A wind turbine tower stress monitoring and early warning method and system
By collecting stress data on wind turbine towers and establishing an autoencoder anomaly detection model, the problem of difficulty in real-time monitoring of tower stress changes in existing technologies has been solved, realizing real-time stress monitoring and early warning of wind turbine towers, and improving the accuracy and stability of early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG MINGYANG WIND POWER IND GRP CO LTD
- Filing Date
- 2023-06-19
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to monitor stress changes in wind turbine towers in real time, especially when wind direction and speed change. This makes it difficult to determine the ultimate stress threshold and monitor the deterioration trend of local stress changes, resulting in insufficient early warning of tower failures.
By collecting stress data from various locations on the wind turbine tower, performing data filtering, frame processing, and outlier processing, an autoencoder anomaly detection model is established. A three-layer neural network is used for feature extraction and early warning, and a threshold is set for anomaly judgment, thereby realizing real-time stress monitoring and early warning of tower components.
It enables scientific monitoring of wind turbine tower stress, improves the accuracy and stability of early warning, can promptly detect potential stress concentration hazards, avoid tower failures, and reduce the risk of loss.
Smart Images
Figure CN116677569B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of stress monitoring and early warning for wind turbine towers, and in particular to a method and system for stress monitoring and early warning of wind turbine towers. Background Technology
[0002] The tower of a wind turbine is its key load-bearing component, and its main components include the tower body, foundation, and tower access gates. Offshore wind turbines also include cable openings. Due to long-term operation in harsh natural environments, wind turbines need to constantly adjust their posture in the face of uncertain wind direction and speed, resulting in complex stress conditions. At the same time, wind turbines must also withstand frequent impact loads from gusts. Although the failure rate of the wind turbine tower accounts for only about 1.5% in wind turbine accidents, such failures generally manifest as tower breakage and collapse. The losses caused are not only the cost of replacing the tower but may also lead to secondary damage to transmission components such as blades, generators, and gearboxes, resulting in further losses.
[0003] Existing tower monitoring methods mostly focus on monitoring tower tilt, foundation settlement, and tower vibration. This is primarily achieved through the installation of accelerometers or satellite remote sensing to monitor and issue alarms for parameters such as wind turbine tilt angle, uneven settlement, and vibration acceleration. Stress monitoring of wind turbine towers is less common, but localized stress monitoring is often the best method to reflect the localized state of the tower. Typical tower stress analysis uses finite element analysis models to calculate the tower's natural frequencies, mode shapes, and stress under ultimate loads through static strength and modal analysis. However, this method cannot achieve real-time monitoring during wind turbine operation. Therefore, a method of installing strain gauges is adopted. Sufficient strain gauges are installed at four fixed locations on the tower—the tower itself, the tower base, the tower entrance, and the cable hole. Signals are transmitted to a data acquisition unit via high-performance, interference-resistant cables, and real-time stress values are calculated.
[0004] Currently, there are two main difficulties in using strain gauge-based tower stress monitoring methods: First, the stress at local locations constantly switches between tensile and compressive stresses as wind direction and speed change, making it difficult to identify patterns. Second, the ultimate load that the tower can withstand during design is often much larger than the actual load it experiences. Monitoring methods based on stress thresholds determined by ultimate load values are unable to detect the deterioration trend of tower foundation stress changes. Furthermore, due to the scarcity of tower collapse accidents and the limited monitoring data, the lack of negative sample input during model building makes it difficult to determine the ultimate stress threshold at the time of failure. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose a method and system for stress monitoring and early warning of wind turbine towers. By collecting stress data from various cross sections of the tower, analyzing and modeling the data, and conducting targeted investigations and maintenance based on the model's identification and early warning status, stress concentration can be avoided.
[0006] The objective of this invention is achieved through the following technical solution: a method for monitoring and early warning of stress on wind turbine towers, comprising the following steps:
[0007] S1. Collect stress data at various measuring points on the wind turbine tower;
[0008] S2. Filter the collected stress data;
[0009] S3. Perform frame segmentation on the extracted stress data;
[0010] S4. Perform outlier processing on each frame of stress data, remove abnormal frames of stress data caused by sensor failure, and normalize the feature vector of each frame of stress data.
[0011] S5. Establish an autoencoder anomaly detection model. Input the processed feature vector matrix into the autoencoder anomaly detection model according to each tower component, and train it to obtain the model of each tower component.
[0012] S6. Statistically analyze the prediction residuals of each data point during the training process, calculate the mean and standard deviation, and set the thresholds for each tower component.
