Amphibious unmanned aerial vehicle drifting flight dynamic path multi-step optimization method

Abstract

The invention discloses a multi-step optimization method for a drifting flight dynamic path of an amphibious unmanned aerial vehicle, and relates to the technical field of unmanned aerial vehicles. Riverway environment information is acquired, and a local map containing obstacle information is constructed; setting an operation starting point and a plurality of target points based on the local map, and generating an optimal global route plan by adopting an ant colony optimization algorithm; according to obstacle information in the local map, using a bidirectional A * algorithm to simulate the preliminary path, and generating a flight path feasible in dynamics; predicting the drifting direction of the unmanned aerial vehicle in real time to obtain a drifting value; establishing a YOLO model to obtain the number of obstacles; objective weighting is carried out through an entropy weight method according to the collected data, a comprehensive score is calculated, an intelligent decision-making takeoff strategy is carried out when a takeoff threshold value is exceeded, and segmented motion conversion is completed. According to the invention, hierarchical path optimization is carried out according to different motion modes, the autonomous intelligent operation capability of the amphibious unmanned aerial vehicle is improved, the measurement task is better completed in a complex environment, and a monitoring basis is provided for digital river construction.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and more specifically, to a multi-step optimization method for the dynamic path of an amphibious unmanned aerial vehicle during drifting flight. Background Art

[0002] Currently, with the increasing demand for river channel measurement, the measurement methods are also constantly updated. The water-air amphibious unmanned aerial vehicle is widely used because it can operate autonomously in the air and on the water, and has characteristics such as high-speed mobility and low cost, and the prospect is very promising. During the mapping process, the water-air amphibious unmanned aerial vehicle often needs to go to multiple points for mapping. Due to the particularity of cross-latitude movement, a path optimization method based on the amphibious unmanned aerial vehicle needs to be studied.

[0003] The invention patent CN118999608A introduces a path planning method for a land-air amphibious unmanned aerial vehicle based on an improved ant colony algorithm. By introducing parameters such as a time decay factor and a mirror gravitational field, the path exploration ability of the amphibious unmanned aerial vehicle is improved. Another invention patent CN114089762 proposes a path planning method based on reinforcement learning. By modeling the Markov decision process (MDP) of the path planning of the water-air amphibious unmanned vehicle and combining the reinforcement learning algorithm, the global path planning is completed.

[0004] However, the existing amphibious path planning uniformly models the air and water movement modes, ignoring the particularity of the two movement modes, which may lead to certain errors in the results obtained by the algorithm.

[0005] Therefore, how to adopt different path optimization methods according to the characteristics of the flight and floating movement modes is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a multi-step optimization method for the dynamic path of an amphibious unmanned aerial vehicle during drifting flight, which fully considers the particularity of the air and water movement modes, can achieve a path planning with low energy consumption at high altitude and almost zero energy consumption on water, and is of great significance for improving the autonomous operation efficiency and intelligent decision-making of the unmanned aerial vehicle.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A multi-step optimization method for the dynamic path of an amphibious unmanned aerial vehicle during drifting flight, comprising:

[0009] Step 1: Obtain river channel environment information, and construct a local map containing obstacle information based on the river channel environment information;

[0010] Based on the local map, set the operation starting point and multiple target points, and use the ant colony optimization algorithm to solve the target point sequence problem and generate an optimal global route plan;

[0011] Step 2: According to the obstacle information in the local map, use the bidirectional A* algorithm to simulate the preliminary path and generate a dynamically feasible flight path;

[0012] Step 3: First, perform drifting forward, and collect data in real time through the fuselage sensors, input it into the improved CNN-BiGRU-Attention model to predict the drifting direction of the UAV in real time, and obtain the drift value;

[0013] Step 4: Establish a YOLO model, feed historical data to train an identification model dedicated to identifying water obstacles, and obtain the number of obstacles;

[0014] Step 5: According to the drift value and the number of obstacles obtained in Step 3 and Step 4, and combined with its own real-time power, preset task time, and target monitoring distance, objectively assign weights through the entropy weight method, calculate the comprehensive score, exceed the takeoff threshold, and implement an intelligent decision-making takeoff strategy to complete the segmented motion conversion.

