Transformer fault diagnosis method and device based on sound and vibration fusion and self-supervised learning, computer equipment and storage medium

Abstract

The invention relates to a transformer fault diagnosis method and device based on sound and vibration fusion and self-supervised learning, computer equipment and a storage medium. The method comprises the steps of obtaining a current operation state of a to-be-analyzed transformer; acquiring a vibration signal and a sound signal of the to-be-analyzed transformer under the condition that the current operation state shows that the to-be-analyzed transformer is in normal operation; respectively performing feature extraction processing on the vibration signal and the sound signal through a trained feature extraction model to obtain a vibration feature vector corresponding to the vibration signal and a sound feature vector corresponding to the sound signal; performing fusion processing on the vibration feature vector and the sound feature vector to obtain a fusion feature vector of the to-be-analyzed transformer; and inputting the fusion feature vector into a trained transformer fault diagnosis model to obtain a fault diagnosis result of the to-be-analyzed transformer. By adopting the method, the accuracy of transformer fault diagnosis can be improved.

Description

Technical Field

[0001] This application relates to the field of power grid technology, and in particular to a transformer fault diagnosis method, device, computer equipment, computer-readable storage medium, and computer program product based on acoustic-vibration fusion and self-supervised learning. Background Technology

[0002] Currently, in order to ensure the safe and stable operation of the power grid, accurate transformer fault diagnosis is of paramount importance.

[0003] In traditional technology, single-mode signal analysis is generally used for transformer fault diagnosis; however, this method has weak anti-interference ability and is difficult to capture weak signs of early faults, resulting in low accuracy of transformer fault diagnosis. Summary of the Invention

[0004] Therefore, it is necessary to provide a transformer fault diagnosis method, device, computer equipment, computer-readable storage medium, and computer program product based on acoustic-vibration fusion and self-supervised learning that can improve the accuracy of transformer fault diagnosis, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning, including:

[0006] Obtain the current operating status of the transformer to be analyzed;

[0007] When the current operating state indicates that the transformer under analysis is operating normally, the vibration signal and sound signal of the transformer under analysis are acquired.

[0008] The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, so as to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal.

[0009] The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed;

[0010] The fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0011] In one embodiment, the trained feature extraction model is obtained by training in the following manner:

[0012] Acquire the first sample vibration signal and the first sample sound signal of the first sample transformer;

[0013] Data augmentation processing is performed on the first sample vibration signal and the first sample sound signal respectively to obtain the enhanced sample vibration signal and the enhanced sample sound signal;

[0014] Based on the enhanced sample vibration signal and the enhanced sample sound signal, positive sample pairs and negative sample pairs are constructed; the positive sample pair includes a first enhanced sample vibration signal and a first enhanced sample sound signal from the enhanced sample vibration signal; the first enhanced sample vibration signal and the first enhanced sample sound signal correspond to the same time point of the same transformer; the negative sample pair includes a second enhanced sample vibration signal and a second enhanced sample sound signal from the enhanced sample vibration signal; the second enhanced sample vibration signal and the second enhanced sample sound signal correspond to different transformers, or to different time points of the same transformer;

[0015] Based on the positive sample pairs and the negative sample pairs, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

[0016] In one embodiment, the step of iteratively training the feature extraction model to be trained based on the positive sample pairs and the negative sample pairs to obtain the trained feature extraction model includes:

[0017] The feature extraction model to be trained is used to perform feature extraction processing on the first enhanced sample vibration signal, the first enhanced sample sound signal, the second enhanced sample vibration signal, and the second enhanced sample sound signal, respectively, to obtain the first vibration feature vector corresponding to the first enhanced sample vibration signal, the first sound feature vector corresponding to the first enhanced sample sound signal, the second vibration feature vector corresponding to the second enhanced sample vibration signal, and the second sound feature vector corresponding to the second enhanced sample sound signal.

[0018] Based on the first vibration feature vector and the first sound feature vector, a first similarity between the first enhanced sample vibration signal and the first enhanced sample sound signal is determined, and based on the second vibration feature vector and the second sound feature vector, a second similarity between the second enhanced sample vibration signal and the second enhanced sample sound signal is determined.

[0019] Based on the first similarity and the second similarity, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

[0020] In one embodiment, the step of iteratively training the feature extraction model to be trained based on the first similarity and the second similarity to obtain the trained feature extraction model includes:

[0021] A first loss value is obtained based on the difference between the first similarity and the first preset similarity, and a second loss value is obtained based on the difference between the second similarity and the second preset similarity.

[0022] The first loss value and the second loss value are fused together to obtain the target loss value;

[0023] Based on the target loss value, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

[0024] In one embodiment, the trained transformer fault diagnosis model is obtained by training in the following manner:

[0025] Acquire the second sample vibration signal and the second sample sound signal of the second sample transformer;

[0026] The trained feature extraction model is used to perform feature extraction processing on the second sample vibration signal and the second sample sound signal respectively to obtain the sample vibration feature vector corresponding to the second sample vibration signal and the sample sound feature vector corresponding to the second sample sound signal.

[0027] The sample vibration feature vector and the sample sound feature vector are fused to obtain the fused sample feature vector of the second sample transformer;

[0028] The feature vectors of the fused samples are split to obtain a training dataset and a validation dataset;

[0029] Based on the training dataset, the transformer fault diagnosis model to be trained is fine-tuned to obtain the fine-tuned transformer fault diagnosis model.

[0030] The fine-tuned transformer fault diagnosis model is validated based on the validation dataset to obtain the validation results of the fine-tuned transformer fault diagnosis model.

[0031] If the verification results meet the preset verification conditions, the trained transformer fault diagnosis model is obtained based on the fine-tuned and trained transformer fault diagnosis model.

[0032] In one embodiment, the trained feature extraction model performs feature extraction processing on the vibration signal and the sound signal respectively to obtain a vibration feature vector corresponding to the vibration signal and a sound feature vector corresponding to the sound signal, including:

[0033] The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, to obtain the first time-frequency diagram, first time-domain feature, first frequency-domain feature and first time-frequency-domain feature corresponding to the vibration signal, and the second time-frequency diagram, second time-domain feature, second frequency-domain feature and second time-frequency-domain feature corresponding to the sound signal.

[0034] Based on the first time-frequency diagram, a first feature vector corresponding to the vibration signal is constructed; based on the first time-domain feature, the first frequency-domain feature, and the first time-frequency-domain feature, a second feature vector corresponding to the vibration signal is constructed; based on the second time-frequency diagram, a third feature vector corresponding to the sound signal is constructed; and based on the second time-domain feature, the second frequency-domain feature, and the second time-frequency-domain feature, a fourth feature vector corresponding to the sound signal is constructed.

