Transformer fault diagnosis method based on voice recognition
Through the joint learning of multimodal data acquisition and deep learning models, combined with self-supervised learning, transfer learning and dynamic parameter adjustment, the problem of insufficient accuracy and adaptability of existing transformer fault diagnosis methods is solved, and more efficient and explainable fault diagnosis effects are achieved.
Patent Information
- Application Number
- CN202311617314.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-29
- Publication Date
- 2025-03-28
AI Technical Summary
The existing transformer fault diagnosis methods based on sound recognition have problems such as environmental noise influence, inability to fully tap sound, vibration and current correlation, poor transfer learning performance, and lack of real-time model parameter adjustment, resulting in inaccurate diagnostic results and insufficient adaptability.
Multimodal data acquisition and preprocessing are adopted to obtain the sound, vibration and current data of the transformer through sensors, and signal filtering, noise reduction and feature extraction are performed to generate a multimodal data set. Then, a multimodal deep neural network is used for joint learning, fusing different modal information, and optimizing the model to suit different working conditions and transformer types through self-supervised learning, transfer learning, and dynamic learning parameter adjustment. Finally, an interpretable machine learning method is introduced to explain and analyze the faults.
Through multimodal data fusion and deep learning model optimization, the accuracy and adaptability of fault diagnosis are improved, the ability to identify faults of different types of transformers is enhanced, and the cause of failure is provided is provided to help technicians better analyze and solve fault problems.
Smart Images
Figure CN119716289A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning and signal processing, and in particular to a transformer fault diagnosis method based on sound recognition. Background Art
[0002] The field of machine learning and signal processing involves processing signal data using computer algorithms, understanding, analyzing and processing signal data by training models or designing algorithms. This includes using machine learning techniques to extract, classify, identify or predict signals to solve practical problems, such as fault diagnosis, speech recognition, image processing, etc. The transformer fault diagnosis method based on sound recognition is a technology that uses sound signals to predict and diagnose transformer faults. Its main purpose is to detect potential fault characteristics and identify possible fault types by analyzing the sound signals generated by the transformer, so as to achieve early fault diagnosis and preventive maintenance. The purpose of this method is to achieve early detection and diagnosis of transformer faults to prevent potential equipment failures and improve the reliability and safety of equipment operation. By analyzing the characteristics in the sound signal, the fault type (such as partial discharge, aging of insulation materials, etc.) can be identified, and maintenance measures can be taken in time to avoid the expansion of the fault and cause equipment damage or shutdown.
[0003] The existing transformer fault diagnosis methods based on sound recognition have the following shortcomings: First, fault diagnosis based solely on sound signals may be affected by environmental noise, resulting in inaccurate diagnosis results. Second, the existing methods may not be able to fully explore the correlation between sound, vibration and current, and thus fail to fully capture fault characteristics. In addition, due to the lack of full consideration of the characteristics of different types of transformers, transfer learning methods may perform poorly on specific types of transformers. Finally, the existing methods lack a real-time adjustment mechanism for model parameters and cannot adapt to changes under different working conditions. In summary, the existing methods have certain limitations in terms of fault diagnosis accuracy, feature extraction capability, model generalization capability and adaptability. Summary of the invention
[0004] In view of the deficiencies in the prior art, the present invention provides a transformer fault diagnosis method based on sound recognition, which solves the problem that fault diagnosis based solely on sound signals may be affected by environmental noise, resulting in inaccurate diagnosis results.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a transformer fault diagnosis method based on sound recognition, characterized in that it includes the following steps:
[0006] S1: Multimodal data acquisition and preprocessing: Use sensors to obtain multimodal data such as sound, vibration and current of the transformer, and perform preprocessing, including signal filtering, noise reduction and feature extraction, to generate a multimodal data set;
[0007] S2: Multimodal fusion and joint learning: Based on multimodal datasets, a multimodal deep neural network is used for joint learning to fuse sound, vibration, and current information, optimize the model to more comprehensively capture fault features, and generate a multimodal fusion deep learning model;
[0008] S3: Self-supervised learning feature extraction, using self-supervised learning methods, automatically learning the features in the transformer sound signal through contrast learning, optimizing the feature extraction network, and generating a self-supervised learning optimized feature extraction model;
[0009] S4: Transfer learning adapts to different types of transformers. The model is trained on a large and diverse dataset, and then fine-tuned on a small dataset of a specific transformer type through transfer learning technology to generate a transfer learning model that adapts to different types of transformers.
