Transformer adaptive fault diagnosis method and system based on AmRMR
By combining AmRMR and DSD-DQN, the problems of insufficient feature selection and difference in the cost of misjudgment in transformer fault diagnosis are solved, and efficient identification and safe diagnosis of serious faults are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-01-08
- Publication Date
- 2026-05-05
AI Technical Summary
Existing transformer fault diagnosis methods lack information in feature selection, making it difficult to fully characterize complex fault modes. They also ignore the differences in the cost of misjudgment and lack consideration for engineering risks, resulting in insufficient ability to identify serious faults.
An adaptive fault diagnosis method based on AmRMR is adopted. By constructing a candidate ratio feature set, standardizing, discretizing and redundancy compression are performed to screen out key ratio features. The method is combined with the DSD-DQN fault diagnosis model for deep learning and introduces priority experience playback and cost-sensitive reward function optimization strategies to improve the identification ability of minority class samples.
It significantly improves the ability to identify serious faults, balances engineering safety and diagnostic accuracy, reduces the risk of misjudgment, and improves the robustness and precision of diagnosis.
Smart Images
Figure CN121980414A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent operation and maintenance technology of power equipment, and relates to a transformer adaptive fault diagnosis method and system based on AmRMR, which is particularly suitable for intelligent identification of multiple types of transformer faults based on oil gas detection data. Background Technology
[0002] Transformers are critical equipment in power systems, and their insulation condition and operating conditions directly affect the safety and reliability of the power grid. Dissolved gas analysis (DGA) in oil has been widely used for transformer condition monitoring and fault diagnosis due to its convenient sampling, high sensitivity, and ability to detect internal defects such as partial discharge and overheating at an early stage. By detecting the content and ratio of characteristic gases such as hydrogen, methane, ethane, ethylene, and acetylene in the oil, the types of faults that may exist inside the transformer can be inferred. Existing DGA fault diagnosis methods mainly include empirical ratio methods, critical gas methods, three-ratio methods, and rule-based methods based on IEC standards. These methods rely on expert experience and preset thresholds, and have limited adaptability to complex operating conditions and new equipment.
[0003] With the development of machine learning and deep learning, an increasing number of studies are modeling transformer DGA fault diagnosis as a multi-classification problem, employing supervised learning models such as support vector machines, neural networks, random forests, ensemble learning, and deep neural networks to improve the automation and accuracy of diagnosis. In these methods, input features are typically selected manually based on experience, such as directly using raw gas concentrations, typical ratios, or a few combinations of features, lacking a systematic evaluation of the correlation between features and fault types and the redundancy between features. Furthermore, these models often employ static classification frameworks, prioritizing overall accuracy or average loss as the main optimization objective, making it difficult to simultaneously reflect class imbalance and the differences in the cost of misclassifying different faults within the same framework. In actual operation, transformer DGA data often exhibits class imbalance: a large number of normal and minor fault samples, while severe faults and malignant defects are relatively scarce; and the risk and cost of misdiagnosing severe faults as normal or low-risk faults are significantly higher than those of general misclassifications. In traditional supervised learning models, although the above problems can be addressed to some extent by means of oversampling, undersampling, class weights, or cost-sensitive reward functions, it usually requires repeated trials to manually set multiple sets of weight parameters. Furthermore, class imbalance and misjudgment costs are easily mixed in the same weight, making it difficult to clearly characterize the relationship between data distribution characteristics and the degree of engineering hazards. Summary of the Invention
[0004] To address the shortcomings of existing transformer fault diagnosis methods, such as the lack of information on single DGA gas component characteristics, difficulty in fully characterizing complex fault modes, neglect of differences in misjudgment costs, and lack of consideration for engineering risks, this invention provides an adaptive transformer fault diagnosis method and system based on AmRMR. This method effectively avoids the highest engineering risk of misjudging faults as normal, significantly improves the ability to identify a minority of samples, such as severe faults, and achieves transformer fault diagnosis that balances engineering safety and diagnostic accuracy.
[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.
[0006] In a first aspect, the present invention proposes a transformer adaptive fault diagnosis method based on AmRMR, comprising the following steps:
[0007] Obtain dissolved gas detection data in transformer oil, obtain key gas concentration characteristics, and construct a candidate ratio feature set based on the gas detection data;
[0008] The key gas concentration features and candidate ratio feature sets are standardized to obtain standardized key gas concentration features and standardized candidate ratio feature sets, respectively; the standardized candidate ratio feature sets are then discretized to obtain discretized candidate ratio feature sets.
[0009] Based on the obtained discretized candidate ratio feature set, the AmRMR algorithm is used to perform redundancy compression and information contribution evaluation on the discretized candidate ratio features, and the most representative key ratio features for fault classification are selected, and then the key ratio feature set is output.
[0010] The standardized key gas concentration features and the obtained key ratio feature set are input into the pre-constructed DSD-DQN fault diagnosis model, and the transformer fault diagnosis results are output.
[0011] In conjunction with the first aspect, the construction of the candidate ratio feature set based on gas detection data further includes:
[0012] The construction form is shown in the following formula:
[0013] ;
[0014] in, The logarithmic ratio eigenvalue represents the constructed logarithmic ratio. Indicates the first Concentration values of the gas; Indicates the first The concentration value of the gas or the concentration value of the total hydrocarbons.
[0015] In conjunction with the first aspect, further, the standardization processing of the key gas concentration features and the candidate ratio feature set to obtain standardized key gas concentration features and standardized candidate ratio feature sets includes:
[0016] The standardization process uses the Z-score method, whose mathematical expression is shown below:
[0017] ;
[0018] in, This represents the standardized candidate ratio eigenvalue; The feature sample value refers specifically to the original gas concentration data of DGA and the logarithmic ratio feature data calculated by the construction formula of the candidate ratio feature set based on gas detection data. This represents the mean of the candidate ratio feature in the training set; This represents the standard deviation of the candidate ratio feature in the training set.