[0035] Based on the first feature vector and the second feature vector, a vibration feature vector corresponding to the vibration signal is obtained, and based on the third feature vector and the fourth feature vector, a sound feature vector corresponding to the sound signal is obtained.

[0036] Secondly, this application also provides a transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning, comprising:

[0037] The status acquisition module is used to acquire the current operating status of the transformer to be analyzed.

[0038] The signal acquisition module is used to acquire the vibration signal and sound signal of the transformer to be analyzed when the current operating state indicates that the transformer to be analyzed is operating normally;

[0039] The feature extraction module is used to perform feature extraction processing on the vibration signal and the sound signal respectively through the trained feature extraction model to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal;

[0040] The feature fusion module is used to fuse the vibration feature vector and the sound feature vector to obtain the fused feature vector of the transformer to be analyzed.

[0041] The fault diagnosis module is used to input the fused feature vector into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0043] Obtain the current operating status of the transformer to be analyzed;

[0044] When the current operating state indicates that the transformer under analysis is operating normally, the vibration signal and sound signal of the transformer under analysis are acquired.

[0045] The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, so as to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal.

[0046] The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed;

[0047] The fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0049] Obtain the current operating status of the transformer to be analyzed;

[0050] When the current operating state indicates that the transformer under analysis is operating normally, the vibration signal and sound signal of the transformer under analysis are acquired.

[0051] The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, so as to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal.

[0052] The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed;

[0053] The fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0054] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0055] Obtain the current operating status of the transformer to be analyzed;

[0056] When the current operating state indicates that the transformer under analysis is operating normally, the vibration signal and sound signal of the transformer under analysis are acquired.

[0057] The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, so as to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal.

[0058] The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed;

[0059] The fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0060] The aforementioned transformer fault diagnosis method, device, computer equipment, storage medium, and computer program product based on acoustic-vibration fusion and self-supervised learning first acquires the current operating state of the transformer to be analyzed. Given that the current operating state indicates the transformer is operating normally, vibration and sound signals of the transformer are acquired. Then, through a trained feature extraction model, feature extraction processing is performed on the vibration and sound signals respectively, yielding vibration feature vectors corresponding to the vibration signals and sound feature vectors corresponding to the sound signals. Next, the vibration and sound feature vectors are fused to obtain a fused feature vector of the transformer to be analyzed. Finally, the fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed. In this way, when diagnosing transformer faults, by focusing on the acquisition of vibration and sound dual-mode signals under normal transformer operation, and relying on the trained feature extraction model, high-dimensional feature vectors of the two types of signals are accurately extracted. Then, through feature fusion processing, the complementary information from both sources is integrated. Based on the trained transformer fault diagnosis model, the more comprehensive and robust representation capability of dual-mode fusion features significantly improves the identification accuracy of early weak faults and complex faults, which is conducive to improving the accuracy of transformer fault diagnosis. Moreover, the entire process adopts a combined analysis method of vibration and sound signals, avoiding the shortcomings of traditional single-mode signal analysis, which has weak anti-interference ability and difficulty in capturing weak signs of early faults, resulting in low accuracy of transformer fault diagnosis. This further improves the accuracy of transformer fault diagnosis. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart illustrating a transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning in one embodiment.

[0063] Figure 2 This is a flowchart illustrating a transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning in another embodiment.

[0064] Figure 3 This is a flowchart illustrating an early mechanical condition diagnosis system for transformers based on acoustic-vibration fusion and self-supervised learning in one embodiment.

[0065] Figure 4 This is a flowchart illustrating a self-supervised pre-trained network structure for acoustic-vibration fusion based on contrastive learning in one embodiment.

[0066] Figure 5 This is a flowchart illustrating the downstream fine-tuning diagnostic module in one embodiment;

[0067] Figure 6 This is a flowchart illustrating the offline preparation steps in one embodiment;

[0068] Figure 7 This is a flowchart illustrating the online diagnosis and early warning process in one embodiment;

[0069] Figure 8 This is a schematic diagram of the time-domain waveform of the vibration signal under the condition of a loose iron core in one embodiment;

[0070] Figure 9 This is a schematic diagram of the time-domain waveform of the vibration signal under winding deformation in one embodiment;

[0071] Figure 10 This is a schematic diagram of the time-domain waveform of the vibration signal under clamping failure in one embodiment;

[0072] Figure 11 This is a schematic diagram of the vibration signal spectrum under the condition of a loose iron core in one embodiment;

[0073] Figure 12 This is a schematic diagram of the vibration signal spectrum under winding deformation in one embodiment;

[0074] Figure 13This is a schematic diagram of the vibration signal spectrum under clamping failure in one embodiment;

[0075] Figure 14 This is a schematic diagram of the time-frequency graph of the vibration signal under the condition of a loose iron core in one embodiment;

[0076] Figure 15 This is a schematic diagram of the time-frequency graph of the vibration signal under winding deformation in one embodiment;

[0077] Figure 16 This is a schematic diagram of the time-frequency graph of the vibration signal under clamping failure in one embodiment;

[0078] Figure 17 This is a structural block diagram of a transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning in one embodiment;

[0079] Figure 18 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0082] In one exemplary embodiment, such as Figure 1 As shown, a transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning is provided. This embodiment illustrates the application of this method to a server; it is understood that this method can also be applied to a terminal, or to a system including a terminal and a server, and is implemented through interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, and tablets; the server can be a standalone server or a server cluster composed of multiple servers. In this embodiment, the method includes the following steps:

[0083] Step S101: Obtain the current operating status of the transformer to be analyzed.

[0084] Among them, the transformer to be analyzed refers to the power transformer that needs to be analyzed.

[0085] The current operating status is used to represent the real-time operating status of the transformer to be analyzed at the current moment, including normal operation, abnormal operation (including fault precursors, minor faults, and serious faults), shutdown, and maintenance.

[0086] For example, the server collects gas data from the oil tank of the transformer under analysis using a gas sensor built into the tank. Then, the gas data is input into an importance prediction model to obtain the predicted importance of the gas data. Gas data with a predicted importance greater than a preset importance are selected as key gas data for the transformer under analysis. Next, feature extraction processing is performed on the key gas data to obtain its feature vector. This feature vector is then input into a trained operating state prediction model to obtain the first predicted operating state of the transformer under analysis, output by the trained model. Next, based on the key gas data, a preset correspondence between gas data and operating states is queried to obtain the second predicted operating state of the transformer under analysis. Finally, based on the first and second predicted operating states, the current operating state of the transformer under analysis is obtained.

[0087] Step S102: Under the condition that the current operating status indicates that the transformer to be analyzed is operating normally, acquire the vibration signal and sound signal of the transformer to be analyzed.

[0088] Among them, vibration signal refers to time-domain signal that reflects the vibration of the internal mechanical structure (such as iron core, winding, clamping parts) of the transformer to be analyzed during operation.

