A fault diagnosis method for wind turbine pitch bearings based on signal coding

Abstract

A fault diagnosis method for wind turbine pitch bearings based on signal coding is disclosed, specifically comprising: acquiring raw vibration signal data of the pitch bearing through the wind turbine condition monitoring system equipped in the wind farm; encoding the bearing vibration signal into a time-frequency diagram and a position matrix diagram respectively using continuous wavelet transform and position matrix transform, thereby enhancing the signal feature information by increasing the visualization dimension of the data; designing global and local feature extraction modules based on global and window attention mechanisms respectively, enabling the model to fully learn feature information at different levels; and constructing a dual-channel parallel fusion network model based on the global and local feature extraction modules to fuse global and local feature information and achieve the purpose of suppressing strong random noise interference. This invention can fully capture the vibration signal characteristics of pitch bearings in different states and effectively solve the problems of poor diagnostic performance and low accuracy of pitch bearings caused by strong random noise interference.

Description

Technical Field

[0001] This invention belongs to the fields of wind power generation technology and bearing fault diagnosis technology, and in particular relates to a fault diagnosis method for wind turbine pitch bearings based on signal coding. Background Technology

[0002] Wind energy, as a clean and pollution-free renewable energy source, is playing an increasingly important role in the power generation sector. According to the latest report released by the Global Wind Energy Council (GWEC), global installed capacity reached 117 GW in 2023, a 50% increase over the previous year, bringing the cumulative installed capacity to 1021 GW. However, due to limitations in wind resource distribution and land resources, wind farms are mostly located in harsh environments with poor transportation, such as deserts, mountains, or offshore areas. Wind turbines, as the main power generation carriers, operate in these harsh environments for extended periods, leading to a high failure rate and severely impacting their operational reliability, thus keeping the operation and maintenance costs of wind farms high.

[0003] Research indicates that the main source of wind turbine failures is related to the pitch control system. The pitch control system regulates the rotor speed and power output by adjusting the blade angle, playing a crucial role in the continuous, safe, and stable operation of the wind turbine. The pitch bearing, a core component of the pitch control system, connects the blades to the pitch drive motor and bears the enormous alternating load impact from the wind on the blades, making it one of the components with a high failure rate in wind turbines. If a pitch bearing failure is not detected and effectively addressed in a timely manner, it will gradually cause the pitch control system to malfunction, triggering a chain reaction and potentially leading to catastrophic accidents such as turbine runaway. Furthermore, repairing these failures faces numerous challenges, including inconvenient transportation and a lack of auxiliary maintenance equipment. The high installation location of the pitch bearing also makes it difficult for maintenance personnel to complete repairs quickly, resulting in significant power generation losses for the wind farm. Timely detection, identification, and location of pitch bearing failures provide maintenance personnel with more time to develop reasonable operation and maintenance plans, minimizing prolonged downtime due to failures and thus improving the stability of wind turbine operation and the economic benefits of the wind farm. Therefore, fault diagnosis research on pitch bearings has significant academic and engineering application value.

[0004] Currently, pitch bearing fault diagnosis methods are mainly divided into two categories: physical model methods and data-driven methods. Physical model methods primarily study the intrinsic relationship between dynamic parameters and fault modes under bearing fault conditions. However, this method relies heavily on prior knowledge and expert experience, making it difficult to accurately establish fault diagnosis models for complex operating conditions. Data-driven methods, on the other hand, learn fault characteristics from monitoring signal data or historical data through a model. This method does not require extensive expert knowledge or reasoning about internal mechanisms.

[0005] High-frequency vibration signals are among the most commonly used monitoring data for many complex mechanical devices, reflecting a wealth of mechanical fault information. For example, standard ISO 16079-2 emphasizes the importance of vibration measurement for early and accurate fault detection. When a pitch bearing fails, the faulty component periodically collides with other contact parts, generating periodic vibration pulses. These vibration signals are collected and transmitted by the wind turbine condition monitoring system (CMS). Therefore, using vibration signal information for fault diagnosis of wind turbine pitch bearings can promptly detect, identify, and locate faults, and is currently the most efficient fault diagnosis method.

