A UAV system capable of carrying a sampling robot and a control method thereof

Abstract

The present invention relates to the field of unmanned aerial vehicle technology, and discloses a unmanned aerial vehicle system capable of carrying a sampling robot and a control method. The system comprises 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. The environmental perception module acquires real-time terrain data through a multimodal sensor; the terrain feature analysis module grid-segments the data and extracts feature tensors using a three-dimensional convolutional neural network; the path planning feature enhancement module optimizes the spatial correlation of the feature tensors; the task scheduling decision module generates a global scheduling decision feature map based on information entropy distribution fusion features; and 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, and instruction generation. The present invention improves the efficiency and accuracy of unmanned aerial vehicles carrying sampling robots in complex terrains, and is suitable for environmental monitoring, geological exploration, and other fields.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles (UAVs), and in particular to a UAV system capable of carrying a sampling robot and a control method thereof. Background Art

[0002] Rapid and accurate sampling in complex terrain areas is crucial for environmental monitoring, geological exploration, and disaster response. Traditional manual sampling methods suffer from low efficiency, high risk, and significant terrain limitations, making them difficult to meet the sampling needs of large-scale, complex environments. The development of drone technology offers new avenues for sampling operations. Its aerial flight capabilities can overcome terrain obstacles and enable rapid access to target areas. However, existing drone systems capable of carrying sampling robots still face numerous technical bottlenecks 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 obtain terrain data comprehensively and accurately. The processing of terrain data mostly remains at the level of simple rasterization, lacking in-depth extraction of terrain features and unable to effectively characterize the local area characteristics of the terrain environment. As a result, subsequent path planning and task scheduling lack reliable environmental information support.

[0004] When it comes to path planning, existing systems often ignore the spatial correlations between local features of the terrain environment. Terrain features include not only local information such as height and slope, but also the spatial relationships between adjacent areas, such as the continuity of the terrain and the correlation of undulations. Traditional path planning methods analyze only local terrain features independently, failing to fully consider the inter-regional influences. This can easily lead to a lack of rationality in the planned path at the global level, potentially resulting in long detours and encountering insurmountable terrain obstacles, impacting the deployment efficiency and safety of the sampling robot.

[0005] The task scheduling decision-making process is crucial for achieving efficient sampling in drone systems. Existing technologies often integrate terrain and environmental features using simple weighted averaging methods, failing to fully consider the information entropy distribution of different regional characteristics—that is, the importance and uncertainty of each region in the sampling task. This results in a lack of scientific rationality in task scheduling decisions, making it impossible to dynamically adjust the sampling robot deployment strategy based on the actual terrain and environment. This can easily lead to irrational allocation of sampling resources, resulting in ineffective coverage of key areas and wasted resources in non-critical areas.

[0006] Traditional control signal generation methods are typically based on fixed rules or simple algorithms, lacking the ability to comprehensively analyze and dynamically respond to global environmental characteristics. In complex and ever-changing terrain, it's difficult to generate precise control commands based on real-time environmental information. This results in inflexible deployment strategies for sampling robots, making them unable to adapt to real-time changes in the terrain and environment, impacting the overall effectiveness of the sampling mission. Summary of the Invention

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

[0008] To achieve the above objectives, the present invention provides the following technical solution: an unmanned aerial vehicle system capable of carrying a sampling robot, the system comprising:

[0009] Environmental perception module, used to obtain real-time terrain data of the target area through multimodal sensors;

[0010] A terrain feature analysis module, configured to perform grid segmentation and terrain feature extraction on the real-time terrain data to obtain a set of feature tensors of a local area of ​​the terrain environment;

[0011] a path planning feature enhancement module, configured to 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;

[0012] A task scheduling decision module is used to perform global feature fusion based on information entropy distribution on the set of optimized feature tensors of the local area of ​​the terrain environment to obtain a global environment scheduling decision feature map;

[0013] A control instruction generation module is used to generate a robot-mounted control signal according to the global environment scheduling decision feature map.

[0014] Preferably, the terrain feature analysis module includes: a terrain data segmentation unit, 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, used to input the data of each unit in the set of terrain data units into a terrain feature encoder based on a three-dimensional convolutional neural network to obtain a set of local area feature tensors of the terrain environment.

[0015] Preferably, the path planning feature enhancement module includes: a feature tensor decomposition unit, used to perform 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 calculation unit, used to calculate the spatial correlation between any two feature vectors in the set of local area feature vectors of the terrain environment to generate a terrain environment regional correlation matrix; a regional spatial constraint correction unit, used to perform distance attenuation correction on the terrain environment regional correlation 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 matrix; and a regional feature optimization unit, used to perform multi-scale convolution fusion on the terrain environment regional constraint correlation 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.

[0016] Preferably, the regional feature correlation calculation unit includes: a feature space mapping subunit, used to map 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; a feature correlation analysis subunit, used to calculate 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 area correlation matrix composed of multiple terrain environment area correlation values.

[0017] Preferably, the regional feature optimization unit is specifically implemented as follows: performing dilated convolution processing on the terrain environment regional constraint correlation matrix to obtain a terrain environment regional constraint correlation feature matrix; inputting the set of terrain environment local area feature vectors and the terrain environment regional constraint correlation feature matrix into the 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 optimization feature tensor.

[0018] Preferably, the task scheduling decision module includes: a local feature dimensionality reduction unit, which is used to perform channel-dimensional maximum pooling processing on each tensor in the set of the local area optimization feature tensors of the terrain environment to obtain a set of local area optimization feature vectors of the terrain environment; a regional feature static information entropy calculation unit, which is used to calculate the information entropy value of each feature vector in the set of the local area optimization feature vectors of the terrain environment to generate a terrain environment static information entropy set; a regional clustering center determination unit, which is used to select the local area optimization feature vector of the terrain environment corresponding to the maximum information entropy value in the terrain environment static information entropy set as the initial scheduling center vector; a regional feature dynamic weight calculation unit, which is used to calculate the dynamic scheduling weight of each feature vector in the set of the local area optimization feature vector of the terrain environment and the initial scheduling center vector based on the spatial distance between each feature vector 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, which is used to use the dynamic scheduling weight set to perform weighted summation on the set of the local area optimization feature vectors of the terrain environment to generate the global environment scheduling decision feature map.

