Vibration-based power transformer defect detection method fused with voiceprint signal

Abstract

The invention relates to a voiceprint signal fused vibration-based power transformer defect detection method, which comprises the following steps: S1, collecting transformer defect types, carrying out a transformer defect test, and collecting vibration signals and sound signals of a transformer in different defect types and under normal conditions, taking the transformer defect state of the defect test as labeling information; s2, time axis alignment is carried out on the collected vibration signals and sound signals, and a vibration-to-sound model, a sound-to-vibration model and a defect detection model are trained based on the aligned vibration signals and sound signals; s3, collecting a vibration signal of the transformer in actual operation to obtain a first feature, and obtaining a sound signal corresponding to the first feature by using a vibration-to-sound model to obtain a second feature; and S4, splicing the first feature and the second feature, and inputting the spliced first feature and second feature into a defect detection model to obtain a defect detection result. Compared with the prior art, the method has the advantages that the accuracy rate of defect detection is guaranteed under the condition that actually measured voiceprint signals are not needed, and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent fault diagnosis of power equipment, in particular to a power transformer defect detection method based on vibration fusing voiceprint signals. BACKGROUND

[0002] Large power transformers are the core equipment of power systems, and play a key role in voltage conversion and energy transmission. Their operating state directly affects the safe and stable operation of the power grid. According to relevant statistics, power outages caused by transformer faults account for 15-20% of total power grid accidents, and about 70% of the faults can be early characterized by vibration or acoustic signals. However, traditional detection methods have significant limitations: In order to solve the limitations of single modal signals, some existing technologies disclose ways that can analyze multi-modal data, for example, Chinese patent CN118779807A discloses a power transformer defect diagnosis method based on multi-modal sound map fusion, the steps include obtaining image data and voiceprint data of the target power transformer; constructing a transformer defect diagnosis model and a transformer classification model; detecting whether the transformer has defects and determining the location of the defects, judging the defect type of the transformer; integrating the results of defect detection and the results of defect type judgment to obtain a prediction score value of the image data; constructing a voiceprint diagnosis model and pre-training the voiceprint diagnosis model; processing the voiceprint data, extracting features and diagnosing the fault type to obtain a prediction score value of the voiceprint data; based on the obtained prediction score value of the image data and the prediction score value of the voiceprint data, adopting an attention mechanism to fuse the prediction score values to make a decision fusion and obtain the final defect diagnosis result.

[0003] And some other existing technologies mention that multi-modal data can also include vibration data, but in practice, in the application link, some multi-modal signals are prone to interference, for example, sound signals can be transmitted through the air, adjacent devices and there is a certain interval, which is also prone to interference. In fact, in some device-dense spaces, it is difficult to collect the sound signal of a single device. For this, although based on the inspiration of existing technologies, multi-modal signals can be used as the superposition of signals that are easy to collect, such as vibration signals and infrared images, but compared to infrared images, sound signals are derived from electromagnetic force and mechanical vibration inside the transformer, and their changes directly reflect the health status of core components such as the core and winding. This makes voiceprint technology particularly good at diagnosing internal mechanical defects such as winding looseness, deformation, and core abnormalities. These defects are one of the main reasons for transformer failure, and the precision of diagnosis will decrease if the sound signal is removed. SUMMARY

[0004] The purpose of this invention is to provide a vibration-based power transformer defect detection method that integrates acoustic signature signals to overcome the shortcomings of the prior art.

[0005] The objective of this invention can be achieved through the following technical solutions: A vibration-based defect detection method for power transformers that integrates acoustic signature signals includes: Step S1: Collect transformer defect types and conduct transformer defect tests. Collect vibration and sound signals of the transformer under different defect types and normal conditions, and use the transformer defect status of the defect test as labeling information. During the test, the transformer is located in a soundproof room. Step S2: Align the collected vibration and sound signals along the time axis, and train the vibration-to-sound model, the sound-to-vibration model, and the defect detection model based on the aligned vibration and sound signals; Step S3: Collect the vibration signal of the voltage transformer in actual operation to obtain the first feature, and use the vibration-to-sound model to obtain the corresponding sound signal to obtain the second feature; Step S4: After concatenating the first feature and the second feature, input them into the defect detection model to obtain the defect detection result.