[0006] With the rapid development of deep learning technology, bearing fault diagnosis methods based on data-driven approaches and neural network models have been widely applied. For example, Li et al. extracted high-level discriminative features from the perspective of sequences using bidirectional LSTM, effectively improving the bearing fault diagnosis performance under varying operating conditions. Wen et al. proposed a Siamese neural network that integrates multi-source features, improving the fault diagnosis performance of bearings with small sample sizes. Chen et al. proposed a parameterless adaptive rectified linear unit network model for fault diagnosis under varying operating conditions. Furthermore, Huo et al. constructed a framework for detecting and identifying rotor-bearing system faults based on a convolutional neural network (CNN) model, using vibration signal data and infrared thermal imaging data. Han Songyu et al. addressed the problem of low diagnostic accuracy in high-speed bearings due to the instability and imbalance of vibration data by proposing an enhanced convolutional neural network based on adaptive weights and multi-scale convolution. However, CNNs are limited by local perception, have weak global feature extraction capabilities, and are highly susceptible to noise interference, making it difficult to capture more comprehensive fault information. Transformer models, on the other hand, have strong feature extraction and long sequence modeling capabilities, helping to overcome the limitations of the aforementioned research.

[0007] In recent years, scholars have gradually explored the application of Transformer models in the field of fault diagnosis. To improve the end-to-end fault diagnosis performance of gearboxes in high-noise environments, Han et al. proposed a robust feature extraction framework that integrates global and local information, leveraging the Transformer model's advantage in capturing global dependencies. Wang et al. proposed a fault diagnosis framework based on the Transformer model, oriented towards high noise and small sample sizes, to improve the performance of small-sample fault diagnosis of mechanical equipment under strong noise interference. Building on this, Yan et al. proposed a lightweight fault diagnosis framework to reduce model complexity. This framework further reduces the parameters and computational cost of model learning through separable multi-scale convolution and broadcast self-attention modules, and experimentally verified the effectiveness and feasibility of this method.

[0008] Due to load disturbances caused by aerodynamic effects, the vibration signal of pitch bearings exhibits strong non-stationarity. Combined with vibration interference from the rotation of wind turbine blades, this results in random noise of varying intensities within the data. Extracting sensitive fault features from non-stationary vibration signals with strong random noise is a challenging task. To address this problem, Liu et al. proposed a multi-scale entropy empirical mode decomposition method based on empirical mode decomposition and applied it to the fault diagnosis of rolling bearings. Wang et al. employed a group decomposition algorithm and extracted the original signal and signal features decomposed at different scale factors using an improved multi-scale inverse discrete entropy algorithm, effectively reducing noise interference to the diagnostic model. However, signal decomposition-based methods suffer from problems such as mode aliasing, imperfect high-order harmonic decomposition, and long decomposition time, significantly limiting their practical application. In recent years, some scholars have attempted to convert time-series signal data into two-dimensional images to enhance key information in signal features, demonstrating significant application value in the field of equipment fault diagnosis. For example, Liang et al. used the Gramian Angular Field (GAF) algorithm to encode one-dimensional vibration signals into image signals and accurately diagnosed bearing fault types using a model based on SVD and a 2-D Residual Neural Network (ResNet). Fu et al. used the Continuous Wavelet Transform (CWT) algorithm to convert the original vibration signal into a two-dimensional time-frequency image for bearing fault diagnosis. Yang et al. combined the GAF and CWT algorithms to convert the original vibration signal of the wind turbine gearbox into a GAF matrix diagram and a time-frequency diagram, respectively, and completed fault type identification through a self-attention parallel fusion network.

[0009] It is worth noting that while current methods have improved fault diagnosis performance to some extent, several challenges remain, specifically: (1) Using a single feature as input limits the feature space perspective and makes it easy to lose key information. (2) The process of encoding signal data into images is complex and suffers from noise aliasing. (3) There is a lack of verification of the model's robustness and generalization ability under non-stationary and strong random noise conditions. Therefore, current methods for pitch bearing fault diagnosis have many shortcomings, and there is an urgent need to explore more advanced fault diagnosis methods to improve the performance of pitch bearing fault diagnosis. Summary of the Invention

[0010] This invention provides a fault diagnosis method for pitch bearings in wind turbines based on signal coding, which can fully capture the vibration signal characteristics of pitch bearings in different states and effectively solve the problems of poor diagnostic performance and low accuracy of pitch bearings caused by strong random noise interference.

