A wind turbine fault early warning method based on SCADA and CMS multi-source data fusion

Abstract

The application discloses a wind turbine fault early warning method based on SCADA and CMS multi-source data fusion, belongs to the technical field of wind turbine state monitoring and fault diagnosis, and comprises the following steps: collecting and preprocessing SCADA operation data and CMS vibration data, extracting multi-dimensional features, screening and optimizing a feature set by adopting a method combining a Pearson correlation coefficient and mutual information, and realizing multi-source feature fusion by an adaptive weighting algorithm. An LSTM-Attention model is constructed for training, real-time fusion features are input into the model to obtain a health state and a fault risk probability, three-level adaptive early warning thresholds are set based on kernel density estimation, an audible and visual early warning is generated when the thresholds are exceeded, and a fault is located by a feature contribution degree. The application can realize early fault early warning of key components of a wind turbine, discover potential faults in advance, improve early warning accuracy, and effectively improve the reliability and safety of wind turbine operation.

Description

Technical Field

[0001] This invention relates to the field of wind turbine condition monitoring and fault diagnosis technology, and in particular to a wind turbine fault early warning method based on the fusion of multi-source data from SCADA and CMS. Background Technology

[0002] With the rapid development of the global wind power industry, the installed capacity and number of wind turbines in operation are constantly increasing. Wind turbines are typically installed in harsh environments such as remote mountainous areas and offshore locations, where they are constantly exposed to strong winds, dust storms, and salt spray. Key components such as gearboxes, generators, and main bearings are prone to failure. Wind turbine failures not only lead to downtime losses but can also trigger safety accidents, causing significant economic losses.

[0003] Currently, early warning systems for wind turbine faults primarily rely on single data sources: one is operational data based on SCADA (Supervisory Control and Data Acquisition) systems, and the other is vibration monitoring data based on CMS (Condition Monitoring System). While SCADA-based early warning methods offer low data acquisition costs and wide coverage, the sampling frequency of SCADA data is typically low (usually once per minute), making it difficult to capture subtle early fault characteristics and resulting in warning lag. CMS-based early warning methods, although with high sampling frequencies and capable of capturing minute changes in vibration signals, suffer from high installation costs, typically requiring sensors only to be installed on some critical components. Furthermore, vibration signals are easily affected by changes in operating conditions, leading to a high false alarm rate.

[0004] Furthermore, most existing fault early warning methods employ traditional machine learning models, which struggle to effectively handle the temporal and nonlinear characteristics of wind turbine operating data, resulting in insufficient ability to identify complex faults. Simultaneously, the warning thresholds are typically fixed values, failing to adapt to the dynamic changes in wind turbine operating conditions, further reducing the accuracy of the warnings. Summary of the Invention

[0005] The purpose of this invention is to propose a wind turbine fault early warning method based on the fusion of multi-source data from SCADA and CMS. By fusing complementary information from the two types of data and combining deep learning technology, the method can achieve early and accurate fault warning and location of key components of wind turbines.

[0006] To achieve the above objectives, this invention provides a wind turbine fault early warning method based on the fusion of multi-source data from SCADA and CMS, comprising the following steps: Step S1: Collect the SCADA system operation data and CMS system vibration monitoring data of the wind turbine generator respectively, and preprocess the collected multi-source heterogeneous data to obtain a standardized time series dataset; Step S2: Extract the first feature set from the preprocessed SCADA data and extract the second feature set from the preprocessed CMS data; Step S3: Based on the method of combining Pearson correlation coefficient and mutual information, calculate the correlation degree between each feature in the first feature set and the second feature set and the health status of key components of the wind turbine. Select features according to the correlation degree to obtain the optimized feature set. Step S4: Use an adaptive weighted fusion algorithm to fuse multi-source features in the optimized feature set to generate a fused feature vector; Step S5: Construct a fault warning model based on the attention mechanism of Long Short-Term Memory Network (LSTM-Attention), and conduct supervised training of the model using historical normal operation data and fault data; Step S6: Input the fused feature vector collected and processed in real time into the trained fault early warning model, and output the health status assessment value and fault risk probability of key components of the wind turbine. Step S7: Adaptively set multi-level early warning thresholds based on the kernel density estimation KDE method. When the probability of fault risk exceeds the corresponding threshold, generate an audible and visual early warning signal of the corresponding level and locate the fault.

