Fan blade vibration prediction and maintenance method based on digital twinning
By combining fiber optic sensor networks and multi-scale digital twin models with edge computing, the accuracy issues of capturing vibration characteristics and predicting damage in wind turbine blades have been resolved. This has enabled efficient fault identification and economical maintenance strategies, reducing operation and maintenance costs and delay risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANENG SHANGHAI GAS TURBINE POWER GENERATION CO LTD
- Filing Date
- 2026-02-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies struggle to fully capture the complex vibration characteristics of wind turbine blades, resulting in low fault identification accuracy and insufficient accuracy and timeliness in damage prediction. Traditional periodic maintenance methods suffer from high costs and detection delays.
A fiber optic sensor network is used to collect blade vibration signals. The signals are preprocessed by edge computing, and feature extraction and online incremental learning are performed by combining wavelet packet decomposition and multi-scale digital twin models. Damage prediction is then performed using an improved crack propagation formula to generate maintenance decisions.
It achieves dynamic assessment of blade health status, vibration prediction error is less than 9.2%, fault identification accuracy is improved, maintenance costs are reduced by 32%, investment payback period is less than 2.3 years, and annual operation and maintenance costs are reduced by 41%.
Smart Images
Figure CN122174377A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent operation and maintenance technology for wind power equipment, specifically to a method for predicting and maintaining wind turbine blade vibration based on digital twins. Background Technology
[0002] Blade vibration is a phenomenon caused by periodic excitation forces generated by uneven airflow during the operation of wind turbine generators. These excitation forces can be categorized into low-frequency and high-frequency types, caused by different airflow anomalies and blade structural characteristics, respectively. Vibration forms include flapping, flaring, and torsional vibrations of individual blades, as well as coupled vibration modes of multiple blades. When the frequency of the excitation force is close to the natural frequency of the blade, resonance may occur, leading to increased amplitude, significant dynamic stress, and even blade damage.
[0003] The global wind power operation and maintenance market exceeds US$15 billion annually, with downtime losses due to blade failure accounting for 28%. Traditional scheduled maintenance has two major drawbacks: maintenance is performed when the blades are in good condition, resulting in 15% to 25% of wasted costs; and when a sudden crack propagates, the average detection delay when it reaches the critical size is as high as 48 hours, which may lead to serious safety accidents and economic losses.
[0004] In vibration monitoring, existing systems mostly use a single sensor type for data acquisition, making it difficult to comprehensively capture the complex vibration characteristics of blades. Although some systems have introduced multi-sensor fusion, data processing is still based on traditional time-domain or frequency-domain analysis methods, which are insufficient in extracting features from non-stationary vibration signals, resulting in low fault identification accuracy.
[0005] In terms of damage prediction, existing methods are mostly based on empirical formulas or statistical models, failing to fully integrate the physical mechanisms of blades and real-time operational data. Single data-driven models have limited applicability when operating conditions change, while purely physical models struggle to reflect the complex factors involved in actual operation. This results in the accuracy and timeliness of damage prediction failing to meet the needs of predictive maintenance.
[0006] Therefore, how to establish a wind turbine blade vibration prediction and maintenance method that integrates physical mechanisms and data-driven approaches and supports real-time updates, so as to achieve early warning of faults and intelligent optimization of maintenance strategies, is a core problem that has not yet been solved in the current wind power operation and maintenance field. Summary of the Invention
[0007] The present invention aims to solve at least one of the technical problems existing in the prior art, and to provide a method for predicting and maintaining wind turbine blade vibration based on digital twins.
