Multi-sensor fusion equipment multi-branch semi-supervised fault diagnosis method and system

Abstract

The invention discloses a multi-sensor fusion equipment multi-branch semi-supervised fault diagnosis method and system, and the method comprises the steps: obtaining transformer fault data: employing a sensor network composed of a plurality of sensors, and synchronously collecting vibration signals of different positions of the surface of a real transformer under different fault defects; based on the transformer fault data, constructing a graph data set, dividing the graph data set into a training set and a test set, and dividing the training set into labeled data and unlabeled data; constructing a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model comprises a supervised learning module and an unsupervised learning module which are arranged in parallel, and acquiring a vibration signal of a to-be-diagnosed transformer, converting the vibration signal of the to-be-diagnosed transformer into graph structure data, and inputting the graph structure data into the trained supervised learning module to obtain a fault identification result.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a multi-branch semi-supervised fault diagnosis method and system for multi-sensor fusion equipment. Background Art

[0002] As a key device for realizing power conversion and power transmission in the power system, the smooth operation of power transformers is a key factor in ensuring the safety and stability of the power grid. However, due to the harsh operating environment and long-term load operation, various faults are prone to occur, threatening the safety and stability of the power grid, and causing loss of life and property. Therefore, in order to ensure the safety and reliability of power grid operation and reduce the operation and maintenance costs of transformers, it is of great practical significance and application value to carry out research on transformer fault diagnosis methods for mechanical failures that are prone to occur during operation.

[0003] At present, in the field of transformer fault detection and diagnosis, the commonly used fault diagnosis methods mainly include frequency response analysis, dissolved gas in oil, vibration signal analysis, etc. Among them, the frequency response analysis method is to apply a sine wave voltage of different frequencies to the winding at one end of the transformer, and measure the output response of the winding at the other end to form a frequency response curve, and compare it with the frequency response curve of the normal winding to determine whether the transformer winding to be tested has deformation and other faults. However, this method belongs to an offline detection method, which can only detect fault information such as winding deformation after the equipment fails, and cannot achieve real-time detection of transformer faults. The dissolved gas in oil method measures the components and content of dissolved gas in the transformer oil, and judges the early faults and potential threats of the transformer according to the relevant ratio standard. However, this method faces the problem that the boundary limit of the threshold judgment is too absolute, and it is only applicable to oil-immersed transformers, and cannot handle the fault diagnosis task of dry transformers. The vibration analysis method is to install acceleration sensors at different positions of the transformer to collect the vibration signals of the transformer during operation, and perform real-time detection and diagnosis of the working status and potential faults of the transformer based on the vibration signals. Different from the above two diagnostic methods, this method is a non-electrical parameter detection method. There is no electrical connection between this method and the transformer to be tested, which will not affect the normal operation of the transformer. It is easy to install, and the diagnostic results are accurate and reliable. It has important research value and practical significance.

[0004] At present, transformer fault diagnosis based on vibration signal analysis has received widespread attention and has made some progress. At the same time, due to the powerful feature extraction and efficient data processing capabilities of deep learning methods, many transformer fault diagnosis methods based on deep learning and vibration signal analysis have emerged. However, traditional fault diagnosis methods based on deep learning and vibration signal analysis still have the following limitations:

[0005] 1) Traditional fault diagnosis methods usually adopt supervised learning strategies, which leads to their heavy reliance on a large amount of labeled data to train the model, which means that it usually takes expensive time and labor costs to label the collected data, which seriously limits its further development and practical application.

[0006] 2) Traditional fault diagnosis methods still mainly rely on data collected by a single sensor. However, due to the complexity of the transformer system itself, relying solely on data collected by a single sensor for fault diagnosis often cannot fully reflect the status information of the transformer and is prone to misjudgment. Summary of the invention

[0007] In order to solve the deficiencies of the prior art, the present invention provides a multi-branch semi-supervised fault diagnosis method and system for multi-sensor fusion equipment; first, the sensor network is used to collect the vibration signals of the box surface at different positions of the transformer under different fault defects, and then the vibration signals between different sensors are collected to construct a graph data set according to the adaptive radius graph, and then a multi-branch semi-supervised fault diagnosis model is built, and the graph data set is used to complete the model training and evaluation, and finally the online monitoring and fault diagnosis of the transformer is realized. It can effectively extract and fuse the fault information of multiple sensors, and make full use of unlabeled data, which significantly improves the effect and accuracy of transformer fault identification.