[0006] In step S1, during the test, the electrical equipment connected to the transformer is located outside the soundproof room.

[0007] Step S2 includes: Step S2-1: Train the vibration-to-sound model and the sound-to-vibration model based on the aligned vibration and sound signals; Step S2-2: Train the defect detection model based on the aligned vibration signal and sound signal combined with the annotation information.

[0008] Step S2-1 includes: Step S2-1-1: Convert the aligned vibration signal into a native vibration spectrum and the aligned sound signal into a native speech spectrum; Step S2-1-2: Convert the original vibration spectrogram into an accompanying speech spectrogram based on the vibration-to-sound model; Step S2-1-3: Convert the accompanying speech spectrogram into a regenerated vibration spectrogram based on the sound-to-vibration model; Step S2-1-4: Calculate the difference between the regenerated vibration spectrum and the original vibration spectrum to obtain the first loss value, and calculate the difference between the accompanying speech spectrum and the original speech spectrum to obtain the second loss value; Step S2-1-5: The first loss value and the second loss value are weighted and summed to obtain the total conversion loss, and the model parameters of the vibration-to-sound model and the sound-to-vibration model are adjusted based on the total conversion loss; Step S2-1-6: Repeat steps S2-1-1 to S2-1-5 until the exit condition is met.

[0009] Step S2-2 includes: Step S2-2-1: Convert the aligned vibration signal into a native vibration spectrum and the aligned sound signal into a native speech spectrum; Step S2-2-2: Convert the original vibration spectrogram into an accompanying speech spectrogram based on the vibration-to-sound model; Step S2-2-3: Extract the first feature based on the original vibration spectrogram, extract the second feature based on the accompanying speech spectrogram, and extract the third feature based on the original speech spectrogram; Step S2-2-4: After concatenating the first feature and the second feature, input them into the defect detection model to obtain the first prediction result output by the defect detection model. After concatenating the first feature and the third feature, input them into the defect detection model to obtain the second prediction result output by the defect detection model. Step S2-2-5: Calculate the difference between the first prediction result and the labeled information to obtain the third loss value, calculate the difference between the second prediction result and the labeled information to obtain the fourth loss value, calculate the total classification loss based on the third loss value and the fourth loss value, and adjust the model parameters of the defect detection model based on the total classification loss; Step S2-2-6: Repeat steps S2-2-1 to S2-2-5 until the exit condition is met.

[0010] The mathematical expression for the total classification loss is: in: For total classification loss, As the first weight, This is the third loss value. As the second weight, This is the fourth loss value. As the third weight, For similarity loss, As the fourth weight, As a complementary loss, For Euclidean distance, For standard information, This is the first prediction result. This is the second prediction result.

[0011] In step S1, each defect type is tested multiple times, and multiple samples are collected.

[0012] In step S1, the number of positive samples is the same as the number of negative samples.

[0013] A vibration-based power transformer defect detection device that integrates acoustic signature signals includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.

[0014] A storage medium having a program stored thereon, which, when executed, implements the method described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By designing vibration-to-sound and sound-to-vibration models, the correlation between vibration and acoustic signature can be explored, enabling the mutual conversion of vibration signals and sound signals. This allows the integration of sound signal characteristics without the need for original sound signal acquisition, thereby improving the accuracy of defect detection.

[0016] 2. Using the spectrum diagram as the standard for the sound signal of the vibration signal can reduce the structural complexity of the vibration-to-sound model and the sound-to-vibration model, thereby reducing the required sample size.

[0017] 3. Compared to simply setting a vibration-to-sound model, generating a regenerated vibration spectrum from the accompanying speech spectrum can avoid the accompanying speech failing to correctly revert to the vibration signal, thus effectively improving the consistency between the accompanying speech spectrum and the vibration spectrum.

