Transformer cooling system fault intelligent diagnosis and cause feature screening method
By employing an interpretable ensemble learning approach, utilizing convolutional autoencoders and the XGBoost-SHAP model, high-precision intelligent diagnosis and feature selection of transformer cooling system faults were achieved. This solves the problems of inaccurate diagnosis and insufficient interpretability in traditional methods, and provides physically traceable decision support.
Patent Information
- Application Number
- CN202511445259.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-11
- Publication Date
- 2026-03-03
AI Technical Summary
Traditional fault diagnosis methods for transformer cooling systems are susceptible to interference from external factors, have difficulty capturing effective characteristic signals, lack automated fault prediction mechanisms, cannot achieve accurate diagnosis and transparent decision-making, and lack interpretability.
We employ an interpretable ensemble learning approach, extracting deep features through a convolutional autoencoder (CAE), and combining XGBoost and SHAP models for fault diagnosis and feature parsing. This constructs a closed-loop self-evolutionary optimization system, enabling automated screening of fault features and visualization of decision-making criteria.
It achieves high-precision intelligent diagnosis of transformer cooling system faults, provides physically traceable decision support, improves the reliability of diagnosis and the visualization capability of features, and solves the problems of insufficient feature mining and lack of model interpretability in traditional methods.
Smart Images
Figure CN121598041A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of transformer equipment fault diagnosis technology, and relates to a method for intelligent diagnosis of transformer faults and screening of causal features, especially a method for intelligent diagnosis of transformer cooling system faults and screening of causal features based on interpretable ensemble learning. Background Technology
[0002] With the rapid development of ultra-high voltage power grids and the large-scale construction of smart substations, the operational reliability of power transformers, as core power transmission and transformation equipment, directly determines the safety of the power grid and the quality of power supply.
[0003] In this crucial process, the transformer's cooling system plays a vital role. By effectively dissipating heat, the cooling system maintains a stable internal temperature, directly slowing down the aging process of the insulation, suppressing mechanical deformation and stress in critical components, and ensuring long-term stability of electrical performance. Especially during long-term operation under complex conditions, the effectiveness of the cooling system has a decisive impact on the overall health of the transformer.
[0004] However, the state evolution of the cooling system itself (such as efficiency decline and component failure) is affected by the complex coupling of multiple physical fields such as electric field, thermal field, and mechanical field. It often exhibits characteristics such as strong nonlinearity, hidden state correlation, and uncertain fault evolution path. Subtle changes in the cooling system may indicate potential major risks to the transformer. Therefore, ensuring the efficient and reliable operation of the transformer cooling system is of paramount importance to ensuring the safe and stable operation of the transformer and even the entire power grid.
[0005] Traditional transformer cooling system fault diagnosis systems primarily rely on operational parameters (such as oil temperature and cooler inlet / outlet temperature difference), oil pump fluctuations, and insulating oil condition for analysis. These methods are susceptible to interference from external factors such as load fluctuations and ambient temperature changes. They lack the ability to deeply analyze the correlations of highly redundant and strongly correlated multidimensional monitoring data, making it difficult to capture effective characteristic signals. Furthermore, they lack automated fault prediction mechanisms, preventing maintenance personnel from quickly analyzing the contribution of key features and generating visual reports, severely limiting the reliability of diagnostic methods. These problems make existing diagnostic methods unable to meet the urgent needs of smart grids for accurate fault diagnosis, transparent decision-making, and feature visualization.
[0006] Therefore, there is an urgent need to construct an intelligent fault diagnosis and causal feature screening method for transformer cooling systems that combines high-precision diagnostic capabilities with inherent interpretability, so as to realize the automated screening of fault features and physical traceability of decision-making basis, and provide intelligent decision support for equipment condition-based maintenance.
[0007] A search revealed no publicly available literature of the same or similar prior art as this invention. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for intelligent fault diagnosis and causal feature screening of transformer cooling systems based on interpretable ensemble learning. This method can achieve in-depth screening of fault features and visual traceability of decision-making basis, providing intelligent decision support for equipment condition-based maintenance.
