Wind turbine generator gearbox coupling fault diagnosis method based on multi-source heterogeneous features and double-branch time-frequency attention

Abstract

The invention discloses a wind turbine generator gearbox coupling fault diagnosis method based on multi-source heterogeneous features and double-branch time-frequency attention, and belongs to the technical field of wind turbine generator gearbox fault diagnosis. The method comprises the following steps: collecting multi-source sensor data for a wind turbine generator gearbox multi-source sensor, and preprocessing the multi-source sensor data to obtain preprocessed data; performing feature extraction on the preprocessed data to obtain multi-source signal depth features; performing scale normalization and weighted fusion on the depth features of the multi-source signal to obtain a multi-source heterogeneous coupling feature vector; and obtaining a cross-domain fusion feature vector based on the multi-source heterogeneous coupling feature vector, and obtaining a gearbox coupling fault diagnosis result based on the cross-domain fusion feature vector. According to the method, the technical problems of single monitoring dimension, simple feature fusion, difficulty in coupling fault separation, high false alarm rate, poor interpretability and the like of the existing fault diagnosis method are solved.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine gearbox fault diagnosis technology, specifically involving a wind turbine gearbox coupling fault diagnosis method based on multi-source heterogeneous characteristics and dual-branch time-frequency attention. Background Technology

[0002] As a crucial component of global clean energy, wind power's long-term stable operation is essential for ensuring power generation efficiency and equipment safety. The gearbox of a wind turbine, as a core component of the transmission system, plays a critical role in converting the low-speed mechanical energy of the rotor into high-speed electrical energy of the generator. Its operating status is closely related to the turbine's power generation efficiency, maintenance costs, and safety risks. Modern wind turbine gearboxes have complex structures, including gears, bearings, planetary carriers, and other key components. They must operate under harsh conditions such as variable loads, strong vibrations, high and low temperature cycles, and sand and dust erosion for extended periods. The complexity of the working environment leads to diverse and coupled failure modes.

[0003] Among various gearbox failures, not only are there single faults such as gear cracks, tooth surface wear, and bearing pitting, but also frequently coupled faults involving multiple components and types, such as gear cracks and planetary gear wear, and bearing failures and tooth surface wear. The characteristic signals of coupled faults overlap and interfere with each other, making it difficult to effectively separate fault characteristics and significantly increasing the diagnostic difficulty compared to single faults. According to wind power industry operation and maintenance data, gearbox failures account for approximately 20%-30% of all wind turbine failures. Although the failure frequency is lower than that of components such as blades, the precision of its internal structure, the difficulty of disassembly and assembly, and the long maintenance cycle mean that the cost of a single failure repair often constitutes a significant portion of the annual operation and maintenance costs of a wind farm. Furthermore, the loss of power generation during downtime due to failure further exacerbates the economic losses.

[0004] Currently, fault diagnosis of wind turbine gearboxes mainly relies on single-signal monitoring and traditional analysis methods. Vibration monitoring is the most widely used technology, but it mainly targets the extraction of single fault features, and in coupled fault scenarios, it is prone to diagnostic errors due to feature confusion. Temperature monitoring can only reflect the thermal anomaly trend of the equipment, has a slow response speed, and is difficult to detect early minor faults. Acoustic fingerprint monitoring, as a non-contact technology, is easily interfered with by wind noise and mechanical background noise, and has low feature recognition. In recent years, multi-source signal fusion diagnostic methods have gradually emerged, but existing technologies still have significant shortcomings: feature fusion methods are relatively rudimentary, mostly involving simple splicing or weighting, failing to fully explore the cross-domain correlation between time-domain and frequency-domain features; there is a lack of quantitative evaluation of the quality of multi-source signal data, and low-quality features can easily interfere with diagnostic results; coupled fault decoupling capabilities are insufficient, making it difficult to effectively separate superimposed fault features; the diagnostic model has a high false alarm rate and insufficient interpretability of the decision-making process, failing to provide maintenance personnel with clear fault location basis.