[0018] 4. By introducing similarity loss and complementarity loss, the model's ability to learn two combinations of features is enhanced. In the similarity loss term, the model is encouraged to maintain consistency between the first and second loss values ​​in the feature space, even if they come from different feature combinations. This allows the model to learn a shared defect representation, whether based on a combination of vibration and accompanying voiceprint features or a combination of vibration and native voiceprint features, thereby enhancing the robustness and stability of the prediction. Simple weighted summation cannot promote this because it ignores possible common patterns between the loss values.

[0019] On the other hand, complementarity loss further enhances the model's utilization of information differences. In practical applications, features between different modalities may not completely overlap, but rather contain their own unique and useful information. By measuring the difference in loss values, complementarity loss allows the model to learn complementarity across different modalities, which is crucial for improving the model's multimodal fusion capability. Weighted summation, however, does not take this into account; it simply adds the two loss values ​​together, which may lead to the model performing worse than expected in certain modalities, especially when there are significant differences between the signals.

[0020] In summary, the innovative loss function, by incorporating similarity and complementarity measures, enables the model to not only focus on reducing the sum of the two losses but also maximize the efficiency of multimodal information utilization by understanding their intrinsic relationships. This approach can more effectively improve the accuracy of transformer defect detection and avoid the risks of information loss or over-reliance on a single mode that may arise from weighted summation. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the main steps of the method of the present invention; Figure 2 This is a schematic diagram illustrating the interconversion of vibration and sound. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] A vibration-based defect detection method for power transformers that integrates acoustic signature signals, such as... Figure 1 As shown, it includes: Step S1: Collect transformer defect types and conduct transformer defect tests. Collect vibration and sound signals of the transformer under different defect types and normal conditions, and use the transformer defect status of the defect test as labeling information. During the test, the transformer is located in a soundproof room. During the test, the electrical equipment connected to the transformer was located outside the soundproof room. The soundproof room, as an acoustic laboratory, had excellent sound insulation properties, ensuring that only the sound of the transformer itself was inside the soundproof room, and there were no other noises.

[0024] For each defect type, multiple experiments were conducted, and multiple samples were collected. The number of positive samples was the same as the number of negative samples. The positive samples were obtained from data collected under normal conditions.

[0025] Step S2: Align the acquired vibration and sound signals along their time axes, and based on the aligned vibration and sound signals, train the vibration-to-sound model, the sound-to-vibration model, and the defect detection model, including: Step S2-1: Train the vibration-to-sound model and the sound-to-vibration model based on the aligned vibration and sound signals, including: Step S2-1-1: Convert the aligned vibration signal into a native vibration spectrum and the aligned sound signal into a native speech spectrum; Step S2-1-2: Convert the original vibration spectrogram into an accompanying speech spectrogram based on the vibration-to-sound model; Step S2-1-3: Convert the accompanying speech spectrogram into a regenerated vibration spectrogram based on the sound-to-vibration model; Step S2-1-4: Calculate the difference between the regenerated vibration spectrum and the original vibration spectrum to obtain the first loss value, and calculate the difference between the accompanying speech spectrum and the original speech spectrum to obtain the second loss value; Step S2-1-5: The first loss value and the second loss value are weighted and summed to obtain the total conversion loss, and the model parameters of the vibration-to-sound model and the sound-to-vibration model are adjusted based on the total conversion loss; Step S2-1-6: Repeat steps S2-1-1 to S2-1-5 until the exit condition is met. Generally, the exit condition is that the model achieves convergence.

[0026] In this application, in addition to the vibration-to-vibration model, a sound-to-vibration model is also used. Unlike the traditional sound-to-vibration model which is used to obtain vibration signals, this application uses a sound-to-vibration model to convert the accompanying speech spectrogram back into the form of a vibration signal. This allows the difference between the regenerated vibration spectrogram and the original vibration spectrogram to obtain a first loss value. Based on this loss value, the reliability of the implicit information in the obtained accompanying speech spectrogram can be determined, thereby effectively improving the integration of the sound signal into the subsequent second feature.

