Wind turbine generator misalignment fault diagnosis method and system based on deep learning technology

Abstract

The invention relates to the technical field of wind turbine generator fault diagnosis, in particular to a wind turbine generator misalignment fault diagnosis method and system based on a deep learning technology. The method comprises the following steps: acquiring original vibration signals of the wind turbine generator set under different working conditions of normal misalignment, parallel misalignment, angle misalignment and comprehensive misalignment; performing standardization processing and parameter-adjustable wavelet threshold de-noising processing on the original vibration signals under different working conditions; extracting multi-scale features of the processed signal based on a lightweight MS-DSCNN residual network model; inputting the multi-scale features into an LSTM network model for processing, and outputting time sequence features; weighted average is carried out on the time sequence features through an attention mechanism, and enhanced features are generated; and sequentially inputting the enhanced features into a full connection layer and a Softmax classification layer, and outputting a fault diagnosis result. According to the invention, the accuracy and timeliness of misalignment fault diagnosis of the wind turbine generator can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine fault diagnosis, and specifically to a wind turbine misalignment fault diagnosis method and system based on deep learning technology, which is suitable for equipment status monitoring and health management in the wind power generation industry. Background Art

[0002] Wind turbines play a vital role in the renewable energy sector, and their stable operation is directly related to the reliability of energy supply and the economical operation costs. However, due to the complex and ever-changing operating environment and the sophisticated internal mechanical structure of wind turbines, wind turbines are susceptible to a variety of internal and external factors during their long-term operation. Misalignment is a common and potentially harmful type of failure. Misalignment can cause additional stress and wear on the mechanical components of wind turbines, thereby affecting the overall performance and lifespan of the equipment.

[0003] In the field of wind turbine operating status monitoring and fault diagnosis, the advancement of signal processing and intelligent diagnostic technology is crucial to ensuring the safe operation of equipment. Existing technologies for monitoring and diagnosing misalignment faults have the following limitations: (1) Although the traditional wavelet transform method can effectively remove noise, it is highly subjective, has high computational complexity, and is not effective when processing non-stationary signals, which makes it difficult to select the noise threshold.

[0004] (2) Deep learning-based methods, such as convolutional neural networks (CNNs), have advantages in feature extraction, but they require a large amount of data support, are time-consuming to train, and are prone to overfitting. Furthermore, the extracted features lack interpretability, and are not ideal for processing time series features. Furthermore, while the end-to-end model combining CNNs and long short-term memory networks (LSTMs) improves diagnostic capabilities, it requires high computational resources, is difficult to converge, lacks a certain degree of adaptability, and is not very real-time. Its dependence on data limits its generalization capabilities.

[0005] In summary, while existing technologies have made some progress in wind turbine misalignment fault diagnosis, they still have many shortcomings in processing complex signals, extracting key features, and improving fault diagnosis accuracy. Current technical approaches exhibit significant limitations when processing non-stationary and noisy vibration signals, making it impossible to accurately extract fault features, which in turn affects the accuracy and timeliness of fault diagnosis. Furthermore, existing fault diagnosis techniques, which mostly rely on fixed thresholds or expert experience, are difficult to adapt to the complex and changing operating environments of wind turbines. When processing long time series data, it is also difficult to effectively resolve long-term dependencies, failing to meet the actual operational requirements of wind turbines. Summary of the Invention

[0006] In order to solve the problems in the prior art, the present invention provides a wind turbine misalignment fault diagnosis method and system based on deep learning technology, which can improve the accuracy and timeliness of wind turbine misalignment fault diagnosis.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for diagnosing misalignment faults of wind turbines based on deep learning technology, comprising the following steps: Obtain the original vibration signals of the wind turbine under different working conditions of normal misalignment, parallel misalignment, angular misalignment and comprehensive misalignment; The original vibration signals under different working conditions are subjected to standardization and wavelet threshold denoising with adjustable parameters. The multi-scale features of the processed signal are extracted based on the lightweight MS-DSCNN residual network model; Input the multi-scale features into the LSTM network model for processing and output the time series features; The time series features are weighted averaged through the attention mechanism to generate enhanced features; The enhanced features are sequentially input into the fully connected layer and the Softmax classification layer to output the fault diagnosis results.