[0019] Preferably, the regional feature static information entropy calculation unit is specifically implemented as follows: calculating the mean vector and variance vector of the local area optimized feature vector of the terrain environment; performing element-by-element difference calculation between the local area optimized feature vector 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.

[0020] Preferably, the regional feature dynamic weight calculation unit is specifically implemented as follows: multiplying the information entropy value of the local area optimization feature vector 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 local area optimization feature vector 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.

[0021] Preferably, the control instruction generation module is specifically implemented as follows: inputting the global environment scheduling decision feature map into a decision tree-based control signal generator to obtain the robot-mounted control signal, which is used to indicate the deployment strategy of the sampling robot.

[0022] Preferably, the present invention also includes a method for controlling a drone that can carry a sampling robot, the method comprising: an environmental perception module, which collects real-time terrain data of a target area through a multimodal sensor; a terrain feature analysis module, which performs 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 performs spatial correlation feature optimization on the set of local area feature tensors of the terrain environment to obtain a set of local area optimized feature tensors of the terrain environment; a task scheduling decision module, which performs global feature fusion based on information entropy distribution on the set of local area optimized feature tensors of the terrain environment to obtain a global environment scheduling decision feature map; and a control instruction generation module, which generates a robot-carrying control signal according to the global environment scheduling decision feature map.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] In terms of environmental perception and terrain feature analysis, multimodal sensors acquire real-time terrain data of the target area, leveraging the strengths of multiple sensors to comprehensively and accurately perceive the multidimensional information of the terrain environment, avoiding the limitations of a single sensor. The terrain feature analysis module grids the real-time terrain data and uses a three-dimensional convolutional neural network to perform feature encoding on each unit of data, enabling in-depth processing of terrain data and efficient extraction of terrain features. This deep learning-based feature extraction method automatically captures complex features in the terrain environment, such as undulations, slope changes, and obstacle distribution, generating a collection of feature tensors for local areas of the terrain environment, providing rich and accurate basic data for subsequent path planning and task scheduling.

[0025] The path planning feature enhancement module effectively improves the rationality and reliability of path planning by optimizing the spatial correlation features of the feature tensor of the local area of ​​the terrain environment. The module first expands the feature tensor along the spatial dimension to obtain a set of feature vectors. Then, by calculating the spatial correlation between the feature vectors and combining the geometric distance to perform distance attenuation correction, it generates a terrain environment regional constraint correlation matrix. On this basis, through operations such as multi-scale convolution fusion and attention encoding network, the spatial correlation of the local area features of the terrain environment is optimized, so that the optimized feature tensor can more accurately reflect the spatial relationship and mutual influence between various regions in the terrain environment. This helps the drone not only consider local terrain features when planning the path, but also comprehensively evaluate the feasibility of the path from a global perspective, avoiding the problem of unreasonable path planning caused by ignoring the correlation between regions, and improving the scientificity and safety of the sampling robot deployment path.

[0026] The task scheduling decision module performs global feature fusion based on information entropy distribution, achieving scientific scheduling of terrain environment features and rational allocation of resources. By calculating the information entropy value of the optimized feature vector of the local area of ​​the terrain environment, it can effectively identify key areas in the terrain environment with rich information and high uncertainty, and use them as the initial scheduling center vector. Then, the dynamic scheduling weight is calculated based on the spatial distance and information entropy value of each feature vector and the initial scheduling center vector, and the feature vectors are weighted summed to generate a 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 areas in the sampling task, realize the dynamic optimization configuration of sampling resources, ensure that key areas are covered and sampled first, improve the pertinence and effectiveness of the sampling task, and avoid the waste of sampling resources.

[0027] The control command generation module feeds the global environmental scheduling decision feature map into a decision-tree-based control signal generator. This module generates precise robot-mounting control signals based on global environmental characteristics, instructing the sampling robot's deployment strategy. The decision tree algorithm boasts excellent interpretability and decision-making capabilities, enabling it to quickly make informed decisions and generate corresponding control commands based on different combinations of environmental characteristics. This global feature-based control command generation method enables the UAV system to dynamically adjust the sampling robot's deployment strategy based on real-time updates of terrain and environmental information, improving the system's adaptability and responsiveness to complex environments and ensuring efficient and accurate completion of sampling tasks.

[0028] Furthermore, through the organic integration and collaborative operation of various modules, this invention forms a complete technical chain from environmental perception, feature analysis, path planning, task scheduling, to control command generation, achieving full-process optimization of the UAV system that can carry sampling robots. The system can automatically and efficiently complete the loading and deployment of sampling robots in complex terrain environments, significantly reducing labor costs and improving the safety and reliability of sampling operations. It has broad application prospects and important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a working principle diagram of the UAV system capable of carrying a sampling robot according to the present invention;

[0030] Figure 2 This is a flow chart of the UAV control method;

[0031] Figure 3 Design diagram of the path planning feature enhancement module. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] See also Figure 1-Figure 3 The present invention relates to an unmanned aerial vehicle system capable of carrying a sampling robot. The system includes an environment perception module, a terrain feature analysis module, a path planning feature enhancement module, a task scheduling decision module, and a control instruction generation module. The specific implementation is as follows:

[0034] Environmental perception module: Uses multimodal sensors (such as lidar, visual cameras, inertial navigation sensors, etc.) to collect terrain data of the target area in real time, covering multi-dimensional information such as terrain height, slope, obstacle distribution, etc., to form a real-time terrain data matrix containing spatial coordinates and physical properties, providing raw data support for subsequent processing.

[0035] Terrain Feature Parsing Module: This module performs grid segmentation and feature extraction on real-time terrain data. It first divides the real-time terrain data into multiple terrain data units according to a preset grid size (e.g., 0.5 m x 0.5 m), forming a terrain data unit set. The data from each unit is then fed into a terrain feature encoder based on a three-dimensional convolutional neural network (3D-CNN). This encoder extracts deep features in the spatial dimension using multiple layers of convolution kernels (e.g., 3×3×3 convolution kernels), capturing local geometric features of the terrain (e.g., ridges, depressions, and flat areas). The encoder then outputs a set of feature tensors for the local area of ​​the terrain environment, representing both spatial location and feature dimensions. Each tensor corresponds to the feature representation of a grid unit.

