Power transformer fault diagnosis method based on improved variational quantum shadow learning
By using an improved variable quantum shadow learning framework, the problems of accuracy and computational complexity in power transformer fault diagnosis are solved, achieving efficient fault identification and diagnosis and ensuring the reliability of the power system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2024-07-04
- Publication Date
- 2026-06-26
AI Technical Summary
Existing machine learning-based power transformer fault diagnosis methods are insufficient to meet the accuracy requirements of fault diagnosis and are difficult to handle complex computational tasks, especially when processing power transformer monitoring data, which faces the problems of high dimensionality and nonlinear relationships.
An improved variable quantum shadow learning (VQSL) framework is adopted, which uses parameter-shared local quantum circuits to act in parallel on a subset of qubits of the same width. Combined with a fully connected neural network, it is used to preprocess and extract features from dissolved gas data in transformer oil, build a fault diagnosis model, and accelerate computation by utilizing quantum parallelism and entanglement.
It improves the accuracy and efficiency of fault diagnosis, can more comprehensively extract the complex nonlinear relationship between oil chromatography data and transformer operating status, reduces computing resource requirements, and ensures the reliable operation of power transformers and the stability of power systems.
Smart Images
Figure CN118861812B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power grid fault diagnosis technology, and in particular to a power transformer fault diagnosis method based on improved variable quantum shadowing learning. Background Technology
[0002] With the significant growth in domestic electricity demand and the increasing demands for power system reliability, ensuring the stable operation of power equipment in the power system has become a crucial research direction. As a key piece of power equipment in the transmission and transformation process, the normal operation of power transformers has a decisive impact on the safety and stability of the entire power system.
[0003] Dissolved gas analysis (DGA) in insulating oil provides crucial information for online monitoring of transformer operating conditions. By analyzing the content and concentration of characteristic gases in insulating oil, researchers have proposed a series of DGA methods for diagnosing transformer operating conditions. These methods include the gas index method, the Duval trigonometric method, the IEC ratio method, and the Rogers ratio method. Existing methods are often limited by imperfect data encoding and absolute thresholds, leading to misdiagnosis of certain fault types or even situations where no diagnosis is possible, failing to meet the requirements for fault diagnosis accuracy. With the continuous development of artificial intelligence theory, machine learning techniques have been widely applied to build fault diagnosis models based on DGA data, including neural networks, support vector machines, clustering, and decision trees. However, these machine learning methods still face a series of difficulties and challenges when processing power transformer monitoring data. Specifically, the monitoring data exhibits highly complex nonlinear relationships and high dimensionality in the feature space. Applying these machine learning algorithms requires large-scale computing resources, and due to limitations in the computing power of classical computers, it may be difficult to effectively handle complex nonlinear relationships. In addition, the high-dimensional feature space of the monitoring data increases the complexity of data processing, usually requiring the use of dimensionality reduction techniques and feature selection to improve computational efficiency. This process may be accompanied by the loss of important information.
[0004] Therefore, a new technical solution is urgently needed to address the technical problems that existing machine learning-based power transformer fault diagnosis methods cannot meet the accuracy requirements of fault diagnosis and cannot handle complex computational tasks. Summary of the Invention
[0005] This invention provides a power transformer fault diagnosis method based on improved variable quantum shadow learning, which solves the technical problems that existing power transformer fault diagnosis methods based on machine learning methods are difficult to meet the fault diagnosis accuracy requirements and are difficult to handle complex computational tasks.
[0006] To achieve the above objectives, this invention provides a power transformer fault diagnosis method based on improved variable quantum shadow learning, comprising:
[0007] Historical data on dissolved gases in transformer oil are collected and combined with preset transformer status information to form a first dataset; the first dataset is preprocessed to obtain a second dataset; the second dataset is divided according to a preset ratio to obtain a training set and a test set;
[0008] The first model is trained a preset number of times based on the training set, and its performance is evaluated based on the test set after each training session. If the performance evaluation result does not meet the preset conditions, the hyperparameters of the first model are adjusted before entering the next training session. The first model includes a model based on an improved VQSL framework, which includes a first module that performs parallel operation of parameter-sharing local quantum circuits on a subset of qubits with the same width as the local quantum circuits.
