Wind turbine generator variable pitch bearing fault diagnosis method based on signal coding

Abstract

A wind turbine generator variable pitch bearing fault diagnosis method based on signal coding specifically comprises the steps that original vibration signal data of a variable pitch bearing are acquired through a wind turbine generator state monitoring system arranged in a wind power plant; bearing vibration signals are respectively coded into a time-frequency diagram and a position matrix diagram by adopting continuous wavelet transformation and position matrix transformation, and the characteristic information of the signals is enhanced by increasing the visual dimension of data; a global feature extraction module and a local feature extraction module are respectively designed based on a global attention mechanism and a window attention mechanism, so that the model can fully learn feature information of different levels; a dual-channel parallel fusion network model is constructed according to global and local feature extraction modules to fuse global and local feature information, and the purpose of suppressing strong random noise interference is achieved. According to the method, the vibration signal characteristics of the variable pitch bearing in different states can be fully captured, and the problems of poor diagnosis performance and low precision of the variable pitch bearing caused by strong random noise interference can be effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of wind power generation and bearing fault diagnosis, and in particular relates to a wind turbine variable pitch bearing fault diagnosis method based on signal coding. Background Art

[0002] As a clean, pollution-free renewable energy source, wind energy plays an increasingly important role in the field of energy generation. According to the latest report released by the Global Wind Energy Council (GWEC), the global installed capacity will be 117GW in 2023, a 50% increase over the previous year, and the cumulative installed capacity has reached 1021GW. However, due to the distribution of wind resources and land resource restrictions, wind farms are mostly located in the Gobi Desert, mountainous areas or sea with harsh environments and inconvenient transportation. As the main carrier of power generation, wind turbines have been operating in harsh environments for a long time, resulting in a high failure rate, which seriously affects the operating reliability of wind turbines and also makes the operation and maintenance costs of wind farms high.

[0003] According to research, the main source of wind turbine failure is related to the pitch system. The pitch system controls the rotor speed and power output by adjusting the windward angle of the wind turbine blades, and plays an important role in the continuous safe and stable operation of the wind turbine. Among them, the pitch bearing, as the core component of the pitch system, connects the blades and the pitch drive motor respectively, and bears the huge alternating load impact of the wind on the blades. It is one of the components with a higher failure rate in the wind turbine. When the pitch bearing fails, if it is not discovered and effectively handled in time, the pitch adjustment system will gradually fail and trigger a chain reaction, and even cause catastrophic accidents such as unit loss of control. In addition, in the process of repairing the fault, there are many unfavorable factors such as inconvenient transportation and lack of auxiliary maintenance equipment. The installation position of the pitch bearing is high from the ground, making it difficult for maintenance personnel to complete the repair in a short time, resulting in great power generation losses in the wind farm. If the pitch bearing fault can be discovered, identified and located in time, it can provide more time for operation and maintenance personnel to formulate a reasonable operation and maintenance plan, and minimize the long-term downtime caused by the fault, thereby improving the stability of wind turbine operation and the economic benefits of the wind farm. Therefore, the fault diagnosis research on pitch bearings has important academic and engineering application value.

[0004] At present, the fault diagnosis methods of variable pitch bearings are mainly divided into two categories: physical model methods and data-driven methods. The physical model method mainly studies the intrinsic relationship between the dynamic parameters and the fault mode under the bearing fault state. This method relies too much on prior knowledge and expert experience, and it is difficult to accurately establish a fault diagnosis model under complex working conditions. The data-driven method uses a model to learn fault characteristics from monitoring signal data or historical data. This method does not require a large amount of expert knowledge and reasoning expression of internal mechanism knowledge.

[0005] Among them, high-frequency vibration signals are the most commonly used monitoring data for many complex mechanical equipment, which can reflect rich mechanical fault information. For example, the standard ISO 16079-2 emphasizes the importance of vibration measurement for early and accurate fault detection. When a variable pitch bearing fails, the faulty component will periodically collide with other contact components, generating periodic vibration pulses, and the wind turbine condition monitoring system (Condition Monitoring System, CMS) will complete the collection and transmission of the generated vibration signals. Therefore, using vibration signal information to diagnose wind turbine variable pitch bearing faults can timely discover, identify and locate faults, which is also the most efficient fault diagnosis method at present.

