A wind turbine gearbox fault identification method and system for class imbalance
By preprocessing and data augmentation of the vibration signal of the wind turbine gearbox, high-fidelity fault samples are generated, and a balanced sample set is constructed for model training. This solves the class imbalance problem in wind turbine gearbox fault identification and improves the robustness and accuracy of fault identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA ENERGY ENG GRP GUANGDONG ELECTRIC POWER DESIGN INST CO LTD
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-21
AI Technical Summary
Existing wind turbine gearbox fault identification methods rely on the distribution of training data, resulting in poor fault identification performance in imbalanced scenarios, especially in the difficulty of identifying early and weak faults, and insufficient robustness and accuracy.
Vibration signals from wind turbine gearboxes under different operating conditions were collected and preprocessed to construct an initial vibration dataset. High-fidelity fault samples were generated using an adversarial training mechanism between the generator and discriminator. Noise was removed by combining filtering techniques and wavelet transform. A balanced sample set was constructed for model training, and a deep learning model was used for fault identification.
It effectively alleviates the problem of scarce fault samples, improves the robustness and accuracy of fault identification, enhances the identification accuracy for different fault types, and solves the problem of fault identification in class imbalance scenarios.
Smart Images
Figure CN122432896A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of electrical equipment testing technology, specifically relating to a method and system for identifying gearbox faults in wind turbines with class-unbalanced operation. Background Technology
[0002] As a critical transmission component of wind turbines, the gearbox's operating condition directly affects the safety and stability of the entire unit. Due to the complex and variable working environment and large fluctuations in operating load of wind turbines, gearboxes are highly susceptible to wear, cracks, and tooth breakage. Once a failure occurs, it not only leads to downtime for maintenance but may also trigger more serious chain reactions, causing huge economic losses.
[0003] Currently, fault identification methods for wind turbine gearboxes mainly rely on vibration signal analysis and deep learning classification. These methods involve real-time acquisition of vibration signals from the gearbox under different operating conditions, followed by feature extraction and classification using a classifier. The output signal category determines the gearbox fault type, such as normal, gear wear, tooth cracks, or tooth breakage. However, these existing technologies have significant shortcomings. Their fault identification performance is highly dependent on the distribution of the training data. In actual wind farms, gearbox fault data is extremely scarce, especially early, weak fault samples, which are even harder to obtain. This results in a distribution where normal samples far outnumber fault samples, leading to severe class imbalance in the training set. This causes the classifier to favor the normal class, making it difficult to identify the few fault patterns, thus reducing the robustness and accuracy of fault identification. Summary of the Invention
[0004] This application proposes a fault identification method and system for wind turbine gearboxes oriented towards class-unbalanced systems, which can solve the problem that the lack of fault data in wind turbine gearboxes in the prior art leads to poor robustness and accuracy of fault identification.
[0005] The first aspect of this application provides a method for identifying gearbox faults in wind turbines with class-unbalanced gearboxes, the method comprising: Within a preset historical period, the original vibration signals of the wind turbine gearbox under normal and fault conditions are collected, and the original vibration signals are preprocessed to obtain effective vibration signals. Based on the fault operation and maintenance records, the effective vibration signals are labeled with corresponding operating condition tags by matching the operating time period. Based on the operating condition tags and the feature values of the effective vibration signals, an initial wind turbine vibration dataset is constructed. A wind turbine fault diagnosis model is obtained by training a pre-defined deep learning model using a pre-defined balanced sample set; wherein, the balanced sample set includes a fault sample set generated based on an initial wind turbine vibration dataset. The effective vibration signal is processed by the wind turbine fault diagnosis model, and the fault identification result of the wind turbine gearbox is output.
[0006] The above scheme collects vibration data of wind turbine gearboxes under various operating conditions. First, the vibration data is preprocessed to effectively remove outliers caused by sensor interference, abnormal acquisition, or instantaneous impacts, preventing these outliers from interfering with the overall signal characteristics and improving data quality. Then, by labeling the vibration signals with operating conditions, the vibration signals inherently carry corresponding category information, eliminating the need for manual secondary labeling in the model and improving fault identification accuracy. Furthermore, data augmentation learning is used to study the temporal variation patterns of vibration signals in the initial wind turbine vibration dataset, generating high-fidelity fault samples that more closely resemble the actual operating state of wind turbines in terms of waveform morphology and evolution trends. This effectively expands the number of fault samples, alleviating the problem of scarce wind turbine fault samples and solving the class imbalance problem. Using a balanced sample set for model training improves the robustness and generalization ability of the wind turbine fault diagnosis model in complex class imbalance scenarios, enhancing the identification accuracy for different fault types.
[0007] In one possible implementation of the first aspect, within a preset historical time period, raw vibration signals of the wind turbine gearbox under normal and fault conditions are collected, and the raw vibration signals are preprocessed to obtain effective vibration signals, specifically as follows: Accelerometers are installed at several locations on the gearbox of the wind turbine to collect the original vibration signals; the historical period is the past six months. By combining filtering techniques and wavelet transform, and by identifying abnormal abrupt changes, environmental noise and invalid vibration frequency bands are removed from the original vibration signal to obtain an effective vibration signal. The abnormal abrupt changes are caused by factors other than gearbox bearing failure, including sudden changes in operating conditions, sensor interference, and unit start-up and shutdown shocks.
[0008] The above scheme purifies the collected raw vibration signals, removes environmental noise interference and invalid vibration frequency bands from the original data, avoids the impact of non-physical abrupt changes on the overall characteristics of the signal, improves the reliability of vibration signals under complex working conditions, and provides more stable and effective input data for subsequent fault identification.
