A wind turbine yawing abnormal sound fault diagnosis method and system based on a sound signal

Abstract

The application discloses a wind turbine yaw abnormal sound fault diagnosis method and system based on a sound signal, and the method comprises the following steps: S1, obtaining the sound signal of the wind turbine, inputting the sound signal into a Resnet model, and obtaining sound characteristics; S2, inputting the sound characteristics into a Transformer model, and obtaining a wind turbine yaw abnormal sound fault diagnosis result. The Resnet-Transformer model based on the sound signal of the yaw system of the application realizes wind turbine yaw abnormal sound fault diagnosis, the local feature extraction of the sound signal is realized through the residual structure of the Resnet model, the global feature of the sound signal is captured by combining the multi-head attention mechanism of the Transformer, the wind turbine yaw abnormal sound fault diagnosis is realized through the local-global feature fusion, and the fault diagnosis precision is high.

Description

Technical Field

[0001] This invention mainly relates to the field of wind power technology, specifically to a method and system for diagnosing yaw noise faults in wind turbine generators based on sound signals. Background Technology

[0002] Wind power, as an important renewable energy technology, plays a crucial role in energy diversification, emission reduction, sustainable development, and energy security. The yaw system is one of the core components of a wind turbine, ensuring efficient directional wind capture and maximizing wind energy utilization. Due to the uncertainty of wind direction and long-term operation in harsh environments, key components of the yaw system are prone to various failures. Furthermore, abnormal noise from the yaw system is particularly prominent, negatively impacting the surrounding ecological environment and residents' lives, and directly identifying yaw faults through external conditions is extremely difficult. The state of abnormal yaw noise generally requires manual judgment, which is time-consuming and labor-intensive, while traditional signal processing methods have low accuracy. Summary of the Invention

[0003] To address the technical problems existing in the prior art, this invention provides a method and system for diagnosing yaw noise faults in wind turbines based on sound signals, which has high fault diagnosis accuracy.

[0004] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0005] A method for diagnosing yaw noise faults in wind turbines based on sound signals, comprising the following steps:

[0006] S1. Acquire the sound signal of the wind turbine and input it into the ResNet model to obtain the sound features;

[0007] S2. Input the sound features into the Transformer model to obtain the yaw noise fault diagnosis results of the wind turbine.

[0008] The specific process of step S1 is as follows:

[0009] S11. Resample the sound signal from the wind turbine to a fixed sampling rate and convert it to a single channel:

[0010] s′(t)=Resample(s(t),f s )

[0011] Where s(t) represents the original sound signal; f s s is the target sampling rate, and s′(t) is the resampled audio signal;

[0012] S12. Then, trim or pad the signal s′(t) to a fixed duration to ensure that all signals have the same length:

[0013] s″(t)=PadTruncate(s′(t),L)

[0014] Where L is the target signal length, and s″(t) is the time-normalized signal;

[0015] S13. Convert the processed audio signal s″(t) into a Mel spectrogram:

[0016] M(f,t)=log(Mel(|STFT(s″(t))| 2 ))

[0017] STFT(s″(t)) represents the short-time Fourier transform of s″(t), Mel() is the Mel filter applied; M(f,t) is the final Mel spectrum; where f and t represent frequency and time, respectively.

[0018] S14. Input the Mel spectrogram M(f,t) into the ResNet model and use ResNet to extract deep features from the Mel spectrogram:

[0019] F l+1 =H l (F l )+F l

[0020] F l and F l+1 H represents the input and output features of the continuous layer, respectively; l It is the residual learning function.

[0021] Preferably, the specific process of step S2 is as follows:

[0022] The features output by the ResNet model are analyzed using the Transformer model:

[0023]

[0024] Q, K, V are the query, key, and value in the Transformer, respectively. k It is the dimension of the key, used to scale the result of the dot product;

[0025] The output of the Transformer model is classified using a fully connected layer:

[0026] y = softmax(W h h+b]

[0027] h is a high-level feature obtained from the Transformer; W hb and y are the weights and biases of the fully connected layer; y is the prediction result of the yaw noise of the wind turbine.