[0089] Among them, sound signal refers to the acoustic signal radiated to the outside when the transformer to be analyzed is running.

[0090] For example, the server determines the current operating status; if the current operating status indicates that the transformer to be analyzed is operating normally, the server collects the vibration signal of the transformer to be analyzed through a high-sensitivity vibration acceleration sensor (installed on key measuring points such as the transformer tank wall and core clamps) associated with the transformer to be analyzed, and collects the sound signal of the transformer to be analyzed through a high-fidelity directional microphone array associated with the transformer to be analyzed.

[0091] Furthermore, if the current operating status indicates that the transformer to be analyzed is operating abnormally, the server generates alarm information corresponding to the transformer to be analyzed and sends the alarm information to the target terminal corresponding to the transformer to be analyzed.

[0092] Step S103: Using the trained feature extraction model, feature extraction is performed on the vibration signal and the sound signal respectively to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal.

[0093] Feature extraction models, also known as feature extractors, are network models capable of extracting features from vibration and sound signals. These include vibration feature extraction networks and sound feature extraction networks. Vibration feature extraction networks can be CNNs (Convolutional Neural Networks) or Transformers; sound feature extraction networks can be ResNets (Residual Neural Networks).

[0094] Among them, the vibration feature vector refers to the feature vector corresponding to the vibration signal.

[0095] Among them, the sound feature vector refers to the feature vector corresponding to the sound signal.

[0096] For example, the server inputs vibration signals and sound signals into the trained feature extraction model, respectively; the vibration feature extraction network in the trained feature extraction model performs feature extraction processing on the vibration signal to obtain the vibration feature vector corresponding to the vibration signal, and the sound feature extraction network in the trained feature extraction model performs feature extraction processing on the sound signal to obtain the sound feature vector corresponding to the sound signal.

[0097] Step S104: The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed.

[0098] Among them, the fused feature vector refers to the feature vector obtained by fusing the vibration feature vector and the sound feature vector.

[0099] For example, the server inputs the vibration feature vector and the sound feature vector into the attention mechanism model respectively to obtain the fusion weights corresponding to the vibration feature vector and the sound feature vector. Then, the fusion weights corresponding to the vibration feature vector and the sound feature vector are normalized to obtain the processed fusion weights corresponding to the vibration feature vector and the sound feature vector. Then, the vibration feature vector and the sound feature vector are fused according to the processed fusion weights corresponding to the vibration feature vector and the sound feature vector to obtain the fused feature vector of the transformer to be analyzed.

[0100] Step S105: Input the fused feature vector into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0101] Among them, the transformer fault diagnosis model, also known as the multilayer perceptron classifier, refers to a network model that can use the fused feature vector of the transformer to be analyzed to obtain the fault diagnosis result of the transformer to be analyzed, such as the MLP (Multi-Layer Perceptron) model.

[0102] Among them, the fault diagnosis results are used to represent the structured conclusions of the early diagnosis of the mechanical condition of the transformer to be analyzed, including normal, loose core, winding deformation, failure of clamping components, etc.

[0103] For example, the server inputs the fused feature vector into the trained transformer fault diagnosis model to obtain the confidence level (i.e. prediction probability) of the transformer to be analyzed under each preset fault diagnosis result; then, from each preset fault diagnosis result, the preset fault diagnosis result with the highest confidence level is selected as the fault diagnosis result of the transformer to be analyzed.

[0104] In the above-mentioned transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning, the current operating state of the transformer to be analyzed is first obtained. Under the condition that the current operating state indicates that the transformer to be analyzed is operating normally, the vibration signal and sound signal of the transformer to be analyzed are obtained. Then, through the trained feature extraction model, the vibration signal and sound signal are processed to extract features respectively, and the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal are obtained. Next, the vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed. Finally, the fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed. In this way, when diagnosing transformer faults, by focusing on the acquisition of vibration and sound dual-mode signals under normal transformer operation, and relying on the trained feature extraction model, high-dimensional feature vectors of the two types of signals are accurately extracted. Then, through feature fusion processing, the complementary information from both sources is integrated. Based on the trained transformer fault diagnosis model, the more comprehensive and robust representation capability of dual-mode fusion features significantly improves the identification accuracy of early weak faults and complex faults, which is conducive to improving the accuracy of transformer fault diagnosis. Moreover, the entire process adopts a combined analysis method of vibration and sound signals, avoiding the shortcomings of traditional single-mode signal analysis, which has weak anti-interference ability and difficulty in capturing weak signs of early faults, resulting in low accuracy of transformer fault diagnosis. This further improves the accuracy of transformer fault diagnosis.

[0105] In an exemplary embodiment, the transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning provided in this application further includes a training process for the trained feature extraction model, specifically including the following: acquiring a first sample vibration signal and a first sample sound signal of a first sample transformer; performing data augmentation processing on the first sample vibration signal and the first sample sound signal respectively to obtain augmented sample vibration signal and augmented sample sound signal; constructing positive sample pairs and negative sample pairs based on the augmented sample vibration signal and the augmented sample sound signal; and iteratively training the feature extraction model to be trained based on the positive sample pairs and the negative sample pairs to obtain the trained feature extraction model.

[0106] The first sample transformer refers to the sample transformer used to train the feature extraction model to be trained.

[0107] The first sample vibration signal refers to the vibration signal of the first sample transformer.

[0108] The first sample audio signal refers to the audio signal of the first sample transformer.

[0109] The enhanced sample vibration signal refers to the set of signals obtained after applying random enhancement processing (such as time-domain stretching, random scaling, adding slight Gaussian noise, time shifting, etc.) to the first sample vibration signal.

[0110] The enhanced sample audio signal refers to the set of signals obtained by applying random enhancement processing (such as volume adjustment, frequency shift, random cropping, slight distortion, etc.) to the first sample audio signal.

[0111] The positive sample pair includes a first enhanced sample vibration signal and a first enhanced sample sound signal in the enhanced sample vibration signal.

[0112] Among them, the vibration signal and the sound signal of the first enhanced sample correspond to the same time point of the same transformer.

[0113] The negative sample pair includes a second enhanced sample vibration signal and a second enhanced sample sound signal from the enhanced sample vibration signal.

[0114] Among them, the vibration signal of the second enhanced sample and the sound signal of the second enhanced sample correspond to different transformers, or to different time points of the same transformer.

