A method for predicting the remaining life probability of key components of large wind turbine units

Abstract

The application belongs to the technical field of wind turbines, and relates to a residual life probability prediction method for key components of large wind turbines. After obtaining wind turbine SCADA data through a data acquisition and monitoring control system, variable screening and preprocessing are first performed, then different periodic changes of the wind turbine SCADA data are preliminarily found through Fourier transform, then multi-dimensional time series data are intercepted according to different periods, then parallel processing is performed through an LSTM neural network, time characteristics are deeply extracted, then output vectors of multiple LSTM neural networks are spliced together and residual life prediction results are obtained through a linear output layer, finally, an uncertainty quantification method based on linear regression and kernel density estimation is used to quantize the uncertainty of the original residual life prediction results, thereby providing a more robust reference for wind turbine maintenance decisions. Compared with other methods, the method can effectively improve the prediction accuracy and reliability of the residual life of wind turbine components.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine technology and relates to a method for predicting the remaining life probability of key components of large wind turbine units. Background Technology

[0002] With the rapid development of the wind power industry, wind turbines have become an important source of clean energy globally. However, due to the long-term operation of their key components (such as gearboxes, bearings, and blades) in complex natural environments, they are susceptible to fatigue, wear, and other factors, leading to failures. Failure of these components can result in equipment downtime and high maintenance costs. Therefore, predicting their remaining lifespan has become crucial for ensuring efficient wind turbine operation and reducing maintenance costs. However, the operating data of wind turbine components is characterized by multiple sources, multiple scales, and nonlinearity. Traditional physical and statistical models struggle to accurately describe their complex degradation processes and effectively capture the dynamic impacts of the environment and workload. Currently, thanks to the rapid development of sensor and communication technologies, data-driven methods are widely used in remaining lifespan prediction. These methods directly extract key information from large amounts of sensor data for condition monitoring to detect changes in system state. Among data-driven methods, machine learning and deep learning have gained significant attention from researchers in recent years, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs). However, in real-world scenarios, time-series data collected during equipment degradation typically exhibits overlap and interaction across multiple periods and is often mixed with noise signals. Many existing models still lack the ability to process such multi-scale mixed noise time-series data, making it difficult to extract comprehensive degradation features from the data. Furthermore, existing models often focus on predicting specific values ​​while ignoring the uncertainties in the prediction results, which can affect the reliability of operational decisions. Summary of the Invention

[0003] To address the aforementioned technical problems in the existing technology, this invention proposes a probabilistic prediction method for the remaining life of key components in large wind turbine units. This method can accurately predict the remaining life of wind turbine components and quantify uncertainties to obtain reliable remaining life prediction results. The specific technical solution is as follows:

[0004] A method for predicting the remaining life probability of key components of a large wind turbine includes the following steps:

[0005] Step S1: Acquire SCADA data of the wind turbine through the data acquisition and monitoring control system, filter out the input variables related to the remaining life of the target components of the wind turbine, and perform data preprocessing to obtain standardized time series data of multiple input variables;

[0006] Step S2: Perform Fast Fourier Transform on the standardized time series data of multiple input variables obtained in Step S1, and cut the time series data into different lengths according to multiple frequency components. Then, input the time series data of different lengths into different LSTM neural networks for parallel processing, extract the feature vectors at different time scales and concatenate and merge them, and then input them into the linear output layer to obtain the predicted value of the remaining life of the wind turbine components.

[0007] Step S3: The uncertainty of the predicted remaining life of the wind turbine components obtained in Step S2 is quantified based on the method of linear regression and kernel density estimation.

[0008] Furthermore, in step S1, the process of selecting input variables related to the remaining life of the target component of the wind turbine is as follows: first, select initial input variables related to the remaining life of the target component based on professional knowledge; then, calculate the Pearson correlation coefficient between other variables and the initial input variables; select variables from the other variables whose correlation coefficient is greater than a preset value p, and use them together with the initial input variables as input variables, thereby obtaining n input variables, where n is a positive integer.

[0009] Furthermore, in step S1, the preprocessing of the data specifically includes:

[0010] First, use a threshold of 5 times the standard deviation to remove extreme outliers in SCADA data, and at the same time use a threshold of active power greater than zero to remove shutdown status data.

[0011] Then, the Z-score normalization method was used to process the time series data of each input variable.

[0012] Finally, a sliding window with a step size of 1 unit time and a width of t unit time is used to segment the standardized input variable time series.

