A method and device for automatically generating a narrow airspace route

Abstract

The application discloses a method and device for automatically generating a flight path in a narrow space, and the method comprises the following steps: preprocessing three-dimensional region data of a target space, and constructing a guide structure according to the three-dimensional region data after preprocessing; constructing and training an initial generation model, and then training the initial generation model according to preset operation constraint parameters to obtain a path generation model; and generating a flight path according to the guide structure and the path generation model. The application can improve the path planning efficiency of a UAV in a narrow space.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a method and apparatus for automatically generating airways in narrow airspace. Background Technology

[0002] With the rapid development of UAV technology, intelligent route planning in low-altitude airspace has become one of the core issues ensuring its safe and efficient operation. In typical complex airspaces such as cities, mountains, and forests, the available space for flight is extremely limited due to factors such as terrain, tall buildings, and radio blind spots, often presenting a "narrow passage" shape, such as urban canyons, highway corridors, and the space under bridges. Existing technologies often employ rule-based mapping-based path search methods or a combination of grid maps and deep reinforcement learning for route planning. However, these methods suffer from shortcomings in narrow spaces, including weak spatial structure representation capabilities, a lack of spatial guidance path generation mechanisms, and an inability to flexibly embed safety margins, resulting in limited intelligence in path generation. Therefore, there is an urgent need for a novel route generation method that can combine complex spatial structure perception, learning-based path generation, and aircraft safety margin control, suitable for UAV route planning tasks in typical narrow passage scenarios such as cities and mountains, to improve the feasibility, safety, and intelligence of routes. Summary of the Invention

[0003] To overcome the shortcomings of the prior art, the present invention provides a method and apparatus for automatically generating routes in narrow airspace, which can improve the efficiency of UAV path planning in narrow airspace.

[0004] An embodiment of the present invention provides a method for automatically generating airways in narrow airspace, comprising the following steps: The three-dimensional region data of the target airspace is preprocessed, and a guiding structure is constructed based on the preprocessed three-dimensional region data; An initial generation model is constructed and trained, and then trained according to preset running constraint parameters to obtain a path generation model; Based on the guidance structure, a flight path is generated through the path generation model.

[0005] Furthermore, the preprocessing of the three-dimensional region data of the target airspace specifically includes: The three-dimensional region data of the target airspace is acquired, and the three-dimensional region data is subjected to projection, resampling and rasterization operations to obtain a three-dimensional voxel mesh model; wherein, the three-dimensional region data includes three-dimensional obstacle data and free space data; Based on the three-dimensional obstacle data, the obstacle model in the three-dimensional voxel mesh model is dilated to obtain an obstacle voxel set; Based on the free space data, a three-dimensional Voronoi diagram is constructed.

[0006] Furthermore, the step of constructing the guiding structure based on the preprocessed three-dimensional region data specifically includes: Determine whether each edge of the Voronoi diagram intersects with the obstacle voxel set, delete the edges that intersect with the obstacle voxel set, and then form the remaining edges into a skeleton edge set. For each skeleton edge, calculate the minimum distance from all points on the skeleton edge to the obstacle voxel set, and delete skeleton edges whose minimum distance is less than a preset safety threshold. Then, form a path edge set with the remaining skeleton edges. Connect the preset starting point and preset ending point to the nearest path edge to obtain the starting point path and ending point path respectively. Merge the starting point path, ending point path and the set of path edges to obtain the path graph structure. The path graph structure is labeled with attributes to obtain the guiding structure; wherein the guiding structure is a sparse spatial graph.

[0007] Furthermore, the construction and training of the initial generative model specifically includes: The initial generation model is constructed based on a preset deep generation model framework. The initial generation model includes a generator and a discriminator. The generator includes a graph embedding encoder and a temporal decoding module, which are used to generate a path point sequence of the route according to the guidance structure. The discriminator is used to evaluate the path point sequence. The generator is trained based on the model input using a preset adversarial loss function; wherein the model input includes a preset starting point, a preset ending point, the operational constraint parameters, and the guiding structure, and the specific formula for the adversarial loss function is:

[0008] in, This represents the real path sample set. For the discriminator, For the generator, Let T be the sequence of path points, and T be the number of path points. This represents the i-th path point.

[0009] Preferably, the operational constraint parameters include, but are not limited to, the UAV's maximum climb angle, maximum pitch angle, minimum turning radius, and maximum rate of change of heading.

