A wind turbine generator unit intelligent inspection method and system

Abstract

A wind turbine intelligent inspection method and system are disclosed. The method realizes continuous uninterrupted phase observation under extreme working conditions through a semantic level and brightness transition double-channel perception system, introduces an adaptive extended Kalman filter model with angular acceleration dimension, solves the problems of easy target missing of phase tracking and easy divergence of filter under non-steady wind speed, guarantees uniform sampling of the whole blade through segmented coupling flight control, improves high-definition imaging quality through predictive shutter triggering, and finally reduces the blade identity confusion rate in a single task through on-site database building identification, greatly reducing the inspection downtime loss and operation and maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of wind power generation operation and maintenance and UAV intelligent inspection technology, specifically to an intelligent inspection method and system for wind turbine generator sets. Background Technology

[0002] With the advancement of the global "carbon peaking and carbon neutrality" strategy, wind power installed capacity continues to grow rapidly. According to the latest data from the Global Wind Energy Council (GWEC), the capacity of wind turbines is rapidly developing towards larger sizes, with onshore wind turbines generally reaching 6MW+ and offshore wind turbines exceeding 16MW, and blade lengths generally exceeding 80 meters or even reaching 120 meters.

[0003] As the core component for capturing wind energy, wind turbine blades are constantly exposed to the harsh environment at high altitudes, making them highly susceptible to damage from lightning strikes, rain erosion, sand and gravel abrasion, and fatigue cracks. Statistics show that blade failures account for approximately 20%-30% of all wind turbine failures, and maintenance costs are high. Therefore, regular, high-frequency inspections are crucial for ensuring the safe operation of the unit and preventing catastrophic accidents such as blade breakage.

[0004] Current blade inspections primarily rely on a shutdown inspection mode, where the turbine is completely locked in place using pitch and braking systems, and inspections are conducted manually or by drones. While drone-based shutdown inspections have significantly improved efficiency, this mode still faces two major industry pain points that are difficult to overcome: 1. Huge economic losses (downtime losses): During routine maintenance, power must be disconnected and the rotor locked. For a 10MW offshore wind turbine, a 4-6 hour shutdown for maintenance at full wind speed can result in a loss of tens of thousands of RMB in power generation alone. In a large wind farm with hundreds of turbines, the annual power generation loss due to routine maintenance is astronomical.

[0005] 2. Mechanical life loss: Frequent start-up and shutdown operations will cause huge mechanical stress and wear on the wind turbine's pitch system, yaw system and main shaft braking system, accelerating equipment aging and shortening the unit's life.

[0006] To address these issues, the industry has begun exploring "non-stop inspection" technology, which aims to complete inspections while the wind turbines are generating electricity normally.

[0007] However, early technological attempts (such as ground telephoto shooting and simple timed snapshots) were limited by problems such as a single perspective and inability to achieve precise synchronization, making it difficult to meet the needs of refined inspection.

[0008] Specifically, existing non-stop inspection technology has the following core shortcomings that need to be addressed: (1) The perception of a single semantic visual channel is interrupted under extreme conditions such as backlight and dense fog, resulting in discontinuous phase observation; (2) The phase prediction method based on the fixed uniform velocity model is prone to phase tracking failure and filter divergence under unsteady wind speeds such as gusts; (3) The lack of coupled control between the flight speed of the UAV and the rotation speed of the blades leads to uneven image sampling density with the rotation speed and a high risk of missed detection; (4) There is a non-negligible link delay between the shooting command and the shutter opening, which produces motion blur and phase deviation on the high-speed rotating blades; (5) In a single task, the blade identification relies on the historical database. After the blade surface is repaired, the identification accuracy drops significantly and blade ID confusion is likely to occur. Summary of the Invention

[0009] To address these issues, this invention proposes an intelligent inspection method and system for wind turbine generator sets.

[0010] According to one aspect of the present invention, an intelligent inspection method for wind turbine generator sets is proposed, comprising the following steps: S1, based on the ambient light intensity and semantic recognition confidence of the captured images of the wind turbine generator, select either the semantic-level visual perception channel or the brightness transition visual perception channel to perform visual observation of the wind turbine generator, and output the visual observation vector. The semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. The brightness transition visual perception channel performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector. S2, Construct the extended Kalman filter model and define the system state vector of the wind turbine generator. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period, This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix The prior error covariance matrix; S3, based on the normalized residual squared The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model.

[0011] By constructing a strongly coupled closed-loop system of dual-channel visual perception and adaptive extended Kalman filter state estimation, the core pain points of existing wind turbine non-stop inspection, such as phase tracking easily missing the target under unsteady wind speed and phase observation interruption under extreme conditions, are solved. The state model is constructed by introducing the angular acceleration dimension, thereby adapting to the nonlinear changes in rotational speed caused by gusts and pitch. The extended Kalman filter update logic is adaptively adjusted based on the observation residual to avoid filter divergence, thus achieving high-precision and highly robust continuous phase tracking.

[0012] Specifically, the semantic-level visual perception channel in S1 identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. This includes setting multiple physical anchor points and corresponding physical anchor point observation values, and setting the position state conditions or shape ratio conditions of the wind turbine blades corresponding to each physical anchor point as physical anchor point events. When the semantic-level visual perception channel observes the occurrence of the corresponding physical anchor point event, it outputs the corresponding physical anchor point observation value as the visual observation vector.

[0013] Specifically, the multiple physical anchor points and their corresponding physical anchor point observations include at least the following three configurations: Using the center of the wind turbine tower as the windward physical anchor point, when the semantic-level visual perception channel observes that the position of the blade coincides with the tower, it determines that a windward physical anchor point event has occurred and outputs a 0° physical anchor point observation value. Using the center of the wind turbine nacelle as the leeward physical anchor point, when the semantic-level visual perception channel observes that the position of the blade is vertically aligned with the nacelle, it determines that a leeward physical anchor point event has occurred and outputs the 180° physical anchor point observation value. Using a preset threshold as the physical anchor point, when the aspect ratio of the leaf image observed by the semantic-level visual perception channel reaches the preset threshold, a side physical anchor point event is determined to have occurred, and a 90° or 270° physical anchor point observation value is output.

