Warehouse unmanned aerial vehicle inspection path planning method based on graph neural network

Abstract

The invention discloses a warehouse unmanned aerial vehicle inspection path planning method based on a graph neural network, and the method comprises the following steps: S1, building warehouse environment data, and carrying out the data preprocessing of the warehouse environment data; s2, modeling by applying a graph convolutional network model in combination with warehouse environment data; s3, the unmanned aerial vehicle starts an inspection task according to the initial path planning model; s4, adopting a water wave optimization algorithm, and dynamically adjusting a path weight by using sensor data; s5, simulating high-priority paths of the wave crest area, and preferentially selecting the paths for routing inspection; s6, generating a final inspection path through multiple iterations; and S7, repeatedly executing the steps S4-S6, and continuously adjusting the inspection path by the unmanned aerial vehicle. According to the method, a graph convolutional network model, a water wave optimization algorithm and a path priority propagation technology are combined, and warehouse unmanned aerial vehicle inspection path planning based on the graph neural network is realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent path planning, and particularly to a method for planning the inspection path of a warehouse drone based on a graph neural network. Background Art

[0002] In the field of modern warehousing logistics, with the continuous improvement of the automation level, the drone inspection technology in warehouses has gradually been widely applied. Drone inspection can effectively improve the efficiency and accuracy of warehouse management, especially in large and complex warehouse environments. Traditional inspection methods usually rely on manual or ground robots for path planning and task execution. However, these methods often have problems such as inaccurate path planning, poor flexibility, and low efficiency. With the continuous progress of artificial intelligence technology, as an emerging deep learning method, the graph neural network has achieved remarkable success in processing graph-structured data, and has demonstrated powerful capabilities especially in fields such as path planning and environmental modeling. Therefore, how to combine the graph neural network with drone inspection technology to achieve more efficient and intelligent inspection path planning has become a hot issue in the current research on warehouse drone inspection technology.

[0003] Existing methods for planning the inspection path of warehouse drones mainly rely on preset fixed paths or traditional optimization algorithms such as ant colony algorithms and genetic algorithms. These algorithms usually perform path calculations based on certain assumptions and empirical rules. Although they can effectively complete tasks in some cases, in the actual complex warehouse environment, these methods usually have problems such as low path planning accuracy, low efficiency, and poor adaptability. Especially in large-scale warehousing environments, due to the complexity, dynamic changes, and uncertainties of the environment, traditional path planning methods often cannot adjust the path in a timely manner, resulting in being unable to adapt to real-time environmental changes, and unable to effectively avoid obstacles and avoid excessive energy consumption.

[0004] In addition, most path planning methods in the prior art do not fully consider the energy consumption of drones and the problem of obstacle avoidance. The results of path planning often only consider the shortest path or the least time, while ignoring the energy consumption of drones and the problem of obstacle avoidance during the actual execution process. There are usually many static and dynamic obstacles in the warehouse environment, such as shelves, transportation equipment, personnel, etc. These obstacles will frequently interfere with the flight path of the drone. Traditional path planning methods do not fully consider the impact of dynamic environmental changes on path selection. Therefore, it is impossible to ensure the smooth completion of drone tasks during the actual execution process.

[0005] In addition, most traditional path planning methods rely on a single algorithm and do not have sufficient flexibility and adaptability to achieve adaptive adjustment in complex environments. There are not only structural complexities in the warehouse, such as diverse shelf and aisle layouts, but also the dynamics of environmental changes, such as temperature, humidity variations, and personnel movement, etc. These all require the drone to have a strong adaptive ability. Traditional algorithms are usually optimized only for static scenarios, lacking the utilization of real-time sensor data and unable to adjust the path according to the real-time changes of the environment.

[0006] As an important development direction in the field of deep learning in recent years, graph convolutional networks have made breakthrough progress in the modeling and reasoning of graph-structured data. Graph convolutional networks aggregate the features of nodes and their neighbor nodes through graph convolution operations, can effectively model the relationships between nodes, and have strong expressive and reasoning abilities. The application of graph convolutional networks in the path planning of warehouse drones can accurately capture the complex relationships between nodes by constructing a graph model of the warehouse environment and achieve the optimization of path planning. Compared with traditional algorithms, graph convolutional networks have stronger adaptability and can dynamically adjust the path planning to cope with different warehouse environments and real-time changes.

[0007] However, the existing path planning methods based on graph convolutional networks still face certain challenges. Although graph convolutional networks can process complex graph-structured data and optimize path planning to a certain extent, traditional graph convolutional network models are usually static, lacking the real-time feedback and adaptive adjustment ability to environmental dynamic changes. The results of path planning are often based on static graph models and do not fully utilize real-time sensor data for dynamic adjustment. Therefore, during the actual execution process, the path planning results of graph convolutional networks may be interfered by real-time environmental changes, leading to a decline in the accuracy and efficiency of path planning.

[0008] To solve these problems, the present invention proposes a drone inspection path planning method that combines graph convolutional networks with the water wave optimization algorithm. By introducing the water wave optimization algorithm on the basis of graph convolutional networks, simulating the process of water wave propagation, and dynamically adjusting the edge weights in the path planning model, the priority of path selection is further optimized. The water wave optimization algorithm simulates the propagation and contraction mechanisms of wave crests and wave troughs, and performs local and global optimizations according to the priority of paths, enabling the optimal path to gradually expand to the entire area of the inspection task, thereby generating a more practical inspection path. This method not only considers the shortest and optimal nature of the path but also can dynamically adapt to environmental changes during the path planning process, avoid obstacles and excessive energy consumption, and has higher flexibility and adaptability.

[0009] In addition, the present invention also combines real-time sensor data to dynamically adjust the path planning model, further improving the accuracy and efficiency of path planning. During the inspection process, the drone continuously monitors environmental changes based on sensor data and adjusts the path selection according to the feedback information, thus ensuring that the inspection task can be successfully completed in a complex dynamic environment. Compared with traditional methods, the method of the present invention can make full use of the powerful modeling ability of the graph convolutional network and combine the dynamic optimization mechanism of the water wave optimization algorithm to enhance the adaptability and robustness of path planning, significantly improving the execution efficiency and accuracy of the warehouse drone inspection task.

[0010] In summary, although the prior art has solved the problem of warehouse drone inspection path planning to a certain extent, there are still defects such as insufficient path planning accuracy, low efficiency, and poor adaptability to dynamic environmental changes. The present invention overcomes the deficiencies of the prior art by combining the graph convolutional network and the water wave optimization algorithm, and proposes a more efficient, accurate, and adaptive path planning method, providing strong technical support for the automation and intelligence of the warehouse drone inspection task.