[0005] Therefore, in response to the technical pain points of existing gearbox fault diagnosis methods, such as single monitoring dimensions, coarse feature fusion, difficulty in separating coupled faults, high false alarm rate, and poor interpretability, there is an urgent need to develop an intelligent diagnostic method that can integrate multi-source heterogeneous signals, accurately decouple coupled fault features, and adaptively fuse high-quality features. This method would enable early, accurate, and robust detection of both single and coupled faults in gearboxes, providing technical support for the efficient operation and safe operation of wind turbine units. Summary of the Invention

[0006] The present invention aims to provide a method for diagnosing coupled faults in wind turbine gearboxes based on multi-source heterogeneous features and dual-branch time-frequency attention, in order to solve the technical problems existing in the prior art, such as single monitoring dimension, rudimentary feature fusion, difficulty in separating coupled faults, high false alarm rate, and poor interpretability.

[0007] To achieve the above objectives, the present invention provides the following solution: a method for diagnosing coupled faults in wind turbine gearboxes based on multi-source heterogeneous features and dual-branch time-frequency attention, comprising the following steps: S1. Install vibration sensors, temperature sensors, and acoustic sensors on the gearbox of the wind turbine generator set, collect multi-source sensor data and preprocess it to obtain preprocessed data; the preprocessed data includes: preprocessed vibration signal, preprocessed acoustic signal, and preprocessed temperature signal. S2. Perform feature extraction on the preprocessed data to obtain multi-source signal depth features; S3. The multi-source signal depth features are scaled and weighted and fused to obtain a multi-source heterogeneous coupling feature vector. S4. Based on the multi-source heterogeneous coupling feature vector, a cross-domain fusion feature vector is obtained, and based on the cross-domain fusion feature vector, a gearbox coupling fault diagnosis result is obtained.

[0008] More preferably, in S1, the preprocessing method includes: removing erroneous, distorted, and invalid data; The spectral kurtosis value of the vibration signal is calculated, and the fault-sensitive frequency band is determined based on the spectral kurtosis value; the vibration signal of the fault-sensitive frequency band is reconstructed to obtain the preprocessed vibration signal. The temperature signal is subjected to moving average filtering to obtain the preprocessed temperature signal; The voiceprint signal is denoised by spectral subtraction to obtain the preprocessed voiceprint signal.

[0009] More preferably, in S2, the method for feature extraction of the preprocessed vibration signal includes: Hilbert envelope demodulation is used to convert the preprocessed vibration signal from the time domain to the envelope domain to obtain the original envelope spectrum; Variational mode decomposition is performed on the original envelope spectrum, and the consistency of the dominant frequency and energy distribution of the envelope spectrum of each MIF component with the original envelope spectrum is compared to obtain the depth characteristics of the vibration signal.

[0010] More preferably, when performing variational mode decomposition on the original envelope spectrum, the method for determining the optimal number of modes includes: ; ; In the formula, EFCI ( k ) indicates the number of patterns is k Time-frequency concentration index; For the first Energy of one IMF component The total energy of all IMF components. For the first The center frequency of each IMF For the first The frequency bandwidth of one IMF; This is an empirical threshold; Decomposition stops when the value is less than the empirical threshold. This represents the optimal number of patterns.

[0011] More preferably, the consistency of the original envelope spectrum includes: ; In the formula, For the first One IMF component; The envelope spectrum of the original vibration signal; For the first The envelope spectrum of an IMF; The Pearson correlation coefficient; The main peak frequency of the original envelope spectrum; The main peak frequency of the IMF envelope spectrum; IMF energy deviation rate; Represented as Kroneck function.

