A fan main shaft component fault prediction method based on wide kernel convolution

Abstract

The application discloses a fan main shaft component fault prediction method based on a wide kernel convolution, which firstly uses a Pearson coefficient to filter original data, extracts strong correlation data of target data, and fuses the strong correlation data of the current moment with the current moment data of the target data and the data of the last moment point as the input of a WCNN-LSTM neural network model; the neural network model in the prediction method introduces the time attribute of a variable, and improves a traditional convolution kernel, so that the problems of preprocessing difficulty and insufficient model precision caused by complex data can be solved, and finally, a sliding window is used for residual error calculation, so that the false detection rate of faults is reduced, and more efficient prediction of main shaft faults is realized.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault detection technology, specifically a fault prediction method for wind turbine main shaft components based on wide kernel convolution. Background Technology

[0002] Wind energy, as a clean energy source, has received widespread attention since the 21st century. However, due to harsh environments and complex mechanical structures, wind turbine failures occur frequently, increasing maintenance costs. Therefore, research on wind turbine fault diagnosis is crucial for reducing losses and lowering costs.

[0003] Currently, the mainstream fault diagnosis methods for wind turbines both domestically and internationally still involve analyzing and studying vibration signals. The main methods include decomposing bearing vibration signals using adaptive variational mode decomposition, support vector machines, convolutional neural networks, and isolated forests to extract fault features from the data. These methods require high-precision acquisition equipment and mostly only diagnose existing faults, lacking the ability to provide early warnings.

[0004] Supervisory Control and Data Acquisition (SCADA) systems for wind turbines, as a mainstream monitoring technology, acquire massive amounts of operational data. Compared to traditional vibration signals, they offer advantages such as ease of acquisition, high visualization, and comprehensive data. Therefore, utilizing SCADA data for fault early warning in wind turbines has become an important direction in wind turbine research. Current mainstream methods include: preprocessing SCADA data and constructing models such as multiple linear regression for temperature data, then evaluating stability through testing methods to determine fault occurrence; or graphically representing SCADA data and using neural network models such as Radial Basis Function (RBF) for fault prediction; or selecting input parameters using algorithms such as Isolation Forest, using a gradient framework (Light Gradient Boosting Machine, LightGBM) for temperature prediction, and using residual thresholds for wind turbine fault detection. The above methods have studied the current SCADA data from different perspectives. However, in practical applications, due to the massive amount of data and the fact that most of the collected data exhibits drastic fluctuations, as well as the data gaps caused by the regular maintenance strategy of wind turbines, missing values ​​will appear. The data preprocessing results are not ideal, and the prediction accuracy is insufficient.

[0005] Research on the application of deep learning in fault prediction reveals that the CNN-LSTM (Convolutional Neural Networks-Long Short Term Memory) model used by Ling Xiang has excellent performance in fault prediction. By studying and analyzing the spatiotemporal properties of the data, it has higher accuracy than other methods. However, it also requires data preprocessing to improve accuracy. Therefore, how to choose a method that does not require data preprocessing while achieving higher accuracy is a problem that still needs to be solved. Summary of the Invention

[0006] This invention addresses the problem of low fault detection rates in wind turbine SCADA data due to excessive outliers and missing values. It proposes a fault prediction method for wind turbine main shaft components based on wide-kernel convolution. This method combines an improved convolutional neural network model with LSTM to construct a wind turbine fault early warning model based on wide-kernel convolution, specifically a model combining a wide-kernel convolutional neural network with a long short-term memory network (WCNN-LSTM). By introducing the temporal attributes of variables into the neural network model and improving the convolution kernel, this invention overcomes the difficulties in preprocessing caused by data complexity and the problem of insufficient model accuracy, achieving more efficient prediction of main shaft faults.

[0007] The technical solution of this invention to solve the aforementioned technical problem is as follows: A fault prediction method for wind turbine main shaft components based on wide kernel convolution is designed, characterized by the following specific implementation steps:

[0008] Step 1: Install temperature sensors at multiple nodes in the main shaft component area inside the fan, collect data points at certain time intervals to obtain historical temperature time series data for each node in the main shaft component area, and the duration of this historical temperature time series data shall be no less than one year.

[0009] Step 2: Data Filtering and Processing

[0010] Let σ be the historical temperature time series data of a certain node in the spindle component region. Then, the dataset τ of the historical temperatures of all nodes in the spindle component region is obtained from step one. σ =(σ1, σ2, ..., σ) h ), where h is the dimension of data collection, representing the number of nodes set during the collection process, and σ x The historical temperature time series data represents the oil temperature on the spindle gearbox side, with the oil temperature on the spindle gearbox side as the target data; correlation coefficient analysis was used to select data related to σ. xNodes with a correlation coefficient greater than 0.8 are considered as strongly correlated data, and q strongly correlated data are obtained with a time scale of T. Each strongly correlated data is divided into two continuous time series data segments with a time scale of T in a ratio of 7:3. Correspondingly, the time scale of the first part is T1, and the time scale of the second part is T2, i.e., T1:T2 = 7:3.