[0006] With the rapid development of deep learning technology, bearing fault diagnosis methods based on data-driven and neural network models have been widely used. For example, Li et al. extracted high-level discriminant features from the perspective of sequence based on bidirectional LSTM, which effectively improved the bearing fault diagnosis performance under variable working conditions. Wen et al. proposed a Siamese neural network that integrates multi-source features to improve the fault diagnosis performance of bearings under small samples. Chen et al. proposed a network model of parameter-free adaptive rectified linear units for fault diagnosis under variable working conditions. In addition, Huo et al. constructed a framework for detecting and identifying rotor-bearing system faults based on the convolutional neural network (CNN) model using vibration signal data and infrared thermal imaging data. Han Songyu et al. proposed an enhanced convolutional neural network based on adaptive weights and multi-scale convolutions to address the problem of low diagnostic accuracy of high-speed bearings due to the unsteadiness and imbalance of vibration data. However, CNN is limited by local perception, has weak global feature extraction capabilities, and is easily interfered by noise, making it difficult to capture more comprehensive fault information. The Transformer model has strong feature extraction and long sequence modeling capabilities, which helps to solve the limitations of the above research.

[0007] In recent years, some scholars have gradually explored the application of Transformer models in the field of fault diagnosis. In order to improve the end-to-end fault diagnosis performance of gearboxes in strong noise environments, Han et al. took advantage of the Transformer model's advantages in capturing global dependencies and proposed a robust feature extraction framework that integrates global and local information. Wang et al. proposed a fault diagnosis framework for strong noise and small samples based on the Transformer model to improve the performance of small sample fault diagnosis of mechanical equipment under strong noise interference. On this basis, Yan et al. proposed a lightweight fault diagnosis framework to reduce the complexity of the model, and further reduced the parameters and computational complexity of model learning through separable multi-scale convolution and broadcast self-attention modules, and verified the effectiveness and feasibility of this method through experiments.

[0008] Affected by the load disturbance caused by the aerodynamic effect, the vibration signal of the variable pitch bearing has a strong non-stationarity. Coupled with the vibration interference caused by the rotation of the wind turbine blades, the data contains random noise of different intensities. How to extract sensitive fault features from non-stationary and strong random noise vibration signals is a challenging task. To address this problem, Liu et al. proposed a multi-scale entropy empirical mode decomposition method based on the empirical mode decomposition method and applied it to the fault diagnosis of rolling bearings. Wang et al. used a group decomposition algorithm and extracted the original signal and the signal features decomposed under different scale factors through an improved multi-scale inverse discrete entropy algorithm, effectively reducing the interference of noise on the diagnosis model. However, the signal decomposition-based method has problems such as mode aliasing, unsatisfactory high-order harmonic decomposition, and long decomposition time, which greatly limits the application of such methods in practice. In recent years, some scholars have tried to convert time series signal data into two-dimensional images to enhance the key information of signal features, which is showing great application value in the field of equipment fault diagnosis. For example, Liang et al. used the Gramian angular field (GAF) algorithm to encode the one-dimensional vibration signal into an image signal, and accurately diagnosed the fault type of the bearing through a model based on SVD and 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 the identification of the fault type through a self-attention parallel fusion network.

[0009] It is worth noting that although the current methods have improved the fault diagnosis performance to a certain extent, there are still challenges in some aspects, specifically: (1) Using a single feature information as input, the feature space perspective is limited, and key information is easily lost. (2) The process of encoding signal data into an image is complex and there is a problem of aliasing with noise. (3) There is a lack of verification of the robustness and generalization ability of the model under non-stationary and strong random noise. Therefore, there are many shortcomings in the current methods for variable pitch bearing fault diagnosis, and it is urgent to explore more advanced fault diagnosis methods to improve the performance of variable pitch bearing fault diagnosis. Summary of the invention

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

[0011] The technical solution adopted by the present invention is:

[0012] A wind turbine pitch bearing fault diagnosis method based on signal coding comprises the following steps:

[0013] Step 1: Data acquisition: obtaining the original vibration signal data of the variable pitch bearing from the wind turbine condition monitoring system equipped in the wind farm;

[0014] Step 2: Signal encoding: Continuous wavelet transform and position matrix transform are used to encode the one-dimensional original vibration signal into a two-dimensional time-frequency diagram and a position matrix diagram, respectively, to enhance the characteristic information of the signal by increasing the visualization dimension of the data;

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

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

[0017] Further:

[0018] The global feature extraction module GFEM in step 3 uses the time-frequency diagram generated by signal coding as input data and is designed based on the global attention mechanism GAM, which includes a convolution layer and an attention mechanism layer;

[0019] The convolution layer consists of Conv2D and BN&PReLU, where Conv2D is used to initially capture spatial features and reduce the complexity of the model; BN&PReLU controls the transmission of model training gradients to avoid gradient explosion and disappearance and speed up the convergence of model training.

[0020] The attention mechanism layer is composed of attention mechanism blocks. By focusing on and recalibrating the important information in the feature map, it adaptively learns the weights of different features, so that the model pays more attention to those features containing rich information. The process is as follows: first, the input feature map is transformed into F1, and then converted to the original dimension after passing through the multi-layer perceptron MLP to obtain F2; secondly, two convolution operations are performed using a convolution kernel of 7×7 to obtain F3; then, the feature map is subjected to maximum and average pooling respectively, and the results are fused to obtain F4, which compresses the number of channels and reduces the amount of calculation; finally, the feature F is output after Sigmoid processing. o ; Assume that 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] Among them, F i ′ represents F i The feature map after dimension conversion, represents element-wise multiplication and σ is the convolution operation.

[0024] The local feature extraction module LFEM described in step 3 takes the position matrix map generated by the conversion as input data and is designed based on the window attention mechanism WAM; the module is divided into four stages: first, the position matrix map is input into the patch segmentation layer in Stage 1 for segmentation, and is segmented 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; then, the extracted feature map is transmitted to the patch merging layer and Wam Transformer Block in Stage 2 and Stage 3 in turn for the same operation, and the sizes of the output feature maps are H / 8×W / 8 and H / 16×W / 16 respectively; finally, the extracted feature map is passed through 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 of the Wam Transformer Block is as follows:

[0026]

[0027] Among them, WAM represents the attention of the current window, S-WAM represents the attention after the window is shifted, and LayerNorm represents the normalization of each sample. represents the output feature of WAM, z l Represents the output features of the multi-layer perceptron MLP.

[0028] The attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model described in step 3 performs multi-feature adaptive fusion on the global spatiotemporal information in the time-frequency diagram extracted by the global feature extraction module GFEM and the local position change information in the position matrix diagram extracted by the local feature extraction module LFEM, and is composed of a GFEM module, a LFEM module, an adaptive fusion module and an output module; the ADCPFN 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 feature information and local feature information, as shown in formula (10), where P represents adaptive average pooling, F G ,F L 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] Among them, 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] The present invention firstly adopts 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 characteristic information of the data. Secondly, based on the image encoded by the signal, the fault characteristics are fully learned through the attention mechanism dual-channel parallel fusion network (ADCPFN) to improve the diagnostic performance of the model. Finally, the diagnosis of the variable pitch bearing fault is completed by the trained ADCPFN model. The present invention can effectively improve the diagnostic performance and accuracy of the variable pitch bearing fault under non-stationary and strong random noise, and its main innovations are as follows:

[0034] CWT and PMT are used to encode the one-dimensional original vibration signal into time-frequency diagram and position matrix diagram respectively. The characteristic information of the fault is enhanced by increasing the visualization dimension of the data. At the same time, the input of multi-dimensional features avoids the loss of key information and effectively improves the diagnostic accuracy of the model.

[0035] Based on the global and window attention mechanisms, a global feature extraction module and a local feature extraction module are constructed respectively to extract the global spatiotemporal information in the fault time-frequency diagram and the local position change information in the position matrix diagram, so that the model can fully learn the feature information at different levels and improve its diagnostic performance.

