Wind turbine generator blade inner cavity fault detection method based on artificial intelligence

Abstract

The invention belongs to the field of wind power operation and maintenance, and mainly relates to a wind turbine generator blade inner cavity fault detection method based on artificial intelligence, and the method comprises the steps: collecting multi-modal data through a sensor and an unmanned plane; the method comprises the following steps: constructing a digital twin body for synchronous simulation, and establishing a self-adaptive abnormal threshold diagnosis mechanism based on an operation condition through dynamic comparison of simulation data and actually measured data so as to realize early recognition and accurate positioning of internal structure damage; moreover, simulation deduction of future working conditions is carried out in the digital twinborn body, the fault evolution trend and the residual life are predicted, and the operation and maintenance mode is improved from passive response to active early warning; an evaluation system including multi-dimensional indexes such as a fault detection rate and a false alarm rate is established, continuous incremental learning and closed-loop optimization are performed on a deep learning model in combination with operation and maintenance feedback data, and a complete intelligent operation and maintenance solution integrating perception, diagnosis, prediction and self-optimization is formed.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of wind power operation and maintenance, and mainly relates to a wind turbine blade inner cavity fault detection method based on artificial intelligence. BACKGROUND

[0002] The wind turbine blade is the core component for capturing wind energy, and its structural health state is directly related to the safe operation and power generation efficiency of the whole machine. In particular, the inner cavity structure of the blade, such as the honeycomb core material, shear beam and adhesive layer, is subjected to complex alternating loads, wind and sand erosion and temperature and humidity changes for a long time, and is prone to hidden faults, such as adhesive layer peeling, honeycomb core shedding and shear beam micro-cracks.

[0003] In a wind turbine blade inner cavity fault detection method, device, equipment and medium with Chinese application number CN202510315357.7, the invention discloses a wind turbine blade inner cavity fault detection method, device, equipment and medium, which is applied to the field of intelligent detection and includes: obtaining blade features corresponding to a blade inner cavity image to be detected, and taking the blade features as shallow features; extracting deep features of the blade features through a reverse bottleneck module; the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, and the heterogeneous convolution module is a module for feature extraction using a large convolution kernel and multiple small convolution kernels in parallel; using a surface fusion module and a high-level fusion module, the shallow features and the deep features are fused to obtain fusion features; the obtained target detection features are identified to determine the target fault category. Compared with the current manual detection, the invention retains more depth and comprehensive information based on the surface fusion module and the high-level fusion module, thereby improving the accuracy and efficiency of fault detection.

[0004] The above-mentioned patent has given a relatively perfect wind turbine blade inner cavity fault detection method, but there are still some problems in the existing technology in the wind turbine blade inner cavity fault detection. The detection dimension of the existing technology is relatively single, mainly relying on visual image data for surface defect recognition, and unable to effectively perceive the hidden structural damage inside the blade, leading to potential risks difficult to discover in time; secondly, the existing technology lacks the ability of forward-looking prediction of the equipment state, so that the operation and maintenance work is in a passive response mode for a long time, and can only issue an alarm after the damage develops to the naked eye or is clearly captured by the camera, and cannot realize early warning and evolution trend prediction of the fault, resulting in missing the best maintenance opportunity; in addition, the intelligent degree of the existing scheme is limited, and the diagnosis model is usually statically set, lacking self-learning and self-adaptive optimization ability; it is difficult to adapt to the performance evolution caused by different wind farm specific operating environment and unit aging.

[0005] In order to solve the above problems, the present application provides a wind turbine blade cavity fault detection method based on artificial intelligence, which comprises the following steps: arranging multiple sensors at key positions of the wind turbine, combining unmanned aerial vehicle visual collection, and establishing a multi-dimensional perception system covering the internal and external states of the blade; then, based on the physical properties and real-time data, a digital twin is driven for synchronous simulation, and through dynamic comparison of simulation data and measured data, an adaptive abnormal threshold diagnosis mechanism based on operating conditions is established to realize early identification and accurate positioning of internal structural damage; further, simulation deduction of future conditions is carried out in the digital twin to predict the fault evolution trend and remaining life, and the operation and maintenance mode is improved from passive response to active warning; through the establishment of an evaluation system including fault detection rate, false alarm rate and other multi-dimensional indicators, and combined with operation and maintenance feedback data, the deep learning model is continuously incrementally learned and closed-loop optimized, so that the system has the ability to continuously evolve from historical experience and new fault modes, thereby adapting to changes brought by different wind field environments and equipment aging, forming a complete intelligent operation and maintenance solution integrating perception, diagnosis, prediction and self-optimization. SUMMARY

[0006] The present application provides a wind turbine blade cavity fault detection method based on artificial intelligence, which aims to solve the problem of single detection dimension in traditional technology, which leads to missed detection of internal damage, passive operation and maintenance, lack of early warning ability, and static model cannot adapt.

[0007] In order to solve the above problems, the present application uses the following technology to realize:

[0008] A wind turbine blade cavity fault detection method based on artificial intelligence:

[0009] Step S1: arranging multiple types of sensors at key positions of the wind turbine, continuously collecting its operating state data; regularly using unmanned aerial vehicles equipped with high-definition camera modules to collect video image data of the wind turbine; uploading the obtained data to the cloud platform through the 5G module;

[0010] Step S2: the cloud platform pre-processes the received data; based on the processed data, a digital twin of the wind power equipment is driven to be built; the real-time video and image data of the unmanned aerial vehicle are fused to verify and dynamically correct the geometry and physical properties of the digital twin;

[0011] Step S3: using deep learning algorithm, the state information reflected by the digital twin is deeply mined to identify the fault in the wind turbine; simulation deduction is carried out on the digital twin to predict the evolution trend of the fault and the potential operation risk of the equipment;

[0012] Step S4: Compare the prediction results of the digital twin with the actual operating status of the wind turbine and the subsequent maintenance records to quantitatively evaluate the accuracy and reliability of the prediction model; based on the evaluation results, optimize and update the parameters of the deep learning algorithm and the risk prediction model.