[0011] Therefore, how to provide a warehouse drone inspection path planning method based on graph neural networks is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0012] An object of the present invention is to propose a warehouse drone inspection path planning method based on graph neural networks. The present invention combines a graph neural network model, a water wave optimization algorithm, and a path priority propagation technique to achieve warehouse drone inspection path planning based on graph neural networks. By constructing a graph model of the warehouse environment and performing data preprocessing, graph convolutional operations are used to aggregate the features of each node and its neighbor nodes to generate a preliminary path planning model. On this basis, the present invention introduces a water wave optimization algorithm to dynamically adjust each edge of the path planning model, simulating the propagation and contraction mechanisms of wave crests and wave troughs, optimizing the path priority, and then generating an optimal path planning model. In addition, the present invention also combines real-time sensor data, which can dynamically adjust the path during the inspection process, avoid obstacles and effectively reduce energy consumption, ensuring the successful completion of the adaptive inspection task of the drone in the warehouse environment. Through multiple iterations of optimization, the method of the present invention can gradually expand the wave crest area, making the path planning result more in line with actual needs, and improving the execution efficiency of the inspection task and the accuracy of path selection.

[0013] The warehouse drone inspection path planning method based on graph neural networks according to an embodiment of the present invention includes the following steps:

[0014] S1. Establish warehouse environment data and perform data preprocessing on the warehouse environment data;

[0015] S2. Apply the graph convolutional network model to build a model in combination with the warehouse environment data. Aggregate the features of each node and its neighbor nodes through graph convolution operations, and construct a preliminary path planning model for the relationships between nodes through the graph convolutional network;

[0016] S3. The drone starts the inspection task according to the initial path planning model and continuously monitors the environmental changes using real-time sensor data;

[0017] S4. Adopt the water wave optimization algorithm to simulate the process of water wave propagation. Initialize the water wave optimization algorithm on each path of the preliminary path planning model. The wave crest is the area with high path priority, and the wave trough is the area with low path priority. Dynamically adjust the path weights using the sensor data;

[0018] S5. Through the wave crest propagation and wave trough contraction mechanisms of the water wave, simulate the high-priority paths in the wave crest area, and preferentially select the high-priority paths for inspection. The low-priority paths in the wave trough area are suppressed;

[0019] S6. Through multiple iterations, further adjust the priority of the paths. The wave crest gradually expands to the optimal path area to generate the final inspection path;

[0020] S7. Repeat S4 - S6, and the drone continuously adjusts the inspection path.

[0021] Optionally, the warehouse environment data specifically includes: spatial location, shelf type, obstacle information, and environmental status.

[0022] Optionally, the features of each node and its neighbor nodes specifically include position coordinates, obstacle status, temperature and humidity, and product type.

[0023] Optionally, S2 specifically includes:

[0024] S21. Construct a warehouse graph structure model according to the preprocessed warehouse environment data;

[0025] S22. Generate a vector containing multi-dimensional features for each node, form the feature matrix of all nodes, generate a path feature vector for each path, and generate a path feature matrix;

[0026] S23. For each node, collect the feature information of all directly connected neighbor nodes, perform weighted aggregation on the features of the neighbor nodes, and then fuse them with its own features to generate a new feature representation of the node. Through the stacking of multi-layer graph convolutional networks, realize the hierarchical update of the features of all nodes in the warehouse graph structure:

[0027]

[0028] Among them, is the The feature representation of node i in the layer is the feature representation of node j in the layer, N(i) is the set of neighbor nodes of node i, and d i is the degree of node i, and d j is the degree of node j, W is the learnable weight matrix in the graph convolutional network, i is the number of the currently updated node, and j is a neighbor node of node i. is the iteration variable used to record the current network layer in the graph convolutional network;

[0029] S24. Through multi-layer graph convolution, obtain the embedding representation of each node in the full graph structure;

[0030] S25. Based on the embedding representation of each node in the full graph structure, sort and filter the possible paths between nodes, and construct a preliminary path planning model for the relationships between nodes.

[0031] Optionally, the S3 specifically includes:

[0032] S31. Set the path order in the preliminary path planning model generated according to the graph convolutional network as the execution path of the initial inspection task of the drone, and send it to the drone navigation module by the control system;

[0033] S32. During the inspection process of the drone, continuously collect real-time sensor data, including obstacle avoidance radar, depth camera, energy monitoring module, and positioning module;

[0034] S33. Process the collected real-time sensor data, map it to the input space of the graph convolutional network, dynamically update the nodes, and adjust the path priority in real time according to environmental changes;

[0035] S34. Set state detection indicators, with the energy consumption rate change less than or equal to 0.5% / s, the obstacle distance greater than 1.0 meter, and the environmental disturbance factor less than or equal to 0.3;

[0036] S35. When the state detection indicators exceed the set thresholds, call the graph convolutional network to adjust the current path planning model, and re-evaluate the path priority by updating the path connection relationship and node status.

[0037] Optionally, the S4 specifically includes:

[0038] S41. In the preliminary path planning model generated by the graph convolutional network, extract all paths and connected nodes, and set path weights for each path in combination with the structural similarity between nodes, path length, passage time, and obstacle information;

[0039] S42. Divide the path graph into a peak region and a trough region according to the path weight distribution. The peak region contains paths with high priorities and is used for the preferred direction of path propagation. The trough region contains paths with low priorities and is divided according to the top 20% and the bottom 20% of the path weights.

[0040] S43. Use the water wave optimization algorithm to dynamically optimize the path weights. Different from the traditional water wave algorithm, introduce the similarity of the nodes generated by the graph convolutional network as an adjustment factor for path propagation, and combine the energy consumption and obstacle avoidance factors to finely adjust the path weights, and output the optimized weights:

[0041]

[0042] Among them, is the optimized weight of path ij after the (t + 1)-th iteration, is the current path weight of path ij at the t-th iteration, e ij is the unit energy consumed for flying from node i to node along path ij, d ij is the distance between path ij and the nearest obstacle, α is the structure enhancement coefficient, β1 is the energy penalty coefficient, β2 is the obstacle avoidance penalty coefficient, is the embedding representation of node i in the (t + 1)-th layer of the graph convolutional network, is the embedding representation of node j in the (t + 1)-th layer of the graph convolutional network, i is the number of the currently updated node, j is a neighbor node of node i, is and is the embedding inner product of, t is the number of iterations in the water wave optimization algorithm;

[0043] S44. Use real-time sensor data to adjust the path propagation process. When the sensor detects that the energy consumption rate change of a certain path is greater than 0.5% / s, the obstacle distance is less than 1.0 meter, and the environmental disturbance factor is greater than 0.3, then weaken the corresponding path weight;

[0044] S45. In multiple rounds of iterations, the paths in the peak region continuously enhance their propagation advantages and are preferentially selected as the backbone paths by the path segments. The priorities of the paths in the trough region are gradually lowered until they are no longer included in the path candidate set;

[0045] S46. Feed back the optimized weights to the preliminary path planning model.