[0008] On the one hand, a multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment is provided, comprising:

[0009] Obtaining transformer fault data: Using a sensor network consisting of multiple sensors, synchronously collect vibration signals at different locations on the surface of the real transformer under different fault defects;

[0010] Based on the transformer fault data, a graph data set is constructed, the graph data set is divided into a training set and a test set, and the training set is divided into labeled data and unlabeled data;

[0011] Constructing a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model includes: a supervised learning module and an unsupervised learning module in parallel; training the multi-branch semi-supervised fault diagnosis model, using labeled data to train the supervised learning module during the training of the model, and at the same time, using unlabeled data to train the unsupervised learning module, and sharing local network structure parameters of the supervised learning module and the unsupervised learning module during the training of the model;

[0012] The vibration signal of the transformer to be diagnosed is obtained, the vibration signal of the transformer to be diagnosed is converted into graph structure data, and the graph structure data is input into a trained supervised learning module to obtain a fault identification result.

[0013] On the other hand, a multi-branch semi-supervised fault diagnosis system for multi-sensor fusion equipment is provided, comprising:

[0014] An acquisition module is configured to: acquire transformer fault data: utilize a sensor network composed of multiple sensors to synchronously collect vibration signals at different positions on the surface of a real transformer under different fault defects;

[0015] A data set construction module is configured to: construct a graph data set based on the transformer fault data, divide the graph data set into a training set and a test set, and divide the training set into labeled data and unlabeled data;

[0016] A model building module is configured to: build a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model includes: a supervised learning module and an unsupervised learning module in parallel; train the multi-branch semi-supervised fault diagnosis model, use labeled data to train the supervised learning module during the model training process, and use unlabeled data to train the unsupervised learning module, and share local network structure parameters of the supervised learning module and the unsupervised learning module during the model training process;

[0017] The fault identification module is configured to: obtain the vibration signal of the transformer to be diagnosed, convert the vibration signal of the transformer to be diagnosed into graph structure data, input the graph structure data into the trained supervised learning module, and obtain the fault identification result.

[0018] On the other hand, there is also provided an electronic device, comprising:

[0019] a memory for non-transitory storage of computer-readable instructions; and

[0020] a processor for executing the computer readable instructions,

[0021] When the computer-readable instructions are executed by the processor, the method described in the first aspect is executed.

[0022] On the other hand, a storage medium is provided, which non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, the method described in the first aspect is executed.

[0023] On the other hand, a computer program product is provided, comprising a computer program, wherein the computer program is used to implement the method described in the first aspect when running on one or more processors.

[0024] The above technical solution has the following advantages or beneficial effects:

[0025] (1) A multi-level feature extraction and deep fusion module for multi-sensor information fusion is proposed. Through cross-connection and downsampling operations, multi-sensor features at different aggregation levels are extracted, thereby effectively avoiding the loss of fault features and improving the accuracy of model fault diagnosis. Compared with the traditional multi-sensor fusion module that extracts features for each single channel independently, the proposed module is more lightweight.

[0026] (2) By constructing a supervised learning module and an unsupervised learning module with local network structure and parameter sharing, and introducing classification loss, reconstruction loss and time-frequency contrast learning loss, the model can not only use the labeled data in the graph dataset for model training, but also make full use of the unlabeled data to assist model training, so that the model can achieve good training results even with limited labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0028] Figure 1 This is a flow chart of the transformer multi-branch semi-supervised fault diagnosis method based on multi-sensor information fusion.

[0029] Figure 2 Schematic diagram of the measurement point location of the IEPE acceleration sensor on the transformer.

[0030] Figure 3 This is the structural diagram of the supervised learning module.

[0031] Figure 4 This is the structural diagram of the multi-level feature extraction and deep fusion module.

[0032] Figure 5 This is the structural diagram of the classification module.

[0033] Figure 6 This is the structural diagram of the unsupervised learning module. DETAILED DESCRIPTION

[0034] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0035] Embodiment 1

[0036] This embodiment provides a multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment;

[0037] like Figure 1 As shown, a multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment includes:

[0038] S101: Obtaining transformer fault data: using a sensor network consisting of multiple sensors to synchronously collect vibration signals at different positions on the surface of a real transformer under different fault defects;

[0039] S102: constructing a graph data set based on the transformer fault data, dividing the graph data set into a training set and a test set, and dividing the training set into labeled data and unlabeled data;

[0040] S103: constructing a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model includes: a supervised learning module and an unsupervised learning module in parallel; training the multi-branch semi-supervised fault diagnosis model, using labeled data to train the supervised learning module during the training process of the model, and at the same time, using unlabeled data to train the unsupervised learning module, and the local network structure parameters of the supervised learning module and the unsupervised learning module are shared during the training process of the model;

[0041] S104: Obtain a vibration signal of the transformer to be diagnosed, convert the vibration signal of the transformer to be diagnosed into graph structure data, input the graph structure data into a trained supervised learning module, and obtain a fault identification result.