[0007] Preferably, in step S1, the preprocessing specifically includes: Step S11: Use linear interpolation to process missing values ​​in the data, and use the 3σ criterion combined with the isolated forest algorithm to identify and remove outliers; Step S12: Timestamp alignment of SCADA data and CMS data is performed, and sliding time window technology is used to unify data with different sampling frequencies to the same time granularity; Step S13: Normalize all data using the Z-score standardization method to unify the dimensions of each feature. The calculation formula is as follows: ; in, For standardized data, This is the original data. This is the mean of this feature in historical normal data. This represents the standard deviation of this feature in historical normal data.

[0008] Preferably, in step S2: The first feature set includes the time-domain statistical features and operating condition features of SCADA data, specifically wind speed, active power, generator speed, gearbox oil temperature, generator stator temperature, and the mean, variance, maximum, minimum, and rate of change of pitch angle. The second feature set includes time-domain features, frequency-domain features, and time-frequency-domain features of CMS data. The time-domain features are mean, root mean square, kurtosis, skewness, waveform factor, and impulse factor. The frequency-domain features are the characteristic frequencies and their amplitudes obtained by Fast Fourier Transform (FFT). The time-frequency-domain features are the energy proportions of each frequency band obtained by wavelet packet decomposition.

[0009] Preferably, in step S3, the correlation degree calculation method is as follows: Calculate the first Features and Health Status Tags Pearson correlation coefficient and mutual information ; The weighted average method is used to obtain the overall correlation degree. The calculation formula is: ; in, This is the weighting coefficient, with a value ranging from 0.4 to 0.6. It is the maximum value of the mutual information of all features; Sort by comprehensive relevance from highest to lowest, and select the top... The optimized feature set is composed of several features. The value ranges from 15 to 50.

[0010] Preferably, in step S4, the adaptive weighted fusion algorithm is specifically as follows: The optimized feature set is divided into SCADA feature subsets. and CMS feature subset ; Calculate the classification accuracy of two feature subsets under different operating conditions. and ; The fusion weights of the two subsets are dynamically adjusted based on the classification accuracy, and the calculation formula is as follows: ; ; in, The weights of the SCADA feature subset, The weights of the CMS feature subsets, and ; Generate fused feature vectors .

[0011] Preferably, in step S5, the method for constructing the LSTM-Attention fault warning model is as follows: Input layer: receiving dimension is The fused feature vectors, with a time step of [value missing]. The input sequence dimension is ; LSTM layer: Contains two LSTM network layers, each with 32 to 128 hidden units, used to extract time series features of the data; Attention layer: Calculates the attention weights of the output at each time step, highlighting the feature contributions of important time steps, and the sum of the attention weights of all time steps is 1; Weighted fusion layer: The LSTM outputs at all time steps are weighted and summed according to the attention weights to obtain the global feature vector; Fully connected layer: maps global feature vectors to fault risk probabilities using the Sigmoid activation function; Output layer: Outputs the health status assessment values ​​and failure risk probabilities of key components of the wind turbine.

[0012] Preferably, in step S7, the method for setting the multi-level early warning threshold is as follows: Collect the residual sequence of model predictions and actual values ​​under normal operating conditions of wind turbine units. ; The probability density function is obtained by estimating the kernel density of the residual sequence using a Gaussian kernel function. ; Three levels of early warning thresholds are set based on cumulative probability density: Level 1 warning threshold The corresponding cumulative probability density is 0.95; Level II warning threshold The corresponding cumulative probability density is 0.99; Level 3 warning threshold The corresponding cumulative probability density is 0.999.

[0013] Preferably, in step S7, the fault location method specifically is as follows: When an early warning signal is generated, the contribution of each feature in the optimized feature set is calculated; Based on the contribution of each feature, determine the top 3 features with the largest contribution; Based on the pre-established feature-component fault association mapping relationship, the key component that failed is located and the fault type is identified.

[0014] Preferably, it also includes step S8: online model update, the specific steps of which include: Step S81: Collect newly added normal operation data and fault data confirmed on-site according to preset time intervals; Step S82: Fine-tune and update the LSTM-Attention model using an incremental learning method, while retaining the knowledge already learned by the model; Step S83: Simultaneously update the weight coefficients and the warning threshold obtained from the kernel density estimation in the adaptive weighted fusion algorithm.

[0015] Preferably, the key components of the wind turbine include the gearbox, generator, main shaft, main bearing, and pitch system.

