A method and device for training and deploying a transformer anomaly detection model

Abstract

The present invention discloses a method and device for training and deploying a transformer anomaly detection model. The method includes the steps of: training a transformer anomaly detection model; performing multiple iterative validations on the trained transformer anomaly detection model to obtain a corresponding number of output results; post-processing the output results to obtain a corresponding number of prediction results, and based on all the prediction results, using a comprehensive judgment strategy to obtain a final prediction result; model deployment and online inference, anomaly annotation and early warning statistics, etc. The method of the present invention is driven by real-time data, can automatically and iteratively train and deploy a transformer anomaly detection model online, can greatly reduce the cost of manual offline model training and deployment, continuously improve the prediction accuracy of the model, and reduce the anomaly false alarm rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer detection, and particularly relates to a method and device for training and deploying a transformer anomaly detection model based on data driving. Background Art

[0002] Traditional transformer fault detection in substations requires regular assignment of maintenance personnel to conduct on-site inspections, which not only results in high maintenance costs but also low efficiency and is unable to detect equipment anomaly signals in a timely manner. With the progress of science and technology, more and more substations have started to use artificial intelligence technology to conduct online real-time monitoring of various equipment in the station. When the AI model detects an abnormal voiceprint signal, it will immediately send a warning message, facilitating maintenance personnel to perform repairs in a timely manner and avoiding losses caused by equipment failures.

[0003] The current transformer anomaly detection models deployed in substations are mainly trained offline. Due to certain differences in the noise of the transformer body and environmental interference noise in each substation, the generalization of the model is poor, and it is difficult to train a general large model applicable to all transformers in substations. To improve the prediction accuracy of each transformer in the substation, it may be necessary to train and deploy a specific model for each transformer. When the number of stations and transformer equipment involved is large, it will consume a huge amount of human costs. Summary of the Invention

[0004] Technical Objective: Aiming at the above technical problems, the present invention proposes a method and device for training and deploying a transformer anomaly detection model based on data driving, which can automatically realize online training and deployment of transformer abnormal voiceprints.

[0005] Technical Solution: To achieve the above technical objective, the present invention adopts the following technical solutions:

[0006] A method for training and deploying a transformer anomaly detection model for an online transformer status monitoring system, characterized by including the steps of:

[0007] For the completed trained transformer anomaly detection model, use the test set for multiple iterative validations to obtain a corresponding number of output results;

[0008] Post-process the output results of the transformer anomaly detection model to obtain a corresponding number of prediction results, and based on all the prediction results, use a comprehensive judgment strategy to obtain the final prediction result.

[0009] Preferably, after the transformer anomaly detection model is verified, model deployment and online inference are performed, that is, the model parameters are automatically loaded into the transformer system, and an inference service interface is provided to the outside world to perform online inference on the transformer voiceprint data collected in real time. If the transformer anomaly detection model detects an abnormal signal, an early warning information is issued and the early warning data is stored.

[0010] Preferably, after completing the model deployment and online inference, manual anomaly correction is performed on the warning data on a regular basis to determine whether the anomaly inferred by the transformer anomaly detection model is a real anomaly and whether the category of the anomaly is true. If the warning data does not match the actual situation, manual classification is performed and the labeled data is updated to the voiceprint database for subsequent iterative training of the transformer anomaly detection model.

[0011] Preferably, the method further includes a step of regularly counting early warning data, that is, automatically counting the false alarm rate of the transformer anomaly detection model within a preset period. When the false alarm rate reaches a preset indicator, the automated training task is automatically triggered, and the model parameters of the previous transformer anomaly detection model are initialized, the parameters and the training set are updated, and iterative training is performed;

[0012] If the false alarm rate does not reach the expected target within a cycle, the training service is restarted through the scheduler to start the next round of iterative training. Otherwise, it will not be restarted.

[0013] Preferably, the transformer anomaly detection model includes a detection model and a classification model, and the post-processing steps include:

[0014] A voiceprint data is divided into several parts, preprocessed, and then input into the detection model and classification model for inference. The detection model and classification model each output an anomaly score. The two anomaly scores are weighted averaged to obtain a total anomaly score. The total anomaly score is compared with the preset benchmark score to obtain the prediction result;

[0015] The total anomaly score is determined by the following formula:

[0016] Score abnormal =(w1*Score det +w2*Score cls ) / (w1+w2)

[0017] Among them, w1 and w2 are weight coefficients, Score det To detect the abnormal score predicted by the model, Score cls Score is the anomaly score predicted by the classification model. cls =1-Confidence normal , Confidence normalThe confidence score of the normal class predicted by the classification model.

