Unmanned aerial vehicle system capable of carrying sampling robot and control method

Abstract

The invention relates to the technical field of unmanned aerial vehicles, and discloses an unmanned aerial vehicle system capable of carrying a sampling robot and a control method, and the system comprises an environment sensing module, a topographic feature analysis module, a path planning feature enhancement module, a task scheduling decision module and a control instruction generation module. The environment sensing module obtains real-time topographic data through a multi-mode sensor; the topographic feature analysis module is used for carrying out grid segmentation on the data and extracting a feature tensor by utilizing a three-dimensional convolutional neural network; the path planning feature enhancement module optimizes the spatial relevance of the feature tensor; the task scheduling decision module generates a global scheduling decision feature graph based on the information entropy distribution fusion features; the control instruction generation module generates a control signal according to the feature map. The method comprises the steps of data acquisition, feature extraction, feature optimization, feature fusion, instruction generation and the like. The working efficiency and precision of the unmanned aerial vehicle carrying the sampling robot under the complex terrain are improved, and the method is suitable for the fields of environment monitoring, geological exploration and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and particularly to an unmanned aerial vehicle system capable of carrying a sampling robot and a control method therefor. Background Art

[0002] In the fields of environmental monitoring, geological exploration, disaster emergency, etc., it is of great significance to perform fast and accurate sampling operations in complex terrain areas. Traditional manual sampling methods face problems such as low efficiency, high danger, and obvious terrain limitations, making it difficult to meet the sampling requirements in large-scale and complex environments. The development of unmanned aerial vehicle technology provides new ideas for sampling operations. Its aerial flight ability can break through terrain obstacles and achieve rapid arrival at the target area. However, there are still many technical bottlenecks in existing unmanned aerial vehicle systems capable of carrying sampling robots in practical applications.

[0003] From the perspective of environmental perception and terrain analysis, complex terrain environments have a high degree of spatial heterogeneity. Traditional single-sensor perception methods are difficult to comprehensively and accurately obtain terrain data, and most of the processing of terrain data stays at simple rasterization division, lacking in-depth extraction of terrain features and being unable to effectively represent the local area characteristics of the terrain environment, resulting in a lack of reliable environmental information support for subsequent path planning and task scheduling.

[0004] In terms of path planning, existing systems often ignore the spatial correlation between local area characteristics of the terrain environment. Terrain features not only include local information such as height and slope, but also involve the spatial relationship between adjacent areas, such as the continuity of the terrain and the correlation of undulation changes. Traditional path planning methods only perform independent analysis based on local terrain features and do not fully consider the mutual influence between regions, easily resulting in the lack of rationality of the planned path at the global level, and problems such as detouring of the path and encountering insurmountable terrain obstacles may occur, affecting the deployment efficiency and safety of the sampling robot.

[0005] The task scheduling decision-making link is the key to the efficient sampling of the unmanned aerial vehicle system. In existing technologies, the fusion of terrain environment characteristics mostly uses simple methods such as weighted average, and fails to fully consider the information entropy distribution of different region characteristics, that is, the importance and uncertainty of different regions in the sampling task. This makes the task scheduling decision lack scientificity and unable to dynamically adjust the deployment strategy of the sampling robot according to the actual situation of the terrain environment, easily resulting in unreasonable allocation of sampling resources, phenomena such as key areas not being effectively covered and waste of resources in non-critical areas.

[0006] In terms of control instruction generation, traditional control signal generation methods are usually based on fixed rules or simple algorithms, lacking the comprehensive analysis of global environmental characteristics and the ability of dynamic response. In a complex and changeable terrain environment, it is difficult to generate accurate control instructions according to the real-time updated environmental information, resulting in an inflexible deployment strategy for the sampling robot, being unable to adapt to the real-time changes of the terrain environment and affecting the overall effect of the sampling task. Summary of the Invention

[0007] The purpose of the present invention is to provide an unmanned aerial vehicle system and a control method that can carry a sampling robot to solve the problems proposed in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: An unmanned aerial vehicle system that can carry a sampling robot, the system includes: An environmental perception module, which is used to obtain real-time terrain data of the target area through multi-modal sensors; A terrain feature analysis module, which is used to perform grid segmentation and terrain feature extraction on the real-time terrain data to obtain a set of local area feature tensors of the terrain environment; A path planning feature enhancement module, which is used to optimize the spatial correlation features of the set of local area feature tensors of the terrain environment to obtain a set of optimized local area feature tensors of the terrain environment; A task scheduling decision module, which is used to perform global feature fusion based on information entropy distribution on the set of optimized local area feature tensors of the terrain environment to obtain a global environmental scheduling decision feature map; A control instruction generation module, which is used to generate a robot carrying control signal according to the global environmental scheduling decision feature map.

[0009] Preferably, the terrain feature analysis module includes: a terrain data segmentation unit, which is used to spatially divide the real-time terrain data according to a preset grid size to obtain a set of terrain data units; a terrain feature encoding unit, which is used to input the unit data in the set of terrain data units into a terrain feature encoder based on a three-dimensional convolutional neural network to obtain the set of local area feature tensors of the terrain environment.

[0010] Preferably, the path planning feature enhancement module includes: a feature tensor decomposition unit for performing tensor expansion on the local area feature tensor of the terrain environment along the spatial dimension to obtain a set of local area feature vectors of the terrain environment; a regional feature correlation degree calculation unit for calculating the spatial correlation degree between any two feature vectors in the set of local area feature vectors of the terrain environment to generate a terrain environment regional correlation degree matrix; a regional space constraint correction unit for performing distance attenuation correction on the terrain environment regional correlation degree matrix according to the geometric distance between any two feature vectors in the set of local area feature vectors of the terrain environment to obtain a terrain environment regional constraint correlation degree matrix; and a regional feature optimization unit for performing multi-scale convolution fusion on the terrain environment regional constraint correlation degree matrix and the set of local area feature vectors of the terrain environment to obtain the optimized feature tensor of the local area of the terrain environment.

[0011] Preferably, the regional feature correlation degree calculation unit includes: a feature space mapping sub-unit for mapping each feature vector in the set of local area feature vectors of the terrain environment to a manifold space to obtain a set of mapped local area feature vectors of the terrain environment; and a feature correlation analysis sub-unit for calculating the cosine similarity between any two mapped feature vectors in the set of mapped local area feature vectors of the terrain environment to generate the terrain environment regional correlation degree matrix composed of multiple terrain environment regional correlation degree values.

[0012] Preferably, the regional feature optimization unit is specifically implemented as: performing dilated convolution processing on the terrain environment regional constraint correlation degree matrix to obtain a terrain environment regional constraint correlation feature matrix; inputting the set of local area feature vectors of the terrain environment and the terrain environment regional constraint correlation feature matrix into an attention encoding network to obtain a set of local area context feature vectors of the terrain environment; and performing tensor reconstruction on the set of local area context feature vectors of the terrain environment to obtain the optimized feature tensor of the local area of the terrain environment.

[0013] Preferably, the task scheduling decision module includes: a local feature dimensionality reduction unit, configured to perform max pooling processing on each tensor in the set of optimized feature tensors of the local area of the terrain environment in the channel dimension to obtain a set of optimized feature vectors of the local area of the terrain environment; a regional feature static information entropy calculation unit, configured to calculate the information entropy value of each feature vector in the set of optimized feature vectors of the local area of the terrain environment to generate a set of terrain environment static information entropy; a regional clustering center determination unit, configured to select the optimized feature vector of the local area of the terrain environment corresponding to the maximum information entropy value in the set of terrain environment static information entropy as the initial scheduling center vector; a regional feature dynamic weight calculation unit, configured to calculate the dynamic scheduling weight of each feature vector according to the spatial distance between each feature vector in the set of optimized feature vectors of the local area of the terrain environment and the initial scheduling center vector and the information entropy value of each feature vector to generate a set of dynamic scheduling weights; a global feature fusion unit, configured to perform weighted summation on the set of optimized feature vectors of the local area of the terrain environment by using the set of dynamic scheduling weights to generate the global environment scheduling decision feature map.