[0011] The training dataset Q is obtained by processing the q strongly correlated data points from the first T1 time scales and the corresponding historical temperature data of the spindle gearbox side oil from the first T1 time scales. σ Specifically: Q σ The first data point corresponds to one data point. σ x1 This is the historical temperature data of the spindle gearbox side oil at the first moment, due to σ at the first moment. x1 Since there is no previous time step, we choose σ here. x1 As data from its previous point in time, These are the data values ​​at the first time point for the q strongly correlated data points;

[0012] Q σ The second time point corresponds to one data point. Here σ x2 The data at the previous time point is σ x1 Therefore, Q is directly used as the value of the data. σ The data corresponding to time point t is

[0013] The q strongly correlated data points at the next T2 time scales and the corresponding historical temperature data of the spindle gearbox side oil at the next T2 time scales were processed in the same way to obtain the test dataset;

[0014] Step 3: Establish the WCNN-LSTM neural network model

[0015] The WCNN-LSTM neural network model includes a WCNN neural network module and an LSTM neural network module. The output of the WCNN neural network module, after being fused with an attention mechanism, is used as the input of the LSTM neural network module, and the output of the LSTM neural network module is the output of the WCNN-LSTM neural network model. The WCNN neural network module includes a wide-kernel convolutional layer, a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, and a first batch normalization (BN) layer. These six parts are sequentially connected, with the output of the previous part serving as the input of the next part. The output of the first BN layer is used as the output of the WCNN neural network module. The LSTM neural network module includes a first LSTM layer, a second LSTM layer, a Fasten layer, a first Dense layer, a second BN layer, a second Dense layer, and an Output layer. These seven parts are sequentially connected, with the output of the previous part serving as the input of the next part. The output of the WCNN neural network module, after being fused with an attention mechanism, is used as the input of the first LSTM layer, and the input of the Output layer is the output of the LSTM neural network module, which is the output of the WCNN-LSTM neural network model. The convolution operation process of the wide-kernel convolutional layer of the WCNN neural network module is as follows: using W... i k and Let x represent the weights and biases of the i-th convolutional operation in the k-th layer, and use x k (j) represents the j-th local region of the k-th layer, and the process is described as follows:

[0016]

[0017] Where * represents the inner product of the kernel and the layout region. The result after convolution is represented by a kernel size of 100*100. Convolutions are performed sequentially along the time direction with a stride of 3. The convolutional result is then batch-normalized. β is used to represent the convolutional result as input data. The batch normalization process is as follows:

[0018]

[0019] Here, parameters γ and d are learned during training and then activated using rectified linear units. Let x represent the input data and y represent the output. The activation process is as follows:

[0020]

[0021] Reassemble the activated y groups in order. This is the output of a wide-kernel convolutional layer;

[0022] Step 4: Train the WCNN-LSTM neural network model using the training dataset.

[0023] 4.1 Training Parameter Settings

[0024] Set the fixed parameter values ​​of the neural network model, set the initial values ​​of the weight parameters to be learned to 1, set the initial values ​​of the bias parameters to 0, set the number of training iterations to 100, set the batch size to 40, and set the optimizer to AdaMod.

[0025] 4.2 Training the Neural Network Model

[0026] The training dataset Q from step two σ As input to the neural network model, the training dataset Q is sequentially... σ Each piece of data in the dataset is input into the neural network model for training, and Q... σt The reference value of the main shaft gearbox side oil temperature at time t is obtained from the output of the neural network model, based on the training dataset Q. σ Reference values ​​for the spindle gearbox side oil temperature at time scale T1 are obtained sequentially. The mean squared error (MSE) is used to calculate the loss value between the reference values ​​for the spindle gearbox side oil temperature obtained in this training round and the historical real data.

[0027]

[0028] Where n represents the population, and here it represents the time scale T1, Y i For historical real data, f(Y) i ) is a reference value output by the network model;

[0029] Based on the loss value, the AdaMod optimizer is used to perform backpropagation to update the weights and bias parameters once, completing one iteration of training. The updated neural network model parameters are then used as the initial parameters, and the training dataset Q is used again. σ The data is sequentially input into the neural network model for the next iteration of training; when the number of training iterations reaches a preset value, the weights and bias parameters of the last update are saved to obtain the trained neural network model.

[0030] 4.3 Testing the Neural Network Model

[0031] The data from the test dataset in step two are sequentially input into the trained neural network model to obtain reference values ​​for the spindle gearbox side oil temperature at the next T2 time scales. The degree of fit R between the reference values ​​and the historical real data is then calculated. 2 When the value is greater than 0.9, the trained neural network model is verified as an effective model;

[0032] Step 5: Spindle Failure Prediction

[0033] The historical data of the main shaft gearbox side oil temperature for the first T1 time scales are compared with the data from the training dataset Q. σ After processing by the trained neural network model, the reference values ​​of the main shaft gearbox side oil temperature obtained in the first T1 time scales are subtracted to obtain the residual sequence. The mean of the residual sequence is calculated using a sliding window of a certain time scale and set as the standard deviation μ.