[0036] A dual-channel parallel fusion network fault diagnosis model based on attention mechanism is proposed. The dual-channel parallel network structure is used to integrate the global and local feature extraction modules, and the captured global and local feature information is fused through the adaptive fusion module. It can effectively suppress the interference of strong random noise on health status recognition and improve the robustness and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0038] Figure 2 This is the flow chart 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 It is the flow chart of ADCPFN model of the present invention;

[0041] Figure 5 The original signal, the encoded time-frequency diagram and the position matrix diagram and the corresponding labels of the present invention;

[0042] Figure 6 is a three-dimensional confusion matrix of the diagnosis results under noises of different intensities of the present invention;

[0043] Figure 7 It is a flow chart of the present invention. Specific implementation methods

[0045] The present invention is further described below in conjunction with the accompanying drawings and experiments.

[0046] Reference Figure 7 A wind turbine pitch bearing fault diagnosis method based on signal coding specifically comprises the following steps:

[0047] Step 1: Data acquisition: obtaining the original vibration signal data of the variable pitch bearing from the wind turbine condition monitoring system equipped in the wind farm;

[0048] Step 2: Signal encoding: Continuous wavelet transform and position matrix transform are used to encode the one-dimensional original vibration signal into a two-dimensional time-frequency diagram and a position matrix diagram, respectively, to enhance the characteristic information of the signal by increasing the visualization dimension of the data;

[0049] Step 3: Construction of attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model. Based on the global and window attention mechanisms, the global feature extraction module and the local feature extraction module are designed respectively, and integrated through the dual-channel parallel network structure and the feature adaptive fusion module to form the attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model;

[0050] Step 4: Training of the attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model. The ADCPFN model is trained based on the two-dimensional time-frequency graph and the position matrix graph. Finally, the trained ADCPFN model completes the diagnosis of the variable pitch bearing fault.

[0051] The following are detailed steps:

[0052] (1) Continuous Wavelet Transform (CWT)

[0053] As an efficient signal analysis method, CWT can perform flexible and sophisticated multi-scale analysis of signals in the time and frequency domains, and plays an important role in engineering fields such as signal denoising, signal analysis, and fault diagnosis. In addition, when processing non-steady-state signals, CWT can effectively separate the different frequency components in the signal by adjusting the scale and translation parameters and selecting appropriate wavelet basis functions, and perform refined analysis of the signal in time and frequency, and accurately identify the main features of the signal even in a noisy background. Compared with time-frequency analysis methods such as Fourier transform, it can better describe the instantaneous characteristics of the signal and reveal the hidden features and patterns in the signal. The specific definition of CWT is as follows:

[0054]

[0055] Where 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 expansion and contraction. represents the continuous wavelet basis function, represents the complex conjugate of the continuous wavelet basis function.

[0056] Selecting a suitable wavelet basis function will significantly improve the effect of wavelet transform. During the transformation process, signals similar to the wavelet basis function will be amplified, and signals with large differences will be suppressed. When selecting the wavelet basis function, the present invention found that the Morlet wavelet has the highest similarity with the vibration signal of the variable pitch bearing, so the present invention uses the Morlet wavelet as the basis function of the continuous wavelet transform. The function expression of the Morlet wavelet is as follows:

[0057]

[0058] Where ω0 is the wavelet frequency.

[0059] (2) Position Matrix Transformation PMT

[0060] PMT is often used to encode time series data into two-dimensional images, and can use advanced computer vision technology to classify and identify tasks. The dark areas in the converted image indicate that the values ​​of the original time series are low, and vice versa, while the solid color areas indicate that the values ​​of the original time series remain unchanged or change slowly. For the waveform of the original time series, PMT is represented by the frequency of color changes from dark to light or from light to dark in the image. This means that a sawtooth waveform will produce high-frequency color changes, and a smooth waveform will produce low-frequency color changes.

[0061] The present invention uses PMT to encode the original vibration signal to enhance the local feature information in the signal, making it easier for the model to identify the differences between each health state from the encoded position matrix diagram. The specific process of PMT is as follows:

[0062] 1) The original vibration signal X = {x t ,t=1,2,…,N} is standardized by formula (3) to obtain a standard normally distributed signal sequence Z=z1,z2,..,z N , where μ and σ are the mean and standard deviation of X, respectively.