[0008] To achieve the above objectives, this invention provides a method for predicting and maintaining wind turbine blade vibration based on digital twins, comprising: A fiber optic sensor network is deployed along the span of the wind turbine blade to collect the vibration signal of the blade at a preset sampling frequency, and the vibration signal is preprocessed by an edge computing unit. The preprocessed vibration signal is decomposed into multiple layers using wavelet packet decomposition. The energy values of each sub-band are calculated and normalized to generate vibration feature vectors. A multi-scale digital twin model containing macroscopic, mesoscopic, and microscopic layers was constructed, and the model parameters of the model were optimized based on a regularized objective function. The vibration feature vector is input into the digital twin model, and the model parameters are updated in real time based on the residual between the predicted and measured values using an online incremental learning strategy based on a sliding window. Based on an improved crack propagation formula that includes stress ratio correction and temperature correction, the crack propagation trend and remaining service life of the blade are predicted according to the stress analysis output of the digital twin model. The predicted crack length is compared with preset warning thresholds and critical thresholds, and warning information, shutdown and maintenance instructions or replacement suggestions are generated based on the comparison results.
[0009] Furthermore, the fiber optic sensor network includes multiple measurement points distributed along the blade spanwise, and the preset sampling frequency is 2000Hz; The vibration signal includes triaxial acceleration data from the leading edge, trailing edge, and middle of the blade.
[0010] Furthermore, the calculation formula for the wavelet packet decomposition is as follows: ; in, These are the wavelet packet decomposition coefficients. The vibration signal, For wavelet basis functions, The number of decomposition layers, This is the sub-band number.
[0011] Furthermore, the vibration feature vector is generated as follows: Calculate the energy value of each sub-band: ; The energy values of each sub-frequency band are normalized to generate the multidimensional vibration feature vector.
[0012] Furthermore, the regularization objective function is: ; in, For the optimized model parameters, These are actual measured values. These are the model's predicted values. As input features, For parameters to be optimized, The regularization coefficient is . For parameter norm; Furthermore, the multi-scale digital twin model includes a macroscopic layer, a mesoscopic layer, and a microscopic layer; The macroscopic layer is used to simulate the overall dynamic response of the blade; The mesoscopic layer is used to simulate the local structural stress distribution of the blade; The microlayer is used to simulate the material damage evolution process.
[0013] Furthermore, the online incremental learning strategy employs a sliding window mechanism, with the window size set to historical data of a preset time length; When new data arrives, update the model parameters and remove old data outside the window; The drift rate of the model parameters is less than 0.3% per month.
[0014] Furthermore, the improved crack propagation formula is as follows: ; in, This represents the crack propagation rate per cycle. and For material constants, This represents the stress intensity factor amplitude. Stress ratio, For ambient temperature, This is the stress ratio correction function. This is a temperature correction function.
[0015] Furthermore, the method for determining the comparison result is as follows: When the predicted crack length reaches the preset warning threshold, an early warning message is generated; When the predicted crack length reaches the preset critical threshold, a shutdown and maintenance command is generated. Replacement recommendations are generated when the predicted remaining useful life is less than a preset useful life threshold.
[0016] Furthermore, it also includes edge computing processing steps: An edge computing unit is deployed inside the cabin to perform real-time preprocessing and feature extraction of the vibration signals; The edge computing unit adopts a multi-core processor architecture and supports deep learning inference acceleration; The preprocessed data is uploaded to a cloud platform for fusion analysis via industrial communication protocols.
[0017] The beneficial effects of this invention are as follows: This invention achieves dynamic assessment of blade health status by constructing a digital twin model coupled with multi-physics fields and combining it with real-time edge data processing, with a vibration prediction error of less than or equal to 9.2%.
[0018] This invention uses wavelet packet decomposition to process non-stationary vibration signals, which can effectively extract time-frequency domain features and significantly improve the accuracy and sensitivity of fault identification.
[0019] This invention integrates physical mechanisms and data-driven models, solving the problem of limited applicability of single models. Through an online incremental learning strategy, it keeps the model parameter drift rate below 0.3% per month.
[0020] This invention predicts damage evolution based on an improved crack propagation formula, reducing maintenance costs by 32%, with an investment payback period of less than 2.3 years and annual operation and maintenance costs reduced by 410,000 yuan per megawatt. Attached Figure Description
[0021] Figure 1 This is a flowchart of the wind turbine blade vibration prediction and maintenance method based on digital twins according to the present invention; Figure 2 This is a schematic diagram of the module composition of the predictive maintenance system of the present invention; Figure 3 This is a frequency band energy distribution diagram of the wavelet packet decomposition of the vibration signal according to the present invention; Figure 4 This is a comparison curve between the crack propagation prediction and actual measurement of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.