[0015] Optionally, Step 1 includes the following contents:

[0016] Hover the amphibious UAV above the area to be measured, adjust the shooting height and parameters according to the size of the measurement area, obtain a two-dimensional grid map of the measurement area, and record the known obstacles in the picture;

[0017] Then, according to the task requirements, set the starting point and multiple measurement points {A, B, C,...} on the map;

[0018] Through the ant colony optimization algorithm, perform the optimal path planning for all target points to determine the order of each target measurement; among them, the ant colony optimization algorithm simulates the behavior of ants searching for the optimal path between different measurement points, and automatically searches for the shortest path or the lowest-cost path under the satisfaction of the constraint conditions; each ant selects a suitable path according to the cost of the path and the pheromone concentration, and gradually approaches the global optimal path through the local and global update processes of the pheromone.

[0019] Optionally, the path selection expression formula is:

[0020]

[0021] where: P ij is the selection probability of the ant from node i to node j, τ ij is the pheromone concentration from node i to node j, η ij is the heuristic information from node i to node j, usually a measure of the feasibility of the path, and α and β are parameters that control the weights of the pheromone and the heuristic information respectively.

[0022] Optionally, global pheromone update occurs after all ants complete a search. By updating the pheromone concentration of the optimal path, the high-quality path is further strengthened. The update formula is as follows:

[0023]

[0024] Where: ρ is the pheromone evaporation coefficient, τ ij (t) is the pheromone concentration from node i to node j at time t; is the pheromone increment calculated according to the optimal path. The increment of the optimal path is proportional to the quality of the path;

[0025]

[0026] Where Q is the pheromone constant, L best represents the distance from each node to the end point.

[0027] Optionally, step 2 includes the following content:

[0028] Step 2.1: Initialize two open lists openlist1 and openlist2, and two closed lists closelist1 and closelist2. Add the starting node S to the open list openlist1, and add the target node G to the open list openlist2 for search expansion from the starting point and the target point respectively;

[0029] Step 2.2: Determine whether the current node n meets the target proximity condition, that is, whether the current node is close enough to the target node G or coincides with the nodes on the reverse search front. If it meets the condition, try to directly connect the current node n to the target node G and detect whether there is a collision with the obstacle. If the path has no collision, set the target node G as the child node of the current node and add it to the closed list closelist1. The search ends and jumps to step 2.6. Otherwise, continue to execute step 2.3;

[0030] Step 2.3: Perform obstacle and boundary detection on the 8-neighborhood nodes of the current node n. If the node is feasible and not in the closed list and not in the open list, add it to the open list, and calculate and update its cost G(n), H(n), and F(n). If it is already in the open list and can be optimized by using the current node as the parent node, update the cost of the node;

[0031] Step 2.4: Add the current node n to the closed list closelist1;

[0032] Step 2.5: Execute the same search process as steps 2.2 to 2.4 in the open list openlist2 in the direction of the target point to ensure that the bidirectional search is synchronized;

[0033] Step 2.6: Starting from the encountered node, trace back along the parent nodes in sequence to generate the complete path P, and finally obtain the optimized preliminary planned path.

[0034] Optionally, it also includes the optimization of the heuristic cost function for search:

[0035] F(n) = G(n) + H(n)

[0036]

[0037] where n represents the current node, G(n) represents the actual cost from the start point of the path to the current node n, H(n) represents the estimated cost from the current node to the target point, and (x n , y n ) represents the position coordinates of the current node, and (x g , y g ) represents the position coordinates of the search target point.