[0010] Furthermore, the step of training the initial generation model according to preset running constraint parameters to obtain a path generation model specifically includes: Based on the preset running constraint parameters and the path point sequence, a dynamic constraint verification function is constructed for each path point pair. And define a dynamic constraint loss term; wherein, the specific formula for the dynamic constraint loss term is:

[0011] in, The preset operating constraint parameters, This is a preset constraint loss function used to generate a positive loss value when the path violates dynamic constraints.

[0012] Define an adversarial training loss term, and calculate the overall loss function of the generator based on the dynamic constraint loss term and the adversarial training loss term; wherein the specific formula for the adversarial training loss term is:

[0013] The specific formula for the overall loss function is as follows:

[0014] in, Preset weighting coefficients; The generator is iteratively trained using the overall loss function, and the path generation model is obtained after training.

[0015] Furthermore, the step of generating a flight path through the path generation model based on the guidance structure specifically includes: The model input is fed into the path generation model, and the spatial features of the guiding structure are extracted by the graph embedding encoder. At the same time, the preset starting point, preset ending point and preset running constraint parameters are embedded into a unified latent space. Finally, the path point sequence that satisfies the spatial topology and constraint conditions is generated by the temporal decoding module and the discriminator. The flight path is generated based on the generated sequence of path points that satisfy the spatial topology and constraints.

[0016] Furthermore, the step of generating the flight path based on the generated path point sequence that satisfies the spatial topology and constraints specifically includes: The path point sequence is scanned to extract a set of continuous path segments; Based on the preset running constraint parameters, determine whether there are any continuous path segments in the set of continuous path segments that do not meet the feasibility conditions, and perform geometric repair on the continuous path segments that do not meet the feasibility conditions through spline interpolation. Global fitting is performed on the path points in the repaired continuous path segment set, and the fitted continuous path segment set is time-parameterized to output the route path.

[0017] Another embodiment of the present invention provides an automatic route generation device for narrow airspace, comprising: a guidance module, a training module, and a generation module; The guidance module is used to preprocess the three-dimensional region data of the target airspace and construct a guidance structure based on the preprocessed three-dimensional region data. The training module is used to construct and train an initial generation model, and then train the initial generation model according to preset running constraint parameters to obtain a path generation model; The generation module is used to generate a flight path based on the guidance structure and the path generation model.

[0018] Furthermore, the guidance module is used to preprocess the three-dimensional region data of the target spatial domain, specifically including: The three-dimensional region data of the target airspace is acquired, and the three-dimensional region data is subjected to projection, resampling and rasterization operations to obtain a three-dimensional voxel mesh model; wherein, the three-dimensional region data includes three-dimensional obstacle data and free space data; Based on the three-dimensional obstacle data, the obstacle model in the three-dimensional voxel mesh model is dilated to obtain an obstacle voxel set; Based on the free space data, a three-dimensional Voronoi diagram is constructed.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention extracts the skeleton of potential passable areas based on a 3D Voronoi diagram, constructs a guidance structure with spatial continuity and safety margin, and generates a flight path that meets dynamic execution requirements based on the guidance structure and pre-collected operational constraint parameters. This can improve the feasibility, safety and intelligence of path planning for UAVs in typical narrow passage scenarios. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an automatic route generation method for narrow airspace provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of an automatic route generation device for narrow airspace provided in another embodiment of the present invention. Detailed Implementation

[0022] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Reference Figure 1 The above is a flowchart illustrating an automatic route generation method for narrow airspace provided in an embodiment of the present invention, comprising the following steps: S1: Preprocess the three-dimensional region data of the target airspace, and construct a guiding structure based on the preprocessed three-dimensional region data; S2: Construct and train an initial generation model, and then train the initial generation model according to preset running constraint parameters to obtain a path generation model; S3: Generate a flight path using the path generation model based on the guidance structure.

[0025] For step S1, specifically, the preprocessing of the three-dimensional region data of the target spatial domain includes: The three-dimensional region data of the target airspace is acquired, and the three-dimensional region data is subjected to projection, resampling and rasterization operations to obtain a three-dimensional voxel mesh model; wherein, the three-dimensional region data includes three-dimensional obstacle data and free space data; Based on the three-dimensional obstacle data, the obstacle model in the three-dimensional voxel mesh model is dilated to obtain an obstacle voxel set; Based on the free space data, a three-dimensional Voronoi diagram is constructed.

