A wind turbine gearbox fault early warning method based on DAE-LSTM-KDE model

Abstract

This invention discloses a wind turbine gearbox fault early warning method based on a DAE-LSTM-KDE model. It employs a deep autoencoder (DAE) to extract temporal features hidden in historical data, constructs a monitoring model of a wind turbine gearbox under normal conditions based on an LSTM neural network, and sets an alarm threshold using kernel density estimation (KDE) based on the obtained output variable prediction residuals. The method includes the following steps: cleaning and normalizing outliers in the raw data from the SCADA system; extracting temporal features hidden in the data using a deep autoencoder; constructing a monitoring model of a wind turbine gearbox under normal conditions based on an LSTM neural network; setting an alarm threshold using kernel density estimation (KDE) based on the obtained output variable prediction residuals; inputting samples from normal operation to the fault occurrence stage into the trained model; and issuing an early warning signal when the prediction residuals exceed the set fault threshold.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation, specifically relating to a wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model. Background Technology

[0002] The gearbox is a crucial component of a wind power generation system. Due to its complex and variable operating environment and long-term operation under alternating load conditions, the probability of failure is significantly increased. Data shows that frequent electrical and control system faults caused by system tripping can generally be resolved through remote reset. While gearbox failures are less frequent, they result in the longest downtime. To reduce equipment maintenance costs and minimize economic losses due to downtime, monitoring technology that can issue early warning signals for gearbox failures is particularly important for ensuring the safe and reliable operation of wind turbines.

[0003] The development of artificial intelligence technology has made data-driven methods based on machine learning an efficient and low-cost solution for gearbox fault early warning. However, most current methods involve human selection of input parameters, which is highly subjective. The models built may not accurately reflect the fault information contained therein. Moreover, most existing studies only focus on the selection of input variables in the prediction process, but ignore the impact of features extracted from the input variables on the model output.

[0004] Currently, deep learning algorithms, represented by convolutional neural networks and deep autoencoders (DAEs), demonstrate advantages in mining deep features of data. However, the data in existing SCADA systems contains information from multiple sensors at a single spatiotemporal scale, which cannot fully extract the essential characteristics that reflect the unit's operating status.

[0005] Purpose of the invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model. Summary of the Invention

[0007] This invention provides a wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model, comprising the following steps:

[0008] Step 1: Use a deep autoencoder (DAE) to extract the temporal features hidden in the historical data, construct a monitoring model of the wind turbine gearbox under normal conditions based on an LSTM neural network, and use the kernel density estimation (KDE) method to set the alarm threshold based on the predicted residuals of the output variables; clean and normalize the outliers in the raw data of the SCADA system.

[0009] Step 2: Feature extraction is completed through the DAE encoding process. The input signal x is mapped to the hidden feature z through a nonlinear relationship. At the same time, the dimension of the hidden layer is limited to be lower than the dimension of the original input data to ensure the effectiveness of the algorithm. The calculation formula is shown in Equation (1):

[0010]

[0011] In the formula, It is a nonlinear function, W is the weight matrix, and b is the bias matrix;

[0012] Step 3: Input the features extracted in Step 2 into the LSTM neural network to construct the DAE-LSTM state prediction model and train it. This further includes the following sub-steps:

[0013] Sub-step 3-1: Use a sliding window to collect the extracted feature data as input to the prediction model, and adjust the window length to capture data changes within a small range;

[0014] Sub-step 3-2: Apply the Particle Swarm Optimization (PSO) algorithm to the optimal combination of hyperparameters such as the number of hidden layer neurons, the sliding time window step size, and the initial learning rate in the LSTM network;

[0015] Sub-step 3-3: Information flow is restricted by the forget gate and the input gate. The forget gate determines the information forgotten from the cell state, while the input gate selectively allows newly input information to pass through, thus retaining the effective information. The calculation formulas are shown in equations (2)-(5):

[0016] f t =σ(W f x t +W f h t-1 +b f (2),

[0017] i t =σ(W i x t +W i h t-1 +b i (3),

[0018]

[0019]

[0020] Among them, f t i t These represent the forget gate and the input gate, respectively. c represents the state of the candidate unit at that moment.t This represents the updated cell state at that moment; x t h represents the input information at the current moment. t-1 c is the input from the previous time step. t-1 The state at the previous moment; W f W i W c and b f b i b c These are the weights and biases of the forget gate and input gate, respectively, and * represents the dot product;

[0021] Sub-steps 3-4: Determine the value of the next hidden state through the output gate, and the calculation formula is shown in equations (6)-(7):

[0022] o t =σ(W p x t +W p h t-1 +b i (6),

[0023] h t =o t *tanh(c t (7),

[0024] Among them, o t Represents the output gate, W p and b p These are the weights and biases of the output gate, respectively.

[0025] Sub-steps 3-5: Set the number of LSTM network layers to 3 and 1 respectively, select ReLU as the activation function, and add a fully connected layer with dropout strategy before the output layer, using the sigmoid function as the activation function of the full layer. Train and update the network weights through the Adam optimizer and train the model.

