Transformer fault diagnosis method and system based on graph mutual mapping transmission network
By using a graph mutual mapping transmission network in transformer fault diagnosis, combined with sample mapping technology and graph generation layer, the problems of insufficient training samples and high computational complexity in the existing technology are solved, and higher diagnostic accuracy and generalization capabilities are achieved.
Patent Information
- Application Number
- CN202411694811.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-04-15
AI Technical Summary
In the transformation fault diagnosis, the prior art has problems such as insufficient training samples, high computational complexity of graph networks and slow convergence speed, which is difficult to meet the needs of modern power systems for real-time monitoring.
A fault diagnosis method based on graph mutual mapping transmission network is proposed. By designing sample mapping technology and graph generation layer, the graph structure is automatically constructed, and data features are extracted using graph convolution networks, and the generalization ability and accuracy of the model are improved through unsupervised domain adaptive training.
It improves the generalization ability, accuracy and convergence speed of the model, can be effectively applied to transformer fault diagnosis, greatly improving the accuracy of the diagnosis.
Smart Images
Figure CN119830090A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of transformer fault diagnosis, and in particular to a transformer fault diagnosis method and system based on a graph mutual mapping transmission network. Background Art
[0002] As the complexity of power systems increases, higher requirements are placed on real-time monitoring of transformer health status. Traditional fault diagnosis methods can no longer meet the needs of modern power systems. The development of MEMS (micro-electromechanical systems) technology has made it possible to realize miniaturized, high-performance sensing devices that can be embedded inside transformers to continuously monitor their operating status. MEMS photoacoustic spectroscopy sensors can detect characteristic gases in transformer oil in real time, such as hydrogen (H2), methane (CH4), acetylene (C2H2), etc., thereby providing important information about the internal condition of the transformer.
[0003] However, data collected by hardware alone is not enough to fully assess the health status of the transformer. Therefore, advanced data analysis algorithms and machine learning techniques, such as artificial neural networks (ANN), support vector machines (SVM), decision trees (DT) and random forests (RF), are introduced to process the massive data provided by MEMS sensors. These algorithms can be trained to identify the association between different gas combinations and specific fault modes, thereby achieving accurate classification and prediction of faults.
[0004] Recent studies have shown that graph networks, as a new machine learning framework, are particularly effective in processing complex relational data. However, graph networks also face some challenges in practical applications. The first is the problem of insufficient training samples, which will weaken the model's migration performance and feature generalization ability. The second is that the convolutional neural network used in graph networks has a high computational complexity, resulting in a slow convergence speed. Summary of the invention
[0005] In order to overcome the shortcomings of the above-mentioned prior art, the present invention proposes a transformer fault diagnosis method and system based on a graph mutual mapping transmission network. First, a sample mapping technology is designed. The source domain and target domain data are processed by an activation function and multiplied with the mapping weight to obtain the new domain data. The weight is adjusted according to the new domain data, and repeated three times to achieve the consistency of feature distribution. Secondly, a graph mutual mapping transmission network is proposed. Each feature vector after sample mapping is regarded as a node, and its value is regarded as a node feature. The graph structure is automatically constructed through the graph generation layer, and finally it is input into the graph convolutional network to extract the data features in the graph. Unsupervised domain adaptation training is achieved by defining a comprehensive objective function including classification loss, domain alignment loss and structural alignment loss.
[0006] Step 1: Use a MEMS photoacoustic sensor with a non-resonant photoacoustic cell to analyze the acoustic wave signal through photoacoustic spectroscopy technology, use FPGA to collect the dissolved gas in the transformer oil to form a parameter sequence, and perform data framing, data arbitration, and data transmission. The data is transmitted to the Feiteng CPU processor through PCIE QDMA for data processing.
[0007] Step 2: A sample mapping technology is designed in the CPU processor. The source domain and target domain data are processed by the activation function and multiplied with the mapping weight to obtain the new domain data. The weight is adjusted according to the new domain data and repeated three times to achieve consistency in feature distribution.
[0008] Step 3: Propose a graph mutual mapping transmission network, which regards each feature vector after sample mapping as a node and its value as the node feature, and automatically constructs the graph structure through the graph generation layer, and finally inputs it into the multi-receptive field graph convolutional network to extract the data features in the graph.