[0019] In conjunction with the first aspect, further, the discretization operation on the standardized candidate ratio feature set to obtain the discretized candidate ratio feature set includes:
[0020] The mathematical definition of the discretization operation is shown in the following equation:
[0021] ;
[0022] in, This represents the discretized candidate ratio feature label, with a value range of [value range missing]. ; This represents the standardized candidate ratio eigenvalue; Indicates the number of discretization levels; They represent the 1st to the 1st. 1 quantile.
[0023] In conjunction with the first aspect, further, based on the obtained discretized candidate ratio feature set, the AmRMR algorithm is used to perform redundancy compression and information contribution evaluation on the discretized candidate ratio features, and the most representative key ratio features for fault classification are selected, thereby outputting a key ratio feature set, including:
[0024] Mutual information is used to measure the dependence between individual candidate ratio features and fault labels, as defined in the following formula:
[0025] ;
[0026] in, It is a joint probability distribution; , They are distributed along their respective edges; Representing characteristic variables With fault labels Mutual information values between them; Representing characteristic variables The specific values of , i.e., the discretized candidate ratio feature labels, have a set of values as follows: ; Indicates fault label variable The specific value is the actual fault label of the transformer;
[0027] Feature selection is performed using an iterative round-by-round approach. Conditional mutual information is used to calculate the net information gain of candidate ratio features. The definition of conditional mutual information is as follows:
[0028] ;
[0029] in, It is a joint probability distribution. Given Z, we define the conditional probability distribution; based on this, we define the candidate ratio feature relative to the selected feature set. Net information gain As shown in the following formula:
[0030] ;
[0031] in, Indicates the first One candidate ratio feature; For fault labels; Indicates the selected feature set;
[0032] In each iteration, the net information gain is first calculated. Calculate the first Contribution rate of single feature information of each candidate ratio feature As shown in the following formula:
[0033] ;
[0034] Define cumulative information contribution rate For the front The ratio of the sum of the net information gains of the candidate ratio features to the total net information gain of the currently selected feature set is used to evaluate the saturation of the feature subset, as shown in the following formula:
[0035] ;
[0036] in: Indicates the currently selected feature set The total number of features in; Indicates the first Candidate ratio features In a given set The net information gain below; Indicates the first One candidate ratio feature; This represents the feature set of remaining candidate ratios; The truncated index variable represents the cumulative summation, where , representing the top [number] after sorting by contribution rate One candidate ratio feature; Indicates the preceding The cumulative information contribution rate of each candidate ratio feature;
[0037] Set dual adaptive constraints during the screening process:
[0038] Constraint 1: The Net information contribution rate of each candidate ratio feature No less than the largest information contribution rate in the first round 0.3 times;
[0039] Constraint 2: The cumulative information contribution rate reaches the preset target threshold;
[0040] If a candidate ratio feature satisfies both of the above constraints, the candidate ratio feature is added to the selected feature set and enters the next round of screening; if the candidate ratio feature does not satisfy either constraint, the iteration stops and the final key ratio feature set is output.
[0041] In conjunction with the first aspect, further, the standardized key gas concentration features and the obtained key ratio feature set are input into the pre-constructed DSD-DQN fault diagnosis model, and the output transformer fault diagnosis results include:
[0042] The DSD-DQN fault diagnosis model includes a convolutional neural network, which performs deep representation learning on the selected features;
[0043] The standardized key gas concentration features and the obtained key ratio features are used as inputs to the DSD-DQN fault diagnosis model. The specific feature extraction process includes, in sequence, a first convolutional layer, a first max pooling, a second convolutional layer, a second max pooling, a flattening operation, and a fully connected layer mapping process. The flattened feature vectors are mapped to a low-dimensional embedding space, which serves as the input to the reinforcement learning part for subsequent state value functions. Advantage function for each action Separate estimates, combined state value function Advantage function for each action Get the final Value function, The definition of a value function is:
[0044] ;
[0045] in, Indicates the parameters of the shared feature extraction layer. , These are the parameters for the advantage function branch and the state value function branch, respectively; This represents the action variables used to calculate the average of the advantage function; This represents the average advantage value of all actions in the current state;
[0046] Target The formula for calculating the value is shown below:
[0047] ;
[0048] in, These represent the parameters of the target network. This is a termination status identifier variable; This indicates the state transitioned to at the next moment after the action is performed; Indicates the current time goal value; Indicates the agent at time t. The immediate reward obtained after performing an action; This represents the overall parameters of the currently online network;
[0049] Based on the calculated target at the current time value Construct a loss function for updating network parameters; the loss function is used to minimize the current online network output. value With the goal value The prediction error between them Indicates time The state; Indicates the agent at time t. The actual actions performed.
[0050] In conjunction with the first aspect, the training and optimization method for the DSD-DQN fault diagnosis model is as follows:
[0051] A priority experience replay mechanism is introduced. By calculating the temporal difference error (TD error) of each state transition sample, a priority is assigned to each state transition sample. A larger error indicates a higher learning value for the current policy. The sampling probability is determined by the following formula, where For the first The TD error of each state transition sample. To control the priority of hyperparameters, non-uniform sampling is performed according to this probability during the empirical replay process, so that the model pays more attention to key states and difficult-to-classify samples, thereby improving the convergence speed and classification accuracy.