[0009] In one possible implementation of the first aspect, by combining filtering techniques and wavelet transform, and by identifying anomalous abrupt changes, environmental noise and invalid vibration frequency bands are removed from the original vibration signal to obtain an effective vibration signal, specifically: The abnormal abrupt change value of the original vibration signal is replaced within a sliding window using median filtering technology, and the original vibration signal after replacement is smoothed using moving average filtering. By using wavelet transform technology, the smoothed original vibration signal is decomposed into multiple scales and denoised by thresholding. Bearing fault-related features are separated from the original vibration signal to obtain an effective vibration signal.
[0010] The above scheme uses a preprocessing method that combines median filtering, moving average filtering, and wavelet transform to ensure that the overall trend of the signal is not destroyed when removing abnormal abrupt values, and to reconstruct a vibration signal with a high signal-to-noise ratio. This avoids feature distortion caused by single filtering and improves the reliability of vibration signals under complex working conditions.
[0011] In one possible implementation of the first aspect, based on the fault operation and maintenance records, the effective vibration signal is labeled with a corresponding operating condition tag by matching the operating time period. An initial wind turbine vibration dataset is constructed based on the operating condition tag and the feature values of the effective vibration signal, specifically: Based on the operating conditions provided by the fault operation and maintenance records, the effective vibration signal is labeled with a corresponding operating condition tag by matching the time period of the operating condition with the running time period of the effective vibration signal; wherein, the operating conditions include normal state, tooth root fracture, tooth root missing, tooth root damaged and tooth surface worn. The RMS value, peak value, peak factor, kurtosis value, waveform factor, impulse factor, and spectral entropy value of the effective vibration signal at each moment are calculated to obtain the characteristic values. An initial wind turbine vibration dataset is constructed based on the operating condition labels and the feature values.
[0012] The above scheme assigns a corresponding operating condition label to each valid vibration signal, thereby identifying the signal type. The calculated feature values are then used in subsequent data augmentation to make the generated fault samples more statistically and physically similar to real wind turbine vibration signals, thus improving the reliability of the generated data.
[0013] In one possible implementation of the first aspect, a wind turbine fault diagnosis model is obtained by training a preset deep learning model using a preset balanced sample set, specifically as follows: The initial wind turbine vibration dataset is input into a preset data augmentation model to construct the potential temporal features of the effective vibration signal through a generator, capture the local waveform features of the effective vibration signal through a discriminator, and generate the fault sample based on the output results of the generator and the discriminator. The fault samples and the initial wind turbine vibration dataset are merged into the balanced sample set. The classifier of the deep learning model is iterated using the balanced sample set to obtain the trained wind turbine fault diagnosis model.
[0014] In one possible implementation of the first aspect, the fault sample is generated based on the outputs of the generator and the discriminator, specifically as follows: Random noise vectors from the initial wind turbine vibration dataset are extracted as input to the generator. Within the generator, based on the learning objective output by the discriminator, the latent temporal features of the effective vibration signal are constructed through a preset temporal network. Then, a fully connected layer is used to map the latent temporal features into a fixed-length vibration time-domain signal to obtain the fault sample. The discriminator captures the local waveform characteristics, impact modes, and frequency-related information of the effective vibration signal. Based on the capture results and preset consistency constraints, the difference between the real vibration signal and the vibration time-domain signal is learned, and the learning target is obtained.
[0015] The above scheme takes into account the characteristics of the vibration signal of the wind turbine gearbox, such as strong time-series correlation, non-stationarity and operating condition dependence. Therefore, a time-series modeling structure is introduced into the data augmentation model, which enables the model to learn the change law of the vibration signal in the time dimension, thereby generating vibration signals that are closer to the actual wind turbine operating state in terms of waveform shape and evolution trend, effectively alleviating the problem of scarce fault samples of wind turbines.
[0016] One possible implementation of the first aspect also includes: The loss function of the data augmentation model is constructed based on adversarial loss, engineering feature loss, frequency domain loss, and envelope spectrum loss. Based on the loss function, the feature values of the effective vibration signal are introduced into the data augmentation model to improve the similarity between the real vibration signal and the vibration time-domain signal.
[0017] The above scheme combines various constraints to construct a loss function, which makes the generated fault samples not only reasonable at the numerical level, but also retains key information related to mechanical structure and fault mechanism, providing high-quality samples for subsequent model training.
[0018] The second aspect of this application provides a fault identification system for wind turbine gearboxes oriented towards class-unbalanced systems. The system includes: a data acquisition module, a data annotation module, a model training module, and a fault identification module. The data acquisition module is used to collect the original vibration signals of the wind turbine gearbox under normal and fault conditions within a preset historical period, and to preprocess the original vibration signals to obtain effective vibration signals. The data labeling module is used to label the effective vibration signal with the corresponding operating condition label by matching the operating time period according to the fault operation and maintenance record, and to construct an initial wind turbine vibration dataset based on the operating condition label and the feature value of the effective vibration signal. The model training module is used to train a preset deep learning model using a preset balanced sample set to obtain a wind turbine fault diagnosis model; wherein, the balanced sample set is generated by data augmentation of the initial wind turbine vibration dataset; The fault identification module is used to process the effective vibration signal through the wind turbine fault diagnosis model and output the fault identification result of the wind turbine gearbox.
[0019] A third aspect of this application provides a terminal device, the device comprising: a terminal device including a processor and a memory, the memory storing a computer program, wherein the processor executes the computer program to implement the steps of the method for identifying gearbox faults in wind turbines oriented towards class imbalance, as described in any one of the embodiments of this application.