[0009] Step 4: Unsupervised domain adaptation training is implemented by defining a comprehensive objective function consisting of classification loss, domain alignment loss, and structural alignment loss.
[0010] Furthermore, the step 1 comprises a MEMS photoacoustic sensor having a non-resonant photoacoustic cell.
[0011] The photoacoustic signal of the photoacoustic cell is given by:
[0012]
[0013] Where P α is the product of the incident power and the absorption coefficient, r is the radius of the photoacoustic cell cavity, ρ is the density, C p is the specific heat at constant pressure, k is the thermal conductivity
[0014] The photoacoustic cell is designed to be in a non-resonant state so that it cannot resonate with the photoacoustic signal, reduce the signal change caused by resonance, and optimize the photoacoustic signal. As the operating frequency of the photoacoustic cell increases, the photoacoustic signal shows an exponential decay trend. The internal dimensions of the photoacoustic cell are optimized to reduce the radius of the cell body to enhance the photoacoustic signal, while the length of the cell body does not affect the signal intensity; the inner wall of the photoacoustic cell uses polytetrafluoroethylene as a coating material to reduce the physical adsorption of gas molecules on the cell wall and improve the detection effect.
[0015] Furthermore, in step 1, the FPGA performs data framing, data arbitration, and data transmission.
[0016] The FPGA chip is KU115. The parallel capability of FPGA is used to collect the concentration of dissolved gas in oil collected by MEMS photoacoustic sensor in parallel, and parallel framing of multi-parameter data is performed in FPGA. The framing format is 16-byte frame header plus 4096-byte data. The frame header is 0x55aaaa55+0x77888877+0xaaaaaaaa+4-byte command number, where 0xaaaaaaaa is the data type format. The data type format of the frame header of gas concentration type 1 to gas concentration type 6 is: 0xaaaaaaaa, 0xbbbbbbbb, 0xcccccccc, 0xdddddddd, 0xeeeeeeee, 0xffffffff. The 4-byte command number is for the CPU to identify the command number to start or stop reading data.
[0017] Data arbitration is performed by inputting each framed data into the fifo for caching, and inputting the framed data of the six gas concentrations into the xilinx AXI interconnect ip core. The data interface is AXI-Stream, and data arbitration is performed, that is, polling and reading the data packets of each type of data to transmit to the next level.
[0018] Data transmission is to use PCIE QDMA to transmit the six gas data through the AXI-Stream interface, and use QDMA's interrupt method to transmit the polled data to the Feiteng CPU for processing.
[0019] Further, the sample mapping technology in step 2.
[0020] The sample mapping technique uses an activation function to process the source domain data, multiplies the processed data with the mapping weights to obtain the new domain data, adjusts the mapping weights according to the new domain data, performs the same process on the target domain data, and the entire mapping process is repeated three times.
[0021] Assume that the input x in the source domain is {x1,x2,...,x t}, where x(i)∈R N ,x∈X, t is the number of working conditions, X is the source domain, and N is the number of samples. The input y of the target domain is {y1,y2,...,y t}, where y(i)∈R N ,y∈Y, Y is the target domain. The new domain space to which the source domain and the target domain are mapped is Z. The results of the first mapping of the source domain and the target domain are a={a1,a2,...,a t} and o={o1,o2,...,o t}, the results of the second mapping are b={b1,b2,...,b t} and p={p1,p2,…,pt}, the new results mapped to Z space are c={c1,c2,…,c t} and q={q1,q2,...,q t The three-way mapping weights of the source domain are A1, A2, and A3, and the three-way mapping weights of the target domain are B1, B2, and B3.
[0022] The activation function f is expressed as
[0023]
[0024] The mapping equation can be written as
[0025] X×A=Z
[0026] Y×B=Z
[0027] Where λ is the penalty parameter. After the source domain and target domain data are processed by the activation function, the mapping weights A and B can be expressed as
[0028] A=(X T X+λI) -1 X T Z
[0029] B=(Y T Y+λI) -1 Y T Z
[0030] Furthermore, the sample mapping technology is characterized by adjustment of mapping weights.
[0031] In order to optimize the mapping weights, that is, to make the mapped data as close to the original data as possible, the optimization is performed by minimizing the residual sum of squares formula:
[0032]
[0033] in is a i , and Q is the residual sum of squares.
