Transformer voiceprint signal fault diagnosis method based on graph attention network
The transformer acoustic signature fault diagnosis method based on graph attention network solves the problems of accuracy and robustness of transformers under complex operating conditions, and realizes efficient fault feature extraction and classification, which is applicable to equipment monitoring and maintenance in power systems.
Patent Information
- Application Number
- CN202510745549.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-05
- Publication Date
- 2025-11-07
AI Technical Summary
Existing transformer fault diagnosis technologies lack accuracy and robustness under complex and variable operating conditions, while acoustic signal monitoring methods have insufficient generalization ability and feature extraction.
A fault diagnosis method for transformer acoustic signature signals based on graph attention networks is adopted. The acoustic signature signals are processed in segments, a spectrum feature association graph is constructed using the K-nearest neighbor algorithm, and feature weighting and aggregation are performed using graph attention networks. Finally, a fully connected neural network is used for fault classification.
It significantly improves the efficiency of fault feature extraction and the robustness of the model, and can accurately identify and classify different transformer fault types, making it suitable for complex and ever-changing power system operating conditions.
Smart Images
Figure CN120913588A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of power equipment state monitoring, and particularly relates to a transformer voiceprint signal fault diagnosis method based on a graph attention network. BACKGROUND
[0002] With the complication of power systems and the increase of power demand, the operation environment and working conditions of transformers become increasingly complex, resulting in an increase in their fault risk. The transformer state monitoring method in the prior art often shows certain limitations when dealing with complex and variable working conditions. Traditional fault diagnosis techniques mainly rely on single features such as vibration, temperature or current signals, and these methods lack accuracy and robustness when dealing with complex and variable working conditions. In recent years, machine learning and deep learning technologies have rapidly developed, providing new solutions for transformer fault diagnosis, among which the voiceprint feature monitoring method based on sound signals has gradually been widely applied due to its non-invasiveness and strong environmental adaptability.
[0003] The voiceprint signal monitoring method can detect the running state and fault characteristics of the equipment by capturing the sound signals emitted by the transformer during operation. Existing researches mainly focus on using convolutional neural networks (CNN) or long short-term memory networks (LSTM) to process voiceprint signals. However, these methods still have problems of insufficient generalization ability and insufficient feature extraction when dealing with complex and variable fault environments. SUMMARY
[0004] In order to solve the problem of insufficient accuracy and robustness of traditional transformer fault diagnosis techniques when dealing with complex and variable working conditions, and to solve the problem of insufficient generalization ability and insufficient feature extraction of existing voiceprint signal monitoring methods, the application provides a transformer voiceprint signal fault diagnosis method based on a graph attention network, which can better capture key features in complex signals, thereby improving the accuracy and robustness of fault diagnosis.
[0005] The technical scheme adopted by the transformer voiceprint signal fault diagnosis method based on a graph attention network of the application is as follows:
[0006] A transformer voiceprint signal fault diagnosis method based on a graph attention network, comprising the following steps,
[0007] S1, dividing the transformer voiceprint signal in time sequence to form a plurality of time sequence samples;
[0008] S2, performing data normalization processing on each time segment, and then converting the time domain signal into a frequency domain signal by using fast Fourier transform to obtain spectral features;
[0009] S3, constructing a correlation graph of the spectrum features by using a K-neighbor algorithm;
[0010] S4, weighting and aggregating the features of the nodes in the correlation graph by using a graph attention network;
[0011] S5, performing transformer fault classification and diagnosis based on the aggregated features; the node aggregation features output by the graph attention network are input into a fully connected neural network, high-order features are extracted, the probability distribution of each category is calculated through a Softmax function, and a diagnosis result is output according to the category corresponding to the maximum probability.
[0012] Further improvement of the technical scheme of the application is that each time sequence sample in the step S1 represents a voiceprint signal in a certain time.
[0013] Further improvement of the technical scheme of the application is that the step S2 is specifically,
[0014] The minimum-maximum normalization processing is performed on each time sequence sample to ensure the comparability of signal amplitudes between different segments; then, the normalized time domain signal is converted into a frequency domain signal by using a fast Fourier transform to obtain the spectrum features, as shown in formula (1):
[0015]
[0016] Wherein, X(f) represents the frequency domain signal, x(n) is the time segment signal, and N is the signal length.
[0017] Further improvement of the technical scheme of the application is that the step S3 is specifically,
[0018] Each spectrum feature is regarded as a node in the graph, the similarity between the nodes is determined by using a K-neighbor algorithm, a connection edge is established, and a correlation graph is formed; wherein, the selection of the K value is determined according to experimental results to balance the accuracy and the calculation complexity.