[0042] Furthermore, the S101: obtaining transformer fault data: using a sensor network composed of multiple sensors to synchronously collect vibration signals at different positions on the surface of a real transformer under different fault defects, including:

[0043] A sensor network consisting of M integrated electronic piezoelectric (IEPE) accelerometers is used to synchronously collect the surface vibration signals of the transformer box under normal operation, winding looseness, core looseness, and combined faults of winding and core looseness at a sampling frequency of 12.8kHz:

[0044]

[0045] Where, X represents a set of vibration signals of the transformer; x m represents a set of vibration signals collected by the mth sensor, x mn It represents the nth sampling point of the vibration signal collected by the mth sensor, 1≤m≤M, 1≤n≤N, M is the number of sensors, here is 12, N is the total number of sampling points of a group of signals, set to 1024, and 300 groups of vibration signals are collected for each fault type. The schematic diagram of the measurement point position of 12 IEPE acceleration sensors on the transformer is as follows: Figure 2 shown.

[0046] Furthermore, the S102: constructing a graph data set is to regard each sensor in the sensor network as a graph node, and use an adaptive radius graph algorithm based on local density to construct a connection relationship between the edges of the nodes, thereby constructing a graph data set.

[0047] It should be understood that a graph is a special data structure consisting of nodes and edges. The present invention abstracts the sensor network into a graph structure, and each sensor is regarded as a node in the graph. m As the node feature of the mth sensor node, the dimension is 1×1024, and the node features of 12 nodes form a node feature matrix, which can still be expressed as X∈R M×N As for the connection relationship between the edges of the nodes, it is determined according to the adaptive radius graph method based on local density proposed by the present invention.

[0048] Furthermore, the use of the local density-based adaptive radius graph algorithm to construct the connection relationship between the edges of the nodes includes:

[0049] (1-1): Use Euclidean distance to calculate the spatial distance d(x) between any two nodes in the graph i ,x j ), the calculation formula is as follows:

[0050]

[0051] Among them, 1≤i,j≤12 and i≠j.

[0052] (1-2): According to the K-nearest neighbor algorithm, find the k nearest neighbor nodes of each node to determine the neighbor space O of the current node, and the evaluation index is based on the Euclidean distance between nodes. Among them, 1≤k≤11, and the value of k can be flexibly selected according to the complexity and accuracy requirements of the actual model.

[0053] (1-3): Calculate the average Euclidean distance between each node and its k nearest neighbor nodes as the adaptive radius of the current node, so that the radius of the radius graph can be dynamically adjusted according to the change of local density. The radius graph assumes that the neighbor nodes within the radius have direct information transmission with this node, thereby constructing an edge connection relationship. For node i, the calculation formula for its adaptive radius is:

[0054]

[0055] Among them, O(i) represents the neighbor space of node i.

[0056] (1-4): The edge connection relationship between nodes is constructed based on the adaptive radius graph. The process is expressed as:

[0057]

[0058] Among them, A ij It is the constructed adjacency matrix, which together with the node feature matrix constitutes the graph structure.

[0059] Through the above method, a smaller radius can be used in node-dense areas and a larger radius can be used in node-sparse areas, so as to dynamically and flexibly adjust the connection relationship between nodes according to the local density changes of the data, avoiding the problem of the graph structure being too sparse or too dense due to a fixed radius.

[0060] Further, the constructed graph data set includes: a plurality of groups of graph data, each group of graph data includes a time domain graph data and a frequency domain graph data;

[0061] The time domain graph data is obtained by processing the vibration signals collected by each sensor using steps (1-1) to (1-4);

[0062] The frequency domain graph data is obtained by first performing fast Fourier transform on the vibration signal to obtain a frequency domain signal, calculating the amplitude of the frequency domain signal, and processing the amplitude of the frequency domain signal collected by each sensor using steps (1-1) to (1-4).

[0063] It should be understood that for each group of vibration signals, in addition to constructing the time domain graph data according to (1-1) to (1-4), the signal is fast Fourier transformed to obtain the frequency domain signal, and the amplitude of the frequency domain signal is calculated. The frequency domain graph data is constructed using the same method. The node feature matrix can be expressed as F∈R M×N In this way, 300 sets of vibration signals of each fault type can generate 300 sets of graph data, each of which contains a time domain graph data and a frequency domain graph data. The generated graph data set is divided into a training set and a test set in a ratio of 8:2.