[0028] Preferably, in step S2, the cross-entropy loss function is used for optimization to calculate the difference between the predicted output y and the actual label c:

[0029]

[0030] c i It is the one-hot encoding of the real tag; p i It is the probability distribution predicted by the model;

[0031] Through repeated iterations, the optimal weights were finally determined, resulting in the most accurate wind turbine yaw noise fault prediction result y:

[0032]

[0033] Preferably, a ResNet-Transformer hybrid model is formed by combining the ResNet model and the Transformer model. During the training of the ResNet-Transformer hybrid model, five-fold cross-validation is used to evaluate the model's generalization ability and stability under different data segmentation conditions.

[0034] Preferably, the evaluation metrics for the ResNet-Transformer hybrid model are training accuracy and training loss.

[0035] Preferably, the sound signal of the wind turbine is acquired using an array microphone, a data acquisition card, and a laptop computer with acquisition software.

[0036] Preferably, after acquiring the sound signal of the wind turbine, it is converted into a sound spectrum signal and displayed.

[0037] The present invention also discloses a computer program product, comprising a computer program that, when executed by a processor, performs the steps of the method described above.

[0038] The present invention further discloses a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0039] The present invention also discloses a wind turbine yaw noise fault diagnosis system based on sound signals, including a memory and a processor connected to each other. The memory stores a computer program, and the computer program executes the steps of the method described above when run by the processor.

[0040] Compared with the prior art, the advantages of the present invention are as follows:

[0041] This invention utilizes a ResNet-Transformer model based on yaw system sound signals to diagnose yaw noise faults in wind turbines. It extracts local features from the sound signal using the residual structure of the ResNet model and captures global features through the Transformer's multi-head attention mechanism. By fusing local and global features, it achieves fault diagnosis for yaw noise in wind turbines. This invention improves the accuracy of yaw noise fault diagnosis in wind turbines (with an identification accuracy exceeding 95%) even with limited data. This is crucial for subsequent rapid and targeted maintenance, improving work efficiency, and provides a new method for fault diagnosis of bearing equipment under actual industrial operation and maintenance conditions.

[0042] This invention uses a non-contact sensor to diagnose faults in key components of wind turbines, such as the yaw system, through sound signals. This is of great significance for improving the service life of wind turbines and reducing the potential harm of yaw noise to surrounding residents. Attached Figure Description

[0043] Figure 1 This is an example diagram of the wind turbine yaw noise fault diagnosis method of the present invention in a specific application.

[0044] Figure 2 This is a time-domain waveform diagram of the normal signal of the wind power yaw system of the present invention.

[0045] Figure 3 This is a time-domain waveform diagram of the abnormal signal of the wind power yaw system of the present invention.

[0046] Figure 4 This is the spectrogram of the normal sound signal according to the present invention.

[0047] Figure 5 This is the spectrogram of the abnormal sound signal in this invention.

[0048] Figure 6 This is a graph showing the accuracy of the model using the five-fold cross-validation method of this invention.

[0049] Figure 7 This is a line graph showing the model loss of the five-fold cross-validation method of this invention.

[0050] Figure 8 This is a confusion matrix diagram of yaw noise fault types according to the present invention.

[0051] Figure 9 This is a comparison chart of the accuracy of yaw noise fault diagnosis for different models in this invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0053] like Figure 1 As shown in the figure, the wind turbine yaw noise fault diagnosis method based on sound signals according to an embodiment of the present invention includes the following steps:

[0054] S1. First, the sound signal from the wind turbine is used as input to the ResNet model. The powerful feature extraction capabilities of the ResNet model are utilized to obtain high-level sound features. The ResNet model, through its unique residual connections, significantly reduces information loss during training and can effectively handle deeper network structures. The specific process is as follows:

[0055] S11. The sound signal acquired by the sensor is first resampled to a fixed sampling rate and converted into a single channel number:

[0056] s′(t)=Resample(s(t),f s )

[0057] Where s(t) represents the original sound signal; f s s is the target sampling rate, and s′(t) is the resampled audio signal;

[0058] S12. Then, trim or pad the signal s′(t) to a fixed duration to ensure that all signals have the same length:

[0059] s″(t)=PadTruncate(s′(t),L)

[0060] Where L is the target signal length, and s″(t) is the time-normalized signal.