[0063]

[0064] 2) Using the piecewise aggregation approximation (PAA) method, select an appropriate reduction factor k and generate a new smooth signal sequence through formula (4): The sequence Z is reduced to m dimensions by calculating the piecewise constant mean while keeping the approximate trend of the original signal sequence.

[0065]

[0066] 3) Calculate the relative position between the two timestamps and convert the processed signal sequence Convert to a two-dimensional matrix M, the conversion equation is as shown in formula (6).

[0067]

[0068] 4) Use minimum-maximum normalization to convert M into a gray value matrix, and finally obtain the relative displacement matrix F through formula (7).

[0069]

[0070] (3) Attention Mechanism

[0071] As a widely used technology in deep learning, the attention mechanism can dynamically assign weights according to the importance of input features during the training process of the neural network model, so as to suppress redundant features and enhance key information. This enables the model to better handle complex tasks. The attention mechanism can be regarded as a combination of query matrix (q), key (k) and value (v). First, the correlation Sim(q,k) between q and different k is calculated. i ), that is, the weight coefficients of different v values, as shown in formula (8). Then, the weight coefficient values ​​are mapped to [0,1], and the probability distribution of the correlation of each element is obtained through the softmax function A = {a1, a2, .., a d}, as shown in formula (9). Finally, the v value is weighted and summed according to the weight coefficient to obtain the final attention value, as shown in formula (10).

[0072]

[0073] Where d represents the dimension of the input vector.

[0074] (4) Global feature extraction module

[0075] In order to obtain global spatiotemporal information from vibration signals with strong random noise, the present invention uses the time-frequency diagram generated by signal encoding as input data and designs a global feature extraction module (GFEM) based on the global attention mechanism (GAM), which includes a convolutional layer and an attention mechanism layer.

[0076] The convolution layer consists of Conv2D and BN&PReLU, where Conv2D is used to initially capture spatial features and reduce the complexity of the model; BN&PReLU controls the transmission of model training gradients to avoid gradient explosion and disappearance and speed up the convergence of model training.

[0077] The attention mechanism layer is the core of GFEM. Based on the design idea of ​​channel and spatial attention mechanism in GAM, this paper reconstructs the attention mechanism block, such as Figure 1As shown. By focusing on and recalibrating the important information in the feature map, the weights of different features are adaptively learned, so that the model pays more attention to those features that contain rich information. First, the input feature map is transformed into F1, and then converted to the original dimension after passing through the multi-layer perceptron (MLP) to obtain F2, so as to amplify the dependency of cross-dimensional channels, highlight important features while suppressing irrelevant or noise features; secondly, two convolution operations are performed with a convolution kernel of 7×7 to obtain F3, which captures information at different spatial positions in the feature map; then the feature map is subjected to maximum and average pooling respectively, and the results are fused to obtain F4, which compresses the number of channels and reduces the amount of calculation; finally, the feature F is output after Sigmoid processing. o Assume that the input feature map F i ∈R C×H×W , the specific feature extraction process is as follows:

[0078] F1=F i ′∈R W×H×C (11)

[0079]

[0080] where F i ′ represents F i The feature map after dimension conversion, represents element-wise multiplication and σ is the convolution operation.

[0081] (5) Local feature extraction module

[0082] In order to capture more comprehensive fault information from vibration signals under strong random noise interference, the acquisition of local features should be strengthened while keeping the focus on global features. To this end, the present invention uses the position matrix map generated by transformation as input data and designs a local feature extraction module (LFEM) based on the window attention mechanism (WAM), such as Figure 2 As shown in the figure, the module is divided into four stages. First, the position matrix map is input into the patch segmentation layer in Stage 1 for segmentation, and is segmented into multiple non-overlapping patches of size 4×4 and linearly embedded. The size of the feature map is reduced while retaining the feature space information. Then the Wam Transformer block outputs a feature map of size H / 4×W / 4. The extracted feature map is then transmitted to the patch merging layer and Wam Transformer block in Stage 2 and Stage 3 in turn for the same operation. The sizes of the output feature maps are H / 8×W / 8 and H / 16×W / 16, respectively. Finally, the extracted feature map passes through the Sparsetoken converter and Transformer blocks in Stage 4 to obtain the final local feature map of size H / 32×W / 32.