[0014] Preferably, the regional feature static information entropy calculation unit is specifically implemented as: calculating the mean vector and variance vector of the optimized feature vectors of the local area of the terrain environment; performing element-wise difference calculation on the optimized feature vectors of the local area of the terrain environment and the mean vector, and performing a fourth-power operation on the difference result to obtain a terrain environment feature difference vector; calculating the overall expected value of the terrain environment feature difference vector; performing a ratio operation on the expected value and the square value of the variance vector, and inputting the result into a normalization function to obtain the information entropy value.

[0015] Preferably, the regional feature dynamic weight calculation unit is specifically implemented as: multiplying the information entropy value of the optimized feature vector of the local area of the terrain environment and the information entropy value of the initial scheduling center vector by a first adjustment coefficient to obtain a first dynamic scheduling factor; multiplying the square value of the Euclidean distance between the optimized feature vector of the local area of the terrain environment and the initial scheduling center vector by a second adjustment coefficient to obtain a second dynamic scheduling factor; performing a ratio operation on the first dynamic scheduling factor and the second dynamic scheduling factor to obtain the dynamic scheduling weight.

[0016] Preferably, the control instruction generation module is specifically implemented as: inputting the global environment scheduling decision feature map into a control signal generator based on a decision tree to obtain the robot-mounted control signal, and the signal is used to indicate the deployment strategy of the sampling robot.

[0017] Preferably, the present invention further includes a method for controlling a drone that can carry a sampling robot, and the method includes: collecting real-time terrain data of a target area through a multi-modal sensor; performing grid segmentation and terrain feature extraction on the real-time terrain data to obtain a set of local terrain environment feature tensors; performing spatial correlation feature optimization on the set of local terrain environment feature tensors to obtain a set of optimized local terrain environment feature tensors; performing global feature fusion based on information entropy distribution on the set of optimized local terrain environment feature tensors to obtain a global environment scheduling decision feature map; generating a control signal for robot carrying according to the global environment scheduling decision feature map.

[0018] Compared with the prior art, the beneficial effects of the present invention are: In terms of environmental perception and terrain feature analysis, by obtaining real-time terrain data of the target area through a multi-modal sensor, the advantages of multiple sensors can be comprehensively utilized to comprehensively and accurately perceive the multi-dimensional information of the terrain environment, avoiding the limitations of a single sensor. The terrain feature analysis module performs grid segmentation on the real-time terrain data and uses a three-dimensional convolutional neural network to perform feature encoding on each unit of data, realizing the in-depth processing of terrain data and the efficient extraction of terrain features. This feature extraction method based on deep learning can automatically capture complex features in the terrain environment, such as terrain undulations, slope changes, obstacle distributions, etc., and generate a set of local terrain environment feature tensors, providing rich and accurate basic data for subsequent path planning and task scheduling.

[0019] The path planning feature enhancement module effectively improves the rationality and reliability of path planning by performing spatial correlation feature optimization on the local terrain environment feature tensors. The module first unfolds the feature tensors along the spatial dimension to obtain a set of feature vectors, and then generates a terrain environment region constraint correlation matrix by calculating the spatial correlation degree between feature vectors and combining geometric distance for distance attenuation correction. On this basis, through operations such as multi-scale convolution fusion and attention encoding network, the spatial correlation optimization of the local terrain environment features is realized, so that the optimized feature tensors can more accurately reflect the spatial relationship and mutual influence between regions in the terrain environment. This helps the drone to not only consider local terrain features but also comprehensively evaluate the feasibility of the path from a global perspective during path planning, avoiding unreasonable path planning problems caused by ignoring the inter-regional correlation, and improving the scientificity and safety of the deployment path of the sampling robot.

[0020] The task scheduling decision module performs global feature fusion based on the information entropy distribution, achieving scientific scheduling of terrain environment features and reasonable allocation of resources. By calculating the information entropy value of the optimized feature vector in the local area of the terrain environment, it can effectively identify key areas with rich information and high uncertainty in the terrain environment, and use them as the initial scheduling center vectors. Then, the dynamic scheduling weights are calculated based on the spatial distance and information entropy value between each feature vector and the initial scheduling center vector, and the feature vectors are weighted and summed to generate the global environment scheduling decision feature map. This global feature fusion method based on information entropy and dynamic weights can fully consider the importance and relevance of different regions in the sampling task, realize the dynamic optimization configuration of sampling resources, ensure that key areas are preferentially covered and sampled, improve the pertinence and effectiveness of the sampling task, and avoid waste of sampling resources.

[0021] The control instruction generation module inputs the global environment scheduling decision feature map into the decision tree-based control signal generator, which can generate accurate robot carrying control signals based on the global environment features to indicate the deployment strategy of the sampling robot. The decision tree algorithm has good interpretability and decision-making ability, and can quickly make reasonable decisions according to different combinations of environmental features and generate corresponding control instructions. This way of generating control instructions based on global features enables the UAV system to dynamically adjust the deployment strategy of the sampling robot according to the real-time updated terrain environment information, improves the adaptability and response ability of the system to complex environments, and ensures that the sampling task can be completed efficiently and accurately.

[0022] In addition, through the organic combination and collaborative work among various modules of the present invention, a complete technical chain from environmental perception, feature analysis, path planning, task scheduling to control instruction generation is formed, realizing the full-process optimization of the UAV system capable of carrying a sampling robot. This system can automatically and efficiently complete the task of carrying and deploying the sampling robot in a complex terrain environment, significantly reducing the labor cost, improving the safety and reliability of the sampling operation, and having broad application prospects and important practical application values. Brief Description of the Drawings

[0023] Figure 1 It is the working principle diagram of the UAV system capable of carrying a sampling robot according to the present invention; Figure 2 It is the flowchart of the UAV control method; Figure 3 It is the design diagram of the path planning feature enhancement module. Detailed Description of the Preferred Embodiment

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0025] Please refer to Figures 1 - 3 , a drone system capable of carrying a sampling robot according to the present invention, the system includes an environmental perception module, a terrain feature analysis module, a path planning feature enhancement module, a task scheduling decision module, and a control instruction generation module, and the specific implementation is as follows: Environmental perception module: Real-time collect terrain data of the target area through multi-modal sensors (such as lidar, visual cameras, inertial sensors, etc.), covering multi-dimensional information such as terrain height, slope, and obstacle distribution, and form a real-time terrain data matrix including spatial coordinates and physical attributes to provide raw data support for subsequent processing.

[0026] Terrain feature analysis module: Perform grid segmentation and feature extraction on the real-time terrain data. First, divide the real-time terrain data into multiple terrain data units according to a preset grid size (such as 0.5 m × 0.5 m) to form a terrain data unit set. Subsequently, input the data of each unit into a terrain feature encoder based on a three-dimensional convolutional neural network (3D-CNN), and extract the depth features in the spatial dimension through multiple convolutional kernels (such as 3×3×3 convolutional kernels) to capture the local geometric features of the terrain (such as protrusions, depressions, flat areas, etc.), and output a set of terrain environment local area feature tensors including spatial positions and feature dimensions, and each tensor corresponds to the feature representation of a grid unit.