[0014] A multi-scenario physical anchor point event was designed for the full-view inspection of wind turbines, which solves the shortcomings of existing technologies that can only achieve phase calibration from the front of the wind turbine and have insufficient alignment accuracy from complex perspectives such as the side / back.

[0015] Specifically, the brightness transition visual perception channel in S1 performs visual observation based on the brightness step event in a specific ROI region and outputs the visual observation vector, which specifically includes: When the observed grayscale change rate in the specific ROI region exceeds the dynamic contrast threshold and the duration of the grayscale change falls within the time window of the wind turbine blades passing through, the brightness transition visual perception channel determines that a brightness transition event has occurred in the specific ROI region. It records the time difference and phase difference between two consecutive brightness transition events to calculate the visual observation vector for output. The phase difference... , The number of blades in a single wind turbine of the aforementioned wind turbine generator set.

[0016] An event-level visual perception brightness step observation mechanism was designed to address the industry pain point of phase observation interruption in scenarios where semantic recognition fails, such as strong backlight, dense fog, and low illumination. It eliminates the need to identify semantic targets such as blades or towers, achieving speed observation solely through pixel-level brightness changes. Dual triggering criteria effectively eliminate false triggers, and the extended Kalman filter observation matrix is ​​dynamically switched upon triggering to calibrate the speed, ensuring phase tracking accuracy in inertial prediction mode and enabling continuous operation of the inspection system.

[0017] Specifically, in S1, based on the ambient light intensity and semantic recognition confidence of the captured images of the wind turbine generator, either a semantic-level visual perception channel or a brightness transition visual perception channel is selected to perform visual observation of the wind turbine generator, and a visual observation vector is output. Specifically, it includes: The average gray value of the central ROI region of the captured wind turbine image is obtained as the ambient light intensity, and the target category probability score output by the built-in target recognition algorithm for the wind turbine image is obtained as the semantic recognition confidence. When the ambient light intensity is within a preset light threshold range and the semantic recognition confidence level is greater than or equal to a preset confidence threshold, the semantic-level visual perception channel is selected for visual observation and outputs a visual observation vector; otherwise, the brightness transition visual perception channel is selected for visual observation and outputs a visual observation vector.

[0018] Using ambient light intensity and semantic recognition confidence level as dual objective judgment criteria, the system achieves fully automatic and seamless switching of sensing channels without manual intervention, making it fully adaptable to unattended, fully automated inspection scenarios. Under normal operating conditions, the semantic channel is activated to ensure phase calibration accuracy, while under extreme conditions, the event channel is automatically switched to ensure observation continuity. This significantly expands the environmental adaptability of the inspection system and improves operational stability and intelligence.

[0019] Specifically, S3 includes: Define preset threshold ,when When the wind turbine generator is in steady-state operation, the noise covariance matrix is ​​defined. Preset value Calculate Kalman gain The state is updated based on the Kalman gain to obtain the a posteriori system operating state of the wind turbine generator. ; when When the wind turbine generator is in a non-steady-state operation, the noise covariance matrix is ​​defined. ， The preset maximum expansion coefficient, and Calculate the Kalman gain based on the phase noise variance and angular velocity noise variance obtained from offline calibration. Based on the Kalman gain, the state is updated to obtain the a posteriori system operating state of the wind turbine generator set. .

[0020] Based on the chi-square test of normalized residual squares, the steady-state and non-steady-state operating conditions of wind turbines can be accurately distinguished, solving the core problems of divergence and phase off-target in non-Gaussian noise scenarios such as gusts of wind for fixed-gain extended Kalman filters. In steady state, strong phase correction is achieved through standard gain to eliminate integral drift; in non-steady state, the inertial prediction mode is smoothly switched through the expansion noise covariance matrix, balancing the absolute accuracy of phase tracking and the robustness against disturbances.

[0021] Specifically, it also includes S4, obtaining the corresponding angular velocity prediction result based on the system state vector updated by the extended Kalman filter model. When the angular velocity prediction result When the vertical climb rate of the inspection drone used for visual observation is greater than or equal to a preset threshold, it is established. Nonlinear coupling control law with the real-time angular velocity of the wind turbine generator ,in, The vertical field of view height of a single frame of the captured image sensor. , This indicates the preset horizontal distance between the drone and the surface of the blade being photographed. This indicates the vertical field of view of the sensor used to capture the image. , This indicates the effective size of the sensor in the vertical direction of the captured image. This indicates the focal length of the sensor used to capture the image. As a high correction factor, The preset spatial overlap ratio is the ratio of the blade area of ​​the wind turbine in two adjacent frames of images captured by the image sensor.

[0022] A nonlinear coupled control law was constructed to integrate the UAV's climb speed with the wind turbine's real-time rotational speed, enabling adaptive synchronous scanning. This addresses the shortcomings of existing technologies that decouple flight speed from wind turbine rotational speed, leading to unstable image overlap and high risk of missed detections. A correction factor that varies with altitude is introduced to ensure a constant sampling density across the entire blade area, reducing the probability of missed detections in high-risk regions.

[0023] Specifically, it also includes S5, which obtains the phase prediction result of the corresponding wind turbine blades based on the updated system state vector of the extended Kalman filter model. With angular velocity prediction results The inspection drone calculates the phase when it detects the blades of the wind turbine generator as the target and takes a picture. The trigger time of the corresponding shooting command. ,in Indicates the current moment. Indicates signal transmission delay. This indicates the mechanical shutter lag of the image sensor. This represents the time consumed by visual algorithm inference and image transmission. This indicates that the phase difference between the calculated phase prediction result and the target image phase is... Normalization and positive value selection operations.