[0061] S13. Convert the processed audio signal s″(t) into a Mel spectrogram:

[0062] M(f,t)=log(Mel(|STFT(s″(t))| 2 ))

[0063] STFT(s″(t)) represents the short-time Fourier transform of s″(t), Mel() is the Mel filter applied; M(f,t) is the final Mel spectrum; where f and t represent frequency and time, respectively.

[0064] S14. Input the Mel spectrogram M(f,t) into the hybrid network model and use ResNet to extract deep features from the Mel spectrogram:

[0065] F l+1 =H l (F l )+F l

[0066] F l and F l+1 H represents the input and output features of the continuous layer, respectively; l It is the residual learning function.

[0067] S2. Further analyze the features output by ResNet using the Transformer model:

[0068]

[0069] Q, K, and V are the query, key, and value in Transformer. k It is the dimension of the key, used to scale the result of the dot product.

[0070] The output of the Transformer is classified through a fully connected layer:

[0071] y = softmax(W h h+b]

[0072] h is a high-level feature obtained from the Transformer; W h b and y are the weights and biases of the fully connected layer; y is the prediction result of the yaw noise of the wind turbine.

[0073] Furthermore, the model is optimized using the cross-entropy loss function to calculate the difference between the predicted output y and the actual label c:

[0074]

[0075] c i It is the one-hot encoding of the real tag; p i It is the probability distribution predicted by the model.

[0076] Through repeated iterations, the optimal weights were finally determined, resulting in the most accurate wind turbine yaw noise fault prediction result y:

[0077]

[0078] Specifically, the acoustic signal analysis of the wind turbine yaw system:

[0079] Normal yaw signals from a wind turbine originate from a complete yaw system, while fault signals originate from yaw systems with brake pad or disc wear issues. Time-domain analysis primarily describes the sound characteristics by observing how the signal changes over time. In diagnosing abnormal yaw sounds, the advantage of time-domain representation lies in its ability to visually display the time-series changes of the sound signal, reflecting both the transient and persistent characteristics of the abnormal noise. For example, time-domain waveforms clearly show the signal's start, peak point, and disappearance process, which is extremely useful for identifying events such as abnormal impacts, friction, or fractures.

[0080] Figure 2 The time-domain waveform of the normal signal shows the sound signal collected by the wind turbine during normal yaw. The time-domain graph shows that the signal amplitude is relatively stable, the waveform is continuous and there are no prominent spikes, indicating that the unit vibrates less during normal operation and there are no sudden impacts or wear phenomena.

[0081] Figure 3 In contrast, the time-domain plot of an abnormal signal shows obvious amplitude fluctuations, especially a sudden increase in amplitude at certain times. This phenomenon may indicate mechanical failures such as loose parts, damaged bearings, or poor lubrication.

[0082] Acoustic spectrum analysis is a visualization tool that combines time-domain and frequency-domain information. It displays the changes in the energy of a sound signal over time and frequency in a graphical form. Its advantage lies in simultaneously presenting the dynamic and static characteristics of the sound signal, allowing us to see the changes in frequency components over different time periods and to grasp the overall spectral structure of the entire audio segment. For diagnosing abnormal noises in yaw systems, acoustic spectrum signals help to identify periodic, nonlinear, or sudden noise patterns, providing an intuitive and comprehensive insight into complex abnormal noise phenomena.

[0083] Figure 4 The spectrogram of a normal sound signal is shown. Figure 4 The energy is low in the low-to-mid frequency region and uniform in the mid-to-high frequency region, with clear spectral lines, reflecting the frequency response of the yaw system under normal conditions.

[0084] Figure 5 The spectrogram of the abnormal signal shows a significant increase in overall spectral energy, particularly a concentration of energy in the low-frequency region, which is typically associated with increased vibrations related to mechanical faults. This change in the spectrum may be caused by non-periodic vibrations resulting from bearing damage or imbalance.

[0085] This invention utilizes a ResNet-Transformer model based on yaw system sound signals to diagnose yaw noise faults in wind turbines. It extracts local features from the sound signal using the residual structure of the ResNet model and captures global features through the Transformer's multi-head attention mechanism. By fusing local and global features, it achieves fault diagnosis for yaw noise in wind turbines. This invention improves the accuracy of yaw noise fault diagnosis in wind turbines (with an identification accuracy exceeding 95%) even with limited data. This is crucial for subsequent rapid and targeted maintenance, improving work efficiency, and provides a new method for fault diagnosis of bearing equipment under actual industrial operation and maintenance conditions.