[0008] Preferably, the specific method of performing wavelet threshold denoising processing on the original vibration signals under different working conditions is: after the original vibration signals under different working conditions are decomposed by wavelet, the wavelet coefficients corresponding to the original vibration signals are obtained, and the original vibration signals whose wavelet coefficients are lower than the preset threshold are removed.

[0009] Preferably, the specific method of the wavelet decomposition is: select a wavelet and determine a wavelet decomposition level N, then perform N-layer wavelet decomposition calculation on the vibration signal; for each high-frequency coefficient from the 1st layer to the Nth layer, select a threshold value. Perform threshold quantization processing; the threshold The calculation formula is:

[0010] Where, is the denoising parameter; is the median of the detail coefficients of the first layer; is the length of the vibration signal; Based on the low-frequency coefficients of the Nth layer of wavelet decomposition and the high-frequency coefficients of the 1st to Nth layers after quantization, the signal is reconstructed by wavelet. The formula is:

[0011] Where, is the inverse discrete wavelet transform.

[0012] Preferably, the lightweight MS-DSCNN residual network model includes multiple groups of depth-separable convolutional layers set in parallel, each group uses convolution kernels of different sizes, and is sequentially connected to a batch normalization layer, an H-Swish activation function and a residual jump connection structure.

[0013] Preferably, L1 regularization is introduced into the lightweight MS-DSCNN residual network model to constrain the weights of the depth-separable convolutional layer.

[0014] Preferably, the specific method of performing weighted averaging on the temporal features through the attention mechanism to generate enhanced features is: Calculate the attention score for the temporal features using the following formula:

[0015] Where, Score for attention; is the attention mechanism weight matrix; is the bias vector of the attention mechanism, is the output of each LSTM channel; Normalize the attention score to generate the attention weight, and the calculation formula is:

[0016] Where, is the attention weight; The calculation formula of the enhanced feature is:

[0017] Where, The enhanced features.

[0018] Preferably, the fully connected layer applies L2 regularization.

[0019] In a second aspect, the present invention provides a wind turbine misalignment fault diagnosis system based on deep learning technology, comprising: Signal acquisition module: acquires the original vibration signals of wind turbines under different working conditions such as normal misalignment, parallel misalignment, angular misalignment and comprehensive misalignment; Pre-processing module: standardizes the original vibration signals under different working conditions and performs wavelet threshold denoising with adjustable parameters; MS-DSCNN network module: extracts multi-scale features of processed signals based on a lightweight MS-DSCNN residual network model; LSTM network model module: inputs multi-scale features into the LSTM network model for processing and outputs time series features; Weighted calculation module: performs weighted averaging of temporal features through the attention mechanism to generate enhanced features; Diagnosis result module: The enhanced features are input into the fully connected layer and the Softmax classification layer in sequence, and the fault diagnosis results are output.

[0020] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and is characterized in that the computer program implements the steps of the above method when executed by a processor.

[0022] Compared with the prior art, the present invention has the following beneficial effects: This paper proposes a novel wind turbine misalignment fault diagnosis method based on a combination of a multi-scale separable convolutional neural network (MS-DSCNN) and an attention-based long short-term memory network (AttLSTM). Through wavelet threshold denoising with adjustable parameters, this method adaptively eliminates noise interference in vibration signals while fully preserving key fault signature information. This effectively addresses the difficulty faced by traditional denoising methods in balancing denoising effectiveness and feature preservation. Secondly, the lightweight MS-DSCNN residual network model structure combines the advantages of multi-scale feature extraction and deep separable convolution. This significantly reduces model complexity while maintaining feature extraction capabilities, avoids gradient explosion or vanishing, and achieves a balance between resources and efficiency. Furthermore, the collaborative work of the LSTM network model and the attention mechanism not only captures long-term temporal dependencies in vibration signals but also automatically identifies and enhances the most diagnostically valuable feature dimensions, significantly improving the model's feature representation capabilities.