[0034] The current actual temperature of the spindle gearbox side oil at the current m-th time point, the corresponding strongly correlated data, and the previous actual temperature of the spindle gearbox side oil are input into the neural network model trained in step three to obtain the reference value of the spindle gearbox side oil temperature at the m-th time point. The reference values ​​of the spindle gearbox side oil temperature at subsequent time points are calculated sequentially, and the mean of the residual sequence segments between the actual temperature sequence segment and the reference value sequence segment is calculated using the aforementioned sliding window at a certain time scale. When the mean ω of the residual sequence segment corresponding to a certain sliding window... M If the value exceeds the threshold by 3μ, it can be determined that the spindle may be malfunctioning.

[0035] Compared with existing technologies, the advantages of this invention are as follows: The prediction method of this invention first uses Pearson coefficients to filter the original data, extracts the strongly correlated data of the target data, and fuses the strongly correlated data at the current time with the current time data of the target data and the data at the previous time point as the input of the WCNN-LSTM neural network model; the neural network model in the prediction method of this invention introduces the temporal attribute of the variable and improves the traditional convolution kernel, which can solve the problems of preprocessing difficulties and insufficient model accuracy caused by data complexity; finally, a sliding window is used to calculate the residual, reducing the false detection rate of faults and achieving more efficient prediction of spindle faults. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the principle of the WCNN-LSTM neural network model in one embodiment of the wind turbine main shaft component fault prediction method based on wide kernel convolution of the present invention.

[0037] Figure 2 This is a comparison chart of the output data of a WCNN-LSTM neural network model, a LightGBM neural network model, a BiLSTM neural network model, and a CNN-LSTM neural network model without adding LMG, based on an embodiment of a fault prediction method for wind turbine main shaft components according to the present invention, without missing data.

[0038] Figure 3This is a comparison chart of the output data of a WCNN-LSTM neural network model, a LightGBM neural network model, a BiLSTM neural network model, and a CNN-LSTM neural network model with LMG added, in one embodiment of a fault prediction method for wind turbine main shaft components based on wide kernel convolution according to the present invention.

[0039] Figure 4 This is a comparison chart of the output data of a WCNN-LSTM neural network model, a LightGBM neural network model, a BiLSTM neural network model, and a CNN-LSTM neural network model without LMG added, in an embodiment of a fault prediction method for wind turbine main shaft components based on wide kernel convolution, according to an embodiment of the present invention.

[0040] Figure 5 This is a comparison chart of the output data of a WCNN-LSTM neural network model, a LightGBM neural network model, a BiLSTM neural network model, and a CNN-LSTM neural network model with LMG added under missing data, representing an embodiment of the fault prediction method for wind turbine main shaft components based on wide kernel convolution of the present invention.

[0041] Figure 6 This invention provides a curve plot of the reference value and the actual value of the oil temperature on the main shaft gearbox side, obtained using a fault prediction method for the main shaft component of a wind turbine based on wide kernel convolution, based on the present invention.

[0042] Specific implementation details

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0044] This invention provides a fault prediction method for wind turbine main shaft components based on wide-kernel convolution (hereinafter referred to as the prediction method). Its key feature is that it omits the preprocessing of the original data, uses wide-kernel convolution to reduce the impact of missing data and fluctuating abnormal data on the model prediction, and employs a sliding window to calculate the mean residual to reduce prediction errors. The specific implementation steps of this prediction method are as follows:

[0045] Step 1: Install temperature sensors at multiple nodes in the main shaft component area inside the fan, collect data points at certain time intervals to obtain historical temperature time series data for each node in the main shaft component area, and the duration of this historical temperature time series data shall be no less than one year.

[0046] This embodiment selects a grid-connected doubly-fed 2MW wind turbine from a wind farm in North China. SCADA (Supervisory Control and Data Acquisition) system monitoring data from March 2021 to June 2022 is used as the turbine's status parameters, and the main shaft gearbox side oil temperature is selected as the target parameter to detect and analyze the main shaft's status. This monitoring data covers: nacelle X and Y axis vibration values; wind speed; wind direction; rotor speed; gearbox input and output shaft oil temperatures; gearbox inlet oil temperature; gearbox side oil temperature; temperatures at multiple points on the generator stator windings; ambient temperature; and generator air-cooling temperature. The required temperature data can be directly obtained from this monitoring data.