[0013] Furthermore, step S2 specifically includes:

[0014] Step S2.1: Process the time-series signal with input length t and number of input variables after window segmentation using Fast Fourier Transform to obtain n feature spectrograms, where n is a positive integer. Then, average the amplitude dimension of all feature spectrograms to obtain a new spectrogram. Based on the conjugate property of the spectrum, only half of the spectrum of the new spectrogram is used for analysis.

[0015]

[0016] In the formula, X i ∈R t×n X iLet A represent a time series of a certain input variable, A represent the corresponding amplitude of all frequency components, FFT(·) represents Fast Fourier Transform, and Amp(·) represents the calculation of amplitude.

[0017] Step S2.2: Select the first k largest amplitude values ​​from A obtained in step S2.1, and find their corresponding frequencies and periods. The calculation formula is shown below:

[0018] [f1,…,f k ] = Fre(Topk(A))

[0019]

[0020] In the formula, Topk(·) represents finding the first k maximum values, Fre(·) represents taking the frequency values, [f1,…,f k [p1,…,p] represents the first k frequency values. k [] indicates the value of the first k cycles corresponding to the frequency value;

[0021] Step S2.3: Based on the k periods obtained from step S2.2, extract k time segments of length denoted by the period value, starting from the last moment of the input time series and moving backward. l∈{1,…,k};

[0022] Step S2.4: Input the k time segments obtained in step S2.3 into different LSTM neural networks to obtain the final output vectors [h1,…,h] of the k LSTM neural networks. k ];

[0023] Step S2.5: Concatenate the k final output vectors obtained in step S2.4 and input them into the linear output layer to obtain the predicted remaining life of the wind turbine components. The expression is:

[0024] y = W l ·cat([h1,…,h k ])

[0025] In the formula, W l y represents the weights of the linear layer, y represents the final prediction result, and the cat(·) function represents concatenating the vectors column by column.

[0026] Furthermore, in step S2.2, when selecting the first k largest amplitude values ​​from A, the frequency components do not include high-frequency components.

[0027] Furthermore, step S3 specifically includes:

[0028] Step S3.1: Process the remaining lifetime prediction value obtained in step S2.5 using a sliding window with a step size of 1 unit time and a width of 1 unit time to obtain the original remaining lifetime prediction value [y1,...,y] within the sliding window. l ], and treat each sliding window as a degradation stage;

[0029] Step S3.2: Calculate the original predicted remaining lifetime values ​​[y1,...,y] within the sliding window. l Performing linear regression, the slope of the fitted regression line is the estimated degradation rate of the component predicted by the neural network during this window period, and the ordinate value at the end of the regression line is the estimated remaining lifetime at the last moment of the current window. The formula is expressed as:

[0030]

[0031] y end =wt end +b

[0032] In the formula, w and b are the slope and intercept of the regression line, respectively. and y represents the mean of the original predicted remaining life and the mean at the corresponding time, respectively. end The ordinate value at the end of the regression line within the window represents the final predicted remaining lifespan, t. end This represents the x-coordinate value at the end of the regression line;

[0033] Step S3.3: Project all the original remaining lifetime predictions obtained in Step 3.2 within the window onto the last moment of the window at the same degradation rate, i.e., along the direction of the regression line, to obtain the remaining lifetime estimate of l at the end of the degradation phase, expressed by the formula:

[0034]

[0035] In the formula, One of the remaining lifetime estimates at the last moment of the window;

[0036] Step S3.4: Estimate the probability density function of the remaining lifetime prediction distribution for the remaining lifetime prediction values ​​within a sliding window obtained in Step S3.3 using the kernel density estimation method. For l remaining lifetime estimates... Let its probability density function be f, and the kernel density estimation expression be:

[0037]

[0038] In the formula, K(·) represents the kernel function, and a represents the bandwidth;

[0039] After estimating the probability density function of the remaining lifetime prediction distribution, the 95% confidence interval of the remaining lifetime prediction distribution is further obtained.