[0009] After training, a second model is obtained; based on the second model, transformer faults are classified to complete transformer fault diagnosis.
[0010] Preferably, before iteratively training the first model based on the training set, the following is included:
[0011] Based on the improved VQSL framework, the fault diagnosis model is constructed by combining the third and fourth modules to obtain the first model; the third module is used for data preprocessing and encoding; and the fourth module is used for outputting the model classification results.
[0012] The first module is used to extract shadow features from the data; the improved VQSL framework also includes a second module; the second module includes a fully connected neural network for post-processing shadow features based on the output of the first module;
[0013] The modules are connected in the following order: Module 3, Module 1, Module 2, and Module 4.
[0014] Preferably, data preprocessing based on the first dataset includes:
[0015] The third module performs amplitude encoding on the historical dissolved gas data and thermal encoding on the preset transformer status information; after data preprocessing, a second dataset is obtained as the input to the first module.
[0016] Preferably, amplitude encoding of dissolved gas historical data includes:
[0017] Assuming there are m types of dissolved gases in transformer oil, then the m-dimensional vector composed of historical data on the concentration of dissolved gases in the transformer oil at the i-th time point is: Amplitude encoding includes the following steps:
[0018] A1, x i Normalization can be expressed by the following relation:
[0019]
[0020] A2. For an m-dimensional vector composed of m gas concentrations, zero-padding is applied to the end of the vector to match the dimension, encoded as 2. N From a vector of dimension ρ, we obtain the pure quantum state |ρ| of N qubits. in (i) This can be represented by the following relation:
[0021]
[0022] Where {|l>} is a set of computational basis for Hilbert space, and N represents the number of encoded qubits. Represents the vector x after expanding its dimensions. i The l-th element.
[0023] Preferably, the method of performing one-hot encoding on the preset transformer status information includes:
[0024] The preset transformer status information includes the transformer's operating status label. If there are n types of operating status label categories for the transformer, the label data at the i-th time point obtained through one-hot encoding can be represented by the following formula:
[0025]
[0026] Where T represents the transpose of a vector.
[0027] Preferably, training the first model a predetermined number of times based on the training set includes:
[0028] After setting the output dimension of the fully connected layer of the first model to correspond to the transformer's operating state label category n, the training set is input into the first model to start training, and the first model outputs a probability distribution.
[0029] Using multi-class cross-entropy loss function As the loss function of the first model, based on the probability distribution and tag data The loss function can be expressed by the following relationship:
[0030]
[0031] Where, N train ρ represents the number of samples in the training set, n represents the number of categories of transformer operating states, and ρ represents the number of samples in the training set. in (i)This represents the training set samples containing quantum states, θ represents the parameters in the first module, and ω and b represent the weights and biases in the second module, respectively; y k (i) and They represent y respectively (i) and The k-th element in;
[0032] Backpropagation based on stochastic gradient descent is used as the optimization algorithm for the first model. A preset number of iterations is set, and through iterative training, the parameters θ in the first module and the weights and biases {ω, b} in the second module are updated. The multi-class cross-entropy loss function is used. Upon convergence, one iteration of training is completed;
[0033] Restart a new iteration of training based on the training set until the preset number of training iterations is reached, then stop training.
[0034] Preferably, performance evaluation of the first model based on the test set includes:
[0035] The first model is used to classify the sample data in the test set sequentially;
[0036] The classification performance of the first model is evaluated using macro-average performance metrics, including accuracy, precision, recall, and F1 score, which can be expressed by the following formula:
[0037]
[0038] Wherein, TP represents the number of positive class samples correctly identified by the model, TN represents the number of other negative class samples correctly identified by the model, FP represents the number of other negative class samples that the model misclassifies as positive class samples, and FN represents the number of positive class samples that the model misclassifies as other negative class samples.
[0039] Preferably, adjusting the hyperparameters of the first model before proceeding to the next training iteration includes:
[0040] The learning rate, batch size, number of qubits, and width and depth of the local quantum circuits of the first model are adjusted, and the first model is retrained.
[0041] Preferred options also include:
[0042] By adjusting the width and depth of the local quantum circuits in the first module, the performance characteristics of the first module in extracting gas concentration can be altered, including:
[0043] By increasing the width and depth of the local quantum circuits in the first module, the performance of the first module in extracting gas concentration characteristics is enhanced.