[0009] The present invention solves its practical problem by adopting the following technical solution: A method for intelligent fault diagnosis and feature screening in a transformer cooling system includes the following steps: Step 1: Collect operational data sets of oil temperature, medium, and fan of the transformer cooling system to construct an operational data system with time-series characteristics; Step 2: Perform data preprocessing on the transformer cooling system operation dataset obtained in Step 1, including missing value imputation, anomaly cleaning, and standard normalization. Step 3: Use a convolutional autoencoder (CAE) to automatically extract deep features of the transformer cooling system from the preprocessed transformer cooling system operation dataset in Step 2. Step 4: Based on the deep features of the transformer cooling system extracted in Step 3, the XGBoost-SHAP two-stage architecture is used to perform fault diagnosis and feature analysis of the transformer cooling system. Step 5: Based on the transformer fault diagnosis and feature analysis results from Step 4, perform model self-evolution optimization.
[0010] Moreover, the specific method of step 1 is as follows: The system collects operational data on three core parameters of the transformer cooling system: oil temperature, medium, and fan. It constructs an operational data system with time-series characteristics by collecting data once per minute. Furthermore, the specific method for step 2 is as follows: First, missing values in the time series are filled using linear interpolation. For any missing point... The interpolation is obtained by linear calculation of adjacent valid data points, as shown in equation (1):
[0011] In the formula, For a moment t The target interpolation, , These are the known observations at the time points adjacent to the missing values; Secondly, the Z-Score method is used to remove outliers. The standardized value of each feature vector is shown in equation (2):
[0012] In the formula, The degree of deviation in the sense of standard deviation, when When this happens, the data point is considered an anomaly and removed. For a single observation data point, This is the historical mean of this feature. The standard deviation of this feature; Finally, after cleaning, all features are subjected to min-max normalization to unify the data scale to the [0,1] interval, as shown in equation (3):
[0013] In the formula, These are the normalized eigenvalues. , These are the maximum and minimum values in the historical data, respectively.
[0014] Furthermore, the specific method for step 3 is as follows: The deep features of the transformer are automatically extracted from the running feature dataset after step 2 by using a convolutional autoencoder CAE. The convolutional autoencoder CAE is composed of stacked convolutional layers and pooling layers, and uses one-dimensional convolution operation. Its operation process is as shown in equation (4).
[0015] in, The output is the i-th convolutional kernel of the l-th layer; and Let be the weights and biases of the i-th convolutional kernel in the l-th layer; f(·) is the output of the j-th convolutional kernel in the (l+1)-th layer; f(·) is the activation function.
[0016] The pooling layer downsamples the output of the convolutional layer to reduce the feature size, and its operation formula is shown in equation (5).
[0017]
[0018] in, For pooling functions, such as max pooling, average pooling, etc. and These are the input and output values for the pooling layer.
[0019] The decoder is composed of stacked transposed convolutional layers, realizing the mapping from low-dimensional features to high-dimensional features. The calculation of transposed convolution is similar to that in equation (4), and the weights of its convolutional kernel are determined by the weights in equation (4). It is obtained by transposing.
[0020] Furthermore, the specific method for step 4 is as follows: First, a high-precision cooling system fault prediction model is constructed using the XGBoost model. The principle of the XGBoost model algorithm is to continuously iterate to train new trees and correct the errors of the previous tree. Assuming that k trees are constructed, the intelligent fault prediction result for the i-th sample is as shown in formula (6);
[0021] in, The space is the set of functions that contains all decision trees. It is the sum of the predictions from the first K-1 trees; This is the prediction result of the k-th decision tree, which is learned from... The residual with the true value This is the deep feature set extracted by CAE.
[0022] We need to define an objective function L(φ) to determine the direction of model training and optimization, as shown in equation (7).
[0023]
[0024] In the formula, These are the predicted values from the XGBoost model. The true value of the sample; For loss functions, such as cross-entropy; It represents the complexity of the k-th tree, and can be used to represent regularization terms.
[0025] Furthermore, Bayesian optimization is employed to adjust the XGBoost parameters. Bayesian optimization constructs a probabilistic surrogate model of parameters and performance based on Gaussian processes, actively exploring potential global optimum regions through a data acquisition function, significantly improving search efficiency. Assuming the model used is... According to the Bayesian parameter optimization method, we have equation (8).