[0064] Furthermore, the dividing of the training set into labeled data and unlabeled data is to perform masking on the labeled data in the training set according to a set ratio, and remove labels of signals in the set ratio to simulate unlabeled data.

[0065] Through the above operations, the independent vibration signals collected by each sensor in the sensor network can be converted into graph structure data with correlation relationships, so as to better utilize and fuse multi-sensor information for feature extraction and fault diagnosis.

[0066] Furthermore, if Figure 3 As shown, the supervised learning module includes:

[0067] a first time domain branch and a first frequency domain branch;

[0068] The first time domain branch inputs labeled time domain graph data;

[0069] The first frequency domain branch inputs labeled frequency domain image data;

[0070] The first time domain branch includes: a first multi-level feature extraction and depth fusion module, a first hidden layer and a first global average pooling layer connected in sequence;

[0071] The first frequency domain branch includes: a second multi-level feature extraction and depth fusion module, a second hidden layer, and a second global average pooling layer connected in sequence;

[0072] The output of the first global average pooling layer is connected to the input of the splicing unit;

[0073] The output of the second global average pooling layer is connected to the input of the splicing unit;

[0074] The output end of the splicing unit outputs a fused feature vector, which is input into the classification module to obtain a fault diagnosis prediction label.

[0075] Furthermore, if Figure 3 As shown, the supervised learning module includes the following working process:

[0076] After the labeled time domain graph data passes through the first multi-level feature extraction and deep fusion module, the time domain graph structure data that aggregates the features of sensor nodes at different levels is obtained;

[0077] After the labeled frequency domain image data passes through the second multi-level feature extraction and deep fusion module, the frequency domain image structure data that aggregates the features of sensor nodes at different levels is obtained;

[0078] Aggregate the temporal graph structure data of sensor node features at different levels, and after processing by the global average pooling layer, obtain the temporal fusion features;

[0079] Aggregate the frequency domain graph structure data of sensor node features at different levels, and after processing by the global average pooling layer, obtain the frequency domain fusion features;

[0080] The time domain fusion features and the frequency domain fusion features are spliced ​​to obtain the time-frequency fusion features;

[0081] The time-frequency fusion features are input into the classification module to obtain the predicted label.

[0082] The specific structure of the supervised learning module is as follows Figure 3 shown.

[0083] Furthermore, the internal structure of the first multi-level feature extraction and depth fusion module is consistent with that of the second multi-level feature extraction and depth fusion module, and the first multi-level feature extraction and depth fusion module includes:

[0084] The first graph convolution layer, the first batch normalization layer, the first activation function layer, the second graph convolution layer, the second batch normalization layer, the second activation function layer, the third graph convolution layer, the third batch normalization layer and the adder are connected in sequence; the input end of the first graph convolution layer is the input end of the first multi-level feature extraction and deep fusion module; the output end of the adder is the output end of the first multi-level feature extraction and deep fusion module;

[0085] The output end of the first graph convolution layer is also connected to the input end of the first downsampling layer, the output end of the first downsampling layer is connected to the input end of the fourth batch normalization layer, and the output end of the fourth batch normalization layer is connected to the input end of the adder;

[0086] The output end of the second graph convolution layer is also connected to the input end of the second downsampling layer, the output end of the second downsampling layer is connected to the input end of the fifth batch normalization layer, and the output end of the fifth batch normalization layer is connected to the input end of the adder.

[0087] It should be understood that the first multi-level feature extraction and deep fusion module aggregates the information between different sensor nodes in turn through three layers of graph convolution layers, so that the model can fully extract and fuse the feature representations of multiple sensors. The detailed structure of this module is as follows Figure 4 As shown in the figure, the graph convolution layer can extract the features of each sensor node while aggregating the feature information of neighboring nodes by performing convolution operations on the graph structure. The graph convolution operation is represented by the following formula:

[0088]

[0089] in, A is the adjacency matrix, which is used to represent the connection relationship between nodes, and I is the identity matrix; yes The degree matrix of is a diagonal matrix, and the calculation formula is shown in formula (5); H (l) is the feature of the lth layer; W (l) is the weight matrix of the lth layer; σ is the nonlinear activation function.

[0090]

[0091] In order to fuse features of different levels and scales, and to avoid the loss of effective information as the number of network layers increases during feature extraction, feature information of different levels is fused by introducing cross-connections. However, since the scales of graph structure data output by different graph convolutional layers are different, in order to facilitate fusion, a downsampling module consisting of a layer of graph convolutional layers is used to reduce the data dimension and unify the data scale.