[0023] Furthermore, the present invention adopts the L1-L2 joint regularization method and adds a penalty mechanism to the lightweight MS-DSCNN residual network module and the fully connected layer, thereby improving the generalization ability of the model and ensuring that misalignment faults can be identified quickly and accurately during the operation of the wind turbine, thereby reducing maintenance costs and improving the operating efficiency and safety of the wind turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is a flow chart of a method for diagnosing misalignment faults of wind turbines based on deep learning technology according to the present invention; Figure 2 This is a structural diagram of the experimental platform in an embodiment of the present invention; Figure 3 (a) Schematic diagram of hard threshold quantization processing method; Figure 3 (b) Schematic diagram of soft threshold quantization processing method; Figure 4 Comparison chart of different threshold selection methods; Figure 5 (a) Schematic diagram of the angular misalignment signal denoising results; Figure 5 (b) Schematic diagram of parallel misalignment signal denoising results; Figure 5 (c) Schematic diagram of the denoising results of the comprehensive misalignment signal; Figure 5 (d) Schematic diagram of the denoising results of normal misalignment signals; Figure 6 Schematic diagram of the MSCNN network model; Figure 7 Schematic diagram of depth-wise separable convolution; Figure 8 Comparison chart of activation functions; Figure 9 Schematic diagram of the lightweight MS-DSCNN residual network model of the present invention; Figure 10 This is the LSTM network model structure diagram; Figure 11 Schematic diagram of the channel attention mechanism; Figure 12 Schematic diagram of training set loss and accuracy; Figure 13 This is a diagram of the confusion matrix of the test set; Figure 14 This is the TSNE distribution diagram of the test set results. DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0028] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0029] In the description of the embodiments of the present invention, it should be noted that if the terms "upper," "lower," "horizontal," "inner," etc. appear, the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the inventive product is typically placed when in use. These terms are merely for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0030] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0031] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0032] The present invention is described in further detail below with reference to the accompanying drawings: The first object of the present invention is to provide a wind turbine misalignment fault diagnosis method based on deep learning technology, such as Figure 1 As shown, the following steps are included: Obtain the original vibration signals of the wind turbine under different working conditions of normal misalignment, parallel misalignment, angular misalignment and comprehensive misalignment; The original vibration signals under different working conditions are subjected to standardization and wavelet threshold denoising with adjustable parameters. The multi-scale features of the processed signal are extracted based on the lightweight MS-DSCNN residual network model; Input the multi-scale features into the LSTM network model for processing and output the time series features; The time series features are weighted averaged through the attention mechanism to generate enhanced features; The enhanced features are sequentially input into the fully connected layer and the Softmax classification layer to output the fault diagnosis results.

[0033] The present invention uses wavelet threshold denoising processing with adjustable parameters to adaptively eliminate noise interference in vibration signals while fully retaining the key information of fault characteristics, effectively solving the problem that traditional denoising methods are difficult to balance denoising effect and feature retention. Secondly, the lightweight MS-DSCNN residual network model structure combines the advantages of multi-scale feature extraction and deep separable convolution, greatly reducing the complexity of the model while ensuring feature extraction capabilities, avoiding gradient explosion or disappearance, and achieving a balance between resources and efficiency. Subsequently, the collaborative work of the LSTM network model and the attention mechanism can not only capture the long-term temporal dependencies in the vibration signal, but also automatically identify and enhance the feature dimensions with the most diagnostic value, significantly improving the feature expression ability of the model. The present invention has good engineering applicability and promotion value, and can meet the stringent requirements of wind turbine fault diagnosis for real-time, accuracy and stability.

[0034] The Long Short Term Mermory network (LSTM) model is a special recursive neural network that can effectively solve the problem of long-term dependency. LSTM has a four-layer structure, such as Figure 10 As shown in Figure 2, the core of the LSTM network model lies in its unique memory module design, which uses a sophisticated gating mechanism to model the long-range dependencies of temporal information. Specifically, each memory unit contains three key gating structures: a forget gate, an input gate, and an output gate, as well as a memory cell that spans the temporal sequence. Figure 10 The horizontal line above the box is called the cell state, which is like a conveyor belt that controls the transmission of information to the next moment. Figure 6 The two tanh layers in the formula correspond to the input and output of the cell respectively. The LSTM can add and delete information to the cell through the gating unit. The passage of information can be selectively decided through the gate, which is composed of a sigmoid neural network layer and a pair of multiplication operations. The output of the layer is a number between 0 and 1. Considering time and accuracy, the number of cells in the application is 128.