[0026] In a preferred embodiment, after acquiring the three-dimensional obstacle data and free space data of the target airspace, all data can be uniformly projected, resampled, and rasterized into a three-dimensional voxel mesh model. , where 1 represents a barrier voxel and 0 represents a free voxel.

[0027] Subsequently, the obstacle model underwent morphological dilation, with the dilation radius set to [value missing]. ,in To obtain the minimum safe redundancy distance for the aircraft, the expanded obstacle voxel set is obtained. ,in, Represents a spherical structural element.

[0028] Finally, a set of Voronoi seed points is set up in free space. This includes boundary points, obstacle surface points, and points within the neighborhood of the start and end points, used to construct a 3D Voronoi diagram. The three-dimensional Voronoi diagram consists of several three-dimensional edges, each edge being an equidistant central axis segment between two Voronoi seed points.

[0029] Furthermore, the step of constructing the guiding structure based on the preprocessed three-dimensional region data specifically includes: Determine whether each edge of the Voronoi diagram intersects with the obstacle voxel set, delete the edges that intersect with the obstacle voxel set, and then form the remaining edges into a skeleton edge set. For each skeleton edge, calculate the minimum distance from all points on the skeleton edge to the obstacle voxel set, and delete skeleton edges whose minimum distance is less than a preset safety threshold. Then, form a path edge set with the remaining skeleton edges. Connect the preset starting point and preset ending point to the nearest path edge to obtain the starting point path and ending point path respectively. Merge the starting point path, ending point path and the set of path edges to obtain the path graph structure. The path graph structure is labeled with attributes to obtain the guiding structure; wherein the guiding structure is a sparse spatial graph.

[0030] In a preferred embodiment, after obtaining the three-dimensional Voronoi diagram, it is necessary to determine whether each edge in the diagram is related to the obstacle voxel set. If any edges intersect, the intersecting edge is removed, and the remaining edges form the skeleton edge set of the passable region. .

[0031] Subsequently, each of the remaining skeleton edges was... Calculate the minimum distance from all its points to the obstacle voxel, and define its safety margin as... and for those below the minimum safety margin threshold By marking or deleting the skeleton edges, we obtain a set of path edges.

[0032] Next, the starting point and the ending point are connected to the nearest points on the path edges, forming connection paths from the starting point to the skeleton and from the ending point to the skeleton. By merging these two connection paths with the set of path edges, a connected path graph structure covering the entire task area is constructed, which contains all traversable path segments and their three-dimensional connection relationships.

[0033] Finally, attribute annotation and data organization are performed on the path graph structure to obtain the guiding structure for the path generation model. The guiding structure is a sparse spatial graph. , where V is the set of discrete coordinate points in three-dimensional space, and E is the set of traversable edges between points. Each edge has attributes such as safety margin and distance, which are used to represent spatial connectivity and local safety.

[0034] For step S2, specifically, the construction and training of the initial generative model includes: The initial generation model is constructed based on a preset deep generation model framework. The initial generation model includes a generator and a discriminator. The generator includes a graph embedding encoder and a temporal decoding module, which are used to generate a path point sequence of the route according to the guidance structure. The discriminator is used to evaluate the path point sequence. The generator is trained based on the model input using a preset adversarial loss function; wherein the model input includes a preset starting point, a preset ending point, the operational constraint parameters, and the guiding structure, and the specific formula for the adversarial loss function is:

[0035] in, This represents the real path sample set. For the discriminator, For the generator, Let T be the sequence of path points, and T be the number of path points. This represents the i-th path point.

[0036] Preferably, the operational constraint parameters include, but are not limited to, the UAV's maximum climb angle, maximum pitch angle, minimum turning radius, and maximum rate of change of heading.

[0037] In a preferred embodiment, a path generative adversarial network model (i.e., the initial generative model) is constructed based on a deep generative model framework, including a generator. and discriminator ,in and Each network represents its own set of parameters. The model input includes the flight mission start point s, end point t, preset operational constraint parameters C, and the guidance structure G. The preset operational constraint parameters C include the maximum climb angle. Maximum pitch angle Minimum turning radius Maximum rate of change of heading wait.

[0038] The generator This is used to extract spatial features from the guiding structure G using a graph embedding encoder, while embedding the starting point s, the ending point t, and the preset running constraint parameters C into the same latent space. Finally, it is combined with a temporal decoding module to generate a path point sequence that satisfies the spatial topology and constraint conditions. , , Where T is the number of path points. This represents the i-th 3D path point.