[0026]
[0027] In the formula, For hyperparameters, and The training dataset is used; the fitness function is obtained from equation (9).
[0028]
[0029] In the formula, This represents the fitness function value; The predicted values obtained after training the model; N represents the number of samples in the validation dataset. This is the actual value.
[0030] Finally, the SHAP (Shapley Additive Explanations) interpretability method is applied to perform attribution analysis on the prediction results of the XGBoost model. The basic concept of SHAP interpretability analysis is a game theory method based on SHAP values. It shows how the output results are distributed among features, can be used to explain the feature importance of the best-performing prediction model, and can quantify the influence and direction of each transformer operating parameter on the occurrence of faults, as shown in Equation (10).
[0031]
[0032] in, For the j-th parameter eigenvalue, These are the sample predictions output by the XGBoost model. This represents the mean of the risk output values.
[0033] Furthermore, the specific method for step 5 is as follows: The XGBoost incremental learning interface dynamically updates the leaf node weights while preserving the original decision tree structure, fully retaining the physical mechanism knowledge of the multi-theory fusion model, and injecting knowledge of newly added fault samples. Combined with monthly feature reports to eliminate the last SHAP features, when there are contradictions in the direction of SHAP attribution and disintegration reports, deviations exceeding the threshold, or continuous missed reports, the CAE-XGBoost joint retraining process is triggered to achieve dynamic co-evolution of feature space and decision logic.
[0034] Advantages and beneficial effects of the present invention: 1. This invention proposes an intelligent fault diagnosis and feature selection method for transformer cooling systems based on interpretable ensemble learning. Addressing the problems of insufficient feature mining, lack of model interpretability, and inefficiency of manual screening inherent in transformer fault diagnosis, this invention uses a convolutional autoencoder (CAE) model to perform unsupervised learning on historical redundant fault signals of the transformer cooling system. Through the output of the encoder, deep-level features of cooling system faults are extracted. Then, by fusing the high-precision classification capability of XGBoost with the interpretability of the SHAP model, deep-level screening of key features of the transformer cooling system and intelligent prediction of fault types are achieved, while providing physically traceable decision-making basis. 2. This invention first collects core data such as cooled oil temperature data, medium data, and fan operation data. To improve the accuracy of data training, the collection frequency is reduced to one minute. Data outliers and abrupt noise values are cleaned and standardized to obtain a preprocessed data set. 3. To address the problem of insufficient deep feature mining in the transformer fault diagnosis process, multi-scale feature extraction is achieved within the data preprocessing set by leveraging the multi-layer convolution and pooling operations of the convolutional autoencoder (CAE). This captures abnormal patterns in different operating data during the operation of the transformer cooling system and naturally learns hierarchical feature representations from low to high levels, enabling efficient unsupervised feature learning.
[0035] 4. To address the shortcomings in automated fault prediction and feature visualization, this paper integrates XGBoost ensemble learning with the SHAP interpretable framework. This ensures diagnostic accuracy while enabling quantifiable analysis of feature contributions. Furthermore, it innovatively establishes a dynamic threshold feature selection feedback mechanism to generate an optimal feature subset with clear physical meaning. 5. This invention constructs a closed-loop self-evolutionary optimization system, which, based on a three-level linkage mechanism of incremental learning, dynamic feature reconstruction, and deviation triggering, realizes the continuous evolution of the diagnostic model and feature library, and completely solves the performance degradation problem of traditional transformer cooling systems caused by equipment aging and the emergence of new fault modes. Attached Figure Description
[0036] Figure 1 This is a flowchart of a transformer fault intelligent diagnosis and feature selection method based on interpretable ensemble learning according to the present invention. Figure 2 This is a framework diagram of a transformer fault intelligent diagnosis and feature selection method based on interpretable ensemble learning according to the present invention. Figure 3 This is a diagram of the CAE network structure of the present invention. Detailed Implementation
[0037] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings: A method for intelligent fault diagnosis and feature selection in transformer cooling systems based on interpretable ensemble learning includes the following steps: Step 1: Collect operational data sets of oil temperature, medium, and fan of the transformer cooling system to construct an operational data system with time-series characteristics; The specific method for step 1 is as follows: The system collects operational data on three core parameters of the transformer cooling system: oil temperature, medium, and fan. It constructs an operational data system with time-series characteristics by collecting data once per minute. In this embodiment, the working principle of step 1 is as follows: Step 1 is used to construct a monitoring data system for the operating status parameters of the transformer cooling system. This is achieved by deploying online monitoring devices, smart sensors, and data interfaces at the top of the transformer cooling system's oil tank, cooler outlet, air-cooling unit, and valves. These devices collect multi-source, heterogeneous operating data of the transformer cooling system, providing a comprehensive and timely input dataset for subsequent intelligent diagnostic model training and real-time analysis. Specifically, the collected data dimensions include oil temperature parameters, pressure parameters, fan operating parameters, oil pump operating parameters, and valve operating parameters. All of the above data is collected synchronously with timestamps, forming a structured input dataset with strict temporal correlation characteristics. The data acquisition device incorporates data validity verification and packet loss retransmission mechanisms to ensure the integrity and consistency of the input dataset during transmission and storage, laying a reliable foundation for subsequent processing.