[0034] After finding the mapping weight matrix, we hope to further optimize the clustering effect of the mapped data in the new domain Z. This requires finding a set of transformation vectors a i , so that the mapped data is not only close to the actual value in the new domain, but also has better clustering characteristics. In order to find the optimal mapping weight, by solving the generalized eigenvalue problem
[0035]
[0036] Among them, A his the transformation matrix of the weight mapping, L is the Laplace matrix, D is the diagonal matrix, λ is the eigenvalue, and the solved A h It can provide better projection, so that the transformed data has better clustering characteristics in the new domain.
[0037] Furthermore, the graph generation layer in step 3 constructs a graph structure.
[0038] Each feature vector z in the mapped feature vector set Z′ i ′, regarded as a node v in the graph i , and use the value of the feature vector as the feature of the node to form a node set V, where V = {v1, v2, …, v n}
[0039] The node features v extracted by sample mapping i The input is fed into a multi-layer perceptron (MLP), and then the adjacency matrix is obtained by matrix multiplication between the MLP features and their transpose, and the nearest node is selected using the TOp-k sorting mechanism. The process can be expressed as
[0040]
[0041] in, is the output of the multilayer perceptron, normalize(·) is the normalization function, and W represents the adjacency matrix formed during the graph generation process. is a sparse adjacency matrix, and Top-k(·) returns the index of the maximum value of each row of W.
[0042] Finally, the constructed graph structure is input into the spectral graph convolutional network, and the data features in the graph are extracted through the convolution kernel expanded by Chebyshev polynomials. The process can be expressed as
[0043]
[0044] Among them, g θ is a graph filter, θ is a learnable parameter, and Λ is the directional eigenvalue matrix of the Laplace moment. K determines the range of the node domain, where in is the eigenvalue of the Laplacian matrix.
[0045] Furthermore, the multi-receptive field graph convolutional network modeling in step 3 is performed.
[0046] The node features of the instance graph are input into the multi-receptive field graph convolutional network, and multiple convolution operations are performed to obtain the fused features. The process can be expressed as
[0047]
[0048] in, is a learnable parameter, and H represents the fused feature.
[0049] Before inputting the fused features into the fully connected layer, feature fusion and learning are performed first. This process can be expressed as
[0050] H0=MRFConv(WXB0)
[0051] H1=MRFConv(WH0B1)
[0052] Among them, H0 is the feature learned for the first time, H1 is the feature learned for the second time, WX and WH0 are the inputs of the first and second convolution operations respectively, and B0 and B1 are trainable weight matrices.
[0053] Further, the unsupervised domain adaptation process in step 4 is described.
[0054] The method includes a classifier for predicting the category labels of source domain and target domain samples; and a domain discriminator for distinguishing the data features of the source domain and the target domain.
[0055] The working assumption of the classifier and domain discriminator is that the source domain D S and the target domain D t The label space is the same, that is, D S =D t , and the feature space X S and X t are different but related, i.e., there is covariate drift. A deep network is built using the maximum mean difference method to predict the labels of target domain samples;
[0056] where the maximum mean difference method is defined as:
[0057]
[0058] in represents a nonlinear mapping function, Ω indicates that the distance is measured in the reproducing kernel Hilbert space (RKHS), and E() represents the mathematical expectation.
[0059] In order to learn transferable features, the overall objective function is set as a combination of classification loss, domain alignment loss, and structure alignment loss to achieve unsupervised domain adaptation;
[0060] The classification loss L C is defined as:
[0061]
[0062] in represents the label classifier for source domain samples, is the prediction result of the category label, represents the cross entropy loss.
[0063] To combat domain covariate shift, a domain discriminator is used to extract features from the source and target domains, and a feature extractor is trained to fool the domain discriminator. The domain alignment loss is calculated as follows: DA is defined as:
[0064]
[0065] in and They represent the extracted features of the i-th source domain sample and the j-th target domain sample respectively. The value of D() is 0 or 1, indicating which domain the sample belongs to.
[0066] Structural alignment loss L MMD is defined as:
[0067]
[0068] The overall objective function L Total is defined as:
[0069] L Total (X s ,Y s ,X t )=L C (X s ,Y s )+rL DA (X s ,X t )+kL MMD (X s ,X t )
[0070] Where r and k are trade-off parameters
[0071] Further, the unsupervised domain adaptation process described.
[0072] An updating module is utilized to update the parameters of the feature extractor, label classifier, and domain discriminator during model operation.