[0046] Optionally, the specific content of S5 includes:

[0047] S51. After the multi-round water wave optimization algorithm iteration is completed, extract the final weights of all paths, and divide the peak region and the trough region according to the path weight distribution. The peak region is the set of paths with high priority, and the trough region is the set of paths with low priority;

[0048] S52. According to the calculated optimized weights, use the optimized weights as the priority index for the inspection path selection, and construct a path propagation probability graph. The paths in the peak region have a higher probability of being selected, and the path selection probability in the trough region decreases:

[0049]

[0050] where p ij is the probability of node i selecting path ij, and ε i is the set of all paths connected by node i, is the final optimized weight of path ij optimized by the water wave optimization algorithm, is the final optimized weight of path ik connected by node i, and e ij is the unit energy consumed for flying from node i to node along path ij. i is the number of the currently updated node, j is a neighbor node of node i, and k is another neighbor node of node i;

[0051] S53. After the path propagation is completed, extract the paths with high path weights and high selection probabilities to form the inspection path;

[0052] S54. In the inspection path, introduce the nodes output by the graph convolutional network, and judge the structural consistency between consecutive nodes of the path. If the embedding inner product is lower than 0.4, the path is eliminated;

[0053] S55. Calculate the comprehensive cost value of each inspection path, and select the one with the minimum cost as the optimal path:

[0054]

[0055] where, is the comprehensive cost value of the path, P is the set of all paths, r ij is the path length, and e ij is the unit energy consumed for flying from node i to node along path ij. i is the number of the currently updated node, j is a neighbor node of node i, α' is the weight coefficient of the path length, β' is the weight coefficient of the energy consumption, γ' is the weight coefficient of the obstacle avoidance penalty, δ' is the reward coefficient of the propagation probability, and e ij is the unit energy consumed for flying from node i to node along path ij, o ij is the obstacle avoidance penalty term of the path, p ij is the probability of node i selecting path ij.

[0056] Optionally, S6 specifically includes:

[0057] S61. Based on the optimal path, extract all path weights in the path, and mark the paths with the top 20% highest weights as the initial peak region, and the bottom 20% as the initial trough region;

[0058] S62. Iteratively propagate on the optimal path, refine and converge the path priorities. In the t-th iteration, further update the path weights:

[0059]

[0060] where i is the number of the currently updated node, j is a neighbor node of node i, and k is another neighbor node of node i, is the updated path weight of path ij in the (t + 1)-th iteration, is the updated path weight of path ij in the t-th iteration, is the local path cost of path ij, t is the number of iterations in the water wave optimization algorithm, is the maximum local path cost of all paths in the current iteration, s ik is the similarity between path ij and path ik, N(ij) is the set of paths adjacent to path ij, ρ is the trough contraction penalty coefficient, is the trough region in the current iteration;

[0061] S63. After each round of iteration is completed, re-sort the path weights to generate a path weight sorting result;

[0062] S64. According to the path weight sorting result and based on the current path weights, delimit new peak and trough regions, and the peak region gradually expands structurally to cover the optimal path.

[0063] Optionally, S64 specifically includes:

[0064] S641. After each round of path weight iteration is completed, according to the latest weight values of all current paths, re-divide the peak region and the trough region, divide the paths with the top 20% weights into the peak region, and divide the paths with the bottom 20% weights into the trough region;

[0065] S642. For each path located in the peak region, find the directly connected adjacent paths in the graph structure;

[0066] S643. If the adjacent path is similar to the current path in terms of structural features and the weight value in the current iteration is at a medium or higher level, then include the adjacent path in the next round of the peak region;

[0067] S644. For the paths in the trough region, restrict the participation in path propagation in the next iteration. The peak region is gradually expanded structurally to cover the optimal path, and the optimal path is output as the final inspection path.

[0068] Optionally, the S7 specifically includes:

[0069] S71. During the process of the drone executing the final inspection path, continuously collect real-time sensor data;

[0070] S72. When the sensor detects that the energy consumption rate change of a certain path is greater than 0.5% / s, or the distance to the obstacle is less than 1.0 meter, or the environmental disturbance factor is greater than 0.3, repeat S3 - S6 to update the weights and path priorities;

[0071] S73. According to the updated path selection probability, re-plan the next path segment, and preferentially select the path with high probability and high weight;

[0072] S74. Connect the updated path segment to the current path execution trajectory to form an adaptive inspection path for the drone.

[0073] The beneficial effects of the present invention are:

[0074] By combining the graph neural network model, the water wave optimization algorithm, and the path priority propagation technology, the present invention significantly improves the efficiency and accuracy of the inspection path planning for warehouse drones. Compared with traditional path planning methods, the present invention can not only handle the complex environmental structure of the warehouse, but also dynamically adjust the path according to real-time sensor data, thereby realizing the adaptive optimization of the path. This method overcomes the limitations of insufficient path planning accuracy, poor environmental adaptability, and inability to respond to dynamic changes in real time in the prior art.

[0075] First, by constructing a graph model of the warehouse environment and using the graph neural network to model the warehouse graph, the present invention can efficiently aggregate the relationships between nodes and handle the complex spatial layout in the warehouse. This process extracts the feature information of the nodes and their neighbor nodes through graph convolution operations, making the path planning more in line with the actual needs of the warehouse environment. Compared with traditional algorithms, the graph neural network adopted by the present invention can more accurately model the relationships between nodes, improving the accuracy and efficiency of path planning.

[0076] Secondly, the water wave optimization algorithm introduced in the present invention optimizes each edge in the path planning model by simulating the propagation and contraction mechanisms of water waves, thereby dynamically adjusting the priority of the path globally and locally. This mechanism effectively avoids the common local optimal solution problem in path planning and ensures the global optimality of path planning. During multiple iterative processes of the water wave optimization algorithm, the crest area is gradually expanded, so that the finally generated path can not only avoid obstacles but also minimize energy consumption.