[0016] Therefore, the wind turbine fault early warning method based on SCADA and CMS multi-source data fusion described above has the following advantages: (1) Multi-source data fusion complements each other: It integrates the global operation data of the SCADA system and the high-precision vibration data of the CMS system, which not only takes advantage of the wide coverage and low cost of SCADA data, but also leverages the sensitivity of CMS data to early faults, effectively improving the accuracy and timeliness of fault warning. (2) Intelligent feature selection and fusion: The Pearson correlation coefficient and mutual information are combined to select features, which can capture both linear and nonlinear correlations between features and health status and eliminate redundant features; an adaptive weighted fusion algorithm is adopted to dynamically adjust the fusion weights according to the classification accuracy of the two types of data under different working conditions, thereby improving the representational ability of the fused features. (3) High-precision time series feature extraction: A bidirectional LSTM network model based on attention mechanism was constructed, which can effectively extract bidirectional time series features of wind turbine operation data. The attention mechanism can automatically focus on the time step that contributes more to fault early warning, further improving the model's ability to identify early faults. (4) Adaptive multi-level early warning: Based on the kernel density estimation method, three-level early warning thresholds are adaptively set, which can adapt to the dynamic changes in the operating conditions of wind turbine units and effectively reduce the false alarm rate and missed alarm rate; accurate fault location is achieved by calculating feature contribution, providing clear maintenance directions for operation and maintenance personnel; (5) Online model update: The incremental learning method is used to update the model online, which can continuously learn new operating data and fault modes, so that the model always maintains the best early warning performance and extends the life of the model.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a wind turbine fault early warning method based on SCADA and CMS multi-source data fusion in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] Example like Figure 1 As shown in the figure, this embodiment proposes a wind turbine fault early warning method based on the fusion of multi-source data from SCADA and CMS. The specific implementation steps are as follows: Step 1: Data Acquisition and Preprocessing Data Acquisition: Operational data from the SCADA system and vibration monitoring data from the CMS system of a 1.5MW doubly-fed induction generator (DFIG) wind turbine at a wind farm were collected. The SCADA data sampling frequency was 1 minute / time, and the collected parameters included wind speed, active power, generator speed, gearbox oil temperature, generator stator temperature, and pitch angle. The CMS data sampling frequency was 20kHz, and acceleration sensors were installed at the high-speed shaft, low-speed shaft of the gearbox, and both ends of the generator to collect vibration signals.

[0022] Missing value handling: Linear interpolation is used to handle missing values ​​in the data. For cases where no more than 3 consecutive data points are missing, linear interpolation between the two consecutive valid data points is used to fill the missing values; for cases where more than 3 consecutive data points are missing, the missing data segments are marked as invalid and removed.

[0023] Outlier handling: Outliers are identified and removed using the 3σ criterion combined with the Isolation Forest algorithm. First, the 3σ criterion is used to initially identify obvious outliers. Then, the Isolation Forest algorithm is used to further identify hidden outliers. Finally, the identified outliers are removed.

[0024] Timestamp alignment and resampling: SCADA and CMS data are timestamped. A sliding time window technique is used to resample CMS data to a 1-minute time granularity, maintaining consistency with the SCADA data. The sliding window size is set to 1 minute, and statistical characteristics are calculated for the CMS data within each window.

[0025] Standardization: The Z-score standardization method is used to normalize all data so that the mean of each feature is 0 and the standard deviation is 1.

[0026] Step 2: Feature Extraction The first feature set was extracted from the preprocessed SCADA data, including: wind speed, active power, generator speed, gearbox oil temperature, generator stator temperature, mean, variance, maximum, minimum and rate of change of pitch angle, a total of 30 features.

[0027] The second feature set is extracted from the preprocessed CMS data, including: Time-domain features: mean, root mean square, kurtosis, skewness, waveform factor, and impulse factor, a total of 6 features; Frequency domain characteristics: The gearbox meshing frequencies, generator rotation frequencies and their harmonic amplitudes are obtained through Fast Fourier Transform (FFT), totaling 12 characteristics; Time-frequency domain features: The vibration signal is decomposed into 3 levels using db4 wavelet packets to obtain 8 frequency bands. The energy proportion of each frequency band is calculated, resulting in a total of 8 features; the second feature set contains a total of 26 features.

[0028] Step 3: Feature Selection Collect historical data of the wind turbine over the past year, including normal operation data and 12 gearbox failure data, and set the health status label to 0 (normal) and 1 (failure).

[0029] Calculate the Pearson correlation coefficient between each feature in the first and second feature sets and the health status label. and mutual information .

[0030] Set weighting coefficients The weighted average method was used to calculate the overall correlation degree. : .

[0031] Sort the features by their overall correlation from highest to lowest, and select the top 30 features to form an optimized feature set.

[0032] Step 4: Multi-source feature fusion: The optimized feature set is divided into SCADA feature subsets. (18 features) and CMS feature subset (12 features).

[0033] Historical data was divided into training and test sets in a 7:3 ratio. Random forest classifiers were trained using SCADA and CMS feature subsets, respectively. The classification accuracy on the test set was then calculated. and .