[0073] The parameters of the feature extractor, label classifier and domain discriminator are respectively F ,θ C , and θ D Indicates that, wherein the update module updates the above parameters through the back propagation algorithm, and the specific expression is:
[0074]
[0075] The present invention proposes a new fault diagnosis method based on a graph mutual mapping transmission network. The method first designs a sample mapping technology, processes the source domain and target domain data through an activation function and multiplies them with the mapping weight to obtain the new domain data, and then adjusts the weight according to the new domain data, repeating three times to achieve the consistency of feature distribution. Secondly, a graph mutual mapping transmission network is proposed, which regards each feature vector after sample mapping as a node, and its value as a node feature, and automatically constructs a graph structure through a graph generation layer, and finally inputs it into a graph convolutional network to extract data features in the graph. Unsupervised domain adaptation training is achieved by defining a comprehensive objective function including classification loss, domain alignment loss and structural alignment loss.
[0076] The system mainly includes:
[0077] Transformer parameter data acquisition module: Use a MEMS photoacoustic sensor with a non-resonant photoacoustic cell to analyze the acoustic wave signal through photoacoustic spectroscopy technology, use FPGA to collect the dissolved gas in the transformer oil to form a parameter sequence, and transmit the data to the Feiteng CPU processor through the PCIEQDMA interface.
[0078] Sample mapping and weight adjustment module, graph structure feature extraction module, unsupervised domain adaptation module
[0079] The present invention has the following advantages:
[0080] The present invention proposes a new fault diagnosis method based on a graph mutual mapping transmission network. The method first designs a sample mapping technology, processes the source domain and target domain data through an activation function and multiplies them with the mapping weight to obtain new domain data, adjusts the weight according to the new domain data, and repeats three times to achieve consistency in feature distribution. Secondly, a graph mutual mapping transmission network is proposed, which regards each feature vector after sample mapping as a node, its value as a node feature, and automatically constructs a graph structure through a graph generation layer, and finally inputs it into a graph convolutional network to extract data features in the graph. Unsupervised domain adaptation training is achieved by defining a comprehensive objective function including classification loss, domain alignment loss and structural alignment loss. This method improves the generalization ability, accuracy and convergence speed of the model. Therefore, the invention can be effectively applied to transformer fault diagnosis, greatly improving the accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0081] Figure 1 It is a system block diagram of the present invention.
[0082] Figure 2 This is a diagram of the FPGA data framing format of the present invention.
[0083] Figure 3 This is a sample mapping structure diagram of the present invention.
[0084] Figure 4 This is the multi-receptive field graph convolutional network diagram of the present invention.
[0085] Figure 5 This is a structural diagram of the present invention.
[0086] Specific implementation cases
[0087] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.
[0088] Example: Transformer various states, including normal (N), partial discharge (PD), low energy discharge (F1), high energy discharge (F2), medium and low temperature overheating (T1) and high temperature overheating (T2). Each fault mode contains 819200 data points, which are organized into 100 samples, each sample contains 8192 data points, 80% of the data is selected as the training set, and the remaining 20% is used as the test set. In the FPGA, the command code is used to stop the data transmission every time 819200 data points are transmitted, waiting for the CPU network to train.
[0089] Parameter setting part: Finally, 3 times were determined as the optimal number of mappings, because more mapping times can slightly improve the accuracy but will significantly increase the amount of calculation.
[0090] 1 is selected as the optimal value of the penalty parameter because in transformer fault diagnosis, when the penalty parameter is 1, the fault diagnosis accuracy reaches the highest 97.92%.
[0091] Based on experimental results, the model width is set to 1024, which can significantly improve the accuracy, although it will increase the running time.
[0092] Step S1: The MEMS photoacoustic sensor of the non-resonant photoacoustic cell adopts the parameter sequence of the dissolved gas in the transformer oil, and uses FPGA to perform data framing, data arbitration, and data transmission, and transmits the data to the Feiteng CPU processor through PCIE QDMA for data analysis.