[0020] A fourth aspect of this application provides a storage medium that stores computer-readable program code, which, when executed, implements the steps of a method for identifying gearbox faults in wind turbines oriented towards class imbalance, as described in any one of the embodiments of this application. Attached Figure Description
[0021] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0022] Figure 1 This is a schematic flowchart of a method for identifying gearbox faults in wind turbines with class-unbalanced gearboxes, provided in one embodiment of this application. Figure 2 This is a time-domain comparison result of a method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems, provided in an embodiment of this application. Figure 3 This is a frequency domain comparison result of a method for identifying gearbox faults in wind turbines with class-unbalanced gearboxes provided in an embodiment of this application; Figure 4 This is a confusion matrix diagram of the classifier for a wind turbine gearbox fault identification method for class imbalance provided in an embodiment of this application on the original dataset; Figure 5This is a confusion matrix diagram of a classifier for a wind turbine gearbox fault identification method for class-imbalanced systems provided in an embodiment of this application on a balanced dataset; Figure 6 This is a structural diagram of a wind turbine gearbox fault identification system for class-unbalanced systems provided in an embodiment of this application; Figure 7 This is a structural diagram of a terminal device provided in an embodiment of this application. Detailed Implementation
[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0024] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.
[0025] First Embodiment In actual wind farm operation, the probability of gearbox failure is relatively low, and the number of fault samples that can be collected, especially early and weak fault samples, is extremely limited, resulting in a severe shortage of training data and making it difficult to support the effective training of deep learning models. One of the objectives of this application is to utilize an adversarial training mechanism between the generator and the discriminator to synthesize fault samples that are highly similar to the real distribution, thereby effectively alleviating the problem of scarce fault data.
[0026] like Figure 1 As shown, to address the problem of scarce fault samples in wind turbine gearboxes in existing technologies, which affects the robustness and accuracy of fault identification, the first embodiment of this application provides a detailed flowchart of a wind turbine gearbox fault identification method for class-unbalanced systems. This embodiment's wind turbine gearbox fault identification method for class-unbalanced systems includes steps S1 to S4, detailed below: Step S1: Within a preset historical time period, collect the original vibration signals of the wind turbine gearbox under normal and fault conditions, and preprocess the original vibration signals to obtain effective vibration signals.
[0027] Accelerometers are installed at several locations on the wind turbine gearbox, such as the bearing housing and the gearbox casing, to collect raw vibration signals of the gearbox during operation.
[0028] Optionally, in this embodiment of the application, the original vibration signals of the wind turbine gearbox under normal and fault conditions over the past six months are selected.
[0029] To address the environmental noise interference and invalid vibration frequency bands in the original vibration signal, noise reduction processing is performed on the original vibration signal using moving average filtering, median filtering, and wavelet transform. Finally, the signal data is normalized to obtain the effective vibration signal.
[0030] For abnormal abrupt changes in the original vibration signal caused by factors other than gearbox bearing failure (such as sudden changes in operating conditions, sensor interference, start-stop impact, etc.), a denoising method combining moving average filtering, median filtering, and wavelet transform is adopted. This method avoids feature distortion caused by single filtering and improves the reliability of vibration signals under complex operating conditions, providing more stable and effective input data for subsequent wind turbine fault diagnosis models.
[0031] Specifically, the original vibration signal is first subjected to median filtering to replace any abnormal abrupt changes identified within the sliding window (replacing the abnormal abrupt changes with the median). This prioritizes the removal of pulse-type and isolated spikes caused by acquisition anomalies or instantaneous impacts, avoiding the impact of non-physical abrupt changes on the overall signal characteristics. Next, moving average filtering is used to smooth the replaced original vibration signal, suppressing high-frequency random vibrations caused by wind condition changes and minor load fluctuations, while maintaining the overall signal trend. Based on the filtered signal, wavelet transform is applied to the smoothed signal. Through multi-scale decomposition and threshold denoising, non-stationary noise and bearing fault-related features are effectively separated from the original signal, and a high signal-to-noise ratio vibration signal is reconstructed to output an effective vibration signal.
[0032] Therefore, the denoising method combining moving average filtering, median filtering, and wavelet transform adopted in this application embodiment is progressive from coarse to fine, which not only improves the signal-to-noise ratio but also preserves the effective information related to mechanical state and fault mechanism to the greatest extent, thereby improving the data quality of vibration signal.
[0033] Step S2: Based on the fault operation and maintenance records, the effective vibration signal is labeled with the corresponding operating condition label by matching the operating time period. Based on the operating condition label and the feature value of the effective vibration signal, an initial wind turbine vibration dataset is constructed.
[0034] In order to classify the effective vibration signals and assist the subsequent wind turbine fault diagnosis model in fault identification, this application embodiment uses the operating conditions provided by the fault operation and maintenance records to match the operating conditions of the corresponding time period and mark the effective vibration signals with the corresponding operating condition labels.
[0035] The operating conditions include normal condition, tooth root fracture, tooth root loss, tooth root damage, and tooth surface wear (four typical faults).
[0036] First, the fault maintenance records of the wind turbine gearbox, recorded by the on-site maintenance personnel of the wind farm, are exported from the auxiliary control system of the wind turbine. Then, based on the operating conditions of the gearbox at different time periods, the effective vibration signal is labeled with the corresponding operating condition tag by matching the time period of the operating condition with the running time period of the effective vibration signal, thereby realizing the category labeling of the vibration data at each moment.
[0037] In addition, the RMS value, peak value, peak factor, kurtosis value, waveform factor, impulse factor and spectral entropy value of the effective vibration signal at each time moment are calculated to obtain the characteristic values of the effective vibration signal.
[0038] Finally, an initial wind turbine vibration dataset is constructed based on the operating condition labels and the feature values.
[0039] Step S3: Train the preset deep learning model using a preset balanced sample set to obtain the wind turbine fault diagnosis model.