[0038] Optionally, step 3 specifically includes the following contents:

[0039] Step 3.1: Data collection. Set the data collected by the UAV sensor within a period of time as the original data set. The collected data includes drift values, longitude, latitude, water speed, water direction, wind speed, wind direction, wind level, and water depth. Then, perform data preprocessing to remove outliers and fill in missing values by the moving average method.

[0040] Step 3.2: Deep learning network module, including the CNN algorithm, BiGRU algorithm, and Attention layer. The CNN algorithm can extract data features and consists of two convolutional layers, a pooling layer, and a ReLU function. The role of the convolutional layer is to perform convolutional operations on the input data to extract local features and help the model identify important patterns in the data. The pooling layer is used to reduce the size of the output of the convolutional layer and reduce the spatial dimension of the data by selecting the maximum value in the local area. The BiGRU algorithm uses bidirectional GRUs to capture the forward and backward information of time series data and can link the upper and lower data in the feature extraction area. The Attention layer re-weights important data so that the formed model can complete ultra-short-term predictions and output predicted drift values.

[0041] Step 3.3: Network optimization module. Introduce the particle swarm optimization algorithm to adjust the hyperparameters of the network model, find the optimal combination of hyperparameters, and improve the accuracy of the prediction model.

[0042] Optionally, step 4 includes the following contents:

[0043] Step 4.1: Historical data collection. Collect data on floating objects and ice floes on water, preprocess the data to conform to the size of the training detection map, label the data set using labelImg, and divide the labeled data set into a training set, a validation set, and a test set;

[0044] Step 4.2: YOLOv8 framework module, which consists of building a Backbone layer, a Neck layer, and a Head layer, integrating a deformable convolution module, a PAN-FPN module, and a multi-head attention mechanism; the Backbone layer is responsible for extracting low-level features and high-level features from the input image; the Neck layer processes the feature maps from the Backbone layer and combines multi-scale information; the Head layer is responsible for generating the final detection results;

[0045] Step 4.3: Train the YOLOv8 water floating object detection model established in Step 4.2;

[0046] Real-time receive the video information obtained by the front camera, and use the trained YOLOv8 water floating object detection model to real-time identify the water floating objects in the video and output the recognition results.

[0047] Optionally, the said Step 5 includes the following content:

[0048] Receive the drift value obtained from the processing in Step 3 and the number of obstacles obtained from the processing in Step 4, combine five data including its own real-time power, preset task time, and target monitoring distance, evaluate the weights of each index through the entropy weight method, and increase the weight of the obstacles, and real-time calculate the comprehensive scores of multiple points on the route, and select the points with scores exceeding the threshold of 0.8 as the take-off points to complete the intelligent take-off decision.

[0049] Optionally, the entropy value calculation formula is as follows:

[0050]

[0051] where H j represents the entropy value of the jth index, p ij represents the normalized probability of the jth index, k is a constant used to normalize the entropy value, and m represents the number of samples.

[0052] Through the above technical solutions, compared with the prior art, the present invention discloses and provides a multi-step optimization method for the dynamic path of an amphibious drone's drift flight, which has the following beneficial effects:

[0053] 1. The present invention makes full use of the amphibious motion characteristics of the amphibious UAV, replaces the flight motion with the zero-energy-consuming part of the drift motion, combines drifting and flying, reduces energy consumption, and improves the working endurance and mapping range. And it optimizes the path according to different motion modes, improves the autonomous intelligent operation ability of the amphibious UAV, and better completes multiple measurement tasks in complex environments, providing a monitoring basis for the digital river construction.

[0054] 2. The present invention makes full use of water power for drifting, enabling the UAV to move on the water surface without additional power drive, thereby significantly reducing energy consumption in this mode, greatly improving the working endurance and mapping range, and establishing a prediction model to predict the future drift value in real time, avoiding the deviation of the unpowered drift from the working route.