[0027] Path planning feature enhancement module: Optimize the spatial correlation of the set of terrain environment local area feature tensors. By expanding the spatial dimension of the feature tensors, calculating the correlation degree, correcting the distance constraint, and performing multi-scale convolution fusion, enhance the spatial dependence relationship between features, and generate a set of optimized terrain environment local area feature tensors to provide a more contextually related feature expression for path planning.

[0028] Task scheduling decision module: Perform global feature fusion on the optimized set of feature tensors based on the information entropy distribution. Determine the information-rich area by calculating the information entropy value of each local feature, use the feature vector corresponding to the maximum information entropy value as the initial scheduling center, calculate the dynamic weight by combining the spatial distance and the information entropy value, perform weighted summation on the local features, and generate a global environment scheduling decision feature map to achieve the global overall planning of the sampling task.

[0029] Control instruction generation module: Input the global environment scheduling decision feature map into the decision tree-based control signal generator, perform logical judgment through preset decision rules (such as node parameters like terrain risk level and sampling priority in the feature map), and output the control signal for the robot carrying. This signal contains instructions such as the deployment position, movement path, and sampling order of the sampling robot, guiding the sampling robot to execute tasks.

[0030] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1:

[0031] This embodiment details the specific structure and implementation process of the terrain feature analysis module. The following describes it from three aspects: module composition, specific implementation of each unit, and data processing flow.

[0032] The terrain feature analysis module includes two core parts: the terrain data segmentation unit and the terrain feature encoding unit. Among them, the function of the terrain data segmentation unit is to convert continuous real-time terrain data into a set of discrete grid cells, realizing the structured division of the terrain space; the terrain feature encoding unit is responsible for extracting and abstractly representing the terrain data in each grid cell, generating a set of tensors containing local terrain features. The two units cooperate in sequence to complete the conversion process from the original terrain data to the feature tensors, providing input data that meets the requirements of algorithm processing for subsequent modules.

[0033] The main operation of the terrain data segmentation unit is to perform spatial division on the real-time terrain data. Specifically, this unit first receives the real-time terrain data output by the environmental perception module. These data usually exist in the form of three-dimensional point clouds or raster maps, containing multi-dimensional information such as the terrain height, surface texture, and obstacle positions in the target area. During the segmentation process, the unit divides the entire terrain data into uniformly distributed grid cells on the horizontal plane (i.e., the X-Y plane) according to a preset grid size (for example, it can be set to 0.5 m × 0.5 m, 1 m × 1 m, etc., and the specific size is determined according to the UAV task accuracy requirements and computing resources). Each grid cell corresponds to a cylinder region perpendicular to the horizontal plane in three-dimensional space, and its height range covers the undulation of the terrain surface, that is, from the lowest altitude to the highest altitude within the grid cell.

[0034] During the partitioning process, the unit conducts statistical processing on the terrain data within each grid cell. For example, it calculates parameters such as the average elevation, mean and variance of terrain slope, number of obstacles, and distribution density within the cell, forming the basic attribute data for each grid cell. These attribute data are stored in the form of multi-dimensional vectors, constituting the terrain data unit set. Through this grid-based segmentation, the continuous terrain space is discretized into computable basic units, facilitating subsequent feature encoding units to independently process the local features of each unit, and at the same time providing a structured data basis for spatial correlation analysis between units.

[0035] The terrain feature encoding unit adopts an architecture based on a three-dimensional convolutional neural network (3D-CNN). Its core objective is to extract representative terrain features from the multi-dimensional attribute data of each grid cell and encode these features into high-dimensional tensors. The processing flow of this unit includes four main links: input layer data preprocessing, multi-layer convolution operations, pooling operations, and output layer feature generation.

[0036] The input layer receives the grid cell attribute data output by the terrain data segmentation unit. Since the data collected by different types of sensors may have different dimensions and ranges, the input layer needs to first normalize the data. For example, it maps parameters such as elevation and slope to the [0, 1] interval to eliminate the impact of dimensional differences on neural network training. The normalized multi-dimensional attribute data constitutes the input tensor, whose dimensions are (grid cell size X, grid cell size Y, number of data channels), where the number of data channels corresponds to the types of terrain attributes, such as elevation, slope, roughness, obstacle density, etc.

[0037] Enter the convolution operation link. The three-dimensional convolutional neural network contains multiple convolutional layers, and each convolutional layer extracts features from the input data through three-dimensional convolutional kernels. Taking three convolutional layers as an example, the first convolutional layer usually uses a convolutional kernel with a smaller size (such as 3×3×3) to extract basic geometric features of the terrain, such as terrain edges, local protrusions or depressions. The convolutional kernel slides in three-dimensional space, performing weighted summation on the attribute data of each local area to generate the first layer of feature maps. The second convolutional layer uses a slightly larger convolutional kernel (such as 5×5×5). Based on the basic features extracted by the first layer, it further captures intermediate terrain features, such as the trends and distribution patterns of ridges and valleys. The third convolutional layer may use a larger convolutional kernel or dilated convolution technology to expand the receptive field and extract high-level features such as the structural relationships between terrain units, such as the terrain undulation correlation between adjacent grid cells.

[0038] Pooling layers are usually inserted between convolutional layers, and max-pooling or average-pooling operations are used to reduce the dimensionality of the feature maps. The pooling operation reduces the size of the feature map and the computational complexity by taking the maximum or average value of a local region in the spatial dimension, while retaining the key features. For example, under a 2×2×2 pooling window, each pooling operation halves the size of the feature map while retaining the most significant eigenvalue in that region.

[0039] After multiple layers of convolution and pooling operations, the data enters the output layer. The output layer converts the feature map into a feature tensor of local regions of the terrain environment through a fully connected layer or a convolutional layer. Each feature tensor corresponds to a grid cell, and its dimensions usually include the spatial coordinate dimensions (X, Y) and the feature channel dimension (C), where the number of feature channels is determined according to the network design and can be set to 64, 128, etc., for example. The value of each element in the feature tensor reflects the feature intensity of the grid cell on the corresponding feature channel. For example, a certain channel may correspond to the "steep terrain" feature, and the larger the value, the higher the probability that the cell belongs to the steep terrain.

[0040] The complete processing flow of the terrain feature parsing module is as follows: First, the terrain data segmentation unit divides the real-time terrain data into multiple grid cells, generating a set containing the attribute data of each cell; then the terrain feature encoding unit sequentially processes the attribute data of each grid cell through a three-dimensional convolutional neural network, extracts multi-layer features from the basic to the advanced level, and encodes these features into a feature tensor of a unified dimension, finally forming a set of feature tensors of local regions of the terrain environment.

[0041] These feature tensors not only retain the spatial position information of the terrain (determined by the coordinates of the grid cells), but also capture various aspects of the terrain attributes through high-dimensional feature channels. For example, the feature tensor of a certain grid cell may have a high value on the "obstacle density" channel, indicating that there are many obstacles in this area and need to be avoided during path planning; while the value on the "gentle terrain" channel is low, indicating that the terrain in this area has large undulations. In this way, the terrain feature parsing module converts the original terrain data into a structured feature representation that can be understood by the computer, providing the necessary data basis for the subsequent path planning feature enhancement module to analyze the spatial correlation between terrain units and the task scheduling decision module to perform global feature fusion. Example 2:

[0042] This embodiment details the composition and implementation process of the path planning feature enhancement module. This module is mainly used to deeply process the local area feature tensors output by the terrain feature parsing module. Through operations such as feature decomposition, correlation degree calculation, spatial constraint correction, and multi-scale convolution fusion, it generates an optimized feature tensor set containing spatial context dependence relationships, providing a more globally relevant feature expression for subsequent path planning. The following describes it from three aspects: module composition, specific implementation of each unit, and data processing logic.