[0077] In addition, the present invention also makes full use of real-time sensor data for dynamic path adjustment, which can monitor environmental changes in real time and make corresponding path optimizations. This enables the UAV to flexibly respond to various environmental challenges according to the actual situation during the inspection process, improving the execution success rate and efficiency of the inspection task. Through multiple rounds of optimization iterations, the priority and selection of the path will be more in line with the requirements in the actual scenario, avoiding interference from obstacles and unnecessary energy waste.

[0078] In summary, by combining the graph neural network and the water wave optimization algorithm, the present invention realizes a more accurate and efficient path planning for warehouse UAV inspections, with strong environmental adaptability and path optimization capabilities. This technology can not only maintain the stability of the path in a complex dynamic environment but also effectively improve the inspection efficiency and accuracy, providing a new solution for the intelligent inspection of warehouse UAVs. Brief Description of the Drawings

[0079] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0080] Figure 1 is a flowchart of the method for path planning of warehouse UAV inspection based on the graph neural network proposed by the present invention;

[0081] Figure 2 is a schematic diagram of the method for path planning of warehouse UAV inspection based on the graph neural network proposed by the present invention;

[0082] Figure 3 is a data flow diagram of the method for path planning of warehouse UAV inspection based on the graph neural network proposed by the present invention. Detailed Description of the Embodiments

[0083] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0084] Refer to Figures 1-3 , the method for path planning of warehouse UAV inspection based on the graph neural network includes the following steps:

[0085] S1. Establish warehouse environment data and perform data preprocessing on the warehouse environment data;

[0086] S2. Apply the graph convolutional network model to model in combination with the warehouse environment data, aggregate the features of each node and its neighbor nodes through graph convolutional operations, and construct a preliminary path planning model for the relationships between nodes through the graph convolutional network;

[0087] S3. The drone starts the inspection task according to the initial path planning model and continuously monitors the environmental changes using real-time sensor data;

[0088] S4. Adopt the water wave optimization algorithm, simulate the water wave propagation process, initialize the water wave optimization algorithm on each path of the preliminary path planning model, with the wave crest being the area with high path priority and the wave trough being the area with low path priority, and dynamically adjust the path weights using the sensor data;

[0089] S5. Through the wave crest propagation and wave trough contraction mechanisms of the water wave, simulate the high-priority paths in the wave crest area, preferentially select the high-priority paths for inspection, and suppress the low-priority paths in the wave trough area;

[0090] S6. Through multiple iterations, further adjust the path priorities, and the wave crest gradually expands to the optimal path area to generate the final inspection path;

[0091] S7. Repeat S4 - S6, and the drone continuously adjusts the inspection path.

[0092] The present invention combines the graph convolutional network and the water wave optimization algorithm, and realizes the intelligent optimization of the drone inspection path by dynamically adjusting the path priorities. By constructing a graph model of the warehouse environment and performing data preprocessing, applying the graph convolutional network to aggregate node features to generate a preliminary path planning model. Combining the real-time sensor data, the drone can adjust the path in real time according to the environmental changes during the inspection process. The water wave optimization algorithm optimizes the selection of the inspection path according to the path priorities by simulating the wave crest propagation and wave trough contraction mechanisms, ensuring that the high-priority paths are preferentially executed. During the multi-round iteration process, the wave crest area gradually expands to the optimal path area, improving the accuracy and efficiency of path selection. At the same time, this method can flexibly adapt to dynamic environmental changes, effectively avoid obstacles and excessive energy consumption, and ensure the efficient execution of the drone inspection task.

[0093] In this embodiment, the warehouse environment data specifically includes: spatial location, shelf type, obstacle information, and environmental status.

[0094] The present invention enhances the accuracy and adaptability of path planning by constructing a multi-dimensional data model of the warehouse environment, including spatial location, shelf type, obstacle information, and environmental status. By comprehensively considering these environmental characteristics, the unmanned aerial vehicle can accurately identify key factors in the path, optimize the inspection path, avoid obstacles, and reduce energy consumption, ensuring the efficient completion of inspection tasks in a complex warehousing environment. This method has the ability to dynamically adjust and adapt to the changing warehouse environment.

[0095] In this embodiment, the characteristics of each node and its neighbor nodes specifically include position coordinates, obstacle status, temperature and humidity, and product type.

[0096] The present invention optimizes the decision-making process of path planning by integrating the multi-dimensional characteristics of each node and its neighbor nodes, including position coordinates, obstacle status, temperature and humidity, and product type. By combining these key environmental parameters, the unmanned aerial vehicle can more accurately select the optimal inspection path, adjust the path priority in real time, effectively avoid obstacles, prevent excessive energy consumption, and improve the inspection efficiency. This method has dynamic adaptability, ensuring flexible and efficient task execution in a complex warehouse environment.

[0097] In this embodiment, S2 specifically includes:

[0098] S21. Construct a warehouse graph structure model based on the preprocessed warehouse environment data;

[0099] S22. Generate a vector containing multi-dimensional characteristics for each node, form the feature matrix of all nodes, generate a path feature vector for each path, and generate a path feature matrix;

[0100] S23. For each node, collect the feature information of all directly connected neighbor nodes, perform weighted aggregation on the features of the neighbor nodes, and then fuse them with its own features to generate a new feature representation of the node. Through the stacking of multi-layer graph convolutional networks, the features of all nodes in the warehouse graph structure are updated layer by layer:

[0101]

[0102] Among them, is the feature representation of node i in the th layer, is the feature representation of node j in the th layer, N(i) is the set of neighbor nodes of node i, d i is the degree of node i, d j is the degree of node j, W is the learnable weight matrix in the graph convolutional network, i is the number of the currently updated node, and j is a neighbor node of node i. is an iterative variable used to record the current network layer in the graph convolutional network;

[0103] S24. Through multi-layer graph convolution, obtain the embedding representation of each node in the full graph structure;

[0104] S25. Based on the embedding representation of each node in the full graph structure, sort and filter the possible paths between nodes, and construct a preliminary path planning model for the relationships between nodes.