[0010] S5: Dynamic learning parameter adjustment, monitoring the transformer operating status, adjusting the deep learning model parameters in real time to adapt to the changes in sound signal characteristics under different working conditions, and generating a deep learning model with dynamic learning parameter adjustment;
[0011] S6: Explainable machine learning fault cause analysis, introduces explainable machine learning methods to explain and analyze faults and generate an explainable analysis report of the fault causes.
[0012] Preferably, multimodal data acquisition and preprocessing uses sensors to obtain multimodal data such as sound, vibration and current of the transformer, and performs preprocessing, including signal filtering, noise reduction and feature extraction, to generate a multimodal data set in the following specific steps:
[0013] S101: sensor data acquisition, using sound, vibration and current sensors to obtain multimodal data of the transformer, using microphones, accelerometers and current sensors to collect sound, vibration and current signals of the transformer, and generating original sound, vibration and current data sets;
[0014] S102: Signal filtering and preprocessing: filtering and preprocessing the collected data to remove noise and outliers. Digital signal processing technology is applied to remove high-frequency noise and outliers to generate a clean data set after filtering and preprocessing.
[0015] S103: Feature extraction, extracting features from the processed data to represent key information of sound, vibration and current, using time-frequency analysis to extract sound frequency features, vibration amplitude features and current spectrum features, and generating sound, vibration and current feature data sets;
[0016] S104: Dataset fusion, integrating sound, vibration and current feature data to form a multimodal data set, integrating the sound, vibration and current feature data according to time series or feature vectors, constructing a multimodal data set, and generating a multimodal data set.
[0017] Preferably, multimodal fusion and joint learning, based on a multimodal data set, adopts a multimodal deep neural network for joint learning, fuses sound, vibration and current information, optimizes the model to more comprehensively capture fault characteristics, and generates a multimodal fusion deep learning model in the following specific steps:
[0018] S201: Multimodal deep neural network construction, design a multimodal deep neural network architecture to jointly learn multimodal data, establish a deep neural network model for jointly processing sound, vibration and current data, and generate a multimodal fusion deep learning model;
[0019] S202: Feature fusion and weight optimization: fuse the features of different modes in the neural network and optimize the weights to maximize the capture of fault features. Use a multi-modal fusion strategy, feature-level fusion or model-level fusion, and optimize the network weights through the back-propagation algorithm to generate an optimized multi-modal fusion deep learning model.
[0020] S203: Model verification and evaluation: verify and evaluate the fusion model to ensure its ability to capture fault features and verify and evaluate the performance of the multimodal fusion deep learning model;
[0021] S204: Model saving and deployment, saving the verified model and preparing to deploy it to the actual system for fault monitoring, saving the optimized model in a deployable format, preparing the resources required for model deployment, and saving the deployable multimodal fusion deep learning model.