[0039] The discriminator Used for generating the path point sequence The evaluation takes the current path point sequence and its contextual information (including motion features and guiding structures) as input, and outputs a feasibility score for the path. This is used to determine whether the path satisfies spatial coherence and physical constraints.

[0040] During the training process of the initial generative model, this preferred embodiment employs an adversarial loss function to guide the generator in continuously optimizing the path structure, thereby enhancing its feasibility and realism under the discriminator's evaluation and satisfying the joint requirements of path shape rationality and physical constraint consistency. The training objective is:

[0041] in, This represents the real path sample set, used to supervise the decision boundary of the discriminator in the early stages of training.

[0042] Furthermore, the step of training the initial generation model according to preset running constraint parameters to obtain a path generation model specifically includes: Based on the preset running constraint parameters and the path point sequence, a dynamic constraint verification function is constructed for each path point pair. And define a dynamic constraint loss term; wherein, the specific formula for the dynamic constraint loss term is:

[0043] in, The preset operating constraint parameters, This is a preset constraint loss function used to generate a positive loss value when the path violates dynamic constraints.

[0044] Define an adversarial training loss term, and calculate the overall loss function of the generator based on the dynamic constraint loss term and the adversarial training loss term; wherein the specific formula for the adversarial training loss term is:

[0045] The specific formula for the overall loss function is as follows:

[0046] in, Preset weighting coefficients; The generator is iteratively trained using the overall loss function, and the path generation model is obtained after training.

[0047] In a preferred embodiment, by introducing a constraint loss term based on the operational constraint parameters, paths that violate the aircraft's dynamic characteristics can be updated, thereby guiding the model to generate a path sequence that meets the dynamic execution requirements. Specific training steps include: Based on the preset running constraint parameter C and the generated path point sequence Based on the spatial relationship between path points, calculate the geometric and motion characteristics of their adjacent point segments; Construct a dynamic constraint check function for each path pair. This is used to measure whether the path violates the aircraft's dynamic constraints on that segment. The overall dynamic constraint loss term for the path is defined as: .in, This is a constraint loss function used to generate a positive loss value when a path violates dynamic constraints.

[0048] The dynamic constraint loss term Compared to the loss items in adversarial training Perform joint optimization to form the overall loss function of the generator. ,in , where is the weighting coefficient, used to adjust the intensity of the influence of dynamic constraints on generator training. The adversarial training loss term is the loss for the generator learning to minimize the generation path that is judged as "false," and the specific formula is: .

[0049] During training, the generator continuously adjusts its parameters based on the joint loss, gradually reducing the degree of violation of dynamic constraints while improving the coherence and spatial rationality of the generated paths. Through iterative training, the model can eventually generate a sequence of path points that meets the actual dynamic execution requirements of the aircraft, demonstrating feasibility and executability, thus obtaining the path generation model.

[0050] For step S3, specifically, generating the flight path through the path generation model based on the guidance structure includes: The model input is fed into the path generation model, and the spatial features of the guiding structure are extracted by the graph embedding encoder. At the same time, the preset starting point, preset ending point and preset running constraint parameters are embedded into a unified latent space. Finally, the path point sequence that satisfies the spatial topology and constraint conditions is generated by the temporal decoding module and the discriminator. The flight path is generated based on the generated sequence of path points that satisfy the spatial topology and constraints.

[0051] Furthermore, the step of generating the flight path based on the generated path point sequence that satisfies the spatial topology and constraints specifically includes: The path point sequence is scanned to extract a set of continuous path segments; Based on the preset running constraint parameters, determine whether there are any continuous path segments in the set of continuous path segments that do not meet the feasibility conditions, and perform geometric repair on the continuous path segments that do not meet the feasibility conditions through spline interpolation. Global fitting is performed on the path points in the repaired continuous path segment set, and the fitted continuous path segment set is time-parameterized to output the route path.

[0052] In a preferred embodiment, after obtaining the pathpoint sequence, a smoothing, repair, and fitting strategy is applied to the pathpoint sequence to correct local violations and add time labels, thereby generating a continuous and executable route. Specific generation steps include: The 3D path point sequence output by the path generative adversarial network is scanned to extract all continuous path segments. Based on the preset operational constraint parameter C, it is determined whether there are any local segments that violate the conditions. For path segments that do not meet the dynamic feasibility conditions, geometric repair is performed using spline interpolation to construct smoothly connected curve segments. to replace the original path segment At the same time, the overall structure of the path remains consistent.

