Transformer hybrid fault diagnosis method and diagnosis system

Abstract

The invention discloses a transformer hybrid fault diagnosis method and system, and the method comprises the steps: carrying out the amplification of a fault type sample with a small number of samples, endowing a higher sample generation weight to a sample located near a classification boundary, and increasing the sample size near the classification boundary; therefore, the model can learn more sample data features of which types are difficult to distinguish near the classification boundary, and the identification capability of complex faults is enhanced. The fault diagnosis model adopts windows of different sizes to slide input data to extract feature information of different scales, each window generates a plurality of sample subsequences, each sample subsequence adopts a random forest classifier and a complete random forest classifier to calculate and output probability vectors respectively, a feature matrix of the window is obtained through splicing, and the feature matrix of the window is obtained; performing global feature fusion and splicing on all window feature matrixes and original features to form an input feature matrix; and a diagnosis result is output after processing is carried out through the cascaded multi-layer forest network, so that a plurality of fault types of the plurality of transformers can obtain relatively good diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to power equipment fault diagnosis technology, specifically a method and system for diagnosing mixed faults in transformers. Background Technology

[0002] Power transformers are crucial equipment in power systems, enabling efficient energy transfer over long distances through voltage conversion. Their normal and safe operation is essential for ensuring a reliable and continuous power supply. Faults in transformers, such as discharge or violent combustion, can severely impact the power system's safety. Therefore, real-time monitoring of power transformer operating status, especially early detection of potential faults, and utilizing artificial intelligence technology for intelligent fault diagnosis are of paramount importance.

[0003] Traditional methods for transformer fault diagnosis based on dissolved gas analysis in insulating oil include the three-ratio method and the Rogers four-ratio method. These methods are simple and intuitive, suitable for preliminary fault diagnosis, and sensitive to local single faults, but they are difficult to effectively distinguish complex fault types. On the other hand, machine learning methods based on support vector machines and artificial neural networks, although the models have pattern recognition capabilities, are prone to overfitting or underfitting due to the scarcity of fault samples in actual transformer operation and the significant differences in the frequency of occurrence of various faults. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defect that sample imbalance affects the accuracy of multiple fault mode identification, and to provide a method and system for diagnosing mixed faults of a compressor.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a hybrid fault diagnosis method for transformers, which uses the dissolved gas concentration in transformer insulating oil as the observed variable and the transformer fault type as the output, acquires data samples to train a fault diagnosis model, uses the trained fault diagnosis model to process the real-time acquired transformer observed variable data to output diagnostic results, transforms and amplifies samples in fault types with a small number of samples to generate new samples, wherein samples with higher proximity density near the classification boundary are given higher sample generation weights to increase the number of samples near the classification boundary; the fault diagnosis model first slides windows of different sizes across the input data to extract feature information at different scales, and each window generates multiple sample subsequences; each sample subsequence of each window uses a random forest classifier and a fully random forest classifier to calculate the output probability vector, and then concatenates the output probability vectors of all samples in the window to obtain the feature matrix of the window, and performs global feature fusion and concatenation of the feature matrices of all windows with the original features to form the input feature matrix; the input feature matrix is ​​processed by a cascaded multi-layer forest network to output the diagnostic results.

[0006] The observed variables are the gas concentrations of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide in the insulating oil, and the output diagnostic result is a fault type code.

[0007] New samples are synthesized from samples in the fault type sample set using the following formula: ,in, For the synthesized new sample, This is a sample of this fault type. for Random nearest neighbors in the sample of this fault type Uniformly distributed random numbers, It is zero-mean Gaussian noise. It is a diagonal matrix formed by the standard deviations of each dimension.