[0022] Preferably, the self-supervised learning feature extraction uses a self-supervised learning method to automatically learn the features in the transformer sound signal through comparative learning, optimize the feature extraction network, and generate a self-supervised learning optimized feature extraction model. The specific steps are:
[0023] S301: self-supervised learning data construction, creating a comparative data set for self-supervised learning, used for training a feature extraction network, using the time series characteristics of the sound signal or the periodicity of the vibration signal to construct a comparative learning data set, forming positive and negative sample pairs, and generating a comparative data set for self-supervised learning;
[0024] S302: Contrastive learning training: using contrastive learning method to train feature extraction network, optimizing network parameters to better capture features in transformer sound signal, using contrastive loss function to train network to distinguish positive and negative sample pairs, and generating feature extraction model optimized by self-supervised learning;
[0025] S303: Feature extractor verification, verifying the effect of the feature extraction model optimized by self-supervised learning on feature extraction, using the verification data set to test the feature extractor, evaluate its performance in capturing transformer sound features, and verify the effectiveness of the feature extraction model optimized by self-supervised learning;
[0026] S304: Feature fusion and model integration: Integrate the feature extractor optimized by self-supervised learning into the deep neural network as one of its modules, and retrain the model to generate a multimodal fusion deep learning model that integrates the self-supervised learning optimized feature extraction model.
[0027] Preferably, transfer learning is adapted to different types of transformers, trained on large and diverse data sets, and then fine-tuned on a small data set of a specific transformer type through transfer learning technology. The specific steps of generating a transfer learning model adapted to different types of transformers are:
[0028] S401: Dataset preparation, filtering sample data related to a specific transformer type from a large and diverse dataset to generate a small dataset of a specific transformer type;
[0029] S402: Transfer learning model selection: Based on the pre-trained model, select a suitable model structure, and fine-tune it according to the characteristics of the small data set to generate a transfer learning model that is suitable for different types of transformers;
[0030] S403: fine-tuning training, using a small data set to train the selected transfer learning model, adjusting the model parameters through a gradient descent algorithm to improve the performance on a specific transformer type, and generating a fine-tuned transfer learning model;
[0031] S404: Performance evaluation, using a test set of a small dataset to evaluate the performance of the transfer learning model on a specific transformer type, and generating a performance evaluation report to confirm whether the model has been successfully adapted to different types of transformers.
[0032] Preferably, the dynamic learning parameter adjustment, monitoring the transformer operating status, and real-time adjustment of the deep learning model parameters to adapt to the changes in sound signal characteristics under different working conditions, the specific steps of generating a deep learning model with dynamic learning parameter adjustment are:
[0033] S501: Real-time data monitoring, collecting sound signal data of the real-time operating status of the transformer and generating real-time monitoring data;
[0034] S502: adjusting parameters of the deep learning model, based on the real-time monitoring data, using the back propagation algorithm to adjust the parameters of the deep learning model in real time to adapt to the changes in the characteristics of the sound signal under different working conditions, and generating a deep learning model that has been adjusted in real time;
[0035] S503: Performance evaluation, by comparing with actual operation data, evaluating the performance of the dynamically adjusted model under different working conditions, and generating a performance evaluation report to confirm the real-time adjustment effect of the model;
[0036] S504: Feedback mechanism: Establish a feedback mechanism to feed back real-time monitoring data and performance evaluation results to the model to further optimize the parameter adjustment strategy and generate feedback results for dynamic optimization of model parameters.
[0037] Preferably, the explainable machine learning fault cause analysis introduces the explainable machine learning method to interpret and analyze the fault, and the specific steps of generating an explainable analysis report of the fault cause are:
[0038] S601: Introduce the attention mechanism. Introduce the attention mechanism into the deep learning model so that the model can focus on the key information in the input data and generate a deep learning model with the attention mechanism;
[0039] S602: Fault sample interpretation and analysis: Use a model with an attention mechanism to interpret and analyze the fault samples, identify the sound signal features that have a significant impact on the fault, and generate a fault sample interpretation and analysis report;
[0040] S603: Generate an interpretability report, combining the fault sample interpretation analysis report to generate an overall interpretability analysis report, including explanations for different fault causes, to generate an interpretability analysis report;
[0041] S604: Feedback to the system: Feedback the explainability analysis report to the monitoring system to improve the system's accuracy in identifying different fault causes and generate feedback results for improving system performance.
