above embodiments, Fig.11 A schematic diagram of the structure of a fault detection device for electrical equipment provided in an embodiment of the present application. Fig.11The fault detection device for electrical equipment provided in this embodiment specifically includes: a multi-source data acquisition module 41, a feature extraction module 42, a feature fusion module 43 and a fault detection module 44.

[0116] The multi-source data acquisition module 41 is configured to acquire time series data, image data and single point data collected by various sensors deployed on the electrical equipment at the current moment;

[0117] a feature extraction module 42 configured to extract a first feature of the time series data and a second feature of the image data;

[0118] A feature fusion module 43 is configured to fuse the first feature, the second feature and the single point data to obtain a first feature vector;

[0119] The fault detection module 44 is configured to determine the real-time fault status of the electrical equipment based on a pre-trained deep autoencoder Gaussian mixture model and a first eigenvector.

[0120] On the basis of the above embodiment, the feature extraction module 42 includes: a two-dimensional matrix conversion unit, configured to perform Fourier transform on the time series data to obtain frequency domain distribution data, and convert the time series data and the frequency domain distribution data into a corresponding two-dimensional matrix; a two-dimensional matrix splicing unit, configured to splice the two-dimensional matrix corresponding to the time series data and the two-dimensional matrix corresponding to the frequency domain distribution data to obtain a time-frequency domain two-dimensional matrix; a multi-dimensional feature extraction unit, configured to perform feature extraction on the time-frequency domain two-dimensional matrix through the Inception network to obtain a multi-dimensional feature matrix; a time series feature extraction unit, configured to perform feature weighted fusion processing on the multi-dimensional feature matrix through the SeNet network to obtain the first feature of the time series data.

[0121] Based on the above embodiments, the deep autoencoding Gaussian mixture model includes a compression network and an estimation network, the compression network includes an encoder and a decoder, and the estimation network includes a perceptron and an activation function; accordingly, the fault detection module 44 includes: an autoencoding unit, configured to input the first feature vector into the encoder to obtain the compression feature output by the encoder, and input the compression feature into the decoder to obtain the reconstruction feature output by the decoder; a perceptron unit, configured to determine the reconstruction error based on the reconstruction feature and the first feature vector, splice the reconstruction error and the compression feature and input them into the perceptron to obtain the second feature vector output by the perceptron; a real-time fault detection unit, configured to calculate the fault probability value through the activation function and the second feature vector, and determine the real-time fault state of the electrical equipment according to the fault probability value.

[0122] On the basis of the above embodiments, the fault detection device of the electrical equipment also includes a first training module, which includes: a first sample acquisition unit, configured to acquire positive sample data, the positive sample data including time series samples, image samples and single point samples collected at the same time by various sensors deployed by the electrical equipment; a first sample feature extraction unit, configured to extract the first sample features of the time series samples through the combined network, extract the second sample features of the image samples through the convolutional neural network, and splice the first sample features, the second sample features and the single point samples to obtain a first sample feature vector; a first sample detection unit, configured to input the first sample feature vector into a deep autoencoder Gaussian mixture model to obtain a sample detection result output by the deep autoencoder Gaussian mixture model; a first parameter optimization unit, configured to calculate a first loss value of the sample detection result through a preset first loss function, and optimize the model parameters of the combined network, the convolutional neural network and the deep autoencoder Gaussian mixture model through the first loss value, and the first loss function includes reconstruction error, likelihood function loss and regularization term.

[0123] On the basis of the above embodiments, the fault detection device of the electrical equipment also includes a testing module, which includes: a second sample acquisition unit, configured to acquire negative sample data after optimizing the model parameters of the combined network, the convolutional neural network and the deep autoencoding Gaussian mixture model through a first loss value, the negative sample data including time series samples, image samples and single point samples collected by various sensors deployed by the electrical equipment at the same time; a sample testing unit, configured to test the combined network, the convolutional neural network and the deep autoencoding Gaussian mixture model based on the negative sample data to obtain the test accuracy of the negative sample data; a retraining unit, configured to retrain the combined network, the convolutional neural network and the deep autoencoding Gaussian mixture model based on positive sample data when the test accuracy is lower than a preset accuracy threshold.