[0040] To address the class imbalance problem caused by scarce fault data, embodiments of this application utilize an adversarial training mechanism between the generator and discriminator to synthesize fault samples that are highly similar to the real distribution. Therefore, the data augmentation model used includes both a generator and a discriminator.
[0041] After inputting the initial wind turbine vibration dataset into a preset data augmentation model, the latent temporal characteristics of the effective vibration signal are modeled through a multi-layer temporal network within the generator. Then, a fully connected layer maps these latent temporal characteristics into a fixed-length vibration time-domain signal, thereby obtaining a highly reliable fault sample. The temporal network is constructed using an LSTM structure, which can effectively learn the implicit temporal correlation and dynamic evolution laws in the vibration signal, making it suitable for describing the non-stationary, strongly temporal characteristics of wind turbines under complex operating conditions. The fully connected layer performs a nonlinear mapping of the high-dimensional temporal characteristics, making the generated signal approximate the real vibration data in terms of amplitude distribution and overall trend.
[0042] Furthermore, the activation function used by the generator can limit the output data to a standardized range, which is beneficial for maintaining a numerical scale consistent with real data. The generator provided in this application not only focuses on the distribution of point values, but also learns the temporal dependencies to make the generated vibration signal have high physical rationality in terms of waveform structure, periodicity, and random disturbance characteristics, making it suitable for vibration data enhancement of wind turbine gearboxes.
[0043] In addition to the generator, the discriminator of the data augmentation model mainly provides learning targets for the generator, further improving the authenticity and reliability of the generated samples.
[0044] The discriminator employs a one-dimensional convolutional neural network structure to extract and discriminate features from the input vibration data layer by layer. Through the convolutional layers of the discriminator, it captures the local waveform features, impact modes, and frequency-related information (collectively referred to as statistical features) of the effective vibration signal, making it particularly suitable for identifying typical patterns in wind turbine vibration caused by structural excitation and mechanical characteristics. As the random convolutional neural network deepens, the discriminator progressively captures and extracts higher-level statistical features, and distinguishes between real and generated signals through preset consistency constraints. It learns the differences between real and time-domain vibration signals, constraining the amplitude distribution of the generated signal and improving the authenticity of the generated vibration signal from a structural feature perspective, thus achieving the learning objective. Furthermore, the inclusion of a gradient penalty mechanism enhances the stability of the discriminator during training, avoiding gradient vanishing or mode collapse problems, thereby providing a more reliable learning objective for the generator.
[0045] Therefore, the data augmentation model provided in this application embodiment has been significantly enhanced in terms of structure and loss design to address the characteristics of vibration signals from wind turbine gearboxes. First, the generator incorporates a time-series network, enabling the model to characterize the time dependence of vibration signals and overcoming the difficulty of traditional models in modeling long-term time-series data. Second, in addition to the adversarial loss of the adversarial training mechanism, the data augmentation model also integrates engineering feature loss, frequency domain loss, and envelope spectrum loss to construct the model's loss function. Based on this loss function, commonly used characteristics in vibration signal analysis, such as RMS, peak value, kurtosis, and spectral properties, are directly introduced into the training process of the data augmentation model. This makes the generated data more statistically and physically similar to real wind turbine vibration signals, improving the similarity between real vibration signals and the vibration time-domain signals.
[0046] The engineering feature loss, frequency domain loss, and envelope spectrum loss are used to measure the differences between the generated vibration signal and the real signal in different feature spaces. Specifically, the engineering feature loss compares the RMS, peak value, kurtosis, and other statistical indicators of the real and generated signals to ensure that the generated data closely approximates the real signal in terms of engineering statistical characteristics such as energy level and impact characteristics. The frequency domain loss constrains the frequency structure and energy distribution of the generated signal by comparing the differences in their spectra (e.g., FFT amplitude spectra). The envelope spectrum loss, based on envelope analysis, constrains the envelope spectrum to ensure that the generated signal retains the impact period and characteristic frequencies common in rotating machinery faults such as bearings. In contrast, the adversarial loss comes from adversarial training between the generator and the discriminator, aiming to make the generated data difficult for the discriminator to distinguish as "fake data" in terms of overall distribution. Therefore, the adversarial loss primarily ensures the authenticity of the data distribution, while the engineering feature loss, frequency domain loss, and envelope spectrum loss are explicit constraints based on signal analysis mechanisms, used to ensure that the generated data is consistent with the real wind turbine vibration signal in both statistical characteristics and physical meaning.
[0047] The fault samples output by the data augmentation model are merged with the initial wind turbine vibration dataset to form a balanced sample set. The balanced sample set is then divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The training set and the validation set are input into a preset deep learning model for model training to obtain a trained wind turbine fault diagnosis model.
[0048] The classifier of the wind turbine fault diagnosis model is constructed using a convolutional neural network based on residual connections. The residual block structure enables deep mapping of fault features, effectively avoiding the vanishing gradient problem in deep network training while significantly improving the extraction accuracy of weak fault features. The classifier iterates its parameters on the balanced sample set, and finally outputs the fault type identification result and the corresponding fault risk probability through a Softmax layer, achieving highly robust diagnosis in imbalanced scenarios.
[0049] Step S4: Process the effective vibration signal using the wind turbine fault diagnosis model and output the fault identification result of the wind turbine gearbox, including the fault type / operating condition and fault risk probability of the wind turbine gearbox.