[0013] S7. Obtain the stress data to be measured at each measuring point of the wind turbine tower on the same day. Repeat steps S2 to S4. Input the processed stress data into the corresponding tower component model for prediction and obtain the residual error.
[0014] S8. Determine the abnormal situation based on the residual error and the preset threshold, and classify the warning level according to the predicted residual threshold obtained in step S6 and the actual mean value of the predicted residual, and issue a warning.
[0015] Furthermore, step S1 includes the following steps:
[0016] Stress data from various locations on the wind turbine tower were collected over the past year. Each wind turbine monitoring point included 6 stress measuring points distributed along the circumference of the tower, 8 stress measuring points distributed along the circumference of the tower base, stress measuring points at the tower gate, and stress measuring points at the submarine cable hole.
[0017] Furthermore, step S2 includes the following steps:
[0018] The collected stress data is screened to identify fault-free stress data, and its continuity is assessed to extract continuous data.
[0019] Furthermore, step S3 includes the following steps:
[0020] The filtered stress data is processed by framing, with each frame consisting of 5 minutes and a frame shift of 90 seconds. Six features are extracted from the data within each 5-minute frame: zero-crossing rate, volatility, out-of-limit frequency, variance, average volatility amplitude, and maximum volatility amplitude.
[0021] The zero-crossing rate ZCR is the number of times the stress crosses zero within a preset period. The stress has positive and negative values, which represent the number of times the stress at the measuring point switches between tensile and compressive stress, respectively.
[0022] Volatility, also known as the coefficient of variation, is a statistical measure of the dispersion of data within a predetermined period.
[0023] The over-limit frequency is the number of times or data points within a preset period where the absolute value of the stress value exceeds a certain set threshold. Here, the stress threshold is set to 64000kN.
[0024] Variance is the degree of dispersion of stress data within a preset period;
[0025] The maximum fluctuation amplitude is the maximum absolute value of the stress amplitude within the preset period;
[0026] The average fluctuation amplitude is the average value of the stress amplitude within the preset period.
[0027] Furthermore, the feature extraction process of the zero-crossing rate (ZCR) is as follows:
[0028]
[0029] Where i represents the frame number, n is the point number in each frame of data, L is the total number of points in a frame of data (here, it is 5 minutes, so L is 300), and y is the stress amplitude per second in each frame of data.
[0030] The definition of the sgn symbolic function is as follows:
[0031]
[0032] Furthermore, step S5 includes the following steps:
[0033] The processed feature vector matrices of each tower component are input into the autoencoder anomaly detection model. Training is then performed on the tower model, tower base model, tower door model, and submarine cable hole model to construct a three-layer neural network. The first layer of the three-layer neural network has 5 nodes as the encoder, the middle layer has 2 nodes as the hidden layer, and the third layer has 5 nodes as the decoder. The activation function is ELU, and the loss function is MSE. 5% of the dataset is used as the validation set to verify the generalization of the model, and the rest is used as training data. During training, the model is trained at a ratio of 8:2 between the training set and the test set to establish the autoencoder anomaly detection model.
[0034] Furthermore, step S6 includes the following steps:
[0035] During the training process of statistical anomaly detection data, the prediction residual of each data point is calculated and the mean μ and standard deviation σ are obtained. Based on the 3σ criterion, the threshold of each component is set as μ+3σ.
[0036] The residuals from encoding and decoding are calculated as follows:
[0037] L(x,x')=||x-x'|| 2 ;
[0038] Where L is the residual, x is the original feature matrix, and x' is the reconstructed feature matrix after encoding and decoding.
[0039] Furthermore, step S8 includes the following steps:
[0040] Anomalies are determined based on the residual error and a preset threshold. If the residual error is greater than the threshold, the prediction result is abnormal; if the residual error is less than the threshold, the prediction result is normal. First, the amount of stress data acquired on the day is determined. If it meets the preset data amount requirement, the duration of abnormal data is determined. If the duration of abnormal data exceeds 30 minutes, a warning is issued. Based on the threshold set in step S6 and the average residual value when the actual prediction is abnormal, the warning is divided into three levels, representing the magnitude of the potential stress concentration in various parts of the tower.