[0024] By compensating for time delays in algorithm inference, signal transmission, and mechanical shutter processes, the problems of mechanical following lag, motion blur, and phase miss during high-speed rotating blade shooting in existing technologies have been solved.

[0025] Specifically, it also includes S6, which collects data on each wind turbine of the wind turbine generator set when it is identified by the semantic-level visual perception channel. Images of the blade's operational status at physical anchor points are collected, and a high-definition reference image library of blades is established. Then, images collected by the inspection drone at the same altitude during a single visual observation are packaged into an image sequence. High-confidence feature anchor frames are selected from the image sequence, and feature matching is performed between the high-confidence feature anchor frames and the high-definition reference image library of blades. Based on the temporal relationship of the images in the same image sequence, the number of the target blade visually observed by the inspection drone is identified.

[0026] A closed-loop blade identification mechanism was constructed for a single inspection mission. Based on the 0° absolute phase anchor point, a dedicated blade reference image library for this mission was built on-site, solving the shortcomings of existing technologies that rely on historical databases and whose identification accuracy drops significantly after blade surface contamination repair. Sequence packaging and high-confidence anchor point frame matching logic were adopted, and the image overlap rate was used to achieve full-sequence identity relay locking, solving the industry pain point of blade ID confusion in non-stop inspection.

[0027] According to one aspect of the present invention, a smart inspection system for wind turbine generators using the method described in any one of the first aspects is proposed, comprising the following modules: The visual perception channel selection module is configured to select either a semantic-level visual perception channel or a brightness transition visual perception channel to perform visual observation of the wind turbine generator based on the ambient light intensity and semantic recognition confidence of the captured images, and output a visual observation vector. The semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. The brightness transition visual perception channel performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector. The residual calculation module is configured to construct an extended Kalman filter model and define the system state vector of the wind turbine generator. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period, This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix The prior error covariance matrix; The system status update module is configured to update the normalized residual squared value. The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model. The flight speed control module is configured to obtain the corresponding angular velocity prediction result based on the system state vector updated by the extended Kalman filter model. When the angular velocity prediction result When the vertical climb rate of the inspection drone used for visual observation is greater than or equal to a preset threshold, it is established. Nonlinear coupling control law with the real-time angular velocity of the wind turbine generator ,in, The vertical field of view height of a single frame of the captured image sensor. , This indicates the preset horizontal distance between the drone and the surface of the blade being photographed. This indicates the vertical field of view of the sensor used to capture the image. , This indicates the effective size of the sensor in the vertical direction of the captured image. This indicates the focal length of the sensor used to capture the image. As a high correction factor, The preset spatial overlap ratio of the blade area of ​​the wind turbine in two adjacent frames captured by the image sensor; The shutter prediction trigger module is configured to obtain the phase prediction result of the corresponding wind turbine blades based on the system state vector updated by the extended Kalman filter model. With angular velocity prediction results The inspection drone calculates the phase when it detects the blades of the wind turbine generator as the target and takes a picture. The trigger time of the corresponding shooting command. ,in Indicates the current moment. Indicates signal transmission delay. This indicates the mechanical shutter lag of the image sensor. This represents the time consumed by visual algorithm inference and image transmission. This indicates that the phase difference between the calculated phase prediction result and the target image phase is... Normalization and positive value selection operation; The blade identification module is configured to collect data on each wind turbine of the wind turbine generator set when it is identified by the semantic-level visual perception channel. Images of the blade's operational status at physical anchor points are collected, and a high-definition reference image library of blades is established. Then, images collected by the inspection drone at the same altitude during a single visual observation are packaged into an image sequence. High-confidence feature anchor frames are selected from the image sequence, and feature matching is performed between the high-confidence feature anchor frames and the high-definition reference image library of blades. Based on the temporal relationship of the images in the same image sequence, the number of the target blade visually observed by the inspection drone is identified.

[0028] The advantages of this invention are as follows: by using a semantic-level and brightness transition dual-channel perception system, continuous and uninterrupted phase observation can be achieved under extreme operating conditions. Furthermore, by introducing an adaptive extended Kalman filter model in the angular acceleration dimension, the problems of phase tracking easily missing the target and filter easily diverging under unsteady wind speeds are solved. At the same time, segmented coupled flight control ensures uniform sampling of the entire blade, and predictive shutter triggering improves the quality of high-definition imaging. Finally, on-site database-based identity recognition reduces the blade identity confusion rate within a single mission, significantly reducing downtime losses and maintenance costs during inspections. Attached Figure Description

[0029] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0030] Figure 1 A flowchart illustrating an intelligent inspection method for wind turbine generator sets according to the present invention is shown. Figure 2 A schematic diagram of a wind turbine intelligent inspection system according to the present invention is shown. Figure 3 A schematic diagram of a computer system architecture suitable for implementing the embodiments of this application is shown. Detailed Implementation

[0031] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0032] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] Figure 1 A smart inspection method for wind turbine generator sets is shown, including the following steps: S1, based on the ambient light intensity and semantic recognition confidence of the captured images of the wind turbine generator, select either the semantic-level visual perception channel or the brightness transition visual perception channel to perform visual observation of the wind turbine generator, and output the visual observation vector. The semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. The brightness transition visual perception channel performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector. S2, Construct the extended Kalman filter model and define the system state vector of the wind turbine generator. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period; This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix The prior error covariance matrix; S3, based on the normalized residual squared The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model.