[0042] The present invention provides a transformer fault diagnosis method based on sound recognition, which has the following beneficial effects:
[0043] The present invention can obtain multimodal data such as sound, vibration and current of the transformer through multimodal data acquisition and preprocessing, and filter, reduce noise and extract features to generate a multimodal data set. In this way, the fault characteristics can be captured more comprehensively, and the accuracy and reliability of fault diagnosis are improved. In multimodal fusion and joint learning, a multimodal deep neural network is used for joint learning, and the sound, vibration and current information are integrated together to optimize the model to capture the fault characteristics more comprehensively. In this way, the learning ability of the model can be improved, and the effect of fault diagnosis can be further improved. The self-supervised learning feature extraction uses the self-supervised learning method to automatically learn the features in the transformer sound signal through comparative learning, and optimize the feature extraction network. In this way, the effect of feature extraction can be improved, and the fault diagnosis ability can be further enhanced. Transfer learning is adapted to different types of transformers, and training can be performed on large and diverse data sets. Then, the model can be fine-tuned on a small data set of a specific transformer type through transfer learning technology. In this way, the generalization ability of the model can be improved, so that it can adapt to the fault diagnosis of different types of transformers. Dynamic learning parameters are adjusted to monitor the operating status of the transformer, and the parameters of the deep learning model are adjusted in real time to adapt to the changes in the sound signal characteristics under different working conditions. This can improve the adaptability and flexibility of the model and further enhance the fault diagnosis effect. Interpretable machine learning fault cause analysis introduces interpretable machine learning methods to interpret and analyze faults and generate interpretable analysis reports of fault causes. This can provide an in-depth understanding of the fault causes and help technicians better analyze and solve fault problems. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 It is a schematic diagram of the main steps of the present invention;
[0045] Figure 2 It is a schematic diagram of the refinement of S1 of the present invention;
[0046] Figure 3 It is a schematic diagram of the refinement of S2 of the present invention;
[0047] Figure 4 It is a schematic diagram of the refinement of S3 of the present invention;
[0048] Figure 5 It is a schematic diagram of the refinement of S4 of the present invention;
[0049] Figure 6 It is a detailed schematic diagram of S5 of the present invention;
[0050] Figure 7 It is a detailed schematic diagram of S6 of the present invention. DETAILED DESCRIPTION
[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0052] Example:
[0053] like Figure 1-7 As shown, an embodiment of the present invention provides a transformer fault diagnosis method based on sound recognition, which is characterized by comprising the following steps:
[0054] S1: Multimodal data acquisition and preprocessing: Use sensors to obtain multimodal data such as sound, vibration and current of the transformer, and perform preprocessing, including signal filtering, noise reduction and feature extraction, to generate a multimodal data set. Through multimodal data acquisition and preprocessing, multimodal data such as sound, vibration and current of the transformer can be obtained, and filtered, noise reduced and feature extracted to generate a multimodal data set. This can capture fault characteristics more comprehensively and improve the accuracy and reliability of fault diagnosis;
[0055] S2: Multimodal fusion and joint learning: Based on multimodal data sets, multimodal deep neural networks are used for joint learning to fuse sound, vibration and current information, optimize the model to more comprehensively capture fault features, and generate a multimodal fusion deep learning model. By using multimodal deep neural networks for joint learning, sound, vibration and current information are fused together, and the model is optimized to more comprehensively capture fault features. This can improve the learning ability of the model and further enhance the effect of fault diagnosis.
[0056] S3: Self-supervised learning feature extraction: using self-supervised learning methods, automatically learning the features in the transformer sound signal through contrast learning, optimizing the feature extraction network, generating a self-supervised learning optimized feature extraction model, automatically learning the features in the transformer sound signal through contrast learning, optimizing the feature extraction network. This can improve the effect of feature extraction and further enhance fault diagnosis capabilities;
[0057] S4: Transfer learning is adapted to different types of transformers. Training is performed on large and diverse datasets. Then, the model is fine-tuned on a small dataset of a specific transformer type through transfer learning technology to generate a transfer learning model adapted to different types of transformers. This can improve the generalization ability of the model and make it adaptable to fault diagnosis of different types of transformers.