[0008] The proportion of new samples generated from each sample is determined by calculating the nearest neighbor density distribution of each sample in a sample set of the same fault type. Number of new samples generated ,in for The proportion of generated samples, This represents the total number of samples that need to be generated for this type of fault. in, , For adjacent density distributions, Denotes the set of k nearest neighbors in the majority class. This represents the sample size of the fault category with the largest number of samples. This represents the original sample size of the minority samples in class c. This is the proportionality coefficient.

[0009] The random forest classifier, for training each subsequence and each tree, randomly selects *a* features from the original feature dimension *D*, uses Gini impurity as the splitting criterion, and selects the optimal features and splitting threshold. For each input sample, each tree outputs a class label. The final probability vector is calculated based on the voting proportions of all trees. ; in, To predict the number of trees of class c, The total number of trees; The fully random forest classifier randomly generates feature selection and splitting thresholds, and outputs a probability vector. ; The feature matrix of each window is the probability vector of all samples. and probability vector splicing.

[0010] In the cascaded multi-layer forest network, each layer contains an independently trained random forest and a completely random forest. The input of the first layer is the input feature matrix, and the input of the next layer is the concatenation vector of the output of the previous layer and the input feature matrix.

[0011] In a cascaded multi-layer forest network, each layer calculates the mean probability output of all forests in that layer to obtain the ensemble probability matrix. As output ,in, Output the class probability matrix for forest k in this layer. This represents the number of forests in this layer.

[0012] The acquired gas concentration data was normalized, and the normalized data sequence was then decomposed using wavelet decomposition for noise reduction.

[0013] A feature matrix composed of multiple observed gas concentration variables is used as training samples. The training samples are then used to generate multiple sample subsequence feature matrices through window sliding. For each sample subsequence feature matrix, the global average pooling of each gas feature is calculated. and global max pooling ,Will and After concatenation, the input is fed into a fully connected layer, and a feature weight vector is generated through an activation function. The sample subsequence feature matrix is ​​multiplied by the generated feature weight vector and then used to calculate the probability vector through a random forest classifier and a fully random forest classifier.

[0014] A diagnostic system used in a transformer hybrid fault diagnosis method includes multiple gas sensors for real-time monitoring of dissolved gases in transformer insulating oil, a data processor for processing the monitoring data from the gas sensors, and a diagnostic server for receiving the output data from the data processor and using it for fault diagnosis. The data processor is installed on the transformer and is communicatively connected to the diagnostic server located in the cloud. The data processor processes the monitoring data from the multiple gas sensors and combines it into a feature vector, which is then sent to the diagnostic server. The diagnostic server deploys a fault diagnosis model and generates diagnostic results based on the received monitoring data.

[0015] The beneficial effects of this invention are: in the process of expanding minority class samples, by assigning higher sample generation weights to samples located near the classification boundary, the model can learn more sample data features that are difficult to distinguish near the classification boundary, thereby enhancing the model's ability to identify complex faults. While expanding minority class samples, the hierarchical structure of the deep forest network is used to improve the model's robustness to imbalanced data, achieving good diagnostic accuracy for various fault types in multi-transformer systems. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the fault diagnosis model of the present invention performing a window sliding scan on the input data.

[0017] Figure 2 This is a schematic diagram of the cascaded multi-layer forest network of this invention.

[0018] Figure 3 This is a diagnostic result confusion matrix. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The specific contents listed in the following embodiments are not limited to the technical features necessary to solve the technical problem described in the claims. Furthermore, the enumerations are merely a part of the present invention, and not all of the embodiments.

[0020] The transformer hybrid fault diagnosis method of the present invention selects the dissolved gas concentration in the transformer insulating oil as the observed variable. When abnormal operating conditions occur inside the transformer, different faults will produce specific combinations of dissolved gas concentrations in the oil, which can be used to determine the fault type. The specific observed variables are shown in Table 1.