[0034] In a specific embodiment, the initial value of the system state vector of the wind turbine generator is determined by the initial observation value of the semantic-level visual perception channel or the brightness transition visual perception channel. The observation value of the semantic-level visual perception channel or the brightness transition visual perception channel at time k. The phase values ​​are obtained through offline calibration. In a controlled environment with known true phase angles (such as a calibration turntable or a target with known angles), the fan or simulation device is driven at a stable speed, and a large number of phase observations are continuously collected using the same visual perception channel of the system (YOLO semantic recognition or event-level brightness detection). As a measured value, The residuals are compared with the true phase angle and angular velocity, and the variance matrix of the residuals is calculated. The two observation dimensions of phase and angular velocity are calibrated separately to obtain the diagonal matrix. It is loaded as a fixed parameter in each inspection task.

[0035] In a specific embodiment, a dual-channel sensing system is constructed as the observation input source for the extended Kalman filter model to ensure that phase observation is not interrupted under extreme conditions such as backlight and dense fog.

[0036] Specifically, in S1, based on the ambient light intensity and semantic recognition confidence of the captured images of the wind turbine generator, either a semantic-level visual perception channel or a brightness transition visual perception channel is selected to perform visual observation of the wind turbine generator, and a visual observation vector is output. Specifically, it includes: The average gray value of the central ROI region of the captured wind turbine image is obtained as the ambient light intensity, and the target category probability score output by the built-in target recognition algorithm for the wind turbine image is obtained as the semantic recognition confidence. When the ambient light intensity is within a preset light threshold range and the semantic recognition confidence level is greater than or equal to a preset confidence threshold, the semantic-level visual perception channel is selected for visual observation and outputs a visual observation vector; otherwise, the brightness transition visual perception channel is selected for visual observation and outputs a visual observation vector.

[0037] Among them, the observation matrix of the extended Kalman filter model It is a deterministic matrix predefined by the system designer, which switches according to the operating type of the visual perception channel. When the semantic-level channel triggers a physical anchor event, It is used to extract phase angle components; when the brightness transition channel triggers a brightness transition... It is used to extract the angular velocity component.

[0038] In a specific embodiment, the system calculates two key indicators in real time: ambient light intensity (…). ) and semantic recognition confidence ( To determine which channel to use: Among them, ambient light intensity ( ) is defined as the ROI region at the center of the calculated image. Average gray value ,in The width and height of the captured image.

[0039] Semantic recognition confidence ( ) is defined as the target category probability score output by the YOLO neural network detection box.

[0040] Only when (The illumination is within the effective dynamic range) , [Within the area, there is no overexposure or underexposure] and Confidence threshold At that time, the system uses the absolute phase anchor point provided by the semantic-level visual perception channel.

[0041] When the environment captured by the inspection drone has characteristics of strong backlight silhouette or dense fog and low light, it leads to... When the effective range is exceeded, or when semantic recognition fails or is unstable, it will result in... When the brightness transition occurs, the system automatically downgrades and switches to the brightness transition visual perception channel, using the brightness transition event to correct the rotation speed.

[0042] The semantic-level visual perception channel in S1 identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. Specifically, this includes setting multiple physical anchor points and corresponding physical anchor point observation values, and setting the position state conditions or shape ratio conditions of the wind turbine blades corresponding to each physical anchor point as physical anchor point events. When the semantic-level visual perception channel observes the occurrence of the corresponding physical anchor point event, it outputs the corresponding physical anchor point observation value as the visual observation vector.

[0043] In one embodiment, under normal lighting conditions, the onboard wide-angle image sensor of the inspection drone and its built-in lightweight YOLO model are used to identify semantic targets such as "tower" and "nacelle".

[0044] The multiple physical anchor points and their corresponding physical anchor point observations include at least the following three configurations: Using the center of the wind turbine tower as the windward physical anchor point, when the semantic-level visual perception channel observes that the position of the blade coincides with that of the tower in the image, it determines that a windward physical anchor point event has occurred and outputs a 0° physical anchor point observation value. Using the center of the wind turbine nacelle as the leeward physical anchor point, when the semantic-level visual perception channel observes that the position of the blade is vertically aligned with the nacelle in the image, it determines that a leeward physical anchor point event has occurred and outputs a 180° physical anchor point observation value. Using a preset threshold as the physical anchor point, when the aspect ratio of the leaf image observed by the semantic-level visual perception channel reaches the preset threshold, a side physical anchor point event is determined to have occurred, and a 90° or 270° physical anchor point observation value is output.

[0045] In strong backlight (silhouette mode) or low contrast scenes, the system detects brightness steps or edge flicker in specific ROI areas of wide-angle images. By calculating the time difference ΔT between two adjacent brightness steps, the instantaneous angular velocity is directly calculated. This ensures that even in extreme cases where semantic features fail (cannot identify specific objects), the system can still maintain the accuracy of phase estimation through precise velocity updates.

[0046] Specifically, the brightness transition visual perception channel in S1 performs visual observation based on the brightness step event in a specific ROI region and outputs the visual observation vector, which specifically includes: When the observed grayscale change rate in the specific ROI region exceeds the dynamic contrast threshold and the duration of the grayscale change falls within the time window of the wind turbine blades passing through, the brightness transition visual perception channel determines that a brightness transition event has occurred in the specific ROI region. It records the time difference and phase difference between two consecutive brightness transition events to calculate the visual observation vector for output. The phase difference... , The number of blades in a single wind turbine of the aforementioned wind turbine generator set.

[0047] The system delineates static regions of interest in the image and calculates the average gray value of these regions in real time. The triggering of a brightness step event requires the simultaneous fulfillment of the following two mathematical conditions: Intensity threshold condition: The grayscale change rate exceeds a preset threshold, i.e. ,in The contrast threshold is dynamically adjusted based on ambient light. Temporal window constraint: the duration of the grayscale change It must fall within a reasonable time window for the physical passage of the blades. This is to eliminate false triggers caused by birds, cloud shadows, or high-frequency jitter of the image sensor.

[0048] In one embodiment, the wind turbine is a three-bladed wind turbine, which records the timestamps of two consecutive event triggers. and Calculate the time difference Solve for the observed angular velocity .