[0052] ;
[0053] in, Indicates the first The probability that a state transition sample is sampled; Indicates the first in the experience replay pool The absolute value of the time difference error of each state transition sample; This represents the index variable of the samples in the experience replay pool; Indicates the normalized denominator. This is the summation index variable for iterating through all stored samples in the experience replay pool;
[0054] A cost-sensitive reward function for the risk prior of fusion engineering is constructed. The expression of the cost-sensitive reward function is shown in the following equation, where... For actual fault categories, To predict the fault category, For the first Number of samples for each fault category The number of samples for the fault category with the most samples;
[0055] ;
[0056] ;
[0057] ;
[0058] in, This represents the immediate reward value of the DSD-DQN fault diagnosis model after performing a diagnostic action in the current state; This represents a cost-sensitive penalty coefficient that reflects the risks involved in the project. This represents the category imbalance weighting factor that reflects the data distribution; Indicates the first The class imbalance ratio of fault-type samples; This represents the average imbalance ratio across all fault categories, used to normalize the weights. This represents the total number of samples in the category with the largest number of samples in the training set. Indicates the first training set The total number of samples of this type of fault;
[0059] In defining the penalty coefficient Previously, based on the actual fault type of the transformer, faults were divided into three main categories: discharge faults, thermal faults, and normal faults; and a penalty coefficient was used to... The degree of engineering hazard is integrated into the learning process of the DSD-DQN fault diagnosis model. Specifically, the degree of hazard of different misjudgment types is classified as follows: false negative > cross-class misclassification > same-class misclassification > false alarm. False negative refers to a fault that is actually predicted as normal. Cross-class misclassification refers to misclassification between different fault categories. False alarm refers to a normal condition that is predicted as a fault. Same-class misclassification refers to misclassification of fault types within the same fault category.
[0060] Secondly, this invention proposes a transformer adaptive fault diagnosis system based on AmRMR, which implements the aforementioned transformer adaptive fault diagnosis method based on AmRMR, including:
[0061] The data preprocessing module is configured to acquire dissolved gas detection data in transformer oil, obtain key gas concentration characteristics, and construct a candidate ratio feature set based on the gas detection data; the key gas concentration characteristics and the candidate ratio feature set are standardized to obtain standardized key gas concentration characteristics and standardized candidate ratio feature sets; the standardized candidate ratio feature set is discretized to obtain a discretized candidate ratio feature set.
[0062] The adaptive feature selection module is configured to perform redundancy compression and information contribution evaluation on the discretized candidate ratio feature set based on the obtained discretized candidate ratio feature set using the AmRMR algorithm, select the key ratio features that are most representative for fault classification, and then output the key ratio feature set.
[0063] The fault identification module is configured to input the standardized key gas concentration features and the obtained key ratio feature set into the pre-built DSD-DQN fault diagnosis model and output the transformer fault diagnosis results.
[0064] Thirdly, the present invention proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described AmRMR-based transformer adaptive fault diagnosis method.
[0065] Fourthly, the present invention provides a computer device comprising:
[0066] Memory, used to store computer programs;
[0067] A processor is used to execute the computer program to implement the steps of the above-described AmRMR-based transformer adaptive fault diagnosis method.
[0068] Fifthly, the present invention proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described AmRMR-based transformer adaptive fault diagnosis method.
[0069] This invention constructs an initial ratio feature set to expand the feature space by referencing the traditional ratio method, and uses the AmRMR algorithm to select the final key feature set from the candidate set. Secondly, it constructs a DSD-DQN fault diagnosis model, integrating a convolutional neural network (CNN) to extract deep features. Furthermore, it decouples state value and action advantage through a dual-stream decoupled deep Q-network structure and a dual Q-learning mechanism, enhancing the DSD-DQN fault diagnosis model's ability to analyze nonlinear fault features. Finally, it designs a cost-sensitive reward function optimization strategy based on priority experience replay and fusion of engineering risk priors. By collaboratively optimizing imbalanced sample weights and differentiated maintenance hazards, it achieves accurate identification of severe transformer faults. This invention can significantly improve the diagnostic accuracy of minority class samples while effectively avoiding high-risk decisions such as missed reports based on maintenance risks, achieving transformer fault diagnosis that balances engineering safety and diagnostic accuracy.
[0070] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0071] (1) The present invention introduces an adaptive maximum correlation minimum redundancy (AmRMR) feature selection mechanism at the feature processing level, which can automatically measure the correlation between each ratio feature and the fault type according to the actual data distribution, suppress redundancy between features, select a subset of features that are highly correlated with the fault mode and have low redundancy from the original feature set, reduce the dependence on manual experience feature screening, and improve the effectiveness of model input and the stability of diagnostic results.
[0072] (2) In the model learning level, the present invention adopts the DSD-DQN fault diagnosis model deep reinforcement learning structure containing convolutional feature extraction layer, and models transformer DGA fault diagnosis as a cost-sensitive decision process. Through branch modeling of state value function and advantage function and Double Q-Learning mechanism, it enhances the ability to express complex nonlinear feature patterns and reduces Q value estimation bias, thereby achieving higher diagnostic accuracy and training convergence stability in multi-fault category identification tasks.
[0073] (3) In the optimization of the learning process, this invention introduces priority experience replay and cost-sensitive reward function, organically combines class imbalance weights and engineering misjudgment costs and achieves decoupling, so that the model pays more attention to minority class and high-risk fault samples during training, and gives differentiated penalties for missed reports, cross-class misclassification, false reports and same-class misclassification, significantly improving the identification rate of minority class serious faults and the overall diagnostic robustness without additional data augmentation. Attached Figure Description
[0074] Figure 1 This is a schematic diagram of the technical route of the transformer adaptive fault diagnosis method based on AmRMR in Embodiment 1 of the present invention;
[0075] Figure 2 This is a schematic diagram of the adaptive maximum relevance minimum redundancy feature screening method in Embodiment 1 of the present invention;
[0076] Figure 3 This is a schematic diagram of the DSD-DQN fault diagnosis model network structure in Embodiment 1 of the present invention. Detailed Implementation
[0077] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.
[0078] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0079] Example 1
[0080] like Figure 1 As shown, the transformer adaptive fault diagnosis method based on AmRMR proposed in this embodiment includes the following steps:
[0081] Step 1: Data preprocessing.
[0082] Step 11: Collect dissolved gas detection data in transformer oil to obtain key gas concentration information (i.e., key gas concentration characteristics) and corresponding fault labels for each monitoring sample under different operating conditions and fault states, i.e., collect samples under different fault states.
[0083] Step 12: Construct candidate ratio features.