[0050] Implementing the embodiments of this application has the following beneficial effects: This application collects vibration data of wind turbine gearboxes under various operating conditions. The vibration data is preprocessed to effectively remove outliers caused by sensor interference, abnormal acquisition, or instantaneous impacts, preventing these outliers from interfering with the overall signal characteristics and improving data quality. Then, by labeling the vibration signals with operating conditions, the vibration signals inherently carry corresponding category information, eliminating the need for manual secondary labeling in the model and improving fault identification accuracy. Furthermore, data augmentation learning is used to study the temporal variation patterns of vibration signals in the initial wind turbine vibration dataset, generating high-fidelity fault samples that more closely resemble the actual operating state of the wind turbine in terms of waveform morphology and evolution trend. This effectively expands the number of fault samples, alleviating the problem of scarce wind turbine fault samples and solving the class imbalance problem. Using a balanced sample set for model training improves the robustness and generalization ability of the wind turbine fault diagnosis model in complex class imbalance scenarios, enhancing the identification accuracy for different fault types.
[0051] Second Embodiment Furthermore, in order to demonstrate the beneficial effects that the above method embodiments can achieve, the embodiments of this application evaluate the fault samples generated by the fault enhancement model to characterize the high fidelity of the fault samples.
[0052] In this embodiment of the application, the generated fault samples are evaluated using three metrics: peak signal-to-noise ratio (PSNR), Pearson correlation coefficient (PCC), and signal-to-noise ratio (SNR) to measure the high fidelity of the fault samples and the similarity between them and the real vibration signals.
[0053] Table 1 below presents the evaluation results for four failure types: tooth root fracture (B1), missing tooth at the tooth root (M1), tooth root damage (R1), and tooth surface wear (W1). Table 1. Evaluation Results of Indicators As shown in the table, the fault enhancement model constructed in this application demonstrates excellent performance in expanding scarce fault data, generating high-fidelity samples that are highly similar to real fault signals with extremely low distortion. The Pearson correlation coefficient (PCC) between the generated data and the real data for all four fault types (R1, B1, M1, W1) reached above 0.96, indicating that the signals generated by the model maintain a high degree of consistency with the real signals in terms of waveform structure and dynamic trends, successfully reproducing key fault characteristics. In terms of peak signal-to-noise ratio (PSNR), the PSNR values of all generated samples are at a high level. A high PSNR value means that the effective information in the generated signal is far stronger than noise and distortion, proving the purity and high quality of the generated data.
[0054] This application also provides embodiments that... Figure 2 The diagram illustrates a time-domain comparison between generated fault samples and real data. The red line represents the vibration signal of the fault samples, the blue line represents the real vibration signal, the vertical axis represents amplitude, and the horizontal axis represents data points at different times. Figure 2 As can be seen, for signals with obvious transient impact characteristics (as shown in the two figures above), the fault samples generated by the fault enhancement model achieve a high degree of consistency with the real vibration signal (blue curve) in terms of amplitude range, signal baseline fluctuation, and the shape and location of key impact events. For more complex and random signals (as shown in the two figures below), the fault samples do not replicate the original signal point by point, but exhibit a very high consistency with the real vibration in terms of overall distribution and dynamic characteristics. The experimental results intuitively demonstrate that the fault enhancement model provided in this application can successfully learn the inherent distribution law and complex temporal dependencies of real time series data, generating high-quality synthetic samples with both fidelity and diversity, verifying the effectiveness and superiority of the model.
[0055] Furthermore, in order to verify the generation effect of the model from the frequency domain perspective, Figure 3 The spectrum comparison between the generated signal and the real signal is shown. Figure 3 x and y coordinates and Figure 2 The results clearly show that the spectrum of the generated signal (red curve) and the spectrum of the real signal (blue curve) exhibit a high degree of consistency in key features. First, the fault enhancement model accurately reproduces the main frequency bands where the signal energy is concentrated and successfully captures several key characteristic frequency peaks in the spectrum. Both the frequency position of the peaks and their relative amplitudes are highly consistent with the real data. Second, the overall envelope shape and spectral structure of the generated spectrum are also very similar to the real spectrum, indicating that the fault enhancement model has learned the deep frequency composition rules of the signal, rather than just the surface time-domain morphology.
[0056] Therefore, the frequency domain analysis result is consistent with... Figure 2 The time-domain waveform analysis results corroborate each other, jointly proving that the proposed fault enhancement model can not only generate signals with realistic morphology in the time domain, but also ensure that it has high fidelity in the frequency domain structure, laying a solid data foundation for subsequent tasks such as feature extraction and fault identification.
[0057] To further verify the effectiveness of the data augmentation model in addressing the common problems of scarce fault samples and imbalanced data classes in practical industrial applications, the classifier of the wind turbine fault diagnosis model was trained and evaluated using both the original dataset and the balanced dataset generated by the model. The final performance of the classifier was determined by... Figure 4 and Figure 5 The confusion matrix is displayed intuitively.
[0058] Figure 4 The performance of a classifier trained on the original dataset is demonstrated. In the confusion matrix shown, the horizontal axis represents the predicted labels of the data, and the vertical axis represents the true labels. The results show that the classifier exhibits severe classification bias. While it achieves a relatively high recognition rate for normal state samples, it almost completely loses its ability to distinguish between the other four fault categories (B1, M1, R1, W1), incorrectly classifying the vast majority of them as normal states. This phenomenon is a typical consequence of data imbalance: because the number of fault samples in the training data is far less than the number of normal samples, the classifier, in order to achieve optimal overall accuracy, tends to predict all unknown inputs as the category with the highest proportion, thus failing to build an effective fault diagnosis model.
[0059] Figure 5 This paper demonstrates the performance of the wind turbine fault diagnosis model provided in this application after data generation. The horizontal axis represents the predicted labels of the data, and the vertical axis represents the true labels of the data. By synthesizing a large number of high-fidelity samples for scarce fault categories, a class-balanced training set is constructed. The classifier trained on this dataset shows improved performance, and the confusion matrix exhibits a diagonal distribution. The model achieves extremely high recognition accuracy for all five categories, including the four previously scarce fault categories.