[0021] Transformer faults can be characterized by various typical patterns, such as high-temperature overheating, medium-temperature overheating, low-energy discharge, high-energy discharge, and mixed faults including high-energy discharge combined with high-temperature overheating, low-energy discharge combined with overheating, and high-temperature overheating combined with high-energy discharge. After determining the observed variables and fault diagnosis types, a stratified sampling method is used to divide the fault sample dataset to ensure a reasonable distribution of each type of fault sample in the training and test sets. Specifically, the dataset is first stratified according to fault type, and then each type of fault sample is independently randomly partitioned from 80% to 20%: for fault categories with more than one sample, 80% of the samples are assigned to the training set and 20% to the test set; if a fault category contains only one sample, all of it is included in the training set. Due to the significant differences in sample size among different transformer fault types, the number of samples for some minority fault categories is severely insufficient, affecting the diagnostic accuracy of these types of faults.

[0022] This invention increases the sample size by amplifying a small number of fault samples. The amplified samples are synthesized based on the original sample set of the fault type. The formula for synthesizing the new samples is as follows: in, For the synthesized new sample, This is a sample of this fault type. for Random nearest neighbors in the sample of this fault type Uniformly distributed random numbers ( ), It is zero-mean Gaussian noise. It is a diagonal matrix formed by the standard deviations of each dimension.

[0023] During the synthesis of new samples, weights are assigned based on the importance of the samples, with a focus on generating difficult samples near the classification boundary to enhance the model's ability to identify complex faults. Specifically, samples with higher proximity density near the classification boundary are given higher sample generation weights. When generating new samples, samples with higher weights generate more new samples, thereby increasing the number of samples near the classification boundary.

[0024] For a small number of fault types with a limited sample size, calculate the total number of samples that need to be generated: in, This represents the sample size of the fault category with the largest number of samples. This represents the original sample size of the minority samples in class c. This is the proportionality coefficient.

[0025] For each sample Calculate its nearest neighbor density distribution: in, Denotes the set of k nearest neighbors in the majority class. is a hyperparameter representing the number of nearest neighbors.

[0026] According to the sample The proportion of new samples generated is: in, , It is a close-proximity density distribution.

[0027] According to the sample The number of new samples generated is: in, for The proportion of generated samples, The total number of samples that need to be generated for this fault type.

[0028] In the expanded sample set, each type of fault sample was independently randomly partitioned from 80% to 20%, with 80% of the samples going to the training set and 20% to the test set. To avoid bias caused by data order, the training and test sets were randomly shuffled after the partitioning. The partitioned training and test sets were then used to train the fault diagnosis model.

[0029] The main body of the fault diagnosis model adopts a cascaded multi-layer forest network model. Multiple gas concentration data are used as inputs to the fault diagnosis model, and the transformer fault type is used as the output.

[0030] Let the feature data matrix of the input fault diagnosis model be... Where N is the number of samples, D is the feature dimension, and the corresponding classification label is... This represents C fault categories as the model's output. The data input to the fault diagnosis model is first processed by sliding windows of different sizes across the data to extract feature information at different scales. For example... Figure 1 As shown, the process is as follows: (1) Generating a sliding window Features Slide your mouse over the data using windows of different sizes, for example, using a window of size... The window slides to generate subsequences, each subsequence being... Each window generates n sample subsequences.

[0031] (2) Subsequence classification prediction For each sample subsequence in each window, the output probability vector is calculated using both a random forest classifier and a fully random forest classifier, where: Random Forest Classifier (RFC): Feature selection: For each subsequence, for the training of each tree, 'a' features are randomly selected from the original feature dimension D (usually...). or ).

[0032] Node splitting: Gini impurity is used as the splitting criterion, and the optimal feature and splitting threshold are selected. It is a node Medium category The sample proportion.

[0033] in, It is a node Medium category The sample proportion.

[0034] Probability calculation: For the input sample Each tree outputs a category label. The final probability vector Calculated based on the voting proportions of all trees: in, To predict the number of trees of class c, This represents the total number of trees.