[0083] LFEM replaces the multi-head self-attention in the Transformer network with WAM and reconstructs the WamTransformerblock. Its structure is as follows Figure 3 As shown. The purpose is to limit the attention calculation to a window, extract local features and significantly reduce the amount of model calculation. In addition, in order to reduce the number of Transformer parameters and make the model more efficient, the present invention introduces a Sparse token converter and combines it with the Transformer block as the fourth stage of LFEM to reduce the number of tokens used in LFEM.

[0084] Compared with the traditional Transformer network architecture, WAM performs independent calculations in a local window in a self-attention manner, which can better solve the huge computational problems caused by the global self-attention mechanism in image intensive classification tasks. The window used divides the image evenly in a non-overlapping manner, so that the Wam Transformer block can capture local visual features more finely while improving computational efficiency. The calculation formula for a continuous Wam Transformer block is as follows:

[0085]

[0086]

[0087] Among them, WAM represents the attention of the current window, S-WAM represents the attention after the window is shifted, and LayerNorm represents the normalization of each sample. represents the output feature of WAM, z l Represents the output features of the multi-layer perceptron MLP.

[0088] (6) Attention Mechanism Dual-Channel Parallel Fusion Network

[0089] If the health status of the variable pitch bearing is to be accurately identified under non-stationary and strong random noise, sufficient signal feature information must be obtained. The present invention comprehensively considers both global and local features and proposes an attention mechanism dual-channel parallel fusion network (ADCPFN) model. The main idea is to adaptively fuse the global spatiotemporal information in the time-frequency diagram extracted by GFEM 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.

[0090] 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 Represent global and local features respectively. The dual-channel input structure of the ADCPFN model can more effectively combine visual information at different levels and suppress the interference of strong random noise on fault identification, thereby improving the model performance and generalization ability.

[0091] F=P (concat (F G , F L )) (20)

[0092] ADCPFN uses the AdamW gradient descent algorithm for supervised training, and the core calculations are shown in formulas (21) to (23).

[0093]

[0094] Among them, 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).

[0095] (7) Experimental verification

[0096] Taking the bearing data of Case Western Reserve University as an example, in order to simulate the actual wind turbine variable pitch bearing operating environment, random Gaussian noise and random scale noise are added to the data set with a probability of 50%, where λ is the intensity of the random noise. According to the present invention, first, CWT and PMT are used to encode the original one-dimensional vibration signal into a time-frequency diagram and a position matrix diagram, respectively. The encoding results are shown in the following example. Figure 5 It can be observed that compared with the one-dimensional vibration signal, the feature difference of the encoded two-dimensional image is more obvious, indicating that increasing the visualization dimension of the data will amplify the difference in vibration signals of bearings in different states and significantly enhance the characteristic information of the fault.

[0097] Table 1 shows the accuracy of the bearing fault diagnosis results of the present invention. The three-dimensional confusion matrix of the diagnosis results of each health state under different intensity random noise is shown in Table 1. Figure 6 shown.

[0098] Table 1 Accuracy of bearing fault diagnosis results of the method of the present invention

[0099]

[0100] It can be observed intuitively that when the random noise intensity is 0.2, the diagnostic accuracy of each health state is higher than 94%, and the diagnostic results of 7 states reach 100%. When the noise intensity increases to 0.4, the diagnostic results of 8 states reach 100%. However, it is worth noting that the accuracy of state 2 is 97% when the noise intensity is 0.2, while the diagnostic accuracy of state 2 reaches 100% when the noise intensity is 0.4. There are two main reasons for this. On the one hand, due to the randomness of noise addition and the independence of noise data sets of different intensities, the diagnostic results have a certain degree of contingency; on the other hand, the noise of 0.4 intensity does not interfere with the signal fluctuation characteristics of state 2, or complements the signal fluctuation of state 2, suppressing the interference of noise. 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, and the diagnostic accuracy decreases, but there are still 6 health states with a diagnostic accuracy of more than 90%. Judging from the results of the diagnosis of various health states under noises of different intensities, the overall diagnostic results of the present invention are not much different when facing medium and low intensity random noise interference, and are less affected by noise, with good noise resistance. When facing strong random noise interference, although the diagnostic accuracy of some health states has decreased, overall, it can still effectively avoid the influence of noise to a certain extent, and has good robustness.