[0047] Step 2: Data Filtering and Processing

[0048] Let σ be the historical temperature time series data of a certain node in the spindle component region. Then, the dataset τ of the historical temperatures of all nodes in the spindle component region is obtained from step one. σ =(σ1, σ2, ..., σ) h ), where h is the dimension of data collection, representing the number of nodes set during the collection process, and σ x This represents the historical temperature time series data of the oil on the main spindle gearbox side, with the oil temperature on the main spindle gearbox side as the target data. Correlation coefficient analysis was used to select data related to σ. x Nodes with a correlation coefficient greater than 0.8 are considered as strongly correlated data, resulting in q strongly correlated data points with a time scale of T. Each strongly correlated data point is then divided into two continuous time series data segments at a time scale of T in a ratio of 7:3. Correspondingly, the time scale of the first segment is T1, and the time scale of the second segment is T2, i.e., T1:T2 = 7:3.

[0049] The training dataset Q is obtained by processing the q strongly correlated data points from the first T1 time scales and the corresponding historical temperature data of the spindle gearbox side oil from the first T1 time scales. σ Specifically: Q σ The first data point corresponds to one data point. σ x1 This is the historical temperature data of the spindle gearbox side oil at the first moment, due to σ at the first moment. x1 Since there is no previous time step, we choose σ here. x1 As data from its previous point in time, These are the data values ​​at the first time point for the q strongly correlated data points.

[0050] Q σ The second time point corresponds to one data point. Here σ x2 The data at the previous time point is σ x1 Therefore, Q is directly used as the value of the data. σThe data corresponding to time point t is

[0051] The test dataset is obtained by performing the same processing on the q strongly correlated data at the next T2 time scales and the corresponding historical temperature data of the spindle gearbox side oil at the next T2 time scales.

[0052] As an example, eight strongly correlated data were selected using the correlation coefficient analysis method, and these eight data are shown in Table 1.

[0053] Table 1. Strong Correlation Data

[0054]

[0055]

[0056] Step 3: Establish the WCNN-LSTM neural network model

[0057] The WCNN-LSTM neural network model comprises a WCNN neural network module and an LSTM neural network module. The output of the WCNN neural network module, after being fused with an attention mechanism, is used as the input to the LSTM neural network module. The output of the LSTM neural network module is the output of the WCNN-LSTM neural network model. The WCNN neural network module includes a wide-kernel convolutional layer (WConv), a first convolutional layer (Conv), a first max-pooling layer (MaxPooling), a second convolutional layer (Conv), a second max-pooling layer (MaxPooling), and a first batch normalization (BN) layer. These six parts are sequentially connected, with the output of the previous part serving as the input to the next part. The output of the first BN layer is the output of the WCNN neural network module. The LSTM neural network module includes a first LSTM layer, a second LSTM layer, a flattened layer, a first dense layer, a second BN layer, a second dense layer, and an output layer. These seven parts are sequentially connected, with the output of the previous part serving as the input to the next part. The output of the WCNN neural network module, after being fused with an attention mechanism, is used as the input to the first LSTM layer. The input to the output layer is the output of the LSTM neural network module, which is the output of the WCNN-LSTM neural network model.

[0058] The convolution operation process of the wide-kernel convolutional layer (WConv) in the WCNN neural network module is as follows: using W... i k and Let x represent the weights and biases of the i-th convolutional operation in the k-th layer, and use x k (j) represents the j-th local region of the k-th layer, and the process is described as follows:

[0059]

[0060] Where * represents the inner product of the kernel and the layout region. The result after convolution is represented by a 100*100 kernel. Convolutions are performed sequentially along the time direction with a stride of 3. The convolutional result is then processed using batch normalization (BN) to accelerate training and improve accuracy. β is used to represent the convolutional result as input data. The batch normalization process is as follows:

[0061]

[0062] The parameters γ and d are learned during training and then activated using a rectified linear unit (ReLU function). Using x to represent the input data and y to represent the output, the activation process is as follows:

[0063]

[0064] Reassemble the activated y groups in order. This is the output of a wide-kernel convolutional layer.

[0065] The output of the wide-kernel convolutional layer is input into the first convolutional layer (Conv). The convolution operation of the first convolutional layer (Conv) is the same as that of the wide-kernel convolutional layer (WConv), but the convolutional kernel is chosen to be 9*9 in size. ReLU is also used for activation, and the activation result is input into the max pooling layer. The pooling region is selected as... The result is P. k The pooling process is as follows:

[0066]

[0067] The pooling region is selected to be 3*3 in size, and pooling is performed sequentially over time with a stride of 3. The pooled result is defined as... The output is then fed into the subsequent second convolutional layer (Conv). The convolution operation of the second convolutional layer (Conv) is the same as that of the first convolutional layer (Conv), but the kernel size is set to 3*3 and the stride is 1. The output of the second convolutional layer (Conv) is fed into the second max pooling layer (MaxPooling), where the pooling region is also selected as 3*3 and the stride is 3. The output of the second max pooling layer (MaxPooling) is fed into the first batch normalization (BN) layer, and Q is converted to Q. β As the output of the first BN layer, Q β The fusion of attention mechanisms enhances the importance of features. The process of fusion of attention mechanisms is as follows: For Q... β For each value Q βn Set a value θ βnθ β As the attention distribution for each value, the model's output is trained with the attention distribution based on different values. The mathematical formula is:

[0068]

[0069] Where q is the query vector, which is the true value of the target data, N is the total number of input values, and V, W, and U are all learnable network parameters.