[0026] Step 4: Calculate the residual statistical characteristics during normal operation using the kernel density estimation KDE method and determine the corresponding alarm threshold. This further includes the following sub-steps:

[0027] Step 4-1: Using the Gaussian function as the kernel function, perform probability statistics on the absolute residual sequence using the nonparametric statistical KDE method, as follows:

[0028]

[0029]

[0030]

[0031] in, Let n be the probability density function of the residual sequence, n be the number of samples, H be the optimal bandwidth parameter of the kernel function, and e be the probability distribution density of the residual sequence. i Let S be the i-th sample of the residual sequence, S be its standard deviation, and Q be the difference between the 75th and 25th percentiles of the sample.

[0032] Step 4-2: Using the probability density distribution of the prediction residuals under the normal model as a reference, obtain the fault threshold e under the condition of confidence level 1-α. r It is represented as shown in equation (11):

[0033] P(ee r )=1-α (11);

[0034] Step 5: Input the online data into the trained DAE-LSTM model to obtain the prediction residuals. Under this model, the absolute residual values ​​of the state variable output predictions are obtained. The specific calculation is expressed as shown in Equation (11):

[0035]

[0036] When the predicted residual exceeds the set fault threshold, it indicates that the gearbox has deviated from its normal operating state, and a warning signal is issued. Attached Figure Description

[0037] Figure 1 This is a flowchart of the wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model described in this invention.

[0038] Figure 2 This is a schematic diagram of the framework of the wind turbine DAE-LSTM-KDE fault early warning system described in this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Those skilled in the art will understand that the step numbers used herein are for ease of description only and are not intended to limit the order in which the steps are performed. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” may include the plural forms unless the context clearly indicates otherwise. The term “and / or” refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. In the description of this invention, “a plurality” means at least two, such as two, three, etc., unless otherwise expressly specified.

[0041] Figure 1 This is a flowchart of the wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model described in this invention. Figure 2 This is a schematic diagram of the framework of the wind turbine DAE-LSTM-KDE fault early warning system of the present invention. As shown in the figure, the early warning method includes the following steps:

[0042] Step 1: Clean and normalize outliers in the raw data of the SCADA system;

[0043] Step 2: Feature extraction is completed through the DAE encoding process, mapping the input signal x to latent features z through a non-linear relationship. Simultaneously, the dimension of the hidden layer is limited to be lower than the dimension of the original input data to ensure the effectiveness of the algorithm. The calculation is as follows:

[0044]

[0045] Step 3: Input the features extracted in Step 2 into the LSTM neural network to construct the DAE-LSTM state prediction model and train it. The specific steps are as follows:

[0046] Step 3-1: Use a sliding window to collect the extracted feature data as input to the prediction model, and adjust the window length to capture data changes within a small range.

[0047] Step 3-2: Apply the particle swarm optimization (PSO) algorithm to the optimal combination of hyperparameters in the LSTM network, such as the number of hidden layer neurons, the sliding time window step size, and the initial learning rate.

[0048] Step 3-3: Information flow is restricted using a forget gate and an input gate. The forget gate determines the information forgotten from the cell state, while the input gate selectively allows newly input information to pass through, thus retaining the valid information. The operation is as follows:

[0049] ft =σ(W f x t +W f h t-1 +b f )#(2)

[0050] i t =σ(W i x t +W i h t-1 +b i )#(3)

[0051]

[0052]

[0053] Among them, f t i t These represent the forget gate and the input gate, respectively. c represents the state of the candidate unit at that moment. t x represents the cell state after the update at that moment. t h represents the input information at the current moment. t-1 c is the input from the previous time step. t-1 W represents the state at the previous moment. f W i W c and b f ,b i ,b c These are the weights and biases of the forget gate and input gate, respectively, and * represents the dot product.

[0054] Steps 3-4: Determine the value of the next hidden state through the output gate, as follows.

[0055] o t =σ(W p x t +W p h t-1 +b i )#(6)

[0056] h t =o t *tanh(c t )#(7)

[0057] Among them, o t Represents the output gate, W p and b p These are the weights and biases of the output gate, respectively.

[0058] Steps 3-5: Set the number of LSTM network layers to 3 and 1 respectively, select ReLU as the activation function, and add a fully connected layer with dropout policy before the output layer, using the sigmoid function as the activation function of the full layer. Train and update the network weights through the Adam optimizer and train the model.

[0059] Step 4: Calculate the residual statistical characteristics during normal operation using the kernel density estimation (KDE) method and determine the corresponding alarm threshold. The specific steps are as follows:

[0060] Step 4-1: Using the Gaussian function as the kernel function, perform probability statistics on the absolute residual sequence using the nonparametric statistical KDE method, as follows:

[0061]

[0062]

[0063]

[0064] in, Let n be the probability density function of the residual sequence, n be the number of samples, H be the optimal bandwidth parameter of the kernel function, and e be the probability distribution density of the residual sequence. i Let S be the i-th sample of the residual sequence, S be its standard deviation, and Q be the difference between the 75th and 25th percentiles of the sample.