[0093] The photoacoustic cell is used to convert light energy into sound wave signals, and its photoacoustic signal is given by the following formula:
[0094]
[0095] Where P α is the product of the incident power and the absorption coefficient, r is the radius of the photoacoustic cell cavity, ρ is the density, Cp is the specific heat at constant pressure, k is the thermal conductivity
[0096] In this embodiment, the photoacoustic cell is designed to be in a non-resonant state so that it cannot resonate with the photoacoustic signal, reduce the signal change caused by resonance, and optimize the photoacoustic signal. As the operating frequency of the photoacoustic cell increases, the photoacoustic signal shows an exponential decay trend. The internal dimensions of the photoacoustic cell are optimized to reduce the radius of the cell body to enhance the photoacoustic signal, while the length of the cell body does not affect the signal intensity; the inner wall of the photoacoustic cell uses polytetrafluoroethylene as a coating material to reduce the physical adsorption of gas molecules on the cell wall and improve the detection effect.
[0097] FPGA performs data processing, including data framing, data arbitration, and data transmission.
[0098] Data framing: The FPGA chip is KU115. The parallel capability of FPGA is used to collect the concentration of dissolved gas in oil collected by MEMS photoacoustic sensor in parallel, and parallel framing of multi-parameter data is performed in FPGA. The framing format is 16-byte frame header and 4096-byte data. Among them, the frame header is 0x55aaaa55+0x77888877+0xaaaaaaaaa+4-byte command number, among which 0xaaaaaaaaa is the data type format, and the data type format of the frame header of gas concentration type 1 to gas concentration type 6 is: 0xaaaaaaaa, 0xbbbbbbbb, 0xcccccccc, 0xdddddddd, 0xeeeeeeee, 0xffffffff; the 4-byte command number is for the CPU to identify the command number to start reading or stop reading data.
[0099] Data arbitration: In order to process data of multiple gas concentrations, each framed data is input to the fifo for buffering. Then, the framed data of the six gas concentrations is input to the xilinx AXI interconnect ip core, and the data interface is AXI-Stream for data arbitration. Data arbitration uses polling to read the data packets of each type of data and transmit it to the next level.
[0100] Data transmission: The six gas data are transmitted to the Feiteng CPU processor through the AXI-Stream interface via PCIE QDMA. The interrupt mode of QDMA ensures the real-time and reliability of the data.
[0101] Step S2: A sample mapping technology is designed in the CPU processor. The source domain and target domain data are processed by the activation function and multiplied with the mapping weight to obtain the new domain data. The weight is adjusted according to the new domain data and repeated three times to achieve consistency in the feature distribution of the source domain and target domain data.
[0102] Specifically: Sample mapping technology maps the source domain data to the new domain space through a combination of activation function and mapping weights, and adjusts the mapping weights according to the new domain data. This process is repeated three times, aiming to gradually approach the optimal mapping result. By minimizing the residual sum of squares and optimizing the clustering effect, the optimal mapping weights and conversion vectors can be found, so that the mapped data has better expression ability and clustering characteristics in the new domain.
[0103] Assumption: The input x in the source domain is {x1,x2,…,x t}, where x(i)∈R N ,x∈X, t is the number of working conditions, X is the source domain, N is the number of samples;
[0104] The input of the target domain y={y1,y2,...,y t}, where y(i)∈R N ,y∈Y, Y is the target domain;
[0105] The source domain and the target domain are mapped to the new domain space Z. The results of the first mapping of the source domain and the target domain are a={a1,a2,...,a t} and o={o1,o2,...,o t}, the results of the second mapping are b={b1,b2,...,b t} and p={p1,p2,...,p t}, the new results mapped to Z space are c={c1,c2,…,c t} and q={q1,q2,…,q t}; The three-way mapping weights of the source domain are A1, A2, and A3, and the three-way mapping weights of the target domain are B1, B2, and B3.
[0106] The activation function f is expressed as:
[0107]
[0108] Mapping equation:
[0109] X×A=Z
[0110] Y×B=Z
[0111] λ is the penalty parameter. After the source domain and target domain data are processed by the activation function, the mapping weights A and B can be expressed as
[0112] A=(X T X+λI) -1 X T Z
[0113] B=(Y T Y+λI)-1 Y T Z
[0114] Adjustment of mapping weights: In order to optimize the mapping weights, i.e., we hope that the mapped data is as close to the original data as possible and has better clustering characteristics in the new domain Z, we need to adjust the weights, including the following two goals:
[0115] Minimize the residual sum of squares: By solving the problem of minimizing the residual sum of squares, we can find the mapping weight that minimizes the difference between the mapped data and the original data. The residual sum of squares represents the sum of the squared errors between the mapped data and the original data.