[0084] In this embodiment, candidate ratio features are constructed by combining five common combustible gas components in transformer oil: H2, CH4, C2H6, C2H4, and C2H2 (i.e., five key gases). The construction of candidate ratio features is based on the mechanism of relative concentration changes of gases in different types of faults, which can more effectively characterize the differences between fault types. Referring to the fault characteristic ratio type used in the traditional ratio method, 16 types of transformer faults were selected as candidate ratio features, that is, 16 candidate ratio features were constructed, and the natural logarithm was taken to suppress the fluctuation caused by the difference in their dimensions. Specifically, they include: C2H2 / H2, C2H2 / CH4, C2H2 / C2H6, C2H4 / H2, C2H4 / CH4, C2H6 / H2, CH4 / C2H6, CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, C2H2 / TH, H2 / TH, C2H4 / TH, CH4 / TH, C2H6 / TH, (CH4+C2H4) / TH, where TH represents the total hydrocarbon concentration value, TH=CH4+C2H6+C2H4+C2H2, and its general construction form is shown in Equation (1).
[0085] (1);
[0086] in, Represents the constructed candidate ratio eigenvalues; Indicates the first Concentration values of the gas; Indicates the first The concentration value of the gas or the concentration value of the total hydrocarbons.
[0087] To eliminate the impact of differences in the dimensions and numerical distribution of the model's input features on subsequent mutual information evaluation and model training stability, the key gas concentration features and constructed candidate ratio features were standardized. The standardization adopted the Z-score method, and its mathematical expression is shown in Equation (2). After standardization, the mean of each ratio feature sample is 0 and the standard deviation is 1.
[0088] (2);
[0089] in, This represents the standardized candidate ratio feature value, which serves as the standard data form for subsequent mutual information calculation and input to the DSD-DQN fault diagnosis model. The characteristic sample value refers to the original gas concentration data of DGA and the candidate ratio characteristic value calculated by formula (1); This represents the mean of the candidate ratio feature in the training set; This represents the standard deviation of the candidate ratio feature in the training set.
[0090] To facilitate the calculation of mutual information between candidate ratio features and fault labels, a quantile discretization method is further employed on the basis of standardization to map continuous features to finite interval labels. The mathematical definition of the discretization process is shown in equation (3), which divides the standardized candidate ratio feature values into several quantiles. The discretization level helps to improve the stability and computational efficiency of mutual information estimation while maintaining a balanced sample distribution.
[0091] (3);
[0092] in, This represents the discretized candidate ratio feature label, with a value range of [value range missing]. ; This represents the standardized candidate ratio eigenvalue; Indicates the number of discretization levels; They represent the 1st to the 1st. 1 quantile.
[0093] Step 2: AmRMR feature selection.
[0094] like Figure 2 As shown, in the adaptive maximum relevance minimum redundancy (AmRMR) feature screening step of this embodiment, after data preprocessing is completed, 16 candidate ratio features derived from the original five key DGA gas features are used as the input basis for subsequent adaptive feature screening.
[0095] Step 21: Calculate the mutual information between candidate ratio features and fault labels. First, the mutual information (MI) is used to measure the dependence between individual candidate ratio features and fault labels, as defined in Equation (4). Mutual information reflects the feature variables. With fault labels The amount of information shared between them. (4);
[0096] in, It is a joint probability distribution; , They are distributed along their respective edges; Representing characteristic variables With fault labels Mutual information values between them; Representing characteristic variables The specific values of are the discretized candidate ratio feature labels obtained after processing by formula (3), and their value set is as follows: ; Indicates fault label variable The specific value is the actual fault label of the transformer.
[0097] Step 22: In this embodiment, feature selection is performed using an iterative round-by-round approach. Conditional mutual information is used to calculate the net information gain of candidate ratio features. This is to accurately evaluate the selected feature set. Given the information, candidate ratio features for fault labels To further contribute, conditional mutual information (CMI) is introduced. The definition of conditional mutual information is shown in equation (5):
[0098] (5);
[0099] in, It is a joint probability distribution. This is the conditional probability distribution given Z. Based on this, we define the candidate ratio feature relative to the selected feature set. Net information gain As shown in equation (6):
[0100] (6);
[0101] in, Indicates the first One candidate ratio feature; For fault labels; This indicates the selected feature set.
[0102] Step 23: To improve the adaptability of the AmRMR method, this embodiment introduces an adaptive feature information contribution rate mechanism. In each iteration, the net information gain calculated above is first... Calculate the first Contribution rate of single feature information of each candidate ratio feature It is used to measure the ratio of the current candidate feature relative to the selected feature set. The net information gain is the relative proportion of all remaining candidate ratio features, as defined in equation (7); meanwhile, the cumulative information contribution rate is defined. For the front The ratio of the sum of the net information gains of the candidate ratio features to the total net information gain of the currently selected feature set is used to evaluate the saturation of the feature subset, as shown in Equation (8).
[0103] (7);
[0104] (8);
[0105] in: Indicates the currently selected feature set The total number of features in; Indicates the first Candidate ratio features In a given set The net information gain below; Indicates the first One candidate ratio feature; Indicates the first The contribution rate of single feature information of each candidate ratio feature; This represents the feature set of remaining candidate ratios; The truncated index variable represents the cumulative summation. ), representing the top [number] after sorting by contribution rate. One candidate ratio feature; Indicates the preceding The cumulative information contribution rate of each candidate ratio feature. This value reflects the previous... The proportion of information carried by each candidate ratio feature relative to the total information of the currently selected feature set is used to evaluate the saturation of the feature subset.
[0106] Set dual adaptive constraints during the screening process:
[0107] Constraint 1: The Net information contribution rate of each candidate ratio feature No less than the largest information contribution rate in the first round 0.3 times, that is ;
[0108] Constraint 2: Cumulative Information Contribution Rate Reaching the preset target threshold, i.e. .
[0109] If a candidate ratio feature satisfies both of the above constraints, the candidate ratio feature is added to the selected feature set and enters the next round of screening; if either constraint is not satisfied (i.e. the feature contribution is too low or the accumulated information is saturated), the iteration stops and the final key ratio feature set is output, thereby achieving an adaptive balance between the number of features and information gain.