[0044] The present invention has the following beneficial effects:
[0045] This invention presents a power transformer fault diagnosis method based on improved variational quantum shadow learning. By preprocessing historical data on dissolved gas concentrations in transformer oil, this data can be input into a model based on the improved VQSL (Variational Quantum Shadow Learning) framework, providing training data for the model. The model based on the improved VQSL framework employs a parallel processing strategy, simultaneously applying parameter-sharing shadow circuits to a subset of qubits with the same width as the shadow circuits, rather than being limited to extracting feature information from adjacent local qubits. This allows the model to more comprehensively extract the complex nonlinear relationship between oil chromatography data and transformer operating conditions, improving model performance and enabling the model to meet the accuracy requirements for fault diagnosis. By leveraging the quantum parallelism and entanglement of quantum computing, quantum neural networks can process large-scale oil chromatographic data in a shorter time, thereby accelerating the execution of complex computational tasks. Applying a special quantum feature mapping technique from quantum algorithms, the high-dimensional features of oil chromatographic data can be mapped to a high-dimensional Hilbert space, helping the model to more effectively handle complex relationships in the high-dimensional feature space without dimensionality reduction. Compared to existing machine learning methods used in transformer fault diagnosis, the significant advantages of quantum computing enable the method of this invention to handle complex computational tasks. The method of this invention can accurately detect and identify potential faults in power transformers, thereby ensuring the continuous and reliable operation of power transformers and improving the reliability of the power system.
[0046] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0047] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0048] Figure 1 This is a flowchart illustrating a preferred embodiment of the present invention.
[0049] Figure 2 The graph shows the model performance results under three encoding methods in the preferred embodiment of the present invention.
[0050] Figure 3 This is a schematic diagram of the normalized confusion matrix of the model of the present invention on the test set according to a preferred embodiment of the present invention. Detailed Implementation
[0051] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.
[0052] See Figure 1 In a preferred embodiment of the present invention, a power transformer fault diagnosis method based on improved variable quantum shadow learning is provided, comprising:
[0053] Step 1: Collect historical data on dissolved gases in transformer oil and combine it with preset transformer status information to form the first dataset; preprocess the first dataset to obtain the second dataset; divide the second dataset according to a preset ratio to obtain the training set and the test set;
[0054] In step one, the data preprocessing based on the first dataset specifically includes:
[0055] The third module performs amplitude encoding on the historical dissolved gas data and thermal encoding on the preset transformer status information; after data preprocessing, a second dataset is obtained as the input to the first module.
[0056] (1) Amplitude encoding of dissolved gas historical data includes:
[0057] Assuming there are m types of dissolved gases in transformer oil, then the m-dimensional vector composed of historical data on the concentration of dissolved gases in the transformer oil at the i-th time point is: Amplitude encoding includes the following steps:
[0058] A1, x i Normalization can be expressed by the following relation:
[0059]
[0060] In a preferred embodiment of the present invention, the main purpose of normalization is to adjust the length of the vector to 1 while preserving the direction of the vector. Normalization can eliminate the influence between different scales and avoid certain features having a disproportionate impact on the result due to their large values.
[0061] A2. For an m-dimensional vector composed of m gas concentrations, zero-padding is applied to the end of the vector to match the dimension, encoded as 2. N From a vector of dimension ρ, we obtain the pure quantum state |ρ| of N qubits. in (i) This can be represented by the following relation:
[0062]
[0063] Where {|l>} is a set of computational basis for the Hilbert space, and N represents the number of encoded qubits. Represents the vector x after expanding its dimensions. i The l-th element.
[0064] For example, to encode a 5-dimensional vector [1,0,0,1,0] into a 3-qubit quantum state, it needs to be normalized first. Then add zeros at the end to make it By matching the 8-dimensional dimensions, the final quantum state is obtained as follows: The dimension of the encoded quantum state depends only on the number of qubits used, and is independent of the number of characteristic gases. However, when the number of characteristic gases changes, the encoding process changes, leading to a change in the final quantum state.