[0013] As a preferred embodiment, the multiple types of sensors arranged at the key parts of the wind turbine specifically include:

[0014] The sensors are installed at the key parts of the wind turbine, which refers to the areas that have a decisive role in the safe operation, power generation efficiency, and structural integrity of the wind turbine, and have a high probability of failure, serious consequences, and are difficult to access during daily inspection.

[0015] The specific types of installed sensors include acoustic sensors, strain sensors, and temperature sensors, as well as other optional sensors.

[0016] The installation position of the sensor is recorded during installation, and a unique ID name is assigned to the sensor based on the installation position; the sensor performs periodic data collection at a low preset frequency.

[0017] A simple control system is set up in the wind turbine to make a simple judgment on the data collected by the sensor, set the threshold value corresponding to the sensor, and when the data collected by the sensor exceeds the threshold value, increase the collection frequency of the sensor and send an alarm through the 5G communication module.

[0018] As a preferred embodiment, the periodic use of unmanned aerial vehicles equipped with high-definition camera modules to collect video image data of the wind turbine specifically includes:

[0019] External and internal image collection of the blades is performed using unmanned aerial vehicle equipment, and a corresponding video image data is collected once before the wind turbine is started after installation;

[0020] During the initial deployment of the system, the wind turbine is periodically shut down to collect video image data; after the stable operation of the system, the collection mode will be changed from periodic inspection to on-demand triggering mode, and only when the system predicts potential risks based on sensor data, will the unmanned aerial vehicle be started for targeted image collection;

[0021] The unmanned aerial vehicle device video image acquisition process of the wind turbine is specifically that an image recognition model is deployed in the unmanned aerial vehicle device, the image recognition model trains a model by recording a manual operation unmanned aerial vehicle acquisition workflow and the unmanned aerial vehicle acquisition workflow as a training set, the model controls the unmanned aerial vehicle to autonomously locate a starting point after detecting the wind turbine, records a video image data of an internal structure of the wind turbine according to a learned video acquisition path, and enters a cavity of the wind turbine from an inspection hole of the wind turbine.

[0022] As a preferred implementation, the cloud platform specifically includes the following in the pre-processing of the received data:

[0023] For the time series data collected by the sensor, a Kalman filter algorithm is used to eliminate noise signals and retain effective components reflecting the true state of the structure.

[0024] For the visual data collected by the unmanned aerial vehicle, corresponding image processing operations are performed to improve the image quality and facilitate subsequent feature recognition; and a uniform timestamp is stamped on all data.

[0025] For the sensor data, a corresponding time series data table is constructed according to the sensor ID name, the obtained data is filled according to the sensor ID, and a variety of position data time series change tables are formed accordingly.

[0026] As a preferred implementation, the digital twin of the wind power equipment is specifically driven by the processed data.

[0027] The cloud platform interfaces with the wind turbine master control system through the OPC UA protocol to obtain operating parameters.

[0028] In the initial stage of operation, the basic information data of the wind turbine is pre-stored and the local installation drawing is entered, and the cloud platform uses the target wind turbine basic information data to use a professional 3D modeling engine 3ds MAX to build an initial wind turbine model.

[0029] The finite element analysis solver is integrated on the digital twin platform to simulate the stress distribution, strain, deformation, vibration mode of the blade under wind load, and the surrounding aerodynamic performance.

[0030] The model is used to simulate the actual wind turbine operation according to the obtained wind turbine operating parameter data, mainly to simulate the rotation of the blade to drive the digital twin to run synchronously.

[0031] Simulate the operation of the wind turbine on the digital twin, generate simulation data of each installation point sensor; in the early stage of deployment, by collecting sensor data of the wind turbine in the fault-free state, establish the digital benchmark of normal operation; compare the simulation data with the actual benchmark data in real time, take the actual sensor data of the current wind turbine as the standard, use the optimization algorithm to fine-tune the physical properties of the digital twin, until the simulation data and the measured data are highly consistent within the allowable error range, thereby constructing a high-fidelity digital twin consistent with the behavior of the physical wind turbine.

[0032] As a preferred embodiment, the real-time video and image data of the fusion unmanned aerial vehicle are used to verify and dynamically correct the geometry and physical properties of the digital twin, which specifically includes:

[0033] Based on the analysis of the video image data obtained periodically at the beginning, the appearance characteristics of the wind turbine are extracted to correct the digital twin, and the starting point trained by the image recognition model in the unmanned aerial vehicle and the real-time flight parameters of the unmanned aerial vehicle are used to determine the spatial mapping range of the current video image in the digital twin, and the influence of the background is removed in the predetermined template.

[0034] The preprocessed multi-angle and multi-time sequence images are input into the structure from motion algorithm SFM, and the attitude parameters of the camera are automatically calculated according to the movement and matching of the feature points in the images, and a dense three-dimensional point cloud is generated.

[0035] According to the spatial mapping range of the video image to the digital model, the corresponding relationship between the point cloud data and the model is determined, and the difference between the point cloud data and the model is analyzed to generate a deviation color spectrum; according to the calculated deviation, the digital twin is deformed to make its geometric shape accurately fit the real point cloud generated by SFM.