[0115] For example, the server selects a first sample transformer (preferably a normal transformer with stable operating conditions) and acquires the first sample vibration signal and the first sample sound signal from the first sample transformer. Then, it performs data enhancement processing on the first sample vibration signal and the first sample sound signal using preset vibration signal enhancement models and preset sound signal enhancement models, respectively, to obtain enhanced sample vibration signals and enhanced sample sound signals. Next, it iterates and combines the enhanced sample vibration signals and enhanced sample sound signals to obtain multiple sample pairs. Then, from the multiple sample pairs, it identifies the first enhanced sample vibration signal and the first enhanced sample sound signal corresponding to the same transformer at the same time point, and constructs positive sample pairs based on the first enhanced sample vibration signal and the first enhanced sample sound signal. Then, from the multiple sample pairs, it identifies the second enhanced sample vibration signal and the second enhanced sample sound signal corresponding to different transformers or different time points corresponding to the same transformer, and constructs negative and positive sample pairs based on the second enhanced sample vibration signal and the second enhanced sample sound signal. Finally, based on the positive sample pairs and negative sample pairs, iteratively trains the feature extraction model to be trained, obtaining the trained feature extraction model.

[0116] In this embodiment, by collecting the vibration and sound signals of the first sample transformer synchronously, and combining data augmentation technology to expand the sample size, the inherent physical correlation of the same source signals is preserved and the data diversity is enriched. Then, by strictly following the preset rules to construct positive and negative sample pairs, high-quality labeled equivalent data is provided for model training, so that the final trained feature extraction model has strong anti-interference ability.

[0117] In an exemplary embodiment, the feature extraction model to be trained is iteratively trained based on positive sample pairs and negative sample pairs to obtain a trained feature extraction model. Specifically, this includes: using the feature extraction model to be trained, performing feature extraction processing on the first enhanced sample vibration signal, the first enhanced sample sound signal, the second enhanced sample vibration signal, and the second enhanced sample sound signal, respectively, to obtain a first vibration feature vector corresponding to the first enhanced sample vibration signal, a first sound feature vector corresponding to the first enhanced sample sound signal, a second vibration feature vector corresponding to the second enhanced sample vibration signal, and a second sound feature vector corresponding to the second enhanced sample sound signal; determining a first similarity between the first enhanced sample vibration signal and the first enhanced sample sound signal based on the first vibration feature vector and the first sound feature vector, and determining a second similarity between the second enhanced sample vibration signal and the second enhanced sample sound signal based on the second vibration feature vector and the second sound feature vector; and iteratively training the feature extraction model to be trained based on the first similarity and the second similarity to obtain a trained feature extraction model.

[0118] The first vibration feature vector refers to the feature vector corresponding to the vibration signal of the first enhanced sample.

[0119] The first sound feature vector refers to the feature vector corresponding to the first enhanced sample sound signal.

[0120] The second vibration feature vector refers to the feature vector corresponding to the vibration signal of the second enhanced sample.

[0121] The second sound feature vector refers to the feature vector corresponding to the second enhanced sample sound signal.

[0122] The first similarity is used to represent the degree of similarity between the vibration signal of the first enhanced sample and the sound signal of the first enhanced sample.

[0123] The second similarity is used to represent the degree of similarity between the vibration signal of the second enhanced sample and the sound signal of the second enhanced sample.

[0124] For example, the server uses the vibration feature extraction network and the sound feature extraction network in the feature extraction model to be trained to perform feature extraction processing on the first enhanced sample vibration signal, the first enhanced sample sound signal, the second enhanced sample vibration signal, and the second enhanced sample sound signal, respectively, to obtain the first vibration feature vector, the first sound feature vector, the second vibration feature vector, and the second sound feature vector corresponding to the first enhanced sample vibration signal. Then, the first vibration feature vector and the first sound feature vector are input into the cosine similarity prediction model to obtain the cosine similarity between the first vibration feature vector and the first sound feature vector, which is used as the first similarity between the first enhanced sample vibration signal and the first enhanced sample sound signal. The second vibration feature vector and the second sound feature vector are input into the cosine similarity prediction model to obtain the cosine similarity between the second vibration feature vector and the second sound feature vector, which is used as the second similarity between the second enhanced sample vibration signal and the second enhanced sample sound signal. Then, based on the first similarity and the second similarity, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

[0125] In this embodiment, the enhanced acoustic and vibration signals in the positive and negative sample pairs are extracted by the feature extraction model to be trained to obtain the corresponding feature vectors. Then, by calculating the first similarity of the positive sample pairs and the second similarity of the negative sample pairs, the dependence on scarce labeled fault data is eliminated, providing high-quality and high-discrimination feature input for subsequent transformer fault diagnosis.

[0126] In an exemplary embodiment, the feature extraction model to be trained is iteratively trained based on a first similarity and a second similarity to obtain a trained feature extraction model. Specifically, this includes: obtaining a first loss value based on the difference between the first similarity and a first preset similarity, and obtaining a second loss value based on the difference between the second similarity and a second preset similarity; fusing the first loss value and the second loss value to obtain a target loss value; and iteratively training the feature extraction model to be trained based on the target loss value to obtain a trained feature extraction model.

[0127] The first preset similarity refers to the similarity threshold set for positive sample pairs (usually a value close to 1, such as 0.95).

[0128] The first loss value refers to the loss value obtained based on the difference between the first similarity and the first preset similarity.

[0129] The second preset similarity refers to the similarity threshold set for negative sample pairs (usually a value close to 0, such as 0.05).

[0130] The second loss value refers to the loss value obtained based on the difference between the second similarity and the second preset similarity.

[0131] The target loss value refers to the comprehensive loss value obtained by fusing the first loss value and the second loss value.

[0132] For example, the server obtains a first loss value based on the difference between a first similarity and a first preset similarity, and a second loss value based on the difference between a second similarity and a second preset similarity. Next, it obtains a first weight corresponding to the first loss value and a second weight corresponding to the second loss value, and sums the first and second loss values ​​according to the first and second weights to obtain a target loss value. Then, the server adjusts the model parameters of the feature extraction model to be trained based on the target loss value, and retrains the feature extraction model with adjusted model parameters until the target loss value obtained by the trained feature extraction model is less than a loss value threshold. Training then stops, and the trained feature extraction model is taken as the completed feature extraction model.

[0133] In this embodiment, the first loss value and the second loss value are obtained by calculating the difference between the first similarity and the first preset similarity, and the second similarity and the second preset similarity, respectively. This accurately quantifies the learning bias of the model in capturing positive sample association and distinguishing negative samples. The target loss value that integrates the dual biases is then obtained through fusion processing, and iterative training is carried out based on this as the core basis, which is conducive to improving the training effect of the model.