[0036] Path Planning Feature Enhancement Module: This module optimizes the spatial correlation of feature tensors in local areas of the terrain environment. By expanding the spatial dimensions of the feature tensors, calculating correlations, correcting distance constraints, and performing multi-scale convolutional fusion, it enhances the spatial dependencies between features and generates optimized feature tensors for local areas of the terrain environment, providing a more contextually relevant feature representation for path planning.

[0037] The task scheduling decision module performs global feature fusion on the optimized feature tensor set based on information entropy distribution. The information-rich regions are determined by calculating the information entropy values ​​of each local feature. The feature vector corresponding to the maximum information entropy value is used as the initial scheduling center. Dynamic weights are calculated by combining spatial distance and information entropy values. A weighted summation of local features is performed to generate a global environment scheduling decision feature map, enabling global planning of sampling tasks.

[0038] Control instruction generation module: The global environment scheduling decision feature map is input into the control signal generator based on the decision tree, and logical judgment is made through preset decision rules (such as terrain risk level, sampling priority and other node parameters in the feature map). The robot-mounted control signal is output. The signal contains instructions such as the sampling robot's placement position, movement path, sampling sequence, etc., to guide the sampling robot to perform tasks.

[0039] The present invention will be further described below in conjunction with Examples 1 to 5: Example 1:

[0040] This embodiment describes in detail 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.

[0041] The terrain feature analysis module consists of two core components: the terrain data segmentation unit and the terrain feature encoding unit. The terrain data segmentation unit converts continuous, real-time terrain data into a discrete set of grid cells, achieving a structured partitioning of the terrain space. The terrain feature encoding unit extracts and abstractly represents the terrain data within each grid cell, generating a set of tensors containing local terrain features. These two units work together to complete the conversion from raw terrain data to feature tensors, providing input data that meets the algorithm processing requirements of subsequent modules.

[0042] The main operation of the terrain data segmentation unit is to perform spatial division on the real-time terrain data. Specifically, the unit first receives the real-time terrain data output by the environmental perception module. These data are usually in the form of a three-dimensional point cloud or a raster map, and contain multi-dimensional information such as the terrain height, surface texture, and obstacle location of the target area. During the segmentation process, the unit divides the entire terrain data into evenly distributed grid cells on the horizontal plane (i.e., the XY plane) according to the preset grid size (for example, it can be set to 0.5 meters × 0.5 meters, 1 meter × 1 meter, etc., and the specific size is determined according to the UAV mission accuracy requirements and computing resources). Each grid cell corresponds to a cylindrical area perpendicular to the horizontal plane in three-dimensional space, and its height range covers the undulations of the terrain surface, that is, from the lowest altitude to the highest altitude within the grid cell.

[0043] During the division process, the unit performs statistical processing on the terrain data within each grid cell, calculating parameters such as the average altitude, the mean and variance of the terrain slope, and the number and density of obstacles within the cell, thereby forming basic attribute data for each grid cell. This attribute data is stored as a multidimensional vector, forming a set of terrain data units. This grid-based segmentation discretizes the continuous terrain space into computable basic units, facilitating the subsequent feature encoding unit to independently process the local features of each unit. It also provides a structured data foundation for spatial correlation analysis between units.

[0044] The terrain feature encoding unit utilizes a three-dimensional convolutional neural network (3D-CNN) architecture. Its core objective is to extract representative terrain features from the multidimensional attribute data of each grid cell and encode these features into a high-dimensional tensor. The unit's processing pipeline includes four main steps: input layer data preprocessing, multi-layer convolution operations, pooling operations, and output layer feature generation.

[0045] The input layer receives grid cell attribute data output by the terrain data segmentation unit. Because data collected by different sensor types may have different dimensions and ranges, the input layer must first normalize the data. For example, parameters such as altitude and slope are mapped to the range [0, 1] to eliminate the impact of dimensional differences on neural network training. The normalized multidimensional 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 type of terrain attribute, such as altitude, slope, roughness, and obstacle density.

[0046] Enter the convolution operation phase. A three-dimensional convolutional neural network contains multiple convolutional layers, each of which extracts features from the input data using a three-dimensional convolution kernel. Taking three convolutional layers as an example, the first convolution layer typically uses a smaller convolution kernel (such as 3×3×3) to extract basic geometric features of the terrain, such as terrain edges, local protrusions or depressions, etc. The convolution kernel slides in three-dimensional space, performing a weighted summation of the attribute data of each local area to generate the first-layer feature map. The second convolution layer uses a slightly larger convolution kernel (such as 5×5×5) to further capture intermediate terrain features such as the direction and distribution patterns of ridges and valleys based on the basic features extracted by the first layer. The third convolution layer may use a larger convolution kernel or dilated convolution technology to expand the receptive field and extract high-level features such as the structural relationship between terrain units, such as the correlation between terrain undulations of adjacent grid units.

[0047] Pooling layers are often inserted between convolutional layers, using either maximum or average pooling to reduce the dimensionality of feature maps. Pooling reduces the size of feature maps by taking the maximum or average value of a local region across the spatial dimensions, reducing computational complexity while preserving key features. For example, with a 2×2×2 pooling window, each pooling operation halves the size of the feature map while retaining the most significant feature values ​​within that region.

[0048] After multiple layers of convolution and pooling, the data enters the output layer. The output layer converts the feature map into a feature tensor representing the local area of ​​the terrain environment through a fully connected or convolutional layer. Each feature tensor corresponds to a grid cell, and its dimensions typically include spatial coordinate dimensions (X, Y) and feature channel dimensions (C). The number of feature channels is determined by the network design and can be set to 64, 128, etc. The value of each element in the feature tensor reflects the characteristic strength of the grid cell in the corresponding feature channel. For example, a channel may correspond to the "steep terrain" feature; a larger value indicates a higher probability that the cell belongs to steep terrain.

[0049] The complete processing flow of the terrain feature analysis module is as follows: first, the terrain data segmentation unit divides the real-time terrain data into multiple grid cells and generates a set of attribute data for each cell; then the terrain feature encoding unit performs three-dimensional convolutional neural network processing on the attribute data of each grid cell in turn, extracting multi-layer features from basic to advanced levels, and encoding these features into feature tensors of uniform dimension, ultimately forming a set of feature tensors for the local area of ​​the terrain environment.