[0105] In the present invention, by constructing a graph structure model based on warehouse environment data, aggregate the multi-dimensional features of each node and its neighbor nodes, and use the graph convolutional network to update the feature representation of the nodes layer by layer. In this way, the features of the nodes can not only reflect their own information, but also fuse the features of neighbor nodes, thereby generating more accurate and rich node embedding representations. Through multi-layer graph convolution, the features of the nodes in the full graph structure are comprehensively updated, realizing a more accurate path planning model. Based on the node embedding representation, the relationships between paths are optimized, and the possible paths between nodes are effectively screened and sorted, thus ensuring the optimal selection of the inspection path and the efficient execution of the overall inspection task. This method can adapt to the complex dynamic changes in the warehouse environment, improving the accuracy and adaptability of path planning.

[0106] In this embodiment, the specific steps of S3 are as follows:

[0107] S31. Set the path order in the preliminary path planning model generated according to the graph convolutional network as the execution path of the initial inspection task of the unmanned aerial vehicle, and send it to the unmanned aerial vehicle navigation module by the control system;

[0108] S32. During the inspection process of the unmanned aerial vehicle, continuously collect real-time sensor data, including obstacle avoidance radar, depth camera, energy monitoring module, and positioning module;

[0109] S33. Process the collected real-time sensor data, map it to the input space of the graph convolutional network, dynamically update the nodes, and adjust the path priority in real time according to environmental changes;

[0110] S34. Set state detection indicators, with the energy consumption rate change less than or equal to 0.5% / s, the obstacle distance greater than 1.0 meter, and the environmental disturbance factor less than or equal to 0.3;

[0111] S35. When the state detection indicators exceed the set threshold, call the graph convolutional network to adjust the current path planning model, and re-evaluate the path priority by updating the path connection relationship and node status.

[0112] The present invention optimizes the inspection path planning of an unmanned aerial vehicle (UAV) by combining a graph convolutional network with real-time sensor data. After the generation of the preliminary path planning model, the path sequence is set as the initial inspection task of the UAV and sent to the UAV navigation module by the control system. During the inspection process, the real-time collected sensor data includes an obstacle avoidance radar, a depth camera, an energy monitoring module, and a positioning module, ensuring that the UAV can dynamically adjust the path according to environmental changes. By mapping the sensor data to the input space of the graph convolutional network, the UAV can update the node and path priorities in real time. When the environmental conditions exceed the set threshold, the graph convolutional network will re-evaluate the path connection relationship to ensure the continuous optimization of the path planning. This method significantly improves the flexibility and adaptability of the path planning, ensuring that the UAV can efficiently avoid obstacles, avoid excessive energy consumption, and ensure the successful completion of the task in a dynamic environment.

[0113] In this embodiment, the S4 specifically includes:

[0114] S41. In the preliminary path planning model generated by the graph convolutional network, extract all paths and connected nodes, and set path weights for each path by combining the structural similarity between nodes, path length, passage time, and obstacle information.

[0115] S42. According to the path weight distribution, divide the path graph into a peak region and a valley region. The peak region contains paths with high priorities and is used for the preferred direction of path propagation. The valley region contains paths with low priorities, which are divided according to the top 20% and bottom 20% of the path weights.

[0116] S43. Use the water wave optimization algorithm to dynamically optimize the propagation of path weights. Different from the traditional water wave algorithm, introduce the similarity of nodes generated by the graph convolutional network as an adjustment factor for path propagation, and refine the adjustment of path weights by combining the energy consumption and obstacle avoidance factors, and output the optimized weights:

[0117]

[0118] Wherein, is the optimized weight of path ij after the (t + 1)-th iteration, is the current path weight of path ij at the t-th iteration, e ij is the unit energy consumed for flying from node i to along path ij, d ij is the distance between path ij and the nearest obstacle, α is the structure enhancement coefficient, β1 is the energy penalty coefficient, β2 is the obstacle avoidance penalty coefficient, is the embedding representation of node i in the (t + 1)-th layer of the graph convolutional network, is the embedding representation of node j in the (t + 1)-th layer of the graph convolutional network, i is the number of the currently updated node, and j is a neighbor node of node i. is and the embedded inner product, where t is the number of iterations in the water wave optimization algorithm;

[0119] S44. Use real-time sensor data to adjust the path propagation process. When the sensor detects that the change rate of the energy consumption rate of a certain path is greater than 0.5% / s, the distance to the obstacle is less than 1.0 meter, and the environmental disturbance factor is greater than 0.3, then weaken the corresponding path weight;

[0120] S45. In multiple rounds of iteration, the paths in the peak region continuously enhance their propagation advantages and are preferentially selected as the backbone paths by the path segments. The priority of the paths in the trough region is gradually reduced until they are no longer included in the path candidate set;

[0121] S46. Feed back the optimized weight to the preliminary path planning model.

[0122] The present invention combines a graph convolutional network and a water wave optimization algorithm to achieve dynamic optimization of the UAV inspection path. In the preliminary path planning model, by analyzing the structural similarity, path length, passage time, and obstacle information between nodes, the path weights are accurately set, and the water wave optimization algorithm is used to dynamically optimize the path weights. Different from the traditional water wave algorithm, the present invention introduces the node similarity generated by the graph convolutional network as the adjustment factor for path propagation, and combines the energy consumption and obstacle avoidance factors for fine adjustment. During the path propagation process, real-time sensor data can further adjust the path weights to ensure that the UAV avoids obstacles, saves energy, and improves the inspection efficiency. Through multiple rounds of iteration, the paths in the peak region are gradually enhanced and preferentially selected as the backbone paths, while the paths in the trough region are gradually suppressed, finally optimizing the path planning to ensure the efficiency and self-adaptability of the path in a dynamic environment.

[0123] In this embodiment, the specific steps of S5 are as follows:

[0124] S51. After multiple rounds of iteration of the water wave optimization algorithm are completed, extract the final weights of all paths, and divide the peak region and the trough region according to the path weight distribution. The peak region is the set of paths with high priority, and the trough region is the set of paths with low priority;

[0125] S52. According to the calculated optimized weight, use the optimized weight as the priority index for selecting the inspection path, and construct a path propagation probability graph. The paths in the peak region have a higher probability of being selected, and the path selection probability in the trough region decreases:

[0126]

[0127] where p ij is the probability of selecting path ij at node i, and ε iThe set of all paths connected to node i The final optimized weight of path ij optimized by the water wave optimization algorithm The final optimized weight of path ik connected to node i, e ij is the unit energy consumed for flying from node i to node along path ij, i is the number of the currently updated node, j is a neighbor node of node i, and k is another neighbor node of node i;

[0128] S53. After completing path propagation, extract the paths with high path weights and large selection probabilities to form the inspection path;

[0129] S54. In the inspection path, introduce the nodes output by the graph convolutional network to judge the structural consistency between consecutive nodes of the path. If the embedding inner product is less than 0.4, then eliminate the path;

[0130] S55. Calculate the comprehensive cost value of each inspection path, and select the one with the minimum cost as the optimal path:

[0131]

[0132] where is the comprehensive cost value of the path, P is the set of all paths, r ij is the path length, e ij The unit energy consumed for flying from node i to node along path ij, i is the number of the currently updated node, j is a neighbor node of node i, α' is the weight coefficient of the path length, β' is the weight coefficient of the energy consumption, γ' is the weight coefficient of the obstacle avoidance penalty, δ' is the reward coefficient of the propagation probability, e ij is the unit energy consumed for flying from node i to node along path ij, o ij is the obstacle avoidance penalty term of the path, p ij is the probability of selecting path ij at node i.