[0038] Step 2: Perform data preprocessing on the transformer cooling system operation dataset obtained in Step 1, including missing value imputation, anomaly cleaning, and standard normalization. The specific method for step 2 is as follows: First, missing values in the time series are filled using linear interpolation. For any missing point... The interpolation is obtained by linear calculation of adjacent valid data points, as shown in equation (1):
[0039] In the formula, For a moment t The target interpolation, , These are the known observations at the time points adjacent to the missing values; Secondly, the Z-Score method is used to remove outliers. The standardized value of each feature vector is shown in equation (2):
[0040] In the formula, The degree of deviation in the sense of standard deviation, when When this happens, the data point is considered an anomaly and removed. For a single observation data point, This is the historical mean of this feature. The standard deviation of this feature; Finally, after cleaning, all features are subjected to min-max normalization to unify the data scale to the [0,1] interval, as shown in equation (3):
[0041] In the formula, These are the normalized eigenvalues. , These are the maximum and minimum values in the historical data, respectively.
[0042] In this embodiment, the working principle of step 2 is as follows: Step 2 is used to receive multi-source raw data uploaded by the data center, and undertakes the functions of data cleaning, unified encoding and unit unification. It is a key foundational step to achieve stable training and inference of the subsequent model.
[0043] Step 3: Use a convolutional autoencoder (CAE) to automatically extract deep features of the transformer cooling system from the preprocessed transformer cooling system operation dataset in Step 2. The specific method for step 3 is as follows: The deep features of the transformer are automatically extracted from the running feature dataset after step 2 by using a convolutional autoencoder CAE. The convolutional autoencoder CAE is composed of stacked convolutional layers and pooling layers, and uses one-dimensional convolution operation. Its operation process is as shown in equation (4).
[0044] in, The output is the i-th convolutional kernel of the l-th layer; and Let be the weights and biases of the i-th convolutional kernel in the l-th layer; f(·) represents the output of the j-th convolutional kernel in the (l+1)-th layer; f(·) represents the activation function such as Sigmoid, ReLU, etc.
[0045] The pooling layer downsamples the output of the convolutional layer to reduce the feature size, and its operation formula is shown in equation (5).
[0046]
[0047] in, For pooling functions, such as max pooling, average pooling, etc. and For the input and output of the pooling layer.
[0048] The decoder is composed of stacked transposed convolutional layers, realizing the mapping from low-dimensional features to high-dimensional features. The calculation of transposed convolution is similar to that in equation (4), and the weights of its convolutional kernel are determined by the weights in equation (4). It is obtained by transposing.