[0116]
[0117] in is a i , and Q is the residual sum of squares.
[0118] Optimizing clustering effect: After finding the mapping weight matrix, in order to further optimize the clustering effect of the mapped data in the new domain Z, a set of transformation vectors needs to be found. These transformation vectors can make the mapped data not only close to the actual value in the new domain, but also have better clustering characteristics. This is usually achieved by solving the generalized eigenvalue problem:
[0119]
[0120] Among them, A h is the transformation matrix of the weight mapping, L is the Laplace matrix, D is the diagonal matrix, λ is the eigenvalue, and the solved A h It can provide better projection, so that the transformed data has better clustering characteristics in the new domain.
[0121] Step S3: Construct a graph mutual mapping transmission network, regard each feature vector after sample mapping as a node, and its value as a node feature, and automatically construct the graph structure through the graph generation layer, and finally input it into the multi-receptive field graph convolutional network to extract feature information of different scales in the graph structure.
[0122] Specific:
[0123] 3.1 In the graph mapping transmission network, the original sample is first mapped to obtain a series of feature vectors:
[0124] 3.2 Each feature vector z in the mapped feature vector set Z′ i ′, regarded as a node v in the graph structure i , and use the value of the feature vector as the feature of the node to form a node set V, where V = {v1, v2, ..., v n};
[0125] 3.3 Automatically determine the connection relationship between nodes based on the similarity, distance or other metrics between node features, thereby generating a complete graph structure:
[0126] In this embodiment, the node feature v extracted by sample mapping i The input is fed into a multi-layer perceptron (MLP), and then the adjacency matrix is obtained by matrix multiplication between the MLP features and their transpose, and the nearest node is selected using the TOp-k sorting mechanism. The process can be expressed as
[0127]
[0128] in, is the output of the multilayer perceptron, normalize(·) is the normalization function, and W represents the adjacency matrix formed during the graph generation process. is a sparse adjacency matrix, and Top-k(·) returns the index of the maximum value of each row of W.
[0129] 3.4 Multi-receptive field graph convolutional network modeling:
[0130] The constructed graph structure is input into the multi-receptive field graph convolutional network, and multiple convolution operations are performed to obtain the fused features. The process is expressed as
[0131]
[0132] in, is a learnable parameter, H represents the fused feature,
[0133] Before inputting the fused features into the fully connected layer, feature fusion and learning are performed first. This process can be expressed as
[0134] H0=MRFConv(WXB0)
[0135] H1=MRFConv(WH0B1)
[0136] Among them, H0 is the feature learned for the first time, H1 is the feature learned for the second time, WX and WH0 are the inputs of the first and second convolution operations respectively, and B0 and B1 are trainable weight matrices.
[0137] Step S4: Unsupervised domain adaptation training is implemented by defining a comprehensive objective function including classification loss, domain alignment loss, and structural alignment loss.
[0138] Specific:
[0139] The method includes a classifier for predicting the category labels of source domain and target domain samples; and a domain discriminator for distinguishing the data features of the source domain and the target domain.
[0140] The working assumption of the classifier and domain discriminator is that the source domain D S and the target domain D t The label space is the same, that is, D S =D t , and the feature space X S and X t are different but related, i.e., there is covariate drift. A deep network is built using the maximum mean difference method to predict the labels of target domain samples;
[0141] where the maximum mean difference method is defined as:
[0142]
[0143] in represents a nonlinear mapping function, Ω indicates that the distance is measured in the reproducing kernel Hilbert space (RKHS), and E() represents the mathematical expectation.
[0144] In order to learn transferable features, the overall objective function is set as a combination of classification loss, domain alignment loss, and structure alignment loss to achieve unsupervised domain adaptation;
[0145] The classification loss L C is defined as:
[0146]
[0147] in represents the label classifier for source domain samples, is the prediction result of the category label, represents the cross entropy loss.
[0148] To combat domain covariate shift, a domain discriminator is used to extract features from the source and target domains, and a feature extractor is trained to fool the domain discriminator. The domain alignment loss is calculated as follows: DA is defined as:
[0149]
[0150] in and They represent the extracted features of the i-th source domain sample and the j-th target domain sample respectively. The value of D() is 0 or 1, indicating which domain the sample belongs to.