[0011] The technical solution adopted in this invention is as follows:

[0012] A fault diagnosis method for wind turbine pitch bearings based on signal coding includes the following steps:

[0013] Step 1: Data acquisition. Obtain the raw vibration signal data of the pitch bearing from the wind turbine condition monitoring system equipped in the wind farm.

[0014] Step 2: Signal encoding. The one-dimensional original vibration signal is encoded into a two-dimensional time-frequency diagram and a position matrix diagram by using continuous wavelet transform and position matrix transform, respectively. The characteristic information of the signal is enhanced by increasing the visualization dimension of the data.

[0015] Step 3: Construction of the ADCPFN fault diagnosis model with dual-channel parallel fusion network based on attention mechanism. A global feature extraction module and a local feature extraction module are designed based on global and window attention mechanisms, respectively. They are then integrated through a dual-channel parallel network structure and a feature adaptive fusion module to form the ADCPFN fault diagnosis model with dual-channel parallel fusion network based on attention mechanism.

[0016] Step 4: Training the dual-channel parallel fusion network ADCPFN fault diagnosis model using the attention mechanism. The ADCPFN model is trained based on the two-dimensional time-frequency diagram and the position matrix diagram. Finally, the trained ADCPFN model completes the diagnosis of pitch bearing faults.

[0017] Further:

[0018] The Global Feature Extraction Module (GFEM) mentioned in step 3 uses the time-frequency map generated by signal encoding as input data. It is designed based on the Global Attention Mechanism (GAM) and includes convolutional layers and attention mechanism layers.

[0019] The convolutional layer consists of Conv2D and BN&PReLU. Conv2D is used to initially capture spatial features and reduce the complexity of the model. BN&PReLU controls the propagation of gradients during model training, avoids gradient explosion and vanishing, and accelerates the convergence speed of model training.

[0020] The attention mechanism layer consists of attention mechanism blocks. By focusing on and recalibrating important information in the feature map, it adaptively learns the weights of different features, making the model pay more attention to features containing rich information. The process is as follows: First, the input feature map is transformed to obtain F1, and then transformed back to its original dimension after passing through a multilayer perceptron (MLP) to obtain F2; second, two convolution operations are performed using a 7×7 kernel to obtain F3; then, max pooling and average pooling are performed on the feature map respectively, and the results are fused to obtain F4, compressing the number of channels and reducing the amount of computation; finally, the output feature F is processed by Sigmoid. o Assume the input feature map F i ∈R C×H×W The specific feature extraction process is as follows:

[0021] F1 = F i ′∈R W×H×C (11)

[0022]

[0023] Where F i ′ represents F i Feature map after dimensionality transformation σ represents element-wise multiplication, and σ represents the convolution operation.

[0024] The Local Feature Extraction (LFEM) module described in step 3 uses the transformed position matrix map as input data and is designed based on the Window Attention (WAM) mechanism. This module consists of four stages: First, the position matrix map is input to the patch segmentation layer in Stage 1 for segmentation into multiple non-overlapping patches of size 4×4 and linearly embedded. Then, the Wam Transformer Block outputs a feature map of size H / 4×W / 4. Next, the extracted feature map is sequentially passed to the patch merging layer and Wam Transformer Block in Stages 2 and 3 for the same operation, and the output feature maps are of sizes H / 8×W / 8 and H / 16×W / 16, respectively. Finally, the extracted feature map is processed by the Sparsetoken converter and Transformer block in Stage 4 to obtain the final local feature map of size H / 32×W / 32.

[0025] The calculation formula for the Wam Transformer Block is as follows:

[0026]

[0027] Where WAM represents the attention of the current window, S-WAM represents the attention after the window is shifted, and LayerNorm represents the normalization process for each sample. This represents the output feature of WAM, z l This represents the output characteristics of a multilayer perceptron (MLP).