[0034] Calculate the fusion weights based on classification accuracy: , In this embodiment , ,therefore , .

[0035] Generate fused feature vectors .

[0036] Step 5: Model Building and Training Constructing an LSTM-Attention fault warning model: Input layer: The input sequence dimension is 10×30 (time step). Feature Dimension ); LSTM layer: Contains two bidirectional LSTM networks, each with 64 hidden units. Dropout is used to prevent overfitting, and the dropout rate is set to 0.2. Attention layer: Calculates the attention weights of the output at each time step, and the sum of the attention weights of all time steps is 1; Weighted fusion layer: The LSTM outputs at all time steps are weighted and summed according to the attention weights to obtain a global feature vector with a dimension of 128; Fully connected layer: Contains two fully connected layers. The first fully connected layer has 64 neurons and uses the ReLU activation function; the second fully connected layer has 1 neuron and uses the Sigmoid activation function. Output layer: Outputs health status assessment values ​​and failure risk probabilities.

[0037] Model training: The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and 100 training epochs. Early stopping was employed to prevent overfitting; training was stopped when the validation set loss did not decrease for 10 consecutive epochs.

[0038] Step 6: Real-time fault warning: Real-time acquisition of SCADA and CMS data from wind turbines, followed by preprocessing, feature extraction, feature selection, and feature fusion according to steps 1-4, to generate a real-time fused feature vector.

[0039] The real-time fused feature vector is input into the trained LSTM-Attention model, which outputs the current health status assessment value and failure risk probability.

[0040] Step 7: Multi-level early warning and fault location: The model output residual sequence of the wind turbine under normal operating conditions for one month was collected, and the kernel density was estimated using the Gaussian kernel function to obtain the probability density function.

[0041] Three levels of early warning thresholds are set based on cumulative probability density: (Corresponding to a cumulative probability density of 0.95) (Corresponding to a cumulative probability density of 0.99) (Corresponding to a cumulative probability density of 0.999).

[0042] When the probability of failure exceeds When the time is reached, a Level 1 warning signal (yellow) is generated to alert maintenance personnel; when the time exceeds [a certain threshold], a Level 1 warning signal (yellow) is generated. When the time is reached, a level-two warning signal (orange) is generated, prompting the scheduling of maintenance; when the time exceeds the warning level... At that time, a Level 3 warning signal (red) will be generated, prompting an immediate shutdown for inspection.

[0043] When an early warning signal is generated, the absolute value of the partial derivative of each feature in the optimized feature set with respect to the probability of failure risk is calculated as the feature contribution. The top three features are determined according to their contribution. In this embodiment, when an early failure occurs in the high-speed shaft bearing of the gearbox, the top three features with the largest contribution are the kurtosis of the CMS high-speed shaft vibration signal, the gearbox oil temperature change rate, and the root mean square of the CMS high-speed shaft vibration signal.

[0044] Based on the pre-established feature-component fault association mapping relationship, the faulty component is located as the high-speed shaft bearing of the gearbox, and the fault type is wear.

[0045] Step 8: Online Model Update New normal operation data and field-confirmed fault data are collected at weekly intervals.

[0046] The LSTM-Attention model was fine-tuned and updated using an incremental learning method, with the learning rate set to 0.0001 and the number of training epochs set to 10.

[0047] Simultaneously update the weight coefficients and early warning thresholds obtained from kernel density estimation in the adaptive weighted fusion algorithm.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for early warning of wind turbine faults based on multi-source data fusion of SCADA and CMS, characterized in that, Includes the following steps: Step S1: Collect the SCADA system operation data and CMS system vibration monitoring data of the wind turbine generator respectively, and preprocess the collected multi-source heterogeneous data to obtain a standardized time series dataset; Step S2: Extract the first feature set from the preprocessed SCADA data and extract the second feature set from the preprocessed CMS data; Step S3: Based on the method of combining Pearson correlation coefficient and mutual information, calculate the correlation degree between each feature in the first feature set and the second feature set and the health status of key components of the wind turbine. Select features according to the correlation degree to obtain the optimized feature set. Step S4: Use an adaptive weighted fusion algorithm to fuse multi-source features in the optimized feature set to generate a fused feature vector; Step S5: Construct a fault warning model based on the attention mechanism of Long Short-Term Memory Network (LSTM-Attention), and conduct supervised training of the model using historical normal operation data and fault data; Step S6: Input the fused feature vector collected and processed in real time into the trained fault early warning model, and output the health status assessment value and fault risk probability of key components of the wind turbine. Step S7: Adaptively set multi-level early warning thresholds based on the kernel density estimation KDE method. When the probability of fault risk exceeds the corresponding threshold, generate an audible and visual early warning signal of the corresponding level and locate the fault.

2. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S1, the preprocessing specifically includes: Step S11: Use linear interpolation to process missing values ​​in the data, and use the 3σ criterion combined with the isolated forest algorithm to identify and remove outliers; Step S12: Timestamp alignment of SCADA data and CMS data is performed, and sliding time window technology is used to unify data with different sampling frequencies to the same time granularity; Step S13: Normalize all data using the Z-score standardization method to unify the dimensions of each feature. The calculation formula is as follows: ； in, For standardized data, The original data, This is the mean of this feature in historical normal data. This represents the standard deviation of this feature in historical normal data.

3. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S2: The first feature set includes the time-domain statistical features and operating condition features of SCADA data, specifically wind speed, active power, generator speed, gearbox oil temperature, generator stator temperature, and the mean, variance, maximum, minimum, and rate of change of pitch angle. The second feature set includes time-domain features, frequency-domain features, and time-frequency-domain features of CMS data. The time-domain features are mean, root mean square, kurtosis, skewness, waveform factor, and impulse factor. The frequency-domain features are the characteristic frequencies and their amplitudes obtained by Fast Fourier Transform (FFT). The time-frequency domain characteristics are the energy proportions of each frequency band obtained through wavelet packet decomposition.

4. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S3, the specific method for calculating the correlation degree is as follows: Calculate the first Features and Health Status Tags Pearson correlation coefficient and mutual information ; The weighted average method is used to obtain the overall correlation degree. The calculation formula is: ； in, This is the weighting coefficient, with a value ranging from 0.4 to 0.

6. The maximum value of the mutual information of all features; Sort by comprehensive relevance from highest to lowest, and select the top... The optimized feature set is composed of several features. The value ranges from 15 to 50.

5. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S4, the adaptive weighted fusion algorithm is as follows: The optimized feature set is divided into SCADA feature subsets. and CMS feature subset ; Calculate the classification accuracy of two feature subsets under different operating conditions. and ; The fusion weights of the two subsets are dynamically adjusted based on the classification accuracy, and the calculation formula is as follows: ； ； in, The weights of the SCADA feature subset, The weights of the CMS feature subsets, and ; Generate fused feature vectors .

6. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S5, the specific method for constructing the LSTM-Attention fault warning model is as follows: Input layer: receiving dimension is The fused feature vectors, with a time step of [value missing]. The input sequence dimension is ; LSTM layer: Contains two LSTM network layers, each with 32 to 128 hidden units, used to extract time series features of the data; Attention layer: Calculates the attention weights of the output at each time step, highlighting the feature contributions of important time steps, and the sum of the attention weights of all time steps is 1; Weighted fusion layer: The LSTM outputs at all time steps are weighted and summed according to the attention weights to obtain the global feature vector; Fully connected layer: maps global feature vectors to fault risk probabilities using the Sigmoid activation function; Output layer: Outputs the health status assessment values ​​and failure risk probabilities of key components of the wind turbine.

7. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S7, the method for setting multi-level early warning thresholds is as follows: Collect the residual sequence of model predictions and actual values ​​under normal operating conditions of wind turbine units. ; The probability density function is obtained by estimating the kernel density of the residual sequence using a Gaussian kernel function. ; Three levels of early warning thresholds are set based on cumulative probability density: Level 1 warning threshold The corresponding cumulative probability density is 0.95; Level II warning threshold The corresponding cumulative probability density is 0.99; Level 3 warning threshold The corresponding cumulative probability density is 0.

999.

8. The wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: In step S7, the fault location method is as follows: When an early warning signal is generated, the contribution of each feature in the optimized feature set is calculated; Based on the contribution of each feature, determine the top 3 features with the largest contribution; Based on the pre-established feature-component fault association mapping relationship, the key component that failed is located and the fault type is identified.

9. A wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: It also includes step S8: online model update, the specific steps of which include: Step S81: Collect newly added normal operation data and fault data confirmed on-site according to preset time intervals; Step S82: Fine-tune and update the LSTM-Attention model using an incremental learning method, while retaining the knowledge already learned by the model; Step S83: Simultaneously update the weight coefficients and the warning threshold obtained from the kernel density estimation in the adaptive weighted fusion algorithm.

10. A wind turbine fault early warning method based on SCADA and CMS multi-source data fusion according to claim 1, characterized in that: Key components of a wind turbine include the gearbox, generator, main shaft, main bearing, and pitch system.

Images