[0053] A unified path smoothing strategy is applied to fit the path points globally, ensuring the continuity of position and velocity. During the fitting process, the selection of interpolation control points takes into account both the original path shape and the boundary conditions of the correction area, avoiding large offsets or newly introduced path anomalies.

[0054] For the repaired and fitted pathpoint sequence, time parameterization is performed according to the aircraft's speed performance and the allowable range of maximum trajectory tracking error, and time labels are assigned to each pathpoint. This makes the path point sequence This forms a time-constrained air route.

[0055] The corrected path data is organized according to a predetermined data structure and output as continuous, dynamically executable three-dimensional route information, thus obtaining the route path.

[0056] Furthermore, the present invention also provides specific application embodiments. To verify the effectiveness of the method of the present invention, a test scenario was simulated, and the specific parameters are as follows: 1. Test area: 1×1×1km 3 The urban environment is characterized by narrow passages formed by typical urban obstacles such as buildings.

[0057] 2. Aircraft operating constraints: maximum speed 15 m / s, maximum climb angle 30°, minimum turning radius 20 m, maximum heading change rate 30° / s.

[0058] 3. Flight mission: Fly from the starting point (coordinates: 100, 100, 50) to the destination (coordinates: 900, 800, 70), traversing a narrow passage.

[0059] 4. Minimum safety redundancy distance: 5 m; Minimum safety margin threshold: 6 m.

[0060] 5. Implementation process: 1) Guiding Structure Construction: Obstacle data is rasterized into a 3D voxel model. Morphological dilation (r = 5 m) was performed to obtain O′.

[0061] A Voronoi seed point set P is set up, including 12,000 points in total, including building edge points and starting / ending point neighborhood sampling points.

[0062] Construct a 3D Voronoi diagram, remove edges that intersect with expansion barriers, and retain the set of safe skeleton edges Esafe, which contains 8342 edges.

[0063] Calculate the safety margin μ(ei) of each edge, delete edges with a margin less than 6m, and finally generate a sparse guided graph G=(V,E) with 4176 nodes and 6210 edges.

[0064] The starting point and the ending point have been successfully connected to the guidance graph, forming a connected path graph.

[0065] 2) Training the Path Generative Adversarial Network: Spatial features of the guiding structure G are extracted using a graph embedding encoder based on a graph attention network. Simultaneously, the starting point s, the ending point t, and the set of running constraint parameters C are embedded into the same latent space. A temporal decoding module based on a long short-term memory artificial neural network is then used to generate a path point sequence that satisfies the spatial topology and constraint conditions. .

[0066] The discriminator For the generated path point sequence The evaluation takes the current path point sequence and its contextual information (including motion features and guiding structures) as input, and outputs a feasibility score for the path. This is used to determine whether the path satisfies spatial coherence and physical constraints.

[0067] Introducing dynamic constraint loss terms Weight .

[0068] Pre-training was performed using a real path sample set, for a total of 5000 training rounds.

[0069] 3) Path post-processing and output: The generated path is smoothed using B-spline and local violation segments are repaired.

[0070] The time parameter is calculated based on the maximum permissible speed, time tags are added, and an executable route is generated.

[0071] 6. Experimental Results (100 tasks)

[0072] Therefore, the method provided by the embodiments of the present invention enables UAVs to stably generate feasible paths even in narrow passages. Compared with the A* algorithm and the unconstrained GAN method, it can ensure that the path is far away from obstacles, with an average minimum distance of 7.1m, which is significantly higher than the safety threshold, and the path transitions with a smooth arc, which is in line with the aircraft's maneuverability.

[0073] Reference Figure 2 The diagram below shows a structural schematic of an automatic route generation device for narrow airspace provided in another embodiment of the present invention, which includes: a guidance module 101, a training module 102 and a generation module 103; The guidance module 101 is used to preprocess the three-dimensional region data of the target airspace and construct a guidance structure based on the preprocessed three-dimensional region data; The training module 102 is used to construct and train an initial generation model, and then train the initial generation model according to preset running constraint parameters to obtain a path generation model; The generation module 103 is used to generate a flight path based on the guidance structure and the path generation model.