[0050] These feature tensors not only retain the spatial location information of the terrain (determined by the coordinates of the grid cells), but also capture the various attributes of the terrain through high-dimensional feature channels. For example, the feature tensor of a grid cell may have a high value on the "obstacle density" channel, indicating that there are many obstacles in the area and they need to be avoided during path planning; while the value on the "flat terrain" channel is low, indicating that the terrain in the area is quite undulating. In this way, the terrain feature parsing module converts the raw terrain data into a structured feature representation that can be understood by a computer, providing the necessary data foundation for the subsequent path planning feature enhancement module to analyze the spatial correlation between terrain cells and for the task scheduling decision module to perform global feature fusion. Example 2:

[0051] This embodiment details the composition and implementation of the path planning feature enhancement module. This module is primarily used to perform in-depth processing on the local area feature tensors output by the terrain feature analysis module. Through operations such as feature decomposition, correlation calculation, spatial constraint correction, and multi-scale convolution fusion, it generates an optimized feature tensor set containing spatial context dependencies, providing a more globally relevant feature expression for subsequent path planning. The following describes the module composition, the specific implementation of each unit, and the data processing logic.

[0052] The path planning feature enhancement module includes four functional units: a feature tensor decomposition unit, a regional feature correlation calculation unit, a regional spatial constraint correction unit, and a regional feature optimization unit. The feature tensor decomposition unit converts three-dimensional feature tensors into two-dimensional feature vectors to facilitate spatial correlation analysis; the regional feature correlation calculation unit is used to explore the semantic correlations between feature vectors; the regional spatial constraint correction unit corrects the correlations based on geometric distance to suppress unreasonable long-distance correlations; and the regional feature optimization unit fuses correlated features with original features through multi-scale convolution and attention mechanisms to generate an optimized feature tensor. These units work together to form a complete processing chain from feature dimensionality reduction to correlation modeling to feature enhancement.

[0053] The main function of the feature tensor decomposition unit is to perform spatial dimension expansion processing on the feature tensor of the local area of ​​the terrain environment. The unit receives the set of three-dimensional feature tensors output by the terrain feature analysis module. The dimension of each tensor is (X, Y, C), where X and Y represent the spatial coordinates of the grid unit and C represents the number of feature channels. During the decomposition process, the unit expands each three-dimensional tensor into a two-dimensional feature vector along the spatial dimension (i.e., the X and Y axis directions). The specific operation is to arrange all the feature channel data of each grid unit in order to form a one-dimensional vector with a dimension of (C,). For example, for the grid unit with coordinates (i, j), its feature tensor After expansion, it is converted into a feature vector , which contains the eigenvalues ​​of the grid cell on all feature channels. Through this operation, the feature tensor in three-dimensional space is converted into a set of two-dimensional feature vectors, so that subsequent feature correlation calculations can be performed in vector space, simplifying the complexity of spatial correlation analysis.

[0054] The regional feature correlation calculation unit is divided into two parts: a feature space mapping subunit and a feature correlation analysis subunit. Both parts are designed to calculate the semantic correlation between feature vectors. First, the feature space mapping subunit uses manifold learning methods (such as Laplace eigenmapping or isometric mapping) to map the feature vector set from the original feature space to the manifold space. The core idea of ​​manifold learning is to exploit the low-dimensional manifold structure implicit in high-dimensional feature data, so that feature vectors with similar geometric distances in the manifold space 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 farther apart may be farther apart due to large differences in terrain types.

[0055] After the mapping is completed, the feature association analysis subunit calculates the cosine similarity between any two feature vectors in the manifold space. Cosine similarity measures the angle between vectors to determine their similarity. The smaller the angle, the higher the similarity. and , the cosine similarity value reflects the similarity of the grid cells they represent in terms of terrain characteristics. For example, the cosine similarity of the eigenvectors of the flat area and the flat area is high, while the similarity between the flat area and the steep area is low. By traversing all the eigenvector pairs, the terrain environment area correlation matrix is ​​generated. The dimension of the matrix is ​​(N, N), where N is the total number of grid cells and the elements in the matrix are Represents the feature correlation between grid cells i and j. The value range is [0,1]. The larger the value, the stronger the correlation.

[0056] The role of the regional spatial constraint correction unit is to correct the correlation matrix based on the geometric distance between grid cells to ensure that the correlation reflects both the semantic similarity of features and the physical constraints of spatial locations. In specific operations, the unit first calculates the Euclidean distance between any two grid cells in geographic space. This distance is calculated based on the center coordinates of the grid cells and reflects the actual physical spacing between the cells. Then, the unit performs an attenuation correction on the elements in the correlation matrix based on the distance. The basic logic of the correction is: the closer the grid cells are, the less their feature correlation is affected by the distance; the farther the grid cells are, the correlation needs to be attenuated in proportion to the distance. For example, even if the features of adjacent cells at close distances are quite different, their spatial proximity may still have an impact on path planning, while if the feature similarity of distant cells is low, their correlation should be significantly suppressed.

[0057] The modified correlation matrix is ​​called the terrain environment area constraint correlation matrix, and the elements in this matrix are This correction incorporates both feature semantic similarity and spatial distance constraints. This avoids the irrational association of distant units based solely on feature similarity, making the association matrix more consistent with the spatial structure of the actual terrain environment and providing more reliable association information for subsequent feature optimization.

[0058] The regional feature optimization unit achieves feature fusion through multi-scale convolution and attention mechanisms, specifically comprising three steps: dilated convolution, attention encoding, and tensor reconstruction. The unit performs dilated convolution on the terrain environment regional constraint correlation matrix. Dilated convolution introduces a dilated convolution parameter to expand the receptive field of the convolution kernel, enabling the capture of multi-scale regional correlation features without increasing computational complexity. For example, using dilated convolution kernels with different dilation rates (e.g., 1, 2, and 3) can extract correlation features at close, medium, and long distances, respectively, thereby generating a terrain environment regional constraint correlation feature matrix containing multi-scale correlation information.

[0059] The unit inputs the set of local area feature vectors of the terrain environment and the constraint-related feature matrix into the attention encoding network. The attention encoding network calculates the dependency 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 target grid cell, the attention mechanism will give higher weights to cells that are spatially adjacent to it and have similar features, while giving lower weights to cells that are distant or have large feature differences. In this way, a set of local area context feature vectors of the terrain environment containing global context information is generated. Each vector not only contains the original features of the cell, but also integrates the associated features of the surrounding area.