[0028] The ADCPPFN fault diagnosis model with attention mechanism described in step 3 uses a multi-feature adaptive fusion of global spatiotemporal information extracted by the global feature extraction module GFEM in the time-frequency map and local position change information extracted by the local feature extraction module LFEM in the position matrix map. It consists of a GFEM module, an LFEM module, an adaptive fusion module, and an output module. This ADCPPFN fault diagnosis model integrates the GFEM module and the LFEM module through a parallel dual-channel network structure, and uses an adaptive fusion module to fuse the acquired global and local feature information, as shown in formula (20), where P represents adaptive average pooling, F G ,F L These represent global and local features, respectively.

[0029] F = P (concat (F) G , F L (20)

[0030] ADCPFN uses the AdamW gradient descent algorithm for supervised training, and the core calculations are shown in formulas (21) to (23).

[0031]

[0032] Where w is the weight coefficient, b is the bias term, α and η represent the learning rate and regularization coefficient, respectively, and E = Loss(w,b,x,y).

[0033] This invention first employs Continuous Wavelet Transform (CWT) and Position Matrix Transform (PMT) to encode the one-dimensional original vibration signal into a two-dimensional time-frequency diagram and a position matrix diagram, respectively, to enhance the feature information of the data. Secondly, based on the signal-encoded images, a dual-channel parallel fusion network with an attention mechanism (ADCPFN) is used to fully learn the fault features, thereby improving the model's diagnostic performance. Finally, the trained ADCPPN model completes the diagnosis of pitch bearing faults. This invention can effectively improve the diagnostic performance and accuracy of pitch bearing faults under non-stationary conditions and strong random noise. Its main innovations are as follows:

[0034] The original one-dimensional vibration signal was encoded into a time-frequency diagram and a position matrix diagram using CWT and PMT, respectively. This increased the visualization dimension of the data to enhance the characteristic information of the fault. At the same time, the input of multi-dimensional features avoided the loss of key information and effectively improved the diagnostic accuracy of the model.

[0035] A global feature extraction module and a local feature extraction module were constructed based on global and window attention mechanisms, respectively, to extract global spatiotemporal information from the fault time-frequency map and local position change information from the position matrix map. This enables the model to fully learn feature information at different levels and improve its diagnostic performance.

[0036] A fault diagnosis model based on an attention mechanism with dual-channel parallel fusion network is proposed. The model integrates global and local feature extraction modules using a dual-channel parallel network structure, and fuses the captured global and local feature information through an adaptive fusion module. This effectively suppresses the interference of strong random noise on health status identification and improves the robustness and generalization ability of the model. Attached Figure Description

[0037] Figure 1 This is a flowchart of the attention mechanism block of the present invention;

[0038] Figure 2 This is a flowchart of the LFEM module of the present invention;

[0039] Figure 3 This is a schematic diagram of the Wam Transformer block structure of the present invention;

[0040] Figure 4 This is a flowchart of the ADCPPFN model of the present invention;

[0041] Figure 5 This invention includes the original signal, the encoded time-frequency diagram and position matrix diagram, and the corresponding tags;

[0042] Figure 6 This is the three-dimensional confusion matrix of the diagnostic results under different noise intensities in this invention;

[0043] Figure 7 This is a flowchart of the present invention. Specific implementation methods

[0044] The invention will be further described below with reference to the accompanying drawings and experiments.

[0045] Reference Figure 7 A method for fault diagnosis of wind turbine pitch bearings based on signal coding, specifically including the following steps:

[0046] Step 1: Data acquisition. Obtain the raw vibration signal data of the pitch bearing from the wind turbine condition monitoring system equipped in the wind farm.

[0047] Step 2: Signal encoding. The one-dimensional original vibration signal is encoded into a two-dimensional time-frequency diagram and a position matrix diagram by using continuous wavelet transform and position matrix transform, respectively. The characteristic information of the signal is enhanced by increasing the visualization dimension of the data.

[0048] Step 3: Construction of the ADCPFN fault diagnosis model with dual-channel parallel fusion network based on attention mechanism. A global feature extraction module and a local feature extraction module are designed based on global and window attention mechanisms, respectively. They are then integrated through a dual-channel parallel network structure and a feature adaptive fusion module to form the ADCPFN fault diagnosis model with dual-channel parallel fusion network based on attention mechanism.

[0049] Step 4: Training the dual-channel parallel fusion network ADCPFN fault diagnosis model using the attention mechanism. The ADCPFN model is trained based on the two-dimensional time-frequency diagram and the position matrix diagram. Finally, the trained ADCPFN model completes the diagnosis of pitch bearing faults.