[0035] The formula for standardizing the original vibration signals under different working conditions is:

[0036] In the formula, is the standardized vibration signal; is the original vibration signal; is the mean of the original vibration signal; is the standard deviation of the original vibration signal.

[0037] The standardization method of the application eliminates the influence of signal dimension, so that the vibration signals under different working conditions have a unified numerical distribution range, effectively solving the problem of inconsistent amplitude of original signals caused by differences in acquisition conditions and sensors.

[0038] The specific method for wavelet threshold denoising of the original vibration signals under different working conditions is: after wavelet decomposition of the original vibration signals under different working conditions, the wavelet coefficients corresponding to the original vibration signals are obtained, and the original vibration signals whose wavelet coefficients are lower than the preset threshold are removed. The core principle of the wavelet threshold denoising method is based on the different performance characteristics of signals and noises in the wavelet domain. Specifically, when the signal is subjected to wavelet transform (using Mallat algorithm), the effective signal components will be concentrated in a small number of significant wavelet coefficients, while the noise components will be dispersed as a large number of small amplitude coefficients. This difference in characteristics makes it possible to separate signal and noise by setting appropriate threshold standards: retaining wavelet coefficients with amplitudes exceeding the threshold, which carry the main signal information, while setting the coefficients below the threshold, which are mainly contributed by noise, to zero. Both effectively suppress the noise interference and better preserve the key features of the original signal, thus achieving the desired denoising effect.

[0039] The specific method for wavelet decomposition is: selecting a wavelet and determining a wavelet decomposition level N, then performing N-layer wavelet decomposition calculation on the vibration signal; for each layer of high-frequency coefficients from the first layer to the Nth layer, a threshold is selected for threshold quantization processing; the threshold is calculated according to the formula:

[0040] In the formula, is the denoising parameter; is the median of the detail coefficients of the first layer; is the length of the vibration signal.

[0041] There are many adaptive methods for selecting the threshold of each layer, mainly hard threshold quantization and soft threshold quantization, such as Figure 3 shown. Figure 3 The hard threshold quantization process in (a) can well preserve local features such as signal edges. Figure 3 The soft threshold quantization process in (b) is relatively smooth, but it will cause distortion such as edge blur. In order to ensure the denoising effect and the smoothness of the threshold function, the present invention (improved method) selects an adjustable denoising function between the soft and hard threshold functions. The function image is as follows Figure 4 shown.

[0042] Here’s how it works: 1) Calculate the median of the detail coefficient using the following formula:

[0043] Where, To calculate the median; is the absolute value; —Detail coefficient of the first layer.

[0044] 2) Estimate the standard deviation of the noise based on the median. The formula is as follows:

[0045] Where, —The calculated absolute value of the median.

[0046] 3) Calculate the threshold using the following formula:

[0047] Where, is the initial threshold; is the length of the vibration signal; is the standard deviation of the noise.

[0048] The present invention adopts parameters To adjust the denoising strength, the threshold calculation formula is as follows:

[0049] The introduction of the detail coefficient median confirmation threshold method ensures the edge characteristics of the signal and prevents signal distortion. is a denoising parameter that is used after the initial threshold is calculated. The final threshold is obtained by multiplication, thereby achieving artificial control of the denoising intensity.

[0050] Based on the low-frequency coefficients of the Nth layer of wavelet decomposition and the high-frequency coefficients of the 1st to Nth layers after quantization, the signal is reconstructed by wavelet. The formula is as follows:

[0051] Where, —Inverse discrete wavelet transform.

[0052] In practical applications, the signal-to-noise ratio (SNR) and the root mean square error (RMSE) between the estimated signal and the original signal can be used to judge the denoising effect.

[0053] The lightweight MS-DSCNN residual network model includes multiple groups of depth-separable convolutional layers set in parallel, each group uses convolution kernels of different sizes, and is sequentially connected to batch normalization layers, H-Swish activation functions and residual jump connection structures. The present invention realizes the collaborative extraction of multi-scale features by setting up depth-separable convolution layers with convolution kernels of different sizes in parallel, and can comprehensively capture fault features of different granularities in vibration signals; combined with batch normalization layers, it effectively stabilizes the network training process, and the H-Swish activation function improves computational efficiency while ensuring nonlinear expression capabilities; the introduced residual jump connection structure not only alleviates the gradient vanishing problem of deep networks, but also enhances the feature reuse capability. This network architecture significantly improves the comprehensiveness and effectiveness of feature extraction while maintaining lightweight characteristics, providing high-quality feature representation for subsequent time series modeling, making the model both high-performance and engineering practical.