[0086] This invention uses a non-contact sensor to diagnose faults in key components of wind turbines, such as the yaw system, through sound signals. This is of great significance for improving the service life of wind turbines and reducing the potential harm of yaw noise to surrounding residents.

[0087] Experimental verification:

[0088] 1. Data Preparation

[0089] In the study of condition monitoring and fault diagnosis of the yaw system, acoustic signal acquisition was conducted. Acoustic data was collected covering normal equipment operation and various known abnormal states, ensuring the sample encompasses a wide range of operating conditions and comprehensively presents the acoustic performance of the yaw system under different health conditions.

[0090] The wind farm area where the abnormal yaw noise data of the wind turbine is located is characterized by high mountains and steep slopes with significant topographic relief, generally ranging from 15° to 35°, and in some areas even steep cliffs. The data acquisition system consists of an array microphone, a data acquisition card, and a laptop computer with LabVIEW acquisition software.

[0091] 2. Model and Evaluation Indicators

[0092] During training, the Adam optimizer was used with a learning rate of 0.0001, and the cross-entropy loss function was chosen, suitable for binary classification tasks. The training run consisted of 50 epochs. Five-fold cross-validation was employed to evaluate the model's generalization ability and stability under different data splits. Five-fold cross-validation is a technique for evaluating model performance. It involves uniformly splitting the original dataset into five mutually exclusive subsets, using one subset as the validation set and the other four subsets for training the model, repeating this cycle five times. This ensures the model is validated on five different subsets, guaranteeing good generalization ability across diverse data subsets, rather than just performing well on a single training set. This evaluation method helps obtain robust estimates of model performance and effectively tests the model's generalization ability on unknown data, thereby reducing the risk of overfitting and improving the model's stability and reliability in real-world scenarios. Therefore, this experiment adopted this method for validation.

[0093] In terms of evaluation metrics, training accuracy and training loss were used as the main performance evaluation standards. The training results for each training epoch were recorded in a file, and the trends of training accuracy and loss at different epochs were displayed as curves. In addition, the model's classification performance on the validation set was visualized by plotting a confusion matrix to further verify the model's accuracy and reliability in identifying normal and abnormal samples.

[0094] 3. Test Results and Analysis

[0095] Figure 6 The model recognition accuracy of wind turbine yaw noise signals after 50 training iterations is shown. The data is divided into five folds for cross-validation. It can be seen that the initial accuracy is very low at different fold numbers, but the accuracy in fold 2 is relatively high compared to other folds. This is because the data was randomly split during training and validation, and there are differences in the data types of different folds. Figure 6 After cross-validation, all the folds shown in the figure have a relatively high accuracy rate in diagnosing yaw noise from sound signals, with the highest being 96.88%. This indicates that the Resnet-Transformer model of the present invention can accurately diagnose yaw noise from wind turbines based on sound signals.

[0096] Figure 7 This is a line graph showing the training loss of the model after five-fold cross-validation. From... Figure 7As can be observed, with the increase of training epochs, the loss values ​​of all five folds show a significant decreasing trend, indicating that the model gradually learns the features of the data and optimizes its performance. Specifically, from the 1st epoch to the 50th epoch, the loss value of each fold gradually decreases from approximately 0.8 to approximately 0.2, with a minimum of 0.0521, demonstrating that the model has good generalization ability and learning efficiency for different data subsets. Furthermore, the loss curves between each fold are relatively close at most time points, indicating that the model's fit to various data subsets is consistent, without overfitting or underfitting. This shows that the ResNet-Transformer-based model of this invention demonstrates good consistency in feature extraction and training of audio data.

[0097] Figure 8 A confusion matrix diagram for identifying yaw noise faults in wind turbines. Figure 8 The image shows the classification results of the wind turbine yaw noise fault diagnosis model. In the matrix, the true and predicted labels are marked as 0 and 1, respectively, where "0" represents the normal state and "1" represents the fault state. Figure 8 As can be seen, the model performs well in fault diagnosis for label 1, correctly identifying 9 fault states as faults, but also misclassifying 3 normal states as faults. For label 0, the model correctly identified 13 normal states, but also misclassified 1 normal state as a fault. Overall, the model has high accuracy in identifying normal states, but also exhibits some misclassification when distinguishing fault states. This may be due to the model's high sensitivity or insufficient training data to cover all fault details.