[0151] Structural alignment loss L MMDis defined as:
[0152]
[0153] The overall objective function L Total is defined as:
[0154] L Total (X s ,Y s ,X t )=L C (X s ,Y s )+rL DA (X s ,X t )+kL MMD (X s ,X t )
[0155] where r and k are trade-off parameters.
[0156] An updating module is utilized to update the parameters of the feature extractor, label classifier, and domain discriminator during model operation.
[0157] The parameters of the feature extractor, label classifier and domain discriminator are respectively F ,θ C , and θ D Indicates that, wherein the update module updates the above parameters through the back propagation algorithm, and the specific expression is:
[0158]
[0159] The experimental results show that the average transformer fault diagnosis accuracy of the graph mutual mapping transmission network is higher than that of other methods, such as MK-MMD, CORAL, DANN, DAGCN and CDANN+E. Therefore, the present invention is not only superior to other methods in terms of accuracy, but also shows its unique advantages in terms of model robustness, generalization ability and computational efficiency. This model provides a new solution for transformer fault diagnosis and has broad application prospects.
Claims
1. A transformer fault diagnosis method based on graph mutual mapping transmission network, characterized in that: The following steps are included. Step 1: Collect dissolved gas in transformer oil through MEMS photoacoustic sensor of non-resonant photoacoustic cell, and use FPGA to perform data framing, data arbitration and data transmission to CPU processor; Step 2: In the CPU processor, the sample mapping technology is used to process the source domain and target domain data. The source domain data is mapped to the new domain space through the combination of activation function and mapping weight, and the mapping weight is adjusted according to the new domain data. This process is repeated three times to achieve consistency in the feature distribution of the source domain and target domain data. Step 3: Construct a graph mutual mapping transmission network, regard each feature vector after sample mapping as a node, and its value as a node feature, automatically construct a graph structure through the graph generation layer, and input the constructed graph structure into the multi-receptive field graph convolution network to extract feature information of different scales in the graph structure; Step 4: Unsupervised domain adaptation training is implemented by defining a comprehensive objective function including classification loss, domain alignment loss and structural alignment loss. The comprehensive objective function uses a classifier to predict the category labels of source domain and target domain samples, uses a domain discriminator to distinguish the data features of the source domain and the target domain, and establishes a deep network through the maximum mean difference method to predict the labels of the target domain samples; and uses an update module to update the parameters of the feature extractor, label classifier and domain discriminator during model operation.
2. A transformer fault diagnosis method based on graph mutual mapping transmission network according to claim 1, characterized in that: The step 1 specifically includes: using a MEMS photoacoustic sensor with a non-resonant photoacoustic cell to analyze the sound wave signal through photoacoustic spectroscopy technology to obtain parameter information of dissolved gases in transformer oil; using FPGA to perform data preprocessing on the parameter information, including data framing and data arbitration, to form a parameter sequence; and transmitting the parameter sequence to the Feiteng CPU processor through the PCIE QDMA interface for further data processing.
3. A transformer fault diagnosis method and system based on graph mutual mapping transmission network according to claim 1, characterized in that: In the step 1, the FPGA chip is KU115, and the parallel capability of FPGA is utilized to parallelly collect the concentration of dissolved gas in oil collected by the MEMS photoacoustic sensor, and parallel framing of multi-parameter data is performed in FPGA.
4. A transformer fault diagnosis method and system based on graph mutual mapping transmission network according to claim 3, characterized in that: The framing format is a 16-byte frame header plus 4096 bytes of data, the frame header is 0x55aaaa55+0x77888877+0xaaaaaaaa+4-byte command number, where 0xaaaaaaaa is the data type format, and the data type format of the frame header of gas concentration type 1 to gas concentration type 6 is: 0xaaaaaaaa, 0xbbbbbbbb, 0xcccccccc, 0xdddddddd, 0xeeeeeeee, 0xffffffff, and the 4-byte command number is for the CPU to identify the command number to start or stop reading data.
5. A transformer fault diagnosis method based on graph mutual mapping transmission network according to claim 1, characterized in that: In the step 1, the data arbitration is performed by inputting the framed data of each channel into the fifo for caching, and the framed data of the six gas concentrations are input into the xilinx AXI interconnect ip core, and the data interface is AXI-Stream, and a data arbitration operation is performed, that is, polling and reading the data packets of each type of data to transmit to the next level; the data transmission, through PCIEQDMA, the six gas data are transmitted through the AXI-Stream interface, and the polled data is transmitted to the Feiteng CPU for processing using the QDMA interrupt method.