[0058] S5: Dynamic learning parameter adjustment, monitoring the transformer operating status, adjusting the deep learning model parameters in real time to adapt to the changes in sound signal characteristics under different working conditions, and generating a deep learning model with dynamic learning parameter adjustment, which can improve the adaptability and flexibility of the model and further enhance the fault diagnosis effect;
[0059] S6: Explainable machine learning fault cause analysis, introduces explainable machine learning methods to explain and analyze faults and generate explainable analysis reports of fault causes. This can provide an in-depth understanding of the causes of faults and help technicians better analyze and solve fault problems.
[0060] Multimodal data acquisition and preprocessing: Use sensors to obtain multimodal data such as sound, vibration and current of the transformer, and perform preprocessing, including signal filtering, noise reduction and feature extraction. The specific steps to generate a multimodal data set are as follows:
[0061] S101: sensor data acquisition, using sound, vibration and current sensors to obtain multimodal data of the transformer, using microphones, accelerometers and current sensors to collect sound, vibration and current signals of the transformer, and generating original sound, vibration and current data sets;
[0062] S102: Signal filtering and preprocessing: filtering and preprocessing the collected data to remove noise and outliers. Digital signal processing technology is applied to remove high-frequency noise and outliers to generate a clean data set after filtering and preprocessing.
[0063] S103: Feature extraction, extracting features from the processed data to represent key information of sound, vibration and current, using time-frequency analysis to extract sound frequency features, vibration amplitude features and current spectrum features, and generating sound, vibration and current feature data sets;
[0064] S104: Dataset fusion, integrating sound, vibration and current feature data to form a multimodal data set, integrating the sound, vibration and current feature data according to time series or feature vectors, constructing a multimodal data set, and generating a multimodal data set.
[0065] Multimodal fusion and joint learning, based on multimodal data sets, adopts multimodal deep neural network for joint learning, integrates sound, vibration and current information, optimizes the model to capture fault characteristics more comprehensively, and generates the specific steps of multimodal fusion deep learning model as follows:
[0066] S201: Multimodal deep neural network construction, design a multimodal deep neural network architecture to jointly learn multimodal data, establish a deep neural network model for jointly processing sound, vibration and current data, and generate a multimodal fusion deep learning model;
[0067] S202: Feature fusion and weight optimization: fuse the features of different modes in the neural network and optimize the weights to maximize the capture of fault features. Use a multi-modal fusion strategy, feature-level fusion or model-level fusion, and optimize the network weights through the back-propagation algorithm to generate an optimized multi-modal fusion deep learning model.
[0068] S203: Model verification and evaluation: verify and evaluate the fusion model to ensure its ability to capture fault features and verify and evaluate the performance of the multimodal fusion deep learning model;
[0069] S204: Model saving and deployment, saving the verified model and preparing to deploy it to the actual system for fault monitoring, saving the optimized model in a deployable format, preparing the resources required for model deployment, and saving the deployable multimodal fusion deep learning model.
