LSTM (Long Short Term Memory)-based fan blade unmanned aerial vehicle routing inspection path optimization method

Abstract

The invention discloses an LSTM (Long Short Term Memory)-based fan blade unmanned aerial vehicle routing inspection path optimization method. The method comprises the following steps: 1) constructing a map model of a fan blade routing inspection area by adopting a grid method; 2) measuring yaw data of the fan in real time; 3) dynamically updating the map model; 4) taking the real-time yaw angle and the yaw speed of the fan and the historical position of the fan blade as the input of the LSTM model, and predicting the trajectory of the fan blade in the future n time steps through the LSTM; and 5) dynamically planning the inspection path of the unmanned aerial vehicle according to a D * algorithm. Compared with a traditional static path planning method, the method has the advantages that higher inspection precision and higher dynamic adaptability can be achieved, dynamic path optimization is achieved, unnecessary repeated flight is avoided, energy consumption is greatly reduced, and meanwhile the coverage rate and efficiency of inspection are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of wind turbine blade inspection, unmanned aerial vehicle (UAV) path planning and optimization, etc., and particularly relates to a method for optimizing the path of a UAV for inspecting wind turbine blades. Background Art

[0002] As an important form of clean energy, wind power has received extensive attention globally. In recent years, with the continuous expansion of the scale of wind farms and the continuous growth of the installed capacity of wind turbines, the inspection of wind turbine blades, as an important link to ensure the safe operation of wind turbine units, has become increasingly important. Wind turbine blades are the core components for converting wind energy into mechanical energy. They operate in a harsh environment for a long time and are easily affected by external factors such as typhoons, ice and snow, and thunderstorms, which can cause problems such as cracks, corrosion, or wear on the blades, thus threatening the operation safety and reliability of wind power equipment. Therefore, the efficient inspection of wind turbine blades is a key link in the operation and maintenance of wind power equipment.

[0003] In recent years, the technology of autonomous UAV inspection has developed rapidly. Through the application of path planning and optimization algorithms, UAVs can achieve efficient and safe automated inspections. However, existing path planning algorithms (such as regular planning methods like the A* algorithm and Dijkstra algorithm, as well as heuristic optimization methods like genetic algorithms and particle swarm algorithms) perform well in small-scale wind farms, but have problems of insufficient adaptability and limited optimization ability when facing the complex geometric structure and dynamic environmental changes of wind turbine blades. At present, a large number of relevant studies and patents at home and abroad have proposed application solutions for UAVs in the inspection of wind turbine blades.

[0004] Zhu Kaihua proposed an automatic flight path planning method based on a UAV in the paper titled "Research on the Automatic Flight Path Planning of UAVs for Wind Turbine Blade Inspection". This method designs a spiral trajectory and a segmented scanning path according to the geometric characteristics of the blade by establishing a mathematical model of the wind turbine blade, and optimizes the path using an improved A* algorithm to improve the inspection coverage rate and reduce the flight time. Although this method can achieve high inspection efficiency and coverage accuracy in a simulation environment, due to the failure to fully consider the changes in the dynamic environment, its adaptability to actual complex scenarios is limited, and it does not solve the impact of UAV hardware limitations on the inspection effect.

[0005] In the patent application with the application number CN202310636073.9 and the title "A Method for Optimizing the Inspection Path of Wind Turbine Blades by UAV", Chen Jun et al. proposed an inspection path optimization method based on image processing. This method optimizes the inspection path by performing grayscale processing and feature extraction on the blade images collected during the UAV flight, reducing path duplication or omission to improve the inspection efficiency. Although this method can effectively reduce the computational burden of path planning, due to the relatively simple path planning algorithm, it is difficult to achieve global optimization in large-scale wind farms, and at the same time, it has insufficient real-time adaptability to dynamic environments.