[0035] Completely Random Forest Classifier (TRFC): A fully random forest classifier randomly generates features and selects split thresholds, without calculating Gini impurity. Probability calculation is similar to RFC, and the output is a probability vector. , indicating that the predicted probability distribution of the i-th completely random tree for the s-th sample is a C-dimensional vector.

[0036] (3) Window-level feature splicing The RFC probability vector of all samples within each window and TRFC probability vector Concatenate into a window feature matrix : Where n is the number of samples covered by the window.

[0037] (4) Feature fusion Compare the feature matrices of all windows with the original features Perform global feature fusion and concatenate the results to form the input feature matrix. : As input to a cascaded forest network.

[0038] The cascaded forest network is the core of the fault diagnosis model of this invention. Figure 2 This is a schematic diagram of a cascaded forest structure, which includes multiple cascaded forest networks. Each layer of the forest network contains multiple independently trained random forests and fully random forests. The input feature matrix is ​​shown above. The input is fed into the first layer of the cascaded forest network. Starting from the second layer of the cascaded forest network, the input to each layer is the predicted vector output by the previous layer and the input feature matrix. The vector is formed by concatenation. The cascading process of the cascaded forest network continues until a set number of layers is reached, or the performance improvement of the model between three consecutive layers is no longer significant.

[0039] Let the number of layers in the cascaded forest network be . Layers, each layer contains Random forest / Fully random forest. Layer input feature matrix Initial layer input , Indicates the first The feature dimension of the input feature matrix for each layer (dynamically increases with the number of cascaded layers). Each forest k outputs a class probability matrix. In a cascaded multi-layer forest network, the mean probability output of all forests in each layer is calculated to obtain the ensemble probability matrix. As the output of this layer.

[0040] in, Output the class probability matrix for forest k in this layer. This represents the number of forests in this layer.

[0041] The output ensemble probability matrix With the input feature matrix That is, the initial layer input Feature concatenation is performed and used as input for the next layer.

[0042] A termination condition is set for the model: training terminates when the classification accuracy on the validation set fails to improve for T consecutive layers. The trained fault diagnosis model is then used to process real-time transformer observation data and outputs diagnostic results based on the data characteristics.

[0043] In the following embodiments of the present invention, the concentrations of seven characteristic gases of transformers—hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), carbon monoxide (CO), and carbon dioxide (CO2)—are used as observed variables. Different gas contents correspond to different fault types. Seven fault types of transformers—single fault low-energy discharge, high-energy discharge, partial discharge, medium-temperature overheating, high-temperature overheating, and mixed faults of low-energy discharge and overheating, and high-temperature overheating and high-energy discharge—are encoded as model outputs. The specific correspondence is shown in Table 2.

[0044] This invention acquired 434 sets of transformer fault data collected on-site, and used a stratified sampling method to divide them into a training set (353 records, 81.3%) and a test set (81 records, 18.7%) at an 8:2 ratio. For minority class samples, the above method was used to expand the sample size so that the sample size for all fault types reached more than 120 records, as shown in Table 3.

[0045] To remove noise from the sensor-acquired data, the acquired gas concentration data is normalized, and then wavelet decomposition is performed on the data sequence for denoising. The Daubechies wavelet (dbN) is chosen for wavelet processing due to its compact support and approximate symmetry, making it suitable for capturing abrupt changes in gas concentration, such as the sudden increase in C2H2 during a discharge fault. In this embodiment, the 4th-order Daubechies wavelet (db4) is selected. For concentration data with significant high-frequency noise, it filters background noise, preserves the macroscopic trend, and does not lose key local abrupt changes. The normalized data is decomposed into four layers using the db4 wavelet, yielding approximation coefficients A4 and detail coefficients D1-D4. A layered thresholding method is used to progressively lower the denoising threshold. Higher thresholds are used for stronger denoising in the high-frequency layers to filter random noise, while lower thresholds are used in the mid-frequency layers to avoid feature loss. Based on testing, the preferred thresholds in this invention are set as follows: , , , ,in The noise standard deviation is estimated using the median of the first-level detail coefficients. D1 represents the detail coefficients of the first layer of wavelet decomposition, and 0.6745 is the distance coefficient from the median to the mean in a standard normal distribution. By reasonably selecting the threshold, the gas features after wavelet decomposition (such as the amplitude of C2H2 mutations and the slope of CO trend) can be made closer to the real fault modes, thereby improving the accuracy of subsequent deep forest models.