[0049] In a specific embodiment, angular acceleration The evolution adopts Describe the process noise interference here. In a physical sense, this directly corresponds to jerk during the update cycle. The integral increment within the system represents unmodeled abrupt dynamic changes in the wind turbine system (such as instantaneous aerodynamic torque changes caused by gusts, drag steps generated by the pitch control system's pitching action, etc.). This direct superposition modeling method avoids complex nonlinear mechanical modeling, simplifying external disturbances into random acceleration steps. When the system detects a significant deviation between the actual observed values ​​and the inertia-based predictions, it means... A non-Gaussian abrupt change (such as a gust of wind) triggers an adjustment of the filter weights, thereby ensuring that the system has inertial prediction capabilities and can maintain phase tracking accuracy for a short period of time even if vision is lost.

[0050] Therefore, to address the filter divergence problem caused by non-Gaussian noise (such as sudden gusts of wind), a dynamic weight adjustment mechanism based on observation residuals is introduced, defining a preset threshold. (e.g., the chi-square test threshold).

[0051] In one embodiment, the normalized squared residuals It follows a chi-square distribution where the degrees of freedom n equals the dimension of the observation vector. Since the observations for both the semantic-level channel and the brightness transition visual perception channel are scalars (n=1),... Based on the chi-square distribution with 1 degree of freedom and the preset significance level, when the significance level α = 0.05, the corresponding... ≈3.84. In engineering practice, the noise level can be adjusted within the range of 3.0 to 7.8 depending on the specific operating conditions of the wind turbine.

[0052] when When the wind turbine generator is in steady-state operation, the high confidence level of the visual observation vectors collected by the aforementioned dual-channel visual sensor is maintained, and the noise covariance matrix is ​​defined. Preset value Calculate Kalman gain , At this point, the value is relatively large, and a state update is performed based on the Kalman gain to obtain the a posteriori system operating state of the wind turbine generator set. Larger This will cause the system state to rapidly change towards the observed value. Convergence. By utilizing the absolute position information of visual anchor points (such as 0 degrees, 180 degrees), the phase drift caused by the accumulation of angular velocity integrals in the state prediction equation is eliminated.

[0053] in, It represents the covariance of the error between the estimated system state and the true state when the state is predicted forward to time k through the state prediction equation after the state update is completed at time k-1. It is used to measure the uncertainty of the extended Kalman filter model's prediction of itself before visual observation fusion. The larger the value, the less reliable the prediction; the smaller the value, the more reliable the prediction. At the initial startup, [the value will be set to - context needed]. Initialize to a large diagonal matrix (e.g.) =diag(1.0,0.1,0.01), where the units are rad², (rad / s)², and (rad / s²)² respectively, indicating high initial uncertainty. Within each extended Kalman filter model update cycle Δt, the prior covariance is calculated using the following formula: .in Let be the state transition matrix (obtained by linearizing the state prediction equation), and Q be the covariance matrix of the process noise interference quantity calibrated offline. After fusing visual observations, Kalman gain is used to... Updated to posterior covariance : .

[0054] when When the wind turbine generator is determined to be in a non-steady-state operation, it may be subject to gust disturbances or visual mismatches. A noise covariance matrix is ​​then defined. ,at this time The value increased dramatically. and Calculate the Kalman gain based on the phase noise variance and angular velocity noise variance obtained from offline calibration. And based on the Kalman gain, the state is updated when When the value is amplified drastically by the above mechanism and approaches infinity, the denominator increases sharply, leading to a change in the calculated gain matrix. Substituting these values ​​into the state update equation yields the posterior system operating state. Therefore, there is The posterior state estimate of the wind turbine generator is directly determined by the prior prediction, thus shielding the system from the interference of abnormal visual observations.

[0055] To prevent numerical calculation overflow, the adaptive expansion coefficient is... Set a saturation upper limit, that is, let , The preset maximum expansion coefficient (can be set according to actual working conditions, such as...) = ),make sure It is numerically bounded, satisfying the requirements for engineering feasibility.

[0056] Specifically, it also includes S4, obtaining the corresponding angular velocity prediction result based on the system state vector updated by the extended Kalman filter model. When the angular velocity prediction result When the vertical climb rate of the inspection drone used for visual observation is greater than or equal to a preset threshold, it is established. Nonlinear coupling control law with the real-time angular velocity of the wind turbine generator .in, The vertical field of view height of a single frame of the captured image sensor. , This indicates the preset horizontal distance between the drone and the surface of the blade being photographed. This indicates the vertical field of view of the sensor used to capture the image. , This indicates the effective size of the sensor in the vertical direction of the captured image. This indicates the focal length of the sensor used to capture the image. As a high correction factor, The preset spatial overlap ratio is the ratio of the blade area of ​​the wind turbine in two adjacent frames of images captured by the image sensor.

[0057] in, As a high correction factor, This refers to the spatial overlap ratio of the target area of ​​the blade (the effective area occupied by the photographed blade in the image) in two consecutive frames of images captured by the image sensor during the vertical climb of the UAV. It is generally taken as 0.4~0.6 (i.e., 40%~60% overlap rate), and can be configured according to the inspection accuracy requirements and flight efficiency requirements. This ensures that no area on the blade is skipped and not captured, guaranteeing a certain area of ​​overlap to provide sufficient common feature points for subsequent blade panoramic stitching and damage localization. For example, η=0.5 (50% overlap rate) means that for every step the UAV takes up... At a height of ×0.5, a single frame is captured, with half of the leaf area overlapping between adjacent frames.

[0058] In a specific embodiment, when the altitude of the inspection drone is within a certain altitude range, set Allows standard speed ascent to improve efficiency; when the inspection drone's altitude exceeds the aforementioned set altitude range, [the following is set:] (e.g., 0.8) Automatically reduce the climbing speed to increase the image sampling density per unit height.