[0050] The following details each step:

[0051] (1) Continuous Wavelet Transform (CWT)

[0052] CWT, as a highly efficient signal analysis method, can perform flexible and precise multi-scale analysis of signals in the time and frequency domains, playing a crucial role in engineering fields such as signal denoising, signal analysis, and fault diagnosis. Furthermore, when processing non-stationary signals, CWT can effectively separate different frequency components by adjusting scaling and translation parameters and selecting appropriate wavelet basis functions, enabling refined analysis of the signal in both time and frequency domains, accurately identifying the main features of the signal even in noisy environments. Compared to time-frequency analysis methods such as Fourier transform, it can better describe the instantaneous characteristics of signals and reveal hidden features and patterns within them. The specific definition of CWT is as follows:

[0053]

[0054] In the formula, x(t) represents the original vibration signal, τ is the translation parameter used to control translation, and α is the scale factor used to control scaling. Represents continuous wavelet basis functions. This represents the complex conjugate of the continuous wavelet basis functions.

[0055] Choosing appropriate wavelet basis functions can significantly improve the wavelet transform effect. During the transform process, signals similar to the wavelet basis function are amplified, while signals with large differences are suppressed. In selecting wavelet basis functions, this invention found that the Morlet wavelet has the highest similarity to the pitch bearing vibration signal; therefore, this invention uses the Morlet wavelet as the basis function for continuous wavelet transform. The functional expression of the Morlet wavelet is as follows:

[0056]

[0057] In the formula, ω0 is the wavelet frequency.

[0058] (2) Position Matrix Transformation PMT

[0059] PMT (Programmable Graph Transformer) is commonly used to encode time-series data into two-dimensional images, enabling task classification and recognition using advanced computer vision techniques. In the converted image, darker areas represent lower values ​​in the original time series, and vice versa, while solid-color areas indicate unchanged or slowly changing values. For the waveform of the original time series, PMT represents it using the frequency of color changes from dark to light or light to dark in the image. This means that jagged waveforms will produce high-frequency color changes, while smooth waveforms will produce low-frequency color changes.

[0060] This invention employs Partial Motion Modeling (PMT) to encode the original vibration signal, thereby enhancing the local feature information within the signal and making it easier for the model to identify the differences between each health state from the encoded position matrix diagram. The specific PMT process is as follows:

[0061] 1) The original vibration signal X = {x} t The sequence Z = z1, z2, ..., zn is standardized using formula (3) to obtain a standard normal distribution signal sequence. N , where μ and σ are the mean and standard deviation of X, respectively.

[0062]

[0063] 2) Using the piecewise aggregation approximation (PAA) method, a suitable reduction factor k is selected to generate a new smooth signal sequence through formula (4). The sequence Z is reduced to m dimensions by calculating the average of the piecewise constants, while maintaining the approximate trend of the original signal sequence.

[0064]

[0065] 3) Calculate the relative position between the two timestamps and process the signal sequence. Convert to a two-dimensional matrix M, and the conversion equation is as shown in formula (6).

[0066]

[0067] 4) Use min-max normalization to convert M into a gray value matrix, and finally obtain the relative displacement matrix F through formula (7).

[0068]

[0069] (3) Attention mechanism

[0070] Attention mechanisms, widely used in deep learning, dynamically allocate weights based on the importance of input features during neural network model training. This suppresses redundant features and enhances key information, enabling the model to better handle complex tasks. An attention mechanism can be viewed as a combination of a query matrix (q), keys (k), and values ​​(v). First, the correlation between q and different k values, Sim(q,k), is calculated. i ), that is, the weight coefficients for different v values, as shown in formula (8). Next, the obtained weight coefficient values ​​are mapped to the range [0,1], and the probability distribution A={a1,a2,..,a...} of the correlation between each element is obtained through the softmax function. d}, as shown in formula (9). Finally, the v values ​​are weighted and summed according to the weight coefficients to obtain the final attention value, as shown in formula (10).

[0071]

[0072] In the formula, d represents the dimension of the input vector.