[0049] In this embodiment, the working principle of step 3 is as follows: Step 3, to overcome the challenges of high data dimensionality, complex structure, and noise interference in transformer cooling systems, employs a convolutional autoencoder (CAE) (such as...). Figure 3As shown, this method automates the extraction of deep features from equipment. This step uses preprocessed, standardized operating data of a multi-source cooling system as input. Through multi-layer convolution and pooling operations at the encoder level, the high-dimensional input signal is compressed layer by layer into a low-dimensional latent feature space containing key information about the equipment's state. The local receptive field of the convolution kernel makes it highly sensitive to minute local anomalies (such as sudden increases in vibration of a single oil pump), while the pooling layer effectively preserves key anomalous signals and suppresses irrelevant noise, significantly improving the feature representation capability. At the decoder level, through transposed convolutional layers and upsampling layers, the feature input signal is reconstructed from the extracted low-dimensional features with the goal of minimizing reconstruction error.
[0050] The core of this process lies in mapping the high-dimensional input to a low-dimensional feature space through convolution operations, and then using a decoder to reconstruct the signal. The low-dimensional feature encoding output by the encoder not only preserves the key information related to the operation of the cooling system in the original signal to the greatest extent, but its compact and abstract characteristics are also more conducive to the subsequent classifier to improve the accuracy and efficiency of fault identification, thereby completing the automated extraction of deep features of the transformer cooling system.
[0051] Step 4: Based on the deep features of the transformer cooling system extracted in Step 3, the XGBoost-SHAP two-stage architecture is used to perform fault diagnosis and feature analysis of the transformer cooling system. The specific method for step 4 is as follows: The XGBoost-SHAP interpretive analysis framework was used to quantify the feature contribution and perform fault attribution analysis on the deep features of the transformer cooling system extracted in step 3, revealing the correlation mechanism between key feature factors and fault modes.
[0052] First, an XGBoost model is used to build a high-precision cooling system fault prediction model. The principle of the XGBoost model algorithm is to use continuous iteration to train new trees, correct the error of the previous tree, and improve the accuracy of the model.
[0053] Assuming that k trees are constructed, the intelligent fault prediction result for the i-th sample is as shown in formula (6).
[0054]
[0055] in, The space is the set of functions that contains all decision trees. It is the sum of the predictions from the first K-1 trees; This is the prediction result of the k-th decision tree, which is learned from... The residual with the true value This is the deep feature set extracted by CAE.
[0056] We need to define an objective function L(φ) to determine the direction of model training and optimization, as shown in equation (7).
[0057]
[0058] In the formula, These are the predicted values from the XGBoost model. The true value of the sample; For loss functions, such as cross-entropy; It represents the complexity of the k-th tree, and can be used to represent regularization terms.
[0059] Furthermore, Bayesian optimization is employed to adjust the XGBoost parameters. Bayesian optimization constructs a probabilistic surrogate model of parameters and performance based on Gaussian processes, actively exploring potential global optimum regions through a data acquisition function, significantly improving search efficiency. Assuming the model used is... According to the Bayesian parameter optimization method, we have equation (8).
[0060]
[0061] In the formula, For hyperparameters, and The training dataset is used; the fitness function is obtained from equation (9).
[0062]
[0063] In the formula, This represents the fitness function value; The predicted values obtained after training the model; N represents the number of samples in the validation dataset. This is the actual value.
[0064] Finally, the SHAP (Shapley Additive Explanations) interpretability method is applied to perform attribution analysis on the prediction results of the XGBoost model. The basic concept of SHAP interpretability analysis is a game theory method based on SHAP values. It shows how the output results are distributed among features, can be used to explain the feature importance of the best-performing prediction model, and can quantify the influence and direction of each transformer operating parameter on the occurrence of faults, as shown in Equation (10).
[0065]
[0066] in, For the j-th parameter eigenvalue, These are the sample predictions output by the XGBoost model. This represents the mean of the risk output values.
[0067] The working principle of step 4 is as follows: Integrating multi-source evidence such as historical maintenance records, maintenance logs, and equipment decommissioning data, the deep feature input dataset extracted in step 3 is labeled and classified into four categories: electrical faults, mechanical faults, operational condition faults, and normal operation. An XGBoost-SHAP two-stage architecture is employed to achieve intelligent fault diagnosis and feature analysis of the transformer cooling system. Using XGBoost as the base classifier, a boosting ensemble strategy is used to learn complex fault modes and output fault types. Bayesian optimization is then applied to optimize the model's hyperparameters. Based on fault prediction, SHAP values are calculated to quantify the contribution of each feature to the diagnostic results, outputting key results such as feature importance maps, feature dependency maps, and feature interaction force maps. Historical attribution maps are output for specific single samples, obtaining fault diagnosis and feature analysis results for the transformer cooling system. This provides crucial information for transformer cooling system status assessment, risk warning, and the development of targeted maintenance strategies.