[0092] Furthermore, if Figure 5 As shown, the classification module includes: a first convolution layer, a second convolution layer and a fully connected layer connected in sequence. The time-frequency fusion feature vector is input into the classification module, and the classification module maps it to the label space, and finally obtains the predicted fault category of the model.

[0093] Furthermore, the unsupervised learning module includes:

[0094] a second time domain branch and a second frequency domain branch;

[0095] The second time domain branch inputs unlabeled time domain graph data;

[0096] The second frequency domain branch inputs unlabeled frequency domain image data;

[0097] The second time domain branch includes: a first graph encoder module, a first hidden layer and a first global average pooling layer connected in sequence; an output end of the first hidden layer is connected to an input end of a first dot product decoder module; an output end of the first dot product decoder module outputs a reconstructed time domain adjacency matrix;

[0098] The second frequency domain branch includes: a second graph encoder module, a second hidden layer and a second global average pooling layer connected in sequence; the output end of the second hidden layer is connected to the input end of the second dot product decoder module; the output end of the second dot product decoder module outputs a reconstructed frequency domain adjacency matrix.

[0099] Furthermore, in the training process of the model, the local network structure parameters of the supervised learning module and the unsupervised learning module are shared with each other, including:

[0100] The structure of the first image encoder module is consistent with the structure of the first multi-level feature extraction and deep fusion module. During the training process, the parameters of the first image encoder module and the first multi-level feature extraction and deep fusion module are shared with each other;

[0101] The structure of the second image encoder module is consistent with the structure of the second multi-level feature extraction and deep fusion module. During the training process, the parameters of the second image encoder module and the second multi-level feature extraction and deep fusion module are shared with each other.

[0102] Furthermore, the working process of the unsupervised learning module includes:

[0103] The unlabeled time domain graph data is mapped to the first hidden feature representation after passing through the first graph encoder module, and then the first hidden feature representation is decoded by the first dot product decoder module to finally obtain the reconstructed time domain adjacency matrix;

[0104] The unlabeled frequency domain graph data is mapped to the second hidden feature representation after passing through the second graph encoder module, and then the second hidden feature representation is decoded by the second dot product decoder module to finally obtain the reconstructed frequency domain adjacency matrix. The specific structure of the unsupervised learning module is as follows: Figure 6 shown.

[0105] Furthermore, the first graph encoder module maps the input time domain graph structure data into a hidden feature representation of a low-dimensional latent space, extracts and retains key information of the input data, and the process can be expressed as:

[0106] Z=f E (A; X)(6)

[0107] Among them, f E (A; X) represents the mapping operation of the first graph encoder module, A represents the adjacency matrix of the unlabeled time domain graph data, X represents the node feature matrix of the unlabeled time domain graph data, and Z represents the hidden feature representation of the low-dimensional latent space, that is, the encoded time domain feature matrix.

[0108] It should be understood that the internal working process of the second image encoder module is consistent with the internal working process of the first image encoder module. The difference is that the first image encoder module inputs time domain image structure data; while the second image encoder module inputs frequency domain image structure data.

[0109] Furthermore, the working process of the first dot product decoder module is consistent with that of the second dot product decoder module, and the first dot product decoder module includes:

[0110] By multiplying the encoded time domain feature matrix with its transposed matrix and then activating it through the activation function, the decoded feature matrix is ​​obtained. The process can be expressed as:

[0111]

[0112] Among them, f D (Z) represents the mapping of the first dot product decoder module, represents the reconstructed time-domain adjacency matrix, σ represents the activation function, usually the Sigmoid function.

[0113] During the model training process, the first dot product decoder module reconstructs an adjacency matrix as close as possible to the original input based on the feature matrix.

[0114] Furthermore, in the training process of the model, the supervised learning module and the unsupervised learning module share network parameters with each other, which means that the first multi-level feature extraction and deep fusion module of the supervised learning module and the first graph encoder module of the unsupervised learning module share network parameters with each other;

[0115] The second multi-level feature extraction and deep fusion module of the supervised learning module and the second graph encoder module of the unsupervised learning module share network parameters with each other.

[0116] It should be understood that S103: The present invention proposes a time-frequency dual multi-level graph autoencoder model as a transformer fault diagnosis model, which is a multi-branch semi-supervised network structure for multi-sensor information fusion, consisting of a supervised learning module and an unsupervised learning module. Among them, the multi-level feature extraction and deep fusion module of the supervised learning module and the graph encoder module of the unsupervised learning module share network structure and parameters. Through such a design, multi-sensor feature information at different aggregation levels can be extracted, and unlabeled data can be fully utilized to complete model training.