[0059] Among them, when When the wind speed is below the preset threshold, the drone quickly maneuvers to a predetermined altitude and hovers, waiting for the blades to sweep by, thus avoiding repeated shooting under low wind speeds.

[0060] Specifically, it also includes S5, which obtains the phase prediction result of the corresponding wind turbine blades based on the updated system state vector of the extended Kalman filter model. With angular velocity prediction results The inspection drone calculates the phase when it detects the blades of the wind turbine generator as the target and takes a picture. The trigger time of the corresponding shooting command. ,in Indicates the current moment. Indicates signal transmission delay. This indicates the mechanical shutter lag of the image sensor. This represents the time consumed by visual algorithm inference and image transmission. This indicates that the phase difference between the calculated phase prediction result and the target image phase is... Normalization and positive value selection operations.

[0061] By taking lead time into account, it is ensured that when the drone is acquiring images of the wind turbine, the blades are captured at the exact moment the shutter is fully open, thus eliminating motion position deviation.

[0062] Specifically, it also includes S6, which collects data on each wind turbine of the wind turbine generator set when it is identified by the semantic-level visual perception channel. Images of the blade's operational status at physical anchor points are collected, and a high-definition reference image library of blades is established. Then, images collected by the inspection drone at the same altitude during a single visual observation are packaged into an image sequence. High-confidence feature anchor frames are selected from the image sequence, and feature matching is performed between the high-confidence feature anchor frames and the high-definition reference image library of blades. Based on the temporal relationship of the images in the same image sequence, the number of the target blade visually observed by the inspection drone is identified.

[0063] In a specific embodiment, filtering high-confidence feature anchor frames specifically includes: using the Laplacian or Tenengrad operator to remove motion-blurred frames caused by the high-speed movement of the blades. For example, in one embodiment, applying a Laplacian convolution kernel to image frame I yields the gradient response map L= ²I, where ² denotes the Laplace operator, used to calculate the variance of L. ,make sure , This indicates the preset sharpness threshold. Greater than... Frames that are clear are identified as clear frames, while those that are not are identified as motion-blurred frames and are discarded.

[0064] Then, the Structural Similarity Index (SSIM) is introduced as a primary screening metric to calculate the global structural similarity between the regions of interest (ROIs) of the images retained after the image sequence screening and the images in the high-resolution leaf reference image library. A similarity threshold is then set. Only candidate frames that are highly consistent with the baseline image in terms of overall structure and texture distribution are retained. This allows for the rapid removal of background interference or non-target leaf frames.

[0065] The homography matrix is ​​then estimated using the RANSAC (Random Sample Consensus) algorithm. The homography matrix is ​​obtained by calculating the geometric projection mapping between two images taken from different viewpoints but containing the same blade surface (e.g., calculating the projection mapping between the coordinates of multiple feature points in the two images). Based on the estimated homography matrix, the feature points of the current image are reprojected onto the reference image coordinate system. The Euclidean distance between the predicted position and the actual matching feature point is then used as the reprojection error. Set a pixel error threshold. Only if the reprojection error is satisfied Number of interior points Exceeding the preset threshold (e.g.) Only when the frame is in the specified state will it be considered a valid identity lock frame.

[0066] In one embodiment, for a three-bladed wind turbine, since the physical order of operation is irreversible, it follows the order from blade A to blade B and then to blade C. If the collected sequence photos are arranged in the order of timestamps, and sequence N is confirmed as A and sequence N+2 is confirmed as C, then the middle sequence N+1 must be B. This can be used to correct the situation of visual rejection or misrecognition.

[0067] According to one aspect of the present invention, a smart inspection system for wind turbine generators using the method described in any one of the first aspects is proposed, comprising the following modules, such as... Figure 2 As shown, it includes the following modules: The visual perception channel selection module 201 is configured to select a semantic-level visual perception channel or a brightness transition visual perception channel to perform visual observation of the wind turbine generator based on the ambient light intensity and semantic recognition confidence of the captured image of the wind turbine generator, and output a visual observation vector. The semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. The brightness transition visual perception channel performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector. The residual calculation module 202 is configured to construct an extended Kalman filter model and define the system state vector of the wind turbine generator. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period, This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix The prior error covariance matrix; System status update module 203 is configured to update the normalized residual squared value based on the normalized residual squared value. The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model. The flight speed control module 204 is configured to obtain the corresponding angular velocity prediction result based on the system state vector updated by the extended Kalman filter model. When the angular velocity prediction result When the vertical climb rate of the inspection drone used for visual observation is greater than or equal to a preset threshold, it is established. Nonlinear coupling control law with the real-time angular velocity of the wind turbine generator ,in, The vertical field of view height of a single frame of the captured image sensor. , This indicates the preset horizontal distance between the drone and the surface of the blade being photographed. This indicates the vertical field of view of the sensor used to capture the image. , This indicates the effective size of the sensor in the vertical direction of the captured image. This indicates the focal length of the sensor used to capture the image. As a high correction factor, The preset spatial overlap ratio of the blade area of ​​the wind turbine in two adjacent frames captured by the image sensor; The shutter prediction trigger module 205 is configured to obtain the phase prediction result of the corresponding wind turbine blades based on the system state vector updated by the extended Kalman filter model. With angular velocity prediction results The inspection drone calculates the phase when it detects the blades of the wind turbine generator as the target and takes a picture. The trigger time of the corresponding shooting command. ,in Indicates the current moment. Indicates signal transmission delay. This indicates the mechanical shutter lag of the image sensor. This represents the time consumed by visual algorithm inference and image transmission. This indicates that the phase difference between the calculated phase prediction result and the target image phase is... Normalization and positive value selection operation; The blade identification module 206 is configured to collect data on each wind turbine of the wind turbine generator set when it is identified by the semantic-level visual perception channel. Images of the blade's operational status at physical anchor points are collected, and a high-definition reference image library of blades is established. Then, images collected by the inspection drone at the same altitude during a single visual observation are packaged into an image sequence. High-confidence feature anchor frames are selected from the image sequence, and feature matching is performed between the high-confidence feature anchor frames and the high-definition reference image library of blades. Based on the temporal relationship of the images in the same image sequence, the number of the target blade visually observed by the inspection drone is identified.