[0134] In an exemplary embodiment, the transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning provided in this application further includes a training process for a trained transformer fault diagnosis model, specifically including the following: acquiring a second sample vibration signal and a second sample sound signal of a second sample transformer; performing feature extraction processing on the second sample vibration signal and the second sample sound signal respectively through a trained feature extraction model to obtain a sample vibration feature vector corresponding to the second sample vibration signal and a sample sound feature vector corresponding to the second sample sound signal; fusing the sample vibration feature vector and the sample sound feature vector to obtain a fused sample feature vector of the second sample transformer; splitting the fused sample feature vector to obtain a training dataset and a validation dataset; fine-tuning the transformer fault diagnosis model to be trained according to the training dataset to obtain a fine-tuned transformer fault diagnosis model; validating the fine-tuned transformer fault diagnosis model according to the validation dataset to obtain a validation result of the fine-tuned transformer fault diagnosis model; and obtaining a trained transformer fault diagnosis model based on the fine-tuned transformer fault diagnosis model when the validation result meets the preset validation conditions.

[0135] The second sample transformer refers to the sample transformer used to train the transformer fault diagnosis model to be trained.

[0136] The second sample vibration signal refers to the vibration signal of the second sample transformer.

[0137] The second sample audio signal refers to the audio signal of the second sample transformer.

[0138] Among them, the sample vibration feature vector refers to the feature vector corresponding to the vibration signal of the second sample.

[0139] Here, the sample sound feature vector refers to the feature vector corresponding to the second sample sound signal.

[0140] Among them, the fused sample feature vector refers to the feature vector obtained by fusing the sample vibration feature vector and the sample sound feature vector.

[0141] The training dataset refers to the feature vectors in the fused sample feature vectors used for fine-tuning the transformer fault diagnosis model to be trained.

[0142] Among them, the feature vectors in the fusion sample feature vectors of the validation dataset are used to validate the transformer fault diagnosis model after fine-tuning training.

[0143] The validation results are used to represent the quantitative evaluation results of the transformer fault diagnosis model performance.

[0144] Among them, the preset verification conditions refer to the pre-set performance threshold standards, including the preset prediction accuracy.

[0145] For example, the server acquires the second sample vibration signal and the second sample sound signal of the second sample transformer; then, through the vibration feature extraction network and sound feature extraction network in the trained feature extraction model, feature extraction processing is performed on the second sample vibration signal and the second sample sound signal respectively to obtain the sample vibration feature vector corresponding to the second sample vibration signal and the sample sound feature vector corresponding to the second sample sound signal; then, the sample vibration feature vector and the sample sound feature vector are fused to obtain the fused sample feature vector of the second sample transformer; next, the fused sample feature vector is split according to a preset splitting ratio to obtain a training dataset and a validation dataset; then, the fused sample feature vector in the training dataset is input into the transformer fault diagnosis model to be trained to obtain the predicted fault diagnosis result of the second sample transformer output by the transformer fault diagnosis model to be trained; finally, the actual fault diagnosis result of the second sample transformer is obtained, and based on the difference between the predicted fault diagnosis result and the actual fault diagnosis result, a... The loss value of the transformer fault diagnosis model to be trained is used to adjust the model parameters. The model is then retrained until the loss value is less than a threshold. Training stops there, and this retrained model is used as the fine-tuned model. Next, the fused sample feature vectors from the validation dataset are input into the fine-tuned model to obtain the predicted fault diagnosis results for the second sample transformer. Based on these results, the validation results are obtained. If the prediction accuracy in the validation results is greater than the preset accuracy, the fine-tuned model is considered the fully trained transformer fault diagnosis model.

[0146] In this embodiment, the model is first fine-tuned based on the training dataset, and then the model performance is verified using the validation dataset. This effectively improves the model's accuracy and generalization ability in identifying transformer faults, while avoiding overfitting or underperformance issues, ensuring that the final output fault diagnosis model has stable and reliable practical performance.

[0147] In an exemplary embodiment, step S103 above, using the trained feature extraction model, performs feature extraction processing on the vibration signal and the sound signal respectively to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal. Specifically, it includes the following: using the trained feature extraction model, performing feature extraction processing on the vibration signal and the sound signal respectively to obtain the first time-frequency diagram, the first time-domain feature, the first frequency-domain feature, and the first time-frequency domain feature corresponding to the vibration signal, and the second time-frequency diagram, the second time-domain feature, the second frequency-domain feature, and the second time-frequency domain feature corresponding to the sound signal; constructing the first feature vector corresponding to the vibration signal based on the first time-frequency diagram; constructing the second feature vector corresponding to the vibration signal based on the first time-domain feature, the first frequency-domain feature, and the first time-frequency domain feature; constructing the third feature vector corresponding to the sound signal based on the second time-frequency diagram; constructing the fourth feature vector corresponding to the sound signal based on the second time-domain feature, the second frequency-domain feature, and the second time-frequency domain feature; obtaining the vibration feature vector corresponding to the vibration signal based on the first feature vector and the second feature vector, and obtaining the sound feature vector corresponding to the sound signal based on the third feature vector and the fourth feature vector.

[0148] The first time-frequency image refers to the two-dimensional image obtained by time-frequency analysis (such as continuous wavelet transform or short-time Fourier transform) corresponding to the vibration signal.

[0149] Among them, the first time-domain feature refers to the time-domain features corresponding to the vibration signal, including root mean square, peak value, peak factor, kurtosis, skewness, etc.

[0150] The first frequency domain feature refers to the frequency domain features corresponding to the vibration signal, including the spectral centroid, spectral kurtosis, harmonic energy ratio, and characteristic frequency amplitude.

[0151] Among them, the first time-frequency domain feature refers to the time-frequency domain feature corresponding to the vibration signal, including time-frequency entropy, area of ​​energy concentration region, frequency and time corresponding to maximum energy, and singular values ​​of time-frequency matrix.

[0152] The second time-frequency diagram refers to the two-dimensional image corresponding to the sound signal obtained through time-frequency analysis (such as continuous wavelet transform or short-time Fourier transform).

[0153] The second time-domain feature refers to the time-domain features corresponding to the sound signal, including root mean square, peak value, peak factor, kurtosis, skewness, etc.

[0154] The second frequency domain feature refers to the frequency domain features corresponding to the sound signal, including the spectral centroid, spectral kurtosis, harmonic energy ratio, and characteristic frequency amplitude.

[0155] The second time-frequency domain feature refers to the time-frequency domain features corresponding to the sound signal, including time-frequency entropy, area of ​​energy concentration region, frequency and time corresponding to maximum energy, and singular values ​​of the time-frequency matrix.

[0156] The first feature vector refers to the feature vector constructed based on the first time-frequency graph.

[0157] The second feature vector refers to the feature vector constructed based on the first time-domain feature, the first frequency-domain feature, and the first time-frequency-domain feature.

[0158] The third feature vector refers to the feature vector constructed based on the second time-frequency graph.

[0159] The fourth feature vector refers to the feature vector constructed based on the second time-domain feature, the second frequency-domain feature, and the second time-frequency-domain feature.