[0060] Therefore, through the above... Figure 4 and Figure 5 The comparison results show that the wind turbine fault diagnosis model provided in this application is an effective solution to the problem of scarce fault samples in the field of wind power fault diagnosis. By generating synthetic samples that are highly consistent with real data in terms of time and frequency domain distribution, the model not only successfully expands the data, but more importantly, effectively solves the problem of imbalanced data categories. This enables the fault identification model to fully learn the unique characteristics of each type of fault from balanced data, thereby significantly improving the model's classification accuracy.
[0061] Third Embodiment Furthermore, in order to implement the wind turbine gearbox fault identification system for class-unbalanced systems corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects, Figure 6 A structural diagram of a wind turbine gearbox fault identification system for class-unbalanced gearboxes is provided. For ease of explanation, only the parts relevant to this embodiment are shown. The wind turbine gearbox fault identification system for class-unbalanced gearboxes provided in this application embodiment includes: The data acquisition module 201 is used to collect the original vibration signals of the wind turbine gearbox under normal and fault conditions within a preset historical period, and to preprocess the original vibration signals to obtain effective vibration signals.
[0062] In this embodiment, acceleration sensors are installed at several locations within the wind turbine gearbox, such as the bearing housing and the gearbox housing. These acceleration sensors collect the raw vibration signals of the gearbox during operation.
[0063] Optionally, in this embodiment of the application, the original vibration signals of the wind turbine gearbox under normal and fault conditions over the past six months are selected.
[0064] To address the environmental noise interference and invalid vibration frequency bands in the original vibration signal, noise reduction processing is performed on the original vibration signal using moving average filtering, median filtering, and wavelet transform. Finally, the signal data is normalized to obtain the effective vibration signal.
[0065] For abnormal abrupt changes in the original vibration signal caused by factors other than gearbox bearing failure (such as sudden changes in operating conditions, sensor interference, start-stop impact, etc.), a denoising method combining moving average filtering, median filtering, and wavelet transform is adopted. This method avoids feature distortion caused by single filtering and improves the reliability of vibration signals under complex operating conditions, providing more stable and effective input data for subsequent wind turbine fault diagnosis models.
[0066] Specifically, the original vibration signal is first subjected to median filtering to replace any abnormal abrupt changes identified within the sliding window (replacing the abnormal abrupt changes with the median). This prioritizes the removal of pulse-type and isolated spikes caused by acquisition anomalies or instantaneous impacts, avoiding the impact of non-physical abrupt changes on the overall signal characteristics. Next, moving average filtering is used to smooth the replaced original vibration signal, suppressing high-frequency random vibrations caused by wind condition changes and minor load fluctuations, while maintaining the overall signal trend. Based on the filtered signal, wavelet transform is applied to the smoothed signal. Through multi-scale decomposition and threshold denoising, non-stationary noise and bearing fault-related features are effectively separated from the original signal, and a high signal-to-noise ratio vibration signal is reconstructed to output an effective vibration signal.
[0067] Therefore, the denoising method combining moving average filtering, median filtering, and wavelet transform adopted in this application embodiment is progressive from coarse to fine, which not only improves the signal-to-noise ratio but also preserves the effective information related to mechanical state and fault mechanism to the greatest extent, thereby improving the data quality of vibration signal.
[0068] The data labeling module 202 is used to mark the effective vibration signal with the corresponding operating condition label by matching the operating time period according to the fault operation and maintenance record, and to construct an initial wind turbine vibration dataset based on the operating condition label and the feature value of the effective vibration signal.
[0069] In this embodiment of the application, in order to classify the effective vibration signal and assist the subsequent wind turbine fault diagnosis model in fault identification, this embodiment of the application uses the operating status provided by the fault operation and maintenance record to match the operating status of the corresponding time period and mark the effective vibration signal with the corresponding operating status label.
[0070] The operating conditions include normal condition, tooth root fracture, tooth root loss, tooth root damage, and tooth surface wear (four typical faults).
[0071] First, the fault maintenance records of the wind turbine gearbox, recorded by the on-site maintenance personnel of the wind farm, are exported from the auxiliary control system of the wind turbine. Then, based on the operating conditions of the gearbox at different time periods, the effective vibration signal is labeled with the corresponding operating condition tag by matching the time period of the operating condition with the running time period of the effective vibration signal, thereby realizing the category labeling of the vibration data at each moment.
[0072] In addition, the RMS value, peak value, peak factor, kurtosis value, waveform factor, impulse factor and spectral entropy value of the effective vibration signal at each time moment are calculated to obtain the characteristic values of the effective vibration signal.
[0073] Finally, an initial wind turbine vibration dataset is constructed based on the operating condition labels and the feature values.
[0074] The model training module 203 is used to train a preset deep learning model based on a preset balanced sample set to obtain a wind turbine fault diagnosis model; wherein the balanced sample set is generated by data augmentation of an initial wind turbine vibration dataset.
[0075] In this embodiment of the application, to address the class imbalance problem caused by scarce fault data, the embodiment utilizes an adversarial training mechanism between the generator and the discriminator to synthesize fault samples that are highly similar to the real distribution. Therefore, the data augmentation model used includes a generator and a discriminator.
[0076] After inputting the initial wind turbine vibration dataset into a preset data augmentation model, the latent temporal characteristics of the effective vibration signal are modeled through a multi-layer temporal network within the generator. Then, a fully connected layer maps these latent temporal characteristics into a fixed-length vibration time-domain signal, thereby obtaining a highly reliable fault sample. The temporal network is constructed using an LSTM structure, which can effectively learn the implicit temporal correlation and dynamic evolution laws in the vibration signal, making it suitable for describing the non-stationary, strongly temporal characteristics of wind turbines under complex operating conditions. The fully connected layer performs a nonlinear mapping of the high-dimensional temporal characteristics, making the generated signal approximate the real vibration data in terms of amplitude distribution and overall trend.