6. A transformer fault diagnosis method based on graph mutual mapping transmission network according to claim 1, characterized in that: In step S2, the combination of the activation function and the mapping weight gradually approaches the optimal mapping result through three iterations, so that the mapped data has better expression ability and clustering characteristics in the new domain.
7. A transformer fault diagnosis method and system based on graph mutual mapping transmission network according to claim 1, characterized in that: The step 3 specifically includes: Each feature vector z in the mapped feature vector set Z′ i ′, regarded as a node v in the graph i , and use the value of the feature vector as the feature of the node to form a node set V, where V = {v1, v2, ..., v n }; The node features v extracted by sample mapping i The input is fed into a multi-layer perceptron (MLP), and then the adjacency matrix is obtained by matrix multiplication between the MLP features and their transpose, and the nearest node is selected using the TOp-k sorting mechanism. The process can be expressed as in, is the output of the multilayer perceptron, normalize(·) is the normalization function, and W represents the adjacency matrix formed during the graph generation process. is a sparse adjacency matrix, and Top-k(·) is the index that returns the maximum value of W row by row; The constructed graph structure is input into the spectral graph convolutional network, and the data features in the graph are extracted through the convolution kernel expanded by Chebyshev polynomials. The process is expressed as Among them, g θ is a graph filter, θ is a learnable parameter, Λ is the directional eigenvalue matrix of the Laplace moment, K determines the range of the node domain, where in is the eigenvalue of the Laplacian matrix; The node features of the graph structure are input into the multi-receptive field graph convolutional network, and multiple convolution operations are performed to obtain the fused features. The process is expressed as in, is a learnable parameter, and H represents the fused feature; Before the fused features are input into the fully connected layer, feature fusion and learning are performed first. This process is expressed as H0=MRFConv(WXB0) H1=MRFConv(WH0B1) Among them, H0 is the feature learned for the first time, H1 is the feature learned for the second time, WX and WH0 are the inputs of the first and second convolution operations respectively, and B0 and B1 are trainable weight matrices.
8. A transformer fault diagnosis method and system based on graph mutual mapping transmission network according to claim 1, characterized in that: The classification loss in step 4 is used to evaluate the prediction accuracy of the label classifier for the source domain samples; the domain alignment loss is used to evaluate the performance of the feature extractor in deceiving the domain discriminator; the structural alignment loss is used to evaluate the structural similarity between the source domain and the target domain data.
9. A transformer fault diagnosis system based on graph mutual mapping transmission network, characterized in that: include: Transformer parameter data acquisition module: It is used to use the MEMS photoacoustic sensor equipped with a non-resonant photoacoustic cell to analyze the sound wave signal through photoacoustic spectroscopy technology to obtain the parameter information of the dissolved gas in the transformer oil, and use FPGA to process the data to form a parameter sequence, and then transmit it to the CPU processor; Sample mapping and weight adjustment module: used to process source domain and target domain data using sample mapping technology in the CPU processor, map the source domain data to a new domain space through a combination of activation function and mapping weight, and adjust the mapping weight according to the new domain data, and repeat this process until the consistency of feature distribution of source domain and target domain data is achieved; The graph structure feature extraction module is used to construct a graph mutual mapping transmission network. The feature vector after sample mapping is regarded as a node, and its value is used as the node feature. The graph structure is automatically constructed based on the similarity, distance and other metrics between node features. The constructed graph structure is input into the multi-receptive field graph convolution network, and multiple convolution operations are performed to extract feature information of different scales in the graph structure. The unsupervised domain adaptation module is used to define a comprehensive objective function including classification loss, domain alignment loss and structural alignment loss, and implement unsupervised domain adaptation training based on the objective function.
10. The transformer fault diagnosis system based on graph mutual mapping transmission network according to claim 9 is characterized in that: The sample mapping and weight adjustment module is also used to minimize the residual sum of squares and optimize the clustering effect to find the optimal mapping result.
11. The transformer fault diagnosis system based on graph mutual mapping transmission network according to claim 9 is characterized in that: The unsupervised domain adaptation module optimizes the performance of the model by updating the parameters of the feature extractor, label classifier, and domain discriminator.