[0068] Step 5: Based on the transformer fault diagnosis and feature analysis results from Step 4, perform model self-evolution optimization.
[0069] The specific method for step 5 is as follows: The XGBoost incremental learning interface dynamically updates the leaf node weights while preserving the original decision tree structure, fully retaining the physical mechanism knowledge of the multi-theory fusion model, and injecting knowledge of newly added fault samples. Combined with monthly feature reports to eliminate the last SHAP features, when there are contradictions in the direction of SHAP attribution and disintegration reports, deviations exceeding the threshold, or continuous missed reports, the CAE-XGBoost joint retraining process is triggered to achieve dynamic co-evolution of feature space and decision logic.
[0070] The working principle of step 5 is as follows: When a new fault sample is added, the model weights are updated using the incremental learning interface of XGBoost. The input feature set is periodically reconstructed based on the feature screening report, removing fault features and incorporating newly derived features. When the SHAP interpretation result deviates significantly from the system's periodic report, the CAE+XGBoost model retraining process is triggered. To achieve continuous intelligent evolution of the system, step 5 establishes a three-level optimization mechanism including online incremental learning, dynamic feature reconstruction, and deviation-triggered retraining. When new fault samples are input, the most recent 100,000 samples are retained through the XGBoost online learning interface, covering a one-year operating cycle, and incremental weight updates are implemented. Adaptive learning rate is used to iteratively adjust model parameters to achieve a balance between knowledge evolution and suppressing catastrophic forgetting. Simultaneously, monthly feature space optimization is performed based on the feature screening report generated in step 4. Features ranking in the bottom 5% of SHAP values for three consecutive periods are closed from the collection channel. At the same time, derived features driven by fault mechanisms are included, and feature library version changes are recorded. When there is a significant deviation between the SHAP interpretation results and the equipment dismantling report, i.e., contradiction in the contribution direction of key features, deviation in single contribution, or five consecutive alarms, the full model retraining process is triggered: the original dataset is rolled back, the CAE feature extraction network structure and XGBoost classifier parameters are updated, the model is updated after verification, and the model self-evolution optimization is completed.
[0071] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.
Claims
1. A method for intelligent fault diagnosis and feature screening in a transformer cooling system, characterized in that: Includes the following steps: Step 1: Collect operational data sets of oil temperature, medium, and fan of the transformer cooling system to construct an operational data system with time-series characteristics; Step 2: Perform data preprocessing on the transformer cooling system operation dataset obtained in Step 1, including missing value imputation, anomaly cleaning, and standard normalization. Step 3: Use a convolutional autoencoder (CAE) to automatically extract deep features of the transformer cooling system from the preprocessed transformer cooling system operation dataset in Step 2. Step 4: Based on the deep features of the transformer cooling system extracted in Step 3, the XGBoost-SHAP two-stage architecture is used to perform fault diagnosis and feature analysis of the transformer cooling system. Step 5: Based on the transformer fault diagnosis and feature analysis results from Step 4, perform model self-evolution optimization.
2. The intelligent fault diagnosis and feature screening method for a transformer cooling system according to claim 1, characterized in that: The specific method for step 1 is as follows: The system collects operational data on three core parameters of the transformer cooling system: oil temperature, medium, and fan. It constructs an operational data system with time-series characteristics by collecting data once per minute.
3. The intelligent fault diagnosis and feature screening method for a transformer cooling system according to claim 1, characterized in that: The specific method for step 2 is as follows: First, missing values in the time series are filled using linear interpolation; for any missing point... The interpolation is obtained by linear calculation of adjacent valid data points, as shown in equation (1): ； In the formula, For a moment t The target interpolation, , These are the known observations at the time points adjacent to the missing values; Secondly, the Z-Score method is used to remove outliers; the standardized value of each feature vector is shown in equation (2): ； In the formula, The degree of deviation in the sense of standard deviation, when When this happens, the data point is considered an anomaly and removed. For a single observation data point, This is the historical mean of this feature. The standard deviation of this feature; Finally, after cleaning, all features are subjected to min-max normalization to unify the data scale to the [0,1] interval, as shown in equation (3): ； In the formula, These are the normalized eigenvalues. , These are the maximum and minimum values in the historical data, respectively.