[0036] As a preferred embodiment, the deep learning algorithm is used to deeply mine the state information reflected by the digital twin and identify the faults in the wind turbine, which specifically includes:

[0037] A wind turbine state analysis model is used, and a deep learning algorithm convolutional neural network model is used as a basic model to analyze the state information of the digital twin;

[0038] In the operation monitoring stage, the simulation data of the digital twin and the actual sensor collected data are continuously compared, and the relationship between the real-time operation parameters of the wind turbine and the sensor data is focused on, and the dynamic abnormal threshold is extracted from which the reasonable deviation value between the simulation data and the sensor is identified;

[0039] When the difference between the simulation data and the actual data from the sensor exceeds the real-time dynamic anomaly threshold, it is determined that the current wind turbine has an abnormal state that the digital twin cannot accurately reflect.

[0040] The system automatically generates alarm information and pushes it to maintenance personnel, suggesting the execution of wind turbine shutdown procedures. At the same time, it initiates drone aerial photography missions to obtain on-site video image data, and corrects the digital twin based on the obtained video image data.

[0041] As a preferred embodiment, the dynamic anomaly threshold specifically includes:

[0042] The core idea of ​​dynamic anomaly thresholds is to adaptively adjust the judgment criteria based on the current operating conditions. For each sensor monitoring point, the residual sequence is calculated using the following formula:

[0043] ;

[0044] in, For actual sensor data, For digital twin simulation data;

[0045] Record operating parameters: speed ,power Wind speed Combined to form the running parameter vector The K-means algorithm is used to cluster historical operating data to determine the optimal number of operating condition categories K, and for each operating condition category... Establish a feature space; during operation, compare the current operating parameters with the corresponding parameters in the feature space to determine the current operating condition category, using the following formula:

[0046] ;

[0047] Where K represents the existing working condition feature space. For the runtime parameter vector, The feature center vector of the kth working condition category;

[0048] The calculation formula is:

[0049] ;

[0050] in, In working conditions Below, the average difference of sensor i, For the residual of sensor i, For the Kth working condition category, is the mathematical expectation operator, representing the calculation of the average value;

[0051] The calculation in a specific working condition The typical offset between the measured value of the sensor and the simulation value is calculated as follows:

[0052]

[0053] Wherein, The residual standard deviation of sensor i in the working condition is the variance operator;

[0054] The final dynamic anomaly threshold setting formula is:

[0055]

[0056] Wherein, The sensitivity coefficient (initial value is 2.0), The threshold value is an interval range.

[0057] As a preferred embodiment, the predicted evolution trend of the fault and the potential operation risk of the equipment specifically include:

[0058] The operation of the wind turbine is predicted based on the digital twin simulation. The long-term wind turbine operation parameters are learned by using the LSTM long short-term memory network model. The historical operation parameters and weather forecast data are used as joint input for training. The non-linear relationship between external wind conditions and internal operation response is deeply mined from multi-modal data through the model, so that the model can simulate the future wind turbine operation parameter data according to the local weather forecast;

[0059] An independent virtual computing space isolated from the real-time operating system is opened on the cloud platform, and the digital twin is copied to the virtual space. In the virtual environment, a simulation time axis independent of the real time is set. The predicted operation parameter data is input into the digital twin in the virtual space as a boundary condition, and the finite element analysis in the digital twin is started. According to the input predicted wind conditions, the morphological changes of the digital twin and the simulation data of each sensor under the current conditions are calculated and simulated;

[0060] The problem of various morphological changes of the wind turbine is analyzed. When there is obvious damage to the elements of the wind turbine in the simulation operation, a risk alarm is immediately generated;

[0061] ​​​For the simulated sensor simulation data, the wind turbine state analysis model for digital twin state analysis is used, the wind turbine operation parameter input simulated by the LSTM long short-term memory network model is input, the theoretical sensor data is simulated, and the sensor simulation data and the theoretical sensor data are compared, and the dynamic abnormal threshold is used for analysis, and the abnormality is identified; when there is data anomaly, the corresponding risk warning is generated;

[0062] The risk warning is automatically classified, and the operation and maintenance decision corresponding thereto is generated and sent to the operation and maintenance personnel.

[0063] As a preferred embodiment, the parameter optimization and iterative updating of the deep learning algorithm and the risk prediction model according to the evaluation result specifically include:

[0064] The simulation simulation data of the digital twin is compared with the actual data at the same time, and the standardized work order is combined; based on the data alignment result, the following core performance indicators are quantitatively calculated: fault detection rate, false alarm rate, missed alarm rate, prediction lead time and positioning accuracy;

[0065] In the case of insufficient prediction accuracy, the model is optimized;

[0066] The verified new fault mode, its feature data and the corresponding repair scheme are updated to the database;

[0067] When the deviation between the model prediction result and the actual data is within the allowable range and the running state is stable, high-quality prediction data is selected to form an enhanced training set, and the model is quickly adapted to the running characteristics of a specific station through incremental training.

[0068] The beneficial effects of the present application are:

[0069] 1. The early discovery and accurate positioning of internal faults of wind turbine blades are realized, a multi-dimensional perception system covering the inside and outside of the blade is constructed by arranging multiple sensors at key positions and combining unmanned aerial vehicle vision, and through the dynamic comparison of simulation data and measured data, the internal hidden structural faults can be accurately identified, so that early discovery and accurate positioning are realized.