[0117] Construct a multi-branch semi-supervised fault diagnosis model; based on the graph autoencoder structure, establish a time-frequency dual multi-level graph autoencoder model, so that it can combine multi-sensor information and make full use of unlabeled data for model training.

[0118] Furthermore, the step S103 of training the multi-branch semi-supervised fault diagnosis model further includes:

[0119] When the total loss function value of the multi-branch semi-supervised fault diagnosis model no longer decreases, the training is stopped to obtain the trained multi-branch semi-supervised fault diagnosis model.

[0120] Furthermore, the total loss function of the multi-branch semi-supervised fault diagnosis model is specifically expressed as:

[0121] L=ω 1 L CE +ω 2 (L CT +L CF )+ω 3 (L TRE +L FRE )(8)

[0122] Among them, ω 1 ,ω 2 and ω 3It is the weight coefficient of different losses, which can be adjusted dynamically during model training.

[0123] When training the model using labeled and unlabeled data, the model is trained and the model parameters are optimized by minimizing formula (8), that is:

[0124]

[0125] Among them, θ represents the parameters of the model.

[0126] For labeled data, the model prediction label is obtained after the classification module, and the classification loss is calculated with the corresponding true label to evaluate the closeness between the probability distribution of the model output and the true label distribution.

[0127] The calculation formula for classification loss is as follows:

[0128]

[0129] Among them, K represents the number of samples, C represents the number of fault categories, and y ic Indicates the true label that the i-th sample belongs to the c-th fault category, It represents the model's predicted probability that the i-th sample belongs to the c-th fault category.

[0130] For unlabeled data, after being mapped to a low-dimensional latent space by the graph encoder module, the time-frequency domain fusion features are obtained by global average pooling, and the time-frequency contrast learning loss L is calculated. CT and L CF , which helps the model learn the similarities and differences of the time-frequency domain features of the input data and improve the model's discriminative power.

[0131] Time-frequency contrastive learning loss L CT and L CF , and its calculation formula is as follows:

[0132]

[0133] For formula (11), t represents the time domain feature representation in the time domain fusion feature, f + is the frequency domain feature representation in the corresponding frequency domain fusion feature, f i is the representation of all frequency domain features in the frequency domain fusion feature, τ is the temperature coefficient, which is used to control the discrimination of negative samples by the graph encoder module in the unsupervised learning module, sim(·) represents the similarity calculation operation, and L CT represents the contrast loss in the time domain, L CF Represents the contrast loss in the frequency domain.

[0134] Formula (12) corresponds to formula (11) and will not be described in detail.

[0135] In addition, the encoded feature matrix is ​​decoded by the dot product decoder module to obtain the reconstructed adjacency matrix, which is used to calculate the reconstruction loss with the original time-frequency domain adjacency matrix to evaluate the reconstruction accuracy and optimize the model. The calculation formula is as follows:

[0136]

[0137] Among them, A ij Represents the adjacency matrix of the input time domain graph data, A i ' j Represents the adjacency matrix of the input frequency domain graph data; represents the reconstructed time domain adjacency matrix, Represents the reconstructed frequency domain adjacency matrix; ε is a very small number, the main function of which is to ensure that the logarithmic operation can be calculated; L TRE represents the time domain reconstruction loss, L FRE represents the frequency domain reconstruction loss.

[0138] It should be understood that the multi-branch semi-supervised fault diagnosis model training: combines classification loss, reconstruction loss and time-frequency contrast learning loss, and uses a graph dataset containing a large amount of unlabeled and a small amount of labeled data to train and evaluate the network model.

[0139] Further, S104: obtain the vibration signal of the transformer to be diagnosed, convert the vibration signal of the transformer to be diagnosed into graph structure data, input the graph structure data into the trained supervised learning module, and obtain the fault identification result, wherein the vibration signal of the transformer to be diagnosed is converted into graph structure data, and the conversion process also regards the sensor of the transformer to be diagnosed as a node, and constructs the connection relationship between the edges of the nodes based on the adaptive radius graph algorithm of local density, and then constructs it.

[0140] It should be understood that the supervised learning module is separated from the trained fault diagnosis model, that is, the unsupervised learning module is removed and only the supervised learning module is retained. The vibration signal of the operating transformer is converted into graph data and then input into the module. The module will automatically extract the fault information contained in the graph data and output the corresponding fault category to complete the fault diagnosis of the transformer.