[0055] 3. The present invention adopts the entropy weight method to calculate the comprehensive score of key parameters in real time, realizes the adaptive comprehensive evaluation and real-time decision-making in multiple scenarios, ensures that the UAV makes the optimal or approximate optimal takeoff decision quickly in complex environments, intelligently divides the drift voyage and the flight voyage, and improves the intelligent planning ability of the autonomous operation of the amphibious UAV.

[0056] 4. The model trained by the present invention has excellent generalization ability, can make full use of the existing prior knowledge, and quickly plan a suitable path in complex environments. There is no need to consume additional search time, thus significantly improving the planning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0058] Figure 1 It is a flowchart of a multi-step optimization method for the drift flight dynamic path of an amphibious UAV provided by the present invention;

[0059] Figure 2.1 It is a graph of the ant colony optimization algorithm provided by the present invention to generate the optimal global route planning graph;

[0060] Figure 2.2 It is a schematic diagram of the average distance and the shortest distance in the ant colony optimization algorithm graph provided by the present invention;

[0061] Figure 3 It is a graph of the two-way A* algorithm provided by the present invention to generate a dynamically feasible flight path;

[0062] Figure 4.1A comparison diagram of the improved CNN-BiGRU-Attention model provided by the present invention and the common model;

[0063] Figure 4.2 The structure diagram of the improved CNN-BiGRU-Attention model provided by the present invention;

[0064] Figure 5 The present invention provides a YOLOv8 training flow chart for identifying floating objects on water. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The embodiment of the present invention discloses a multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle, such as Figure 1 As shown, including:

[0067] Step 1: Obtain river environment information and build a local map containing obstacle information based on the river environment information; based on the local map, set the operation starting point and multiple target points, use the ant colony optimization algorithm to solve the target point sequence problem, and generate the optimal global route planning;

[0068] Specifically, pictures of the measured river channel are taken, multiple target points are selected, and the ant colony optimization algorithm is used to find the waypoint sequence with the shortest trajectory time.

[0069] Drone preparation: Choose an amphibious drone that has the ability to float on water and fly in the air, and make sure it is equipped with sensors such as a single-beam depth sounder and camera equipment (Hikvision camera).

[0070] Amphibious UAV operation: The operator controls the UAV to fly over the area to be tested through the remote control, sets the flight altitude to 100 meters, sets the resolution of the camera to 1080p, the frame rate to 30fps, uses RTK+GPS precise positioning, uses the camera to shoot the river map during the flight, and records the obstacles in the area. The acquired pictures are uploaded to the control system of the UAV, and a two-dimensional grid map is generated based on the river pictures taken, and the coordinates of the obstacle positions are given.

[0071] Global waypoint planning: Set the starting point and operation point {A, B, C, ...} on the map, initialize the ant colony algorithm path planning parameters, the corresponding parameters include the ant colony population number PopNum, the maximum number of iterations iter max, the expected heuristic factor β, the pheromone heuristic factor α, the pheromone constant Q, the time decay factor β t , the gravitational constant G 0 etc. Simulate the behavior of ants searching for the optimal path between different measurement points. Randomly place ants at measurement points {A, B, C, …}, randomly select a route through the roulette wheel algorithm, and update the pheromone in the route. Through the process of local and global pheromone updates, the ant colony will gradually optimize the search process and finally converge to the globally optimal path, as Figure 2.1 and Figure 2.2 shown for the path planning implemented by the algorithm. The state transition (path selection) formula of the ant colony algorithm:

[0072]

[0073] where: P ij is the selection probability of the ant from node i to node j, τ ij is the pheromone concentration from node i to node j, η ij is the heuristic information from node i to node j, usually a measure of the feasibility of the path, and α and β are parameters that control the weights of the pheromone and heuristic information respectively.