[0070] 2. It has the ability of fault evolution trend prediction and active warning. The present application can simulate and deduce future working conditions in the digital twin, predict the development trend of faults, and change the operation and maintenance work from passive response to active warning.

[0071] 3. An intelligent system with continuous self-learning and adaptive optimization capabilities is formed. By establishing an evaluation system containing multiple dimensions such as fault detection rate and false alarm rate, and combining with actual operation and maintenance feedback data, the deep learning model used is continuously incrementally learned and closed-loop optimized, which can continuously evolve from historical experience and new fault patterns, and adapt to changes brought by different wind farm environments and equipment aging. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a method flowchart of the present application;

[0073] Figure 2 is an effect comparison chart of the present application compared with the traditional method. DETAILED DESCRIPTION

[0074] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application will be further described below in combination with specific embodiments, but the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application. In the following examples, the experimental methods are conventional methods, and the materials and reagents used in the following examples are commercially available unless otherwise specified.

[0075] Example 1 as Figure 1 The method flowchart of the present application is shown. This embodiment provides a wind turbine blade inner cavity fault detection method based on artificial intelligence, which specifically includes the following steps:

[0076] Step S1: multiple types of sensors are arranged at key parts of the wind turbine to continuously collect operation state data; a drone equipped with a high-definition camera module is used to collect video image data of the wind turbine at regular intervals; and the obtained data is uploaded to a cloud platform through a 5G module;

[0077] Specifically, sensors are installed at key parts of the wind turbine, which refer to components or areas that have a decisive role in the safe operation, power generation efficiency and structural integrity of the wind turbine, and have a high probability of failure, serious consequences of failure and difficulty in daily inspection, such as honeycomb core area, shear beam connection part, adhesive layer contact surface, etc. The specific types of sensors installed include acoustic sensors, strain sensors and temperature sensors, and optionally humidity sensors, air pressure sensors, etc. The installation position of the sensor is recorded when the sensor is installed, and a unique ID name is assigned to the sensor in the current wind turbine based on the installation position; the sensor performs periodic data collection at a low preset frequency such as 1 Hz, establishes a baseline of the normal operating state of the equipment, and reduces power consumption; at the same time, a simple control system is set up in the wind turbine to make a simple judgment on the data collected by the sensor, set the threshold value corresponding to the sensor, and the threshold value is set as the normal range of the data collected by the sensor (the maximum data and minimum data collected by the sensor obtained through local testing), when the data collected by the sensor is detected to exceed the threshold value, the collection frequency of the sensor is increased, and an alarm is sent through the 5G communication module at the same time;

[0078] In addition, a UAV device is used to collect external and internal image data of the blade. This collection behavior is performed once before the wind turbine is installed and started, and the corresponding video image data is collected. Subsequently, at the initial stage of system deployment, the wind turbine is periodically stopped to collect video image data. After the system is stably deployed and runs, the collection mode will be changed from periodic inspection to on-demand triggering mode, i.e. only when the system predicts potential risks based on sensor data, the UAV will be started to collect images accordingly. The process of video image collection of the UAV device on the wind turbine is as follows: an image recognition model is deployed in the UAV device, the image recognition model is trained by recording the manual operation of the UAV collection workflow and the UAV collection workflow as a training set, so that the model can control the UAV to autonomously locate the starting point after detecting the wind turbine, record the external video of the blade according to the learned video collection path, and record the internal structure video image data from the access hole of the wind turbine equipment.

[0079] After data collection is completed, the collected sensor data, sensor adjustment behavior and video image data are uploaded to the cloud platform through the 5G communication module.

[0080] Step S2: the cloud platform pre-processes the received data; based on the processed data, a digital twin of the wind power equipment is driven to be built; the real-time video and image data of the UAV are fused to verify and dynamically correct the geometric and physical properties of the twin;

[0081] Specifically, after receiving the uploaded data, the cloud platform adopts Kalman filtering algorithm for the time series data collected by the sensors to eliminate noise signals caused by electromagnetic interference, vibration impact, etc. and retain the effective components reflecting the true state of the structure; for the visual data collected by the unmanned aerial vehicle, image enhancement, distortion correction and defogging are performed to improve the image quality and facilitate subsequent feature recognition; at the same time, a uniform timestamp is stamped for all data; for the sensor data, a corresponding time series data table is constructed according to the sensor ID name, and the obtained data is filled according to the sensor ID, thereby forming a data time series change table of multiple positions;

[0082] In addition, the cloud platform interfaces with the wind turbine main control system through the OPC UA protocol to obtain operating parameters such as rotational speed, power and power generation.

[0083] In the initial stage of operation, the basic information data of the specific model, material, etc. of the wind turbine is pre-stored and the local installation drawing is recorded, and the initial wind turbine model is constructed using the professional 3D modeling engine 3ds MAX in the cloud platform using the basic information data of the target wind turbine. Different material elements are divided and labeled by color, and corresponding material properties such as density, Poisson's ratio, shear strength, peel strength, creep properties, etc. are assigned to different elements according to the material; the corresponding sensor installation positions are labeled on the constructed wind turbine model according to the initial recorded sensor installation positions; the finite element analysis solver is integrated on the digital twin platform to simulate the stress distribution, strain, deformation, vibration mode of the blade under wind load and the surrounding aerodynamic performance; the actual wind turbine operation is simulated using the model according to the obtained wind turbine operating parameter data, mainly simulating the rotation of the blade to drive the digital twin to run synchronously;