[0141] The present invention constructs a more lightweight multi-level feature extraction and deep fusion module for multi-sensor information fusion, and introduces cross-connection and downsampling to extract multi-sensor features at different aggregation levels, effectively avoiding the loss of fault features in the convolution aggregation process, and improving the accuracy of model fault diagnosis. By constructing a supervised learning module and an unsupervised learning module with local network structure and parameter sharing in the fault diagnosis model, and using multiple loss functions to jointly train the network model, the model can obtain good training results even with limited label data, ensuring the accuracy of model fault identification.

[0142] Embodiment 2

[0143] This embodiment provides a multi-branch semi-supervised fault diagnosis system for multi-sensor fusion equipment, including:

[0144] An acquisition module is configured to: acquire transformer fault data: utilize a sensor network composed of multiple sensors to synchronously collect vibration signals at different positions on the surface of a real transformer under different fault defects;

[0145] A data set construction module is configured to: construct a graph data set based on the transformer fault data, divide the graph data set into a training set and a test set, and divide the training set into labeled data and unlabeled data;

[0146] A model building module is configured to: build a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model includes: a supervised learning module and an unsupervised learning module in parallel; train the multi-branch semi-supervised fault diagnosis model, use labeled data to train the supervised learning module during the model training process, and use unlabeled data to train the unsupervised learning module, and share local network structure parameters of the supervised learning module and the unsupervised learning module during the model training process;

[0147] The fault identification module is configured to: obtain the vibration signal of the transformer to be diagnosed, convert the vibration signal of the transformer to be diagnosed into graph structure data, input the graph structure data into the trained supervised learning module, and obtain the fault identification result.

[0148] It should be noted that the acquisition module, data set construction module, model construction module and fault identification module described above correspond to steps S101 to S104 in Embodiment 1, and the examples and application scenarios implemented by the modules and corresponding steps are the same, but are not limited to the contents disclosed in Embodiment 1 described above. It should be noted that the modules described above as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0149] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0150] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0151] Embodiment 3

[0152] This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory so that the electronic device executes the method described in the above embodiment one.

[0153] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0154] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type. In the implementation process, each step of the above method may be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software.

[0155] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0156] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0157] Embodiment 4: This embodiment further provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the method described in Embodiment 1 is completed.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment, characterized in that: include: Obtaining transformer fault data: Using a sensor network consisting of multiple sensors, synchronously collect vibration signals at different locations on the surface of the real transformer under different fault defects; Based on the transformer fault data, a graph data set is constructed, the graph data set is divided into a training set and a test set, and the training set is divided into labeled data and unlabeled data; Constructing a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model includes: a supervised learning module and an unsupervised learning module in parallel; The multi-branch semi-supervised fault diagnosis model is trained. During the training process of the model, the supervised learning module is trained with labeled data, and the unsupervised learning module is trained with unlabeled data. During the training process of the model, the local network structure parameters of the supervised learning module and the unsupervised learning module are shared with each other. The vibration signal of the transformer to be diagnosed is obtained, the vibration signal of the transformer to be diagnosed is converted into graph structure data, and the graph structure data is input into a trained supervised learning module to obtain a fault identification result.

2. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment as claimed in claim 1, characterized in that: To construct a graph dataset, each sensor in the sensor network is regarded as a graph node, and the connection relationship between the edges of the nodes is constructed using an adaptive radius graph algorithm based on local density, thereby constructing a graph dataset.

3. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment as claimed in claim 1, characterized in that: The supervised learning module includes: a first time domain branch and a first frequency domain branch; the first time domain branch inputs labeled time domain image data; the first frequency domain branch inputs labeled frequency domain image data; the first time domain branch includes: a first multi-level feature extraction and deep fusion module, a first hidden layer and a first global average pooling layer connected in sequence; the first frequency domain branch includes: a second multi-level feature extraction and deep fusion module, a second hidden layer and a second global average pooling layer connected in sequence; the output end of the first global average pooling layer is connected to the input end of the splicing unit; the output end of the splicing unit outputs a fused feature vector, and the fused feature vector is input into the classification module to obtain a fault diagnosis prediction label.

4. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment as claimed in claim 3, characterized in that: The supervised learning module has a working process including: after the labeled time domain graph data passes through the first multi-level feature extraction and deep fusion module, the time domain graph structure data that aggregates the features of sensor nodes at different levels is obtained; after the labeled frequency domain graph data passes through the second multi-level feature extraction and deep fusion module, the frequency domain graph structure data that aggregates the features of sensor nodes at different levels is obtained; after the time domain graph structure data that aggregates the features of sensor nodes at different levels is processed by the global average pooling layer, the time domain fusion features are obtained; after the frequency domain graph structure data that aggregates the features of sensor nodes at different levels is processed by the global average pooling layer, the frequency domain fusion features are obtained; the time domain fusion features and the frequency domain fusion features are spliced ​​to obtain the time-frequency fusion features; the time-frequency fusion features are input into the classification module to obtain the predicted labels.

5. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment as claimed in claim 3, characterized in that: The internal structure of the first multi-level feature extraction and deep fusion module is consistent with that of the second multi-level feature extraction and deep fusion module. The first multi-level feature extraction and deep fusion module includes: a first graph convolution layer, a first batch of normalization layers, a first activation function layer, a second graph convolution layer, a second batch of normalization layers, a second activation function layer, a third graph convolution layer, a third batch of normalization layers and an adder connected in sequence; the input end of the first graph convolution layer is the input end of the first multi-level feature extraction and deep fusion module; the output end of the adder is the output end of the first multi-level feature extraction and deep fusion module; the output end of the first graph convolution layer is also connected to the input end of the first downsampling layer, the output end of the first downsampling layer is connected to the input end of the fourth batch of normalization layer, and the output end of the fourth batch of normalization layer is connected to the input end of the adder; the output end of the second graph convolution layer is also connected to the input end of the second downsampling layer, the output end of the second downsampling layer is connected to the input end of the fifth batch of normalization layer, and the output end of the fifth batch of normalization layer is connected to the input end of the adder.

6. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment as claimed in claim 1, characterized in that: The unsupervised learning module includes: a second time domain branch and a second frequency domain branch; the second time domain branch inputs unlabeled time domain graph data; the second frequency domain branch inputs unlabeled frequency domain graph data; the second time domain branch includes: a first graph encoder module, a first hidden layer and a first global average pooling layer connected in sequence; the output end of the first hidden layer is connected to the input end of the first dot product decoder module; the output end of the first dot product decoder module outputs a reconstructed time domain adjacency matrix; the second frequency domain branch includes: a second graph encoder module, a second hidden layer and a second global average pooling layer connected in sequence; the output end of the second hidden layer is connected to the input end of the second dot product decoder module; the output end of the second dot product decoder module outputs a reconstructed frequency domain adjacency matrix; The working process of the unsupervised learning module includes: the unlabeled time domain graph data is mapped to the first hidden feature representation after passing through the first graph encoder module, and then the first hidden feature representation is decoded by the first dot product decoder module to finally obtain the reconstructed time domain adjacency matrix; the unlabeled frequency domain graph data is mapped to the second hidden feature representation after passing through the second graph encoder module, and then the second hidden feature representation is decoded by the second dot product decoder module to finally obtain the reconstructed frequency domain adjacency matrix.

7. A multi-branch semi-supervised fault diagnosis method for multi-sensor fusion equipment as claimed in claim 6, characterized in that: The first graph encoder module maps the input time-domain graph structure data into a hidden feature representation in a low-dimensional latent space, extracting and retaining key information of the input data; The first dot product decoder module includes: obtaining a decoded feature matrix by multiplying the encoded time domain feature matrix and its transposed matrix, and then activating it through an activation function.

8. A multi-branch semi-supervised fault diagnosis system for multi-sensor fusion equipment, characterized in that: include: An acquisition module is configured to: acquire transformer fault data: utilize a sensor network composed of multiple sensors to synchronously collect vibration signals at different positions on the surface of a real transformer under different fault defects; A data set construction module is configured to: construct a graph data set based on the transformer fault data, divide the graph data set into a training set and a test set, and divide the training set into labeled data and unlabeled data; A model building module is configured to: build a multi-branch semi-supervised fault diagnosis model; the multi-branch semi-supervised fault diagnosis model includes: a supervised learning module and an unsupervised learning module in parallel; train the multi-branch semi-supervised fault diagnosis model, use labeled data to train the supervised learning module during the model training process, and use unlabeled data to train the unsupervised learning module, and share local network structure parameters of the supervised learning module and the unsupervised learning module during the model training process; The fault identification module is configured to: obtain the vibration signal of the transformer to be diagnosed, convert the vibration signal of the transformer to be diagnosed into graph structure data, input the graph structure data into the trained supervised learning module, and obtain the fault identification result.

9. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; as well as a processor for executing the computer readable instructions, When the computer-readable instructions are executed by the processor, the method described in any one of claims 1 to 7 is executed.

10. A storage medium, characterized in that: The computer-readable instructions are non-transitory stored, wherein when the non-transitory computer-readable instructions are executed by a computer, the method according to any one of claims 1 to 7 is performed.

Images