[0060] The unit performs a tensor reconstruction operation on the set of contextual feature vectors. Tensor reconstruction is the inverse process of feature tensor decomposition. It rearranges the two-dimensional feature vectors into a three-dimensional tensor according to the spatial coordinates of the original grid cells, restoring their spatial dimensionality and generating an optimized feature tensor for the local area of ​​the terrain environment. The dimensions of the optimized feature tensor are the same as the original feature tensor (X, Y, C), but the eigenvalues ​​at each location have incorporated multi-scale spatial correlation information, which can more accurately reflect the global structural characteristics of the terrain environment and provide richer contextual clues for path planning algorithms.

[0061] The complete processing flow of the path planning feature enhancement module is as follows: first, the feature tensor decomposition unit expands the three-dimensional feature tensor into a two-dimensional feature vector; then the regional feature correlation calculation unit generates a correlation matrix through manifold mapping and cosine similarity calculation; then the regional spatial constraint correction unit combines the geometric distance to attenuate the correlation matrix; finally, the regional feature optimization unit generates an optimized feature tensor set through dilated convolution, attention mechanism and tensor reconstruction.

[0062] The technical significance of this module lies in that it makes up for the lack of global context in traditional local feature extraction methods by explicitly modeling the spatial correlation of terrain features. For example, in complex terrain, the passability of a grid cell depends not only on its own characteristics (such as slope), but also on the terrain continuity of adjacent cells (such as the presence of continuous steep slopes). Through the feature enhancement module, the UAV system can perceive the global structure of the terrain, avoiding path planning errors caused by relying solely on local features (such as falling into the local optimal path), thereby improving the path planning efficiency and safety of the sampling robot in complex environments. In addition, the module realizes the adaptive fusion of correlated features at different distance scales through the combination of multi-scale convolution and attention mechanism, enhancing the system's adaptability to diverse terrain environments. Example 3:

[0063] This embodiment explains in detail the working mechanism of the task scheduling decision module. This module is mainly used to perform in-depth processing on 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, a decision feature map reflecting the scheduling priority of the global environment is generated, providing a global decision basis for the generation of control instructions. The following is a description from three aspects: module composition, specific implementation of each unit and data processing logic. The process involves a formula and character meaning explanation.

[0064] The task scheduling decision module consists of five functional units: local feature dimensionality reduction, regional feature static information entropy calculation, regional cluster center determination, regional feature dynamic weight calculation, and global feature fusion. The local feature dimensionality reduction unit simplifies feature dimensions to improve computational efficiency; the static information entropy calculation unit exploits 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; and the global feature fusion unit generates a global decision feature map through weighted summation. These units work together to achieve cross-level information integration from local features to global decisions.

[0065] The main function of the local feature dimensionality reduction unit is to reduce the channel dimension of the optimized feature tensor set for the local area of ​​the terrain environment. This unit receives the three-dimensional feature tensors output by the path planning feature enhancement module. Each tensor has the dimension (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 achieved through maximum pooling. Specifically, the maximum value of the channel dimension is taken for each grid cell's feature tensor position by position. That is, for the grid cell with coordinates (i, j), the maximum value of the C channel values ​​corresponding to each spatial position (i, j) in its feature tensor is taken as the feature value after dimensionality reduction.

[0066] After max pooling, the three-dimensional feature tensor (X, Y, C) is reduced to a set of two-dimensional feature vectors, each of dimension (X, Y, 1). This means that each grid cell corresponds to a one-dimensional feature vector that retains the cell's most significant eigenvalue across all feature channels. For example, if the eigenvalue of a cell in the "obstacle density" channel is the maximum among all values ​​in that channel, the reduced eigenvector value reflects that the obstacle density of that cell is the most significant local feature. This dimensionality reduction not only reduces the data dimensionality and subsequent computational complexity, but also highlights the dominant terrain attributes of each grid cell by preserving the extreme features of each channel.

[0067] The regional characteristic static information entropy calculation unit is used to quantify the uncertainty and information richness of the feature vector of each grid unit. The calculation process includes four steps: mean and variance calculation, feature difference vector generation, expected value calculation and normalization. First, the unit performs statistical analysis on the feature vectors of all grid units and calculates the mean vector of the feature vector set. and variance vector . Mean vector Each element is the average value of all unit eigenvalues ​​on the corresponding feature dimension, and the variance vector Each element of is the variance of the eigenvalue on the corresponding feature dimension, reflecting the degree of discreteness of the feature in that dimension.

[0068] Next, the feature vector of each grid cell With the mean vector Perform element-by-element difference calculation to obtain the difference vector , which reflects the difference between the unit feature and the global average feature. In order to amplify the influence of feature difference, 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 units with significant feature differences as information-rich regions.

[0069] The unit then calculates the overall expected value of the characteristic difference vectors of all grid cells , the expected value represents the average level of feature differences in the entire terrain environment. Finally, the expected value is compared with the square value of the variance vector corresponding to each unit, and the result is mapped to the [0,1] interval through the normalization function to obtain the information entropy value The calculation formula of information entropy is:

[0070]

[0071] in, For the The information entropy value of a grid cell reflects the uncertainty of the cell's features: the larger the value, the more complex the features and the richer the information, and the more priority is needed for sampling or path planning; For the The variance value of the dimension corresponding to the grid unit feature vector is used to measure the degree of global discreteness of the unit feature; It is a normalization function (such as Min-Max normalization) to ensure that the information entropy values ​​are distributed in a uniform interval for comparison.

[0072] The role of the regional cluster center determination unit is to select the initial scheduling center from all grid cells as the priority starting point for the sampling task. This unit analyzes the information entropy value set output by the regional characteristic 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 area with the largest information entropy value usually has the most complex terrain features or the highest information richness, such as transition zones containing multiple terrain types, areas with dense obstacles, or potential areas for target sampling points. Prioritizing scheduling in this area can maximize the information acquisition efficiency of the sampling task and provide a key anchor point for global path planning.

[0073] For example, in mountainous terrain, the area with the largest information entropy value may be located at the junction of the foot of the mountain and the plain. This area also contains steep slopes, flat terrain and possible obstacles. It is the area where the terrain characteristics change most dramatically. Prioritizing this area helps the UAV system quickly establish global terrain cognition and optimize the layout of subsequent sampling points.