[0084] Subsequently, the operation of the wind turbine is simulated on the digital twin to generate simulation data of each installed point sensor; in the initial deployment stage, the sensor data of the wind turbine under fault-free operation state is collected to establish a digital benchmark for normal operation; the simulation data is compared with the actual benchmark data in real time, and the actual sensor data of the current wind turbine is used as the standard. The parameter inversion and other optimization algorithms are used to finely adjust the physical properties (such as material parameters and boundary conditions) of the digital twin, until the simulation data and the measured data are highly consistent within the allowable error range, thereby constructing a high-fidelity digital twin consistent with the behavior of the physical wind turbine;

[0085] Based on the analysis of the video image data obtained periodically at the initial stage, the appearance characteristics of the wind turbine generator are extracted to correct the digital twin. Specifically, the starting point trained by the image recognition model in the unmanned aerial vehicle and the real-time flight parameters of the unmanned aerial vehicle are used to determine the spatial mapping range of the current video image in the digital twin, and the influence of the background is removed in the predetermined template; the preprocessed multi-angle and multi-time images are input into the structure from motion (SFM) algorithm, SFM automatically calculates the camera pose parameters according to the movement and matching of the feature points in the image, and generates a dense three-dimensional point cloud, which accurately reflects the geometric shape of the blade in the real world, including the slight deformation caused by manufacturing, installation or long-term operation; according to the spatial mapping range of the digital model by the video image, the corresponding relationship between the point cloud data and the model is determined, and the difference analysis is performed on the point cloud data and the model according to the corresponding relationship, and a deviation color spectrum is generated to show which areas have depressions, bulges or twists; according to the calculated deviation, the digital twin is deformed to accurately fit the real point cloud generated by SFM, and the possible correction content includes: blade twist angle, airfoil profile, front edge corrosion depth, rear edge glue line protrusion, crack and other geometric features; when there are features that are harmful to the wind turbine generator, such as cracks and abnormal twisting, an alarm is generated and sent to the operation and maintenance personnel.

[0086] After the digital twin is constructed and its data can be fitted with the actual situation, the video acquisition work of the unmanned aerial vehicle enters the on-demand triggering mode.

[0087] Step S3: using a deep learning algorithm, the state information reflected by the digital twin is deeply mined to identify the faults in the wind turbine generator; simulation and deduction are performed on the digital twin to predict the evolution trend of the fault and the potential operation risk of the equipment;

[0088] Specifically, a wind turbine generator state analysis model is used, and a convolutional neural network model based on deep learning algorithm is used as a basic model to analyze the state information of the digital twin. In the operation monitoring stage, the simulation data of the digital twin and the actual sensor data are continuously compared, and the relationship between the real-time operation parameters (such as speed and power output) of the wind turbine generator and the sensor data is focused on. When the actual operation parameters are running in the digital twin, the relationship between the simulation data and the actual sensor data is learned, the dynamic abnormal threshold is extracted, the reasonable deviation value between the simulation data and the sensor is identified, and the dynamic abnormal threshold is extracted based on the current operation parameters of the wind turbine generator. Its actual significance is that the speed and power output of the wind turbine generator vary greatly at different times and in different weather conditions, and the dynamic abnormal threshold based on different operation parameters can better identify the current abnormal situation;

[0089] When the difference between the simulation data and the actual data of the sensor exceeds the real-time dynamic anomaly threshold, it is determined that the current wind turbine has an abnormal state that the digital twin cannot accurately reflect. Such anomalies usually represent unmodeled physical damage to the blade structure, such as sudden failures caused by flying stone impact, twisting, cracking, etc. At this time, the system will automatically generate an alarm information push to the operation and maintenance personnel, and suggest to execute the wind turbine shutdown program, and start the unmanned aerial vehicle aerial task to obtain the on-site video image data. According to the obtained video image data, the step of modifying the digital model in step two is repeated, so that the model reflects the true situation;

[0090] Further, based on the simulation running of the digital twin, the long-term wind turbine running parameters are learned by using the LSTM long short-term memory network model, and the historical running parameters and the weather forecast data, especially the wind speed data, are used as joint input for training. The wind speed data and the running parameters are aligned, and the complex nonlinear relationship between the external wind conditions and the internal running response (such as power, main shaft speed, key point strain) is deeply mined from the multi-modal data through the model, so that the model can simulate the running parameter data of the wind turbine in the future period according to the local weather forecast;

[0091] Accordingly, an independent virtual computing space isolated from the real-time operating system is opened on the cloud platform, and the digital twin is copied to the virtual space. In the virtual environment, a simulation time axis independent of the real time is set, the predicted wind speed, wind direction and other running parameter data are input into the digital twin in the virtual space as boundary conditions, and the finite element analysis in the digital twin is started. According to the input predicted wind conditions, the digital twin morphological changes such as blade dynamic load and stress distribution, key position strain and deformation, structure vibration response and fatigue damage accumulation, and each sensor simulation data under the current condition are calculated and simulated;

[0092] First, analyze the various morphological changes of the wind turbine. When there is obvious damage to the blade or other elements in the simulation running, a risk alarm is immediately generated;

[0093] For the simulated sensor simulation data, the wind turbine state analysis model mentioned above for digital twin state analysis is used. The LSTM long short-term memory network model simulates the wind turbine operating parameters. The wind turbine state analysis model does not analyze the simulated physical performance, but only generates theoretical sensor data according to the relationship between the current operating parameters learned from long-term operating parameters and the sensor data under normal conditions. Further comparison of the sensor simulation data and the theoretical sensor data is made, and the dynamic abnormal threshold is used for analysis to identify abnormalities. This process is used to identify wind turbine damage that is not obvious, which is not directly considered by the model to affect the operation of the wind turbine, but has caused a big problem in actual operation. When there is data anomaly, the corresponding risk warning is generated;

[0094] The risk warning is automatically classified, and the corresponding operation and maintenance decision is generated, including triggering the unmanned aerial vehicle inspection task to obtain visual evidence, or issuing a power reduction operation instruction to avoid risks. All decisions are pushed to the operation and maintenance personnel in the form of executable task work orders to ensure that the warning information is closed loop processed.