[0077] Furthermore, the activation function used by the generator can limit the output data to a standardized range, which is beneficial for maintaining a numerical scale consistent with real data. The generator provided in this application not only focuses on the distribution of point values, but also learns the temporal dependencies to make the generated vibration signal have high physical rationality in terms of waveform structure, periodicity, and random disturbance characteristics, making it suitable for vibration data enhancement of wind turbine gearboxes.
[0078] In addition to the generator, the discriminator of the data augmentation model mainly provides learning targets for the generator, further improving the authenticity and reliability of the generated samples.
[0079] The discriminator employs a one-dimensional convolutional neural network structure to extract and discriminate features from the input vibration data layer by layer. Through the convolutional layers of the discriminator, it captures the local waveform features, impact modes, and frequency-related information (collectively referred to as statistical features) of the effective vibration signal, making it particularly suitable for identifying typical patterns in wind turbine vibration caused by structural excitation and mechanical characteristics. As the random convolutional neural network deepens, the discriminator progressively captures and extracts higher-level statistical features, and distinguishes between real and generated signals through preset consistency constraints. It learns the differences between real and time-domain vibration signals, constraining the amplitude distribution of the generated signal and improving the authenticity of the generated vibration signal from a structural feature perspective, thus achieving the learning objective. Furthermore, the inclusion of a gradient penalty mechanism enhances the stability of the discriminator during training, avoiding gradient vanishing or mode collapse problems, thereby providing a more reliable learning objective for the generator.
[0080] Therefore, the data augmentation model provided in this application embodiment has been significantly enhanced in terms of structure and loss design to address the characteristics of vibration signals from wind turbine gearboxes. First, the generator incorporates a time-series network, enabling the model to characterize the time dependence of vibration signals and overcoming the difficulty of traditional models in modeling long-term time-series data. Second, in addition to the adversarial loss of the adversarial training mechanism, the data augmentation model also integrates engineering feature loss, frequency domain loss, and envelope spectrum loss to construct the model's loss function. Based on this loss function, commonly used characteristics in vibration signal analysis, such as RMS, peak value, kurtosis, and spectral properties, are directly introduced into the training process of the data augmentation model. This makes the generated data more statistically and physically similar to real wind turbine vibration signals, improving the similarity between real vibration signals and the vibration time-domain signals.
[0081] The fault samples output by the data augmentation model are merged with the initial wind turbine vibration dataset to form a balanced sample set. The balanced sample set is then divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The training set and the validation set are input into a preset deep learning model for model training to obtain a trained wind turbine fault diagnosis model.
[0082] The classifier of the wind turbine fault diagnosis model is constructed using a convolutional neural network based on residual connections. The residual block structure enables deep mapping of fault features, effectively avoiding the vanishing gradient problem in deep network training while significantly improving the extraction accuracy of weak fault features. The classifier iterates its parameters on the balanced sample set, and finally outputs the fault type identification result and the corresponding fault risk probability through a Softmax layer, achieving highly robust diagnosis in imbalanced scenarios.
[0083] The fault identification module 204 is used to process the effective vibration signal through the wind turbine fault diagnosis model and output the fault identification result of the wind turbine gearbox, including the fault type / operating condition and fault risk probability of the wind turbine gearbox.
[0084] Implementing the embodiments of this application has the following beneficial effects: This application collects vibration data of wind turbine gearboxes under various operating conditions. The vibration data is preprocessed to effectively remove outliers caused by sensor interference, abnormal acquisition, or instantaneous impacts, preventing these outliers from interfering with the overall signal characteristics and improving data quality. Then, by labeling the vibration signals with operating conditions, the vibration signals inherently carry corresponding category information, eliminating the need for manual secondary labeling in the model and improving fault identification accuracy. Furthermore, data augmentation learning is used to study the temporal variation patterns of vibration signals in the initial wind turbine vibration dataset, generating high-fidelity fault samples that more closely resemble the actual operating state of the wind turbine in terms of waveform morphology and evolution trend. This effectively expands the number of fault samples, alleviating the problem of scarce wind turbine fault samples and solving the class imbalance problem. Using a balanced sample set for model training improves the robustness and generalization ability of the wind turbine fault diagnosis model in complex class imbalance scenarios, enhancing the identification accuracy for different fault types.
[0085] Furthermore, Figure 7 This is a structural diagram of a terminal device provided in one embodiment of this application. Figure 7 As shown, the terminal device 3 of this embodiment includes: at least one processor 30 (in... Figure 7 (Only one is shown in the image) and a memory 31 and a computer program 32 stored in the memory 31 and executable on the at least one processor, wherein when the processor 30 executes the computer program 32, it can implement the steps of the method for identifying wind turbine gearbox faults oriented to class imbalance as described in any one of the embodiments of this application.
[0086] The terminal device 3 may be a computing device such as a desktop computer, a cloud server, or a laptop computer, and the computing device may include, but is not limited to, a processor 30 and a memory 31. Figure 7This is merely an example of terminal device 3 and does not constitute a limitation on terminal device 3. It may include more or fewer components than those shown in the figure.
[0087] This application provides a storage medium that stores computer-readable program code. When the computer-readable program code is executed, it implements the steps of the above-described error correction method for an open-gap current transformer.