[0070] The output of the first Batch Normalization (BN) layer is combined with the attention mechanism to obtain the output Q. γ .

[0071] Output Q γ As the input to the first LSTM layer of the LSTM neural network module, the first LSTM layer analyzes the temporal characteristics of the current data. The first LSTM layer consists of F LSTM units, which mainly consist of an input gate, an output gate, and a forget gate. The forget gate f t The formula is described as follows:

[0072] f t =∈(W f [h t-1 X t ]+b f )

[0073] In the formula h t-1 x is the output of the previous LSTM unit. t Let be the input of the current unit, ∈ be the hyperbolic tangent activation function, and W and b be the weight matrix and bias matrix, respectively. Its input gate mainly consists of the following two parts: i t It determines which information is updated, and To update the information, ReLU is chosen as the activation function, and the formula is described below:

[0074] i t =∈(W i [h t-1 X t ]+b i )

[0075]

[0076] The output gate mainly consists of long-term stored information C t Output gate layer t and output result h t The composition and formula are described as follows:

[0077]

[0078] o t=∈(W o [h t-1 X t ]+b o )

[0079] h t =o t *ReLU(C t )

[0080] The number of units is typically set to 128.

[0081] Q γ After processing by the first LSTM layer, the result is then input into the second LSTM layer; the number of units in the second LSTM layer is typically set to 64, and the output result at this time is... Will Dimensionality reduction is performed using a Flatten layer. The reduced result is then input into the first Dense layer to further reduce the dimensionality of the features. The dimension parameter is set to 64 at this stage. The output of the first Dense layer is then input into the second Batch Normalization (BN) layer, and the output of the second BN layer is input into the second Dense layer. The dimension parameter of the second Dense layer is set to 1, reducing the final result to a 1-dimensional Q-value. δ Q δ It serves as both the input to the Output layer and a reference value for the target data output by the WCNN-LSTM neural network model.

[0082] The WCNN neural network module first uses wide-kernel convolutional layers to extract features from the data. The convolutional layers weight the input data matrix according to the kernel size and perform sequential convolutions according to the stride, thereby extracting relevant features by adjusting the weights and biases. However, this often requires data cleaning to obtain relatively stable and comprehensive data. Therefore, a larger convolutional kernel and a longer stride are used to prevent missing data and drastic changes from affecting feature extraction. After feature extraction in the wide-kernel convolutional layers, the extracted features are input into a normal-sized convolutional layer for further feature extraction. Then, the features are input into a max-pooling layer for feature compression. Similar to the convolutional layers, the max-pooling layer selects the maximum value in the pooling region, reducing the amount of data while preserving features. The results are then input into the convolutional-max-pooling layer for training, and simultaneously into a Batch Normalization (BN) layer to accelerate training. The BN layer mainly improves the model's accuracy and speed by normalizing and setting its own learning parameters, ultimately yielding the trained feature matrix.

[0083] The feature matrix is ​​combined with the attention mechanism and input into the LSTM neural network module. The LSTM layer sets weights for different positions of the matrix to increase the importance of key features. The LSTM analyzes the relationship between features to find potential relationships in the time series. After training, the data is input into the LSTM layer again for feature learning. At this time, the data is still a data matrix.

[0084] The data matrix is ​​input into the Flatten layer for dimensionality reduction, reducing the data to 1 dimension. Then, the data is input into the Dense layer to extract the correlation between feature variables. The Dense layer undergoes non-linear transformation to extract the correlation between these features. Finally, it is mapped onto the output space and input into the Batch Normalization (BN) layer to accelerate model training. Finally, it is input into the Dense layer again for dimensionality reduction of the model, and finally, the estimated reference value is obtained.

[0085] Step 4: Train the WCNN-LSTM neural network model using the training dataset.

[0086] 4.1 Training Parameter Settings

[0087] Set fixed parameter values ​​for the neural network model, set the initial values ​​of the weight parameters to be learned to 1, set the initial values ​​of the bias parameters to 0, set the number of training iterations to 100, set the batch size to 40, and set the optimizer to AdaMod.

[0088] Table 2 Fixed parameter settings for neural network models

[0089]

[0090] 4.2 Training the Neural Network Model

[0091] The training dataset Q from step two σ As input to the neural network model, the training dataset Q is sequentially... σ Each piece of data in the dataset is input into the neural network model for training, and Q... σt The reference value of the main shaft gearbox side oil temperature at time t is obtained from the output of the neural network model, based on the training dataset Q. σ Reference values ​​for the spindle gearbox side oil temperature at time scale T1 are obtained sequentially. The mean squared error (MSE) is used to calculate the loss value between the reference values ​​for the spindle gearbox side oil temperature obtained in this training round and the historical real data.