[0068] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system 300 suitable for implementing electronic devices according to embodiments of the present application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0069] like Figure 3 As shown, the computer system 300 includes a central processing unit (CPU) 301, which performs various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 309 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the system 300. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0070] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a liquid crystal display (LCD) and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 310 as needed so that computer programs read from it can be installed into storage section 308 as needed.

[0071] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts are implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program is downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application.

[0072] It should be noted that the computer-readable storage medium of this application is a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium is, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium is any tangible medium that contains or stores a program used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium includes a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals take various forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium or any computer-readable storage medium other than a computer-readable storage medium may transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0073] Computer program code for performing the operations of this application is written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code executes entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer is connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or connected to an external computer (e.g., via the Internet using an Internet service provider).

[0074] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram represents a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually execute substantially in parallel, and they may sometimes execute in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, is implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0075] The modules described in the embodiments of this application are implemented in software or hardware.

[0076] In another aspect, this application also provides a computer-readable storage medium, which is included in the electronic device described in the above embodiments; it also exists independently and is not assembled into the electronic device. The computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: S1, based on the ambient light intensity and semantic recognition confidence of the captured image of the wind turbine generator, select a semantic-level visual perception channel or a brightness transition visual perception channel to perform visual observation of the wind turbine generator, and output a visual observation vector. S1, the semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm to perform visual observation and output the visual observation vector; the brightness transition visual perception channel performs visual observation and outputs the visual observation vector based on brightness step events in a specific ROI region; S2, construct an extended Kalman filter model and define the system state vector of the wind turbine generator set. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period, This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix S3 is the prior error covariance matrix; S4 is based on the normalized residual squared value. The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model.

[0077] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for intelligent inspection of wind turbine generator sets, characterized in that, Includes the following steps: S1, based on the ambient light intensity and semantic recognition confidence of the captured images of the wind turbine generator, select either the semantic-level visual perception channel or the brightness transition visual perception channel to perform visual observation of the wind turbine generator, and output the visual observation vector. The semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. The brightness transition visual perception channel performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector. S2, Construct the extended Kalman filter model and define the system state vector of the wind turbine generator. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period, This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix The prior error covariance matrix; S3, based on the normalized residual squared The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model.

2. The intelligent inspection method for wind turbine generator sets according to claim 1, characterized in that, The semantic-level visual perception channel in S1 identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. Specifically, this includes setting multiple physical anchor points and corresponding physical anchor point observation values, and setting the position state conditions or shape ratio conditions of the wind turbine blades corresponding to each physical anchor point as physical anchor point events. When the semantic-level visual perception channel observes the occurrence of the corresponding physical anchor point event, it outputs the corresponding physical anchor point observation value as the visual observation vector.

3. The intelligent inspection method for wind turbine generator sets according to claim 2, characterized in that, The multiple physical anchor points and their corresponding physical anchor point observations include at least the following three configurations: Using the center of the wind turbine tower as the windward physical anchor point, when the semantic-level visual perception channel observes that the position of the blade coincides with the tower, it determines that a windward physical anchor point event has occurred and outputs a 0° physical anchor point observation value. Using the center of the wind turbine nacelle as the leeward physical anchor point, when the semantic-level visual perception channel observes that the position of the blade is vertically aligned with the nacelle, it determines that a leeward physical anchor point event has occurred and outputs the 180° physical anchor point observation value. Using a preset threshold as the physical anchor point, when the aspect ratio of the leaf image observed by the semantic-level visual perception channel reaches the preset threshold, a side physical anchor point event is determined to have occurred, and a 90° or 270° physical anchor point observation value is output.

4. The intelligent inspection method for wind turbine generator sets according to claim 1, characterized in that, The brightness transition visual perception channel described in S1 performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector, which specifically includes: When the observed grayscale change rate in the specific ROI region exceeds the dynamic contrast threshold and the duration of the grayscale change falls within the time window of the wind turbine blades passing through, the brightness transition visual perception channel determines that a brightness transition event has occurred in the specific ROI region. It records the time difference and phase difference between two consecutive brightness transition events to calculate the visual observation vector for output. The phase difference... , The number of blades in a single wind turbine of the aforementioned wind turbine generator set.

5. The intelligent inspection method for wind turbine generator sets according to claim 1, characterized in that, In S1, based on the ambient light intensity and semantic recognition confidence of the captured images of the wind turbine generator, either a semantic-level visual perception channel or a brightness transition visual perception channel is selected to perform visual observation of the wind turbine generator, and a visual observation vector is output. Specifically, it includes: The average gray value of the central ROI region of the captured wind turbine image is obtained as the ambient light intensity, and the target category probability score output by the built-in target recognition algorithm for the wind turbine image is obtained as the semantic recognition confidence. When the ambient light intensity is within a preset light threshold range and the semantic recognition confidence level is greater than or equal to a preset confidence threshold, the semantic-level visual perception channel is selected for visual observation and outputs a visual observation vector; otherwise, the brightness transition visual perception channel is selected for visual observation and outputs a visual observation vector.