[0160] For example, the server uses a trained feature extraction model to perform feature extraction processing on vibration signals and sound signals respectively, obtaining a first time-frequency map, a first time-domain feature, a first frequency-domain feature, and a first time-frequency domain feature corresponding to the vibration signal, and a second time-frequency map, a second time-domain feature, a second frequency-domain feature, and a second time-frequency domain feature corresponding to the sound signal. Next, the first time-frequency map is convolved, activated, and pooled to obtain a first feature vector corresponding to the vibration signal. The first time-domain feature, the first frequency-domain feature, and the first time-frequency domain feature are standardized and concatenated according to a preset concatenation order to obtain a second feature vector corresponding to the vibration signal. The second time-frequency map is convolved, activated, and pooled to obtain a third feature vector corresponding to the sound signal. The second time-domain feature, the second frequency-domain feature, and the second time-frequency domain feature are standardized and concatenated according to a preset concatenation order to obtain a fourth feature vector corresponding to the sound signal. Finally, the first and second feature vectors are fused to obtain a vibration feature vector corresponding to the vibration signal, and the third and fourth feature vectors are fused to obtain a sound feature vector corresponding to the sound signal.

[0161] In this embodiment, the trained feature extraction model extracts multi-dimensional features from vibration and sound signals, so that the final vibration and sound feature vectors retain the dynamic change details of the signals and integrate core statistical information, effectively avoiding the loss of information in a single feature dimension, and providing high-quality and comprehensive feature input for the subsequent transformer fault diagnosis model.

[0162] In one exemplary embodiment, such as Figure 2 As shown, another transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning is provided. Taking the application of this method to a server as an example, the specific steps include:

[0163] Step S201: Obtain the current operating status of the transformer to be analyzed.

[0164] Step S202: Under the condition that the current operating status indicates that the transformer to be analyzed is operating normally, acquire the vibration signal and sound signal of the transformer to be analyzed.

[0165] Step S203: Using the trained feature extraction model, feature extraction processing is performed on the vibration signal and the sound signal respectively to obtain the first time-frequency diagram, first time-domain feature, first frequency-domain feature and first time-frequency domain feature corresponding to the vibration signal, and the second time-frequency diagram, second time-domain feature, second frequency-domain feature and second time-frequency domain feature corresponding to the sound signal.

[0166] Step S204: Based on the first time-frequency diagram, construct the first feature vector corresponding to the vibration signal; based on the first time-domain feature, the first frequency-domain feature, and the first time-frequency-domain feature, construct the second feature vector corresponding to the vibration signal; based on the second time-frequency diagram, construct the third feature vector corresponding to the sound signal; based on the second time-domain feature, the second frequency-domain feature, and the second time-frequency-domain feature, construct the fourth feature vector corresponding to the sound signal.

[0167] Step S205: Based on the first feature vector and the second feature vector, obtain the vibration feature vector corresponding to the vibration signal, and based on the third feature vector and the fourth feature vector, obtain the sound feature vector corresponding to the sound signal.

[0168] Step S206: The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed.

[0169] Step S207: Input the fused feature vector into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

[0170] In the aforementioned transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning, during transformer fault diagnosis, the method focuses on acquiring dual-mode vibration and sound signals under normal transformer operation. It then uses a trained feature extraction model to accurately extract high-dimensional feature vectors from both types of signals. Furthermore, it integrates complementary information from both sources through feature fusion processing. Based on the trained transformer fault diagnosis model, the dual-mode fusion features provide a more comprehensive and robust representation, significantly improving the accuracy of identifying early, subtle, and complex faults, thus enhancing the overall accuracy of transformer fault diagnosis. Moreover, the entire process employs a combined analysis of vibration and sound signals, avoiding the shortcomings of traditional single-mode signal analysis methods, which suffer from weak anti-interference capabilities and difficulty in capturing subtle early signs of faults, leading to lower accuracy in transformer fault diagnosis. This further improves the overall accuracy of transformer fault diagnosis.

[0171] In an exemplary embodiment, to more clearly illustrate the transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning provided in this application, the following specific embodiment will be used to describe the transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning. In one embodiment, as... Figure 3 As shown, this application also provides an early diagnosis system for the mechanical condition of transformers based on acoustic-vibration fusion and self-supervised learning. Specifically, it includes the following:

[0172] 1. Sound-Vibration Heterogeneous Synchronous Acquisition Module: This module is deployed at the transformer site and includes a high-sensitivity vibration acceleration sensor and a high-fidelity directional microphone array. Equipped with a high-precision synchronization trigger unit, it ensures that sound and vibration signals are strictly synchronized in the time domain, laying the data foundation for subsequent cross-modal fusion. The sampling rate is configured differently according to the signal characteristics, and time alignment accuracy is guaranteed at the microsecond level through hardware clock synchronization.

[0173] 2. Signal Preprocessing and Feature Construction Module: This module performs DC removal, filtering, and noise reduction on the acquired raw sound pressure and vibration acceleration signals. Continuous wavelet transforms are applied to a synchronized segment of sound and vibration signals to generate high-quality time-frequency maps. This time-frequency map, as a two-dimensional image, serves as the feature input for subsequent convolutional neural networks. From the synchronized sound and vibration signal segments, time-domain features (such as root mean square, peak value, and kurtosis), frequency-domain features (such as spectral centroid and frequency variance), and time-frequency-domain features (such as wavelet packet energy entropy) are extracted to collectively construct a high-dimensional multimodal fusion feature vector.

[0174] 3. Sound-Vibration Fusion Self-Supervised Pre-training Module: This module adopts a multimodal self-supervised pre-training network based on contrastive learning. The schematic diagram of this network structure is shown below. Figure 4 As shown: For any synchronously acquired data pair, random augmentation is applied to each pair. The two augmented versions, along with the original data pair, constitute multiple positive sample pairs. Acoustic and vibration signals acquired from different time points or different transformers are randomly paired to form negative sample pairs. A dual-tower twin network structure is adopted, with one tower for vibration feature extraction and the other for sound feature extraction. The training objective is to maximize the similarity of acoustic and vibration representation vectors in positive sample pairs while minimizing the similarity in negative sample pairs. The loss function uses normalized temperature-scaled cross-entropy loss. The above network is pre-trained using only a large amount of unlabeled synchronous acoustic and vibration data acquired under normal transformer conditions.

[0175] 4. Downstream fine-tuning diagnostic module, its workflow is as follows: Figure 5As shown: The parameters of the pre-trained sound feature extraction network and vibration feature extraction network are frozen or used as initialization, and then a trainable multilayer perceptron classifier is connected to them. Supervised fine-tuning is performed on this classifier using a very small amount of labeled data. During fine-tuning, the fused features obtained after the input sound and vibration signals pass through the pre-trained feature extraction network are input into the classifier to determine the current state.