[0006] In the patent with the patent number US20150274258A1 and the title "Autonomous Drone Inspection System for Wind Turbines", T.A. Johansen et al. proposed an UAV inspection system based on 3D lidar. This system realizes the automatic inspection of wind turbine blades by generating a point cloud map of the wind turbine blades and their surrounding environment and combining geometric modeling and dynamic path planning algorithms (such as particle swarm algorithm or genetic algorithm). Although this method can adapt to complex inspection environments through high-precision modeling and effectively improve the inspection efficiency, since lidar provides instant point cloud data, although it can very accurately describe the current environmental state, the information it can provide is only the static or approximately static state at the current moment. During the rotation and yaw of the wind turbine blades, lidar cannot predict the specific position or movement trajectory of the future blades in advance, nor can it foresee the instantaneous changes when the wind turbine undergoes yaw. Therefore, relying solely on lidar for obstacle avoidance will cause the UAV to be unable to predict the future obstacle positions in a timely manner, resulting in a collision risk, especially in dynamic environments. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to provide an LSTM-based method for optimizing the inspection path of wind turbine blades by UAV to solve the technical problems that the existing inspection path planning methods for wind turbine blades have obvious deficiencies in aspects such as adaptability to dynamic environments, path optimization efficiency, and energy consumption control.

[0008] The LSTM-based method for optimizing the inspection path of wind turbine blades according to the present invention includes the following steps:

[0009] 1) The map model of the inspection area of the wind turbine blade is constructed by using the grid method, so as to discretize the three-dimensional space of the inspection area of the wind turbine blade into regular two-dimensional grids; the accessibility of the area is expressed by the binary state of each grid node: G(x, y) = 1 indicates passable, and G(x, y) = 0 indicates obstacles or dangerous areas; at the same time, each node records the cumulative cost g(x, y) and the total cost f(x, y) required for path planning; the selection of the grid size G is determined by the formula G = total area of the inspection area / total number of grids; the fixed part of the wind turbine in the map model is marked as a static obstacle, and the dynamic part of the wind turbine is marked as an updatable dynamic area in the initial model;

[0010] 2) The yaw data of the wind turbine is measured in real time, including:

[0011] The nacelle and blades of the wind turbine are scanned in real time by the lidar carried by the UAV to obtain the point cloud data of the wind turbine, and the coordinates (x current , y current ) of the key feature points are extracted, and the real-time yaw angle θ of the nacelle is calculated by the following formula yaw ;

[0012]

[0013] where x reference is the abscissa of the reference point, and y reference is the ordinate of the reference point;

[0014] In a continuous time interval, by comparing the current yaw angle θ yaw,current and the previous yaw angle θ yaw,previous , the yaw speed of the nacelle is calculated:

[0015]

[0016] where Δt is the measurement time interval;

[0017] 3) Dynamically update the map model, including:

[0018] Based on the real-time obtained yaw angle θ yaw and yaw speed ω yaw , dynamically adjust the spatial position and orientation of the wind turbine blade in the map model to generate a dynamic obstacle area; through the rotation matrix R ( θ yaw) transform the initial coordinates (x, y) of the wind turbine blade into the coordinates (x′, y′) in the yaw state:

[0019]

[0020] Combined with the length R blade of the wind turbine blade and the end velocity v blade= ω blade ·R blade , ω blade is the angular velocity of the fan blade, and calculate the radius of the dynamic obstacle area covered by the fan blade:

[0021] R safe = R blade + v blade ·t reaction + d margin (4)

[0022] where t reaction is the response time of the UAV, and d margin is the safety margin; Map the dynamic obstacle area determined according to formula (4) to the map model through rasterization, and perform regional reachability marking according to formula (5):

[0023]

[0024] 4) Use the real-time yaw angle θ yaw , yaw rate ω yaw , and the historical position (x blade , y blade ) of the fan blade as the input of the LSTM model, and predict the trajectory of the fan blade in the next n time steps through LSTM:

[0025]

[0026] Based on the prediction results, generate the dynamic obstacle area formed by the fan blade at the predicted trajectory according to formula (4), and then perform reachability marking according to (5);

[0027] 5) Dynamically plan the inspection path of the UAV according to the D* algorithm. The total cost function Cost(n) in the D* algorithm is defined as:

[0028] Cost(n) = w 1 × Distance(n) + w 2 × Time(n) + w 3 × Safety(n) + w yaw × Yaw - Risk(n) (7)

[0029] where: Distance(n) is the distance cost of the UAV path, and its calculation formula is as follows:

[0030]

[0031] where: x i is the abscissa of the current point, y i is the ordinate of the current point, x jis the abscissa of the next target point, y j is the ordinate of the next target point;

[0032] Time(n) is the path flight time cost, and its calculation formula is as follows:

[0033]

[0034] where, v current = max(v min , v default - k·d risk ); v min is the minimum safe speed of the UAV, d risk is the minimum distance between the UAV and the obstacle area, v default is the default flight speed of the UAV, and k is the speed reduction coefficient;

[0035] Safety(n) is the safety distance cost between the UAV and the dynamic obstacle, and its calculation formula is as follows:

[0036] Safety(n) = max(0, d obstacle - d risk ) (10)

[0037] where, d obstacle = v blade ·t reaction + d margin ;

[0038] Yaw - Risk(n) is the yaw risk cost of whether node n is located in the safe yaw area of the fan, and its calculation formula is as follows:

[0039] Yaw - Risk(n) = LSTM([θ yaw , ω yaw , x blade , y blade ) (11)

[0040] w 1 , w 2 , w 3 , w yaw are the weight coefficients adjusted according to the inspection requirements;

[0041] When the dynamic obstacle area changes, the D* algorithm optimizes the path through the incremental update mechanism, and only recalculates the affected path segments. The update formula is:

[0042] f new (n) = min(f current (n), f neighbor (n)+ c(S neighbor , Scurrent )) (12)

[0043] Among them, f neighbor (n) is the total cost value of neighbor nodes, and its calculation formula is as follows:

[0044] f neighbor (n) = Cost(S neighbor ) (13)

[0045] c(S neighbor , S current ) is the cost from the neighbor node to the current node, and its calculation formula is as follows:

[0046] c(S neighbor , S current ) = Distance(S neighbor , S current ) + ΔCost dynamic (14)

[0047]

[0048] Among them, S neighbor is the neighbor node, S current is the current node, x current is the abscissa of the current node, y current is the ordinate of the current node, x neighbor is the abscissa of the neighbor node, y neighbor is the ordinate of the neighbor node, α is the distance obstacle avoidance cost parameter, β is the yaw risk cost parameter, and d safe is the static safety threshold;

[0049] ΔP = P actual - P predicted (17)

[0050] Among them, P actual is the blade position measured in real time by the lidar carried by the UAV, and P predicte is the predicted wind turbine blade trajectory by the LSTM model; if the deviation |ΔP| exceeds the set threshold, path adjustment is triggered, and by comparing the updated total cost value f new (n) and the current cost value f current (n), the path segment affected by the dynamic obstacle is adjusted, and the dynamic obstacle area is updated again.

[0051] Furthermore, the above-mentioned method for optimizing the inspection path of the wind turbine blade by the UAV based on LSTM also includes optimizing the parameters input into the LSTM model through the following steps to improve the prediction accuracy of the LSTM model for the wind turbine blade trajectory:

[0052] The parameters measured by the lidar and IMU sensors carried by the UAV are fused using Kalman filtering. The state prediction formula is as follows:

[0053]

[0054] Among them, is the state prediction value of the parameter to be optimized at the current moment, x k-1 is the state value of the parameter to be optimized at the previous moment, A is the state transition matrix, u k-1 is the control input, B is the control input matrix, and the parameter to be optimized includes θ yaw and ω yaw ;

[0055] The state update formula is

[0056]

[0057] Among them: z k is the measured value of the parameter to be optimized, H is the observation matrix, K k is the Kalman gain, and its calculation formula is:

[0058]

[0059] Among them, is the prediction error covariance matrix, and R is the measurement noise covariance matrix.

[0060] Furthermore, the method for optimizing the inspection path of the wind turbine blade UAV based on LSTM further includes dynamically adjusting the current flight speed of the UAV through a real-time feedback mechanism. The current flight speed of the UAV is calculated by the following formula:

[0061]

[0062] Advantages of the present invention:

[0063] 1. The method for optimizing the inspection path of the wind turbine blade UAV based on LSTM of the present invention can capture the dynamic state changes of the wind turbine blade in real time through the time series data prediction ability of LSTM, and perform intelligent path optimization for the complex geometric structure and changing wind field environment of the wind turbine blade; compared with the traditional static path planning method, the present invention can achieve higher inspection accuracy and stronger dynamic adaptability.