[0040] Compared with the prior art, the technological advancements achieved by this invention are:

[0041] The method of this invention decomposes multidimensional SCADA time series to decouple the changes within different periods and between periods in complex time patterns caused by different cycles. This overcomes the problem that the multiple periods of wind turbine SCADA time series data overlap and influence each other, making variable modeling difficult to handle. It improves the accuracy of wind turbine component remaining life prediction and provides a more robust reference for wind turbine maintenance decisions by quantifying the uncertainty of the prediction results. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a method for predicting the remaining life probability of key components of a large wind turbine according to an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the multi-scale temporal feature extraction neural network MH-LSTM according to an embodiment of the present invention;

[0044] Figure 3a This is a linear regression analysis graph of the original predicted remaining lifetime values ​​within the sliding window in this embodiment;

[0045] Figure 3b yes Figure 3a The probability density analysis diagram of the corresponding remaining lifetime prediction distribution;

[0046] Figures 4a-4d The diagrams show the results of predicting the remaining life of wind turbine gearboxes and quantifying uncertainties using LSTM, BiLSTM, CNN-LSTM-AM, and MH-LSTM neural network methods, respectively. Detailed Implementation

[0047] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0048] like Figure 1 As shown in this embodiment, a method for predicting the remaining life probability of key components of a large wind turbine includes the following steps:

[0049] Step S1: Acquire SCADA data (multivariate time series data) of the wind turbine through the Supervisory Control and Data Acquisition (SCADA) system. Filter input variables that may be related to the remaining lifespan of the target component, and preprocess the data to obtain multiple standardized input variable time series data, which will serve as training and testing samples for the neural network. Specifically, this includes:

[0050] Step S1.1: Based on professional knowledge, initially select initial input variables that may be related to the remaining life of the target component. Then, calculate the Pearson correlation coefficient between other variables and the initial input variables. Additionally, select variables with correlation coefficients greater than the preset value p and use them together with the initial input variables as input variables, finally obtaining n input variables.

[0051] Step S1.2: Use a threshold of 5 times the standard deviation to remove extreme outliers in SCADA data, and use an active power greater than zero as a threshold to remove shutdown status data;

[0052] Step S1.3: The Z-score normalization method is used to process the time series data of each input variable. The calculation formula is shown below:

[0053]

[0054] Where μ is the mean of the sample data and σ is the standard deviation of the sample data;

[0055] Step S1.4: Use a sliding window with a step size of 1 unit time and a width of t unit time to segment the obtained standardized input variable time series, which can be used as a sample dataset for neural network training and testing.

[0056] Step S2: Perform a Fast Fourier Transform on the time-series data of the multiple standardized input variables obtained in Step S1, and prune the data into time-series data of different lengths according to multiple frequency components. Then, input the time-series data of different lengths into different LSTM neural networks in parallel, extract the feature vectors at different time scales, and concatenate and merge them. Finally, input the result into the linear output layer to obtain the predicted remaining life of the wind turbine components, such as... Figure 2 As shown, it specifically includes:

[0057] Step S2.1: For the time-series signal X obtained after window segmentation with input length t and number of input variables n... i ∈R t×nThe Fast Fourier Transform (FFT) is used to process the data, resulting in n feature spectra. To comprehensively consider the influence of each input variable, the amplitude dimension of all feature spectra is averaged to obtain a new spectra. Due to the conjugate property of the spectra, only the half-spectrum of the new spectra is used for analysis.

[0058]

[0059] In the formula, X i Let A represent a time series of a certain input variable, A represent the corresponding amplitude of all frequency components, FFT(·) represents Fast Fourier Transform, and Amp(·) represents the calculation of amplitude.

[0060] Step S2.2: Select the k largest amplitude values ​​from A obtained in step S2.1, and find their corresponding frequencies and periods. Note that, in order to remove the influence of noise, high-frequency components are generally not considered when selecting frequency components. The calculation formula is as follows:

[0061] [f1,…,f k ] = Fre(Topk(A))

[0062]

[0063] In the formula, Topk(·) represents finding the first k maximum values, Fre(·) represents taking the frequency values, and [f1,...,f k [p1,...,p] represents the first k frequency values. k The ] indicates the period value corresponding to the aforementioned frequency value;

[0064] Step S2.3: Based on the k periods obtained from step S2.2, extract k time segments of length denoted by the period value, starting from the last moment of the input time series and moving backward. l∈{1,…,k};

[0065] Step S2.4: Input the k time segments obtained in step S2.3 into different LSTM neural networks to obtain the final output vectors [h1,…,h] of the k LSTM neural networks. k ];

[0066] Step S2.5: Concatenate the k final output vectors obtained in step S2.4 and input them into the linear output layer to obtain the predicted remaining life of the wind turbine components. The expression is:

[0067] y = W l ·cat([h1,…,h k ])

[0068] In the formula, W ly represents the weights of the linear layer, y represents the final prediction result, and the cat(·) function represents concatenating the vectors column by column.