[0070] Self-supervised learning feature extraction, using the self-supervised learning method, automatically learn the features of the transformer sound signal through comparative learning, optimize the feature extraction network, and generate the self-supervised learning optimized feature extraction model. The specific steps are:
[0071] S301: self-supervised learning data construction, creating a comparative data set for self-supervised learning, used for training a feature extraction network, using the time series characteristics of the sound signal or the periodicity of the vibration signal to construct a comparative learning data set, forming positive and negative sample pairs, and generating a comparative data set for self-supervised learning;
[0072] S302: Contrastive learning training: using contrastive learning method to train feature extraction network, optimizing network parameters to better capture features in transformer sound signal, using contrastive loss function to train network to distinguish positive and negative sample pairs, and generating feature extraction model optimized by self-supervised learning;
[0073] S303: Feature extractor verification, verifying the effect of the feature extraction model optimized by self-supervised learning on feature extraction, using the verification data set to test the feature extractor, evaluate its performance in capturing transformer sound features, and verify the effectiveness of the feature extraction model optimized by self-supervised learning;
[0074] S304: Feature fusion and model integration: Integrate the feature extractor optimized by self-supervised learning into the deep neural network as one of its modules, and retrain the model to generate a multimodal fusion deep learning model that integrates the self-supervised learning optimized feature extraction model.
[0075] Transfer learning adapts to different types of transformers, trains on large and diverse data sets, and then fine-tunes the model on a small data set of a specific transformer type through transfer learning technology. The specific steps to generate a transfer learning model that adapts to different types of transformers are:
[0076] S401: Dataset preparation, filtering sample data related to a specific transformer type from a large and diverse dataset to generate a small dataset of a specific transformer type;
[0077] S402: Transfer learning model selection: Based on pre-trained models, such as BERT, ResNet, etc., select a suitable model structure, and fine-tune it according to the characteristics of the small data set to generate a transfer learning model that is suitable for different types of transformers;
[0078] S403: fine-tuning training, using a small data set to train the selected transfer learning model, adjusting the model parameters through a gradient descent algorithm to improve the performance on a specific transformer type, and generating a fine-tuned transfer learning model;
[0079] S404: Performance evaluation, using a test set of a small data set to evaluate the performance of the transfer learning model on a specific transformer type, including indicators such as accuracy and recall, and generate a performance evaluation report to confirm whether the model has successfully adapted to different types of transformers.
[0080] Dynamic learning parameter adjustment, monitoring the transformer operating status, and real-time adjustment of deep learning model parameters to adapt to the changes in sound signal characteristics under different working conditions. The specific steps for generating a deep learning model with dynamic learning parameter adjustment are:
[0081] S501: Real-time data monitoring, collecting sound signal data of the real-time operating status of the transformer and generating real-time monitoring data;
[0082] S502: adjusting parameters of the deep learning model, based on the real-time monitoring data, using the back propagation algorithm to adjust the parameters of the deep learning model in real time to adapt to the changes in the characteristics of the sound signal under different working conditions, and generating a deep learning model that has been adjusted in real time;
[0083] S503: Performance evaluation, by comparing with actual operation data, evaluating the performance of the dynamically adjusted model under different working conditions, and generating a performance evaluation report to confirm the real-time adjustment effect of the model;
[0084] S504: Feedback mechanism: Establish a feedback mechanism to feed back real-time monitoring data and performance evaluation results to the model to further optimize the parameter adjustment strategy and generate feedback results for dynamic optimization of model parameters.
[0085] Explainable machine learning fault cause analysis introduces explainable machine learning methods to explain and analyze faults. The specific steps to generate an explainable analysis report of the fault cause are as follows:
[0086] S601: Introduce the attention mechanism. Introduce the attention mechanism into the deep learning model so that the model can focus on the key information in the input data and generate a deep learning model with the attention mechanism;
[0087] S602: Fault sample interpretation and analysis: Use a model with an attention mechanism to interpret and analyze the fault samples, identify the sound signal features that have a significant impact on the fault, and generate a fault sample interpretation and analysis report;
[0088] S603: Generate an interpretability report, combining the fault sample interpretation analysis report to generate an overall interpretability analysis report, including explanations for different fault causes, to generate an interpretability analysis report;
[0089] S604: Feedback to the system: Feedback the explainability analysis report to the monitoring system to improve the system's accuracy in identifying different fault causes and generate feedback results for improving system performance.