[0098] Figure 9 Box plots show the accuracy of wind turbine yaw noise diagnosis results for Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and ResNet-Transformer after multiple training iterations. From... Figure 9 As can be seen, the method proposed in this invention has a significant advantage in model stability compared to CNN and LSTM. In terms of fault diagnosis accuracy, the method of this invention also achieves higher accuracy compared to the other two methods. Therefore, it can be used as a robust model for fault diagnosis of wind turbine yaw systems.

[0099] This invention also discloses a computer program product, including a computer program that, when run by a processor, performs the steps of the method described above. This invention further discloses a computer-readable storage medium storing a computer program that, when run by a processor, performs the steps of the method described above. This invention also discloses a wind turbine yaw noise fault diagnosis system based on sound signals, including an interconnected memory and a processor, wherein the memory stores a computer program that, when run by a processor, performs the steps of the method described above. The products, media, and systems of this invention, corresponding to the methods described above, also possess the advantages described above.

[0100] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0101] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for diagnosing yaw noise faults in wind turbines based on sound signals, characterized in that, Including the following steps: S1. Acquire the sound signal of the wind turbine and input it into the ResNet model to obtain the sound features; S2. Input the sound features into the Transformer model to obtain the yaw noise fault diagnosis results of the wind turbine. The specific process of step S1 is as follows: S11. Resample the sound signal from the wind turbine to a fixed sampling rate and convert it to a single channel: in Represents the original sound signal; It is the target sampling rate. It is the resampled sound signal; S12, then the signal Trim or pad to a fixed duration to ensure all signals have the same length: in It is the length of the target signal. It is a signal after duration standardization; S13, Process the audio signal Convert to Mel spectrogram: Indicates to The short-time Fourier transform was performed. It uses a Mel filter; This is the final Mel spectrum; in which and Representing frequency and time respectively; S14, Mel spectrum diagram As input to the ResNet model, ResNet is used to extract deep features from the Mel spectrogram: and These represent the input and output features of the continuous layer, respectively. It is the residual learning function; The specific process of step S2 is as follows: The features output by the ResNet model are analyzed using the Transformer model: , , It refers to queries, keys, and values ​​in Transformer. It is the dimension of the key, used to scale the result of the dot product; The output of the Transformer model is classified using a fully connected layer: These are high-level features obtained from the Transformer; and These are the weights and biases of the fully connected layer; This is the predicted result of abnormal yaw noise from the wind turbine.

2. The method for diagnosing yaw noise faults in wind turbines based on sound signals according to claim 1, characterized in that, In step S2, the cross-entropy loss function is used for optimization to calculate the predicted output. With actual label Differences between them: It is the unique hot encoding of the real label; It is the probability distribution predicted by the model; Through repeated iterations, the optimal weights were finally determined, resulting in the most accurate wind turbine yaw noise fault prediction result y: 。 3. The method for diagnosing yaw noise faults in wind turbines based on sound signals according to claim 1 or 2, characterized in that, A ResNet-Transformer hybrid model is constructed by combining the ResNet and Transformer models. During the training of the ResNet-Transformer hybrid model, five-fold cross-validation is used to evaluate the model's generalization ability and stability under different data partitioning conditions.

4. The method for diagnosing yaw noise faults in wind turbines based on sound signals according to claim 3, characterized in that, The evaluation metrics for the ResNet-Transformer hybrid model are training accuracy and training loss.

5. The method for diagnosing yaw noise faults in wind turbines based on sound signals according to claim 1 or 2, characterized in that, The sound signals of the wind turbine are acquired using an array microphone, a data acquisition card, and a laptop with acquisition software.

6. The method for diagnosing yaw noise faults in wind turbines based on sound signals according to claim 1 or 2, characterized in that, After acquiring the sound signal from the wind turbine, it is converted into a sound spectrum signal and displayed.

7. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor to perform the steps of the method as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-6.

9. A wind turbine yaw noise fault diagnosis system based on sound signals, comprising a memory and a processor interconnected, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-6.

Images