[0043] The path planning feature enhancement module includes four functional units: a feature tensor decomposition unit, a regional feature correlation degree calculation unit, a regional spatial constraint correction unit, and a regional feature optimization unit. Among them, the feature tensor decomposition unit is responsible for converting the three-dimensional feature tensor into a two-dimensional feature vector for easy spatial correlation analysis; the regional feature correlation degree calculation unit is used to mine the semantic correlation between feature vectors; the regional spatial constraint correction unit corrects the correlation degree in combination with the geometric distance to suppress unreasonable long-distance correlations; the regional feature optimization unit fuses the correlated features and the original features through multi-scale convolution and attention mechanism to generate an optimized feature tensor. Each unit collaborates in sequence to form a complete processing link from feature dimensionality reduction to correlation modeling and then to feature enhancement.

[0044] The main function of the feature tensor decomposition unit is to perform a spatial dimension expansion process on the local area feature tensors of the terrain environment. This unit receives the three-dimensional feature tensor set output by the terrain feature parsing module. The dimension of each tensor is (X, Y, C), where X and Y represent the spatial coordinates of the grid cells, and C represents the number of feature channels. During the decomposition process, the unit unfolds each three-dimensional tensor along the spatial dimensions (i.e., the X and Y axis directions) into two-dimensional feature vectors. The specific operation is to arrange the data of all feature channels of each grid cell in sequence to form a one-dimensional vector with a dimension of (C). For example, for a grid cell with coordinates (i, j), its feature tensor After unfolding, it is converted into a feature vector , which contains the feature values of this grid cell on all feature channels. Through this operation, the feature tensors in the three-dimensional space are converted into a two-dimensional feature vector set, enabling subsequent feature correlation degree calculations to be performed in the vector space and simplifying the complexity of spatial correlation analysis.

[0045] The regional feature correlation calculation unit is divided into a feature space mapping subunit and a feature correlation analysis subunit, aiming to calculate the semantic correlation between feature vectors. First, the feature space mapping subunit uses manifold learning methods (such as Laplacian eigenmaps or isometric mapping) to map the set of feature vectors from the original feature space to the manifold space. The core idea of manifold learning is to mine the underlying low-dimensional manifold structure in high-dimensional feature data, so that in the manifold space, feature vectors with close geometric distances have higher semantic similarity. For example, for terrain features, the feature vectors of adjacent grid cells are usually closer in the manifold space, while grid cells that are far apart may be far away due to significant differences in terrain types.

[0046] After the mapping is completed, the feature correlation analysis subunit calculates the cosine similarity between any two feature vectors in the manifold space. Cosine similarity judges the similarity between vectors by measuring the angle between them, and the smaller the angle, the higher the similarity. When calculating specifically, for any two feature vectors and , the value of their cosine similarity reflects the similarity degree of the grid cells they represent in terms of terrain features. For example, the cosine similarity of feature vectors between flat areas is relatively high, while the similarity between flat areas and steep areas is relatively low. By traversing all pairs of feature vectors, a terrain environment regional correlation matrix is generated. The dimension of this matrix is (N, N), where N is the total number of grid cells. The element in the matrix represents the feature correlation degree between grid cells i and j, and the numerical range is [0, 1]. The larger the value, the stronger the correlation.

[0047] The role of the regional space constraint correction unit is to correct the correlation matrix by combining the geometric distances between grid cells, so as to ensure that the correlation degree reflects both the feature semantic similarity and the physical constraints of the spatial position. In the specific operation, the unit first calculates the Euclidean distance between any two grid cells in the geographical space. This distance is calculated based on the central coordinates of the grid cells and reflects the actual physical interval between the cells. Then, the unit performs attenuation correction on the elements in the correlation matrix according to the distance. The basic logic of the correction is that for grid cells that are closer, the influence of the distance on their feature correlation degree is smaller; for grid cells that are farther apart, their correlation degree needs to be attenuated according to the distance ratio. For example, for adjacent cells at a short distance, even if their features are quite different, their spatial proximity may still affect path planning, while for cells at a long distance with low feature similarity, their correlation degree should be significantly suppressed.

[0048] The corrected correlation matrix is called the terrain environment regional constraint correlation matrix, and the element It contains dual information of both feature semantic similarity and spatial distance constraint. Through this correction, the unreasonable association of distant units caused by only feature similarity is avoided, making the association degree matrix more conform to the spatial structure characteristics of the actual terrain environment and providing more reliable association information for subsequent feature optimization.

[0049] The regional feature optimization unit realizes feature fusion through multi-scale convolution and attention mechanism, which specifically includes three steps: dilated convolution processing, attention encoding, and tensor reconstruction. The unit performs dilated convolution processing on the terrain environment regional constraint association degree matrix. The dilated convolution expands the receptive field of the convolution kernel by introducing a dilation rate parameter, enabling the capture of multi-scale regional association features without increasing the computational amount. For example, by using dilated convolution kernels with different dilation rates (such as 1, 2, and 3), the association features at close range, medium range, and long range can be extracted respectively, thus generating a terrain environment regional constraint association feature matrix containing multi-scale association information.

[0050] The unit inputs the terrain environment local region feature vector set and the constraint association feature matrix into the attention encoding network. The attention encoding network calculates the dependence relationship between each feature vector and all other feature vectors through the self-attention mechanism, automatically learning the importance weights of different regional features in the global context. For example, for a certain target grid unit, the attention mechanism will assign higher weights to the units that are spatially adjacent and feature-similar to it, while assigning lower weights to the distant or feature-different units. In this way, a terrain environment local region context feature vector set containing global context information is generated, and each vector not only contains the original features of the unit but also fuses the association features of the surrounding regions.

[0051] The unit performs a tensor reconstruction operation on the context feature vector set. Tensor reconstruction is the inverse process of feature tensor decomposition, that is, rearranging the two-dimensional feature vectors according to the spatial coordinates of the original grid units into a three-dimensional tensor to restore its spatial dimension information and generate a terrain environment local region optimized feature tensor. The dimension of the optimized feature tensor is the same as that of the original feature tensor (X, Y, C), but the feature value at each position has fused multi-scale spatial association information, which can more accurately reflect the global structure characteristics of the terrain environment and provide richer context clues for the path planning algorithm.

[0052] The complete processing flow of the path planning feature enhancement module is as follows: First, the feature tensor decomposition unit unfolds the three-dimensional feature tensor into two-dimensional feature vectors; then the regional feature association degree calculation unit generates an association degree matrix through manifold mapping and cosine similarity calculation; next, the regional space constraint correction unit performs attenuation correction on the association degree matrix by combining geometric distance; finally, the regional feature optimization unit generates an optimized feature tensor set through dilated convolution, attention mechanism, and tensor reconstruction.

[0053] The technical significance of this module lies in that by explicitly modeling the spatial correlation of terrain features, it makes up for the lack of global context in traditional local feature extraction methods. For example, in complex terrains, the passability of a certain grid cell not only depends on its own features (such as slope), but also needs to consider the terrain coherence of adjacent cells (such as the existence of continuous steep slopes). Through the feature enhancement module, the UAV system can perceive the global structure of the terrain, avoid path planning mistakes caused by relying solely on local features (such as getting stuck in local optimal paths), and thus improve the path planning efficiency and safety of sampling robots in complex environments. In addition, the module realizes the adaptive fusion of associated features at different distance scales through the combination of multi-scale convolution and attention mechanism, enhancing the system's adaptability to diverse terrain environments. Embodiment 3:

[0054] This embodiment details the working mechanism of the task scheduling decision module. This module is mainly used to deeply process the set of optimized feature tensors of the local area of the terrain environment output by the path planning feature enhancement module. Through operations such as feature dimensionality reduction, information entropy calculation, cluster center determination, dynamic weight generation, and global feature fusion, it generates a decision feature map reflecting the scheduling priority of the global environment, providing a global decision basis for control instruction generation. The following describes it from three aspects: module composition, specific implementation of each unit, and data processing logic. A formula and an explanation of character meanings are involved in the process.