[0074] Furthermore, the guidance module 101 is used to preprocess the three-dimensional region data of the target airspace, specifically including: The three-dimensional region data of the target airspace is acquired, and the three-dimensional region data is subjected to projection, resampling and rasterization operations to obtain a three-dimensional voxel mesh model; wherein, the three-dimensional region data includes three-dimensional obstacle data and free space data; Based on the three-dimensional obstacle data, the obstacle model in the three-dimensional voxel mesh model is dilated to obtain an obstacle voxel set; Based on the free space data, a three-dimensional Voronoi diagram is constructed.

[0075] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for automatically generating flight paths in narrow airspace, characterized in that, The steps include the following: The three-dimensional region data of the target airspace is preprocessed, and a guiding structure is constructed based on the preprocessed three-dimensional region data; wherein, the preprocessing of the three-dimensional region data of the target airspace specifically includes: The three-dimensional region data of the target airspace is acquired, and the three-dimensional region data is subjected to projection, resampling and rasterization operations to obtain a three-dimensional voxel mesh model; wherein, the three-dimensional region data includes three-dimensional obstacle data and free space data; Based on the three-dimensional obstacle data, the obstacle model in the three-dimensional voxel mesh model is dilated to obtain an obstacle voxel set; Based on the free space data, construct a three-dimensional Voronoi diagram; The construction of the guiding structure based on the preprocessed three-dimensional region data specifically includes: Determine whether each edge of the Voronoi diagram intersects with the obstacle voxel set, delete the edges that intersect with the obstacle voxel set, and then form the remaining edges into a skeleton edge set. For each skeleton edge, calculate the minimum distance from all points on the skeleton edge to the obstacle voxel set, and delete skeleton edges whose minimum distance is less than a preset safety threshold. Then, form a path edge set with the remaining skeleton edges. Connect the preset starting point and preset ending point to the nearest path edge to obtain the starting point path and ending point path respectively. Merge the starting point path, ending point path and the set of path edges to obtain the path graph structure. The path graph structure is labeled with attributes to obtain the guiding structure; wherein, the guiding structure is a sparse spatial graph; An initial generation model is constructed and trained, and then trained according to preset running constraint parameters to obtain a path generation model; wherein, the construction and training of the initial generation model specifically includes: The initial generation model is constructed based on a preset deep generation model framework. The initial generation model includes a generator and a discriminator. The generator includes a graph embedding encoder and a temporal decoding module, used to generate a path point sequence of the flight path according to the guidance structure. The discriminator is used to evaluate the path point sequence. The generator is trained based on the model input using a preset adversarial loss function; wherein the model input includes a preset starting point, a preset ending point, the operational constraint parameters, and the guiding structure, and the specific formula for the adversarial loss function is: in, This represents the real path sample set. For the discriminator, For the generator, Let T be the sequence of path points, and T be the number of path points. Represents the i-th path point; The step of training the initial generation model according to preset running constraint parameters to obtain the path generation model specifically includes: Based on the preset running constraint parameters and the path point sequence, a dynamic constraint verification function is constructed for each path point pair. And define a dynamic constraint loss term; wherein, the specific formula for the dynamic constraint loss term is: in, The preset operating constraint parameters, This is a preset constraint loss function used to generate a positive loss value when a path violates dynamic constraints. Define an adversarial training loss term, and calculate the overall loss function of the generator based on the dynamic constraint loss term and the adversarial training loss term; wherein the specific formula for the adversarial training loss term is: The specific formula for the overall loss function is as follows: in, Preset weighting coefficients; The generator is iteratively trained using the overall loss function, and the path generation model is obtained after training. Based on the guidance structure, a flight path is generated through the path generation model.

2. The method for automatically generating airways in narrow airspaces as described in claim 1, characterized in that, The operational constraints include, but are not limited to, the UAV's maximum climb angle, maximum pitch angle, minimum turning radius, and maximum rate of change of heading.

3. The method for automatically generating airways in narrow airspaces as described in claim 1, characterized in that, The step of generating a flight path using the path generation model based on the guidance structure specifically includes: The model input is fed into the path generation model, and the spatial features of the guiding structure are extracted by the graph embedding encoder. At the same time, the preset starting point, preset ending point and preset running constraint parameters are embedded into a unified latent space. Finally, the path point sequence that satisfies the spatial topology and constraint conditions is generated by the temporal decoding module and the discriminator. The flight path is generated based on the generated sequence of path points that satisfy the spatial topology and constraints.