[0110] Step 3: Fault identification based on the DSD-DQN fault diagnosis model.
[0111] like Figure 3 As shown, in the fault identification step based on Dual-Stream Decoupled DeepQ-Network (DSD-DQN) in this embodiment, a convolutional neural network (CNN) is first used to perform deep representation learning on the selected features.
[0112] The numerical feature vector formed by combining the final key ratio feature selected by the adaptive feature selection module with the key gas concentration feature is used as the input to the DSD-DQN fault diagnosis model. The specific feature extraction process includes:
[0113] 1) First convolutional layer: 32 one-dimensional convolutional kernels are used to perform convolution operations on the input feature vector, with a kernel size of k. size =3, step size is s tride =1, fill size is p ad =1, this layer combines the ReLU activation function to extract local correlation patterns between input features, and enhances the DSD-DQN fault diagnosis model's ability to express feature combination structures;
[0114] 2) First max pooling: Max pooling is applied to the output of the first convolutional layer with a pooling window size of 2 and a stride of 2 to reduce feature dimensionality, reduce computational complexity and enhance the generalization ability of the DSD-DQN fault diagnosis model.
[0115] 3) Second convolutional layer: 64 one-dimensional convolutional kernels are applied to further extract deeper, higher-order features. The kernel size is k. size =3, step size is s tride =1, fill size is p ad =1, consistent with the convolution parameters of the first convolutional layer, to enhance the DSD-DQN fault diagnosis model's ability to perceive complex nonlinear features;
[0116] 4) Second max pooling: Max pooling is used to further reduce dimensionality, reduce the risk of overfitting, and retain the most discriminative feature patterns.
[0117] 5) Flattening operation: Flatten the multidimensional tensor feature output into a one-dimensional vector to facilitate subsequent processing by fully connected layers;
[0118] 6) Fully Connected Layer Mapping: A fully connected layer with ReLU activation is used to map the flattened feature vectors to a low-dimensional embedding space. This embedding space serves as the input to the reinforcement learning part, used for the subsequent estimation of the state value function and the advantage function. The advantage function branch output has a dimension of 64×6, where A represents the action advantage function value output by this branch; the state value branch output has a dimension of 64×1, where V represents the state value function output by this branch. The outputs of these two branches are fused, i.e., Q(V, A) represents the final synthesized state-action value function, with a dimension of 64×6, corresponding to the final score for each diagnostic action.
[0119] In the fault identification process, the multi-class classification task is modeled as a Markov decision process (MDP), which includes a state space. Action space Reward function State transition probability and discount factor :
[0120] state It consists of key gas characteristic data in the transformer oil to be diagnosed and key ratio characteristics after screening;
[0121] action For diagnostic decision-making, corresponding to each transformer fault category;
[0122] award The reward function is given by the cost-sensitive reward function in the optimization step of the learning process, which will be described later.
[0123] State transition: Implemented by traversing samples or constructing in batches;
[0124] Discount factor Used to weigh current rewards against future rewards.
[0125] This embodiment employs the DSD-DQN fault diagnosis model structure to improve the estimation accuracy of the state value function and action advantage function. The neural network architecture splits into two parallel branches after the shared feature extraction layer: one branch is used to estimate the state value function. Another branch is used to estimate the advantage function for each action. The two were combined to form the final one. The value function is defined as shown in equation (9). This structure helps to alleviate the problems of low training efficiency and estimation bias in traditional DQN in high-dimensional action spaces.
[0126] (9);
[0127] in, Indicates the parameters of the shared feature extraction layer. , These are the parameters for the advantage function branch and the state value function branch, respectively; This represents the action variables used to calculate the average of the advantage function; This represents the average advantage value of all actions in the current state.
[0128] To further reduce To address the value estimation problem, a Double Q-Learning mechanism is introduced. Specifically, action selection is performed by the current online network, while action value evaluation is completed by the target network, thereby decoupling selection and evaluation. Its objective is... The value calculation formula is shown in Equation (10). When the DSD-DQN fault diagnosis model incorrectly identifies the minority class, the cumulative reward for this round is terminated to strengthen the penalty.
[0129] (10);
[0130] in, These represent the parameters of the target network. This is a termination status identifier variable; This indicates the state transitioned to at the next moment after the action is performed; Indicates the current time goal value; Indicates the agent at time t. The immediate reward obtained after performing an action; This represents the overall parameters of the currently online network.
[0131] The target at the current time obtained based on the above calculations value We construct a loss function for updating network parameters. The loss function aims to minimize the current online network output. value With the goal value The prediction error between them, where, Indicates time The state; Indicates the agent at time t. The actual actions performed. Specifically, this embodiment uses mean squared error (MSE) as the loss function and updates the overall network parameters through backpropagation using the gradient descent algorithm, so that the model's predictions are accurate. The value gradually approaches the true target value.
[0132] During training, the agent is based on - Greedy strategies select actions based on probability. Randomly select actions to explore, with a probability of 1- Select Current Utilize the action with the highest value and transfer all state transition tuples. The samples are stored in an experience replay pool. When the number of samples stored in the experience replay pool reaches the threshold for extracting empirical values, parameters are updated through mini-batch sampling. After multiple rounds of iterative training, the DSD-DQN fault diagnosis model can learn a near-optimal fault classification strategy.
[0133] In the optimization steps of the learning process of the DSD-DQN fault diagnosis model, this embodiment adopts a combination of priority experience replay mechanism and cost-sensitive reward function to improve the model training efficiency, the ability to identify minority fault samples, and the suppression of high-risk misjudgment decision-making behavior, thereby achieving the optimization goal of balancing diagnostic accuracy and operational safety.