[0074] The regional feature dynamic weight calculation unit is used to generate a dynamic scheduling weight for each grid unit. This weight comprehensively considers the unit's information entropy value and the spatial distance to the initial scheduling center to achieve adaptive weighting of local features. In the specific calculation process, the unit will calculate the information entropy value of each grid unit. The information entropy value of the initial scheduling center vector Multiply by the first adjustment coefficient , and 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 the unit, the higher the basic value of its scheduling priority weight.

[0075] The unit calculates the Euclidean distance between each grid cell and the initial dispatch center in spatial coordinates , and square the distance and multiply it by the second adjustment coefficient , and obtain the second dynamic scheduling factor. The second dynamic scheduling factor is positively correlated with the spatial distance, that is, the farther the unit is from the initial scheduling center, the stronger its weight suppression effect.

[0076] The unit performs a ratio operation on the first dynamic scheduling factor and the second dynamic scheduling factor to obtain the dynamic scheduling weight , the calculation formula is:

[0077]

[0078] in, For the The dynamic scheduling weight of each grid cell, and This is an adjustable parameter determined through a pre-set parameter tuning process (for example, different adjustment coefficients are set based on terrain complexity and mission type). The physical meaning of this weight is that units with high information entropy and close proximity to the initial dispatch center receive higher weights and are prioritized for global feature fusion; units with low information entropy or distant distances receive lower weights and have less influence in global decision-making. This dynamic weighting mechanism enables the module to automatically adjust the importance of each unit based on the actual distribution of terrain features, balancing the dual requirements of information richness and spatial proximity.

[0079] The core function of the global feature fusion unit is to use the dynamic scheduling weight set to perform weighted summation on the optimized feature vector set of the local area of ​​the terrain environment to generate a global environment scheduling decision feature map. In the specific operation, the unit traverses the feature vectors of all grid cells. And the corresponding dynamic scheduling weight , perform weighted summation according to the following formula:

[0080]

[0081] in, It is the global environment scheduling decision feature map with dimensions of (X, Y, 1). Each element value represents the comprehensive priority and risk level of the corresponding grid unit in the global scheduling. The weighted summation results in cells with high information entropy and close proximity to the initial scheduling center making greater contributions to the global feature map, thereby highlighting areas that require priority processing in the feature map.

[0082] For example, in the generated global feature map, high-weight areas may appear as highlighted areas, indicating that the drone should prioritize deploying sampling robots or planning paths through them; low-weight areas may appear as low-brightness areas, indicating that the terrain is simple or the information value is low, so processing can be postponed or selected as a secondary path. In this way, the global feature map provides an intuitive global scheduling basis for the drone system, enabling mission planning to be upgraded from local feature analysis to global optimal decision-making.

[0083] 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 value of each unit through statistical analysis; then the regional cluster center determination unit selects the unit with the largest information entropy as the initial scheduling center; the regional feature dynamic weight calculation unit combines information entropy and spatial distance to generate dynamic weights; finally, the global feature fusion unit generates a global decision feature map through weighted summation.

[0084] The technical value of this module lies in its intelligent screening and global integration of terrain and environmental features through the combination of information entropy theory and a dynamic weighting mechanism. Traditional global feature fusion methods often use average weighting or fixed weights, which are difficult to adapt to the dynamic changes in feature importance in complex terrain. This module, however, quantifies feature uncertainty through information entropy and utilizes 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 automatically increases the weight of high-entropy areas, guiding the drone to prioritize exploring complex terrain; in areas with flat terrain, the weight is reduced to reduce computing resource consumption. This intelligent global feature fusion mechanism significantly improves the efficiency of task scheduling and the scientific nature of decision-making for drone systems under different terrain conditions, laying a solid foundation for the efficient deployment and path planning of sampling robots. Example 4:

[0085] This example describes the specific implementation of the control instruction generation module in detail. This module's primary function is to convert the global environment scheduling decision feature map output by the task scheduling decision module into executable robot-mounted control signals, guiding the deployment, movement, and sampling operations of the sampling robot. The following describes this in detail, combining specific application scenarios and processing logic.

[0086] The control instruction generation module receives the global environment scheduling decision feature map output by the task scheduling decision module. This feature map is a two-dimensional matrix, with each element value corresponding to the global scheduling priority of a grid cell (e.g., larger values ​​indicate higher priority or higher risk level). The module's core component is a decision tree-based control signal generator, which parses the feature map using predefined decision rules to generate structured control instructions. The decision tree is constructed based on prior knowledge of terrain characteristics and sampling tasks. For example, the values ​​in the feature map are divided into different risk level intervals and a corresponding control strategy is associated with each interval.

[0087] The root node of a decision tree typically determines the terrain risk level. Based on numerical thresholds in the feature map, the terrain is divided into low-risk, medium-risk, and high-risk areas. For example, a feature value less than 0.3 is considered low-risk, 0.3-0.7 is considered medium-risk, and greater than 0.7 is considered high-risk. Each child node corresponds to a more specific terrain feature or mission requirement, forming a hierarchical decision branch.

[0088] Taking low-risk areas as an example, a child node can be directly associated with the instruction "Prioritize deployment of sampling robots" because the terrain in this area is flat and there are few obstacles, making it suitable for robots to quickly deploy operations. For example, when the value of a certain area in the feature map is 0.2, which is in the low-risk range, the decision tree traverses to this branch and generates the instruction "Deploy the sampling robot to coordinates (X1, Y1, Z1)", where the coordinates are determined by the center position of the grid cell corresponding to the feature map, and Z1 is the current flight altitude of the drone minus the safe deployment distance.

[0089] For medium-risk areas, the decision tree further subdivides nodes, for example, judging terrain complexity based on information entropy (which can be obtained through the intermediate output of the task scheduling decision module). If the information entropy value of a medium-risk area is greater than 0.6, indicating complex terrain features (such as localized bumps or sparse obstacles), a "plan a circuitous path to avoid obstacles" instruction is generated. Specifically, the control signal generator determines the location of obstacles based on the area's feature vectors, instructing the robot to move in an "S-shaped" or "zigzag" path, and activating obstacle avoidance sensors to adjust its direction in real time. If the information entropy value is less than 0.6, indicating that the terrain, while risky, is relatively low in complexity (such as a single gentle slope), a "move in a straight line to the sampling point" instruction is generated, instructing the robot to move at a moderate speed and maintain real-time sensor scanning.