[0176] 5. Status Early Warning and Visualized Human-Computer Interaction Module: This module receives diagnostic results, calculates health indices, and triggers multi-level early warnings when abnormal or fault trends are detected. It provides a web (network) or client interface to visually display real-time / historical acoustic and vibration signals, time-frequency graphs, model diagnostic results, confidence levels, and early warning information.

[0177] The workflow of this system is as follows: pre-training and model optimization, the steps are as follows: Figure 6 As shown. This stage requires no fault data; model pre-training is completed using only normal state data, followed by fine-tuning with a small amount of labeled fault data, providing a "highly robust feature extractor + high-precision classification model" for online diagnosis. The online diagnosis and early warning process is as follows: Figure 7 As shown. This stage utilizes the model trained in the first stage to perform automated diagnosis and early warning of real-time data. This is achieved through time-domain, frequency-domain, and real-time frequency-domain analysis of the acoustic and vibration signals of three types of mechanical faults, such as... Figure 8 , Figure 9 , Figure 10 , Figure 11 , Figure 12 , Figure 13 , Figure 14 , Figure 15 , Figure 16 As shown.

[0178] In the above embodiments, when diagnosing transformer faults, the system focuses on acquiring dual-mode vibration and sound signals during normal transformer operation. High-dimensional feature vectors of both signals are accurately extracted using a trained feature extraction model. Then, complementary information from both sources is integrated through feature fusion processing. Based on the trained transformer fault diagnosis model, the system leverages the more comprehensive and robust representation capabilities of the dual-mode fusion features to significantly improve the accuracy of identifying early, subtle, and complex faults, thus enhancing the overall accuracy of transformer fault diagnosis. Furthermore, the entire process employs a combined analysis of vibration and sound signals, avoiding the shortcomings of traditional single-mode signal analysis methods, which suffer from weak anti-interference capabilities and difficulty in capturing subtle early signs of faults, leading to lower accuracy in transformer fault diagnosis. This further improves the overall accuracy of transformer fault diagnosis.

[0179] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0180] Based on the same inventive concept, this application also provides a transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning for implementing the transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning provided below can be found in the limitations of the transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning described above, and will not be repeated here.

[0181] In one exemplary embodiment, such as Figure 17 As shown, a transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning is provided, including: a state acquisition module 1701, a signal acquisition module 1702, a feature extraction module 1703, a feature fusion module 1704, and a fault diagnosis module 1705, wherein:

[0182] The status acquisition module 1701 is used to acquire the current operating status of the transformer to be analyzed.

[0183] The signal acquisition module 1702 is used to acquire the vibration signal and sound signal of the transformer to be analyzed when the current operating status indicates that the transformer to be analyzed is operating normally.

[0184] The feature extraction module 1703 is used to perform feature extraction processing on vibration signals and sound signals respectively through the trained feature extraction model to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal.

[0185] The feature fusion module 1704 is used to fuse the vibration feature vector and the sound feature vector to obtain the fused feature vector of the transformer to be analyzed.

[0186] The fault diagnosis module 1705 is used to input the fused feature vector into the trained transformer fault diagnosis model to obtain the fault diagnosis results of the transformer to be analyzed.

[0187] In an exemplary embodiment, the transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning further includes a first training module, used to acquire a first sample vibration signal and a first sample sound signal of a first sample transformer; perform data augmentation processing on the first sample vibration signal and the first sample sound signal respectively to obtain augmented sample vibration signal and augmented sample sound signal; construct positive sample pairs and negative sample pairs based on the augmented sample vibration signal and the augmented sample sound signal; the positive sample pair includes a first augmented sample vibration signal and a first augmented sample sound signal from the augmented sample vibration signal; the first augmented sample vibration signal and the first augmented sample sound signal correspond to the same time point of the same transformer; the negative sample pair includes a second augmented sample vibration signal and a second augmented sample sound signal from the augmented sample vibration signal; the second augmented sample vibration signal and the second augmented sample sound signal correspond to different transformers, or to different time points of the same transformer; and iteratively train the feature extraction model to be trained based on the positive sample pair and the negative sample pair to obtain the trained feature extraction model.

[0188] In an exemplary embodiment, the first training module is further configured to perform feature extraction processing on the first enhanced sample vibration signal, the first enhanced sample sound signal, the second enhanced sample vibration signal, and the second enhanced sample sound signal respectively using the feature extraction model to be trained, to obtain a first vibration feature vector corresponding to the first enhanced sample vibration signal, a first sound feature vector corresponding to the first enhanced sample sound signal, a second vibration feature vector corresponding to the second enhanced sample vibration signal, and a second sound feature vector corresponding to the second enhanced sample sound signal; determine a first similarity between the first enhanced sample vibration signal and the first enhanced sample sound signal based on the first vibration feature vector and the first sound feature vector, and determine a second similarity between the second enhanced sample vibration signal and the second enhanced sample sound signal based on the second vibration feature vector and the second sound feature vector; and iteratively train the feature extraction model to be trained based on the first similarity and the second similarity to obtain the trained feature extraction model.

[0189] In an exemplary embodiment, the first training module is further configured to obtain a first loss value based on the difference between the first similarity and the first preset similarity, and to obtain a second loss value based on the difference between the second similarity and the second preset similarity; to fuse the first loss value and the second loss value to obtain a target loss value; and to iteratively train the feature extraction model to be trained based on the target loss value to obtain a trained feature extraction model.

[0190] In an exemplary embodiment, the transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning further includes a second training module, used to acquire second sample vibration signals and second sample sound signals of a second sample transformer; perform feature extraction processing on the second sample vibration signals and second sample sound signals respectively through the trained feature extraction model to obtain sample vibration feature vectors corresponding to the second sample vibration signals and sample sound feature vectors corresponding to the second sample sound signals; fuse the sample vibration feature vectors and sample sound feature vectors to obtain fused sample feature vectors of the second sample transformer; split the fused sample feature vectors to obtain training datasets and validation datasets; fine-tune the transformer fault diagnosis model to be trained according to the training dataset to obtain a fine-tuned transformer fault diagnosis model; validate the fine-tuned transformer fault diagnosis model according to the validation dataset to obtain the validation result of the fine-tuned transformer fault diagnosis model; and, if the validation result meets the preset validation conditions, obtain the trained transformer fault diagnosis model based on the fine-tuned transformer fault diagnosis model.