[0124] On the basis of the above embodiments, the fault detection device of the electrical equipment also includes: a fault prediction module, the fault prediction module includes: a first feature storage unit, configured to store the first feature vector into a feature library, the feature library stores the first feature vector at the previous moment and the first feature vector at the current moment; a feature prediction unit, configured to input the first feature vector combination stored in the feature library into a pre-trained gated recurrent network to obtain the first feature vector at the next moment output by the gated recurrent network; a future fault prediction unit, configured to input the first feature vector at the next moment into a pre-trained deep autoencoder Gaussian mixture model to obtain the future fault state of the electrical equipment output by the deep autoencoder Gaussian mixture model.

[0125] On the basis of the above-mentioned embodiment, the fault detection device of the electrical equipment also includes: a second training module, the second training module includes: a third sample acquisition unit, configured to acquire multiple positive sample data, the multiple positive sample data are time series samples, image samples and single-point samples collected by various sensors of the same electrical equipment at multiple consecutive moments; a second sample feature extraction unit, configured to extract the third sample feature of the time series sample through the trained combined network, extract the fourth sample feature of the image sample through the trained convolutional neural network, and splice the third sample feature, the fourth sample feature and the single-point sample to obtain a second sample feature vector; a sample feature prediction unit, configured to store the second sample feature vector into a feature library, input the second sample feature vector stored in the feature library into the gated recurrent network, and obtain the third sample feature vector output by the gated recurrent network; a third parameter optimization unit, configured to calculate the second loss value through a preset second loss function and the third sample feature vector, and optimize the model parameters of the gated recurrent network through the second loss value.

[0126] As mentioned above, the fault detection device for electrical equipment provided by the embodiment of the present application obtains the time series data, image data and single point data collected by each sensor deployed by the electrical equipment at the current moment; extracts the first feature of the time series data, extracts the second feature of the image data; fuses the first feature, the second feature and the single point data to obtain the first feature vector; and determines the real-time fault state of the electrical equipment based on the pre-trained deep self-encoding Gaussian mixture model and the first feature vector. Through the above technical means, the time series data, image data and single point data collected by various types of sensors are obtained as multi-source data for detecting the fault state of the electrical equipment, so as to comprehensively detect the fault state of the electrical equipment and improve the detection accuracy of the fault state. The features of the multi-source data are fused, and the abnormal features identified by the deep self-encoding Gaussian mixture model in the fused features are used to determine the real-time fault state of the electrical equipment based on the abnormal features. In the absence of negative sample data, the deep self-encoding Gaussian mixture model can also accurately detect the fault state of the electrical equipment, and performs well in the fault detection task, reducing the dependence on negative sample data, solving the problem that the model performance of the fault detection model is affected by insufficient negative sample data in the prior art, and improving the accuracy and reliability of fault detection.

[0127] The fault detection device for electrical equipment provided in the embodiment of the present application can be used to execute the fault detection method for electrical equipment provided in the above-mentioned embodiment, and has corresponding functions and beneficial effects.

[0128] Fig.12 is a schematic diagram of a fault detection device for an electrical device provided in an embodiment of the present application, with reference to Fig.12The fault detection device of the electrical equipment includes: a processor 51, a memory 52, a communication device 53, an input device 54 and an output device 55. The number of processors 51 in the fault detection device of the electrical equipment can be one or more, and the number of memories 52 in the fault detection device of the electrical equipment can be one or more. The processor 51, the memory 52, the communication device 55, the input device 54 and the output device 55 of the fault detection device of the electrical equipment can be connected through a bus or other methods.

[0129] The memory 52, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the fault detection method of the electrical equipment of any embodiment of the present application (for example, multi-source data acquisition module 41, feature extraction module 42, feature fusion module 43 and fault detection module 44). The memory 52 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the device, etc. In addition, the memory 52 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0130] The communication device 53 is used for data transmission.

[0131] The processor 51 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 52, that is, implements the above-mentioned fault detection method of the electrical device.