[0027] Step S2-2: Train the defect detection model based on the aligned vibration signal and sound signal combined with the annotation information.

[0028] Step S2-2 includes: Step S2-2-1: Convert the aligned vibration signal into a native vibration spectrum and the aligned sound signal into a native speech spectrum; Step S2-2-2: Convert the original vibration spectrogram into an accompanying speech spectrogram based on the vibration-to-sound model; Step S2-2-3: Extract the first feature based on the original vibration spectrogram, extract the second feature based on the accompanying speech spectrogram, and extract the third feature based on the original speech spectrogram; Step S2-2-4: After concatenating the first feature and the second feature, input them into the defect detection model to obtain the first prediction result output by the defect detection model. After concatenating the first feature and the third feature, input them into the defect detection model to obtain the second prediction result output by the defect detection model. Step S2-2-5: Calculate the difference between the first prediction result and the labeled information to obtain the third loss value, calculate the difference between the second prediction result and the labeled information to obtain the fourth loss value, calculate the total classification loss based on the third loss value and the fourth loss value, and adjust the model parameters of the defect detection model based on the total classification loss; The mathematical expression for the total classification loss is: in: For total classification loss, As the first weight, This is the third loss value. As the second weight, This is the fourth loss value. As the third weight, For similarity loss, As the fourth weight, As a complementary loss, For Euclidean distance, For standard information, This is the first prediction result. This is the second prediction result.

[0029] By introducing similarity and complementarity losses, the model's ability to learn two combinations of features is enhanced. In the similarity loss term, the model is encouraged to maintain consistency between the first and second loss values ​​in the feature space, even if they come from different feature combinations. This allows the model to learn a shared defect representation, whether based on a combination of vibration and accompanying voiceprint features or a combination of vibration and native voiceprint features, thereby enhancing the robustness and stability of the prediction. Simple weighted summation cannot promote this because it ignores possible common patterns between the loss values.

[0030] On the other hand, complementarity loss further enhances the model's utilization of information differences. In practical applications, features between different modalities may not completely overlap, but rather contain their own unique and useful information. By measuring the difference in loss values, complementarity loss allows the model to learn complementarity across different modalities, which is crucial for improving the model's multimodal fusion capability. Weighted summation, however, does not take this into account; it simply adds the two loss values ​​together, which may lead to the model performing worse than expected in certain modalities, especially when there are significant differences between the signals.

[0031] In summary, the innovative loss function, by incorporating similarity and complementarity measures, enables the model to not only focus on reducing the sum of the two losses but also maximize the efficiency of multimodal information utilization by understanding their intrinsic relationships. This approach can more effectively improve the accuracy of transformer defect detection and avoid the risks of information loss or over-reliance on a single mode that may arise from weighted summation.

[0032] Step S2-2-6: Repeat steps S2-2-1 to S2-2-5 until the exit condition is met. Generally, the exit condition is that the model achieves convergence.

[0033] Step S3: Collect the vibration signal of the voltage transformer in actual operation to obtain the first feature, and use the vibration-to-sound model to obtain the corresponding sound signal to obtain the second feature; Step S4: After concatenating the first feature and the second feature, input them into the defect detection model to obtain the defect detection result.

[0034] The above defect detection model can be implemented using a convolutional neural network model. The first and second features can be implemented using an image encoder, and the vibration-to-sound model and sound-to-vibration model can be implemented using a generative adversarial network.

[0035] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A vibration-based defect detection method for power transformers that integrates acoustic signature signals, characterized in that, include: Step S1: Collect transformer defect types and conduct transformer defect tests. Collect vibration and sound signals of the transformer under different defect types and normal conditions, and use the transformer defect status of the defect test as labeling information. During the test, the transformer is located in a soundproof room. Step S2: Align the collected vibration and sound signals along the time axis, and train the vibration-to-sound model, the sound-to-vibration model, and the defect detection model based on the aligned vibration and sound signals; Step S3: Collect the vibration signal of the voltage transformer in actual operation to obtain the first feature, and use the vibration-to-sound model to obtain the corresponding sound signal to obtain the second feature; Step S4: After concatenating the first feature and the second feature, input them into the defect detection model to obtain the defect detection result.