[0041] A wind turbine tower stress monitoring and early warning system, used to implement the above-mentioned wind turbine tower stress monitoring and early warning method, comprising:
[0042] The data cleaning and preprocessing module is used to filter out fault-free operation data of the wind turbine;
[0043] The feature extraction module is used to extract features from each segment of continuous data;
[0044] The autoencoder modeling module is used to build an autoencoder model for anomaly detection, implementing an anomaly detection method based on the backpropagation and gradient descent algorithms. It uses input data as supervision to guide the neural network to learn a mapping relationship, thereby obtaining a reconstructed output feature.
[0045] Furthermore, the autoencoder model includes an encoder, a hidden layer, and a decoder.
[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0047] 1. The model of this invention extracts the time-scale features of the data, ensuring the continuity of the data in time series. Compared with directly inputting the collected data into the model for training, the processed features contain more comprehensive information, making stress monitoring more scientific.
[0048] 2. The anomaly detection model implemented in this invention is an unsupervised learning model, which does not require fault data input for model training or data labeling, while ensuring the model's stability and strong generalization ability. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the measuring points at various locations on the wind turbine tower.
[0050] Figure 2 Schematic diagram for modeling an autoencoder.
[0051] Figure 3 This is a flowchart of the present invention.
[0052] Figure 4 This is a graph showing the convergence of the MSE loss during the training of the tower model. Detailed Implementation
[0053] The present invention will be further described below with reference to specific embodiments.
[0054] Example 1
[0055] See Figures 1 to 4 As shown, this embodiment provides a method for monitoring and early warning of wind turbine tower stress, including the following steps:
[0056] S1. Collect stress data at various measuring points on the wind turbine tower, including the following steps:
[0057] Stress data from various locations on the wind turbine tower were collected over the past year. Each wind turbine monitoring point included 6 stress measuring points distributed along the circumference of the tower, 8 stress measuring points distributed along the circumference of the tower base, stress measuring points at the tower gate, and stress measuring points at the submarine cable hole.
[0058] S2. Filter the collected stress data, filter out the stress data without faults, and make a continuity judgment to extract the continuous data.
[0059] S3. Perform frame segmentation processing on the extracted stress data, including the following steps:
[0060] The filtered stress data is processed by framing, with each frame consisting of 5 minutes and a frame shift of 90 seconds. Six features are extracted from the data within each 5-minute frame: zero-crossing rate, volatility, out-of-limit frequency, variance, average volatility amplitude, and maximum volatility amplitude.
[0061] The zero-crossing rate (ZCR) is the number of times the stress crosses zero within a preset period. The stress has positive and negative values, representing the number of times the stress at the measuring point switches between tensile and compressive stress, respectively, as shown in the following formula:
[0062]
[0063] Where i represents the frame number, n is the point number in each frame of data, L is the total number of points in a frame of data (here, it is 5 minutes, so L is 300), and y is the stress amplitude per second in each frame of data.
[0064] The definition of the sgn symbolic function is as follows:
[0065]
[0066] Volatility, also known as the coefficient of variation, is a statistical measure of the dispersion of data within a predetermined period.
[0067] The over-limit frequency is the number of times or data points within a preset period where the absolute value of the stress value exceeds a certain set threshold. Here, the stress threshold is set to 64000kN.
[0068] Variance is the degree of dispersion of stress data within a preset period;
[0069] The maximum fluctuation amplitude is the maximum absolute value of the stress amplitude within the preset period;
[0070] The average fluctuation amplitude is the average value of the stress amplitude within the preset period.
[0071] S4. Perform outlier processing on each frame of stress data, remove abnormal frames of stress data caused by sensor failure, and normalize the feature vector of each frame of stress data.
[0072] S5. Establish an autoencoder anomaly detection model. Input the processed feature vector matrix into the autoencoder anomaly detection model for each tower component, and train it to obtain the model for each tower component. This includes the following steps:
[0073] The processed feature vector matrices of each tower component are input into the autoencoder anomaly detection model. Training is then performed on the tower model, tower base model, tower door model, and submarine cable hole model to construct a three-layer neural network. The first layer of the three-layer neural network has 5 nodes as the encoder, the middle layer has 2 nodes as the hidden layer, and the third layer has 5 nodes as the decoder. The activation function is ELU, and the loss function is MSE. 5% of the dataset is used as the validation set to verify the model's generalization ability, and the remainder is used as training data. During training, the model is trained at an 8:2 training-to-test ratio to establish the autoencoder anomaly detection model. The convergence of the MSE loss for the tower components during training is shown below. Figure 4 As shown in Table 1, the training accuracy and loss results for each tower component are as follows:
[0074] Table 1 Training Results
[0075] part Tower Tower base Tower door Submarine cable hole accurancy 91.54％ 85.28％ 88.05％ 91.83％ mse_loss 0.0056 0.0081 0.0168 0.0144
[0076] S6. Statistically analyze the prediction residuals of each data point during the training process, calculate the mean and standard deviation, and set thresholds for each tower component, including the following steps:
[0077] During the training process of statistical anomaly detection data, the prediction residual of each data point is calculated and the mean μ and standard deviation σ are obtained. Based on the 3σ criterion, the threshold of each component is set as μ+3σ.