[0132] The input device 54 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the device. The output device 55 may include a display device such as a display screen.

[0133] The fault detection device for electrical equipment provided above can be used to execute the fault detection method for electrical equipment provided in the above embodiment, and has corresponding functions and beneficial effects.

[0134] An embodiment of the present application also provides a storage medium containing computer executable instructions, which, when executed by a computer processor, are used to execute a fault detection method for electrical equipment. The fault detection method for electrical equipment includes: obtaining time series data, image data, and single-point data collected by each sensor deployed on the electrical equipment at the current moment; extracting a first feature of the time series data, and extracting a second feature of the image data; fusing the first feature, the second feature, and the single-point data to obtain a first feature vector; and determining the real-time fault status of the electrical equipment based on a pre-trained deep autoencoder Gaussian mixture model and the first feature vector.

[0135] Storage medium - any of various types of memory devices or storage devices. The term "storage medium" is intended to include: installation media, such as CD-ROM, floppy disk or tape device; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media (such as hard disk or optical storage); registers or other similar types of memory elements, etc. Storage media may also include other types of memory or combinations thereof. In addition, the storage medium may be located in the first computer system in which the program is executed, or may be located in a different second computer system, which is connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term "storage medium" may include two or more storage media residing in different locations (for example, in different computer systems connected by a network). The storage medium may store program instructions (for example, embodied as a computer program) that can be executed by one or more processors.

[0136] Of course, the storage medium containing computer executable instructions provided in the embodiment of the present application is not limited to the above-mentioned fault detection method for electrical equipment, and the computer executable instructions can also execute related operations in the fault detection method for electrical equipment provided in any embodiment of the present application.

[0137] The fault detection device of electrical equipment, the fault detection system of electrical equipment, the storage medium and the fault detection device of electrical equipment provided in the above embodiments can execute the fault detection method of electrical equipment provided in any embodiment of the present application. For technical details not described in detail in the above embodiments, please refer to the fault detection method of electrical equipment provided in any embodiment of the present application.

[0138] The above are only preferred embodiments of the present application and the technical principles used. The present application is not limited to the specific embodiments herein, and various obvious changes, readjustments and substitutions that can be made by those skilled in the art will not deviate from the protection scope of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may also include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the claims.

Claims

1. A method for detecting a fault of an electrical device, characterized in that: include: Obtain the time series data, image data and single point data collected by each sensor deployed on the electrical equipment at the current moment; Extracting a first feature of the time series data and extracting a second feature of the image data; Fusing the first feature, the second feature and the single point data to obtain a first feature vector; Based on a pre-trained deep autoencoder Gaussian mixture model and the first feature vector, the real-time fault status of the electrical equipment is determined.

2. The fault detection method of electrical equipment according to claim 1, characterized in that: The extracting the first feature of the time series data includes: Performing Fourier transformation on the time series data to obtain frequency domain distribution data, and converting the time series data and the frequency domain distribution data into corresponding two-dimensional matrices; The two-dimensional matrix corresponding to the time series data and the two-dimensional matrix corresponding to the frequency domain distribution data are concatenated to obtain a two-dimensional matrix in the time-frequency domain; Extracting features from the two-dimensional matrix in the time-frequency domain by using an Inception network to obtain a multi-dimensional feature matrix; The multi-dimensional feature matrix is ​​subjected to feature weighted fusion processing through the SeNet network to obtain the first feature of the time series data.

3. The fault detection method of electrical equipment according to claim 1, characterized in that: The deep autoencoder Gaussian mixture model includes a compression network and an estimation network, the compression network includes an encoder and a decoder, and the estimation network includes a perceptron and an activation function; Accordingly, the method of determining the real-time fault state of the electrical device based on the pre-trained deep autoencoder Gaussian mixture model and the first eigenvector at the current moment includes: Inputting the first feature vector into the encoder to obtain a compressed feature output by the encoder, and inputting the compressed feature into the decoder to obtain a reconstructed feature output by the decoder; Determine a reconstruction error based on the reconstruction feature and the first feature vector, concatenate the reconstruction error and the compression feature, and input the concatenated features into the perceptron to obtain a second feature vector output by the perceptron; A fault probability value is calculated by using the activation function and the second eigenvector, and a real-time fault state of the electrical equipment is determined according to the fault probability value.