2. The method for detecting defects in a vibration-based power transformer by fusing acoustic signature signals according to claim 1, characterized in that, In step S1, during the test, the electrical equipment connected to the transformer is located outside the soundproof room.

3. The method for detecting defects in a vibration-based power transformer by fusing acoustic signature signals according to claim 1, characterized in that, Step S2 includes: Step S2-1: Train the vibration-to-sound model and the sound-to-vibration model based on the aligned vibration and sound signals; Step S2-2: Train the defect detection model based on the aligned vibration signal and sound signal combined with the annotation information.

4. The vibration-based power transformer defect detection method based on fused acoustic signature signals according to claim 3, characterized in that, Step S2-1 includes: Step S2-1-1: Convert the aligned vibration signal into a native vibration spectrum and the aligned sound signal into a native speech spectrum; Step S2-1-2: Convert the original vibration spectrogram into an accompanying speech spectrogram based on the vibration-to-sound model; Step S2-1-3: Convert the accompanying speech spectrogram into a regenerated vibration spectrogram based on the sound-to-vibration model; Step S2-1-4: Calculate the difference between the regenerated vibration spectrum and the original vibration spectrum to obtain the first loss value, and calculate the difference between the accompanying speech spectrum and the original speech spectrum to obtain the second loss value; Step S2-1-5: The first loss value and the second loss value are weighted and summed to obtain the total conversion loss, and the model parameters of the vibration-to-sound model and the sound-to-vibration model are adjusted based on the total conversion loss; Step S2-1-6: Repeat steps S2-1-1 to S2-1-5 until the exit condition is met.

5. The method for detecting defects in a vibration-based power transformer by fusing acoustic signature signals according to claim 3, characterized in that, Step S2-2 includes: Step S2-2-1: Convert the aligned vibration signal into a native vibration spectrum and the aligned sound signal into a native speech spectrum; Step S2-2-2: Convert the original vibration spectrogram into an accompanying speech spectrogram based on the vibration-to-sound model; Step S2-2-3: Extract the first feature based on the original vibration spectrogram, extract the second feature based on the accompanying speech spectrogram, and extract the third feature based on the original speech spectrogram; Step S2-2-4: After concatenating the first feature and the second feature, input them into the defect detection model to obtain the first prediction result output by the defect detection model. After concatenating the first feature and the third feature, input them into the defect detection model to obtain the second prediction result output by the defect detection model. Step S2-2-5: Calculate the difference between the first prediction result and the labeled information to obtain the third loss value, calculate the difference between the second prediction result and the labeled information to obtain the fourth loss value, calculate the total classification loss based on the third loss value and the fourth loss value, and adjust the model parameters of the defect detection model based on the total classification loss; Step S2-2-6: Repeat steps S2-2-1 to S2-2-5 until the exit condition is met.

6. The method for detecting defects in a vibration-based power transformer by fusing acoustic signature signals according to claim 3, characterized in that, The mathematical expression for the total classification loss is: in: For total classification loss, As the first weight, This is the third loss value. As the second weight, This is the fourth loss value. As the third weight, For similarity loss, As the fourth weight, As a complementary loss, For Euclidean distance, For standard information, This is the first prediction result. This is the second prediction result.

7. The method for detecting defects in a vibration-based power transformer by fusing acoustic signature signals according to claim 1, characterized in that, In step S1, each defect type is tested multiple times, and multiple samples are collected.

8. The method for detecting defects in a vibration-based power transformer by fusing acoustic signature signals according to claim 1, characterized in that, In step S1, the number of positive samples is the same as the number of negative samples.

9. A vibration-based power transformer defect detection device integrating acoustic signature signals, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1-8.

Images