[0065] If the m-dimensional vector composed of the concentrations of m gases is exactly 2 N If the encoding is done with N qubits, no dimension padding is needed. If the encoding is done with more than N qubits, such as N+1 qubits, then dimension padding is needed up to 2. N+1 .
[0066] After amplitude encoding, the characteristic gas concentration is stored as quantum bit information in the form of amplitude.
[0067] In a preferred embodiment of the present invention, by preprocessing the historical data of dissolved gas concentration in transformer oil, the historical data of dissolved gas concentration in transformer oil can be input into a model based on the improved VQSL framework, providing a data source for model training and testing.
[0068] (2) The preset transformer status information is encoded using a single-hot encoding method, including:
[0069] The preset transformer status information includes the transformer's operating status label. If there are n types of operating status label categories for the transformer, the label data at the i-th time point obtained through one-hot encoding can be represented by the following formula:
[0070]
[0071] Where T represents the transpose of a vector.
[0072] In a preferred embodiment of the present invention, the operating status labels of the transformer include: normal (NS), low temperature overheating fault (LT), medium temperature overheating fault (MT), high temperature overheating fault (HT), low energy discharge fault (LD), partial discharge fault (PD), and high energy discharge fault (HD).
[0073] The unique heat vectors corresponding to the transformer's operating status are shown in Table 1.
[0074] Table 1 shows the transformer operating status corresponding to the label values.
[0075]
[0076] In a preferred embodiment of the present invention, the transformer's operating status label is transformed into a one-hot vector by performing one-hot encoding, thus providing a data source for the calculation of the loss function.
[0077] Before step two iteratively trains the first model based on the training set, the following is also included:
[0078] Based on the improved VQSL framework, the fault diagnosis model is constructed by combining the third and fourth modules to obtain the first model; the third module is used for data preprocessing and encoding; and the fourth module is used for outputting the model classification results.
[0079] The improved VQSL framework consists of a first module and a second module;
[0080] The first module includes a parameter-shared variational local quantum circuit for extracting data shadow features, and the parameter-shared variational local quantum circuit acts in parallel on a subset of qubits with the same width as the local quantum circuit.
[0081] The variational local quantum circuit topology in the first module mainly consists of single-qubit rotation gates Ry and Rz, as well as controlled quantum gates. The entire circuit is repeated multiple times to achieve more complex quantum state transformations. This structural design aims to optimize the quantum state by adjusting the rotation angle to achieve fault diagnosis tasks.
[0082] The second module includes a fully connected neural network for post-processing shadow features based on the output of the first module.
[0083] The modules are connected in the following order: Module 3, Module 1, Module 2, and Module 4.
[0084] Step 2: Train the first model a preset number of times based on the training set, and evaluate the performance of the first model based on the test set after each training session; if the performance evaluation result does not meet the preset conditions, adjust the hyperparameters of the first model before proceeding to the next training session; the first model includes a model based on the improved VQSL framework, which includes a first module that performs parallel operation of parameter-sharing local quantum circuits on a subset of qubits with the same width as the local quantum circuits;
[0085] In a preferred embodiment of the present invention, parameter sharing in quantum computing and quantum machine learning typically refers to using the same set of parameters across different parts of a quantum circuit. In the improved VQSL framework, the local quantum circuit with parameter sharing plays the following roles:
[0086] (1) Reduce the number of parameters: By sharing parameters, the number of parameters that need to be optimized can be significantly reduced, thereby simplifying the optimization process, which is especially important when optimizing complex quantum circuits.
[0087] (2) Improve the generalization ability of the model: Sharing parameters can enable the model to learn more general features. This is also widely used in convolutional neural networks (CNNs) in classic machine learning. This approach can help the model perform more robustly when dealing with different input data.
[0088] (3) Accelerate the training process: After reducing the number and complexity of parameters, the optimization algorithm may converge faster, thereby accelerating the training process.
[0089] (4) Reduce resource consumption: Optimizing fewer parameters requires less computational resources and time, which is especially important when quantum computing resources are limited.
[0090] In step two, training the first model a predetermined number of times based on the training set specifically includes:
[0091] After setting the output dimension of the fully connected layer of the first model to correspond to the transformer's operating state label category n, the training set is input into the first model to start training, and the first model outputs a probability distribution.