[0055] The task scheduling decision module includes five functional units: a local feature dimensionality reduction unit, a regional feature static information entropy calculation unit, a regional cluster center determination unit, a regional feature dynamic weight calculation unit, and a global feature fusion unit. Among them, the local feature dimensionality reduction unit is used to simplify the feature dimension to improve the calculation efficiency; the static information entropy calculation unit mines the uncertainty and information richness of regional features; the cluster center determination unit selects key scheduling starting points; the dynamic weight calculation unit generates adaptive weights by combining information entropy and spatial distance; the global feature fusion unit generates a global decision feature map through weighted summation. Each unit works together to achieve cross-level information integration from local features to global decisions.

[0056] The main function of the local feature dimensionality reduction unit is to perform dimensionality reduction processing on the set of optimized feature tensors of the local area of the terrain environment in the channel dimension. This unit receives the three-dimensional feature tensors output by the path planning feature enhancement module. The dimension of each tensor is (X, Y, C), where X and Y are the spatial coordinates of the grid cell, and C is the number of feature channels (such as 64 or 128 channels). The dimensionality reduction operation is implemented through max pooling processing. Specifically, for each grid cell's feature tensor, a maximum value operation is performed position by position in the channel dimension. That is, for the grid cell with coordinates (i, j), among the C channel values corresponding to each spatial position (i, j) in its feature tensor, the maximum value is taken as the feature value after dimensionality reduction.

[0057] After max pooling, the three-dimensional feature tensor (X, Y, C) is reduced to a set of two-dimensional feature vectors, with each vector having a dimension of (X, Y, 1), that is, each grid cell corresponds to a one-dimensional feature vector, which retains the most significant eigenvalue of that cell across all feature channels. For example, if the eigenvalue of a certain cell in the "obstacle density" channel is the maximum among all values in that channel, the value of the reduced feature vector reflects that the obstacle density of that cell is the most significant local feature. The dimensionality reduction not only reduces the data dimension and subsequent computational complexity but also highlights the dominant terrain attributes of each grid cell by retaining the extreme value features of each channel.

[0058] The regional feature static information entropy calculation unit is used to quantify the uncertainty and information richness of the feature vector of each grid cell. Its calculation process includes four steps: mean and variance calculation, generation of feature difference vectors, calculation of expected values, and normalization processing. First, the unit performs statistical analysis on the feature vectors of all grid cells to calculate the mean vector of the feature vector set and the variance vector . Each element of the mean vector is the average of the feature values of all cells in the corresponding feature dimension, and each element of the variance vector

[0059] is the variance of the feature values in the corresponding feature dimension, reflecting the degree of dispersion of the features in that dimension.

[0059] Next, the unit performs element-wise difference calculation on the feature vector of each grid cell and the mean vector to obtain the difference vector , which reflects the difference between the features of that cell and the global average features. To amplify the impact of feature differences, the unit performs a fourth-power operation on the difference result to obtain the terrain environment feature difference vector . The role of the fourth-power operation is to enhance the weight of abnormal features, making it easier to identify cells with significant feature differences as information-rich regions.

[0060] Then, the unit calculates the overall expected value of the feature difference vectors of all grid cells, which represents the average level of feature differences in the entire terrain environment. Finally, the unit performs a ratio operation on the expected value and the square value of the variance vector corresponding to each cell, and maps the result to the [0, 1] interval through a normalization function to obtain the information entropy value . The calculation formula for the information entropy value is:

[0061] where is the The information entropy value of a grid cell reflects the uncertainty of the characteristics of that cell: the larger the value, the more complex the characteristics, the richer the information, and the more urgent it is to prioritize sampling or path planning; is the variance value of the corresponding dimension of the feature vector of the th grid cell, which is used to measure the degree of dispersion of the cell's features in the global context; is a normalization function (such as Min - Max normalization) to ensure that the information entropy values are distributed within a unified interval for comparison.

[0062] The role of the regional clustering center determination unit is to select an initial scheduling center from all grid cells as the priority starting point for the sampling task. This unit analyzes the set of information entropy values output by the regional feature static information entropy calculation unit and selects the feature vector corresponding to the grid cell with the largest information entropy value as the initial scheduling center vector. The logic behind this selection is that the region with the largest information entropy value usually has the most complex terrain features or the highest information richness, such as areas containing transition zones of multiple terrain types, dense obstacle areas, or potential areas for target sampling points. Prioritizing the scheduling of this region can maximize the information acquisition efficiency of the sampling task and provide key anchor points for global path planning.

[0063] For example, in mountainous terrain, the region with the largest information entropy value may be located at the junction of the mountain foot and the plain, which contains steep slopes, gentle terrains, and possible obstacles, and is the area with the most drastic terrain feature changes. Prioritizing the scheduling of this region helps the UAV system quickly establish a global terrain awareness and optimize the layout of subsequent sampling points.

[0064] The regional feature dynamic weight calculation unit is used to generate dynamic scheduling weights for each grid cell. This weight comprehensively considers the information entropy value of the cell and the spatial distance to the initial scheduling center to achieve adaptive weighting of local features. In the specific calculation process, the unit multiplies the information entropy value of each grid cell by the information entropy value of the initial scheduling center vector and multiplies it by the first adjustment coefficient to obtain the first dynamic scheduling factor. The first dynamic scheduling factor is positively correlated with the information entropy value, that is, the higher the information entropy value of a cell, the higher the basic value of its scheduling priority weight.

[0065] The unit calculates the Euclidean distance between each grid cell and the initial scheduling center in spatial coordinates, squares the distance, and then multiplies it by the second adjustment coefficient to obtain the second dynamic scheduling factor. The second dynamic scheduling factor is positively correlated with the spatial distance, that is, the farther a cell is from the initial scheduling center, the stronger the weight suppression effect.

[0066] The unit performs a ratio operation on the first dynamic scheduling factor and the second dynamic scheduling factor to obtain the dynamic scheduling weight , and the calculation formula is:

[0067] Among them, is the dynamic scheduling weight of the th grid cell, and are adjustable parameters determined through a preset parameter tuning process (for example, setting different adjustment coefficients according to terrain complexity and task type). The physical meaning of this weight is that cells with high information entropy values and close to the initial scheduling center will obtain higher weights and participate in global feature fusion preferentially; cells with low information entropy values or far distances will have lower weights and have less influence in global decision-making. This dynamic weighting mechanism enables the module to automatically adjust the importance of each cell according to the actual distribution of terrain features, balancing the dual needs of information richness and spatial proximity.

[0068] The core function of the global feature fusion unit is to perform weighted summation on the optimized feature vector set of the local area of the terrain environment using the dynamic scheduling weight set to generate a global environment scheduling decision feature map. In specific operations, the unit traverses the feature vectors of all grid cells and the corresponding dynamic scheduling weights , and performs weighted summation according to the following formula:

[0069] Among them, is the global environment scheduling decision feature map, with a dimension of (X, Y, 1), and each element value represents the comprehensive priority and risk level of the corresponding grid cell in global scheduling; is the total number of grid cells. The result of weighted summation makes cells with high information entropy values and close to the initial scheduling center contribute more to the global feature map, thus highlighting the areas that need to be processed preferentially in the feature map.