[0018] Preferably, the comprehensive judgment strategy is as follows: calculate the variance var of the total abnormal scores of N groups of prediction results, the empirical benchmark variance var', and the number n of abnormal predictions in N groups of classification model prediction results, and determine the final prediction result according to the following formula:

[0019]

[0020] y represents the prediction result. If the final prediction result obtained above is abnormal, that is, the class label is abnormal, the abnormal class is the type with the largest number of classes in N groups of classification results; otherwise, it is the normal class normal or the unknown class unknown.

[0021] Preferably, when verifying the transformer abnormal detection model, if the preset performance index is not reached within the preset maximum number of iterative training times, select the abnormal detection model with the best performance among them and proceed to the next step.

[0022] Preferably, the steps of training the transformer abnormal detection model include:

[0023] Obtain transformer voiceprint data including normal voiceprint data and abnormal voiceprint data as the original data. The abnormal voiceprint data is from the voiceprint database;

[0024] Preprocess the original data;

[0025] Based on the extracted features, use audio detection and classification algorithms to train a group of detection models and classification models in sequence. The transformer abnormal detection model includes a feature extraction model, a detection model, and a classification model. The detection model is used to detect whether there is an abnormality in the transformer voiceprint data, and the classification model is used to identify the type of the transformer voiceprint data;

[0026] After the transformer abnormal detection model is trained, use the test set to perform multiple iterative validations on the transformer abnormal detection model. If the performance reaches the preset index, proceed to the next step; otherwise, return to the step of preprocessing the original data, and adjust the hyperparameters of the transformer abnormal detection model and retrain with the training data.

[0027] A device for training and deploying a transformer abnormal detection model, which is used for a transformer system, and is characterized by including:

[0028] The transformer abnormal detection model is used to detect and identify abnormal data of the transformer voiceprint, and perform model training according to the updated voiceprint database of the system;

[0029] A verification module for iteratively verifying the transformer anomaly detection model that has completed training multiple times using a test set to obtain a corresponding number of output results;

[0030] A post - processing module that post - processes the output results of the transformer anomaly detection model to obtain a corresponding number of prediction results. Based on all the prediction results, using a comprehensive judgment strategy, the final prediction result is obtained;

[0031] A model deployment and online inference module for performing model deployment and online inference after the transformer anomaly detection model is verified, that is, automatically loading the model parameters into the transformer system, providing an inference service interface externally, performing online inference on the real - time collected transformer voiceprint data. If the transformer anomaly detection model detects an abnormal signal, a warning message is sent and the warning data is stored;

[0032] An anomaly annotation module for manually and regularly correcting the warning data by humans after the model deployment and online inference are completed to determine whether the anomalies inferred by the transformer anomaly detection model are real anomalies and whether the categories of the anomalies are true. If the warning data does not match the actual situation, the category is manually marked, and at the same time, the marked data is updated to the voiceprint database for subsequent iterative training of the transformer anomaly detection model;

[0033] A warning number regular statistics module for automatically counting the false alarm rate of the transformer anomaly detection model within a preset period. When the false alarm rate reaches the preset index, an automated training task is automatically triggered, initialized based on the model parameters of the previous transformer anomaly detection model, the training set is updated, and iterative training is performed; If the false alarm rate does not reach the expected target within a period, the training service is restarted through a scheduler to start the next round of iterative training, otherwise it is not restarted.

[0034] A computer - readable storage medium, characterized in that the computer - readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method.

[0035] Beneficial effects: Due to the adoption of the above - mentioned technical solutions, the present invention has the following beneficial effects:

[0036] A method for automated model online training and deployment proposed by the present invention can use real - time data to drive, automatically and iteratively perform online training and deployment of the model, which can greatly reduce the cost of manual offline model training and deployment. At the same time, it can also continuously improve the prediction accuracy of the model and reduce the abnormal false alarm rate. Description of the Drawings

[0037] Figure 1It is a flowchart of a method for training and deploying a transformer anomaly detection model based on data-driven;

[0038] Figure 2 It is a flowchart of data preprocessing;

[0039] Figure 3 The noise spectrogram of the voiceprint before noise reduction in;

[0040] Figure 4 The noise spectrogram of the voiceprint after noise reduction;

[0041] Figure 5 It is a flowchart of post-processing. Detailed implementation manners

[0042] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Embodiment 1

[0044] This embodiment provides a method for training and deploying a transformer anomaly detection model for an on-line monitoring system of transformer states, including the steps of:

[0045] For the completed trained transformer anomaly detection model, use the test set for multiple iterative validations to obtain a corresponding number of output results;

[0046] Perform post-processing on the output results of the transformer anomaly detection model to obtain a corresponding number of prediction results, and based on all the prediction results, use a comprehensive judgment strategy to obtain the final prediction result.