[0046] The feature matrix composed of denoised gas concentration data is used as training samples. Multiple sample sub-sequence feature matrices are generated by sliding the window through these training samples. To enhance the model's sensitivity to key gas features, channel attention is added to the gas features of each sub-sequence after the sliding window generates the sub-sequences. For each sample sub-sequence feature matrix, global average pooling is calculated for each gas feature. and global max pooling ,Will and After concatenation, a weight vector is generated through a fully connected layer. , where D is the gas characteristic dimension.

[0047] in, , This is the weight matrix. This is the ReLU activation function.

[0048] weight matrix vector The sample subsequence feature matrix is ​​used to highlight key gas features. After attention is added, the output probability vectors of the sample subsequences are calculated using random forest classifiers and fully random forest classifiers respectively. These vectors are concatenated into a window feature matrix, globally fused, and then input into a cascaded multilayer forest network for processing.

[0049] In the multi-layer forest network, each layer contains 10 forests, and each forest consists of 200 trees. During training, the model achieved its highest validation accuracy of 88.581% in the first layer, followed by a performance decline in the next two layers (86.967% for the second layer and 86.159% for the third layer), triggering an early stopping mechanism. Ultimately, the optimal number of layers was determined to be two. The test set feature variables were input into the trained fault diagnosis model for fault diagnosis, and the transformer fault diagnosis results are shown in Table 4. Figure 3 As shown.

[0050] According to the test results, the method of the present invention can achieve a good fault diagnosis accuracy for various fault types and can be applied to the real-time detection of transformer faults.

[0051] The transformer hybrid fault diagnosis system of this invention can adopt an edge device combined with cloud service architecture. Multiple gas sensors are installed on the transformer to monitor the concentration of various dissolved gases in the transformer insulating oil in real time. A data processor is set up at the transformer site to process the monitoring data from the gas sensors in a timely manner, such as data normalization and noise reduction. The diagnostic server, which deploys the fault diagnosis model, is located in the cloud and communicates with the data processor. The data processor processes the monitoring data from multiple gas sensors and combines it into a feature vector, which is then sent to the diagnostic server. The diagnostic server uses the fault diagnosis model to extract data features and outputs diagnostic results.

[0052] The above description of specific embodiments is only for the purpose of helping to understand the technical concept and core idea of ​​the present invention. Although specific preferred embodiments have been used to describe and illustrate the technical solutions, they should not be construed as limiting the present invention itself. Those skilled in the art can make various changes in form and detail without departing from the technical concept of the present invention. These easily conceived changes or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A hybrid fault diagnosis method for transformers, comprising using the dissolved gas concentration in transformer insulating oil as an observed variable, the transformer fault type as an output, acquiring data samples to train a fault diagnosis model, and using the trained fault diagnosis model to process real-time acquired transformer observed variable data to output diagnostic results, characterized in that: For fault types with a small sample size, samples are transformed and amplified to generate new samples. Samples with higher proximity density near the classification boundary are assigned higher generation weights to increase the sample size near the classification boundary. The fault diagnosis model first slides windows of different sizes across the input data to extract feature information at different scales, with each window generating multiple sample subsequences. For each sample subsequence of a window, output probability vectors are calculated using both random forest and fully random forest classifiers. Then, the output probability vectors of all samples in that window are concatenated to obtain the feature matrix of that window. The feature matrices of all windows are then globally fused with the original features to form the input feature matrix. The input feature matrix is ​​processed by a cascaded multi-layer forest network to output the diagnostic results.