[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, or improvements made by those skilled in the art within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems, characterized in that, include: Within a preset historical period, the original vibration signals of the wind turbine gearbox under normal and fault conditions are collected, and the original vibration signals are preprocessed to obtain effective vibration signals. Based on the fault operation and maintenance records, the effective vibration signals are labeled with corresponding operating condition tags by matching the operating time period. Based on the operating condition tags and the feature values of the effective vibration signals, an initial wind turbine vibration dataset is constructed. A wind turbine fault diagnosis model is obtained by training a pre-defined deep learning model using a pre-defined balanced sample set; wherein, the balanced sample set includes a fault sample set generated based on an initial wind turbine vibration dataset. The effective vibration signal is processed by the wind turbine fault diagnosis model, and the fault identification result of the wind turbine gearbox is output.
2. The method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems according to claim 1, characterized in that, Within a preset historical time period, raw vibration signals of the wind turbine gearbox under normal and fault conditions are collected. These raw vibration signals are then preprocessed to obtain effective vibration signals. Specifically: Accelerometers are installed at several locations on the gearbox of the wind turbine to collect the original vibration signals; the historical period is the past six months. By combining filtering techniques and wavelet transform, and by identifying abnormal abrupt changes, environmental noise and invalid vibration frequency bands are removed from the original vibration signal to obtain an effective vibration signal. The abnormal abrupt changes are caused by factors other than gearbox bearing failure, including sudden changes in operating conditions, sensor interference, and unit start-up and shutdown shocks.
3. The method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems according to claim 2, characterized in that, The combination of filtering techniques and wavelet transform, by identifying abnormal abrupt changes, removes environmental noise and invalid vibration frequency bands from the original vibration signal to obtain an effective vibration signal, specifically: The abnormal abrupt change value of the original vibration signal is replaced within a sliding window using median filtering technology, and the original vibration signal after replacement is smoothed using moving average filtering. By using wavelet transform technology, the smoothed original vibration signal is decomposed into multiple scales and denoised by thresholding. Bearing fault-related features are separated from the original vibration signal to obtain an effective vibration signal.
4. The method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems according to claim 1, characterized in that, The step involves labeling the valid vibration signals with corresponding operating condition tags based on fault maintenance records and matching the operating time periods. Then, based on the operating condition tags and the feature values of the valid vibration signals, an initial wind turbine vibration dataset is constructed. Specifically: Based on the operating conditions provided by the fault operation and maintenance records, the effective vibration signal is labeled with a corresponding operating condition tag by matching the time period of the operating condition with the running time period of the effective vibration signal; wherein, the operating conditions include normal state, tooth root fracture, tooth root missing, tooth root damaged and tooth surface worn. The RMS value, peak value, peak factor, kurtosis value, waveform factor, impulse factor, and spectral entropy value of the effective vibration signal at each moment are calculated to obtain the characteristic values. An initial wind turbine vibration dataset is constructed based on the operating condition labels and the feature values.
5. The method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems according to claim 1, characterized in that, The process of training a pre-defined deep learning model using a pre-defined balanced sample set to obtain a wind turbine fault diagnosis model is as follows: The initial wind turbine vibration dataset is input into a preset data augmentation model to construct the potential temporal features of the effective vibration signal through a generator, capture the local waveform features of the effective vibration signal through a discriminator, and generate the fault sample based on the output results of the generator and the discriminator. The fault samples and the initial wind turbine vibration dataset are merged into the balanced sample set. The classifier of the deep learning model is iterated using the balanced sample set to obtain the trained wind turbine fault diagnosis model.
6. The method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems according to claim 5, characterized in that, The generation of the fault sample based on the outputs of the generator and the discriminator specifically involves: Random noise vectors from the initial wind turbine vibration dataset are extracted as input to the generator. Within the generator, based on the learning objective output by the discriminator, the latent temporal features of the effective vibration signal are constructed through a preset temporal network. Then, a fully connected layer is used to map the latent temporal features into a fixed-length vibration time-domain signal to obtain the fault sample. The discriminator captures the local waveform characteristics, impact modes, and frequency-related information of the effective vibration signal. Based on the capture results and preset consistency constraints, the difference between the real vibration signal and the vibration time-domain signal is learned, and the learning target is obtained.
7. The method for identifying gearbox faults in wind turbines oriented towards class-unbalanced systems according to claim 6, characterized in that, Also includes: The loss function of the data augmentation model is constructed based on adversarial loss, engineering feature loss, frequency domain loss, and envelope spectrum loss. Based on the loss function, the feature values of the effective vibration signal are introduced into the data augmentation model to improve the similarity between the real vibration signal and the vibration time-domain signal.
8. A fault identification system for wind turbine gearboxes oriented towards class-unbalanced systems, characterized in that, include: Data acquisition module, data annotation module, model training module, and fault identification module; The data acquisition module is used to collect the original vibration signals of the wind turbine gearbox under normal and fault conditions within a preset historical period, and to preprocess the original vibration signals to obtain effective vibration signals. The data labeling module is used to label the effective vibration signal with the corresponding operating condition label by matching the operating time period according to the fault operation and maintenance record, and to construct an initial wind turbine vibration dataset based on the operating condition label and the feature value of the effective vibration signal. The model training module is used to train a preset deep learning model using a preset balanced sample set to obtain a wind turbine fault diagnosis model; wherein, the balanced sample set is generated by data augmentation of the initial wind turbine vibration dataset; The fault identification module is used to process the effective vibration signal through the wind turbine fault diagnosis model and output the fault identification result of the wind turbine gearbox.
9. A terminal device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method for identifying gearbox faults in wind turbines oriented towards class imbalance, as described in any one of claims 1 to 7.
10. A storage medium, characterized in that, The storage medium stores computer-readable program code, which, when executed, implements the steps of the method for identifying gearbox faults in wind turbines oriented towards class imbalance, as described in any one of claims 1 to 7.