[0070] For example, in the generated global feature map, high-weight areas may appear as highlighted areas, indicating that the drone should preferentially deploy sampling robots or plan the path to pass through this area; low-weight areas may appear as low-brightness areas, indicating that the terrain in this area is simple or the information value is low, and it can be postponed or used as a secondary path option. In this way, the global feature map provides an intuitive global scheduling basis for the drone system, enabling task planning to be upgraded from local feature analysis to global optimal decision-making.

[0071] The complete processing flow of the task scheduling decision module is as follows: First, the local feature dimensionality reduction unit performs channel dimensionality reduction on the optimized feature tensor to obtain a set of two-dimensional feature vectors; then the regional feature static information entropy calculation unit generates the information entropy values of each unit through statistical analysis; next, the regional clustering center determination unit selects the unit with the maximum information entropy as the initial scheduling center; the regional feature dynamic weight calculation unit generates dynamic weights by combining information entropy and spatial distance; finally, the global feature fusion unit generates a global decision feature map through weighted summation.

[0072] The technical value of this module lies in the combination of information entropy theory and dynamic weighting mechanism, realizing the intelligent screening and global integration of terrain environment features. Traditional global feature fusion methods often use average weighting or fixed weights, which are difficult to adapt to the dynamic changes in the importance of features in complex terrains. However, this module quantifies the uncertainty of features through information entropy, uses dynamic weights to balance spatial distance and information value, enabling the global feature map to adaptively reflect the actual needs of the terrain environment. For example, in areas with fragmented terrain, the module will automatically increase the weight of high-information entropy regions, guiding the drone to explore complex terrains preferentially; in areas with gentle terrain, the weight will be reduced to reduce computational resource consumption. This intelligent global feature fusion mechanism significantly improves the task scheduling efficiency and decision-making scientificity of the drone system under different terrain conditions, laying a solid foundation for the efficient deployment and path planning of sampling robots. Example 4:

[0073] This example details the specific implementation process of the control instruction generation module. The main function of this module is to convert the global environment scheduling decision feature map output by the task scheduling decision module into executable robot-mounted control signals to guide the deployment, movement, and sampling operations of the sampling robot. The following is an explanation in combination with specific application scenarios and processing logics.

[0074] The control instruction generation module receives the global environment scheduling decision feature map output by the task scheduling decision module. This feature map is in the form of a two-dimensional matrix, and each element value corresponds to the global scheduling priority of the grid unit (for example, the larger the value, the higher the priority or the higher the risk level). The core component of the module is the decision tree-based control signal generator, which parses the feature map through predefined decision rules to generate structured control instructions. The construction of the decision tree is based on the prior knowledge of terrain features and sampling tasks. For example, the values in the feature map are divided into different risk level intervals, and corresponding control strategies are associated with each interval.

[0075] The root node of the decision tree is usually the judgment of the terrain risk level. According to the numerical threshold in the feature map, the terrain is divided into low-risk, medium-risk, and high-risk areas. For example, it is set that when the feature value is less than 0.3, it is a low-risk area; when it is between 0.3 and 0.7, it is a medium-risk area; and when it is greater than 0.7, it is a high-risk area. Each child node corresponds to more specific terrain features or task requirements, forming a hierarchical decision-making branch.

[0076] Taking the low-risk area as an example, the child node can be directly associated with the instruction of "prioritize deploying sampling robots" because the terrain in this area is flat and there are few obstacles, which is suitable for the robot to quickly start operations. For example, when the value of a certain area in the feature map is 0.2, which belongs to the low-risk range, when the decision tree traverses to this branch, an instruction of "deploy the sampling robot to coordinates (X1, Y1, Z1)" is generated, where the coordinates are determined by the center position of the grid unit corresponding to the feature map, and Z1 is the current flight altitude of the UAV minus the safe deployment distance.

[0077] For the medium-risk area, the decision tree further subdivides the nodes. For example, the terrain complexity is judged according to the information entropy value (the information entropy value can be obtained from the intermediate output of the task scheduling decision module). If the information entropy value of a certain medium-risk area is greater than 0.6, it indicates that the terrain features are complex (such as the existence of local protrusions or sparse obstacles), then an instruction of "plan a detour path to avoid obstacles" is generated. In the specific implementation, the control signal generator determines the position of the obstacles according to the feature vector of this area, instructs the robot to move in an "S-shaped" or "zigzag" path, and turns on the obstacle avoidance sensor to adjust the direction in real time. If the information entropy value is less than 0.6, it indicates that although the terrain has a certain risk but the complexity is low (such as a single gentle slope), then an instruction of "move straight to the sampling point" is generated, and at the same time, the robot is instructed to move at a medium speed and keep the sensor scanning in real time.

[0078] The high-risk area (such as the feature value is 0.8) triggers the instruction of "suspend deployment and re-plan". For example, when the feature map shows that a certain area is a steep cliff or a dense obstacle group, the decision tree determines that the risk of directly deploying a robot in this area is too high, generates an instruction to let the UAV hover over this area, and at the same time adjusts the sensor parameters (such as the lidar scanning frequency) for secondary detection. After obtaining more detailed terrain data, it is re-input to the environmental perception module to start a new round of feature analysis and path planning process.

[0079] The control signal carried by the robot output by the control instruction generation module is structured data, which usually includes fields such as instruction type, target coordinates, motion parameters, and sensor configuration. For example: Instruction type: including "deploy", "move", "sample", "hover", "return", etc., which clarify the operation type of the robot.

[0080] Target coordinates: Represented in three-dimensional coordinates (X, Y, Z), e.g., (100m, 200m, 5m), where the Z-axis is the height from the ground and is dynamically adjusted according to the terrain undulation.

[0081] Motion parameters: Include traveling speed (such as low speed 0.5m / s, medium speed 1m / s, high speed 2m / s), steering angle (such as turning left 30°), path mode (straight line, circuitous, etc.).

[0082] Sensor configuration: Specify the types and parameters of sensors to be enabled, such as "vision camera + lidar, scanning frequency 10Hz", or "turn off lidar and only use inertial sensors" to save power.

[0083] Taking a specific sampling task as an example, assume that the global feature map shows that the area with coordinates (50m, 50m) is a low-risk high-priority area (feature value 0.25, information entropy value 0.7). After parsing, the control signal generator generates the following instructions: Instruction type: Delivery Target coordinates: (50m, 50m, 2m) (The Z-axis is set to 2 meters to avoid hitting the ground) Motion parameters: Descend at low speed and stop after reaching the ground Sensor configuration: Turn on the vision camera and soil composition sensor When the sampling robot completes the delivery, the feature map updates to show that there is a medium-risk area (feature value 0.5, information entropy value 0.4) near this area (60m, 50m). The control signal generator generates a "move" instruction: Instruction type: Move Target coordinates: (60m, 50m, 0.5m) (Move slowly 0.5 meters above the ground) Motion parameters: Move in a straight line, speed 0.8m / s, obstacle avoidance mode enabled Sensor configuration: Lidar scans in real time to detect obstacles within 2 meters ahead The control instruction generation module supports the sequential generation of multiple segments of instructions to handle complex task processes. For example, when the sampling robot needs to visit multiple sampling points in sequence, the module generates an instruction sequence including "delivery → move to point A → sampling → move to point B → sampling → return" according to the priority sorting of the global feature map. During the execution process, if there are significant differences between the terrain data real-time feedback by the sensor and the feature map (such as sudden obstacles), the module can trigger the dynamic adjustment mechanism, pause the current instruction and re-call the task scheduling decision module to generate a new feature map, thereby updating the control instruction.