[0092] Using multi-class cross-entropy loss function As the loss function of the first model, based on the probability distribution and tag data The loss function can be expressed by the following relationship:
[0093]
[0094] Where, N train ρ represents the number of samples in the training set, n represents the number of categories of transformer operating states, and ρ represents the number of samples in the training set. in (i) This represents the training set samples containing quantum states, θ represents the parameters in the first module, and ω and b represent the weights and biases in the second module, respectively; y k (i) and They represent y respectively (i) and The k-th element in;
[0095] Backpropagation based on stochastic gradient descent is used as the optimization algorithm for the first model. A preset number of iterations is set, and through iterative training, the parameters θ in the first module and the weights and biases {ω, b} in the second module are updated. The multi-class cross-entropy loss function is used. Upon convergence, one iteration of training is completed;
[0096] Restart a new iteration of training based on the training set until the preset number of training iterations is reached, then stop training.
[0097] In step two, the performance evaluation of the first model based on the test set includes:
[0098] The first model is used to classify the sample data in the test set sequentially;
[0099] The classification performance of the first model is evaluated using macro-average performance metrics, including accuracy, precision, recall, and F1 score, which can be expressed by the following formula:
[0100]
[0101] Wherein, TP represents the number of positive class samples correctly identified by the model, TN represents the number of other negative class samples correctly identified by the model, FP represents the number of other negative class samples that the model misclassifies as positive class samples, and FN represents the number of positive class samples that the model misclassifies as other negative class samples.
[0102] In a preferred embodiment of the present invention, the specific TP, TN, FP, and FN are obtained by calculating the confusion matrix based on the prediction results and the real labels.
[0103] In a preferred embodiment of this invention, precision is the ratio of correctly classified samples to the total number of samples. Accuracy is the ratio of correctly classified positive samples to all samples predicted as positive; it is suitable for class-balanced datasets and provides an overall overview of classification performance. Recall is the ratio of correctly classified positive samples to all true positive samples. The F1 score is the harmonic mean of precision and recall. In cases of class imbalance, the F1 score is a better overall evaluation metric. In specific situations, either precision or recall can be emphasized to ensure the model meets practical needs. High precision indicates good overall model performance, but class balance in the dataset must be ensured. High precision indicates accurate model predictions and a low false positive rate, suitable for scenarios with high false positive costs. High recall indicates strong ability to identify positive classes and a low false negative rate, suitable for scenarios with high false negative costs. A high F1 score indicates good overall performance and is suitable for scenarios requiring a balance between precision and recall.
[0104] In step two, adjusting the hyperparameters of the first model before proceeding to the next training iteration includes:
[0105] The learning rate, batch size, number of qubits, and width and depth of the local quantum circuits of the first model are adjusted, and the first model is retrained.
[0106] Step 3: After training is completed, a second model is obtained; based on the second model, the transformer is classified for faults, and the fault diagnosis of the transformer is completed.
[0107] In a preferred embodiment of the present invention, by adjusting the width and depth of the local quantum circuit in the first module, the performance characteristics of the gas concentration extraction feature of the first module can be changed, including:
[0108] By increasing the width and depth of the local quantum circuits in the first module, the performance of the first module in extracting gas concentration characteristics is enhanced.
[0109] Verification section:
[0110] Experiments were conducted using 1080 sets of dissolved gas concentration data (hydrogen, methane, ethane, ethylene, and acetylene) collected from five transformer oils. The DGA data were quantum encoded using amplitude encoding, angle encoding, and IQP encoding, respectively. Models were established and trained according to the method of this invention, and the model performance under the three encoding methods was compared in detail. Experimental results are available in [reference needed]. Figure 2 .
[0111] according to Figure 2 It can be seen that in the improved VQSL model of this invention, the amplitude encoding strategy exhibits superior performance, followed by angle encoding, while IQP encoding shows a lower performance level. This indicates that in quantum neural networks, the encoding method determines the quality of mapping classical data to quantum states and the required quantum computing resources, thus directly affecting the model performance.