[0064] 2. The method for optimizing the inspection path of a wind turbine blade UAV based on LSTM of the present invention uses a path planning method that combines LSTM with the D* algorithm, and redesigns the total cost function of the D* algorithm, achieving dynamic path optimization, avoiding unnecessary repeated flights, significantly reducing energy consumption, and simultaneously improving the inspection coverage rate and efficiency. Compared with the high-energy consumption and inefficient path planning problems existing in the prior art, the present invention not only effectively improves the inspection efficiency but also reduces the operation cost.

[0065] 3. The method for optimizing the inspection path of a wind turbine blade UAV based on LSTM of the present invention combines the forward-looking prediction ability of LSTM with real-time data feedback, enabling the UAV to make more accurate autonomous decisions in the complex inspection environment of wind turbine blades. Compared with the traditional inspection method that relies on manual operation, the present invention greatly reduces manual intervention, improves the inspection safety, and avoids potential safety hazards caused by manual operation errors.

[0066] 4. The method for optimizing the inspection path of a wind turbine blade UAV based on LSTM of the present invention optimizes the blade position estimation and yaw angle measurement by introducing Kalman filtering to fuse lidar and IMU sensor data, further improving the accuracy of the inspection path planning and adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of a grid map and possible movement directions of the UAV.

[0068] Figure 2 It is a flowchart of the D* algorithm.

[0069] Figure 3 It is the overall process of optimizing the path of a wind turbine blade inspection UAV based on LSTM. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] The present invention will be further described below with reference to the drawings and embodiments.

[0071] The method for optimizing the inspection path of a wind turbine blade UAV based on LSTM in this embodiment includes the following steps:

[0072] 1) The grid method is used to construct a map model of the inspection area of the wind turbine blade, so as to discretize the three-dimensional space of the inspection area of the wind turbine blade into regular two-dimensional grids. The accessibility of the area is expressed by the binary state of each grid node: G(x, y) = 1 indicates passable, and G(x, y) = 0 indicates obstacles or dangerous areas; at the same time, each node records the cumulative cost g(x, y) and the total cost f(x, y) required for path planning, which is convenient for path optimization calculation. The selection of the grid size G is determined by the formula G = total area of the inspection area / total number of grids. The selection of the grid size G needs to balance the modeling accuracy and the computational complexity. The fixed part of the wind turbine in the map model is marked as a static obstacle, and the dynamic part (blade) of the wind turbine is marked as an updatable dynamic area in the initial model.

[0073] In addition, the initial position and orientation of the wind turbine can be used as the reference direction for modeling. After the flight trajectory of the UAV is discretized by the grid, the motion information on each grid characterizes the flight direction of the UAV on this grid. In theory, there are countless motions of the UAV on each grid, but considering the complexity of model establishment, generally only 8 directions of motion are defined in actual applications, namely: forward, backward, left, right, right front, right rear, left front, left rear, as Figure 1 shown.

[0074] 2) The yaw data of the wind turbine is measured in real time, including:

[0075] The nacelle and blades of the wind turbine are scanned in real time by the lidar carried by the UAV to obtain the point cloud data of the wind turbine, and the coordinates (x current , y current ) of the key feature points are extracted, and the real-time yaw angle θ of the nacelle is calculated by the following formula yaw . The wind turbine yaw angle is the rotation angle of the wind turbine nacelle around the vertical axis relative to the reference direction (such as north).

[0076]

[0077] where x reference is the abscissa of the reference point, and y reference is the ordinate of the reference point;

[0078] Within a continuous time interval, by comparing the current yaw angle θ yaw,current and the yaw angle θ at the previous moment yaw,previous , the yaw speed of the nacelle is calculated:

[0079]

[0080] where Δt is the measurement time interval.

[0081] 3) Dynamically update the map model, including:

[0082] Based on the yaw angle θ obtained in real time yaw and the yaw rate ω yaw , dynamically adjust the spatial position and orientation of the fan blades in the map model to generate a dynamic obstacle area. Through the rotation matrix R ( θ yaw) transform the initial coordinates (x, y) of the fan blade into the coordinates (x′, y′) in the yaw state:

[0083]

[0084] Combined with the length R of the fan blade blade and the tip speed v blade = ω blade ·R blade ω blade is the angular velocity of the fan blade, calculate the radius of the dynamic obstacle area covered by the fan blade:

[0085] R safe = R blade + v blade ·t reaction + d margin (4)