[0023] The wind turbine blade vibration prediction and maintenance method based on digital twins of the present invention is implemented based on a predictive maintenance system.
[0024] See Figure 2 The predictive maintenance system comprises a physical sensing layer, a digital twin layer, a simulation analysis layer, and a decision output layer. The physical sensing layer collects blade vibration data from a fiber optic sensor network and performs real-time preprocessing and feature extraction via edge computing units. The digital twin layer constructs multi-scale blade models, achieving real-time synchronization between the physical blades and the virtual model. The simulation analysis layer performs vibration response simulation and damage evolution prediction based on the digital twin model. The decision output layer generates maintenance decisions based on the prediction results and pushes them to maintenance personnel.
[0025] Example 1 This embodiment uses blade vibration monitoring and predictive maintenance of a 2MW offshore wind turbine as an application scenario. The turbine blade is 45 meters long and made of glass fiber reinforced composite material. A fiber optic sensor network is installed on the blade, with sensors evenly distributed along the blade span, spaced 5 meters apart, for a total of 9 measurement points. The edge computing unit is installed inside the nacelle, using a 32-core processor architecture to support deep learning inference acceleration.
[0026] See Figure 1 The predictive maintenance method in this embodiment includes the following steps: Step S1: Vibration data acquisition.
[0027] A fiber optic sensor network continuously acquires blade vibration signals at a sampling frequency of 2000Hz. Each measurement point collects triaxial acceleration data, corresponding to the blade's flapping, oscillation, and torsional directions. The sensors employ fiber Bragg grating technology, characterized by electromagnetic interference resistance, corrosion resistance, and high sensitivity. The acquired raw data is first preprocessed in an edge computing unit, including noise reduction filtering, signal amplification, and data compression. A Butterworth low-pass filter is used, with a cutoff frequency set to 800Hz. After preprocessing, the data volume is compressed to 20% of the original data and uploaded to the cloud platform via industrial Ethernet protocol.
[0028] Step S2: Vibration feature extraction.
[0029] The cloud platform receives the preprocessed vibration data and performs time-frequency domain analysis using wavelet packet decomposition. The calculation formula for wavelet packet decomposition is as follows: ,in It is a vibration signal. The basis functions are db4 wavelet functions. Number of decomposition levels. It is set to 5 layers, with a corresponding number of 16 sub-bands and a frequency resolution of 62.5Hz.
[0030] The system calculates the energy value of each sub-band. The energy values are then normalized to generate a 16-dimensional vibration feature vector.
[0031] In this embodiment, under normal operating conditions, the main energy of blade vibration is concentrated in the 3rd to 5th sub-bands, corresponding to a frequency range of 125Hz to 312.5Hz; when early cracks appear on the blade, the energy proportion of the 8th to 10th sub-bands increases significantly, which can be used as a characteristic indicator for fault identification.
[0032] Step S3: Digital twin modeling.
[0033] The system constructs a multi-scale digital twin model of the wind turbine blade, developed using the ANSYS Twin Builder software platform. The model comprises three layers: the macroscopic layer uses beam elements to establish the overall dynamic model of the blade, containing 150 nodes, for calculating the blade's natural frequencies and mode shapes; the mesoscopic layer uses shell elements to establish a local structural model with a mesh size of 50 mm, for analyzing the stress distribution in stress concentration areas; and the microscopic layer uses cohesive elements to establish a material damage model for simulating delamination and crack propagation in composite materials.
[0034] Model parameter optimization is based on a regularized objective function: The regularization coefficient Set to 0.01, use the gradient descent algorithm for iterative optimization, and set the convergence threshold to 0.001.
[0035] Step S4: Model synchronization update.