[0054] Specifically, firstly, MSCNN as a multi-scale convolutional neural network can more effectively extract features from high-dimensional complex inputs than traditional machine learning algorithms, such as Figure 6 As shown in Figure 2, MSCNN is an improved convolutional neural network that mines feature information from multiple scales through convolution kernels of different sizes. It effectively solves the adaptive selection problem of convolution kernels in traditional CNN models and improves the convergence speed of the network to a certain extent.

[0055] Multi-scale CNN extracts local and near-global features from time series data. Each layer is followed by batch normalization and ReLU activation function to enhance the nonlinear expression ability of the model and stabilize the training process. The formula is as follows:

[0056] in, represents a set of features extracted from multiple scales, is the convolution kernel, is the input data, is the bias term, is the set of convolution kernels, represents the convolution operation, Indicates feature fusion.

[0057] However, the problem is that it requires a lot of device memory and requires a lot of computation, which makes it impossible to run on mobile devices and embedded devices.

[0058] In traditional convolution, the number of convolution kernel channels is equal to the number of input feature matrix channels, and the number of output feature matrix channels is equal to the number of convolution kernels. Separable convolution (Depthwise Separable Conv) consists of depthwise convolution (DW) convolution (DepthwiseConv) and pointwise convolution (PW). The PW network is a common convolution with a convolution kernel size of 1. The input feature matrix channel is equal to the number of convolution kernels and the output feature matrix channel is equal to the number of convolution kernels. That is, the depth of the feature convolution does not change. The convolution process of depthwise separable convolution is as follows: Figure 7 shown.

[0059] The computational cost of traditional convolution C:

[0060] The computational cost of depth-wise separable convolution is S:

[0061] The ratio of the two:

[0062] in, is the size of the convolution kernel, is the size of the input feature matrix, is the depth of the input feature matrix, is the depth of the output feature matrix.

[0063] Therefore, the computational complexity of traditional convolution is times that of depthwise separable convolution. In the present invention, when other network structures are consistent, the number of parameters of the ordinary convolution layer is 20544, the number of parameters of the MSCNN layer is 24312, and the number of parameters of the MS-DSCNN layer is 4052. It can be seen that the present invention uses the depthwise separable convolution in MobileNet instead of the original ordinary convolution and multi-scale convolution to greatly reduce the amount of computation.

[0064] The present invention adopts the activation function of MobileNetV3, which is expressed as:

[0065] The core idea is to add a constant offset to the input and then make corrections. The ReLU6 function truncates the input value to the interval [0,6] to ensure that the output range is bounded, thereby improving numerical stability. Figure 8 As shown in Figure 2, the advantage of H-Swish is that it retains some of the nonlinear characteristics of ReLU while achieving a smoother transition, thereby reducing the possibility of gradient disappearance. This design is particularly important in lightweight models because it optimizes the model's forward propagation time and memory usage.

[0066] For example, the present invention adds an attention layer on the basis of LSTM, identifies key information, calculates the importance weight of each channel output, automatically identifies the most critical information, prioritizes those channel data with high abnormal probability and reweights them, and finally sums and reduces the dimension of the data points of each channel to input the final fully connected layer. The principle of channel attention is as follows Figure 11 shown.

[0067] The specific method of performing weighted averaging of temporal features through the attention mechanism to generate enhanced features is as follows: Calculate the attention score for the temporal features using the following formula:

[0068] Where, Score for attention; is the attention mechanism weight matrix; is the bias vector of the attention mechanism, is the output of each LSTM channel; Normalize the attention score to generate the attention weight, and the calculation formula is:

[0069] Where, is the attention weight; The calculation formula of the enhanced feature is:

[0070] Where, The enhanced features.