[0047] Specifically, after the transformer anomaly detection model is completed and verified, model deployment and on-line inference are performed, that is, the model parameters are automatically loaded into the transformer system, an inference service interface is provided externally, on-line inference is performed on the real-time collected transformer voiceprint data. If the transformer anomaly detection model detects an abnormal signal, a warning message is sent and the warning data is stored.

[0048] After the model deployment and on-line inference are completed, the warning data is manually corrected for anomalies regularly by personnel to determine whether the anomalies inferred by the transformer anomaly detection model are real anomalies and whether the categories of the anomalies are true. If the warning data does not match the real situation, the category is manually marked, and at the same time, the marked data is updated to the voiceprint database for subsequent iterative training of the transformer anomaly detection model.

[0049] Specifically, the transformer anomaly detection model includes a detection model and a classification model, and the steps of post-processing include:

[0050] Split a voiceprint data into several parts, perform preprocessing, and input them into a detection model and a classification model for inference in sequence. The detection model and the classification model respectively output an anomaly score. Perform a weighted average on the two anomaly scores to obtain a total anomaly score. Compare the total anomaly score with a preset benchmark score to obtain a prediction result.

[0051] The comprehensive judgment strategy is as follows: Calculate the variance var of the total anomaly scores of N groups of prediction results, the benchmark variance var' based on experience, and the number n of abnormal predictions in the prediction results of N groups of classification models to determine the final prediction result. The prediction results include three categories: abnormal, normal, and unknown. For the abnormal category, take the type with the largest number of categories in the N groups of classification results.

[0052] Among them, the steps of training the transformer anomaly detection model include:

[0053] Obtain transformer voiceprint data including normal voiceprint data and abnormal voiceprint data as the original data. The abnormal voiceprint data is from the voiceprint database;

[0054] Perform preprocessing on the original data;

[0055] Based on the extracted features, use audio detection and classification algorithms to train a group of detection models and classification models in sequence. The transformer anomaly detection model includes a feature extraction model, a detection model, and a classification model. The detection model is used to detect whether there is an anomaly in the transformer voiceprint data, and the classification model is used to identify the type of the transformer voiceprint data;

[0056] After the training of the transformer anomaly detection model is completed, use the test set to perform multiple iterative validations on the transformer anomaly detection model. If the performance reaches the preset index, proceed to the next step. Otherwise, return to the step of performing preprocessing on the original data, and adjust the hyperparameters of the transformer anomaly detection model and retrain with the training data.

[0057] Embodiment 2

[0058] As Figures 1 to 5 shown, this embodiment provides a data-driven method for training and deploying a transformer anomaly detection model, including the following steps 1-13:

[0059] 1. Obtain the original data

[0060] Obtain the transformer voiceprint data collected in real time by the voiceprint sensor, and its storage format is an audio file in wav format. Generally, at least one week of original voiceprint data needs to be obtained to ensure that the model can reach the preset index and has a certain degree of generalization. In addition to the normal voiceprint data, it is also necessary to obtain the abnormal voiceprint data stored in the voiceprint database.

[0061] 2. Set the model parameters

[0062] Start the automated model training and deployment task with model parameters such as learning rate, sampling rate, training batch size, number of iterations, etc.

[0063] 3. Data preprocessing

[0064] Cut the saved original voiceprint data to generate several wav files of the same duration (usually 10 seconds). Then, perform simple random sampling on these audio files to generate a training set and a test set in a ratio of 8:2. By default, the above training and test samples are all normal samples.

[0065] 4. Voiceprint data noise reduction

[0066] The transformer voiceprint data contains environmental interference sounds, which has a certain impact on the prediction accuracy and generalization of the model. Therefore, noise data needs to be denoised during data preprocessing. Common voice noise reduction and separation algorithms include spectral subtraction, wavelet decomposition, Wiener filtering, and deep learning-related algorithms. As Figure 3 and Figure 4 shown.