[0112] To verify the accuracy and reliability of the method of this invention, the performance of VQSL, QNN, GNN, CNN, MLP, SVM, and the model of this invention based on the improved VQSL were compared. Among them, GNN, CNN, MLP, and SVM are models based on classical computation, while VQSL, QNN, and the model of this invention based on the improved VQSL are quantum-classical hybrid models. The model performance comparison results are shown in Table 2.
[0113] Table 2. Model Performance Comparison Results
[0114]
[0115]
[0116] As shown in Table 2, the model based on improved VQSL exhibits superior performance without considering noise interference. Compared to the unimproved VQSL model, the improved VQSL model shows a significant performance improvement, with accuracy increasing by 19.4%, precision by 22%, recall by 23.6%, and F1-score by 23%. This indicates that the improved VQSL model, by increasing the number of shadow features, can more completely extract the feature information between qubits, thus significantly improving model performance. The improved VQSL model also slightly outperforms QNN. This suggests that compared to extracting features from the entire space, extracting multiple shadow features from the subspace through convolution can more efficiently extract gas concentration features, thereby improving model performance. Compared to the GNN model, which ranks second in overall performance, the improved VQSL model shows improvements in accuracy, precision, recall, and F1-score, with increments of 4.6%, 4%, 6.3%, and 5.3%, respectively. Even under certain noise interference, the improved VQSL model still performs best. This demonstrates the potential advantages of quantum machine learning algorithms, which use quantum feature mapping to map gas concentration data into high-dimensional quantum states, thereby better extracting data features.
[0117] The normalized confusion matrix of the model based on the improved VQSL in this invention on the test set is shown in [reference]. Figure 3 ,Depend on Figure 3 It can be seen that the model based on the improved VQSL of this invention achieves high accuracy in classifying each transformer operating type. The diagnostic accuracy of the model based on the improved VQSL of this invention for NS, LT, MT, HT, PD, LD, and HD conditions is 96%, 94%, 95%, 91%, 96%, 93%, and 96%, respectively. This demonstrates that in fault diagnosis tasks, the model based on the improved VQSL of this invention can accurately distinguish different fault types.
[0118] In summary, the power transformer fault diagnosis method based on improved variable quantum shadowing learning in the preferred embodiment of this invention preprocesses historical data on dissolved gas concentrations in transformer oil, enabling this data to be input into a model based on the improved VQSL framework, thus providing training data for the model. The model based on the improved VQSL framework of this invention employs a parallel processing strategy, simultaneously applying parameter-sharing shadow circuits to a subset of qubits with the same width as the shadow circuits, rather than being limited to extracting feature information from adjacent local qubits. This allows the model in this invention to more comprehensively extract the complex nonlinear relationship between oil chromatography data and transformer operating status, improving model performance and enabling the model to meet the accuracy requirements for fault diagnosis. By leveraging the quantum parallelism and entanglement of quantum computing, quantum neural networks can process large-scale oil chromatographic data in a shorter time, thereby accelerating the execution of complex computational tasks. Applying a special quantum feature mapping technique from quantum algorithms, the high-dimensional features of oil chromatographic data can be mapped to a high-dimensional Hilbert space, helping the model to more effectively handle complex relationships in the high-dimensional feature space without dimensionality reduction. Compared to existing machine learning methods used in transformer fault diagnosis, the significant advantages of quantum computing enable the method of this invention to handle complex computational tasks. The method of this invention can accurately detect and identify potential faults in power transformers, thereby ensuring the continuous and reliable operation of power transformers and improving the reliability of the power system.
[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A power transformer fault diagnosis method based on improved variable quantum shadow learning, characterized in that, include: Historical data on dissolved gases in transformer oil are collected and combined with preset transformer status information to form the first dataset. The first dataset is preprocessed to obtain the second dataset; The second dataset is divided into a training set and a test set according to a preset ratio; The first model is trained a preset number of times based on the training set, and the performance of the first model is evaluated based on the test set after each training. When the performance evaluation results do not meet the preset conditions, the hyperparameters of the first model are adjusted before entering the next training; the first model includes a model based on the improved VQSL framework, which includes a first module that performs parallel operation of parameter-sharing local quantum circuits on a subset of qubits with the same width as the local quantum circuits. After training, a second model is obtained; based on the second model, transformer faults are classified to complete transformer fault diagnosis. Before iteratively training the first model based on the training set, the following is included: Based on the improved VQSL framework, the first model is obtained by constructing a fault diagnosis model by combining the third and fourth modules; the third module is used for data preprocessing and encoding; the fourth module is used for outputting the model classification results. The first module is used to extract shadow features from the data; the improved VQSL framework also includes a second module; the second module includes a fully connected neural network for post-processing shadow features based on the output of the first module; The modules are connected in the following order: Module 3, Module 1, Module 2, and Module 4. Preprocessing the first dataset includes: The third module performs amplitude encoding on the historical dissolved gas data and thermal encoding on the preset transformer status information; after data preprocessing, a second dataset is obtained as the input to the first module.