[0012] More preferably, in S4, the method for obtaining the cross-domain fusion feature vector includes: inputting the multi-source heterogeneous coupling feature vector into a dual-branch time-frequency feature extraction network to obtain the cross-domain fusion feature vector; The dual-branch time-domain feature extraction network includes a time-domain branch and a frequency-domain branch; The temporal branch takes the temporal heterogeneous features in the multi-source heterogeneous coupling feature vector as input, and uses a multi-scale one-dimensional convolutional neural network combined with data quality weights to perform weighted modeling of temporal information to obtain the comprehensive encoding of the multi-source heterogeneous coupling feature vector. The frequency domain branch takes the frequency domain heterogeneous features in the multi-source heterogeneous coupling feature vector as input, formats the frequency domain heterogeneous features into a time-frequency graph, and uses a two-dimensional convolutional neural network to extract cross-field frequency domain collaborative features of different frequency regions from the time-frequency graph. Construct a multi-source domain bias vector based on the depth features of the multi-source signals; Max pooling and average pooling operations are performed on the comprehensive coding and the frequency domain collaborative features to obtain time domain pooling features and frequency domain pooling features; The temporal pooling features, the frequency pooling features, and the multi-source domain bias vector are concatenated and input into a fully connected layer. After passing through channel compression and expansion operations, adaptive channel attention weights are generated at the output through Sigmoid activation. The adaptive channel attention weights are then applied to the temporal heterogeneous features and the frequency heterogeneous features in the multi-source heterogeneous coupling feature vector to obtain the cross-domain fusion feature vector.

[0013] More preferably, in S4, a gearbox diagnostic model is used to process the cross-domain fusion feature vector to obtain a gearbox coupling fault diagnosis result; the method for processing the cross-domain fusion feature vector using the gearbox diagnostic model includes: Support Vector Machine (SVM) uses a basis kernel function to perform a preliminary nonlinear partitioning of the cross-domain fusion feature vectors, mapping the original low-dimensional space to a high-dimensional space. The deep neural network uses the Softmax function to calculate the probability distribution of multi-class faults on the cross-domain fusion feature vector processed by the support vector machine, and obtains the gearbox coupling fault diagnosis results.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention deploys multiple types of sensors to collect vibration, temperature, and acoustic signature signals, which are then purified through differentiated preprocessing. Deep features of these three signal types are extracted and combined with Data Quality Index (DQI) weights to construct a multi-source heterogeneous coupled feature vector. Time-domain and frequency-domain features are extracted through a dual-branch time-frequency network, and adaptively fused using a Domain-biased Multi-Effective Channel Attention (MECA) module to generate a cross-domain fused feature vector. Finally, a hybrid classifier combining support vector machines and deep neural networks is used to complete gearbox coupled fault diagnosis. This invention achieves early, accurate, and robust detection of single faults such as gear cracks, tooth surface wear, and bearing pitting, as well as coupled faults such as gear cracks and planetary gear wear, and bearing failures and tooth surface wear. It solves the technical problems of existing fault diagnosis methods, such as single monitoring dimensions, rudimentary feature fusion, difficulty in separating coupled faults, high false alarm rates, and poor interpretability. It significantly improves the accuracy and robustness of single and coupled fault diagnosis in gearboxes, providing strong support for the safe and stable operation of wind turbine units. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of the wind turbine gearbox coupling fault diagnosis method based on multi-source heterogeneous features and dual-branch time-frequency attention provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the construction process of multi-source heterogeneous coupling feature vectors in an embodiment of the present invention; Figure 3 This is a schematic diagram of the cross-domain fusion feature vector construction process in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1: like Figure 1As shown, this embodiment provides a method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention, including the following steps: S1. Install vibration sensors, temperature sensors, and acoustic sensors on the gearbox of the wind turbine generator set, collect multi-source sensor data and preprocess it to obtain preprocessed data; the preprocessed data includes: preprocessed vibration signal, preprocessed acoustic signal, and preprocessed temperature signal.

[0020] Vibration sensors, temperature sensors, and acoustic sensors are installed at key detection points in the wind turbine gearbox. All sensors use a unified time reference to record sampling time, ensuring time alignment for subsequent time-domain / time-frequency joint analysis. Channels exhibiting significant distortion, saturation, or disconnection are removed in real time, and sampling descriptions are added to signals with different sampling rates in preparation for subsequent preprocessing.

[0021] In this embodiment, the historical vibration, acoustic signature, and temperature data collected during the operation of the wind turbine are cleaned and invalid signals are removed, including removing data with acquisition errors, distorted segments, and long-term invalid data, to ensure the availability and continuity of the input data; then, differentiated preprocessing is performed on the physical characteristics of the three types of signals.