4. The fault detection method of electrical equipment according to claim 1, characterized in that: The method further comprises: Acquire positive sample data, where the positive sample data includes time series samples, image samples, and single point samples collected at the same time by various sensors deployed on the electrical equipment; Extracting a first sample feature of the time series sample through a combined network, extracting a second sample feature of the image sample through a convolutional neural network, and concatenating the first sample feature, the second sample feature and the single point sample to obtain a first sample feature vector; Inputting the first sample feature vector into a deep autoencoder Gaussian mixture model to obtain a sample detection result output by the deep autoencoder Gaussian mixture model; A first loss value of the sample detection result is calculated by a preset first loss function, and the model parameters of the combined network, the convolutional neural network and the deep autoencoder Gaussian mixture model are optimized by the first loss value. The first loss function includes reconstruction error, likelihood function loss and regularization term.

5. The fault detection method of electrical equipment according to claim 4, characterized in that: After optimizing the model parameters of the combined network, the convolutional neural network, and the deep autoencoder Gaussian mixture model by the first loss value, the method further includes: Acquire negative sample data, where the negative sample data includes time series samples, image samples, and single point samples collected at the same time by various sensors deployed on the electrical equipment; Testing the combined network, the convolutional neural network, and the deep autoencoder Gaussian mixture model based on the negative sample data to obtain a test accuracy rate of the negative sample data; When the test accuracy is lower than a preset accuracy threshold, the combined network, the convolutional neural network and the deep autoencoder Gaussian mixture model are retrained based on positive sample data.

6. The fault detection method of electrical equipment according to claim 1, characterized in that: After fusing the first feature, the second feature and the single point data to obtain a first feature vector, the method further includes: Storing the first feature vector in a feature library, the feature library storing the first feature vector at the previous moment and the first feature vector at the current moment; Inputting the first feature vector combination stored in the feature library into a pre-trained gated recurrent network to obtain the first feature vector output by the gated recurrent network at the next moment; The first feature vector at the next moment is input into a pre-trained deep autoencoder Gaussian mixture model to obtain the future fault state of the electrical equipment output by the deep autoencoder Gaussian mixture model.

7. The fault detection method of electrical equipment according to claim 6, further comprising: Acquire a plurality of positive sample data, wherein the plurality of positive sample data are time series samples, image samples and single point samples collected by various sensors of the same electrical device at a plurality of consecutive moments; Extracting a third sample feature of the time series sample through the trained combined network, extracting a fourth sample feature of the image sample through the trained convolutional neural network, and concatenating the third sample feature, the fourth sample feature and the single point sample to obtain a second sample feature vector; The second sample feature vector is stored in a feature library, and the second sample feature vector stored in the feature library is input into a gated recurrent network to obtain a third sample feature vector output by the gated recurrent network; A second loss value is calculated using a preset second loss function and the third sample feature vector, and the model parameters of the gated recurrent network are optimized using the second loss value.

8. A fault detection device for electrical equipment, characterized in that: include: A multi-source data acquisition module is configured to acquire time series data, image data and single point data collected by various sensors deployed on the electrical equipment at the current moment; a feature extraction module, configured to extract a first feature of the time series data and a second feature of the image data; a feature fusion module, configured to fuse the first feature, the second feature and the single point data to obtain a first feature vector; The fault detection module is configured to determine the real-time fault status of the electrical equipment based on a pre-trained deep autoencoder Gaussian mixture model and the first feature vector.

9. A fault detection device for electrical equipment, characterized in that: include: one or more processors; A storage device stores one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the fault detection method for electrical equipment as described in any one of claims 1-7.

10. A storage medium containing computer executable instructions, characterized in that: The computer executable instructions are used to execute the fault detection method of the electrical equipment as described in any one of claims 1 to 7 when executed by a computer processor.

Images