[0067] 5. Simulate abnormal data

[0068] Since there is little abnormal voiceprint data of transformers in the real scenario, a batch of abnormal voiceprint data needs to be simulated for training. Mix pure abnormal voiceprint data of transformers such as overcurrent, no-load, partial discharge, and high-voltage overexcitation with normal transformer voiceprint data according to a certain signal-to-noise ratio (-5 - 20 db) to simulate a batch of abnormal voiceprint data, and set abnormal class labels for each abnormal data as negative samples for subsequent model training and verification.

[0069] 6. Extract time-frequency features

[0070] Generally, the wav file saves the time-domain information of the voice signal. One can only see a time-domain signal that changes with time and cannot see how many sine and cosine functions are contained behind the signal. Therefore, it is necessary to perform Fourier transform to convert the time-domain signal into a frequency-domain signal. Then, extract log Mel spectrogram or MFCC coefficients, etc. as feature data based on the time-frequency signal, and use anomaly detection and classification algorithms to train the model on these features.

[0071] The Fourier transform formula used is as follows:

[0072]

[0073] Among them, the function F(w) in the frequency domain is expressed as an integral form of the function f(t) in the time domain.

[0074] 7. Model training

[0075] The transformer anomaly detection model includes a detection model and a classification model. An unsupervised (no labels required) neural network model such as Autoencoder can be used as the detection model, and supervised (labels required) models such as VGG16 and Resnet can be used as the classification model. Model training is performed based on the voiceprint features extracted in the previous step. First, a detection model and then a classification model are trained in sequence. The detection model is used to detect anomalies, and the classification model is used to identify the type of the voiceprint data, which can be a normal category or a certain anomaly category.

[0076] 8. Model Performance Verification

[0077] After the detection or classification model training is completed, the performance of the model needs to be verified. Only when the performance requirements are met can subsequent deployment be carried out. The trained model is used to perform class prediction on the labeled test samples, and the corresponding metric values are calculated. For example, the F1 score is calculated for the detection model, and the accuracy is calculated for the classification model. If the performance reaches the preset metrics (such as F1 score > 0.85 or accuracy > 0.9), then proceed to the next step; otherwise, return to data preprocessing, update the training data, and adjust the model hyperparameters (such as learning rate, mini-batch, number of iterations, optimization algorithm, etc.), and then perform iterative training again until the termination condition is met.

[0078] The maximum number of iterative training times can be set to N times. If none of the N times reaches the preset metrics, then select the model with the best performance among the N times and proceed to the next step.

[0079] Among them, the F1 score is used to evaluate the performance of the detection model. It takes into account both the precision and recall of the classification model and can be regarded as a harmonic mean of precision and recall. Its value range is 0 to 1, and the larger the value, the better the model performance. The formula for the model performance evaluation metric is as follows:

[0080]

[0081] precision represents the classification accuracy, and recall represents the recall rate. For the classification model, the classification accuracy (accuracy) is used as the performance measurement metric. The classification accuracy is expressed as the number of correctly classified samples divided by the total number of samples. The higher the classification accuracy, the better the model performance.

[0082] 9. Post-processing

[0083] Perform a trade-off on the output results of the unsupervised detection model and the supervised classification model. The unsupervised algorithm only trains on normal samples and learns the features of normal samples. When the new data has a large difference in features from the normal data, it will predict a large anomaly score. The supervised classification algorithm can only identify a limited number of anomaly types in the training set. Combining the two types of algorithms can reduce the false alarm rate and at the same time discover some unknown anomalies. As Figure 5 shown

[0084] (1) Calculate the anomaly score and its variance

[0085] Cut a voiceprint data into N parts (N 10-second audio files), and after preprocessing, input them into the detection model and the classification model for inference in turn. Among them, the detection model will directly output an anomaly score, the classification model will output a predicted category and a confidence score, and an anomaly score is obtained through conversion. Then, the two anomaly scores are weighted and averaged to obtain a weighted anomaly score.

[0086] Score cls = 1 - Confidence normal

[0087] where Confidence normal is the confidence score predicted by the classification model as the normal category, and its value range is a floating point number between 0 and 1.

[0088] Score abnormal = (w1 * Score det + w2 * Score cls ) / (w1 + w2)

[0089] where w1 and w2 are the weight coefficients respectively, Score det is the anomaly score predicted by the detection model, and Score cls is the score of the anomaly category predicted by the classification model.