[0036] The digital twin model employs an online incremental learning strategy to achieve real-time synchronization with the physical blade. The system uses a 7-day historical data sliding window for real-time correction of model parameters. When a new vibration feature vector arrives, the system calculates the residual between the model's predicted and measured values. If the residual exceeds a preset threshold of 0.05, a parameter update is triggered. Parameter updates utilize stochastic gradient descent with a learning rate of 0.001. After online learning optimization, the monthly drift rate of the model parameters remains below 0.3%, effectively adapting to the gradual changes in blade performance over service time. Simultaneously, the system automatically removes old data outside the window to prevent historical data from interfering with current predictions.
[0037] Step S5: Damage evolution prediction.
[0038] The system predicts crack propagation based on an improved Paris formula. The improved formula is as follows: ,in and These are the experimental calibration constants for glass fiber composite materials. Set as , Set it to 3.5.
[0039] Stress intensity factor amplitude The stress ratio correction function was calculated based on the stress analysis results from the digital twin model. Temperature correction function ,in Temperature is in Celsius.
[0040] The system performs crack propagation calculations every hour to predict the crack length trend over the next 30 days.
[0041] See Figure 4 The comparison between the predicted curve and the actual measurement data shows that the prediction error is within 10%, which meets the requirements of engineering applications.
[0042] Step S6: Maintain decision generation.
[0043] The system generates maintenance decisions based on damage evolution prediction results. When the predicted crack length reaches a preset warning threshold of 15 mm, the system automatically generates an early warning message, notifying maintenance personnel to increase monitoring frequency. When the predicted crack length reaches a preset critical threshold of 30 mm, the system generates a shutdown and maintenance instruction, requiring on-site inspection to be completed within 72 hours. When the predicted remaining service life is less than a preset service life threshold of 6 months, the system generates blade replacement recommendations and provides the optimal replacement time window. Maintenance decisions are pushed to maintenance personnel via mobile terminal applications and simultaneously synchronized to the wind farm operation and maintenance management system.
[0044] Example 2 This embodiment uses predictive maintenance of the blades of a 3MW wind turbine in an onshore wind farm as an application scenario. The wind farm is located in the Northwest Plateau region at an altitude of 2200 meters, with an average annual wind speed of 8.5 m / s and an annual temperature range of -25°C to 35°C. The wind turbine blades are 56 meters long, made of carbon fiber reinforced composite materials, and each blade weighs 12 tons. A fiber optic sensor network is installed on the blades, with sensors evenly distributed along the blade span, spaced 7 meters apart, for a total of 8 measurement points. The edge computing unit is installed inside the nacelle and equipped with an industrial-grade temperature control system to adapt to the extreme temperature environment.
[0045] The predictive maintenance method in this embodiment is the same as the steps in Embodiment 1, except that the parameter configuration and application environment are different.
[0046] In step S1, since vibration interference on land is relatively less than in the marine environment, but there are still special weather conditions such as sandstorms and lightning, the filter cutoff frequency is adjusted to 600Hz to filter out high-frequency noise. The sampling frequency of the fiber optic sensor remains at 2000Hz, but data transmission adopts a redundant dual-channel design. The main channel transmits through fiber optics, and the backup channel transmits through a wireless network to ensure data reliability under adverse weather conditions. The preprocessing algorithm of the edge computing unit adds a sandstorm interference identification module. When a characteristic sandstorm interference waveform is detected in the sensor signal, the median filtering algorithm is automatically activated for noise reduction. The median filtering window length is set to 15 sampling points.
[0047] In step S2, the number of decomposition layers is determined. The system maintains a 5-layer design with 16 sub-bands. Considering the vibration characteristics of the blades in a 3MW onshore turbine, the natural frequency of the blades is lower than that of the 2MW unit. The focus is on the energy variations in sub-bands 5 to 8, corresponding to a frequency range of 250Hz to 500Hz. This frequency range covers the first and second bending modes of the carbon fiber composite blades, exhibiting high sensitivity to early delamination damage. The vibration characteristic vector is normalized using the minimum-maximum normalization method, with the normalization formula as follows: This ensures the comparability of feature vectors under different operating conditions.