[0134] On the one hand, a priority experience replay mechanism is introduced. By calculating the temporal difference error (TD error) of each state transition sample, a priority is assigned to each state transition sample; the larger the error, the higher the learning value of the state transition sample for the current policy. (State transition samples) The sampling probability can be determined according to equation (11), where For the first The TD error of each state transition sample. To control the priority of hyperparameters, non-uniform sampling is performed according to this sampling probability during empirical replay. This allows the DSD-DQN fault diagnosis model to pay more attention to critical states and hard-to-classify samples, thereby improving convergence speed and classification accuracy.
[0135] (11);
[0136] in, Indicates the first The probability that a state transition sample is sampled; Indicates the first in the experience replay pool The absolute value of the time difference error of each state transition sample; This represents the index variable of the samples in the experience replay pool; Indicates the normalized denominator. This is the summation index variable for iterating through all stored samples in the experience replay pool.
[0137] On the other hand, a cost-sensitive reward function that integrates prior engineering risks is constructed. Specifically, based on the different proportions of sample classes in the dataset, a class balancing weight is introduced, so that a higher reward is given when the minority class is correctly identified; based on the difference in actual risks caused by different fault misjudgments to transformer operation and maintenance, a cost-sensitive penalty mechanism is introduced, so that a greater penalty is given when a serious fault is misjudged as a normal state, thereby guiding the DSD-DQN fault diagnosis model to converge toward the optimal strategy that balances diagnostic accuracy and engineering safety. The basic form of the cost-sensitive reward function is shown in Equation (12), where For actual fault categories, To predict the fault category, For the first Number of samples for each fault category The number of samples for the fault category with the most samples.
[0138] (12);
[0139] (13);
[0140] (14);
[0141] in, This represents the immediate reward value of the DSD-DQN fault diagnosis model after performing a diagnostic action in the current state; This represents a cost-sensitive penalty coefficient that reflects the risks involved in the project. This represents the category imbalance weighting factor that reflects the data distribution; Indicates the first The class imbalance ratio of fault-type samples; This represents the average imbalance ratio across all fault categories, used to normalize the weights. This represents the total number of samples in the category with the largest number of samples in the training set. Indicates the first training set The total number of samples of this type of fault.
[0142] In defining the penalty coefficient Previously, based on the actual fault type of the transformer, faults were divided into three main categories: discharge faults, thermal faults, and normal faults. A penalty coefficient was then used to... The severity of engineering hazards is integrated into the learning process of the DSD-DQN fault diagnosis model. Specifically, the severity of different misjudgment types is classified as follows: false negative > cross-class misclassification > same-class misclassification > false positive. False negative refers to a situation where a fault is actually predicted as normal; cross-class misclassification refers to misclassification between different fault categories; false positive refers to a situation where a normal condition is actually predicted as a fault; and same-class misclassification refers to misclassification of a fault type within the same fault category. Penalty coefficients are assigned to each of these four types of errors, with a relative ratio of 3:2:1.5:1.
[0143] By implementing the above settings, the problem of low serious fault identification rate caused by data class imbalance is addressed by adjusting the class imbalance weights that reflect data distribution. Cost-sensitive penalty coefficient reflecting engineering risk By decoupling the reward function, the DSD-DQN fault diagnosis model can simultaneously consider the sample distribution characteristics and the severity of fault misjudgment during training, thereby improving the recognition rate of minority classes such as severe faults and the overall diagnostic reliability.
[0144] Example 2
[0145] Based on the fault diagnosis method in Example 1, this example provides a specific engineering application method.
[0146] In practical applications, several in-service or experimental oil-immersed transformers were selected, and DGA detection data were collected under various operating conditions, including normal operation, partial discharge, high-energy discharge, low-energy discharge, and overheating at different temperature levels. Experts or experimental records provided corresponding fault labels. The sample set was divided into training, validation, and test sets. Following the steps in Example 1, the ratio feature construction, standardization and discretization, AmRMR adaptive feature selection, DSD-DQN fault diagnosis model construction, and learning process optimization were completed. The network structure and hyperparameters were adjusted on the validation set, and accuracy, various recall rates, and F1 scores were evaluated on the test set. Experimental results show that the method of this invention can significantly improve the recognition rate of severe fault categories without additional data augmentation, verifying the effectiveness of this method in processing multi-class imbalanced fault data.
[0147] Example 3
[0148] Based on the same inventive concept as Embodiment 1, this embodiment introduces a transformer adaptive fault diagnosis system based on AmRMR, used to implement the transformer adaptive fault diagnosis method based on AmRMR in Embodiment 1, including:
[0149] The data preprocessing module first acquires the dissolved gas detection data in the oil of the transformer to be diagnosed, and constructs a candidate ratio feature set based on the characteristics of five key gases. At the same time, it is also responsible for standardizing the original gas detection data (i.e., key gas concentration characteristics) and the candidate ratio feature set, and discretizing the standardized candidate ratio feature set to obtain the discretized candidate ratio feature set. The continuous data is converted into discrete level labels to provide standardized input data for subsequent feature selection.
[0150] The adaptive feature selection module receives the candidate ratio feature set after discretization and calculates the net information gain and information contribution rate of the candidate ratio features based on the adaptive maximum correlation minimum redundancy (AmRMR) algorithm logic. By evaluating the correlation between the candidate ratio features and the fault label and the redundancy between features, the module uses dual adaptive constraints to select the key ratio features that are most representative of fault classification, and then outputs the final key ratio feature set to reduce data dimensionality and improve feature quality.
[0151] The fault identification module constructs a fault diagnosis model based on a dual-stream decoupled deep Q-network (DSD-DQN). It integrates a convolutional neural network (CNN) to extract deep features, and concatenates key gas concentration features and a set of selected key ratio features as input to the DSD-DQN fault diagnosis model. At the same time, the state value function and action advantage function are independently modeled through the dual-stream decoupled network structure, and finally fused to generate the state-action value (Q value) to achieve intelligent identification of fault categories.
[0152] The learning process optimization module optimizes the training process of the DSD-DQN fault diagnosis model. It introduces a priority experience replay mechanism to improve learning efficiency on difficult samples and constructs a cost-sensitive reward function that incorporates engineering risk priors to address issues such as class imbalance and misclassification risk asymmetry. The network parameters are updated using the backpropagation algorithm, and the final transformer fault diagnosis result is output using the trained and optimized DSD-DQN fault diagnosis model.