[0022] Specifically, the spectral kurtosis value of the vibration signal is calculated, and the fault-sensitive frequency band is determined based on the spectral kurtosis value; the vibration signal of the fault-sensitive frequency band is reconstructed to obtain the preprocessed vibration signal.

[0023] The formulas for spectral kurtosis include: ; In the formula, This represents the expectation operation; Represents the continuous wavelet transform coefficients of the signal; The scale parameter represents the wavelet transform. This represents the target frequency point to be analyzed; Represents a time variable. The larger the spectral kurtosis value, the stronger the impact of the corresponding frequency band. The vibration signal is used to lock the fault-sensitive frequency band through the multi-scale dynamic spectral kurtosis method, and the impact energy is focused by Butterworth bandpass filtering. Then, wavelet transform is used to perform threshold noise reduction on the decomposed components. A soft threshold function is set to reconstruct a clean vibration signal, i.e., the pre-processed vibration signal.

[0024] The soft threshold function is: ; In the formula, Represents a symbolic function; For adaptive threshold: ; For signal length, This represents the noise standard deviation.

[0025] The temperature signal is processed by moving average filtering to smooth its slow change trend and enhance the recognizability of the temperature rise pattern, thus obtaining the preprocessed temperature signal.

[0026] The acoustic signature signal is denoised by spectral subtraction, and then enhanced by spectral enhancement to improve the energy distribution of the acoustic signature, making it more prominent in gear meshing and structural feature changes. The preprocessed acoustic signature signal is obtained by using the spectral subtraction formula: ; In the formula, The noise power spectrum, The power spectrum of the signal containing noisy ripples. For the floor threshold, For over-subtraction factor, This is the power spectrum of the clean audio waveform after noise reduction.

[0027] The multi-source signals that have undergone differentiated preprocessing are uniformly normalized and divided into training, validation and test sets according to a specified ratio to construct a standardized multi-source dataset with consistent structure and scale. This dataset aims to ensure that the fault modes and feature distributions are consistent among different subsets, providing high-quality input for subsequent deep feature extraction and coupled fault diagnosis.

[0028] S2. Perform feature extraction on the preprocessed data to obtain multi-source signal depth features.

[0029] Specifically, methods for feature extraction from preprocessed vibration signals include: Hilbert envelope demodulation is employed to convert the preprocessed vibration signal from the time domain to the envelope domain, yielding the original envelope spectrum. The generated envelope spectrum amplifies fault-related frequency components such as gear meshing frequencies, bearing characteristic frequencies, their harmonics, and sidebands, providing a clear view of the characteristics of impact-related faults. Variational mode decomposition (VMD) is then performed on the original envelope spectrum, and the consistency of the dominant frequency and energy distribution of each MIF component with the original envelope spectrum is compared to obtain the depth characteristics of the vibration signal.

[0030] When performing variational mode decomposition (VMD) on the original envelope spectrum, the number of modes in the variational mode decomposition is adaptively determined using the energy-frequency concentration index (EFCI). k The method includes: EFCI continuously monitors the energy contribution and frequency distribution changes of each newly added IMF (Intrinsic Mode Function). When the energy contribution of the newly added IMF to the fault characteristics is lower than the normal threshold, and its center frequency deviates from the fault-related frequency range, the decomposition is stopped and the optimal value is determined. kValues ​​should be set to avoid energy dilution due to excessive decomposition or omission of fault characteristics due to insufficient decomposition.

[0031] The formula for the Energy-Frequency Concentration Index (EFCI) is: ; In the formula, EFCI ( k ) indicates the number of patterns is k Time-frequency concentration index; For the first Energy of one IMF component The total energy of all IMF components. For the first The center frequency of each IMF For the first The frequency bandwidth of the IMF.

[0032] when Stop decomposition when the current time is reached. The optimal number of patterns; where, This is an empirical threshold.