[0073] (4) Global Feature Extraction Module

[0074] In order to obtain global spatiotemporal information from vibration signals with strong random noise, this invention uses the time-frequency map generated by signal encoding as input data and designs a global feature extraction module (GFEM) based on the global attention mechanism (GAM). This module includes convolutional layers and attention mechanism layers.

[0075] The convolutional layer consists of Conv2D and BN&PReLU. Conv2D is used to initially capture spatial features and reduce the complexity of the model. BN&PReLU controls the propagation of gradients during model training, avoids gradient explosion and vanishing, and accelerates the convergence speed of model training.

[0076] As the core of GFEM, the attention mechanism layer in this invention is based on the design concept of channel and spatial attention mechanisms in GAM, and reconstructs the attention mechanism block, such as... Figure 1As shown, by focusing on and recalibrating important information in the feature map, the model adaptively learns the weights of different features, making the model pay more attention to those features containing rich information. First, the input feature map is transformed to obtain F1, and then transformed back to its original dimension after passing through a multilayer perceptron (MLP) to obtain F2, so as to amplify the dependence of cross-dimensional channels, highlighting important features while suppressing irrelevant or noisy features; secondly, two convolution operations are performed using a 7×7 convolution kernel to obtain F3, capturing information from different spatial locations in the feature map; then, max pooling and average pooling are performed on the feature map respectively, and the results are fused to obtain F4, compressing the number of channels and reducing the amount of computation; finally, the output feature F is processed by Sigmoid. o Assume the input feature map F i ∈R C×H×W The specific feature extraction process is as follows:

[0077] F1 = F i ′∈R W×H×C (11)

[0078]

[0079] Where F i ′ represents F i Feature map after dimensionality transformation σ represents element-wise multiplication, and σ represents the convolution operation.

[0080] (5) Local Feature Extraction Module

[0081] To capture more comprehensive fault information from vibration signals subjected to strong random noise interference, while maintaining attention to global features, it is also necessary to enhance the acquisition of local features. Therefore, this invention uses the transformed position matrix map as input data and designs a Local Feature Extraction (LFEM) module based on the Window Attention (WAM) mechanism, as follows... Figure 2 As shown, this module consists of four stages. First, the position matrix map is input into the patch segmentation layer in Stage 1 for segmentation into multiple non-overlapping patches of size 4×4, which are then linearly embedded. This reduces the size of the feature map while preserving feature space information. The Wam Transformer block then outputs a feature map of size H / 4×W / 4. Next, the extracted feature map is sequentially passed to the patch merging layer and Wam Transformer block in Stages 2 and 3” for the same operation, resulting in output feature maps of sizes H / 8×W / 8 and H / 16×W / 16, respectively. Finally, the extracted feature map is processed by the Sparsetoken converter and Transformer block in Stage 4 to obtain the final local feature map, with a size of H / 32×W / 32.

[0082] LFEM replaces the multi-head self-attention in the Transformer network with WAM and reconstructs the WamTransformerblock, as shown in the following structure. Figure 3 As shown. The goal is to limit attention computation to a single window, extracting local features while significantly reducing model computation. Furthermore, to reduce the number of Transformer parameters and make the model more efficient, this invention introduces a sparse token converter, which is combined with the Transformer block as the fourth stage of LFEM, to reduce the number of tokens used in LFEM.

[0083] Compared to the traditional Transformer network architecture, WAM performs computation independently within local windows using self-attention, effectively addressing the significant computational challenges posed by global self-attention mechanisms in dense image classification tasks. The windows used uniformly divide the image in a non-overlapping manner, allowing the Wam Transformer block to capture more detailed local visual features while improving computational efficiency. The calculation formula for consecutive Wam Transformer blocks is as follows:

[0084]

[0085]

[0086] Where WAM represents the attention of the current window, S-WAM represents the attention after the window is shifted, and LayerNorm represents the normalization process for each sample. This represents the output feature of WAM, z l This represents the output characteristics of a multilayer perceptron (MLP).

[0087] (6) Attention mechanism dual-channel parallel fusion network

[0088] Accurate identification of the health status of pitch bearings under non-stationary conditions and strong random noise requires sufficient signal feature information. This invention comprehensively considers both global and local features, proposing an attention mechanism dual-channel parallel fusion network (ADCPFN) model. Its main idea is to adaptively fuse multi-feature information from the global spatiotemporal information extracted by GFEM in the time-frequency diagram and the local position change information in the position matrix diagram extracted by LFEM to obtain sufficient fault feature information. Its structure is as follows: Figure 4 As shown, it includes GFEM, LFEM, adaptive fusion module and output module.