[0090] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A transformer fault diagnosis method based on sound recognition, characterized in that The following steps are involved: S1: Acquire the transformer sound, vibration and current multimodal data, perform preprocessing, and generate a multimodal data set; S2: Based on the multimodal dataset, a multimodal deep neural network is used for joint learning to fuse the sound, vibration and current information to generate a multimodal fusion deep learning model; S3: Using self-supervised learning methods, the features in the transformer sound signal are automatically learned through contrastive learning methods to generate a feature extraction model optimized by self-supervised learning; S4: Train on large and diverse datasets, and then use transfer learning techniques to generate transfer learning models that adapt to different types of transformers; S5: Monitor the operating status of the transformer and adjust the deep learning model parameters in real time; S6: Introduce interpretable machine learning methods to explain and analyze faults.
2. A transformer fault diagnosis method based on sound recognition according to claim 1, characterized in that: Step S1 includes: S101: Collect sound, vibration and current signals of the transformer using a microphone, an accelerometer and a current sensor; S102: filtering and preprocessing the collected signal; S103: extracting features from the processed data to represent key information of sound, vibration, and current; S104: Integrate the sound, vibration and current feature data according to time series or feature vectors to generate a multimodal data set.
3. A transformer fault diagnosis method based on sound recognition according to claim 1, characterized in that: Step S2 includes: S201: Build a multimodal deep neural network architecture to jointly process sound, vibration, and current data; S202: adopt a multimodal fusion strategy and optimize the network weights through a back-propagation algorithm to generate a multimodal fusion deep learning model; S203: Verify and evaluate the performance of the multimodal fusion deep learning model; S204: Prepare resources required for model deployment and save the deployable multimodal fusion deep learning model.
4. A transformer fault diagnosis method based on sound recognition according to claim 1, characterized in that: Step S3 includes: S301: constructing a data set for contrastive learning by using the time series characteristics of the sound signal or the periodicity of the vibration signal; forming positive and negative sample pairs, and generating a contrastive data set for self-supervised learning; S302: Train the feature extraction network using a contrastive learning method to generate a feature extraction model optimized by self-supervised learning; S303: Verify the effectiveness of the feature extraction model optimized by self-supervised learning; S304: Integrate the feature extractor optimized by self-supervised learning into the deep neural network to generate a multimodal fusion deep learning model that integrates the self-supervised learning optimized feature extraction model.
5. The transformer fault diagnosis method based on sound recognition according to claim 1 is characterized in that: Step S4 includes: S401: Filtering sample data related to a specific transformer type from a large and diverse data set to generate a small data set of the specific transformer type; S402: Based on the pre-trained model, fine-tune according to the characteristics of the small data set to generate a transfer learning model that is suitable for different types of transformers; S403: training the selected transfer learning model using the small data set to generate a fine-tuned transfer learning model; S404: Evaluate the performance of the transfer learning model on a specific transformer type using a test set of a small dataset.
6. A transformer fault diagnosis method based on sound recognition according to claim 1, characterized in that: Step S5 includes: S501: Collecting sound signal data of the real-time operating status of the transformer to generate real-time monitoring data; S502: Based on the real-time monitoring data, use a back propagation algorithm to adjust the parameters of the deep learning model in real time to generate a deep learning model that has been adjusted in real time; S503: Confirming the real-time adjustment effect of the model by comparing with the actual operation data; S504: Feedback the real-time monitoring data and performance evaluation results to the model to optimize the parameter adjustment strategy.
7. The transformer fault diagnosis method based on sound recognition according to claim 1 is characterized in that: Step S6 includes: S601: Introduce an attention mechanism into a deep learning model to generate a deep learning model with an attention mechanism; S602: Use a model with an attention mechanism to interpret and analyze the fault sample, identify the sound signal features that have a significant impact on the fault, and generate a fault sample interpretation and analysis report; S603: Generate an overall explainability analysis report based on the fault sample explanation analysis report, including explanations of different fault causes; S604: Feedback the explainability analysis report to the monitoring system.