[0048] In step S3, the macroscopic layer of the digital twin model is discretized with 200 nodes to accommodate the geometry of a 56-meter-long blade. The mesh size of the mesoscopic layer is adjusted to 40 mm, with mesh refinement at the blade root transition zone and blade tip region, where the mesh size is 20 mm. The damage model of the microscopic layer is optimized for the properties of carbon fiber composite materials, increasing the simulation capability for both fiber fracture and matrix cracking damage modes. Regularization coefficient. The learning rate is kept at 0.01, but the learning rate of the gradient descent algorithm is adjusted to 0.0008 to improve convergence stability in the high-dimensional parameter space.
[0049] In step S4, the sliding window size is adjusted to 14 days of historical data to accommodate the long wind condition variation cycle of onshore wind farms. The residual threshold for model parameter updates is adjusted to 0.08; parameter updates are triggered when the residual exceeds this threshold. The parameter update frequency is limited to a maximum of three updates per day to avoid model instability caused by frequent updates. A momentum optimization mechanism is introduced during online learning, with the momentum coefficient set to 0.9 to accelerate parameter convergence and reduce oscillations.
[0050] In step S5, due to the large fluctuations in the land environment temperature, the temperature correction function is adjusted to... The temperature correction factor was increased from 0.02 to 0.025 to more accurately reflect the effect of large temperature difference environments on crack propagation rate.
[0051] Material constants of carbon fiber composites Adjusted to m is adjusted to 3.2. The stress ratio correction function remains unchanged. The crack propagation prediction cycle has been adjusted to perform calculations every 2 hours to predict the crack length trend over the next 45 days.
[0052] In step S6, considering the ease of operation and maintenance and traffic conditions of onshore wind farms, the preset warning threshold is adjusted to 20 mm, the preset critical threshold is adjusted to 40 mm, and the preset lifespan threshold is adjusted to 9 months. When an early warning is triggered, maintenance personnel are required to complete on-site inspections within 7 days; when a shutdown maintenance order is triggered, a detailed inspection and maintenance plan must be completed within 96 hours. Maintenance decisions also consider seasonal factors, appropriately relaxing maintenance response time requirements during the low-temperature winter months, but increasing the frequency of remote monitoring.
[0053] Example 3 This embodiment illustrates the verification of the prediction accuracy and the economic benefit analysis of the method of the present invention. The verification experiment was conducted in an offshore wind farm, with an experimental period of 18 months, covering two complete seasonal operating cycles.
[0054] The experiment involved 60 blades from 20 2MW wind turbine units, with each blade equipped with the fiber optic sensor network and predictive maintenance system of this invention.
[0055] To verify the prediction accuracy, a double-blind comparison method was used in the experiment, comparing the crack propagation prediction results of the digital twin model with the results of periodic non-destructive testing. Non-destructive testing employed ultrasonic phased array technology, with a testing cycle of once every 30 days and a testing accuracy of 0.5 mm. A total of 4320 sets of valid prediction data were recorded during the experiment, covering operating conditions under different wind speeds, temperatures, and humidity levels.
[0056] Statistical analysis results show that the average absolute error between the predicted crack length and the actual measured crack length using the method of this invention is 0.82 mm, the relative error is 8.7%, the maximum relative error is 12.3%, and the prediction error is less than or equal to 9.2% within the 90% confidence interval. Compared with the traditional empirical formula prediction method, the prediction accuracy is improved by 47%.
[0057] Regarding the timeliness of fault early warning, a total of 23 early crack damage events were detected during the experiment, of which 19 were detected by the system of this invention, with a success rate of 82.6%. The average early warning time was 17 days, the longest early warning time was 32 days, and the shortest early warning time was 8 days. Of the 19 successful early warning events, 16 were repaired before the cracks extended to the warning threshold, preventing further damage to the blades. Compared with the traditional periodic maintenance mode, the average fault detection time was shortened by 22 days, effectively avoiding secondary damage caused by detection delays.