[0074] Global pheromone update occurs after all ants complete a search. By updating the pheromone concentration of the optimal path, the high-quality path is further strengthened; the update formula is as follows:

[0075]

[0076] where: ρ is the pheromone evaporation coefficient, τ ij (t) is the pheromone concentration from node i to node j at time t; is the pheromone increment calculated according to the optimal path, and the increment of the optimal path is proportional to the quality of the path (such as the reciprocal of the path length);

[0077]

[0078] where, Q is the pheromone constant, L best represents the distance from each node to the end point.

[0079] Step 2: According to the obstacle information in the local map, use the bidirectional A* algorithm to simulate the preliminary path and generate a dynamically feasible flight path;

[0080] Specifically, on each flight path, mark the obstacle points recorded in Step 1, and use the bidirectional A* algorithm to generate each feasible flight track as Figure 3 .

[0081] Determine the starting node of the flight path (map ij) = S and the target node (map i j) = G, directly connect and perform collision detection. If there is no obstacle, the search ends and step 2.6 is executed; otherwise, step 2.1 is executed.

[0082] Step 2.1: Initialize two open lists (openlist1), (openlist2) and two closed lists (closelist1), (closelist2). Add the starting node S to the open list (openlist1) and the target node G to the open list (openlist2) for search expansion from the starting point and the target point respectively.

[0083] Step 2.2: Determine whether the current node n meets the target proximity condition, that is, whether the current node is close enough to the target point G or coincides with the nodes on the reverse search front. If it meets the condition, try to directly connect the current node n to the target node G and detect whether there is a collision with an obstacle; if the path has no collision, set the target node G as the child node of the current node and add it to the closed list (closelist1), the search ends, and jump to step 2.6; otherwise, continue to execute step 2.3.

[0084] Step 2.3: Perform obstacle and boundary detection on the 8-neighborhood nodes of the current node n. If the node is feasible and not in the closed list and not in the open list, add it to the open list, calculate and update its cost G(n), H(n) and F(n); if it is already in the open list and can be optimized by using the current node as the parent node, update the cost of this node.

[0085] Step 2.4: Add the current node n to the closed list (closelist1).

[0086] Step 2.5: Execute the same search process as steps 2.2 to 2.4 in the open list (openlist2) in the direction of the target point to ensure that the bidirectional search is synchronized.

[0087] Step 2.6: Starting from the meeting node, trace back along the parent nodes in turn to generate the complete path P, and finally obtain the optimized preliminary planning path.

[0088] Among them, the optimization of the heuristic cost function for search:

[0089] F(n) = G(n) + H(n)

[0090]

[0091] Among them, n represents the current node, G(n) represents the actual cost from the starting point of the path to the current node n, H(n) represents the estimated cost from the current node to the target point, (xn , y n ) represents the coordinates of the current node position, (x g , y g ) represents the coordinates of the search target point.

[0092] Step 3: First, perform drifting forward, and collect data in real time through the fuselage sensors, and input it into the improved CNN-BiGRU-Attention model to predict the drifting direction of the UAV in real time, obtaining the drift value;

[0093] Step 3.1: Data collection. Set the data collected by the UAV sensors within a period of time as the original data set. The collected data includes drift values, longitude, latitude, water speed, water direction, wind speed, wind direction, wind level, and water depth. Among them, high-altitude cameras and floating methods are used to obtain river flow velocity and direction data. A single-beam sounding instrument is used to collect water depth data. Wind direction, wind speed, and wind level data are collected through wind speed sensors. Drift values, longitude, and latitude data are obtained through RTK and GPS. All data is collected and uploaded every 5 seconds.

[0094] Then, perform data preprocessing. Remove outliers and fill in missing values through the moving average method, and eliminate the influence of differences between different dimensions through the data normalization unit.

[0095] Establish an improved CNN-BiGRU-Attention drift prediction model, and input the drift values, longitude, latitude, water speed, water direction, wind speed, wind direction, wind level, and water depth information obtained during the operation of the amphibious UAV in real time for the first 300s, and output the drift value for the next 30s. Predict the future drift value of the amphibious UAV in the river through historical data and predict the drifting direction of the UAV.