[0033] Determining the optimal number of decomposition patterns k Then, the consistency of the original envelope spectrum is calculated using the "Envelope Spectrum Dominant Frequency Consistency Constraint (ESMC)" strategy, including: ; In the formula, For the first One IMF component; The envelope spectrum of the original vibration signal; For the first The envelope spectrum of an IMF; The Pearson correlation coefficient; The main peak frequency of the original envelope spectrum; The main peak frequency of the IMF envelope spectrum; IMF energy deviation rate; Represented as Kroneck function.

[0034] The envelope spectrum of each IMF (Inductively Coupled Fault) was compared with the original envelope spectrum in terms of dominant frequency, energy distribution, and consistency. Only effective components related to fault characteristics were retained, while background noise, structural transmission modes, or irrelevant frequency bands were removed. Next, a fixed-length window sequence-based impact motif retrieval was performed on the retained IMF signals. The signals were segmented, and the similarity distance between segments was calculated to filter recurring fault feature patterns. Combined with envelope spectrum features, a preliminary distinction between single and coupled faults was made. Repeated impact segments were extracted and their occurrence period, energy, and evolution trend were identified to further enhance the sensitivity to early meshing anomalies, localized abrasion, or crack initiation. Finally, the envelope spectrum, VMD-selected IMFs, and various time-frequency features of the impact motifs were combined to constitute the deep physical interpretable features of the vibration channel.

[0035] Preprocessed voiceprint signal Perform a Short-Time Fourier Transform (STFT) to obtain its time-frequency distribution characteristics. The STFT formula is as follows: ; In the formula, For window functions, It is the time frame number. It is the sampling point number. It's frequency. for Time, frequency The time-frequency amplitude at a given location corresponds to the time-varying frequency distribution of the voiceprint.

[0036] Further calculation of Mel frequency cepstral coefficients (MFCC) is used to characterize the energy distribution pattern of the acoustic signature signal under different harmonic structures. At the same time, based on the STFT spectrum, the depth features of the acoustic signature signal, such as the spectral centroid, spectral extension, spectral slope, and bandwidth change rate, which reflect the dynamic changes of the spectral morphology, are extracted, so that the subtle changes in the acoustic signature caused by abnormal gear meshing, lubrication deterioration, and structural wear can be effectively captured.

[0037] For the preprocessed temperature signal, the sliding window is used to analyze its change trend over time, and characteristic indicators such as the temperature rise rate per unit time, the temperature gradient difference between different monitoring points, and the consistency of gradient change are calculated to characterize the changes in the thermal characteristics of the gearbox under load fluctuation, increased friction, or decreased heat dissipation performance. Finally, a temperature signal depth feature that can reflect the acoustic signature and slow temperature change characteristics and state accumulation effect is formed, providing a reliable input for subsequent multi-source coupling feature construction.

[0038] S3. The depth features of the multi-source signals are scaled and weighted and fused to obtain the multi-source heterogeneous coupling feature vector.

[0039] like Figure 2As shown, based on multi-source coupling theory and data quality weighting (DQI), the quality of depth features of three types of signals—vibration, acoustic signature, and temperature—is evaluated. First, corresponding quality indices are constructed based on the feature clarity, stability, and reliability exhibited by the three types of signals during acquisition, preprocessing, and feature extraction. Among them, the depth features of vibration signals are evaluated for their effectiveness in characterizing fault impact by combining the dominant frequency prominence of the envelope spectrum, the peak value of spectral kurtosis, and the energy concentration of the impact feature. The depth features of acoustic signature signals are evaluated for their sensitivity to meshing abnormalities and wear signs based on the signal-to-noise ratio, spectral centroid shift rate, feature band energy stability, and time-frequency mode continuity. The depth features of temperature signals are evaluated for their reliability in characterizing slow-heating faults by the temperature rise rate, the gradient difference at monitoring points, consistency changes, and trend stability. After normalization, the above indicators yield Data Quality Indicators (DQIs) corresponding to the depth features of vibration, acoustic signature, and temperature signals. Subsequently, scale normalization is performed on each of the three signal depth features to ensure comparability in both dimensions and amplitude. The feature sub-vectors are then weighted and fused using their respective DQI weights, ensuring that features with higher quality and stronger representational capabilities occupy a higher proportion in the final fused structure. Finally, the three weighted depth features are concatenated in a fixed order to construct a multi-source heterogeneous coupling feature vector. The multi-source heterogeneous coupling feature vector is represented as follows: ; in, ; In the formula, This represents the depth feature vector of the vibration signal. This is the depth feature vector of the temperature signal; This is the depth feature vector of the voiceprint signal; The DQI weights for the depth features of the vibration signal. The DQI weights for the depth features of the temperature signal. The DQI weights are the depth features of the voiceprint signal.