[0092]

[0093] Where n represents the population, and here it represents the time scale T1, Y i For historical real data, f(Y) i ) is the reference value output by the network model.

[0094] Based on the loss value, the AdaMod optimizer is used to perform backpropagation to update the weights and bias parameters once, completing one iteration of training. The updated neural network model parameters are then used as the initial parameters, and the training dataset Q is used again. σ The data is sequentially input into the neural network model for the next iteration of training; when the number of training iterations reaches a preset value, the weights and bias parameters of the last update are saved to obtain the trained neural network model.

[0095] 4.3 Testing the Neural Network Model

[0096] The data from the test dataset in step two are sequentially input into the trained neural network model to obtain reference values ​​for the spindle gearbox side oil temperature at the next T2 time scales. The degree of fit R between the reference values ​​and the historical real data is then calculated. 2 When the value is greater than 0.9, the trained neural network model is verified as an effective model.

[0097] To more thoroughly examine the accuracy of the reference values ​​output by the neural network model, this embodiment sets up four different test conditions, all of which are compared with the outputs of the algorithms LightGBM, BiLSTM, and CNN-LSTM. The four test conditions are: no missing data and no LMG variable added to the previous time step (i.e., no missing data and no added LMG variable), no missing data and added LMG variable, missing data and no added LMG variable, and missing data and added LMG variable. The model is trained using q strongly correlated data points from the first T1 time scales and the corresponding historical temperature data of the spindle gearbox side oil from the first T1 time scales. The loss is calculated using the mean squared error (MSE). Then, the model is tested using q strongly correlated data points from the next T2 time scales and the corresponding historical temperature data of the spindle gearbox side oil from the next T2 time scales. The data of the reference values ​​and true values ​​output by each neural network model under different test conditions are as follows: Figure 2-5 The relevant configurations for each neural network model are the same during training and testing within the same group.

[0098] To more intuitively demonstrate the comparison results of different neural network models, this embodiment introduces the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and goodness of fit (R). 2 Four metrics are used to evaluate different neural network models, as shown in the following formula:

[0099]

[0100]

[0101]

[0102]

[0103] Where n represents the population, and here it is the time scale T2, Y i Based on historical data, f(Y) is the average of historical real data. i The value is the reference value output by the network model. Table 3-6 compares the accuracy of the reference values ​​output by different neural network models under four different test conditions.

[0104] Table 3 shows the LMG variable without missing data.

[0105]

[0106]

[0107] Table 4 shows the addition of the LMG variable when there are no missing data.

[0108]

[0109] Table 5 shows missing data where the LMG variable was not added.

[0110]

[0111] Table 6: Adding the LMG variable for missing data.

[0112]

[0113]

[0114] The above comparative tests show that the WCNN-LSTM neural network model designed in this invention has better accuracy and wider applicability.

[0115] Step 5: Spindle Failure Prediction

[0116] The historical data of the main shaft gearbox side oil temperature for the first T1 time scales are compared with the data from the training dataset Q. σ After processing by the trained neural network model, the reference values ​​of the main shaft gearbox side oil temperature obtained in the first T1 time scales are subtracted to obtain the residual sequence. The mean of the residual sequence is calculated using a sliding window of a certain time scale and set as the standard deviation μ.

[0117] The current actual temperature of the spindle gearbox side oil at the current m-th time point, the corresponding strongly correlated data, and the previous actual temperature of the spindle gearbox side oil at the previous time point are input into the neural network model trained in step three to obtain the reference value of the spindle gearbox side oil temperature at the m-th time point. The reference values ​​of the spindle gearbox side oil temperature at subsequent time points are calculated sequentially, and the mean of the residual sequence segments between the actual temperature sequence segment and the reference value sequence segment is calculated using the aforementioned sliding window at a certain time scale. When the mean ω of the residual sequence segment corresponding to a certain sliding window... M If the value exceeds the threshold by 3μ, it can be determined that the spindle may be malfunctioning. The following are the residuals under normal conditions:

[0118] |ω M |≤3μ

[0119] To verify the fault identification capability of the method of the present invention, data was collected from a wind turbine with missing data and frequent faults. The data was then used for model training and fault diagnosis. The resulting curve of the reference value and the actual value of the oil temperature on the main shaft gearbox side is shown in the figure. Figure 6 As can be seen from the figure, the method of the present invention can predict the occurrence of a fault at least 16 hours in advance (the x-axis in the figure is the time point number, the interval between two adjacent time points is 10 minutes, the fault can be predicted to occur near the 100th time point to the left of the x-axis, and the fault actually occurs near the 200th time point to the left).

[0120] This document illustrates the principles and implementation of the present invention through specific examples. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of the present invention. Furthermore, based on the ideas of the present invention, those skilled in the art will find some variations in specific implementations and application scope. In conclusion, the content of this specification should not be construed as limiting the present invention.