[0095] Step S4: Compare the prediction results of the digital twin with the actual operating state of the wind turbine and the subsequent maintenance records to quantitatively evaluate the accuracy and reliability of the prediction model; according to the evaluation results, the deep learning algorithm and the risk prediction model are parameter optimized and iteratively updated;

[0096] Specifically, the simulation data of the digital twin is compared with the actual data at the same time, and when the operation and maintenance personnel perform power adjustment, blade maintenance or switching, etc., a standardized work order is automatically generated to record the operation and maintenance content and results; based on the data alignment result, the following core performance indicators are quantitatively calculated: fault detection rate, false alarm rate, missed alarm rate, prediction lead time and positioning accuracy.

[0097] For the case of insufficient prediction accuracy, the main factor in prediction is that the dynamic threshold is too sensitive, which may be that the relationship of the wind turbine normal operation is not clearly learned in the training stage. The reason may be that the environment of the deployment site is complex, and the wind turbine operating data changes greatly, resulting in errors in model learning. The same site model migration or retraining can be used to lengthen the training period;

[0098] The verified new fault mode, its feature data, and the corresponding maintenance scheme are updated to the database to expand the cognitive range;

[0099] When the model prediction result deviates from the actual data within the allowable range and the running state is stable, high-quality prediction data is screened to form an enhanced training set, and the model is quickly adapted to the running characteristics of a specific station through incremental training, so as to accelerate the construction of an expert model with rich local experience.

[0100] As Figure 2 The effect comparison chart of the present application compared with the traditional method is shown in the figure, the black column is the effect of the present application, and the gray column is the effect of the traditional application, by adopting the present application scheme, the fault detection rate and operation and maintenance response efficiency are greatly improved, and the alarm accuracy is good.

[0101] Embodiment 2: The dynamic abnormal threshold mentioned in embodiment 1, the specific content is: the core idea of the dynamic abnormal threshold is to adaptively adjust the judgment standard according to the current running condition, for each sensor monitoring point, calculate the residual sequence, the calculation formula is:

[0102] ;

[0103] Among them, is the actual sensor data, is the digital twin simulation data;

[0104] Record the running parameters: rotating speed , power , wind speed , combine to form the running parameter vector , use K-means algorithm to cluster the historical running data, determine the optimal working condition category number K, for each working condition category Establish a feature space; during operation, compare the current running parameters with the corresponding parameters of the feature space to confirm the current working condition category, the formula is:

[0105] ;

[0106] Among them, K is the existing working condition feature space, is the running parameter vector, the feature center vector of the kth working condition category;

[0107] The calculation formula of is:

[0108] ;

[0109] Among them, is the average difference of sensor i under the working condition , is the residual of sensor i, is the Kth working condition category, is the mathematical expectation operator, denoting the average value;

[0110] The calculation of the specific working condition The typical offset between the measured value of the sensor and the simulation value is calculated as follows:

[0111] ;

[0112] Wherein, is the residual standard deviation of sensor i under the working condition is the variance operator;

[0113] The final dynamic anomaly threshold setting formula is:

[0114] ;

[0115] Wherein, is the sensitivity coefficient (initial value is set to 2.0), denotes that the threshold is an interval range.

[0116] Example 3: In the mountainous wind farm with complex terrain and variable climate, considering the particularity of local wind condition disorder caused by complex terrain, inconvenient transportation and poor communication conditions, the implementation method is adjusted as follows:

[0117] Optimized deployment of sensor network:

[0118] On the basis of the sensor described in example 1, three-dimensional ultrasonic anemometer is added to monitor special wind conditions such as turbulence and shear wind caused by complex terrain;

[0119] Increase the sensor deployment density in the specific area where the blade is easily affected by terrain (such as downwind blade), and focus on the dynamic load change caused by turbulence;

[0120] Adaptive improvement of communication transmission:

[0121] Adopt edge-cloud collaborative processing architecture, deploy edge computing nodes on the wind farm side, realize local preprocessing and caching of sensor data, real-time anomaly detection based on lightweight model, and autonomous running capability (automatic adjustment of operating power of wind turbine) when communication is interrupted;

[0122] Targeted enhancement of digital twin modeling:

[0123] Integrate terrain wind field model, integrate computational fluid dynamics (CFD) model in digital twin, simulate wind field distribution under complex terrain; Establish the mapping relationship between local terrain and wind turbine load to accurately restore the unique operating environment of each site;

[0124] ​And consider the influence of large temperature difference, icing and other mountain unique environmental factors in the material property model, add ultraviolet aging correction coefficient in the model according to the characteristics of strong ultraviolet in high altitude area;

[0125] Intelligent diagnosis and operation and maintenance strategy optimization:

[0126] Warning optimization based on terrain features, establish the correlation knowledge base of terrain-wind condition-fault, identify the typical fault mode under specific terrain conditions; For mountain valley wind, slope flow and other special wind conditions, individualize the dynamic abnormal threshold of different positions;

[0127] Operation and maintenance decision adaptability improvement, combined with mountain traffic conditions, establish an intelligent operation and maintenance scheduling system based on weather forecast and road conditions; When it is predicted that severe weather may cause traffic interruption, trigger preventive inspection in advance and appropriately increase the warning level.