[0086] where t reaction is the response time of the UAV, and d margin is the safety margin; Map the dynamic obstacle area determined according to formula (4) to the map model through rasterization, and perform regional reachability marking according to formula (5):

[0087]

[0088] 4) To cope with the dynamic changes brought about by the yaw of the fan and the rotation of the blades, use the real-time yaw angle θ of the fan yaw , the yaw rate ω yaw , and the historical position (x blade , y blade ) of the fan blade as the input of the LSTM model, and predict the trajectory of the fan blade in the next n time steps through LSTM:

[0089]

[0090] Based on the prediction results, generate the dynamic obstacle area formed by the fan blade at the predicted trajectory according to formula (4), and then perform reachability marking according to (5). Accurately predict the dynamic blade trajectory and the dynamic obstacle area through LSTM, providing key support for UAV path planning and real-time obstacle avoidance.

[0091] 5) Dynamically plan the inspection path of the UAV according to the D* algorithm. The specific process of the D* algorithm is as Figure 2As shown, the G value in the figure is the cost value calculated by the following formula (7), and the F value in the figure is the cost value calculated by the following formula (12).

[0092] In this embodiment, the total cost function Cost(n) in the D* algorithm is defined as:

[0093] Cost(n) = w 1 ×Distance(n) + w 2 ×Time(n) + w 3 ×Safety(n) + w yaw ×Yaw - Risk(n) (7)

[0094] Where: Distance(n) is the distance cost of the UAV path, and its calculation formula is as follows:

[0095]

[0096] Where: x i is the abscissa of the current point, y i is the ordinate of the current point, x j is the abscissa of the next target point, y j is the ordinate of the next target point;

[0097] Time(n) is the path flight time cost, and its calculation formula is as follows:

[0098]

[0099] Where, v current = max(v min , v default - k·d risk ); v min is the minimum safe speed of the UAV, d risk is the minimum distance between the UAV and the obstacle area, v default is the default flight speed of the UAV, and k is the speed reduction coefficient;

[0100] Safety(n) is the safety distance cost between the UAV and the dynamic obstacle, and its calculation formula is as follows:

[0101] Safety(n) = max(0, d obstacle - d risk ) (10)

[0102] Where, d obstacle = v blade ·t reaction + d margin ;

[0103] Yaw - Risk(n) is the yaw risk cost indicating whether node n is located in the yaw safety area of the wind turbine. Its calculation formula is as follows:

[0104] Yaw - Risk(n) = LSTM([θ yaw , ω yaw , x blade , y blade ) (11)

[0105] w 1 , w 2 , w 3 , w yaw are the weight coefficients adjusted according to the inspection requirements;

[0106] When the dynamic obstacle area changes, the D* algorithm optimizes the path through the incremental update mechanism, only recalculating the affected path segments. The update formula is:

[0107] f new (n) = min(f current (n), f neighbor (n) + c(S neighbor , S current )) (12)

[0108] Among them, f neighbor (n) is the total cost value of the neighbor node, and its calculation formula is as follows:

[0109] f neighbor (n) = Cost(S neighbor ) (13)

[0110] c(S neighbor , S current ) is the cost from the neighbor node to the current node, and its calculation formula is as follows:

[0111] c(S neighbor , S current ) = Distance(S neighbor , S current ) + ΔCost dynamic (14)

[0112]

[0113] Among them, S neighbor is the neighbor node, S current is the current node, x current is the abscissa of the current node, y current is the ordinate of the current node, x neighbor is the abscissa of the neighbor node, y neighboris the ordinate of the neighbor node, α is the distance obstacle avoidance cost parameter, β is the yaw risk cost parameter, and d safe is the static safety threshold;

[0114] ΔP = P actual - P predicted (17)

[0115] where P actual is the blade position measured in real time by the lidar carried by the UAV, and P predicte is the predicted wind turbine blade trajectory by the LSTM model; if the deviation |ΔP| exceeds the set threshold, the path adjustment is triggered. By comparing the updated total cost value f new (n) and the current cost value f current (n), the path segment affected by the dynamic obstacle is adjusted, and the dynamic obstacle area is updated. The updated path planning will preferentially avoid the dynamic obstacle area to ensure the flight safety of the UAV.