[0090] Then calculate the variance var of the N anomaly scores Score abnormal .

[0091]

[0092] where u is the mean of the N Score abnormal .

[0093] (2) Calculate the final prediction result

[0094] After calculating the variance var based on N anomaly scores, we then set an empirically based baseline variance var'. Assuming that there are N sets of classification model predictions, with n predicted anomalies (the normal category is Nn, 0 <= n <= N), the final prediction result is calculated using the following formula.

[0095]

[0096] If the final prediction result is abnormal, the abnormal category is the category with the largest number of prediction categories in the N groups of classification prediction results (such as overcurrent, no-load, partial discharge, etc.). Otherwise, the normal or unknown category is output.

[0097] 10. Model deployment and online inference

[0098] After both the detection and classification models pass performance verification, the scheduler automatically deploys the task. The scheduler loads the two models into memory and provides an inference service interface. It performs online inference on the transformer voiceprint data collected and transmitted in real time by the sensor, outputting a prediction result. If the model identifies the audio data as abnormal, an alert is sent to the backend service.

[0099] The standard for a model to pass performance verification is as follows: the trained model is tested on the test set, and the overall F1 score or classification accuracy is calculated. If the performance index of the model on the test set is higher than the preset value, the model is considered to have passed performance verification.

[0100] Among them, online inference is as follows: after the model deployment is completed, the scheduler will start the online inference service, and it will be able to receive the transformer voiceprint data (usually a 60-second wav file) sent by the sensor end at a regular interval (such as every 20 seconds). The inference service will call the data preprocessing module and the anomaly detection and classification model to perform data processing and online inference on the voiceprint data respectively, and then input the prediction results into the post-processing module to output a final prediction result, such as: Is the audio file sent by the sensor normal or abnormal? If it is abnormal, what type of abnormality it belongs to.

[0101] 11. Anomaly marking

[0102] Operations and maintenance personnel regularly manually calibrate the warning data to determine whether the anomalies inferred by the model are true anomalies and whether the anomaly categories are accurate. If the warning results do not match the actual situation, manual classification is performed. The labeled data is also updated to the voiceprint database to facilitate subsequent iterative training.

[0103] 12. Regular statistics of warning numbers

[0104] After the model is deployed, the false alarm rate is automatically counted within a certain period (such as one month). When the false alarm rate reaches the preset index, an automated training task is automatically triggered, initialized based on the weight parameters of the previous model, and the model parameters and training samples are updated for iterative training.

[0105] Among them, the formula for calculating the false alarm rate is:

[0106] Rate = 1 - Num1 / Num2

[0107] Among them: Num1 represents the number of real anomalies after manual verification, and Num2 represents the number of anomaly warnings;

[0108] 13. If the false alarm rate does not reach the expected target (such as Rate > 0.05) within one cycle, the scheduler will restart the training service to start the next round of iterative training, otherwise it will not be restarted.

[0109] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for training and deploying a transformer anomaly detection model for an on-line monitoring system of transformer status, characterized in that Including steps: The trained transformer anomaly detection model is iterated and verified multiple times using the test set to obtain the corresponding number of output results. Post-process the output results of the transformer anomaly detection model to obtain the corresponding number of prediction results. Based on all the prediction results, a comprehensive judgment strategy is used to obtain the final prediction result. After the transformer anomaly detection model is verified, it is deployed and online inference is performed. This means that the model parameters are automatically loaded into the transformer system, and an inference service interface is provided to the outside world. Online inference is performed on the real-time collected transformer voiceprint data. If the transformer anomaly detection model detects an abnormal signal, an early warning message is issued and the early warning data is stored. After model deployment and online inference are completed, manual anomaly correction is performed on the warning data on a regular basis to determine whether the anomaly inferred by the transformer anomaly detection model is a real anomaly and whether the anomaly category is true. If the warning data does not match the actual situation, manual classification is performed and the annotated data is updated to the voiceprint database for subsequent iterative training of the transformer anomaly detection model. The system also includes a regular statistical step for early warning data, which automatically counts the false alarm rate of the transformer anomaly detection model within a preset period. When the false alarm rate reaches a preset indicator, it automatically triggers an automated training task, initializes the model parameters based on the previous transformer anomaly detection model, updates the parameters and training set, and performs iterative training. If the false alarm rate does not reach the expected target within a cycle, the training service is restarted through the scheduler to start the next round of iterative training. Otherwise, it will not be restarted.