[0071] In order to avoid the algorithm from falling into the local optimal space and overfitting, the present invention optimizes the model from multiple aspects. L1 regularization is introduced in the lightweight MS-DSCNN residual network model to constrain the weights of the depth-separable convolution layer, and L2 regularization is applied to the fully connected layer to penalize the parameters. Due to the non-smooth penalty mechanism of L1 regularization, the network will return unimportant features to 0 to a certain extent at the beginning of training. By adding L1 regularization to the MS-DSCNN multi-scale convolution layer, the training process is simplified and features are effectively selected. At the same time, L2 regularization is added to the fully connected layer. Its smooth penalty mechanism encourages small but non-zero coefficients, prevents any one feature from dominating the prediction of the model and promotes overall weight shrinkage. This can effectively retain the optimal parameters, thereby improving the accuracy and model generalization ability. This method has achieved good results in all compared networks. In addition, when the number of iterations reaches a certain stable initial stage, the present invention reduces the learning rate and enters the current network parameter structure to ensure rapid convergence of accuracy. If the test set data is accurate and no longer increases, the network iteration process is stopped immediately to retain the optimal parameters. Finally, Dropout is used to randomly ignore a certain ratio of parameters, so that it can go deeper into the current optimal space during gradient descent and continue to find the optimal solution.

[0072] In summary, the lightweight MS-DSCNN residual feature network described in this paper combines the multi-scale features of MSCNN with the H-Swish activation function and depthwise separable convolution of MobileNet to reduce parameters and computational complexity, improve detection accuracy, prevent gradient explosion or vanishing, and optimize resource utilization. Data fault diagnosis is performed using the improved LSTM network model, namely the MS-DSCNN-AttLSTM network model structure. This structural design diversifies data features and enhances the adaptability of the network model. The LSTM network model can effectively handle the changing trends of data in time series. The introduction of the attention mechanism can avoid the influence of irrelevant features, further improving the network's targeted capabilities. Experiments have demonstrated the effectiveness of the model.

[0073] Example 1. Vibration data (i.e. vibration signal) collection Since the actual fault signal of wind farm accounts for a small proportion, in order to study various misalignment faults, this embodiment builds an experimental platform in the laboratory to simulate the operation of actual wind turbines and create various misalignment phenomena. The overall structure of the experimental platform is as follows: Figure 2 shown.

[0074] The piezoelectric accelerometer CA-YD-186 was used to extract vibration data under four working conditions: normal, parallel misalignment, angular misalignment, and combined misalignment. The data was further expanded using a sliding window. Finally, the total data set had a total of 1624 sample points. The ratio of the training set to the test set was 8:2, so the number of sample points in the training set was 1299, and the number of sample points in the test set was 325.

[0075] 2. Data preprocessing The vibration data were standardized using Z-score; Select a wavelet and determine a wavelet decomposition level N, then perform N-layer wavelet decomposition calculations on the vibration data; for each high-frequency coefficient from the 1st layer to the Nth layer, select a threshold Perform threshold quantization processing; the threshold The calculation formula is:

[0076] Where, is the denoising parameter; is the median of the detail coefficients of the first layer; is the length of the vibration signal.

[0077] When the denoising parameters When the original data of the experimental platform is denoised by taking 0.2, the signal-to-noise ratio is improved by 7dB. Figure 5 (a) is a schematic diagram of the denoising results of the angle misalignment signal. Figure 5 (b) is a schematic diagram of the parallel misalignment signal denoising results. Figure 5 (c) is a schematic diagram of the denoising results of the comprehensive misalignment signal. Figure 5 (d) is a schematic diagram of the denoising results of the normal misalignment signal.

[0078] 3. Lightweight MS-DSCNN residual network model feature extraction The present invention uses convolution kernels of 3x1 and 5x1 sizes to construct multi-scale separable convolution layers, such as Figure 9 As shown, the network first uses a multi-scale convolutional neural network (MSCNN) to extract features from multiple dimensions. A specific convolution kernel is inserted into each channel, implementing a depthwise separable convolution (DW) operation. At the intermediate output stage, the network uses point convolution techniques from a depthwise separable convolutional neural network (PW) to process features. The number of output convolution kernels is determined by the required number of output channels. This invention effectively combines the advantages of multi-scale feature extraction and depthwise separable convolution, aiming to improve the richness of feature representation and the computational efficiency of the network.

[0079] 4. LSTM network model learning features The LSTM network model is used to continue processing the multi-scale features, capture long-term dynamic changes in the time sequence, and output time sequence features by adding regularization to improve the generalization ability of the model and prevent network overfitting.