[0078] The residuals from encoding and decoding are calculated as follows:
[0079] L(x,x')=||x-x'|| 2 ;
[0080] Where L is the residual, x is the original feature matrix, and x' is the reconstructed feature matrix after encoding and decoding.
[0081] S7. Obtain the stress data to be measured at each measuring point of the wind turbine tower on the same day. Repeat steps S2 to S4. Input the processed stress data into the corresponding tower component model for prediction and obtain the residual error.
[0082] S8. Determine the abnormal situation based on the residual error and the preset threshold, and classify the warning level according to the predicted residual threshold obtained in step S6 and the actual mean value of the predicted residual, and issue a warning, including the following steps:
[0083] Anomalies are determined based on the residual error and a preset threshold. If the residual error is greater than the threshold, the prediction result is abnormal; if the residual error is less than the threshold, the prediction result is normal. First, the amount of stress data acquired on the day is determined. If it meets the preset data amount requirement, the duration of abnormal data is then determined. If the duration of abnormal data exceeds 30 minutes, a warning is issued. Based on the threshold set in step S6 and the average residual value when the actual prediction is abnormal, the warning is divided into three levels, representing the magnitude of the potential stress concentration in various parts of the tower. The threshold range and warning level for each tower component are shown in Table 2 below.
[0084] Table 2 Warning Levels
[0085] part Tower Tower base Tower door Submarine cable hole Normal threshold x<0.15 x<0.22 x<0.30 x<0.28 Level 1 warning 0.20>x≥0.15 0.35>x≥0.22 0.35>x≥0.30 0.35>x≥0.28 Level II Warning 0.25>x≥0.20 0.45>x≥0.35 0.45>x≥0.35 0.45>x≥0.35 Level III Warning x≥0.25 x≥0.45 x≥0.45 x≥0.45
[0086] Example 2
[0087] This embodiment provides a wind turbine tower stress monitoring and early warning system to implement the wind turbine tower stress monitoring and early warning method described in Embodiment 1, which includes:
[0088] The data cleaning and preprocessing module is used to filter out fault-free operation data of the wind turbine;
[0089] The feature extraction module is used to extract features from each segment of continuous data;
[0090] The autoencoder modeling module is used to establish an autoencoder model for anomaly detection, implementing an anomaly detection method. Based on the backpropagation algorithm and gradient descent algorithm, it uses input data as supervision to guide the neural network to learn a mapping relationship, thereby obtaining a reconstructed output feature. The autoencoder model includes an encoder, a hidden layer, and a decoder.
[0091] Example 3
[0092] This embodiment discloses a non-transitory computer-readable medium storing instructions that, when executed by a processor, perform the steps of the wind turbine tower stress monitoring and early warning method according to Embodiment 1.
[0093] In this embodiment, the non-transitory computer-readable medium can be a disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), USB flash drive, portable hard drive, etc.
[0094] Example 4
[0095] This embodiment discloses a computing device, including a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the wind turbine tower stress monitoring and early warning method described in Embodiment 1.
[0096] The computing device described in this embodiment may be a desktop computer, laptop computer, smartphone, PDA handheld terminal, tablet computer, programmable logic controller (PLC), or other terminal device with processor function.
[0097] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.