[0040] S4. Based on the multi-source heterogeneous coupling feature vector, a cross-domain fusion feature vector is obtained, and based on the cross-domain fusion feature vector, a gearbox coupling fault diagnosis result is obtained.

[0041] The method for obtaining the cross-domain fusion feature vector includes: inputting the multi-source heterogeneous coupling feature vector into a dual-branch time-frequency feature extraction network to obtain the cross-domain fusion feature vector. Specifically, the dual-branch time-domain feature extraction network includes a time-domain branch and a frequency-domain branch; The temporal branch takes the temporal heterogeneous features in the multi-source heterogeneous coupling feature vector as input, and uses a multi-scale one-dimensional convolutional neural network (1D-CNN) combined with data quality weights (DQI) to perform weighted modeling of temporal information from different sources. Through multi-scale convolutional kernels, it captures the short-period repeatability of vibration and impact features, the mesoscale stability of acoustic temporal drift features, and the long-term accumulation of temperature gradual change trends, thereby realizing the comprehensive encoding of multi-source temporal features in the same network structure.

[0042] The frequency domain branch takes the frequency domain heterogeneous features in the multi-source heterogeneous coupling feature vector as input, and formats the frequency domain heterogeneous features into a time-frequency diagram. The method includes: taking the frequency domain feature subset of the multi-source coupling feature vector as time-frequency conversion and format adaptation to form multi-channel input data, wherein the first channel is a time-frequency diagram generated by continuous wavelet transform of vibration signal, the second channel is a soundprint spectrum obtained by matrix reconstruction of the frequency domain features of soundprint signal, and the third channel is a frequency domain diagram generated by normalization of the frequency domain features of temperature signal. The three channels of data reflect the fault characteristics from different physical dimensions and are spliced ​​in a fixed order to form heterogeneous time-frequency input. A two-dimensional convolutional neural network (2D-CNN) is used to extract cross-domain frequency-domain collaborative features from the time-frequency map. The time-frequency map is composed of wavelet time-frequency maps of vibration signals, reconstructed spectrograms of acoustic signature signals, and normalized frequency-domain maps of temperature signals superimposed in a fixed order. The 2D-CNN uses multi-scale convolutional units to extract cross-domain frequency-domain collaborative features in different frequency regions, and combines DQI weights to adjust the contribution of different feature channels during convolution and pooling, so that the frequency-domain branch pays more attention to high-quality feature sources. Finally, both branches output representation vectors of the same dimension for subsequent cross-domain attention fusion, providing a structured time-frequency joint deep feature representation for the fault diagnosis model.

[0043] The 1D-CNN network in the time domain and the 2D-CNN network in the frequency domain both consist of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation function. The 1D-CNN network captures temporal coupling features across different time dimensions through multi-scale extraction. The 2D-CNN network transforms a subset of frequency-domain features from the multi-source coupling feature vector into a multi-channel time-frequency / spectral graph, and then extracts frequency-domain collaborative features within different frequency ranges using multi-scale convolutional kernels, accurately capturing fault correlation information in the frequency domain distribution of the signal.