[0069] Step S3: Quantify the uncertainty of the predicted remaining life of wind turbine components obtained in Step S2 using linear regression and kernel density estimation methods. Specifically, this includes:

[0070] Step S3.1: Process the remaining lifetime prediction value obtained in step S2.5 using a sliding window with a step size of 1 unit time and a width of 1 unit time to obtain the original remaining lifetime prediction value [y1,…,y] within the sliding window. l Each sliding window is treated as a small degradation stage;

[0071] Step S3.2: To estimate the component degradation rate within the window, it is necessary to process the original predicted values ​​of remaining lifetime [y1,…,y] obtained in step S3.1 within the sliding window. l Perform linear regression, such as Figure 3a As shown, the slope of the fitted regression line is the estimated degradation rate of the component predicted by the neural network during this window period, and the ordinate value at the end of the regression line is the estimated remaining lifetime at the last moment of the current window. This process can be expressed by the following formula:

[0072]

[0073] y end =wt end +b

[0074] In the formula, w and b are the slope and intercept of the regression line, respectively. and y represents the mean of the original predicted remaining life and the mean at the corresponding time, respectively. end The ordinate value at the end of the regression line within the window represents the final predicted remaining lifespan, t. end This represents the x-coordinate value at the end of the regression line;

[0075] Step S3.3: Project all the original remaining lifetime predictions obtained in Step 3.2 within the window onto the last moment of the window at the same degradation rate, i.e., along the direction of the regression line. This yields all possible remaining lifetime predictions at the end of this degradation phase, mathematically expressed as:

[0076]

[0077] In the formula, One of the remaining lifetime estimates at the last moment of the window;

[0078] Step S3.4: Reference Figure 3bFor all possible remaining lifetime predictions obtained in step S3.3 within a sliding window, the probability density function of the remaining lifetime prediction distribution is estimated using the kernel density estimation method. For l remaining lifetime estimates... Let its probability density function be f, and the kernel density estimation expression be:

[0079]

[0080] In the formula, K(·) represents the kernel function, usually a Gaussian kernel is chosen, and a represents the bandwidth; after estimating the probability density function of the remaining lifetime prediction distribution, the 95% confidence interval of the distribution can be obtained.

[0081] To verify the effectiveness of the method of this invention, this embodiment uses a SCADA dataset of a 4MW wind turbine from a wind farm in Jiangsu Province, China, to evaluate the proposed method. The fault records in the dataset show that due to a gearbox failure, wind turbines No. 42 and No. 47 underwent maintenance shutdowns at 12:00 on July 8, 2023, and 13:18 on July 8, 2023, respectively. This embodiment selects SCADA data from the two wind turbines from four months before the fault occurred until the fault occurred for training and testing. Detailed dataset information is shown in Table 1. The proposed method, MH-LSTM, is then compared with other methods such as LSTM, BiLSTM, and CNN-LSTM-AM.

[0082] Table 1 Dataset Details

[0083]

[0084] In this embodiment, the gearbox degradation data of wind turbine No. 42 was used as training data to train the multi-scale temporal feature extraction neural network MH-LSTM proposed in this invention. The hyperparameter settings of the neural network are shown in Table 2. During the training process of the neural network, in order to accelerate the convergence rate and reduce the training time, a learning rate decay strategy and an early stopping strategy were adopted.

[0085] Table 2. Neural Network Hyperparameter Settings

[0086]

[0087] Gearbox degradation data from wind turbine No. 47 was used as test data. To verify the superiority of the method of this invention, it was compared with several other neural network models, including LSTM, BiLSTM, and CNN-LSTM-AM. Figures 4a-4dAs shown, compared with other methods, the remaining lifetime prediction obtained by the method of this invention is more accurate at various time points, the probability density function is more concentrated, and it better conforms to the changing patterns. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used as the benchmarks. 2 The interval coverage probability (PICP), average width (PINAW), and comprehensive index (CWC) are used as evaluation indicators for the prediction of the remaining life of the gearbox and the quantification of uncertainty. The calculation results are shown in Table 3. It can be seen that the method of this invention achieves optimal results in all five evaluation indicators: RMSE, MAE, R², PINAW, and CWC. In summary, the remaining life probability prediction method of this invention has stronger identification ability for different stages of the degradation process and exhibits high prediction accuracy and stability throughout the entire lifespan degradation process of wind turbine gearboxes.