[0191] In an exemplary embodiment, the feature extraction module 1703 is further configured to perform feature extraction processing on the vibration signal and the sound signal respectively using the trained feature extraction model, to obtain a first time-frequency diagram, a first time-domain feature, a first frequency-domain feature, and a first time-frequency domain feature corresponding to the vibration signal, and a second time-frequency diagram, a second time-domain feature, a second frequency-domain feature, and a second time-frequency domain feature corresponding to the sound signal; construct a first feature vector corresponding to the vibration signal based on the first time-frequency diagram, construct a second feature vector corresponding to the vibration signal based on the first time-domain feature, the first frequency-domain feature, and the first time-frequency domain feature, construct a third feature vector corresponding to the sound signal based on the second time-frequency diagram, and construct a fourth feature vector corresponding to the sound signal based on the second time-domain feature, the second frequency-domain feature, and the second time-frequency domain feature; obtain a vibration feature vector corresponding to the vibration signal based on the first feature vector and the second feature vector, and obtain a sound feature vector corresponding to the sound signal based on the third feature vector and the fourth feature vector.

[0192] The modules in the aforementioned transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0193] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 18 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data such as vibration signals and sound signals. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning.

[0194] Those skilled in the art will understand that Figure 18 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0195] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0196] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above-described method embodiments.

[0197] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0198] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0199] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0200] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A transformer fault diagnosis method based on acoustic-vibration fusion and self-supervised learning, characterized in that, The method includes: Obtain the current operating status of the transformer to be analyzed; When the current operating state indicates that the transformer under analysis is operating normally, the vibration signal and sound signal of the transformer under analysis are acquired. The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, so as to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal. The vibration feature vector and the sound feature vector are fused to obtain the fused feature vector of the transformer to be analyzed; The fused feature vector is input into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

2. The method according to claim 1, characterized in that, The trained feature extraction model is obtained through the following method: Acquire the first sample vibration signal and the first sample sound signal of the first sample transformer; Data augmentation processing is performed on the first sample vibration signal and the first sample sound signal respectively to obtain the enhanced sample vibration signal and the enhanced sample sound signal; Based on the enhanced sample vibration signal and the enhanced sample sound signal, positive sample pairs and negative sample pairs are constructed; the positive sample pair includes a first enhanced sample vibration signal and a first enhanced sample sound signal from the enhanced sample vibration signal; the first enhanced sample vibration signal and the first enhanced sample sound signal correspond to the same time point of the same transformer; the negative sample pair includes a second enhanced sample vibration signal and a second enhanced sample sound signal from the enhanced sample vibration signal; the second enhanced sample vibration signal and the second enhanced sample sound signal correspond to different transformers, or to different time points of the same transformer; Based on the positive sample pairs and the negative sample pairs, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

3. The method according to claim 2, characterized in that, The step of iteratively training the feature extraction model to be trained based on the positive sample pairs and the negative sample pairs to obtain the trained feature extraction model includes: The feature extraction model to be trained is used to perform feature extraction processing on the first enhanced sample vibration signal, the first enhanced sample sound signal, the second enhanced sample vibration signal, and the second enhanced sample sound signal, respectively, to obtain the first vibration feature vector corresponding to the first enhanced sample vibration signal, the first sound feature vector corresponding to the first enhanced sample sound signal, the second vibration feature vector corresponding to the second enhanced sample vibration signal, and the second sound feature vector corresponding to the second enhanced sample sound signal. Based on the first vibration feature vector and the first sound feature vector, a first similarity between the first enhanced sample vibration signal and the first enhanced sample sound signal is determined, and based on the second vibration feature vector and the second sound feature vector, a second similarity between the second enhanced sample vibration signal and the second enhanced sample sound signal is determined. Based on the first similarity and the second similarity, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

4. The method according to claim 3, characterized in that, The step of iteratively training the feature extraction model to be trained based on the first similarity and the second similarity to obtain the trained feature extraction model includes: A first loss value is obtained based on the difference between the first similarity and the first preset similarity, and a second loss value is obtained based on the difference between the second similarity and the second preset similarity. The first loss value and the second loss value are fused together to obtain the target loss value; Based on the target loss value, the feature extraction model to be trained is iteratively trained to obtain the trained feature extraction model.

5. The method according to claim 1, characterized in that, The trained transformer fault diagnosis model was obtained through the following method: Acquire the second sample vibration signal and the second sample sound signal of the second sample transformer; The trained feature extraction model is used to perform feature extraction processing on the second sample vibration signal and the second sample sound signal respectively to obtain the sample vibration feature vector corresponding to the second sample vibration signal and the sample sound feature vector corresponding to the second sample sound signal. The sample vibration feature vector and the sample sound feature vector are fused to obtain the fused sample feature vector of the second sample transformer; The feature vectors of the fused samples are split to obtain a training dataset and a validation dataset; Based on the training dataset, the transformer fault diagnosis model to be trained is fine-tuned to obtain the fine-tuned transformer fault diagnosis model. The fine-tuned transformer fault diagnosis model is validated based on the validation dataset to obtain the validation results of the fine-tuned transformer fault diagnosis model. If the verification results meet the preset verification conditions, the trained transformer fault diagnosis model is obtained based on the fine-tuned and trained transformer fault diagnosis model.

6. The method according to any one of claims 1 to 5, characterized in that, The trained feature extraction model performs feature extraction processing on the vibration signal and the sound signal respectively, to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal, including: The trained feature extraction model is used to extract features from the vibration signal and the sound signal respectively, to obtain the first time-frequency diagram, first time-domain feature, first frequency-domain feature and first time-frequency-domain feature corresponding to the vibration signal, and the second time-frequency diagram, second time-domain feature, second frequency-domain feature and second time-frequency-domain feature corresponding to the sound signal. Based on the first time-frequency diagram, a first feature vector corresponding to the vibration signal is constructed; based on the first time-domain feature, the first frequency-domain feature, and the first time-frequency-domain feature, a second feature vector corresponding to the vibration signal is constructed; based on the second time-frequency diagram, a third feature vector corresponding to the sound signal is constructed; and based on the second time-domain feature, the second frequency-domain feature, and the second time-frequency-domain feature, a fourth feature vector corresponding to the sound signal is constructed. Based on the first feature vector and the second feature vector, a vibration feature vector corresponding to the vibration signal is obtained, and based on the third feature vector and the fourth feature vector, a sound feature vector corresponding to the sound signal is obtained.

7. A transformer fault diagnosis device based on acoustic-vibration fusion and self-supervised learning, characterized in that, The device includes: The status acquisition module is used to acquire the current operating status of the transformer to be analyzed. The signal acquisition module is used to acquire the vibration signal and sound signal of the transformer to be analyzed when the current operating state indicates that the transformer to be analyzed is operating normally; The feature extraction module is used to perform feature extraction processing on the vibration signal and the sound signal respectively through the trained feature extraction model to obtain the vibration feature vector corresponding to the vibration signal and the sound feature vector corresponding to the sound signal; The feature fusion module is used to fuse the vibration feature vector and the sound feature vector to obtain the fused feature vector of the transformer to be analyzed. The fault diagnosis module is used to input the fused feature vector into the trained transformer fault diagnosis model to obtain the fault diagnosis result of the transformer to be analyzed.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to 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 a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images