[0153] Example 4
[0154] Based on the same inventive concept as other embodiments, this embodiment introduces a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described AmRMR-based transformer adaptive fault diagnosis method.
[0155] Example 5
[0156] Based on the same inventive concept as other embodiments, this embodiment introduces a computer device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described AmRMR-based transformer adaptive fault diagnosis method.
[0157] Example 6
[0158] Based on the same inventive concept as other embodiments, this embodiment introduces a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described AmRMR-based transformer adaptive fault diagnosis method.
[0159] In summary, this invention innovatively combines AmRMR feature selection with the DSD-DQN deep reinforcement learning model to construct an overall method of "adaptive feature selection + reinforcement learning decision-making" for the diagnosis of dissolved gas faults in transformer oil. Through logarithmic ratio feature construction and AmRMR adaptive selection, this invention suppresses feature redundancy while considering the correlation between features and fault categories. Compared to traditional methods that rely on manual experience to select features, it can more fully explore fault modes in DGA data, improving the effectiveness and stability of the diagnostic model's input features.
[0160] This invention employs convolutional feature extraction and a DSD-DQN network structure to model multi-class fault diagnosis as a cost-sensitive decision-making process. It jointly estimates the Q-values of each fault class through state-value functions and advantage function branches, and incorporates Double Q-Learning to mitigate Q-value overestimation, thereby improving the model's convergence and diagnostic reliability under multi-class, imbalanced data conditions. By constructing a cost-sensitive reward function that simultaneously considers class imbalance weights and engineering misjudgment costs, and introducing a priority experience replay mechanism to highlight minority and hard-to-classify samples, this invention significantly improves the recognition rate and overall diagnostic performance of key categories such as severe faults under limited and unevenly distributed real-world DGA data conditions.
[0161] This invention provides a novel technical solution for intelligent fault diagnosis of transformers, which can be integrated into transformer online monitoring systems, condition assessment platforms, or cloud-based diagnostic services. By improving the accuracy and timeliness of fault identification, it helps reduce the risk of sudden failures and operation and maintenance costs, thereby enhancing the safety and economy of power grid operation. The adaptive feature selection and cost-sensitive reinforcement learning framework proposed in this invention also has application value for other equipment condition assessment and fault diagnosis problems involving data imbalance and differences in misjudgment costs.
[0162] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0163] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0165] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0166] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention, and these modifications are all within the protection scope of the present invention.
Claims
1. A transformer adaptive fault diagnosis method based on AmRMR, characterized in that, Includes the following steps: Obtain dissolved gas detection data in transformer oil, obtain key gas concentration characteristics, and construct a candidate ratio feature set based on the gas detection data; The key gas concentration features and candidate ratio feature sets are standardized to obtain standardized key gas concentration features and standardized candidate ratio feature sets, respectively; the standardized candidate ratio feature sets are then discretized to obtain discretized candidate ratio feature sets. Based on the obtained discretized candidate ratio feature set, the AmRMR algorithm is used to perform redundancy compression and information contribution evaluation on the discretized candidate ratio features, and the most representative key ratio features for fault classification are selected, and then the key ratio feature set is output. The standardized key gas concentration features and the obtained key ratio feature set are input into the pre-constructed DSD-DQN fault diagnosis model, and the transformer fault diagnosis results are output.
2. The transformer adaptive fault diagnosis method based on AmRMR according to claim 1, characterized in that: The construction of the candidate ratio feature set based on gas detection data includes: The construction form is shown in the following formula: ； in, The logarithmic ratio eigenvalue represents the constructed logarithmic ratio. Indicates the first Concentration values of the gas; Indicates the first The concentration value of the gas or the concentration value of the total hydrocarbons.
3. The transformer adaptive fault diagnosis method based on AmRMR according to claim 1, characterized in that: The discretization operation on the standardized candidate ratio feature set to obtain the discretized candidate ratio feature set includes: The mathematical definition of the discretization operation is shown in the following equation: ； in, This represents the discretized candidate ratio feature label, with a value range of [value range missing]. ; This represents the standardized candidate ratio eigenvalue; Indicates the number of discretization levels; They represent the 1st to the 1st. 1 quantile.
4. The transformer adaptive fault diagnosis method based on AmRMR according to claim 1, characterized in that: Based on the obtained discretized candidate ratio feature set, the AmRMR algorithm is used to perform redundancy compression and information contribution evaluation on the discretized candidate ratio features, and the most representative key ratio features for fault classification are selected, thereby outputting a key ratio feature set, including: Mutual information is used to measure the dependence between individual candidate ratio features and fault labels, as defined in the following formula: ； in, It is a joint probability distribution; , They are distributed along their respective edges; Representing characteristic variables With fault labels Mutual information values between them; Representing characteristic variables The specific values of , i.e., the discretized candidate ratio feature labels, have a set of values as follows: ; Indicates fault label variable The specific value is the actual fault label of the transformer; Feature selection is performed using an iterative round-by-round approach. Conditional mutual information is used to calculate the net information gain of candidate ratio features. The definition of conditional mutual information is as follows: ； in, It is a joint probability distribution. Given Z, we define the conditional probability distribution; based on this, we define the candidate ratio feature relative to the selected feature set. Net information gain As shown in the following formula: ； in, Indicates the first One candidate ratio feature; For fault labels; Indicates the selected feature set; In each iteration, the net information gain is first calculated. Calculate the first Contribution rate of single feature information of each candidate ratio feature As shown in the following formula: ； Define cumulative information contribution rate For the front The ratio of the sum of the net information gains of the candidate ratio features to the total net information gain of the currently selected feature set is used to evaluate the saturation of the feature subset, as shown in the following formula: ； in: Indicates the currently selected feature set The total number of features in; Indicates the first Candidate ratio features In a given set The net information gain below; Indicates the first One candidate ratio feature; This represents the feature set of remaining candidate ratios; The truncated index variable represents the cumulative summation, where , representing the top [number] after sorting by contribution rate One candidate ratio feature; Indicates the preceding The cumulative information contribution rate of each candidate ratio feature; Set dual adaptive constraints during the screening process: Constraint 1: The Net information contribution rate of each candidate ratio feature No less than the largest information contribution rate in the first round 0.3 times; Constraint 2: The cumulative information contribution rate reaches the preset target threshold; If a candidate ratio feature satisfies both of the above constraints, the candidate ratio feature is added to the selected feature set and enters the next round of screening; if the candidate ratio feature does not satisfy either constraint, the iteration stops and the final key ratio feature set is output.