2. The transformer hybrid fault diagnosis method as described in claim 1, characterized in that: The observed variables are the gas concentrations of hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide, and carbon dioxide in the insulating oil, and the output diagnostic result is a fault type code.

3. The transformer hybrid fault diagnosis method as described in claim 1, characterized in that: New samples are synthesized from samples in the fault type sample set using the following formula: ,in, For the synthesized new sample, This is a sample of this fault type. for Random nearest neighbors in the sample of this fault type Uniformly distributed random numbers, It is zero-mean Gaussian noise. It is a diagonal matrix formed by the standard deviations of each dimension.

4. The transformer hybrid fault diagnosis method as described in claim 3, characterized in that: The proportion of new samples generated from each sample is determined by calculating the nearest neighbor density distribution of each sample in a sample set of the same fault type. Number of new samples generated ,in for The proportion of generated samples, This represents the total number of samples that need to be generated for this type of fault. in, , For adjacent density distributions, Denotes the set of k nearest neighbors in the majority class. This represents the sample size of the fault category with the largest number of samples. This represents the original sample size of the minority samples in class c. This is the proportionality coefficient.

5. The transformer hybrid fault diagnosis method as described in claim 1, characterized in that: The random forest classifier, for training each subsequence and each tree, randomly selects *a* features from the original feature dimension *D*, uses Gini impurity as the splitting criterion, and selects the optimal features and splitting threshold. For each input sample, each tree outputs a class label. The final probability vector is calculated based on the voting proportions of all trees. ; in, To predict the number of trees of class c, The total number of trees; The fully random forest classifier randomly generates feature selection and splitting thresholds, and outputs a probability vector. ; The feature matrix of each window is the probability vector of all samples. and probability vector splicing.

6. The transformer hybrid fault diagnosis method as described in claim 1, characterized in that: In the cascaded multi-layer forest network, each layer contains an independently trained random forest and a completely random forest. The input of the first layer is the input feature matrix, and the input of the next layer is the concatenation vector of the output of the previous layer and the input feature matrix.

7. The transformer hybrid fault diagnosis method as described in claim 6, characterized in that: In a cascaded multi-layer forest network, each layer calculates the mean probability output of all forests in that layer to obtain the ensemble probability matrix. As output ,in, Output the class probability matrix for forest k in this layer. This represents the number of forests in this layer.

8. The transformer hybrid fault diagnosis method as described in claim 1, characterized in that: The acquired gas concentration data was normalized, and the normalized data sequence was then decomposed using wavelet decomposition for noise reduction.

9. The transformer hybrid fault diagnosis method as described in claim 1, characterized in that: A feature matrix composed of multiple observed gas concentration variables is used as training samples. The training samples are then used to generate multiple sample subsequence feature matrices through window sliding. For each sample subsequence feature matrix, the global average pooling of each gas feature is calculated. and global max pooling ,Will and After concatenation, the input is fed into a fully connected layer, and a feature weight vector is generated through an activation function. The sample subsequence feature matrix is ​​multiplied by the generated feature weight vector and then used to calculate the probability vector through a random forest classifier and a fully random forest classifier.

10. The diagnostic system used in the transformer hybrid fault diagnosis method as described in claim 1, characterized in that: The system includes multiple gas sensors for real-time monitoring of dissolved gases in transformer insulating oil, a data processor for processing the monitoring data from the gas sensors, and a diagnostic server for receiving the output data from the data processor and using it for fault diagnosis. The data processor is installed on the transformer and communicates with the diagnostic server located in the cloud. The data processor processes the monitoring data from the multiple gas sensors and combines it into a feature vector, which is then sent to the diagnostic server. The diagnostic server deploys a fault diagnosis model and generates diagnostic results based on the received monitoring data.

Images