Claims

1. A wind turbine variable pitch bearing fault diagnosis method based on signal coding, characterized in that: The following steps are involved: Step 1: Data acquisition: obtaining the original vibration signal data of the variable pitch bearing from the wind turbine condition monitoring system equipped in the wind farm; Step 2: Signal encoding: Continuous wavelet transform and position matrix transform are used to encode the one-dimensional original vibration signal into a two-dimensional time-frequency diagram and a position matrix diagram, respectively, to enhance the characteristic information of the signal by increasing the visualization dimension of the data; Step 3: Construction of attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model. Based on the global and window attention mechanisms, the global feature extraction module and the local feature extraction module are designed respectively, and integrated through the dual-channel parallel network structure and the feature adaptive fusion module to form the attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model; Step 4: Training of the attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model. The ADCPFN model is trained based on the two-dimensional time-frequency graph and the position matrix graph. Finally, the trained ADCPFN model completes the diagnosis of the variable pitch bearing fault.

2. A wind turbine variable pitch bearing fault diagnosis method based on signal coding according to claim 1, characterized in that: The global feature extraction module GFEM in step 3 uses the time-frequency diagram generated by signal coding as input data and is designed based on the global attention mechanism GAM, which includes a convolution layer and an attention mechanism layer; The convolution layer consists of Conv2D and BN&PReLU, where Conv2D is used to initially capture spatial features and reduce the complexity of the model; BN&PReLU controls the transfer of model training gradients to avoid gradient explosion and disappearance and speed up the convergence of model training; The attention mechanism layer is composed of attention mechanism blocks. By focusing on and recalibrating the 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 transformed into F1, and then converted to its original dimension after passing through the multi-layer perceptron MLP to obtain F2; secondly, two convolution operations are performed using a 7×7 convolution kernel to obtain F3; then, the feature map is subjected to maximum and average pooling respectively, and the results are fused to obtain F4, compressing the number of channels and reducing the amount of calculation; finally, the feature map is processed by Sigmoid to output F o ; Assume that 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) Among them, F i ′ represents F i The feature map after dimension conversion, represents element-wise multiplication and σ is the convolution operation.

3. A wind turbine variable pitch bearing fault diagnosis method based on signal coding according to claim 1, characterized in that: The local feature extraction module LFEM described in step 3 takes the position matrix map generated by the conversion as input data and is designed based on the window attention mechanism WAM; the module is divided into four stages: first, the position matrix map is input into the patch segmentation layer in Stage 1 for segmentation, and is segmented 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; then, the extracted feature map is transmitted to the patch merging layer and Wam Transformer Block in Stage 2 and Stage 3 in turn for the same operation, and the sizes of the output feature maps are H / 8×W / 8 and H / 16×W / 16 respectively; finally, the extracted feature map is passed through the Sparse tokenconverter and Transformer blocks in Stage 4 to obtain the final local feature map of size H / 32×W / 32.

4. A wind turbine pitch bearing fault diagnosis method based on signal coding according to claim 3, characterized in that: The calculation formula of the Wam Transformer Block is as follows: Among them, WAM represents the attention of the current window, S-WAM represents the attention after the window is shifted, and LayerNorm represents the normalization of each sample. represents the output feature of WAM, z l Represents the output features of the multi-layer perceptron MLP.

5. The wind turbine variable pitch bearing fault diagnosis method based on signal coding according to claim 1, characterized in that: The attention mechanism dual-channel parallel fusion network ADCPFN fault diagnosis model described in step 3 performs multi-feature adaptive fusion on the global spatiotemporal information in the time-frequency diagram extracted by the global feature extraction module GFEM and the local position change information in the position matrix diagram extracted by the local feature extraction module LFEM, and is composed of a GFEM module, a LFEM module, an adaptive fusion module and an output module; the ADCPFN 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 feature information and local feature information, as shown in formula (10), where P represents adaptive average pooling, F G ,F L 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). Among them, 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