6. The intelligent inspection method for wind turbine generator sets according to claim 1, characterized in that, S3 specifically includes: Define preset threshold ,when When the wind turbine generator is in steady-state operation, the noise covariance matrix is ​​defined. Preset value Calculate Kalman gain The state is updated based on the Kalman gain to obtain the a posteriori system operating state of the wind turbine generator. ; when When the wind turbine generator is in a non-steady-state operation, the noise covariance matrix is ​​defined. ， The preset maximum expansion coefficient, and Calculate the Kalman gain based on the phase noise variance and angular velocity noise variance obtained from offline calibration. Based on the Kalman gain, the state is updated to obtain the a posteriori system operating state of the wind turbine generator set. 。 7. The intelligent inspection method for wind turbine generator sets according to claim 1, characterized in that, It also includes S4, which obtains the corresponding angular velocity prediction result based on the system state vector updated by the extended Kalman filter model. When the angular velocity prediction result When the vertical climb rate of the inspection drone used for visual observation is greater than or equal to a preset threshold, it is established. Nonlinear coupling control law with the real-time angular velocity of the wind turbine generator ,in, The vertical field of view height of a single frame of the captured image sensor. , This indicates the preset horizontal distance between the drone and the surface of the blade being photographed. This indicates the vertical field of view of the sensor used to capture the image. , This indicates the effective size of the sensor in the vertical direction of the captured image. This indicates the focal length of the sensor used to capture the image. As a high correction factor, The preset spatial overlap ratio is the ratio of the blade area of ​​the wind turbine in two adjacent frames of images captured by the image sensor.

8. The intelligent inspection method for wind turbine generator sets according to claim 7, characterized in that, It also includes S5, which obtains the phase prediction result of the corresponding wind turbine blades based on the updated system state vector of the extended Kalman filter model. With angular velocity prediction results The inspection drone calculates the phase when it detects the blades of the wind turbine generator as the target and takes a picture. The trigger time of the corresponding shooting command. ,in Indicates the current moment. Indicates signal transmission delay. This indicates the mechanical shutter lag of the image sensor. This represents the time consumed by visual algorithm inference and image transmission. This indicates that the phase difference between the calculated phase prediction result and the target image phase is... Normalization and positive value selection operations.

9. The intelligent inspection method for wind turbine generator sets according to claim 3, characterized in that, It also includes S6, which collects data on each wind turbine of the wind turbine generator set when it is identified by the semantic-level visual perception channel. Images of the blade's operational status at physical anchor points are collected, and a high-definition reference image library of blades is established. Then, images collected by the inspection drone at the same altitude during a single visual observation are packaged into an image sequence. High-confidence feature anchor frames are selected from the image sequence, and feature matching is performed between the high-confidence feature anchor frames and the high-definition reference image library of blades. Based on the temporal relationship of the images in the same image sequence, the number of the target blade visually observed by the inspection drone is identified.

10. A wind turbine generator intelligent inspection system applying the method as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The visual perception channel selection module is configured to select either a semantic-level visual perception channel or a brightness transition visual perception channel to perform visual observation of the wind turbine generator based on the ambient light intensity and semantic recognition confidence of the captured images, and output a visual observation vector. The semantic-level visual perception channel identifies preset physical anchor point events through a target detection algorithm and outputs the visual observation vector. The brightness transition visual perception channel performs visual observation based on brightness step events in a specific ROI region and outputs the visual observation vector. The residual calculation module is configured to construct an extended Kalman filter model and define the system state vector of the wind turbine generator. Construct state prediction equations The prior estimated system state vector is obtained by updating the system state vector based on the state prediction equation. Calculate the observation residual between the prior estimated system state vector and the visual observation vector. The normalized squared residuals are calculated based on the observed residuals. ,in, To extend the Kalman filter model update period, This refers to the amount of process noise interference. Let k be the blade phase angle. Let k be the blade angular velocity. Let k be the blade angular acceleration at time k. To extend the observation matrix of the Kalman filter model, values ​​are assigned based on the selection type of the visual perception channel. Based on the measurement of the noise covariance matrix The prior error covariance matrix; The system status update module is configured to update the normalized residual squared value. The system determines the operating status of the wind turbine generator by checking whether the value exceeds a preset threshold, and updates the system state vector under different operating conditions using the extended Kalman filter model. The flight speed control module is configured to obtain the corresponding angular velocity prediction result based on the system state vector updated by the extended Kalman filter model. When the angular velocity prediction result When the vertical climb rate of the inspection drone used for visual observation is greater than or equal to a preset threshold, it is established. Nonlinear coupling control law with the real-time angular velocity of the wind turbine generator ,in, The vertical field of view height of a single frame of the captured image sensor. , This indicates the preset horizontal distance between the drone and the surface of the blade being photographed. This indicates the vertical field of view of the sensor used to capture the image. , This indicates the effective size of the sensor in the vertical direction of the captured image. This indicates the focal length of the sensor used to capture the image. As a high correction factor, The preset spatial overlap ratio of the blade area of ​​the wind turbine in two adjacent frames captured by the image sensor; The shutter prediction trigger module is configured to obtain the phase prediction result of the corresponding wind turbine blades based on the system state vector updated by the extended Kalman filter model. With angular velocity prediction results The inspection drone calculates the phase when it detects the blades of the wind turbine generator as the target and takes a picture. The trigger time of the corresponding shooting command. ,in Indicates the current moment. Indicates signal transmission delay. This indicates the mechanical shutter lag of the image sensor. This represents the time consumed by visual algorithm inference and image transmission. This indicates that the phase difference between the calculated phase prediction result and the target image phase is... Normalization and positive value selection operation; The blade identification module is configured to collect data on each wind turbine of the wind turbine generator set when it is identified by the semantic-level visual perception channel. Images of the blade's operational status at physical anchor points are collected, and a high-definition reference image library of blades is established. Then, images collected by the inspection drone at the same altitude during a single visual observation are packaged into an image sequence. High-confidence feature anchor frames are selected from the image sequence, and feature matching is performed between the high-confidence feature anchor frames and the high-definition reference image library of blades. Based on the temporal relationship of the images in the same image sequence, the number of the target blade visually observed by the inspection drone is identified.

Images