[0096] Step 3.2: The deep learning network module such as Figure 4.1 , Figure 4.2 , includes the CNN algorithm, BiGRU algorithm, and Attention layer. The CNN algorithm can extract data features and consists of two convolutional layers, a pooling layer, and a ReLU function. The role of the convolutional layer is to perform convolutional operations on the input data to extract local features and help the model identify important patterns in the data; while the pooling layer is used to reduce the size of the output of the convolutional layer, and reduces the spatial dimension of the data by selecting the maximum value (max pooling) in the local area, making the model more efficient in processing and helping to prevent overfitting. The BiGRU algorithm uses bidirectional GRUs to capture the forward and backward information of time series data, and can link the upper and lower data in the feature extraction area to better understand the context information of the data. The Attention layer re-weights important data to improve accuracy, enabling the formed model to complete ultra-short-term predictions and output predicted drift values.

[0097] Step 3.3: The network optimization module introduces the particle swarm optimization algorithm to adjust the hyperparameters of the network model, such as the learning rate, the number of GRU neurons, the key-value size of the attention mechanism, and the convolution kernel size, etc., to find the optimal combination of hyperparameters and improve the accuracy of the prediction model.

[0098] Step 4: Establish a YOLO model, feed historical data to train an identification model dedicated to identifying water obstacles, and obtain the number of obstacles.

[0099] Specifically, in Step 4.1: Historical data collection, a total of 2,400 images of waterborne floating objects and ice floes are collected, and the labeled dataset is divided into a training set, a validation set, and a test set. The data is preprocessed to conform to the size of the training detection map of 416 * 416 pixels, and the labellmg tool is used to label the dataset to record information such as the position, size, and attributes of the targets in the images.

[0100] Step 4.2: The YOLOv8 framework module is composed of building a Backbone layer, a Neck layer, and a Head layer, integrating a deformable convolution module, a PAN-FPN module, and a multi-head attention mechanism. The Backbone layer is mainly responsible for extracting low-level features and high-level features from the input image. The main task of the Neck layer is to further process the feature maps from the Backbone layer and combine multi-scale information. The design of this layer usually uses feature fusion technology to enable the network to better perform object detection at different scales. The Head layer is the last layer of the YOLOv8 model, mainly responsible for generating the final detection results. In this layer, YOLOv8 adopts Decoupled-Head, separating the classification and detection heads, making the training and inference of the network more efficient.

[0101] Step 4.3: Input the dataset obtained in Step 4.1 into the YOLOv8 model for training, set 100 training epochs, and the batch size is 16 to obtain a YOLOv8 waterborne floating object detection model. Real-time receive the video information obtained by the front-view camera, and use the trained YOLOv8 waterborne floating object detection model to identify the waterborne floating objects in the video in real time, output the recognition results, draw the filtered detection boxes on the video frames, and label the object category, confidence, and spatial position of the object, etc. in the upper left corner of each box to display the results of object detection. Finally, output these video frames with detection boxes and annotation information and play them in real time for display. The schematic diagram is as Figure 5 shown.

[0102] Step 5: According to the drift values and the number of obstacles obtained in Step 3 and Step 4, and in combination with its own real-time power, preset task time, and target monitoring distance, objectively assign weights through the entropy weight method, calculate the comprehensive score, and if it exceeds the takeoff threshold, implement an intelligent decision-making takeoff strategy to complete the segmented motion conversion.

[0103] Specifically, receive the drift value obtained from the processing in Step 3 and the number of obstacles obtained from the processing in Step 4, combine the five data of its own real-time power, preset task time, and target monitoring distance, evaluate the weights of each index through the entropy weight method, and introduce λ j adjustment coefficient to increase the weight of obstacles, calculate the comprehensive score of the next 30s in the route in real time, select the point with a score exceeding the threshold of 0.8 as the takeoff point, realize intelligent takeoff decision-making, and complete the drift-flight conversion.