2. A method for training and deploying a transformer anomaly detection model according to claim 1, characterized in that, The transformer anomaly detection model includes a detection model and a classification model, and the post-processing steps include: A voiceprint data is divided into several parts, preprocessed, and then input into the detection model and classification model for inference. The detection model and classification model each output an anomaly score. The two anomaly scores are weighted averaged to obtain a total anomaly score. The total anomaly score is compared with the preset benchmark score to obtain the prediction result; The total anomaly score is determined by the following formula: Score abnormal =(w1*Score det +w2*Score cls ) / (w1+w2) Among them, w1 and w2 are weight coefficients respectively, Score det is the anomaly score predicted by the detection model, Score cls is the anomaly score predicted by the classification model, Score cls = 1-Confidence normal , Confidence normal is the confidence score of the normal category predicted by the classification model.

3. A method for training and deploying a transformer anomaly detection model according to claim 2, characterized in that The comprehensive judgment strategy is: calculate N the variance of the total anomaly scores of the var group of prediction results ， u For N a Score abnormal mean value; Empirical-based benchmark variance var' , and the number of anomalies predicted in the prediction results of N sets of classification models n , determine the final prediction result according to the following formula: ， y represents the prediction result. If the final prediction result obtained according to the above is abnormal, that is, the class label is abnormal , the abnormal category takes the type with the largest number of categories in the N-group classification results. Otherwise, it is the normal category normal or the unknown category unknown.

4. A method for training and deploying a transformer anomaly detection model according to claim 1, characterized in that: When verifying the transformer anomaly detection model, if the preset performance indicators are not achieved within the preset maximum number of iterative training times, the anomaly detection model with the best performance is selected and proceed to the next step.

5. A method for training and deploying a transformer anomaly detection model according to claim 1, characterized in that The steps to train a Transformer anomaly detection model include: Acquire transformer voiceprint data including normal voiceprint data and abnormal voiceprint data as original data, wherein the abnormal voiceprint data comes from a voiceprint database; Preprocess the raw data; Based on the extracted features, a set of detection models and classification models are trained in sequence using audio detection and classification algorithms. The transformer anomaly detection model includes a feature extraction model, a detection model, and a classification model. The detection model is used to detect whether the transformer voiceprint data has anomalies, and the classification model is used to identify the type of transformer voiceprint data. After the transformer anomaly detection model training is completed, the test set is used to perform multiple iterations of the transformer anomaly detection model verification. If the performance reaches the preset indicators, it proceeds to the next step. Otherwise, it returns to the step of preprocessing the original data, and adjusts the transformer anomaly detection model hyperparameters and training data for retraining.

6. A device for training and deploying a transformer anomaly detection model, which is used for a transformer system, is characterized in that, include: Transformer anomaly detection model, used to detect and identify abnormal transformer voiceprint data and perform model training based on the system-updated voiceprint database; The verification module is used to perform multiple iterations of the trained transformer anomaly detection model using the test set to obtain a corresponding number of output results; The post-processing module processes the output results of the transformer anomaly detection model to obtain a corresponding number of prediction results. Based on all the prediction results, a comprehensive judgment strategy is used to obtain the final prediction result. The model deployment and online reasoning module is used to deploy the model and perform online reasoning after the transformer anomaly detection model is verified. This module automatically loads the model parameters into the transformer system, provides an external reasoning service interface, and performs online reasoning on the real-time collected transformer voiceprint data. If the transformer anomaly detection model detects an abnormal signal, it issues an early warning and stores the early warning data. The anomaly labeling module is used to manually perform anomaly correction on the warning data on a regular basis after model deployment and online inference are completed. This module determines whether the anomaly inferred by the transformer anomaly detection model is a real anomaly and whether the anomaly category is true. If the warning data does not match the actual situation, the category is manually labeled and the labeled data is updated to the voiceprint database for subsequent iterative training of the transformer anomaly detection model. The warning number periodic statistics module is used to automatically count the false alarm rate of the transformer anomaly detection model within a preset period. When the false alarm rate reaches the preset indicator, the automated training task is automatically triggered. The model parameters of the previous transformer anomaly detection model are initialized, the training set is updated, and iterative training is performed. If the false alarm rate does not reach the expected target within a period, the training service is restarted through the scheduler to start the next round of iterative training. Otherwise, it will not be restarted.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method according to any one of claims 1 to 5.

Images