[0019] Further improvement of the technical scheme of the application is that the step S4 is specifically,
[0020] The graph attention network can dynamically adjust the weights of the nodes by calculating the attention coefficients between the center node and the neighbor nodes, and then focus on the key feature area in the voiceprint signal; formula (2) and formula (3) represent the calculation method of the attention coefficients:
[0021]
[0022] e ij =a([Wh i ||Wh j ]) (3)
[0023] wherein, alpha ij is the attention coefficient between node i and node j, W is a learnable weight matrix, a is a vector parameter in the attention mechanism, h i and h j are node feature vectors.
[0024] Thanks to the above technical solutions, the technical progress achieved by the present application includes:
[0025] The present application first applies the graph attention network (GAT) to the transformer voiceprint signal fault diagnosis, uses the feature weighting mechanism of GAT-TASFDM, significantly improves the extraction efficiency of fault features and the robustness of the model, and through segmented processing of the voiceprint signal and using the K nearest neighbor algorithm to construct a correlation graph, the method of the present application can better capture the complex feature changes in the voiceprint signal; at the same time, the method of the present application converts the time domain signal into frequency domain features through fast Fourier transform (FFT), and in combination with the training of the GAT model, can accurately identify and classify different transformer fault types. BRIEF DESCRIPTION OF DRAWINGS
[0026] Figure 1 is a diagnostic flowchart of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application;
[0027] Figure 2 is a process diagram of establishing a correlation graph of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application;
[0028] Figure 3 is a diagram of calculating the attention coefficient in the GAT model of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application;
[0029] Figure 4 is a GAT-TASFDM model structure diagram of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application;
[0030] Figure 5 is an accuracy curve of the GAT-TASFDM model of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application in the training process;
[0031] Figure 6 is a t-SNE-based GAT-TASFDM feature visualization diagram of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application;
[0032] Figure 7 is a comparison diagram of the test set accuracy of a transformer voiceprint signal fault diagnosis method based on a graph attention network of the present application and other four common fault diagnosis methods.
[0033] Figure 8 is a transformer voiceprint signal fault diagnosis method based on a graph attention network. The false positive (FP) and false negative (FN) values of the four common fault diagnosis methods and the transformer voiceprint signal fault diagnosis method based on the graph attention network are compared for different fault types. DETAILED DESCRIPTION
[0034] To make the objects, technical solutions, and advantages of the present application clearer, further detailed description will be given below in conjunction with specific embodiments and with reference to the accompanying drawings. In the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present application.
[0035] As shown in Figure 1 The embodiment provides a transformer voiceprint signal fault diagnosis method based on a graph attention network. The method comprises the following steps:
[0036] S1, segmenting the transformer voiceprint signal in time sequence to form a plurality of time sequence samples.
[0037] Specifically, first, the transformer voiceprint signal is segmented in time sequence to form a plurality of independent small samples. Each sample represents the voiceprint signal in a certain time. This segmentation method can capture the state change of the transformer at different time points, which helps to improve the accuracy of fault feature extraction.
[0038] S2, performing data normalization processing on each time segment, and then converting the time domain signal into a frequency domain signal by using fast Fourier transform to obtain the spectrum feature.
[0039] Specifically, the minimum-maximum normalization processing is performed on each time segment to ensure the comparability of the signal amplitudes between different segments. Subsequently, the normalized time domain signal is converted into a frequency domain signal by using fast Fourier transform (FFT) to obtain the spectrum feature, as shown in formula (1):
[0040]
[0041] Wherein, X(f) represents the frequency domain signal, x(n) is the time segment signal, and N is the signal length. Through frequency domain conversion, the frequency components in the voiceprint signal can be more clearly analyzed, and the characteristic frequencies of different fault types are revealed.
[0042] S3, constructing a correlation graph of the spectrum feature by using a K-nearest neighbor algorithm.
[0043] As shown in Figure 2As shown, this step reveals the intrinsic connections between audioprint signals by constructing a correlation graph between spectral features. Each spectral feature is treated as a node in the graph, and the similarity between nodes is determined using the K-Nearest Neighbors (KNN) algorithm, establishing connecting edges to form the correlation graph. The value of K is chosen based on experimental results to achieve a balance between accuracy and computational complexity.
[0044] S4. Use a graph attention network to perform feature weighting and aggregation on the nodes in the association graph.
[0045] like Figure 3 As shown, the GAT model performs feature weighting and aggregation on the nodes in the association graph. Specifically, by calculating the attention coefficient between the central node and its neighboring nodes, GAT can dynamically adjust the weights of each node, thereby focusing on the key feature regions in the voiceprint signal. Equations (2) and (3) represent the calculation method of the attention coefficient:
[0046]
[0047] e ij =a([Wh i ||Wh j (3)
[0048] Where, α ij Let be the attention coefficient between node i and node j, W be the learnable weight matrix, a be the vector parameters in the attention mechanism, and h be the vector parameters in the attention mechanism. i and h j These are the node feature vectors.