[0104] The entropy value calculation formula is as follows:

[0105]

[0106] where H j represents the entropy value of the j-th index, and p ij represents the normalized probability of the j-th index. k is a constant used to normalize the entropy value, and m represents the number of samples.

[0107]

[0108] where ω j is the weight of the j-th index.

[0109] Through the above detailed model establishment, the present invention improves the intelligent planning ability of autonomous operation of the amphibious UAV, greatly reduces the working energy consumption and expands the working area, providing convenience for river monitoring and management.

[0110] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0111] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-step optimization method for the dynamic path of an amphibious unmanned aerial vehicle drift flight, characterized in that: include: Step 1: Obtain river environment information and build a local map containing obstacle information based on the river environment information; Based on the local map, the operation starting point and multiple target points are set, and the ant colony optimization algorithm is used to solve the target point sequence problem and generate the optimal global route planning; Step 2: Based on the obstacle information in the local map, use the bidirectional A* algorithm to simulate the preliminary path and generate a dynamically feasible flight path; Step 3: Drift forward first, collect data in real time through the fuselage sensor, input it into the improved CNN-BiGRU-Attention model to predict the drift direction of the drone in real time and obtain the drift value; Step 4: Establish a YOLO model, feed historical data to train a recognition model specifically for identifying water obstacles, and obtain the number of obstacles; Step 5: Based on the drift value and number of obstacles obtained in Steps 3 and 4, and combined with its own real-time power, preset mission time, and target monitoring distance, objective weighting is performed through the entropy weight method, and the comprehensive score is calculated. When the takeoff threshold is exceeded, an intelligent decision-making takeoff strategy is implemented to complete the segmented motion conversion.

2. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 1 is characterized in that: The step 1 includes the following contents: Hover the amphibious drone above the area to be measured, adjust the shooting height and parameters according to the size of the measurement area, obtain a two-dimensional grid map of the measurement area, and record known obstacles in the picture; Then, according to the task requirements, set the starting point and multiple measurement points {A, B, C, ...} in the map; The ant colony optimization algorithm is used to plan the optimal path for all target points and determine the measurement order of each target. The ant colony optimization algorithm simulates the behavior of ants in finding the optimal path between different measurement points, and automatically searches for the shortest path or the lowest cost path under the constraints. Each ant selects a suitable path based on the cost of the path and the pheromone concentration, and gradually approaches the global optimal path through the local and global update process of pheromones.

3. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 2 is characterized in that: The path selection expression formula is: Where: P ij is the probability of an ant choosing from node i to node j, τ ij is the pheromone concentration from node i to node j, η ij is the heuristic information from node i to node j, usually a feasibility measure of the path, α and β are parameters that control the weights of pheromone and heuristic information, respectively.

4. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 2 is characterized in that: Global pheromone update occurs after all ants complete a search. By updating the pheromone concentration of the optimal path, the high-quality path is further strengthened. The update formula is as follows: Where: ρ is the pheromone volatility coefficient, τ ij (t) is the pheromone concentration from node i to node j at time t; It is the pheromone increment calculated based on the optimal path. The increment of the optimal path is proportional to the quality of the path. Among them, Q is the pheromone constant, L best Indicates the distance from each node to the end point.

5. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 1 is characterized in that: The step 2 includes the following contents: Step 2.1: Initialize two open lists openlist1 and openlist2 and two closed lists closelist1 and closelist2, add the starting point node S to the open list openlist1, and add the target node G to the open list openlist2, for searching and expanding from the starting point and the target point respectively; Step 2.2: Determine whether the current node n meets the target approach condition. If so, try to directly connect the current node n with the target node G and detect whether there is a collision with an obstacle; If there is no collision on the path, set the target node G as the child node of the current node and add it to the close list closelist1. The search ends and jumps to step 2.