[0090] High-risk areas (e.g., an eigenvalue of 0.8) trigger a "hold deployment, replan" command. For example, if the feature map shows a steep cliff or dense obstacle cluster in a certain area, the decision tree determines that it is too risky to directly deploy the robot there. It then generates a command for the drone to hover over the area while adjusting sensor parameters (such as the lidar scanning frequency) for a secondary survey. Once more detailed terrain data is obtained, it is re-input into the environmental perception module, initiating a new round of feature analysis and path planning.

[0091] The robot control signal output by the control command generation module is structured data, which usually contains fields such as command type, target coordinates, motion parameters, and sensor configuration. For example:

[0092] Command type: including "drop", "move", "sample", "hover", "return", etc., to clarify the operation type of the robot.

[0093] Target coordinates: Expressed in three-dimensional coordinates (X, Y, Z), for example (100m, 200m, 5m), where the Z axis is the height above the ground and is dynamically adjusted based on the terrain.

[0094] Motion parameters: including travel speed (such as low speed 0.5m / s, medium speed 1m / s, high speed 2m / s), steering angle (such as 30° left turn), path mode (straight line, circuitous, etc.).

[0095] Sensor configuration: Specifies the type of sensors to enable and their parameters, such as "visual camera + lidar, scanning frequency 10Hz", or "turn off lidar and use only inertial navigation sensor" to save power.

[0096] Taking a specific sampling task as an example, assuming that the global feature map shows that the area with coordinates (50m, 50m) is a low-risk, high-priority area (eigenvalue 0.25, information entropy value 0.7), the control signal generator generates the following instructions after parsing:

[0097] Command type: Delivery

[0098] Target coordinates: (50m, 50m, 2m) (Z axis is set to 2 meters to avoid collision with the ground)

[0099] Movement parameters: descend at low speed and stop after reaching the ground

[0100] Sensor configuration: Turn on the visual camera and soil composition sensor

[0101] After the sampling robot is deployed, the feature map is updated to show that there is a medium-risk area (feature value 0.5, information entropy value 0.4) near the area (60m, 50m). The control signal generator generates a "move" command:

[0102] Instruction Type: Move

[0103] Target coordinates: (60m, 50m, 0.5m) (moving at a low speed of 0.5 meters above the ground)

[0104] Movement parameters: linear movement, speed 0.8m / s, obstacle avoidance mode on

[0105] Sensor configuration: LiDAR real-time scanning, detecting obstacles within 2 meters ahead

[0106] The control instruction generation module supports the generation of multi-segment instruction sequences to handle complex task flows. For example, when a sampling robot needs to visit multiple sampling points in sequence, the module prioritizes the global feature map and generates an instruction sequence consisting of "drop → move to point A → sample → move to point B → sample → return." During execution, if the real-time terrain data fed back by the sensor differs significantly from the feature map (e.g., due to an unexpected obstacle), the module triggers a dynamic adjustment mechanism, pausing the current instruction and re-invoking the task scheduling decision module to generate a new feature map, thereby updating the control instructions.

[0107] For example, when the robot is moving towards point A, the lidar detects a deep pit suddenly appearing in front of it (not reflected in the initial feature map). After receiving the real-time data, the control instruction generation module immediately generates a "hover" instruction, and at the same time inputs the new data into the environmental perception module, starting the rapid terrain resampling process. After the updated feature map is generated, the path to bypass the deep pit is replanned and a new "move" instruction is generated.

[0108] The control command generation module uses a decision tree model to transform abstract feature maps into concrete control actions. Its core advantage lies in the interpretability and flexibility of its decision logic. Decision tree node parameters (such as risk level thresholds and adjustment coefficients) can be quickly adjusted to meet the needs of different tasks (such as environmental monitoring, disaster relief, and agricultural sampling). For example, in an agricultural sampling scenario, the "crop density" feature can be used as a key node in the decision tree, prioritizing the robot to sample areas with high crop density. In disaster relief scenarios, "obstacle density" and "terrain flatness" are used as key judgment criteria to ensure the robot's safe passage.

[0109] This module enables the UAV system to dynamically adjust its control strategy based on real-time terrain characteristics and mission objectives, achieving a complete closed loop from "environmental perception → feature analysis → decision generation → mission execution." Whether rapidly deploying samples in simple plains or meticulously planning paths in complex mountainous environments, the control command generation module ensures the sampling robot completes its mission efficiently and safely through a structured decision-making process, demonstrating the technological advancement of UAV systems in the field of intelligent control. Example 5:

[0110] This example describes the complete implementation process of a drone system and control method capable of carrying a sampling robot, based on specific application scenarios. Assume the target task is to collect soil samples in a hilly area with gentle slopes, steep slopes, gullies, and scattered vegetation. The drone must dynamically plan the sampling path and deploy the robot based on real-time terrain data. The following describes the entire process, from data collection, feature processing, decision generation, to command execution.

[0111] Equipped with multimodal sensors (such as lidar, visual cameras, and an inertial navigation system), the drone enters the target area and conducts an initial cruise at an altitude of 5 meters and a speed of 10 meters per second. The lidar scans the terrain in real time, generating point cloud data containing three-dimensional coordinates (X, Y, Z) and reflection intensity. The visual camera simultaneously captures RGB images, extracting texture features such as vegetation cover and surface color. The inertial navigation system records the drone's position, speed, and attitude data. The data from these three sensors is fused to form real-time terrain data. For example, the point cloud data for a certain area indicates an altitude between 100 and 120 meters. The visual image identifies a mixed area of ​​brown soil and green vegetation, and the inertial navigation data indicates that this area is 50 meters northeast of the drone's current position.

[0112] The terrain feature analysis module first grids the real-time terrain data, setting the grid size to 1 meter by 1 meter based on the mission's accuracy requirements. With the coordinates (0,0) as the origin, the target area is divided into 200 by 200 grid cells, each corresponding to a 1-square-meter rectangular area on the ground. The terrain data segmentation unit calculates statistical parameters for each cell, such as the average elevation of cell (50,50) of 110 meters, the mean slope of 15°, the vegetation coverage of 30%, and the obstacle density (e.g., the number of rocks) of 2 per cell.