[0044] like Figure 3As shown, to emphasize the differentiated contribution of features from different physical fields to fault modes, an adaptive feature weighting mechanism based on the fault sensitivity of multi-source signals is introduced to optimize the cross-domain feature fusion effect by combining the fault representation characteristics of signals from each physical domain. The initial attention bias is constructed using the fault sensitivity index obtained from the deep feature extraction stage of multi-source signals. In this model, the spectral kurtosis peak of the vibration signal is used to characterize the sensitivity to impact-type faults; the spectral centroid shift rate and bandwidth change rate of the acoustic signature signal are used to characterize the sensitivity to acoustic mode changes; and the temperature rise gradient of the temperature signal is used to characterize the sensitivity to thermal anomalies. These three types of sensitivity are normalized and then weighted and fused according to their fault characterization reliability to generate a multi-source domain bias vector. Subsequently, the features output from the time-domain and frequency-domain branches are subjected to max pooling and average pooling, respectively, to obtain time-domain pooled features and frequency-domain pooled features, which are used to extract significant fault features and overall background information. The two types of pooled features are then concatenated with the multi-source domain bias vector and input into a fully connected layer. Channel compression and expansion operations are used to learn the importance of different feature channels, and adaptive channel attention weights are generated at the output through Sigmoid activation. Finally, attention weights are applied to the original time-domain and frequency-domain features using element-wise multiplication, making the model more focused on heterogeneous feature channels with high fault relevance, thereby achieving deep fusion and enhancement of cross-domain features and outputting a cross-domain fused feature vector, providing discriminative, high-quality coupled features for subsequent classification models.

[0045] A gearbox diagnostic model is trained using cross-domain fused feature vectors and historical data including normal, single fault, and coupled faults to learn the mapping relationship between multi-source heterogeneous coupling features and fault modes. During training, the Adam optimizer and cross-entropy loss function are used to iteratively update the network parameters, and attention weight visualization technology is combined to dynamically verify the model's attention to key features, improving the model's interpretability and reliability. Finally, a hybrid classifier based on support vector machine (SVM) and deep neural network (DNN) is constructed. The gearbox diagnostic model is then used to process the cross-domain fused feature vectors to obtain gearbox coupled fault diagnosis results. The method for processing the cross-domain fused feature vectors in the gearbox diagnostic model includes: Support Vector Machines (SVMs) use a basis kernel function to perform preliminary nonlinear partitioning on the cross-domain fused feature vectors, mapping the original low-dimensional space to a high-dimensional space; enabling different categories to achieve preliminary separability in this space; the formula for the radial basis kernel function is: ; in, The kernel function value; These are kernel function parameters; , These are two samples from the multi-source heterogeneous coupling feature vector set.

[0046] Deep neural networks further perform deep abstraction and pattern recognition on cross-domain fusion features; specifically, the Softmax function is used to activate the output layer to generate probability distributions for multiple types of faults, thus completing the final identification of single and coupled faults in the gearbox; the formula for the Softmax function is: ; in, For the output layer The original output of each neuron, This represents the total number of fault categories. The sum of the exponents of the original output for all fault types.

[0047] Finally, the cross-domain fusion features of the gearbox to be diagnosed are input into the SVM-DNN hybrid classifier that has been trained and validated. The classifier enables gearbox coupling fault diagnosis, providing accurate and robust intelligent diagnostic capabilities for wind turbine gearboxes.

[0048] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for diagnosing coupled faults in wind turbine gearboxes based on multi-source heterogeneous features and bi-branch time-frequency attention, characterized in that, Includes the following steps: S1. Install vibration sensors, temperature sensors and acoustic sensors on the gearbox of the wind turbine, collect multi-source sensor data and preprocess it to obtain preprocessed data. The preprocessed data includes: preprocessed vibration signal, preprocessed acoustic signature signal, and preprocessed temperature signal; S2. Perform feature extraction on the preprocessed data to obtain multi-source signal depth features; S3. The multi-source signal depth features are scaled and weighted and fused to obtain a multi-source heterogeneous coupling feature vector. S4. Based on the multi-source heterogeneous coupling feature vector, a cross-domain fusion feature vector is obtained, and based on the cross-domain fusion feature vector, a gearbox coupling fault diagnosis result is obtained.