[0121] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for predicting faults in wind turbine main shaft components based on wide kernel convolution, characterized in that, The specific implementation steps of this prediction method are as follows: Step 1: Install temperature sensors at multiple nodes in the main shaft component area inside the fan, collect data points at certain time intervals to obtain historical temperature time series data for each node in the main shaft component area, and the duration of this historical temperature time series data shall be no less than one year. Step 2: Data Filtering and Processing Let σ be the historical temperature time series data of a certain node in the spindle component region. Then, the dataset τ of the historical temperatures of all nodes in the spindle component region is obtained from step one. σ =(σ1, σ2, ..., σ) h ), where h is the dimension of data collection, representing the number of nodes set during the collection process, and σ x The historical temperature time series data represents the oil temperature on the spindle gearbox side, with the oil temperature on the spindle gearbox side as the target data; correlation coefficient analysis was used to select data related to σ. x Nodes with a correlation coefficient greater than 0.8 are considered as strongly correlated data, and q strongly correlated data are obtained with a time scale of T. Each strongly correlated data is divided into two continuous time series data segments with a time scale of T in a ratio of 7:

3. Correspondingly, the time scale of the first part is T1, and the time scale of the second part is T2, i.e., T1:T2 = 7:

3. The training dataset Q is obtained by processing the q strongly correlated data points from the first T1 time scales and the corresponding historical temperature data of the spindle gearbox side oil from the first T1 time scales. σ Specifically: Q σ The first data point corresponds to one data point. σ x1 This is the historical temperature data of the spindle gearbox side oil at the first moment, due to σ at the first moment. x1 Since there is no previous time step, we choose σ here. x1 As data from its previous point in time, These are the data values ​​at the first time point for the q strongly correlated data points; Q σ The second time point corresponds to one data point. Here σ x2 The data at the previous time point is σ x1 Therefore, Q is directly used as the value of the data. σ The data corresponding to time point t is The q strongly correlated data points at the next T2 time scales and the corresponding historical temperature data of the spindle gearbox side oil at the next T2 time scales were processed in the same way to obtain the test dataset; Step 3: Establish the WCNN-LSTM neural network model The WCNN-LSTM neural network model includes a WCNN neural network module and an LSTM neural network module. The output of the WCNN neural network module, after being fused with an attention mechanism, is used as the input of the LSTM neural network module, and the output of the LSTM neural network module is the output of the WCNN-LSTM neural network model. The WCNN neural network module includes a wide-kernel convolutional layer, a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, and a first batch normalization (BN) layer. These six parts are sequentially connected, with the output of the previous part serving as the input of the next part. The output of the first BN layer is used as the output of the WCNN neural network module. The LSTM neural network module includes a first LSTM layer, a second LSTM layer, a Fasten layer, a first Dense layer, a second BN layer, a second Dense layer, and an Output layer. These seven parts are sequentially connected, with the output of the previous part serving as the input of the next part. The output of the WCNN neural network module, after being fused with an attention mechanism, is used as the input of the first LSTM layer, and the input of the Output layer is the output of the LSTM neural network module, which is the output of the WCNN-LSTM neural network model. The convolution operation process of the wide-kernel convolutional layer of the WCNN neural network module is as follows: using W... i k and Let x represent the weights and biases of the i-th convolutional operation in the k-th layer, and use x k (j) represents the j-th local region of the k-th layer, and the process is described as follows: Where * represents the inner product of the kernel and the layout region. The result after convolution is represented by a kernel size of 100*100. Convolutions are performed sequentially along the time direction with a stride of 3. The convolutional result is then batch-normalized. β is used to represent the convolutional result as input data. The batch normalization process is as follows: Here, parameters γ and d are learned during training and then activated using rectified linear units. Let x represent the input data and y represent the output. The activation process is as follows: Reassemble the activated y groups in order. This is the output of a wide-kernel convolutional layer; Step 4: Train the WCNN-LSTM neural network model using the training dataset. 4.1 Training Parameter Settings Set the fixed parameter values ​​of the neural network model, set the initial values ​​of the weight parameters to be learned to 1, set the initial values ​​of the bias parameters to 0, set the number of training iterations to 100, set the batch size to 40, and set the optimizer to AdaMod. 4.2 Training the Neural Network Model The training dataset Q from step two σ As input to the neural network model, the training dataset Q is sequentially... σ Each piece of data in the dataset is input into the neural network model for training, and Q... σt The reference value of the main shaft gearbox side oil temperature at time t is obtained from the output of the neural network model, based on the training dataset Q. σ Reference values ​​for the spindle gearbox side oil temperature at time scale T1 are obtained sequentially. The mean squared error (MSE) is used to calculate the loss value between the reference values ​​for the spindle gearbox side oil temperature obtained in this training round and the historical real data. Where n represents the population, and here it represents the time scale T1, Y i For historical real data, f(Y) i ) is a reference value output by the network model; Based on the loss value, the AdaMod optimizer is used to perform backpropagation to update the weights and bias parameters once, completing one iteration of training. The updated neural network model parameters are then used as the initial parameters, and the training dataset Q is used again. σ The data is sequentially input into the neural network model for the next iteration of training; when the number of training iterations reaches a preset value, the weights and bias parameters of the last update are saved to obtain the trained neural network model. 4.3 Testing the Neural Network Model The data from the test dataset in step two are sequentially input into the trained neural network model to obtain reference values ​​for the spindle gearbox side oil temperature at the next T2 time scales. The degree of fit R between the reference values ​​and the historical real data is then calculated. 2 When the value is greater than 0.9, the trained neural network model is verified as an effective model; Step 5: Spindle Failure Prediction The historical data of the main shaft gearbox side oil temperature for the first T1 time scales are compared with the data from the training dataset Q. σ After processing by the trained neural network model, the reference values ​​of the main shaft gearbox side oil temperature obtained in the first T1 time scales are subtracted to obtain the residual sequence. The mean of the residual sequence is calculated using a sliding window of a certain time scale and set as the standard deviation μ. The current actual temperature of the spindle gearbox side oil at the current m-th time point, the corresponding strongly correlated data, and the previous actual temperature of the spindle gearbox side oil are input into the neural network model trained in step three to obtain the reference value of the spindle gearbox side oil temperature at the m-th time point. The reference values ​​of the spindle gearbox side oil temperature at subsequent time points are calculated sequentially, and the mean of the residual sequence segments between the actual temperature sequence segment and the reference value sequence segment is calculated using the aforementioned sliding window at a certain time scale. When the mean ω of the residual sequence segment corresponding to a certain sliding window... M If the value exceeds the threshold by 3μ, it can be determined that the spindle may be malfunctioning.