2. The power transformer fault diagnosis method based on improved variable quantum shadow learning according to claim 1, characterized in that, Amplitude encoding of the dissolved gas historical data includes: Assuming there are m types of dissolved gases in the transformer oil, then the first... An m-dimensional vector composed of historical data on dissolved gas concentrations in transformer oil at various time points is given by The amplitude encoding includes the following steps: A1, the above Normalization can be expressed by the following relation: ； A2. For an m-dimensional vector composed of m gas concentrations, zero-padding is applied to the end of the vector to match the dimension, encoded as 2. N From a dimensional vector, we obtain a pure quantum state of N qubits. It can be represented by the following relation: ； in, It is a set of computational bases in Hilbert space, where N represents the number of encoded qubits. Represents the vector after expanding the dimensions. The Each element.
3. The power transformer fault diagnosis method based on improved variable quantum shadow learning according to claim 2, characterized in that, The process of performing one-hot encoding on the preset transformer state information includes: The preset transformer status information includes the transformer's operating status label. If there are n types of operating status label categories for the transformer, then the nth type obtained through one-hot encoding... The label data at each time point can be represented by the following formula: ； Where T represents the transpose of a vector.
4. The power transformer fault diagnosis method based on improved variable quantum shadow learning according to claim 3, characterized in that, Training the first model a preset number of times based on the training set includes: After setting the output dimension of the fully connected layer of the first model to correspond to the transformer's operating state label category n, the training set is input into the first model to start training, and the first model outputs a probability distribution. ; Using multi-class cross-entropy loss function As the loss function of the first model, based on the probability distribution and the tag data The loss function can be expressed by the following relationship: ； in, This indicates the number of samples in the training set. The number of categories representing the transformer's operating status. This represents a training set sample containing quantum states. This represents the parameters in the first module. and These represent the weights and biases in the second module, respectively. and They represent and The k-th element in; Backpropagation based on stochastic gradient descent is used as the optimization algorithm for the first model. A preset number of iterations is set, and the parameters in the first module are updated through iterative training. and the weights and biases in the second module When the multi-class cross-entropy loss function Upon convergence, one iteration of training is completed; A new iteration of training will begin based on the training set until a preset number of training iterations are reached, at which point training will stop.
5. The power transformer fault diagnosis method based on improved variable quantum shadow learning according to claim 4, characterized in that, The performance evaluation of the first model based on the test set includes: The first model is used to classify the sample data in the test set sequentially; The classification performance of the first model is evaluated using macro-average performance metrics, including accuracy, precision, recall, and F1 score, which can be expressed by the following relationship: ； Wherein, TP represents the number of positive class samples correctly identified by the model, TN represents the number of other negative class samples correctly identified by the model, FP represents the number of other negative class samples that the model misclassifies as positive class samples, and FN represents the number of positive class samples that the model misclassifies as other negative class samples.
6. The power transformer fault diagnosis method based on improved variable quantum shadow learning according to claim 5, characterized in that, The process of adjusting the hyperparameters of the first model before proceeding to the next training iteration includes: The learning rate, batch size, number of qubits, and width and depth of the local quantum circuit of the first model are adjusted, and the first model is retrained.
7. The power transformer fault diagnosis method based on improved variable quantum shadowing learning according to any one of claims 1 to 6, characterized in that, Also includes: By adjusting the width and depth of the local quantum circuits in the first module, the performance characteristics of the gas concentration extraction feature of the first module can be altered, including: By increasing the width and depth of the local quantum circuits in the first module, the performance of the first module in extracting gas concentration characteristics is enhanced.