6. Otherwise, continue to step 2.

3. Step 2.3: Perform obstacle and boundary detection on the 8 neighboring nodes of the current node n. If the node is feasible and is not in the closed list and is not in the open list, add it to the open list, calculate and update its cost G(n), H(n) and F(n); if it is already in the open list and can be better optimized by using the current node as a parent node, update the cost of the node; Step 2.4: Add the current node n to the close list closelist1; Step 2.5: Execute the same search process as steps 2.2 to 2.4 in the open list openlist2 in the direction of the target point to ensure that the two-way search is performed synchronously; Step 2.6: Starting from the node where we meet, trace back along the parent nodes one by one to generate a complete path P, and finally obtain the optimized preliminary planning path.

6. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 5 is characterized in that: Also included is a heuristic cost function optimization for searching: F(n)=G(n)+H(n) Where n represents the current node, G(n) represents the actual cost from the starting point of the path to the current node n, H(n) represents the estimated cost from the current node to the target point, (x n ,y n ) represents the current node position coordinates, (x g ,y g ) represents the search target point location coordinates.

7. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 1 is characterized in that: The step 3 specifically includes the following contents: Step 3.1: Data collection: The data collected by the drone sensor over a period of time is set as the original data set. The collected data includes drift value, longitude, latitude, water speed, water direction, wind speed, wind direction, wind level, and water depth. Then, data preprocessing is performed to remove outliers and fill in missing values ​​through the average moving method. Step 3.2: Deep learning network module, including CNN algorithm, BiGRU algorithm and Attention layer; CNN algorithm can extract data features, which consists of two convolutional layers, pooling layer and ReLU function. The function of convolutional layer is to perform convolution operation on input data, extract local features from it, and help the model identify important patterns in the data; the pooling layer is used to reduce the size of the output of convolutional layer, and reduce the spatial dimension of data by selecting the maximum value in the local area; BiGRU algorithm uses bidirectional GRU to capture the forward and backward information of time series data, and can link the upper and lower data of feature extraction area; Attention layer reweights important data, so that the formed model can complete ultra-short-term prediction and output prediction drift value; Step 3.3: In the network optimization module, the particle swarm optimization algorithm is introduced to adjust the hyperparameters of the network model, find the optimal hyperparameter combination, and improve the accuracy of the prediction model.

8. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 1 is characterized in that: The step 4 comprises the following contents: Step 4.1: Historical data collection, collect floating objects and ice data on the water, and pre-process the data to meet the training test image size, use labellmg to annotate the data set, and divide the annotated data set into training set, validation set and test set; Step 4.2: YOLOv8 framework module, which is composed of Backbone layer, Neck layer and Head layer, integrates variable convolution module, PAN-FPN module and multi-head attention mechanism; Backbone layer is responsible for extracting low-level features and high-level features from the input image; The Neck layer processes the feature maps from the Backbone layer and combines multi-scale information; the Head layer is responsible for generating the final detection results; Step 4.3: Train the YOLOv8 floating object detection model established in step 4.2; Receive video information obtained by the front-view camera in real time, identify floating objects in the video in real time through the trained YOLOv8 floating object detection model, and output the recognition results.

9. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 1 is characterized in that: The step 5 comprises the following contents: Receive the drift value obtained in step 3 and the number of obstacles obtained in step 4, combine the real-time power, preset mission time, target monitoring distance and other five data, evaluate the weight of each indicator through the entropy weight method, increase the obstacle weight, calculate the comprehensive score of multiple points in the route in real time, select the point with a score exceeding the threshold of 0.8 as the take-off point, and complete the intelligent take-off decision.

10. The multi-step optimization method for the dynamic path of drift flight of an amphibious unmanned aerial vehicle according to claim 9 is characterized in that: The entropy calculation formula is as follows: Among them, H j represents the entropy value of the jth indicator, p ij represents the normalized probability of the jth indicator, k is a constant used to normalize the entropy value, and m represents the number of samples.

Images