2. The method for predicting faults in wind turbine main shaft components based on wide kernel convolution as described in claim 1, characterized in that, The output of the wide-kernel convolutional layer is input into the first convolutional layer. The convolution operation of the first convolutional layer is the same as that of the wide-kernel convolutional layer, but the convolutional kernel is chosen to be 9*9 in size. ReLU activation is also used, and the activation result is input into the max pooling layer. The pooling region is selected as... The result is P. k The pooling process is as follows: The pooling region is selected to be 3*3 in size, and pooling is performed sequentially over time with a stride of 3. The pooled result is defined as... The data is then input into the subsequent second convolutional layer. The convolution operation process of the second convolutional layer is the same as that of the first convolutional layer, but the kernel size is set to 3*3 and the stride is 1. The output of the second convolutional layer is input into the second max pooling layer, and the pooling region is also selected as 3*3 with a stride of 3. The output of the second max pooling layer is input into the first BN layer.

3. The method for predicting faults in wind turbine main shaft components based on wide kernel convolution as described in claim 1, characterized in that, Q β As the output of the first BN layer, Q β The process of integrating attention mechanisms is as follows: For Q β For each value Q βn Set a value θ βn θ β As the attention distribution for each value, the model's output is trained with different values ​​using the attention distribution, and the mathematical formula is: Where q is the query vector, which is the true value of the target data, N is the total number of input values, and V, W, and U are all learnable network parameters.

4. The method for predicting faults in wind turbine main shaft components based on wide kernel convolution as described in claim 1, characterized in that, The output Q obtained by fusing the output of the first BN layer with the attention mechanism γ This is the input to the first LSTM layer of the LSTM neural network module. The first LSTM layer analyzes the temporal characteristics of the current data. The first LSTM layer consists of 128 LSTM units, which mainly consist of an input gate, an output gate, and a forget gate. The forget gate f... t The formula is described as follows: f t =∈(W f [h t-1 ,X t ]+b f ) In the formula h t-1 X is the output of the previous LSTM unit. t For the input of the current unit, ∈ is the hyperbolic tangent activation function, and W and b are the weight matrix and bias matrix, respectively; its input gate mainly consists of the following two parts: i t It determines which information is updated, and To update the information, ReLU is chosen as the activation function, and the formula is described below: i t =∈(W i [h t-1 ,X t ]+b i ) The output gate mainly consists of long-term stored information C t Output gate layer t and output result h t The composition and formula are described as follows: o t =∈(W o [h t-1 ,X t ]+b o ) h t =one t *ReLU(C t ) Q γ After processing by the first LSTM layer, the result is then input into the second LSTM layer; the number of cells in the second LSTM layer is set to 64, and the output result at this time is... Will Dimensionality reduction is performed using a Flatten layer. The reduced result is then input into the first Dense layer to further reduce the dimensionality of the features. The dimension parameter is set to 64 at this stage. The output of the first Dense layer is then input into the second Batch Normalization (BN) layer, and the output of the second BN layer is input into the second Dense layer. The dimension parameter of the second Dense layer is set to 1, reducing the final result to a 1-dimensional Q-value. δ Q δ It serves as both the input to the Output layer and a reference value for the target data output by the WCNN-LSTM neural network model.

Images