[0058] Regarding the update performance of the digital twin model, experiments verified the effectiveness of the online incremental learning strategy. In the initial three months of system operation, the average prediction error was 11.2%. After online learning optimization, the prediction error decreased to 9.1% in the fourth to sixth months; and stabilized between 8.5% and 9.0% in the seventh to eighteenth months. Statistical results of the monthly drift rate of model parameters showed an average monthly drift rate of 0.23% and a maximum monthly drift rate of 0.28%, both lower than the design target of 0.3%. The sliding window mechanism effectively avoided interference from historical data on current predictions, and the model's adaptation time to new operating conditions averaged 5 days.
[0059] In terms of economic benefit analysis, a full life-cycle cost model is used to evaluate the return on investment of the predictive maintenance system. The initial investment cost of the system includes the following components: the hardware procurement cost of the fiber optic sensor network is 80,000 yuan per unit, and the installation and commissioning cost is 70,000 yuan per unit, totaling 150,000 yuan per unit; the procurement cost of the edge computing unit is 60,000 yuan per unit, and the installation cost is 20,000 yuan per unit, totaling 80,000 yuan per unit; the cloud platform construction cost includes server procurement of 300,000 yuan, software platform development of 150,000 yuan, and network communication equipment of 50,000 yuan, totaling 500,000 yuan; the digital twin model development cost includes model construction of 500,000 yuan, parameter calibration of 200,000 yuan, and system integration of 100,000 yuan, totaling 800,000 yuan. Taking a wind farm with 20 wind turbines as the calculation object, the total investment in sensors and edge computing units is 4.6 million yuan, plus the platform and model cost of 1.3 million yuan, the total initial investment is approximately 5.9 million yuan.
[0060] The annual operation and maintenance benefits of adopting the method of this invention include the following aspects: Regarding reducing unplanned downtime losses, three serious blade failures were avoided during the experiment due to early warnings. Based on an average downtime of 15 days per failure and a daily power generation loss of 80,000 yuan, the annual reduction in downtime losses is approximately 2.4 million yuan. Regarding reducing ineffective maintenance costs, traditional periodic maintenance requires four blade inspections per wind turbine per year, with a single inspection costing 30,000 yuan. With the condition monitoring method of this invention, the number of inspections is reduced to two per year, with additional inspections only required when the system issues a warning, resulting in annual maintenance cost savings of approximately 1 million yuan. Regarding extending blade lifespan, through early damage intervention, the average blade lifespan is extended from 18 years to 21 years. Based on a replacement cost of 1.2 million yuan per blade and 60 blades in a wind farm, the lifespan cost savings for blade replacement are approximately 12 million yuan, equivalent to an annual saving of approximately 800,000 yuan.
[0061] The system's annual operating costs include: cloud platform maintenance costs of 150,000 yuan, sensor calibration and maintenance costs of 100,000 yuan, and software upgrade costs of 50,000 yuan, totaling 300,000 yuan annually. The comprehensive annual net benefit is calculated as 2.4 million yuan + 1 million yuan + 800,000 yuan - 300,000 yuan, equaling 3.9 million yuan. The investment payback period is 5.9 million yuan divided by 3.9 million yuan, equaling 1.51 years, significantly better than the industry average of 2.3 years. Based on a 20-year operating cycle, after deducting initial investment and annual operating costs, the method of this invention can generate a net benefit of approximately 66.1 million yuan for the wind farm. Based on a total installed capacity of 40MW for the wind farm, the average annual net benefit per megawatt of installed capacity is 97,500 yuan, with an annual maintenance cost reduction rate of 41%, equivalent to an annual saving of approximately 410,000 yuan per megawatt.
[0062] In terms of sensitivity analysis, sensitivity tests were conducted on key parameters. When the failure rate of the fiber optic sensor increased from the design value of 2% to 5%, the annual net benefit decreased by 12%, and the investment payback period was extended to 1.72 years, still demonstrating good economic feasibility. When the wind farm scale increased from 20 to 50 wind turbines, the initial investment cost per unit installed capacity decreased by 35% due to the sharing of platform and model costs, and the investment payback period was shortened to 1.12 years. When the electricity price subsidy was reduced by 20%, the annual net benefit decreased by 18%, and the investment payback period was extended to 1.84 years. The sensitivity analysis results show that the method of the present invention has robust economic benefits under different operating conditions.