[0065] Step 4-2: Using the probability density distribution of the prediction residuals under the normal model as a reference, obtain the fault threshold e under the condition of confidence level 1-α. r :

[0066] P(ee r )=1-α#(11)

[0067] Step 5: Input the online data into the trained DAE-LSTM model to obtain the prediction residuals. Under this model, obtain the absolute residual values ​​of the predicted state variables. The specific calculation is as follows:

[0068]

[0069] When the predicted residual exceeds the set fault threshold, it indicates that the gearbox has deviated from its normal operating state, and a warning signal is issued.

[0070] To address the issue of low accuracy in wind turbine condition monitoring models caused by the difficulty of traditional machine learning methods in extracting multidimensional temporal features from massive SCADA data, this invention establishes a DAE-LSTM-KDE fault early warning model for wind turbines, with gearbox oil temperature as the target output, based on historical operating data of the SCADA system. This model can detect abnormal gearbox conditions in a timely and effective manner, providing sufficient time for the inspection and maintenance of wind turbine gearboxes and further improving the prediction accuracy of important parameter models for wind turbines, enabling accurate judgment of abnormal gearbox operating conditions.

Claims

1. A wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model, characterized in that, Includes the following steps: Step 1: Use a deep autoencoder (DAE) to extract the temporal features hidden in the historical data, construct a monitoring model of the wind turbine gearbox under normal conditions based on an LSTM neural network, and use the kernel density estimation (KDE) method to set the alarm threshold based on the predicted residuals of the output variables; clean and normalize the outliers in the raw data of the SCADA system. Step 2: Feature extraction is completed through the DAE encoding process. The input signal x is mapped to the hidden feature z through a nonlinear relationship. At the same time, the dimension of the hidden layer is limited to be lower than the dimension of the original input data to ensure the effectiveness of the algorithm. The calculation formula is shown in Equation (1): In the formula, It is a nonlinear function, W is the weight matrix, and b is the bias matrix; Step 3: Input the features extracted in Step 2 into the LSTM neural network to construct the DAE-LSTM state prediction model and train it. Step 4: Calculate the residual statistical characteristics during normal operation using the kernel density estimation KDE method and determine the corresponding alarm threshold; Step 4 further includes the following sub-steps: Step 4-1: Using the Gaussian function as the kernel function, perform probability statistics on the absolute residual sequence using the nonparametric statistical KDE method, as follows: in, Let n be the probability density function of the residual sequence, n be the number of samples, H be the optimal bandwidth parameter of the kernel function, and e be the probability distribution density of the residual sequence. i Let S be the i-th sample of the residual sequence, S be its standard deviation, and Q be the difference between the 75th and 25th percentiles of the sample. Step 4-2: Using the probability density distribution of the predicted residuals under the normal model as a reference, obtain the fault threshold e under the condition of confidence level 1-α. r It is represented as shown in equation (11): P(yes r )=1-α (11), Step 5: Input the online data into the trained DAE-LSTM model to obtain the prediction residuals. Under this model, the absolute residual values ​​of the state variable output predictions are obtained. The specific calculation is expressed as shown in Equation (11): When the predicted residual exceeds the set fault threshold, it indicates that the gearbox has deviated from its normal operating state, and a warning signal is issued.

2. The wind turbine gearbox fault early warning method based on the DAE-LSTM-KDE model according to claim 1, characterized in that, Step 3 further includes the following sub-steps: Sub-step 3-1: Use a sliding window to collect the extracted feature data as input to the prediction model, and adjust the window length to capture data changes within a small range; Sub-step 3-2: Apply the Particle Swarm Optimization (PSO) algorithm to the optimal combination of hyperparameters for the number of hidden layer neurons, sliding time window step size, and initial learning rate in the LSTM network; Sub-step 3-3: Information flow is restricted by the forget gate and the input gate. The forget gate determines the information forgotten from the cell state, while the input gate selectively allows newly input information to pass through, thus retaining the effective information. The calculation formulas are shown in equations (2)-(5): f t =σ(W f x t +W f h t-1 +b f ) (2), i t =σ(W i x t +W i h t-1 +b i ) (3), Among them, f t i t These represent the forget gate and the input gate, respectively. c represents the state of the candidate unit at that moment. t This represents the updated cell state at that moment; x t h represents the input information at the current moment. t-1 c is the input from the previous time step. t-1 The state at the previous moment; W f W i W c and b f b i b c These are the weights and biases of the forget gate and input gate, respectively, and * represents the dot product; Sub-steps 3-4: Determine the value of the next hidden state through the output gate, and the calculation formula is shown in equations (6)-(7): o t =σ(W o x t +W o h t-1 +b i ) (6), h t =o t *fishy(c) t ) (7), Among them, o t Represents the output gate, W o and b o These are the weights and biases of the output gate, respectively. Sub-steps 3-5: Set the number of LSTM network layers to 3 and 1 respectively, select ReLU as the activation function, and add a fully connected layer with dropout strategy before the output layer. Use the sigmoid function as the activation function of the fully connected layer, train and update the network weights through the Adam optimizer, and train the model.

Images