[0116] In the actual path planning, the D* algorithm preferentially selects the grid node with the smallest total cost value and optimizes the search efficiency through the dynamic sorting of the priority queue (i.e., Figure 2 the OpenList in it); at the same time, the coverage range of the blade trajectory predicted by the LSTM is marked as an impassable area, and the cost value of the risk area is increased through the Yaw-Risk(n) in the path cost function to force the path to avoid the high-risk area. This process significantly improves the safety and dynamic adaptability of the path planning.

[0117] As an improvement to the above embodiment, the method for optimizing the inspection path of the wind turbine blade UAV based on LSTM further includes optimizing the parameters input into the LSTM model through the following steps to improve the prediction accuracy of the LSTM model for the wind turbine blade trajectory:

[0118] The Kalman filter is used to fuse the parameter data measured by the lidar and IMU sensors carried by the UAV. The state prediction formula is:

[0119]

[0120] where is the state prediction value of the parameter to be optimized at the current moment, x k-1 is the state value of the parameter to be optimized at the previous moment, A is the state transition matrix, u k-1 is the control input, B is the control input matrix, and the parameters to be optimized include θ yaw and ω yaw ;

[0121] The state update formula is

[0122]

[0123] where: z k is the measured value of the parameter to be optimized, H is the observation matrix, and K k is the Kalman gain, and its calculation formula is:

[0124]

[0125] where is the predicted error covariance matrix, and R is the measurement noise covariance matrix.

[0126] Through the multi-sensor data fusion of Kalman filtering, the UAV can more accurately estimate the dynamic obstacle area, improving the real-time performance and robustness of path adjustment.

[0127] As an improvement to the above embodiment, the method for optimizing the inspection path of the wind turbine blade UAV based on LSTM further includes dynamically adjusting the current flight speed of the UAV through a real-time feedback mechanism. The current flight speed of the UAV is calculated by the following formula:

[0128]