[0084] For example, during the process of the robot moving towards point A, the lidar detects a deep pit that suddenly appears in front (not reflected in the initial feature map). After receiving this real-time data, the control instruction generation module immediately generates a "hover" instruction. At the same time, it inputs the new data into the environmental perception module and starts the fast terrain resampling process. After the updated feature map is generated, it re-plans the path to bypass the deep pit and generates a new "move" instruction.

[0085] The control instruction generation module converts the abstract feature map into concrete control actions through a decision tree model. Its core advantages lie in the interpretability and flexibility of the decision logic. The node parameters of the decision tree (such as the risk level threshold, adjustment coefficient) can be quickly adjusted according to different task requirements (such as environmental monitoring, disaster rescue, agricultural sampling, etc.). For example, in the agricultural sampling scenario, the "crop density" feature can be used as an important node of the decision tree, and the robot can be preferentially instructed to sample in areas with high crop density; in the disaster rescue scenario, the "obstacle density" and "terrain flatness" are used as key judgment conditions to ensure the safe passage of the robot.

[0086] The application of this module enables the UAV system to dynamically adjust the control strategy according to the real-time terrain features and task objectives, realizing a complete closed-loop from "environmental perception → feature analysis → decision generation → task execution". Whether it is quickly deploying sampling in a plain area with simple terrain or performing fine path planning in a complex mountain environment, the control instruction generation module can ensure that the sampling robot efficiently and safely completes the task through a structured decision-making process, reflecting the technological advancement of the UAV system in the field of intelligent control. Embodiment 5:

[0087] This embodiment combines a specific application scenario to detail the complete implementation process of a UAV system and a control method that can carry a sampling robot. Assume that the target task is to collect soil samples in a mountainous and hilly area. The terrain of this area includes gentle slopes, steep slopes, gullies, and scattered vegetation coverage. The UAV needs to dynamically plan the sampling path and deploy the robot according to the real-time terrain data. The following will explain the entire process from data collection, feature processing, decision generation to instruction execution.

[0088] The unmanned aerial vehicle (UAV) carrying multi-modal sensors (such as lidar, visual cameras, and inertial navigation systems) enters the airspace above the target area and conducts an initial cruise at a height of 5 meters and a speed of 10 m / s. The lidar scans the terrain in real time to generate point cloud data containing three-dimensional coordinates (X, Y, Z) and reflection intensity; the visual camera synchronously captures RGB images and extracts texture features such as vegetation coverage and surface color; the inertial navigation system records the position, speed, and attitude data of the UAV. After the data of the three types of sensors are fused, real-time terrain data is formed. For example, the point cloud data of a certain area shows that the altitude is between 100 and 120 meters, the visual image identifies a mixed area of brown soil and green vegetation, and the inertial navigation data marks that this area is 50 meters northeast of the current position of the UAV.

[0089] The terrain feature analysis module first conducts grid segmentation on the real-time terrain data and sets the grid size to 1 meter × 1 meter according to the task accuracy requirements. Taking the coordinate (0, 0) as the origin, the target area is divided into 200 × 200 grid cells, and each cell corresponds to a rectangular area of 1 square meter on the ground. The terrain data segmentation unit calculates the statistical parameters of each cell. For example, the average altitude of cell (50, 50) is 110 meters, the average slope is 15°, the vegetation coverage rate is 30%, and the obstacle density (such as the number of rocks) is 2 per cell.

[0090] Subsequently, the terrain feature encoding unit inputs the multi-dimensional data (altitude, slope, vegetation coverage rate, obstacle density, etc.) of each cell into a three-dimensional convolutional neural network. After three layers of convolutional processing, the first layer extracts edge features and identifies the boundary between soil and vegetation in cell (50, 50); the second layer extracts intermediate features and determines that this cell belongs to a gentle slope terrain; the third layer combines the features of adjacent cells and determines that this area is the edge zone where the gentle slope transitions to the steep slope, and outputs a tensor containing 64-channel features. For example, the value of a certain channel represents the feature intensity of "transition terrain" as 0.8.

[0091] The path planning feature enhancement module processes the feature tensor of the local area of the terrain environment. First, the feature tensor decomposition unit unfolds the three-dimensional tensor into a two-dimensional feature vector. For example, the feature vector of cell (50, 50) contains 64 feature values corresponding to the feature intensities of each channel. The regional feature correlation calculation unit maps all feature vectors to the manifold space and calculates the cosine similarity between any two cells. For example, the similarity between cell (50, 50) and (51, 50) is 0.9 (both are gentle slope transition terrains), and the similarity with (60, 50) is 0.3 (the latter is a steep slope).

[0092] The regional spatial constraint correction unit corrects the correlation degree according to the geometric distance. The Euclidean distance between the units (50, 50) and (51, 50) is 1 meter, and the correlation degree remains 0.9; the distance from (60, 50) is 10 meters, and the correlation degree decays to 0.3×e^(-10² / σ²) (σ is the preset attenuation parameter). The regional feature optimization unit uses dilated convolution and attention mechanism to fuse multi-scale correlation features and generate an optimized feature tensor, so that the features of the unit (50, 50) contain the terrain coherence information within 10 meters adjacent, for example, it is recognized that there is a steep slope 10 meters to the right of this unit, and going straight needs to be avoided in path planning.

[0093] The task scheduling decision module processes the optimized feature tensor. The local feature dimensionality reduction unit retains the most significant features of each unit through max pooling. For example, the maximum eigenvalue of the unit (50, 50) corresponds to the "transition terrain" channel (value 0.8). The regional feature static information entropy calculation unit analyzes and finds that the expected value of the fourth power of the feature difference vector of this unit is relatively high and the variance is small. After normalization, the information entropy value is 0.7, indicating that the terrain features are complex and rich in information, and it needs to be scheduled preferentially.

[0094] The regional clustering center determination unit selects the unit with the maximum information entropy (60, 60, information entropy 0.9) as the initial scheduling center. This area is the edge of the gully and the terrain changes violently. The regional feature dynamic weight calculation unit calculates the weight of the unit (50, 50) according to the formula (α, β are preset coefficients): its information entropy is 0.7, and the distance from the center is 10 meters, and the weight is (α×0.7) / (β×100). The global feature fusion unit generates a decision feature map by weighted summation. The unit (60, 60) and its adjacent areas show high brightness, indicating the priority sampling area.

[0095] The control instruction generation module generates instructions based on the decision feature map. For the initial scheduling center (60, 60), the feature map shows a high-risk and high-priority area (feature value 0.8), and the decision tree determines it as a medium-risk terrain (information entropy 0.9, feature value 0.8 is between 0.3 - 0.7), and generates the following instructions: Instruction type: Drop Target coordinates: (60, 60, 2 meters) (safe height 2 meters from the ground) Motion parameters: Descend at low speed. After the robot lands, turn on the lidar to scan the surrounding 5-meter range Sensor configuration: Lidar + soil sensor, scanning frequency 5Hz After the sampling robot landed, the real-time feedback data showed that there was a 1-meter-diameter pit in the (60, 60) cell. The module immediately triggered dynamic adjustment: generated an instruction to "move to (59, 60)", and the robot bypassed the pit circuitously at a speed of 0.5 m / s. After reaching the new position, it performed the soil sampling operation. After the sampling was completed, the updated feature map showed that the information entropy of the (55, 55) cell rose to 0.8 (newly identified rock distribution). The module generated an instruction to "move to (55, 55)", and the robot continued to operate according to the planned path.