4. The intelligent fault diagnosis and feature screening method for a transformer cooling system according to claim 1, characterized in that: The specific method for step 3 is as follows: The deep features of the transformer are automatically extracted from the running feature dataset after step 2 by using a convolutional autoencoder CAE. The convolutional autoencoder CAE is composed of stacked convolutional layers and pooling layers, and uses one-dimensional convolution operation. Its operation process is as shown in equation (4). ； in, The output is the i-th convolutional kernel of the l-th layer; and Let be the weights and biases of the i-th convolutional kernel in the l-th layer; f(·) represents the output of the j-th convolutional kernel in the (l+1)-th layer; f(·) represents the activation function such as Sigmoid, ReLU, etc. The pooling layer downsamples the output of the convolutional layer to reduce the feature size, and its operation formula is shown in equation (5). ； in, For pooling functions, such as max pooling, average pooling, etc. and For the input and output of the pooling layer; The decoder is composed of stacked transposed convolutional layers, realizing the mapping from low-dimensional features to high-dimensional features. The calculation of transposed convolution is similar to that in equation (4), and the weights of its convolutional kernel are determined by the weights in equation (4). It is obtained by transposing.
5. The intelligent fault diagnosis and feature screening method for a transformer cooling system according to claim 1, characterized in that: The specific method for step 4 is as follows: First, a high-precision cooling system fault prediction model is constructed using the XGBoost model. The principle of the XGBoost model algorithm is to continuously iterate to train new trees and correct the errors of the previous tree. Assuming that k trees are constructed, the intelligent fault prediction result for the i-th sample is as shown in formula (6); ； in, The space is the set of functions that contains all decision trees. It is the sum of the predictions from the first K-1 trees; This is the prediction result of the k-th decision tree, which is learned from... The residual with the true value This is a deep feature set extracted via CAE; We need to define an objective function L(φ) to determine the direction of model training and optimization, as shown in equation (7); ； In the formula, These are the predicted values from the XGBoost model. The true value of the sample; For loss functions, such as cross-entropy; It represents the complexity of the k-th tree, and can be used to represent regularization terms; Furthermore, Bayesian optimization is employed to adjust the XGBoost parameters. Bayesian optimization constructs a probabilistic surrogate model of parameters and performance based on Gaussian processes, actively exploring potential global optimum regions through a data acquisition function, significantly improving search efficiency. Assuming the model used is... According to the Bayesian parameter optimization method, we have equation (8); ； in For hyperparameters, and For the training dataset; the fitness function is obtained from equation (9); ； in The value predicted by the trained model; Finally, the SHAP (Shapley Additive Explanations) interpretability method is applied to perform attribution analysis on the prediction results of the XGBoost model. The basic concept of SHAP interpretability analysis is a game theory method based on SHAP values. It shows how the output results are distributed among features, which can be used to explain the feature importance of the best-performing prediction model and can quantify the influence and direction of each transformer operating parameter on the occurrence of faults, as shown in Equation (10). ； in, For the j-th parameter eigenvalue, These are the sample predictions output by the XGBoost model. This represents the mean of the risk output values.
6. The intelligent fault diagnosis and feature screening method for a transformer cooling system according to claim 1, characterized in that: The specific method for step 5 is as follows: The XGBoost incremental learning interface dynamically updates the leaf node weights while preserving the original decision tree structure, fully retaining the physical mechanism knowledge of the multi-theory fusion model, and injecting knowledge of newly added fault samples. Combined with monthly feature reports to eliminate the last SHAP features, when there are contradictions in the direction of SHAP attribution and disintegration reports, deviations exceeding the threshold, or continuous missed reports, the CAE-XGBoost joint retraining process is triggered to achieve dynamic co-evolution of feature space and decision logic.