Claims
1. A method for monitoring and early warning of stress on wind turbine towers, characterized in that, Includes the following steps: S1. Collect stress data at various measuring points on the wind turbine tower, including the following steps: Stress data from various locations on the wind turbine tower were collected over the past year. Each wind turbine monitoring point included 6 stress measuring points distributed along the circumference of the tower, 8 stress measuring points distributed along the circumference of the tower base, stress measuring points at the tower gate, and stress measuring points at the submarine cable hole. S2. Filter the collected stress data, including the following steps: The collected stress data is filtered to identify fault-free stress data, and its continuity is assessed to extract continuous data. S3. Perform frame segmentation processing on the extracted stress data, including the following steps: The filtered stress data is processed by framing, with each frame consisting of 5 minutes and a frame shift of 90 seconds. Six features are extracted from the data within each 5-minute frame: zero-crossing rate, volatility, out-of-limit frequency, variance, average volatility amplitude, and maximum volatility amplitude. The zero-crossing rate (ZCR) is the number of times the stress crosses zero within a preset period. The stress can have positive or negative values, representing the number of times the stress at the measuring point switches between tensile and compressive stresses, respectively. The feature extraction process for the ZCR is as follows: ； Where i represents the frame number, n is the point number in each frame of data, L is the total number of points in a frame of data (here, it is 5 minutes, so L is 300), and y is the stress amplitude per second in each frame of data. The definition of the sgn symbolic function is as follows: ； Volatility, also known as the coefficient of variation, is a statistical measure of the dispersion of data within a predetermined period. The over-limit frequency is the number of times or data points within a preset period when the absolute value of the stress value exceeds a certain set threshold. Variance is the degree of dispersion of stress data within a preset period; The maximum fluctuation amplitude is the maximum absolute value of the stress amplitude within the preset period; The average fluctuation amplitude is the average value of the stress amplitude within a preset period. S4. Perform outlier processing on each frame of stress data, remove abnormal frames of stress data caused by sensor failure, and normalize the feature vector of each frame of stress data. S5. Establish an autoencoder anomaly detection model. Input the processed feature vector matrix into the autoencoder anomaly detection model according to each tower component, and train it to obtain the model of each tower component. S6. Statistically analyze the prediction residuals of each data point during the training process, calculate the mean and standard deviation, and set the thresholds for each tower component. S7. Obtain the stress data to be measured at each measuring point of the wind turbine tower on the same day. Repeat steps S2 to S4. Input the processed stress data into the corresponding tower component model for prediction and obtain the residual error. S8. Determine the abnormal situation based on the residual error and the preset threshold, and classify the warning level according to the predicted residual threshold obtained in step S6 and the actual mean value of the predicted residual, and issue a warning.
2. The method for monitoring and early warning of wind turbine tower stress according to claim 1, characterized in that, Step S5 includes the following steps: The processed feature vector matrices of each tower component are input into the autoencoder anomaly detection model. Training is then performed on the tower model, tower base model, tower door model, and submarine cable hole model to construct a three-layer neural network. The first layer of the three-layer neural network has 5 nodes as the encoder, the middle layer has 2 nodes as the hidden layer, and the third layer has 5 nodes as the decoder. The activation function is ELU, and the loss function is MSE. 5% of the dataset is used as the validation set to verify the model's generalization ability, and the rest is used as training data. During training, the model is trained at a ratio of 8:2 between the training set and the test set to establish the autoencoder anomaly detection model.
3. The method for monitoring and early warning of wind turbine tower stress according to claim 1, characterized in that, Step S6 includes the following steps: The prediction residuals of each data point during the training process of statistical anomaly detection data are calculated, and the mean is obtained. and standard deviation ,according to The criteria set the threshold values for each component as follows: ; The residuals from encoding and decoding are calculated as follows: ； Where L is the residual, x is the original feature matrix, and x' is the reconstructed feature matrix after encoding and decoding.
4. The method for monitoring and early warning of wind turbine tower stress according to claim 1, characterized in that, Step S8 includes the following steps: Anomalies are determined based on the residual error and a preset threshold. If the residual error is greater than the threshold, the prediction result is abnormal; if the residual error is less than the threshold, the prediction result is normal. First, the amount of stress data acquired on the day is determined. If it meets the preset data amount requirement, the duration of abnormal data is determined. If the duration of abnormal data exceeds 30 minutes, a warning is issued. Based on the threshold set in step S6 and the average residual value when the actual prediction is abnormal, the warning is divided into three levels, representing the magnitude of the potential stress concentration in various parts of the tower.
5. A wind turbine tower stress monitoring and early warning system, characterized in that, The method for monitoring and early warning of wind turbine tower stress according to any one of claims 1-4 includes: The data cleaning and preprocessing module is used to filter out fault-free operation data of the wind turbine; The feature extraction module is used to extract features from each segment of continuous data; The autoencoder modeling module is used to build an autoencoder model for anomaly detection, implementing an anomaly detection method based on the backpropagation and gradient descent algorithms. It uses input data as supervision to guide the neural network to learn a mapping relationship, thereby obtaining a reconstructed output feature.
6. The wind turbine tower stress monitoring and early warning system according to claim 5, characterized in that: The autoencoder model includes an encoder, a hidden layer, and a decoder.