[0088] Table 3 Comparison of Evaluation Indicators for Different Methods

[0089]

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life probability of a key component of a large wind turbine, characterized in that, The method comprises the following steps: Step S1: obtaining SCADA data of a wind turbine through a data acquisition and monitoring control system, screening input variables related to the remaining life of a target component of the wind turbine, and pre-processing the data to obtain standardized time series data of the input variables; In step S1, the screening of the input variables related to the remaining life of the target component of the wind turbine comprises: first selecting initial input variables related to the remaining life of the target component according to professional knowledge, then calculating the Pearson correlation coefficients of other variables and the initial input variables, and selecting variables with correlation coefficients greater than a preset value p from the other variables as input variables together with the initial input variables, thereby obtaining n input variables, n being a positive integer; Step S2: performing fast Fourier transform on the standardized time series data of the input variables obtained in step S1, and cutting the frequency components into time series data of different lengths, then inputting the time series data of different lengths into different LSTM neural networks for parallel processing, extracting feature vectors at different time scales and splicing and merging them, and then inputting them into a linear output layer to obtain a wind turbine component remaining life prediction value; Step S3: quantifying the uncertainty of the wind turbine component remaining life prediction value obtained in step S2 based on a linear regression and kernel density estimation method; In step S1, the data preprocessing specifically comprises: First, remove extreme abnormal values of the SCADA data using a threshold of 5 times the standard deviation, and remove shutdown state data by taking active power greater than zero as a threshold; Then, perform data processing on the time series of each input variable using a Z-score standardization method; Finally, segment the standardized input variable time series using a sliding window with a step size of 1 unit of time and a width of t units of time; The step S2 specifically comprises: Step S2.1: processing the time series signal with an input length of t and an input variable number of n after window segmentation using fast Fourier transform to obtain n feature spectrum graphs, n being a positive integer, then averaging the amplitude dimensions of all feature spectrum graphs to obtain a new spectrum graph, and only taking half of the new spectrum for analysis based on the conjugate property of the spectrum: ， wherein , denotes a certain input variable time series, A denotes the corresponding amplitude of the respective frequency component, denotes a fast Fourier transform, denotes the calculation of the amplitude; Step S2.2: selecting the first k maximum amplitudes from A obtained in step S2.1, finding their corresponding frequencies and periods, and the calculation formula is as follows: ， ， In the formula, represents the first k maximum values, represents the frequency values, represents the first k frequency values, represents the first k period values corresponding to the frequency values; Step S2.3: According to the k periods obtained from step S2.2, k time segments with the length of the period value are respectively obtained by cutting forward from the last time of the input time sequence , ; Step S2.4: inputting the k time segments obtained in step S2.3 into different LSTM neural networks respectively to obtain final output vectors of the k LSTM neural networks ; Step S2.5: splicing the k final output vectors obtained in step S2.4 together and inputting them into a linear output layer to obtain a wind turbine component remaining life prediction value, and the expression is: ， In the formula, is the weight of the linear layer, y is the final prediction result, The function represents the concatenation of vectors by column.

2. The wind turbine key component remaining life probability prediction method of claim 1, wherein, In step S2.2, when selecting the first k maximum amplitudes from A, the frequency components do not include high frequency components.

3. The wind turbine key component remaining life probability prediction method of claim 1, wherein, The step S3 specifically comprises: Step S3.1: process the remaining life prediction value obtained in step S2.5 in a sliding window manner with a step of 1 unit time and a width of l unit time to obtain the remaining life original prediction value in the sliding window and each sliding window is regarded as a degradation stage; Step S3.2: residual life raw prediction value within the sliding window A linear regression is performed, and the slope of the resulting regression line is the estimated degradation rate of the component predicted by the neural network during this window, and the y-intercept of the regression line is the residual life estimate at the end of the current window, expressed as: ， ， ， where w and b are the slope and intercept of the regression line, respectively, and are the mean of the original prediction of the remaining life and the mean of the corresponding time, respectively, is the ordinate value of the end of the regression line in the window, i.e. the final prediction of the remaining life, is the abscissa value of the end of the regression line. Step S3.3: projecting all remaining life original prediction values in the window obtained in step 3.2 to the last time of the window at the same degradation rate, i.e. along the direction of the regression line, to obtain l remaining life estimation values at the end of the degradation stage, and the formula expression is: ， In the formula, is one of the remaining lifetime estimates for the last time instant of the window; Step S3.4: Estimate the probability density function of the remaining life prediction distribution for the remaining life prediction values in one sliding window obtained in step S3.3 by kernel density estimation method, for the l remaining life estimation values Let its probability density function be f, and the kernel density estimation expression be: ， wherein represents a kernel function, and a represents a bandwidth; After estimating the probability density function of the remaining life prediction distribution, further obtain the 95% confidence interval of the remaining life prediction distribution.

Images