2. The method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention as described in claim 1, characterized in that, In S1, the preprocessing method includes: removing erroneous, distorted, and invalid data; The spectral kurtosis value of the vibration signal is calculated, and the fault-sensitive frequency band is determined based on the spectral kurtosis value; the vibration signal of the fault-sensitive frequency band is reconstructed to obtain the preprocessed vibration signal. The temperature signal is subjected to moving average filtering to obtain the preprocessed temperature signal; The voiceprint signal is denoised by spectral subtraction to obtain the preprocessed voiceprint signal.

3. The method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention as described in claim 1, characterized in that, In S2, the methods for feature extraction of the preprocessed vibration signal include: Hilbert envelope demodulation is used to convert the preprocessed vibration signal from the time domain to the envelope domain to obtain the original envelope spectrum; Variational mode decomposition is performed on the original envelope spectrum, and the consistency of the dominant frequency and energy distribution of the envelope spectrum of each MIF component with the original envelope spectrum is compared to obtain the depth characteristics of the vibration signal.

4. The method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention as described in claim 3, characterized in that, When performing variational mode decomposition on the original envelope spectrum, the method for determining the optimal number of modes includes: ； ； In the formula, EFCI ( k ) indicates the number of patterns is k Time-frequency concentration index; For the first Energy of one IMF component The total energy of all IMF components. For the first The center frequency of each IMF For the first The frequency bandwidth of one IMF; This is an empirical threshold; Decomposition stops when the value is less than the empirical threshold. This represents the optimal number of patterns.

5. The method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention as described in claim 3, characterized in that, The consistency of the original envelope spectrum includes: ； In the formula, For the first One IMF component; The envelope spectrum of the original vibration signal; For the first The envelope spectrum of an IMF; The Pearson correlation coefficient; The main peak frequency of the original envelope spectrum; The main peak frequency of the IMF envelope spectrum; IMF energy deviation rate; Represented as Kroneck function.

6. The method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention as described in claim 1, characterized in that, In S4, the method for obtaining the cross-domain fusion feature vector includes: inputting the multi-source heterogeneous coupling feature vector into a dual-branch time-frequency feature extraction network to obtain the cross-domain fusion feature vector; The dual-branch time-domain feature extraction network includes a time-domain branch and a frequency-domain branch; The temporal branch takes the temporal heterogeneous features in the multi-source heterogeneous coupling feature vector as input, and uses a multi-scale one-dimensional convolutional neural network combined with data quality weights to perform weighted modeling of temporal information to obtain the comprehensive encoding of the multi-source heterogeneous coupling feature vector. The frequency domain branch takes the frequency domain heterogeneous features in the multi-source heterogeneous coupling feature vector as input, formats the frequency domain heterogeneous features into a time-frequency graph, and uses a two-dimensional convolutional neural network to extract cross-field frequency domain collaborative features of different frequency regions from the time-frequency graph. Construct a multi-source domain bias vector based on the depth features of the multi-source signals; Max pooling and average pooling operations are performed on the comprehensive coding and the frequency domain collaborative features to obtain time domain pooling features and frequency domain pooling features; The temporal pooling features, the frequency pooling features, and the multi-source domain bias vector are concatenated and input into a fully connected layer. After passing through channel compression and expansion operations, adaptive channel attention weights are generated at the output through Sigmoid activation. The adaptive channel attention weights are then applied to the temporal heterogeneous features and the frequency heterogeneous features in the multi-source heterogeneous coupling feature vector to obtain the cross-domain fusion feature vector.

7. The method for diagnosing gearbox coupling faults in wind turbines based on multi-source heterogeneous features and dual-branch time-frequency attention as described in claim 1, characterized in that, In S4, a gearbox diagnostic model is used to process the cross-domain fusion feature vector to obtain the gearbox coupling fault diagnosis result; the method of the gearbox diagnostic model processing the cross-domain fusion feature vector includes: Support Vector Machine (SVM) uses a basis kernel function to perform a preliminary nonlinear partitioning of the cross-domain fusion feature vectors, mapping the original low-dimensional space to a high-dimensional space. The deep neural network uses the Softmax function to calculate the probability distribution of multi-class faults on the cross-domain fusion feature vector processed by the support vector machine, and obtains the gearbox coupling fault diagnosis results.

Images