[0089] ADCPFN integrates GFEM and LFEM through a parallel dual-channel network structure design, and uses an adaptive fusion module to fuse the acquired global and local feature information, as shown in formula (20), where P represents adaptive average pooling, F G ,F L These represent global and local features, respectively. The dual-channel input structure of the ADCPFN model can more effectively combine visual information from different levels, suppress the interference of strong random noise on fault identification, and thus improve model performance and generalization ability.

[0090] F = P (concat (F) G , F L (20)

[0091] ADCPFN uses the AdamW gradient descent algorithm for supervised training, and the core calculations are shown in formulas (21) to (23).

[0092]

[0093] Where w is the weight coefficient, b is the bias term, α and η represent the learning rate and regularization coefficient, respectively, and E = Loss(w,b,x,y).

[0094] (7) Experimental verification

[0095] Taking bearing data from Case Western Reserve University as an example, to simulate the real operating environment of wind turbine pitch bearings, random Gaussian noise and random scale noise are sequentially added to the dataset with a 50% probability, where λ is the intensity of the random noise. According to the present invention, firstly, CWT and PMT are used to encode the original one-dimensional vibration signal into a time-frequency diagram and a position matrix diagram, respectively. An example of the encoding results is shown below. Figure 5 As shown, it can be observed that the differences in features of the encoded two-dimensional image are more obvious compared to the one-dimensional vibration signal, indicating that increasing the visualization dimension of the data amplifies the differences in vibration signals of bearings under different states and can significantly enhance the characteristic information of the fault.

[0096] Table 1 shows the accuracy of the bearing fault diagnosis results of this invention. The three-dimensional confusion matrix of the health state diagnosis results under different intensities of random noise is as follows: Figure 6 As shown.

[0097] Table 1. Accuracy of bearing fault diagnosis results using the method of the present invention

[0098]

[0099] It can be intuitively observed that when the random noise intensity is 0.2, the diagnostic accuracy for all health states is higher than 94%, and the diagnostic results for 7 states reach 100%. When the noise intensity increases to 0.4, the diagnostic results for 8 states reach 100%. However, it is worth noting that the accuracy for state 2 is 97% when the noise intensity is 0.2, while the diagnostic accuracy for state 2 reaches 100% when the noise intensity is 0.4. This is mainly due to two reasons: firstly, the randomness of the noise addition and the independence between datasets with different noise intensities lead to a certain degree of randomness in the diagnostic results; secondly, the noise intensity of 0.4 does not interfere with the signal fluctuation characteristics of state 2, or it complements the signal fluctuations of state 2, suppressing the noise interference. When the noise intensity increases to 0.6, it can be significantly observed that strong random noise interferes with the diagnosis of most health states, reducing the diagnostic accuracy, but the diagnostic accuracy for 6 health states still remains above 90%. Based on the diagnostic results of various health states under different noise intensities, the present invention exhibits relatively consistent diagnostic results when facing low to medium intensity random noise interference, showing minimal impact from noise and demonstrating good noise resistance. While the diagnostic accuracy of some health states decreases when facing strong random noise interference, overall, it can still effectively avoid the influence of noise to a certain extent, demonstrating good robustness.

Claims

1. A method for fault diagnosis of pitch bearings in wind turbine units based on signal coding, characterized in that, Includes the following steps: Step 1: Data acquisition. Obtain the raw vibration signal data of the pitch bearing from the wind turbine condition monitoring system equipped in the wind farm. Step 2: Signal encoding. The one-dimensional original vibration signal is encoded into a two-dimensional time-frequency diagram and a position matrix diagram by using continuous wavelet transform and position matrix transform, respectively. The characteristic information of the signal is enhanced by increasing the visualization dimension of the data. Step 3: Construction of the ADCPFN fault diagnosis model with dual-channel parallel fusion network based on attention mechanism. A global feature extraction module and a local feature extraction module are designed based on global and window attention mechanisms, respectively. They are then integrated through a dual-channel parallel network structure and a feature adaptive fusion module to form the ADCPFN fault diagnosis model with dual-channel parallel fusion network based on attention mechanism. Step 4: Training the dual-channel parallel fusion network ADCPFN fault diagnosis model using the attention mechanism. The ADCPFN model is trained based on the two-dimensional time-frequency diagram and the position matrix diagram. Finally, the trained ADCPFN model completes the diagnosis of pitch bearing faults.