[0129] Through the above real-time feedback mechanism, the dynamic adaptation ability of the UAV to complex inspection tasks is further improved. Especially in high-risk scenarios such as wind turbine yaw and blade rotation, it can provide reliable safety guarantees.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A wind turbine blade UAV inspection path optimization method based on LSTM, characterized by: The following steps are involved: 1) The grid method is used to construct a map model of the wind turbine blade inspection area, thereby discretizing the three-dimensional space of the wind turbine blade inspection area into a regular two-dimensional grid; the binary state of each grid node is used to express the accessibility of the area: G(x,y)=1 indicates passable, and G(x,y)=0 indicates an obstacle or a dangerous area; at the same time, each node records the cumulative cost g(x,y) and the total cost f(x,y) required for path planning; the selection of the grid size G is determined by the formula G=total area of ​​the inspection area / total number of grids; the fixed part of the wind turbine in the map model is marked as a static obstacle, and the dynamic part of the wind turbine is marked as an updateable dynamic area in the initial model; 2) Real-time measurement of wind turbine yaw data, including: The laser radar carried by the drone scans the wind turbine cabin and blades in real time to obtain the point cloud data of the wind turbine and extract the coordinates of the key feature points (x current ,y current ), and calculate the real-time yaw angle θ of the cabin by the following formula yaw ; Among them, x reference is the horizontal coordinate of the datum reference point, y reference is the ordinate of the datum reference point; In continuous time intervals, by comparing the current yaw angle θ yaw,current and the yaw angle θ at the previous moment yaw,previous , calculate the yaw speed of the cabin: Wherein, Δt is the measurement time interval; 3) Dynamically update the map model, including: Based on the yaw angle θ obtained in real time yaw and yaw speed ω yaw , dynamically adjust the spatial position and orientation of the wind turbine blades in the map model to generate a dynamic obstacle area; by rotating the matrix R ( θ yaw) Transform the initial coordinates (x, y) of the wind turbine blade into the coordinates (x′, y′) in the yaw state: Combined with the length E of the fan blade blade and terminal velocity v blade =ω blade ·R blade ,ω blade is the angular velocity of the fan blades, and the radius of the dynamic obstacle area covered by the fan blades is calculated: R safe =R blade +v blade ·t reaction +d margin (4) Among them, t reaction is the drone reaction time, d margin is the safety margin; the dynamic obstacle area determined by formula (4) is mapped to the map model through rasterization, and the regional accessibility is marked according to formula (5): 4) Based on the real-time yaw angle θ of the wind turbine yaw , yaw speed ω yaw , the historical position of the fan blades (x blade ,y blade ) is used as the input of the LSTM model to predict the trajectory of the wind turbine blades in the next n time steps: Based on the prediction results, the dynamic obstacle area formed by the wind turbine blades at the predicted trajectory is generated according to formula (4), and then the accessibility is marked according to (5); 5) Dynamically plan the inspection path of the drone according to the D* algorithm. The total cost function Cost(n) in the D* algorithm is defined as: Cost(n)=w1×Distance(n)+w2×Time(n)+w3×Safety(n)+w yaw ×Yaw-Risk(n) (7) Where: Distance(n) is the distance cost of the drone path, and its calculation formula is as follows: Where: x i is the horizontal coordinate of the current point, y i is the ordinate of the current point, x j is the horizontal coordinate of the next target point, y j is the ordinate of the next target point; Time(n) is the path flight time cost, and its calculation formula is as follows: Among them, v current =max(v min , v default -k·d risk );v min is the minimum safe speed of the drone, d risk is the minimum distance between the UAV and the obstacle area, v default is the default flight speed of the drone, and k is the speed reduction coefficient; Safety(n) is the safety distance cost between the drone and the dynamic obstacle, and its calculation formula is as follows: Safety(n)=max(0,d obstacle -d risk ) (10) Among them, d obstacle =v blade ·t reaction +d margin ; Yaw-Risk(n) is the yaw risk cost of whether node n is located in the wind turbine yaw safety zone, and its calculation formula is as follows: Yaw-Risk(n)=LSTM([θ yaw ,ω yaw ,x blade ,y blade ]) (11) w1,w2,w3,w yaw is the weight coefficient adjusted according to inspection requirements; When the dynamic obstacle area changes, the D* algorithm optimizes the path through an incremental update mechanism and only recalculates the affected path segments. The update formula is: f new (n)=min(f current (n),f neighbor (n)+c(S neighbor ,S current )) (12) Among them, f neighbor (n) is the total cost of neighbor nodes, which is calculated as follows: f neighbor (n)=Cost(S neighbor ) (13) c(S neighbor ,S current ) is the cost from the neighbor node to the current node, and its calculation formula is as follows: c(S neighbor ,S current )=Distance(S neighbor ,S current )+ΔCost dynamic (14) Among them, S neighbor is the neighbor node, S current is the current node, x current is the horizontal coordinate of the current node, y current is the ordinate of the current node, x neighbor is the horizontal coordinate of the neighbor node, y neighbor is the ordinate of the neighbor node, α is the distance obstacle avoidance cost parameter, β is the yaw risk cost parameter, d safe is the static safety threshold; ΔP=P actual -P predicted (17) Among them, P actual is the blade position measured in real time by the laser radar carried by the UAV, P predicte The wind turbine blade trajectory predicted by the LSTM model; if the deviation |ΔP| exceeds the set threshold, the path adjustment is triggered, and the updated total cost value f is compared new (n) and the current cost value f current (n), adjust the path segments affected by dynamic obstacles and re-update the dynamic obstacle area.

2. The LSTM-based wind turbine blade UAV inspection path optimization method according to claim 1 is characterized in that: It also includes optimizing the parameters of the input LSTM model through the following steps to improve the prediction accuracy of the LSTM model for the wind turbine blade trajectory: Kalman filtering is used to fuse the parameter data measured by the laser radar and IMU sensor carried by the drone, and the state prediction formula is: in, is the state prediction value of the optimized parameter at the current moment, x k-1 is the state value of the optimized parameter at the previous moment, A is the state transfer matrix, u k-1 is the control input, B is the control input matrix, and the optimized parameters include θ yaw and ω yaw ; The state update formula is: Where: z k is the measured value of the optimized parameter, H is the observation matrix, K k is the Kalman gain, and its calculation formula is: in, is the prediction error covariance matrix, and R is the measurement noise covariance matrix.

3. The LSTM-based wind turbine blade UAV inspection path optimization method according to claim 1 or 2 is characterized in that: It also includes dynamically adjusting the current flight speed of the drone through a real-time feedback mechanism. The current flight speed of the drone is calculated by the following formula:

Images