4. The method for automatically generating airways in narrow airspaces as described in claim 3, characterized in that, The step of generating the flight path based on the generated path point sequence that satisfies the spatial topology and constraints specifically includes: The path point sequence is scanned to extract a set of continuous path segments; Based on the preset running constraint parameters, determine whether there are any continuous path segments in the set of continuous path segments that do not meet the feasibility conditions, and perform geometric repair on the continuous path segments that do not meet the feasibility conditions through spline interpolation. Global fitting is performed on the path points in the repaired continuous path segment set, and the fitted continuous path segment set is time-parameterized to output the route path.

5. An automatic route generation device for narrow airspace, characterized in that, include: The module consists of a bootstrapping module, a training module, and a generation module. The guidance module is used to preprocess the three-dimensional region data of the target spatial domain and construct a guidance structure based on the preprocessed three-dimensional region data; wherein, the preprocessing of the three-dimensional region data of the target spatial domain specifically includes: The three-dimensional region data of the target airspace is acquired, and the three-dimensional region data is subjected to projection, resampling and rasterization operations to obtain a three-dimensional voxel mesh model; wherein, the three-dimensional region data includes three-dimensional obstacle data and free space data; Based on the three-dimensional obstacle data, the obstacle model in the three-dimensional voxel mesh model is dilated to obtain an obstacle voxel set; Based on the free space data, construct a three-dimensional Voronoi diagram; The construction of the guiding structure based on the preprocessed three-dimensional region data specifically includes: Determine whether each edge of the Voronoi diagram intersects with the obstacle voxel set, delete the edges that intersect with the obstacle voxel set, and then form the remaining edges into a skeleton edge set. For each skeleton edge, calculate the minimum distance from all points on the skeleton edge to the obstacle voxel set, and delete skeleton edges whose minimum distance is less than a preset safety threshold. Then, form a path edge set with the remaining skeleton edges. Connect the preset starting point and preset ending point to the nearest path edge to obtain the starting point path and ending point path respectively. Merge the starting point path, ending point path and the set of path edges to obtain the path graph structure. The path graph structure is labeled with attributes to obtain the guiding structure; wherein, the guiding structure is a sparse spatial graph; The training module is used to construct and train an initial generation model, and then train the initial generation model according to preset running constraint parameters to obtain a path generation model; wherein, the construction and training of the initial generation model specifically includes: The initial generation model is constructed based on a preset deep generation model framework. The initial generation model includes a generator and a discriminator. The generator includes a graph embedding encoder and a temporal decoding module, used to generate a path point sequence of the flight path according to the guidance structure. The discriminator is used to evaluate the path point sequence. The generator is trained based on the model input using a preset adversarial loss function; wherein the model input includes a preset starting point, a preset ending point, the operational constraint parameters, and the guiding structure, and the specific formula for the adversarial loss function is: in, This represents the real path sample set. For the discriminator, For the generator, Let T be the sequence of path points, and T be the number of path points. Represents the i-th path point; The step of training the initial generation model according to preset running constraint parameters to obtain the path generation model specifically includes: Based on the preset running constraint parameters and the path point sequence, a dynamic constraint verification function is constructed for each path point pair. And define a dynamic constraint loss term; wherein, the specific formula for the dynamic constraint loss term is: in, The preset operating constraint parameters, This is a preset constraint loss function used to generate a positive loss value when a path violates dynamic constraints. Define an adversarial training loss term, and calculate the overall loss function of the generator based on the dynamic constraint loss term and the adversarial training loss term; wherein the specific formula for the adversarial training loss term is: The specific formula for the overall loss function is as follows: in, Preset weighting coefficients; The generator is iteratively trained using the overall loss function, and the path generation model is obtained after training. The generation module is used to generate a flight path based on the guidance structure and the path generation model.

6. The automatic route generation device for narrow airspace as described in claim 5, characterized in that, The generation module is used to generate a flight path based on the guidance structure and the path generation model, specifically including: The model input is fed into the path generation model, and the spatial features of the guiding structure are extracted by the graph embedding encoder. At the same time, the preset starting point, preset ending point and preset running constraint parameters are embedded into a unified latent space. Finally, the path point sequence that satisfies the spatial topology and constraint conditions is generated by the temporal decoding module and the discriminator. The flight path is generated based on the generated sequence of path points that satisfy the spatial topology and constraints.

Images