5. The transformer adaptive fault diagnosis method based on AmRMR according to claim 1, characterized in that: The standardized key gas concentration features and the obtained key ratio feature set are input into the pre-constructed DSD-DQN fault diagnosis model, and the output transformer fault diagnosis results include: The DSD-DQN fault diagnosis model includes a convolutional neural network, which performs deep representation learning on the selected features; The standardized key gas concentration features and the obtained key ratio features are used as inputs to the DSD-DQN fault diagnosis model. The specific feature extraction process includes, in sequence, a first convolutional layer, a first max pooling, a second convolutional layer, a second max pooling, a flattening operation, and a fully connected layer mapping process. The flattened feature vectors are mapped to a low-dimensional embedding space, which serves as the input to the reinforcement learning part for subsequent state value functions. Advantage function for each action Separate estimates, combined state value function Advantage function for each action Get the final Value function, The definition of a value function is: ； in, Indicates the parameters of the shared feature extraction layer. , These are the parameters for the advantage function branch and the state value function branch, respectively; This represents the action variables used to calculate the average of the advantage function; This represents the average advantage value of all actions in the current state; Target The formula for calculating the value is shown below: ； in, These represent the parameters of the target network. This is a termination status identifier variable; This indicates the state transitioned to at the next moment after the action is performed; Indicates the current time goal value; Indicates the agent at time t. The immediate reward obtained after performing an action; This represents the overall parameters of the currently online network; Based on the calculated target at the current time value Construct a loss function for updating network parameters; the loss function is used to minimize the current online network output. value With the goal value The prediction error between them Indicates time The state; Indicates the agent at time t. The action performed.
6. The transformer adaptive fault diagnosis method based on AmRMR according to claim 1, characterized in that: The training and optimization method for the DSD-DQN fault diagnosis model is as follows: A priority experience replay mechanism is introduced. By calculating the temporal difference error (TD error) of each state transition sample, a priority is assigned to each state transition sample. A larger error indicates a higher learning value for the current policy. The sampling probability is determined by the following formula, where For the first The TD error of each state transition sample. Hyperparameters for controlling priority; During the experience replay process, non-uniform sampling is performed according to this probability, so that the model pays more attention to key states and difficult-to-classify samples, thereby improving the convergence speed and classification accuracy. ； in, Indicates the first The probability that a state transition sample is sampled; Indicates the first in the experience replay pool The absolute value of the time difference error of each state transition sample; This represents the index variable of the samples in the experience replay pool; Indicates the normalized denominator. This is the summation index variable for iterating through all stored samples in the experience replay pool; A cost-sensitive reward function for the risk prior of fusion engineering is constructed. The expression of the cost-sensitive reward function is shown in the following equation, where... For actual fault categories, To predict the fault category, For the first Number of samples for each fault category The number of samples for the fault category with the most samples; ； ； ； in, This represents the immediate reward value of the DSD-DQN fault diagnosis model after performing a diagnostic action in the current state; This represents a cost-sensitive penalty coefficient that reflects the risks involved in the project. This represents the category imbalance weighting factor that reflects the data distribution; Indicates the first The class imbalance ratio of fault-type samples; This represents the average imbalance ratio across all fault categories, used to normalize the weights. This represents the total number of samples in the category with the largest number of samples in the training set. Indicates the first training set The total number of samples of this type of fault; In defining the penalty coefficient Previously, based on the actual fault type of the transformer, faults were divided into three main categories: discharge faults, thermal faults, and normal faults; and a penalty coefficient was used to... The degree of engineering hazard is integrated into the learning process of the DSD-DQN fault diagnosis model. Specifically, the degree of hazard of different misjudgment types is classified as follows: false negative > cross-class misclassification > same-class misclassification > false alarm. False negative refers to a fault that is actually predicted as normal. Cross-class misclassification refers to misclassification between different fault categories. False alarm refers to a normal condition that is predicted as a fault. Same-class misclassification refers to misclassification of fault types within the same fault category.
7. A transformer adaptive fault diagnosis system based on AmRMR, characterized in that, Implementing the transformer adaptive fault diagnosis method based on AmRMR as described in any one of claims 1 to 6, comprising: The data preprocessing module is configured to acquire dissolved gas detection data in transformer oil, obtain key gas concentration characteristics, and construct a candidate ratio feature set based on the gas detection data; the key gas concentration characteristics and the candidate ratio feature set are standardized to obtain standardized key gas concentration characteristics and standardized candidate ratio feature sets; the standardized candidate ratio feature set is discretized to obtain a discretized candidate ratio feature set. The adaptive feature selection module is configured to perform redundancy compression and information contribution evaluation on the discretized candidate ratio feature set based on the obtained discretized candidate ratio feature set using the AmRMR algorithm, select the key ratio features that are most representative for fault classification, and then output the key ratio feature set. The fault identification module is configured to input the standardized key gas concentration features and the obtained key ratio feature set into the pre-built DSD-DQN fault diagnosis model and output the transformer fault diagnosis results.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AmRMR-based transformer adaptive fault diagnosis method as described in any one of claims 1 to 6.
9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the AmRMR-based transformer adaptive fault diagnosis method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the AmRMR-based transformer adaptive fault diagnosis method as described in any one of claims 1 to 6.