[0063] In summary, the embodiments disclosed herein have at least the following technical effects: This invention achieves real-time synchronization and dynamic updating of the physical blades and the virtual model by constructing a digital twin model with multi-physics coupling.
[0064] This invention uses wavelet packet decomposition to process non-stationary vibration signals, effectively extracting fault features.
[0065] This invention integrates physical mechanisms with data-driven models, improving the accuracy and adaptability of predictions. It predicts damage evolution based on an improved crack propagation formula, providing a scientific basis for maintenance decisions.
[0066] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.
Claims
1. A method for predicting and maintaining wind turbine blade vibration based on digital twins, characterized in that, include: A fiber optic sensor network is deployed along the span of the wind turbine blade to collect the vibration signal of the blade at a preset sampling frequency, and the vibration signal is preprocessed by an edge computing unit. The preprocessed vibration signal is decomposed into multiple layers using wavelet packet decomposition. The energy values of each sub-band are calculated and normalized to generate vibration feature vectors. A multi-scale digital twin model containing macroscopic, mesoscopic, and microscopic layers was constructed, and the model parameters of the model were optimized based on a regularized objective function. The vibration feature vector is input into the digital twin model, and the model parameters are updated in real time based on the residual between the predicted and measured values using an online incremental learning strategy based on a sliding window. Based on an improved crack propagation formula that includes stress ratio correction and temperature correction, the crack propagation trend and remaining service life of the blade are predicted according to the stress analysis output of the digital twin model. The predicted crack length is compared with preset warning thresholds and critical thresholds, and warning information, shutdown and maintenance instructions or replacement suggestions are generated based on the comparison results.
2. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The fiber optic sensor network includes multiple measurement points distributed along the blade span, and the preset sampling frequency is 2000Hz; The vibration signal includes triaxial acceleration data from the leading edge, trailing edge, and middle of the blade.
3. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The calculation formula for wavelet packet decomposition is as follows: ； in, These are the wavelet packet decomposition coefficients. The vibration signal, For wavelet basis functions, The number of decomposition layers, This is the sub-band number.
4. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 3, characterized in that, The vibration feature vector is generated as follows: Calculate the energy value of each sub-band: ; The energy values of each sub-frequency band are normalized to generate the multidimensional vibration feature vector.
5. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The regularization objective function is: ； in, The optimized model parameters, These are actual measured values. These are the model's predicted values. As input features, For parameters to be optimized, The regularization coefficient is . Let be the parameter norm.
6. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The multi-scale digital twin model includes a macroscopic layer, a mesoscopic layer, and a microscopic layer; The macroscopic layer is used to simulate the overall dynamic response of the blade; The mesoscopic layer is used to simulate the local structural stress distribution of the blade; The microlayer is used to simulate the material damage evolution process.
7. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The online incremental learning strategy employs a sliding window mechanism, with the window size set to historical data of a preset time length. When new data arrives, update the model parameters and remove old data outside the window; The drift rate of the model parameters is less than 0.3% per month.
8. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The improved crack propagation formula is as follows: ； in, This represents the crack propagation rate per cycle. and For material constants, This represents the stress intensity factor amplitude. Stress ratio, For ambient temperature, This is the stress ratio correction function. This is a temperature correction function.
9. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to claim 1, characterized in that, The method for determining the comparison result is as follows: When the predicted crack length reaches the preset warning threshold, an early warning message is generated; When the predicted crack length reaches the preset critical threshold, a shutdown and maintenance command is generated. Replacement recommendations are generated when the predicted remaining useful life is less than a preset useful life threshold.
10. The method for predicting and maintaining wind turbine blade vibration based on digital twins according to any one of claims 1 to 9, characterized in that, It also includes edge computing processing steps: An edge computing unit is deployed inside the cabin to perform real-time preprocessing and feature extraction of the vibration signals; The edge computing unit adopts a multi-core processor architecture and supports deep learning inference acceleration; The preprocessed data is uploaded to a cloud platform for fusion analysis via industrial communication protocols.