[0080] 5、ATTENTION network feature enhancement An attention layer is added to the LSTM to identify key information, calculate the importance weight of each channel output, automatically identify the most critical information, and preferentially process channel data with high abnormal probability and reweight.

[0081] Specifically, the specific method of the above reweighting is: The attention score is calculated for the time sequence feature, and the calculation formula is:

[0082] In the formula, is the attention score; is the attention mechanism weight matrix; is the bias vector of the attention mechanism, is the output of each LSTM channel; The attention score is normalized to generate the attention weight, and the calculation formula is:

[0083] In the formula, is the attention weight; The calculation formula of the enhanced feature is:

[0084] In the formula, is the enhanced feature.

[0085] 6、Data output The data points of each channel are summed and reduced in dimension, sequentially input into the fully connected layer and the Softmax classification layer, and the fault diagnosis result is output.

[0086] 7、Model optimization The present application adds L1 regularization to the light MS-DSCNN residual network model to effectively select the features, adds L2 regularization to the fully connected layer to punish the parameters. At the same time, when the number of iterations reaches a certain stable initial stage, the learning rate is reduced to enter the current network parameter structure to ensure the accuracy of rapid convergence. If the test set data accuracy no longer rises, the network iteration process is immediately stopped, and the optimal parameters are retained. Finally, the Dropout randomly ignores a certain percentage of parameters, so that it can go deep into the current optimal space when gradient descent, and continue to find the optimal solution.

[0087] 8、Model testing The MS-DSCNN-AttLSTM fault diagnosis algorithm proposed in this invention is used, and the performance of the model is evaluated using the accuracy indicator. The specific formula is as follows:

[0088] Among them, TP stands for True Positive, TN stands for True Negative, FP stands for False Positive, and FN stands for False Negative. Figure 12 As shown, it can be seen that the training set loss decreases with the increase of the number of iterations, and the rate of change slows down as the learning rate decreases, and eventually tends to be stable. The same is true for the accuracy.

[0089] The network of the present invention was used on four fault data sets of a wind turbine test bench running at 100 rpm. The training set achieved an accuracy of 98.15%, and the test set achieved an accuracy of 99.07%. This shows that the network has a highly reliable detection capability. The confusion matrix is ​​mainly used to evaluate the performance of the classification model. By showing the correspondence between the model prediction results and the actual categories, it provides information about the prediction accuracy of the model in each category and the common error types. The t-SNE distribution map is used to explore the structure and relationship of data in high-dimensional space. By mapping high-dimensional data to low-dimensional space, it shows the clustering structure of data points, outliers, and intra-class and inter-class relationships. Through Figure 13 and Figure 14 It can be seen that the present invention can diagnose the misalignment fault of the wind turbine generator set with high accuracy.

[0090] A second object of the present invention is to provide a wind turbine misalignment fault diagnosis system based on deep learning technology, comprising: Signal acquisition module: used to obtain the original vibration signals of wind turbines under different working conditions such as normal misalignment, parallel misalignment, angular misalignment and comprehensive misalignment; Preprocessing module: used to perform standardization processing and wavelet threshold denoising with adjustable parameters on the original vibration signals under different working conditions; MS-DSCNN network module: used to extract multi-scale features of processed signals based on the lightweight MS-DSCNN residual network model; LSTM network model module: used to input multi-scale features into the LSTM network model for processing and output time series features; Weighted calculation module: used to perform weighted averaging of temporal features through the attention mechanism to generate enhanced features; Diagnosis result module: used to input the enhanced features into the fully connected layer and the Softmax classification layer in sequence, and output the fault diagnosis results.

[0091] The signal acquisition module ensures the acquisition of raw vibration data that comprehensively covers all types of fault conditions; the preprocessing module effectively improves data quality through standardization and adaptive denoising; the innovative MS-DSCNN network module uses a lightweight structure to achieve efficient multi-scale feature extraction; the LSTM network module accurately captures the dynamic characteristics of time series signals; the weighted calculation module intelligently focuses on key features through an attention mechanism; and finally, the diagnosis module achieves accurate fault classification.

[0092] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory, wherein the memory is configured to store a computer program, the computer program including program instructions, and the processor is configured to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal and is suitable for implementing one or more instructions, specifically loading and executing one or more instructions in the computer storage medium to implement a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to operate a wind turbine misalignment fault diagnosis method based on deep learning technology.