[0133] The present invention optimizes the selection and adjustment of the UAV inspection path by combining the water wave optimization algorithm and the graph convolutional network. After multiple rounds of optimization iterations, according to the optimized weights of the paths, the paths are divided into the peak region and the trough region, and the high-priority paths are preferentially selected for inspection. During the path propagation process, based on the optimized path weights and selection probabilities, a path propagation probability graph is constructed, effectively improving the accuracy of path selection. By calculating the comprehensive cost of the path, considering multi-dimensional factors such as path length, energy consumption, obstacle avoidance penalty, and propagation probability, the path with the minimum total cost is selected as the optimal path. At the same time, the node embedding representation output by the graph convolutional network is introduced to ensure the structural consistency between nodes in the path, further improving the accuracy and robustness of path planning. This method not only improves the efficiency of path planning but also can dynamically respond to environmental changes to ensure the efficient execution of the inspection task.

[0134] In this embodiment, S6 specifically includes:

[0135] S61. Based on the optimal path, extract all path weights in the path, and mark the first 20% of the paths with the highest weights as the initial peak region, and the last 20% as the initial trough region;

[0136] S62. Iteratively propagate on the optimal path, refine and converge the path priorities. In the t-th iteration, further update the path weights:

[0137]

[0138] where i is the number of the currently updated node, j is a neighbor node of node i, and k is another neighbor node of node i. is the updated path weight of path ij in the (t + 1)-th iteration. is the updated path weight of path ij in the t-th iteration. is the local path cost of path ij, t is the number of iterations in the water wave optimization algorithm. is the maximum local path cost of all paths in the current iteration, s ik is the similarity between path ij and path ik, N(ij) is the set of paths adjacent to path ij, and ρ is the trough contraction penalty coefficient. is the trough region in the current iteration;

[0139] S63. After each round of iteration, re-sort the path weights to generate the path weight sorting result;

[0140] S64. According to the path weight sorting result and based on the current path weights, delimit new peak and trough regions. The peak region gradually expands structurally to cover the optimal path.

[0141] The present invention iteratively updates the path weights through the water wave optimization algorithm to gradually refine the path priorities. Based on the optimal path, by marking the first 20% of the paths with the highest weights as the initial peak region and the last 20% as the initial trough region, and updating the path weights in each iteration. The local cost and path similarity of the paths are comprehensively considered to accurately adjust the path priorities. Through continuous iteration, the path weights are re-sorted, the path propagation is optimized, and the peak region gradually expands to cover the optimal path, thereby ensuring that the path selection not only conforms to the global optimality but also can adapt to the changes in the dynamic environment. This method can effectively avoid local optimal solutions and improve the flexibility and execution efficiency of path planning.

[0142] In this embodiment, S64 specifically includes:

[0143] S641. After each round of path weight iteration is completed, re-divide the peak region and the trough region according to the latest weight values of all current paths. Divide the paths with the top 20% of the weights into the peak region, and divide the paths with the bottom 20% of the weights into the trough region;

[0144] S642. For each path located in the peak region, find the directly connected adjacent paths in the graph structure;

[0145] S643. If the adjacent path is similar to the current path in terms of structural features and the weight value in the current iteration is at a medium or higher level, include the adjacent path in the next round of the peak region;

[0146] S644. For the paths located in the trough region, restrict their participation in path propagation in the next iteration. The peak region gradually expands structurally to cover the optimal path, and output the optimal path as the final inspection path.

[0147] The present invention optimizes the UAV inspection path selection by dynamically adjusting the path weights. After each round of path weight iteration is completed, re-divide the peak region and the trough region according to the path weights to ensure that the paths with high weights are preferentially selected for inspection. By finding adjacent paths and judging their structural similarity, the potential paths are successfully included in the peak region to further optimize path propagation. The peak region gradually expands to cover the optimal path, while the paths in the trough region are restricted from participating in propagation during iteration, ensuring that the finally selected path can maximize the inspection efficiency and obstacle avoidance ability. In addition, the iteration mechanism ensures the flexibility and adaptability of path selection and can cope with dynamic environmental changes.

[0148] In this embodiment, the specific steps of S7 are as follows:

[0149] S71. During the process of the UAV executing the final inspection path, continuously collect real-time sensor data;

[0150] S72. When the sensor detects that the energy consumption rate change of a certain path is greater than 0.5% / s or the obstacle distance is less than 1.0 meter or the environmental disturbance factor is greater than 0.3, repeat steps S3 - S6 to update the weights and the path priorities;

[0151] S73. According to the updated path selection probability, re-plan the next path segment and preferentially select the paths with high probability and high weight;

[0152] S74. Connect the updated path segment to the current path execution trajectory to form an adaptive inspection path for the UAV.

[0153] The present invention continuously monitors the inspection path of the drone through real-time sensor data and can dynamically adjust the path planning. When the sensor detects that the energy consumption rate change, the obstacle distance, or the environmental disturbance factor exceeds the preset threshold, the system will automatically return to the path planning and optimization stage. By updating the path weights and priorities, it ensures that the drone can avoid obstacles and reduce energy consumption. By re-planning the path segments and preferentially selecting the paths with high probability and high weight, it ensures the efficiency and safety of the inspection path. The finally formed adaptive inspection path can flexibly respond to complex environmental changes, ensure the smooth completion of the inspection task, and improve the accuracy and efficiency of the path planning.

[0154] Embodiment 1:

[0155] To verify the feasibility of the present invention in implementation, the present invention is applied to the inspection path planning of a drone in a large warehouse. The warehouse is located in Chaoyang District, Beijing, with an area of about 10,000 square meters, including multiple shelf areas, multiple import and export channels, and complex stacking areas. The environment also includes dynamic personnel and equipment flows. Traditional drone inspection methods usually adopt preset fixed paths or perform path planning based on simple optimization algorithms. However, these methods often perform poorly in dynamic environments, especially when encountering obstacles, personnel activities, or other emergencies, and cannot adjust the path in a timely manner, resulting in low inspection efficiency and even possibly affecting the completion quality of the task.