[0096] During the entire operation process, the UAV system achieved closed-loop control through multi-module collaboration: sensors continuously collected data → feature parsing and correlation analysis updated the terrain expression in real time → task scheduling dynamically adjusted the priority → control instructions precisely drove the robot to execute. For example, when the UAV detected a sudden increase in the vegetation coverage rate in a certain area (such as detecting a forest through visual images), the environmental perception module supplemented data in real time, the terrain feature parsing module re-extracted features, the path planning module adjusted the correlation matrix, the task scheduling module updated the sampling priority, and the control instruction module re-planned a path to bypass the forest.

[0097] This full-process dynamic adaptability enables the UAV system to cope with uncertainties in complex terrains, such as sudden obstacles and terrain feature changes, ensuring that the sampling robot can efficiently complete tasks in unstructured environments such as mountainous and hilly areas. Each link from data collection to instruction execution is closely connected, reflecting the integrated design of the system in aspects of environmental perception, feature modeling, decision-making generation, and control execution, providing a reliable technical solution for automated sampling in complex scenarios.

[0098] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0099] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An unmanned aerial vehicle system capable of carrying a sampling robot, characterized in that, Comprising: An environmental perception module, configured to obtain real-time terrain data of a target area through a multi-modal sensor; A terrain feature parsing module, configured to perform grid segmentation and terrain feature extraction on the real-time terrain data to obtain a set of local terrain environment region feature tensors; A path planning feature enhancement module, configured to optimize spatial correlation features of the set of local terrain environment region feature tensors to obtain a set of optimized local terrain environment region feature tensors; A task scheduling decision module, configured to perform global feature fusion based on information entropy distribution on the set of optimized local terrain environment region feature tensors to obtain a global environment scheduling decision feature map; A control instruction generation module, configured to generate a robot-mounted control signal according to the global environment scheduling decision feature map.

2. The unmanned aerial vehicle system capable of carrying a sampling robot according to claim 1, characterized in that, The terrain feature parsing module includes: a terrain data segmentation unit, configured to spatially divide the real-time terrain data according to a preset grid size to obtain a set of terrain data units; a terrain feature encoding unit, configured to input each unit data in the set of terrain data units into a terrain feature encoder based on a three-dimensional convolutional neural network to obtain the set of local terrain environment region feature tensors.

3. The drone system capable of carrying a sampling robot according to claim 2, wherein, The path planning feature enhancement module includes: a feature tensor decomposition unit, configured to perform tensor unfolding on the local terrain environment region feature tensors along the spatial dimension to obtain a set of local terrain environment region feature vectors; a region feature correlation degree calculation unit, configured to calculate the spatial correlation degree between any two feature vectors in the set of local terrain environment region feature vectors to generate a terrain environment region correlation degree matrix; a region space constraint correction unit, configured to perform distance attenuation correction on the terrain environment region correlation degree matrix according to the geometric distance between any two feature vectors in the set of local terrain environment region feature vectors to obtain a terrain environment region constraint correlation degree matrix; a region feature optimization unit, configured to perform multi-scale convolutional fusion on the terrain environment region constraint correlation degree matrix and the set of local terrain environment region feature vectors to obtain the optimized local terrain environment region feature tensors.

4. The unmanned aerial vehicle system capable of carrying a sampling robot according to claim 3, characterized in that, The region feature correlation degree calculation unit includes: a feature space mapping sub-unit, configured to map each feature vector in the set of local terrain environment region feature vectors to a manifold space to obtain a set of mapped local terrain environment region feature vectors; a feature correlation analysis sub-unit, configured to calculate the cosine similarity between any two mapped feature vectors in the set of mapped local terrain environment region feature vectors to generate the terrain environment region correlation degree matrix composed of multiple terrain environment region correlation degree values.

5. The drone system capable of carrying a sampling robot according to claim 4, characterized in that, The area feature optimization unit is specifically implemented as follows: performing dilated convolution processing on the terrain environment area constraint correlation matrix to obtain a terrain environment area constraint correlation feature matrix; inputting the set of terrain environment local area feature vectors and the terrain environment area constraint correlation feature matrix into an attention encoding network to obtain a set of terrain environment local area context feature vectors; performing tensor reconstruction on the set of terrain environment local area context feature vectors to obtain the terrain environment local area optimized feature tensor.

6. The drone system capable of carrying a sampling robot according to claim 5, characterized in that, The task scheduling decision module includes: a local feature dimensionality reduction unit, configured to perform max pooling processing on each tensor in the set of terrain environment local area optimized feature tensors in the channel dimension to obtain a set of terrain environment local area optimized feature vectors; an area feature static information entropy calculation unit, configured to calculate the information entropy value of each feature vector in the set of terrain environment local area optimized feature vectors to generate a terrain environment static information entropy set; an area clustering center determination unit, configured to select the terrain environment local area optimized feature vector corresponding to the maximum information entropy value in the terrain environment static information entropy set as the initial scheduling center vector; an area feature dynamic weight calculation unit, configured to calculate the dynamic scheduling weight of each feature vector according to the spatial distance between each feature vector in the set of terrain environment local area optimized feature vectors and the initial scheduling center vector and the information entropy value of each feature vector to generate a dynamic scheduling weight set; a global feature fusion unit, configured to perform weighted summation on the set of terrain environment local area optimized feature vectors by using the dynamic scheduling weight set to generate the global environment scheduling decision feature map.

7. The drone system capable of carrying a sampling robot according to claim 6, wherein The area feature static information entropy calculation unit is specifically implemented as follows: calculating the mean vector and variance vector of the terrain environment local area optimized feature vectors; performing element-wise difference calculation between the terrain environment local area optimized feature vectors and the mean vector, and performing a fourth-power operation on the difference result to obtain a terrain environment feature difference vector; calculating the overall expected value of the terrain environment feature difference vector; performing a ratio operation on the expected value and the square value of the variance vector, and inputting the result into a normalization function to obtain the information entropy value.

8. The unmanned aerial vehicle system capable of carrying a sampling robot according to claim 7, characterized in that, The area feature dynamic weight calculation unit is specifically implemented as follows: multiplying the information entropy value of the terrain environment local area optimized feature vector and the information entropy value of the initial scheduling center vector by a first adjustment coefficient to obtain a first dynamic scheduling factor; multiplying the squared Euclidean distance value between the terrain environment local area optimized feature vector and the initial scheduling center vector by a second adjustment coefficient to obtain a second dynamic scheduling factor; performing a ratio operation on the first dynamic scheduling factor and the second dynamic scheduling factor to obtain the dynamic scheduling weight.

9. The unmanned aerial vehicle system capable of carrying a sampling robot according to claim 8, characterized in that, The control instruction generation module is specifically implemented as follows: inputting the global environment scheduling decision feature map into a control signal generator based on a decision tree to obtain the robot-mounted control signal, and the signal is used to indicate the deployment strategy of the sampling robot.

10. A method for controlling an unmanned aerial vehicle capable of carrying a sampling robot, characterized in that, Including: Collect real-time terrain data of the target area through multi-modal sensors; perform grid segmentation and terrain feature extraction on the real-time terrain data to obtain a set of local area feature tensors of the terrain environment; perform spatial correlation feature optimization on the set of local area feature tensors of the terrain environment to obtain a set of optimized local area feature tensors of the terrain environment; perform global feature fusion based on information entropy distribution on the set of optimized local area feature tensors of the terrain environment to obtain a global environment scheduling decision feature map; generate a robot-mounted control signal according to the global environment scheduling decision feature map.

Images