[0128] Model continuous learning mechanism enhancement:

[0129] Adopt the hybrid architecture of federated learning + incremental learning: each unit in the wind field shares knowledge through federated learning; Single unit quickly adapts to the location characteristics through incremental learning; Establish seasonal pattern learning mechanism, automatically identify and adapt to the change of environmental characteristics in different seasons.

[0130] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, the above examples and descriptions in the specification are only preferred examples of the present application, and are not intended to limit the present application, various changes and improvements can be made without departing from the spirit and scope of the present application, these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based wind turbine blade cavity fault detection method, characterized by: Step S1: Multiple types of sensors are arranged at key positions of the wind turbine to continuously collect operating state data; a drone equipped with a high-definition camera module is used to collect video image data of the wind turbine at regular intervals; the obtained data is uploaded to a cloud platform through a 5G module; Step S2: The cloud platform pre-processes the received data; based on the processed data, a digital twin of the wind power equipment is constructed; the real-time video and image data of the drone are fused to verify and dynamically correct the geometric and physical properties of the twin; Step S3: Using deep learning algorithms, the state information reflected by the digital twin is deeply mined to identify faults in the wind turbine; Simulation and deduction are performed on the digital twin to predict the evolution trend of the fault and the potential operation risk of the equipment; Step S4: The prediction results of the digital twin are compared with the actual operating state of the wind turbine and subsequent maintenance records to quantitatively evaluate the accuracy and reliability of the prediction model; based on the evaluation results, the deep learning algorithm and the risk prediction model are parameterized and iteratively updated.

2. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The multiple types of sensors arranged at key positions of the wind turbine specifically include: Sensors are installed at key positions of the wind turbine, and the key positions refer to areas that have a decisive role in the safe operation, power generation efficiency, and structural integrity of the wind turbine, and are difficult to access for daily inspection due to high probability of failure, serious consequences of failure, and other factors; The specific types of installed sensors include acoustic sensors, strain sensors, temperature sensors, and other optional sensors; The installation position of the sensor is recorded, and a unique ID name is assigned to the sensor in the current wind turbine based on the installation position; the sensor performs periodic data collection at a low preset frequency; A simple control system is installed in the wind turbine to make a simple judgment on the data collected by the sensor, set the threshold value corresponding to the sensor, and when the data collected by the sensor exceeds the threshold value, increase the collection frequency of the sensor and send an alarm through the 5G communication module.

3. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The drone equipped with a high-definition camera module collects video image data of the wind turbine specifically includes: External and internal image collection of the blade is performed using a drone device; once after installation of the wind turbine, start the corresponding video image data collection; During the initial deployment of the system, the wind turbine is periodically shut down for video image data collection; after the system is stably deployed and operated, the collection mode will change from periodic inspection to on-demand triggering mode, and the drone will only be started for targeted image collection when the system predicts potential risks based on sensor data; The unmanned aerial vehicle device collects video images of the wind turbine, specifically, an image recognition model is deployed in the unmanned aerial vehicle device, the image recognition model records the manual operation of the unmanned aerial vehicle collection workflow and the unmanned aerial vehicle collection workflow as a training set, the model is trained to control the unmanned aerial vehicle to autonomously locate the starting point after detecting the wind turbine, to record video images of the outside of the blade according to the learned video collection path, and to record video image data of the internal structure from the inspection hole of the wind turbine device into the internal cavity.

4. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The cloud platform specifically includes the following steps for pre-processing the received data: For time series data collected by sensors, Kalman filtering algorithm is used to remove noise signals and retain effective components reflecting the true state of the structure; For visual data collected by the unmanned aerial vehicle, corresponding image processing operations are performed to improve image quality and facilitate subsequent feature recognition; at the same time, a uniform timestamp is stamped on all data; For sensor data, a corresponding time series data table is constructed according to the sensor ID name, and the obtained data is filled according to the sensor ID, thereby forming a data time series change table of multiple positions.

5. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: Based on the processed data, the digital twin of the wind power equipment is driven to be built, specifically including the following steps: The cloud platform connects the wind turbine main control system through the OPC UA protocol to obtain operating parameters; In the initial stage of operation, the basic information data of the wind turbine is pre-stored and the local installation drawing is recorded, and the cloud platform uses the target wind turbine basic information data to build an initial wind turbine model using a professional 3D modeling engine 3ds MAX; Integrate the finite element analysis solver on the digital twin platform to simulate the stress distribution, strain, deformation, vibration mode of the blade under wind load and the surrounding aerodynamic performance; According to the obtained wind turbine operating parameter data, the model simulates the actual operation of the wind turbine, mainly simulating the rotation of the blade to drive the digital twin to run synchronously; Simulate the operation of the wind turbine on the digital twin to generate simulation data for each installation point sensor; in the initial deployment stage, the sensor data of the wind turbine in the fault-free operation state is collected to establish a digital benchmark for normal operation; compare the simulation data with the actual benchmark data in real time, use the actual sensor data of the current wind turbine as the standard, use the optimization algorithm to finely iterate and adjust the physical properties of the digital twin, until the simulation data and the measured data are highly consistent within the allowable error range, thereby constructing a high-fidelity digital twin consistent with the behavior of the physical wind turbine.

6. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The fusion of real-time video and image data of the unmanned aerial vehicle and the verification and dynamic correction of the geometry and physical properties of the twin specifically include: Based on the analysis of the video image data obtained periodically in the initial stage, the appearance characteristics of the wind turbine are extracted to correct the digital twin, and the spatial mapping range of the current video image in the digital twin is determined according to the starting point obtained by the image recognition model in the unmanned aerial vehicle and the real-time flight parameters of the unmanned aerial vehicle, and the influence of the background is removed in the predetermined template; The pretreated multi-angle and multi-time images are input into a structure from motion algorithm SFM, and according to the movement and matching of feature points in the images, the attitude parameters of the camera are automatically solved, and a dense three-dimensional point cloud is generated; According to the spatial mapping range of the video image on the digital model, the corresponding relationship between the point cloud data and the model is determined, and the difference analysis is performed on the point cloud data and the model, and the deviation color spectrum is generated; according to the calculated deviation, the digital twin is deformed to accurately fit the geometric shape of the real point cloud generated by the SFM.

7. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The state information reflected by the digital twin is deeply mined by using a deep learning algorithm, and the specific fault in the wind turbine is identified, which specifically includes: A wind turbine state analysis model is adopted, and a deep learning algorithm convolutional neural network model is used as a basic model to analyze the state information of the digital twin; In the operation monitoring stage, the simulation data of the digital twin and the actual sensor collected data are continuously compared, and the relationship between the real-time operation parameters of the wind turbine and the sensor data is mainly learned, and the dynamic abnormal threshold value is extracted, and the reasonable deviation value between the simulation data and the sensor is identified; When the difference between the simulation data and the actual data of the sensor exceeds the real-time dynamic abnormal threshold value, it is determined that the current wind turbine has an abnormal state that cannot be accurately reflected by the digital twin; Automatic generation of alarm information is pushed to the operation and maintenance personnel, and it is suggested to execute the wind turbine shutdown program, and a unmanned aerial vehicle aerial photography task is started to obtain video image data, and the digital twin is corrected according to the obtained video image data.

8. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 7, characterized in that: The dynamic abnormal threshold value specifically includes: The core idea of the dynamic abnormal threshold value is to adaptively adjust the judgment standard according to the current operation condition, and for each sensor monitoring point, the residual sequence is calculated, and the calculation formula is: ； wherein, is actual sensor data, is digital twin simulation data; Recorded operating parameters: rotational speed , power , wind speed , combined to form an operating parameter vector , cluster the historical operating data using the K-means algorithm to determine the optimal number of operating condition categories K, and determine the feature space for each operating condition category ; during operation, compare the current operating parameters with the corresponding parameters in the feature space to determine the current operating condition category, and the formula is: ； wherein K is an existing operating condition characteristic space, is an operating parameter vector, is a characteristic center vector of the kth operating condition class. The calculation formula is: ； wherein, is the average difference of the sensor i, under the working condition is the residual of the sensor i, is the Kth working condition class, is the mathematical expectation operator, indicating the average value; The calculation is made in a specific working condition The typical offset of the difference between the measured value of the sensor and the simulation value is calculated by the following formula: ； wherein, is the residual standard deviation of sensor i at operating condition is the residual standard deviation of sensor i at operating condition is the variance operator; The final dynamic abnormal threshold value setting formula is: ； wherein, is a sensitivity coefficient (initial value is set to 2.0), The threshold value is an interval range.

9. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The evolution trend of the predicted fault and the potential operation risk of the equipment specifically include: The operation of the wind turbine is predicted based on the simulation operation of the digital twin, a LSTM long short-term memory network model is used to deeply learn the long-term wind turbine operation parameters, historical operation parameters and weather forecast data are used as joint input for training, and the non-linear relationship between external wind conditions and internal operation response is deeply mined from the model, so that the model can simulate the future operation parameter data of the wind turbine according to the local weather forecast; An independent virtual computing space isolated from the real-time operating system is opened on the cloud platform, the digital twin is copied to the virtual space, in the virtual environment, a simulation time axis independent of the real time is set, the predicted operation parameter data is input into the digital twin in the virtual space, the finite element analysis in the digital twin is started, and the shape change of the digital twin and the simulation data of each sensor under the current condition are calculated and simulated according to the input predicted wind condition; When the wind turbine elements have obvious damage in the simulation operation, a risk alarm is generated. For the simulated sensor simulation data, the wind turbine state analysis model for digital twin state analysis is used, the LSTM long short-term memory network model simulates the wind turbine operation parameter input, simulates the theoretical sensor data, and further compares the sensor simulation data with the theoretical sensor data, and accordingly analyzes the dynamic abnormal threshold to identify the abnormality; when there is data anomaly, the corresponding risk warning is generated; The risk warning is automatically classified, and the corresponding operation and maintenance decision is generated and sent to the operation and maintenance personnel.

10. The wind turbine blade internal cavity fault detection method based on artificial intelligence according to claim 1, characterized in that: The parameter optimization and iterative update of the deep learning algorithm and the risk prediction model according to the evaluation result specifically include: The simulation data of the digital twin is compared with the actual data at the same time, combined with the standardized work order; based on the data alignment result, the following core performance indicators are quantitatively calculated: fault detection rate, false alarm rate, missed alarm rate, prediction lead time and positioning accuracy; In the case of insufficient prediction accuracy, the model is optimized; The verified new fault mode, its feature data and the corresponding maintenance scheme are updated to the database; When the deviation between the model prediction result and the actual data is within the allowable range and the running state is stable, high-quality prediction data is selected to form an enhanced training set, and the model is quickly adapted to the operation characteristics of a specific station through incremental training.

Images