[0093] The present invention also provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device, used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space, which stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the wind turbine misalignment fault diagnosis method based on deep learning technology in the above-mentioned embodiment.

[0094] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0096] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1The function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A wind turbine misalignment fault diagnosis method based on deep learning technology, characterized in that: The following steps are involved: Obtain the original vibration signals of the wind turbine under different working conditions of normal misalignment, parallel misalignment, angular misalignment and comprehensive misalignment; The original vibration signals under different working conditions are subjected to standardization and wavelet threshold denoising with adjustable parameters. The multi-scale features of the processed signal are extracted based on the lightweight MS-DSCNN residual network model; Input the multi-scale features into the LSTM network model for processing and output the time series features; The time series features are weighted averaged through the attention mechanism to generate enhanced features; The enhanced features are sequentially input into the fully connected layer and the Softmax classification layer to output the fault diagnosis results.

2. The method for diagnosing misalignment faults of wind turbines based on deep learning technology according to claim 1, characterized in that: The specific method of performing wavelet threshold denoising on the original vibration signals under different working conditions is as follows: after decomposing the original vibration signals under different working conditions by wavelet, the wavelet coefficients corresponding to the original vibration signals are obtained, and the original vibration signals whose wavelet coefficients are lower than a preset threshold are removed.

3. The method for diagnosing misalignment faults of wind turbines based on deep learning technology according to claim 2, characterized in that: The specific method of wavelet decomposition is as follows: select a wavelet and determine a wavelet decomposition level N, then perform N-layer wavelet decomposition calculation on the vibration signal; for each high-frequency coefficient from the 1st layer to the Nth layer, select a threshold Perform threshold quantization processing; the threshold The calculation formula is: Where, is the denoising parameter; is the median of the detail coefficients of the first layer; is the length of the vibration signal; Based on the low-frequency coefficients of the Nth layer of wavelet decomposition and the high-frequency coefficients of the 1st to Nth layers after quantization, the signal is reconstructed by wavelet. The formula is: Where, is the inverse discrete wavelet transform.

4. The method for diagnosing misalignment faults of wind turbines based on deep learning technology according to claim 1, characterized in that: The lightweight MS-DSCNN residual network model includes multiple groups of depth-wise separable convolutional layers set in parallel, each group uses convolution kernels of different sizes, and is sequentially connected to a batch normalization layer, an H-Swish activation function, and a residual skip connection structure.

5. The method for diagnosing misalignment faults of wind turbines based on deep learning technology according to claim 4, characterized in that: L1 regularization is introduced into the lightweight MS-DSCNN residual network model to constrain the weights of the depthwise separable convolutional layer.

6. The method for diagnosing misalignment faults of wind turbines based on deep learning technology according to claim 1, characterized in that: The specific method of performing weighted averaging of temporal features through the attention mechanism to generate enhanced features is as follows: Calculate the attention score for the temporal features using the following formula: Where, Score for attention; is the attention mechanism weight matrix; is the bias vector of the attention mechanism, is the output of each LSTM channel; Normalize the attention score to generate the attention weight, and the calculation formula is: Where, is the attention weight; The calculation formula of the enhanced features is: Where, The enhanced features.

7. The method for diagnosing misalignment faults of wind turbines based on deep learning technology according to claim 1, characterized in that: The fully connected layers apply L2 regularization.

8. A wind turbine misalignment fault diagnosis system based on deep learning technology, characterized in that: include: Signal acquisition module: acquires the original vibration signals of wind turbines under different working conditions such as normal misalignment, parallel misalignment, angular misalignment and comprehensive misalignment; Pre-processing module: standardizes the original vibration signals under different working conditions and performs wavelet threshold denoising with adjustable parameters; MS-DSCNN network module: used to extract multi-scale features of processed signals based on the lightweight MS-DSCNN residual network model; LSTM network model module: used to input multi-scale features into the LSTM network model for processing and output time series features; Weighted calculation module: used to perform weighted averaging of temporal features through the attention mechanism to generate enhanced features; Diagnosis result module: used to input the enhanced features into the fully connected layer and the Softmax classification layer in sequence, and output the fault diagnosis results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images