2. The method for fault diagnosis of wind turbine pitch bearings based on signal coding according to claim 1, characterized in that, The Global Feature Extraction Module (GFEM) mentioned in step 3 uses the time-frequency map generated by signal encoding as input data. It is designed based on the Global Attention Mechanism (GAM) and includes convolutional layers and attention mechanism layers. The convolutional layer consists of Conv2D and BN&PReLU. Conv2D is used to initially capture spatial features and reduce the complexity of the model. BN&PReLU controls the propagation of gradients during model training, avoids gradient explosion and vanishing, and speeds up the convergence of model training. The attention mechanism layer consists of attention mechanism blocks. By focusing on and recalibrating important information in the feature map, it adaptively learns the weights of different features, making the model pay more attention to those features that contain rich information. The process is as follows: First, the input feature map is dimensionality-transformed to obtain F1, and then converted back to its original dimensions after passing through a multilayer perceptron (MLP) to obtain F2; second, two convolution operations are performed using a 7×7 kernel to obtain F3; then, max pooling and average pooling are performed on the feature map respectively, and the results are fused to obtain F4, compressing the number of channels and reducing computational cost; finally, the output feature map is processed by a sigmoid function to output F. o Assume the input feature map F i ∈R C×H×W The specific feature extraction process is as follows: F1=F i ′∈R W×H×C (11) Where F i ′ represents F i Feature map after dimensionality transformation σ represents element-wise multiplication, and σ represents the convolution operation.

3. The method for fault diagnosis of wind turbine pitch bearings based on signal coding according to claim 1, characterized in that, The Local Feature Extraction (LFEM) module described in step 3 uses the transformed position matrix map as input data and is designed based on the Window Attention (WAM) mechanism. This module consists of four stages: First, the position matrix map is input to the patch segmentation layer in Stage 1 for segmentation into multiple non-overlapping patches of size 4×4 and linearly embedded. Then, the WamTransformer Block outputs a feature map of size H / 4×W / 4. Next, the extracted feature map is sequentially passed to the patch merging layer and WamTransformer Block in Stages 2 and 3 for the same operation, and the output feature maps are of sizes H / 8×W / 8 and H / 16×W / 16, respectively. Finally, the extracted feature map is processed by the Sparse tokenconverter and Transformer block in Stage 4 to obtain the final local feature map, which is of size H / 32×W / 32.

4. The method for fault diagnosis of wind turbine pitch bearings based on signal coding according to claim 3, characterized in that, The calculation formula for the Wam Transformer Block is as follows: Where WAM represents the attention of the current window, S-WAM represents the attention after the window is shifted, and LayerNorm represents the normalization process for each sample; Z represents the output feature of WAM. l This represents the output characteristics of a multilayer perceptron (MLP).

5. The method for fault diagnosis of wind turbine pitch bearings based on signal coding according to claim 1, characterized in that, The ADCPPFN fault diagnosis model with attention mechanism described in step 3 uses a multi-feature adaptive fusion of global spatiotemporal information extracted by the global feature extraction module GFEM in the time-frequency map and local position change information extracted by the local feature extraction module LFEM in the position matrix map. It consists of a GFEM module, an LFEM module, an adaptive fusion module, and an output module. This ADCPPFN fault diagnosis model integrates the GFEM module and the LFEM module through a parallel dual-channel network structure, and uses an adaptive fusion module to fuse the acquired global and local feature information, as shown in formula (20), where P represents adaptive average pooling, F G ,F L These represent global and local features, respectively. F=P(concat(F G ,F L )) (20) ADCPFN uses the AdamW gradient descent algorithm for supervised training, and the core calculations are shown in formulas (21) to (23): Where w is the weight coefficient, b is the bias term, α and η represent the learning rate and regularization coefficient, respectively, and E = Loss(w,b,x,y).

Images