[0049] S5. Transformer fault classification and diagnosis based on aggregated features: Input the node aggregated features output by the graph attention network into the fully connected neural network, extract high-order features, calculate the probability distribution of each category using the Softmax function, and output the diagnosis result according to the category corresponding to the highest probability.
[0050] Model training and performance validation
[0051] like Figure 4 As shown, by training the GAT-TASFDM, this invention can effectively identify and classify different transformer fault types. During the training process, the model parameters are continuously optimized to maximize the model's classification accuracy on the test data. Figure 5 The accuracy curves during training are shown, indicating that the model gradually converges during training and eventually reaches a high level of accuracy.
[0052] 1) Visual analysis of fault characteristics:
[0053] like Figure 6As shown, the t-SNE algorithm is used to visualize the fault features extracted by the GAT model. The t-SNE algorithm can reduce high-dimensional data to two dimensions, thereby clearly showing the feature distribution of different fault types. The results show that the voiceprint features of different fault categories form an obvious clustering structure in the two-dimensional space, further verifying the effectiveness of the method.
[0054] 2) Performance comparison and result analysis:
[0055] As shown in Figure 7 and Figure 8 , the performance of the method of the present application is compared with other common fault diagnosis methods. The results show that the GAT-TASFDM method is superior to the traditional method in terms of classification accuracy and false classification rate (FP and FN), demonstrating its excellent performance in complex fault diagnosis tasks.
[0056] In the above embodiments, the present application provides a transformer voiceprint signal fault diagnosis method based on a graph attention network. By combining the GAT model and the extraction of spectral features, the present application successfully realizes accurate classification and diagnosis of transformer faults. The present application not only improves the accuracy and robustness of fault diagnosis, but also demonstrates its advantages in processing complex and diverse signals; it is suitable for equipment monitoring and maintenance tasks under various complex working conditions in power systems.
[0057] The above-described embodiments are merely preferred embodiments of the present application and do not limit the concept and scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design concept of the present application shall fall within the scope of protection of the present application, and the technical content claimed by the present application has been fully recorded in the claims.
Claims
1. A transformer voiceprint signal fault diagnosis method based on a graph attention network, characterized by: The method comprises the following steps, S1, dividing the transformer voiceprint signal according to time sequence to form a plurality of time sequence samples; S2, performing data normalization processing on each time segment, and then converting the time domain signal into a frequency domain signal by using fast Fourier transform to obtain a spectrum feature; S3, constructing a correlation graph of the spectrum feature by using a K nearest neighbor algorithm; S4, performing feature weighting and aggregation on the nodes in the correlation graph by using a graph attention network; S5, performing transformer fault classification diagnosis based on the aggregated features; inputting the node aggregation features output by the graph attention network into a fully connected neural network, extracting high-order features, calculating the probability distribution of each category by using a Softmax function, and outputting a diagnosis result according to the category corresponding to the maximum probability.
2. The transformer voiceprint signal fault diagnosis method based on a graph attention network according to claim 1, characterized in that: Each time sequence sample in the step S1 represents a voiceprint signal in a certain time.
3. The transformer voiceprint signal fault diagnosis method based on a graph attention network according to claim 1, characterized in that: The step S2 is specifically, performing minimum-maximum normalization processing on each time sequence sample to ensure the comparability of signal amplitudes between different segments; then, converting the normalized time domain signal into a frequency domain signal by using fast Fourier transform to obtain a spectrum feature, as shown in formula (1): Wherein, X(f) represents the frequency domain signal, x(n) is the time segment signal, and N is the signal length.
4. The transformer voiceprint signal fault diagnosis method based on a graph attention network according to claim 1, characterized in that: The step S3 is specifically, each spectrum feature is regarded as a node in the graph, the similarity between the nodes is determined by using a K nearest neighbor algorithm, and a connection edge is established to form a correlation graph; wherein, the selection of the K value is determined according to the experimental results to balance the accuracy and the calculation complexity.
5. The transformer voiceprint signal fault diagnosis method based on a graph attention network according to claim 1, characterized in that: The step S4 is specifically, by calculating the attention coefficients between the center node and its neighbor nodes, the graph attention network can dynamically adjust the weights of the nodes, and then focus on the key feature area in the voiceprint signal; formula (2) and formula (3) represent the calculation method of the attention coefficients: e ij = a([Wh i || Wh j ]) (3) wherein a ij is the attention coefficient between node i and node j, W is a learnable weight matrix, a is a vector parameter in the attention mechanism, h i and h j are node feature vectors.