[0156] In this embodiment, we combine the graph neural network model with the water wave optimization algorithm and use sensor data for real-time path adjustment, thereby greatly improving the accuracy and efficiency of the inspection path planning of the warehouse drone. Specifically, we first model the warehouse environment through the graph neural network and generate a graph model based on each key position in the warehouse. The features of each node include position coordinates, shelf types, obstacle information, temperature and humidity, and product types, etc. These information helps the graph neural network generate a more accurate path planning model in subsequent calculations.

[0157] Subsequently, we introduce the water wave optimization algorithm to perform dynamic adjustment on each edge of the path planning model, simulating the mechanism of the wave crest and wave trough of the water wave. Through the propagation of the wave crest and the contraction of the wave trough of the water wave, we can perform local and global optimization according to the path priorities and gradually expand to the optimal path area. This optimization process ensures that during the actual execution, the drone can avoid obstacles and reduce energy consumption, while improving the global optimality of the path planning.

[0158] During the inspection process, by using the real-time sensor data of the drone, we can dynamically monitor environmental changes and adjust the path in a timely manner. The drone continuously feeds back the status of the current path through sensors and fine-tunes the path weights according to the sensor data. For example, when an obstacle is detected, the weight of the path changes, and the peak area expands to the vicinity of the obstacle to ensure that the drone can avoid the obstacle and select a suitable path to continue the inspection.

[0159] To verify the effectiveness of the present invention, we conducted a one-week test in the above warehouse environment. During the test, multiple key data such as the inspection time, path accuracy, obstacle avoidance rate, and energy consumption of the drone inspection were recorded. The following is the relevant data table during the test, which shows the comparison of the optimization effect of the drone inspection path and the traditional path planning method under different conditions.

[0160] Table 1 Comparison table of the optimization effect of the warehouse drone inspection path planning method based on the graph neural network

[0161]

[0162]

[0163] Table 1 shows that the path planning method based on the graph neural network and the water wave optimization algorithm has obvious advantages in many aspects. First, the inspection time is significantly shortened. The shortest inspection time is 40 minutes, while the inspection time of the traditional path planning method can reach up to 60 minutes at most. Second, the path accuracy is greatly improved. The path planning method based on the present invention can ensure that the path accuracy is within 2 meters in the optimal path planning, while the path accuracy of the traditional method is generally poor, reaching up to 7 meters at most. Most importantly, the obstacle avoidance rate is greatly improved. The drone based on the present invention can avoid more than 98% of the obstacles, while the avoidance rate of the traditional method is relatively low, only 75% to 80%. In addition, the energy consumption is also significantly reduced. The energy consumption of the traditional path planning method is usually relatively high, reaching up to 90 Wh at most, while the energy consumption of the path planning method based on the graph neural network and the water wave optimization algorithm usually remains below 60 Wh.

[0164] As mentioned above, the above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A method for warehouse drone inspection path planning based on graph neural network, characterized in that, It includes the following steps: S1. Establish warehouse environment data and perform data preprocessing on the warehouse environment data; S2. Apply a graph convolutional network model to build a model in combination with the warehouse environment data. Aggregate the features of each node and its neighbor nodes through graph convolutional operations, and construct a preliminary path planning model for the relationships between nodes through the graph convolutional network; S3. The drone starts the inspection task according to the initial path planning model and continuously monitors the environmental changes using real-time sensor data; S4. Adopt the water wave optimization algorithm to simulate the water wave propagation process. Initialize the water wave optimization algorithm on each path of the preliminary path planning model. The wave crest is the area with high path priority, and the wave trough is the area with low path priority. Dynamically adjust the path weights using the sensor data; S5. Through the wave crest propagation and wave trough contraction mechanisms of the water wave, simulate the high-priority paths in the wave crest area, and preferentially select the high-priority paths for inspection. The low-priority paths in the wave trough area are suppressed; S6. Through multiple iterations, further adjust the path priorities. The wave crest gradually expands to the optimal path area to generate the final inspection path; S7. Repeat S4 - S6, and the drone continuously adjusts the inspection path.

2. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, wherein The warehouse environment data specifically includes: spatial location, shelf type, obstacle information, and environmental status.

3. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, wherein The features of each node and its neighbor nodes specifically include position coordinates, obstacle status, temperature and humidity, and product type.

4. The method for planning the inspection path of warehouse drones based on graph neural network according to claim 1, wherein, S2 specifically includes: S21. According to the preprocessed warehouse environment data, construct a warehouse graph structure model; S22. Generate a vector containing multi-dimensional features for each node, form the node feature matrix from the features of all nodes, generate a path feature vector for each path, and generate a path feature matrix; S23. For each node, collect the feature information of all directly connected neighbor nodes, perform weighted aggregation on the features of the neighbor nodes, and then fuse them with its own features to generate a new feature representation of the node. Through the stacking of multiple-layer graph convolutional networks, realize the layer-by-layer update of the features of all nodes in the warehouse graph structure: Among them, is the feature representation of node i in the layer, is the feature representation of node j in the layer, N(i) is the set of neighbor nodes of node i, and d i is the degree of node i, d j is the degree of node j, W is the learnable weight matrix in the graph convolutional network, i is the number of the currently updated node, j is a neighbor node of node i, is the iteration variable used in the graph convolutional network to record the current network level; S24. Through multiple-layer graph convolution, obtain the embedding representation of each node in the full graph structure; S25. Based on the embedding representation of each node in the full graph structure, sort and screen the possible paths between nodes, and construct a preliminary path planning model for the relationships between nodes.

5. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, characterized in that, S3 specifically includes: S31. Set the path order in the preliminary path planning model generated according to the graph convolutional network as the execution path of the drone's initial inspection task, and send it to the drone navigation module by the control system; S32. During the drone's inspection process, continuously collect real-time sensor data, including obstacle avoidance radar, depth camera, energy monitoring module, and positioning module; S33. Process the collected real-time sensor data, map it to the input space of the graph convolutional network, dynamically update the nodes, and the path priorities are adjusted in real time according to the environmental changes; S34. Set the state detection indicators, the change rate of energy consumption is less than or equal to 0.5% / s, the obstacle distance is greater than 1.0 meter, and the environmental disturbance factor is less than or equal to 0.3; S35. When the status detection index exceeds the set threshold, call the graph convolutional network to adjust the current path planning model. By updating the path connection relationship and node status, re-evaluate the path priority.

6. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, wherein The specific steps of S4 are as follows: S41. In the preliminary path planning model generated by the graph convolutional network, extract all paths and connecting nodes. Combine the structural similarity between nodes, path length, passage time, and obstacle information to set path weights for each path. S42. According to the path weight distribution, divide the path graph into peak regions and valley regions. The peak regions contain paths with high priority, which are used as the preferred directions for path propagation. The valley regions contain paths with low priority, and they are divided according to the top 20% and bottom 20% of the path weights. S43. Use the water wave optimization algorithm to dynamically optimize the propagation of path weights. Different from the traditional water wave algorithm, introduce the similarity of nodes generated by the graph convolutional network as an adjustment factor for path propagation, and combine the energy consumption and obstacle avoidance factors to finely adjust the path weights, and output the optimized weights. Among them, is the optimized weight of path ij after the (t + 1)-th iteration, is the current path weight of path ij at the t-th iteration, e ij is the unit energy consumed by flying from node i to node along path ij, d ij is the distance between path ij and the nearest obstacle, α is the structure enhancement coefficient, β1 is the energy penalty coefficient, and β2 is the obstacle avoidance penalty coefficient. is the embedding representation of node i in the (t + 1)-th layer of the graph convolutional network, is the embedding representation of node j in the (t + 1)-th layer of the graph convolutional network, i is the number of the currently updated node, and j is a neighbor node of node i. is and is the inner product of the embeddings, and t is the number of iterations in the water wave optimization algorithm; S44. Use real-time sensor data to adjust the path propagation process. When the sensor detects that the energy consumption rate change of a certain path is greater than 0.5% / s, the obstacle distance is less than 1.0 meter, and the environmental disturbance factor is greater than 0.3, then weaken the corresponding path weight. S45. In multiple rounds of iteration, the paths in the peak regions continuously enhance their propagation advantages and are preferentially selected as the backbone paths by the path segments. The priority of the paths in the valley regions is gradually suppressed until they are no longer included in the path candidate set. S46. Feed back the optimized weights to the preliminary path planning model.

7. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, wherein The specific steps of S5 are as follows: S51. After multiple rounds of iteration of the water wave optimization algorithm are completed, extract the final weights of all paths. Divide the peak regions and valley regions according to the path weight distribution. The peak regions are the set of paths with high priority, and the valley regions are the set of paths with low priority. S52. According to the calculated optimized weights, use the optimized weights as the priority index for inspection path selection to construct a path propagation probability graph. The paths in the peak regions have a higher probability of being selected, and the path selection probability in the valley regions decreases. where p ij is the probability that node i selects path ij, ε i is the set of all paths connected to node i, is the final optimized weight of path ij optimized by the water wave optimization algorithm, is the final optimized weight of path ik connected to node i, e ij is the unit energy consumed for flying from node i to node along path ij, i is the number of the currently updated node, j is a neighbor node of node i, and k is another neighbor node of node i; S53. After path propagation is completed, extract the paths with high path weights and large selection probabilities to form the inspection path. S54. In the inspection path, introduce the nodes output by the graph convolutional network to judge the structural consistency between consecutive nodes of the path. If the embedded inner product is lower than 0.4, then eliminate the path. S55. Calculate the comprehensive cost value of each inspection path and select the one with the minimum cost as the optimal path. Among them, is the comprehensive path cost value, P is the set of all paths, r ij is the path length, e ij is the unit energy consumed for flying from node i to node along path ij. i is the number of the currently updated node, j is a neighbor node of node i, α' is the weight coefficient of the path length, β' is the weight coefficient of the energy consumption, γ' is the weight coefficient of the obstacle avoidance penalty, δ' is the reward coefficient of the propagation probability, e ij is the unit energy consumed for flying from node i to node along path ij, o ij is the obstacle avoidance penalty term of the path, p ij is the probability of selecting path ij at node i.

8. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, characterized in that, The specific steps of S6 are as follows: S61. Based on the optimal path, extract all path weights in the path and mark the paths with the top 20% highest weights as the initial peak regions, and the bottom 20% as the initial valley regions. S62. Iteratively propagate on the optimal path to refine and converge the path priority. In the t-th iteration, further update the path weights. Wherein, i is the node number currently being updated, j is a neighbor node of node i, and k is another neighbor node of node i. is the updated path weight of path ij in the (t + 1)-th iteration. is the updated path weight of path ij in the t-th iteration. is the local path cost of path ij, and t is the iteration round in the water wave optimization algorithm. is the maximum local path cost of all paths in the current iteration, s ik is the similarity between path ij and path ik, N(ij) is the set of paths adjacent to path ij, and ρ is the wave trough contraction penalty coefficient. is in the wave trough area of the current iteration; S63. After each round of iteration is completed, re-sort the path weights to generate the path weight sorting result. S64. Based on the path weight sorting result and according to the current path weight, delimit new peak and valley regions, and the peak region gradually expands structurally to cover the optimal path.

9. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, wherein The specific content of S64 includes: S641. After each round of path weight iteration is completed, according to the latest weight values of all current paths, re-divide the peak region and the valley region. Divide the paths with the top 20% of the weights into the peak region, and divide the paths with the bottom 20% of the weights into the valley region; S642. For each path in the peak region, find the directly connected adjacent paths in the graph structure; S643. If the adjacent path is similar to the current path in terms of structural features and the weight value in the current iteration is above the medium level, include the adjacent path in the next round of the peak region; S644. For the paths in the valley region, restrict their participation in path propagation in the next iteration. The peak region gradually expands structurally to cover the optimal path, and output the optimal path as the final inspection path.

10. The method for planning the inspection path of a warehouse drone based on a graph neural network according to claim 1, wherein The specific content of S7 includes: S71. During the process of the drone executing the final inspection path, continuously collect real-time sensor data; S72. When the sensor detects that the change rate of the energy consumption rate of a certain path is greater than 0.5% / s, or the distance to the obstacle is less than 1.0 meter, or the environmental disturbance factor is greater than 0.3, repeat S3 - S6 to update the weights and update the path priorities; S73. According to the updated path selection probability, re-plan the next path segment, and preferentially select the paths with high probability and high weight; S74. Connect the updated path segment to the current path execution trajectory to form an adaptive inspection path for the drone.

Images