[0113] The terrain feature encoding unit then feeds each cell's multidimensional data (altitude, slope, vegetation coverage, obstacle density, etc.) into a three-dimensional convolutional neural network. After three layers of convolution, the first layer extracts edge features, identifying the boundary between soil and vegetation at the cell (50,50). The second layer extracts intermediate features, determining that the cell belongs to gently sloping terrain. The third layer, combining features from adjacent cells, identifies the area as a transition zone from gentle to steep slopes. The output is a tensor containing 64 channel features. For example, a channel value indicates a "transition terrain" feature strength of 0.8.

[0114] The path planning feature enhancement module processes the local area feature tensor of the terrain environment. First, the feature tensor decomposition unit expands the three-dimensional tensor into a two-dimensional feature vector. For example, the feature vector of the unit (50,50) contains 64 eigenvalues, corresponding to the feature strength 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 units. For example, the similarity between the unit (50,50) and (51,50) is 0.9 (both are gently sloping transition terrain), and the similarity with (60,50) is 0.3 (the latter is a steep slope).

[0115] The regional spatial constraint correction unit modifies the correlation based on geometric distance. The Euclidean distance between units (50,50) and (51,50) is 1 meter, maintaining a correlation of 0.9. The correlation with (60,50) is attenuated to 0.3×e^(-10² / σ²) (σ is a preset attenuation parameter) at a distance of 10 meters. The regional feature optimization unit uses dilated convolution and an attention mechanism to fuse multi-scale correlation features and generate an optimized feature tensor. This ensures that the features of unit (50,50) include terrain continuity information within the adjacent 10 meters. For example, it identifies a steep slope 10 meters to the right of the unit, necessitating avoiding straight travel during path planning.

[0116] The task scheduling decision module processes the optimized feature tensor. The local feature dimensionality reduction unit uses maximum pooling to retain the most significant features of each unit. For example, the maximum eigenvalue of the unit (50,50) corresponds to the "transition terrain" channel (value 0.8). Analysis of the regional feature static information entropy calculation unit revealed that the fourth power expected value of the feature difference vector of this unit is high and the variance is small. The normalized information entropy value is 0.7, indicating that the terrain features are complex and rich in information, requiring priority scheduling.

[0117] The regional cluster center determination unit selects the unit with the largest information entropy (60, 60, information entropy 0.9) as the initial dispatch center. This area is at the edge of a gully and has a drastic terrain change. The regional feature dynamic weight calculation unit uses the formula (α and β are preset coefficients). The weight of cell (50,50) is calculated: its information entropy is 0.7, and its distance from the center is 10 meters, so the weight is (α × 0.7) / (β × 100). The weighted summation of the global feature fusion units generates a decision feature map. Cell (60,60) and its adjacent area appear high in brightness, indicating that it is a priority sampling area.

[0118] The control instruction generation module generates instructions based on the decision feature map. For the initial dispatch center (60,60), the feature map shows a high-risk, high-priority area (eigenvalue 0.8), and the decision tree determines it as a medium-risk terrain (information entropy 0.9, eigenvalue 0.8 is between 0.3-0.7). The following instructions are generated:

[0119] Command type: Delivery

[0120] Target coordinates: (60,60,2 meters) (2 meters above the ground for safety)

[0121] Movement parameters: Slow descent, the robot turns on the lidar to scan the surrounding 5 meters after landing

[0122] Sensor configuration: LiDAR + soil sensor, scanning frequency 5Hz

[0123] After the sampling robot landed, real-time feedback data indicated a 1-meter-diameter pit at the (60,60) cell. The module immediately triggered a dynamic adjustment: a "move to (59,60)" command was generated. The robot circumvented the pit at a speed of 0.5 m / s and, upon reaching the new location, began soil sampling. After completing the sampling operation, the feature map updated to show that the information entropy of the (55,55) cell had risen to 0.8 (newly identified rock distribution). The module then generated a "move to (55,55)" command, and the robot continued along the planned path.

[0124] Throughout the entire operation, the drone system achieves closed-loop control through the collaboration of multiple modules: sensors continuously collect data → feature parsing and correlation analysis update the terrain representation in real time → task scheduling dynamically adjusts priorities → control commands precisely drive robot execution. For example, if the drone detects a sudden increase in vegetation coverage in a certain area (such as a forest detected in a visual image), the environmental perception module replenishes data in real time, the terrain feature parsing module re-extracts features, the path planning module adjusts the correlation matrix, the task scheduling module updates sampling priorities, and the control command module replans a path to avoid the forest.

[0125] This dynamic adaptability throughout the entire process enables the UAV system to cope with uncertainties in complex terrain, such as unexpected obstacles and changing terrain features, ensuring that the sampling robot can efficiently complete its mission in unstructured environments such as mountains and hills. The seamless integration of every link from data acquisition to command execution demonstrates the system's integrated design across environmental perception, feature modeling, decision generation, and control execution, providing a reliable technical solution for automated sampling in complex scenarios.

[0126] It should be noted that, in this document, relational terms such as first and second, etc., are used only 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 terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0127] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A drone system capable of carrying a sampling robot, characterized in that: include: Environmental perception module, used to obtain real-time terrain data of the target area through multimodal sensors; A terrain feature analysis module, configured to perform grid segmentation and terrain feature extraction on the real-time terrain data to obtain a set of feature tensors of a local area of ​​the terrain environment; a path planning feature enhancement module, configured to 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; A task scheduling decision module is used to perform global feature fusion based on information entropy distribution on the set of optimized feature tensors of the local area of ​​the terrain environment to obtain a global environment scheduling decision feature map; A control instruction generation module, configured to generate a robot-mounted control signal based on the global environment scheduling decision feature map; The task scheduling decision module includes: a local feature dimensionality reduction unit, which is used to perform channel-dimensional maximum pooling processing on each tensor in the set of the local area optimization feature tensors of the terrain environment to obtain a set of local area optimization feature vectors of the terrain environment; a regional feature static information entropy calculation unit, which is used to calculate the information entropy value of each feature vector in the set of the local area optimization feature vectors of the terrain environment to generate a terrain environment static information entropy set; a regional clustering center determination unit, which is used to select the local area optimization feature vector of the terrain environment corresponding to the maximum information entropy value in the terrain environment static information entropy set as the initial scheduling center vector; a regional feature dynamic weight calculation unit, which is used to calculate the dynamic scheduling weight of each feature vector in the set of the local area optimization feature vector of the terrain environment and the initial scheduling center vector based on the spatial distance between each feature vector 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, which is used to use the dynamic scheduling weight set to perform weighted summation on the set of the local area optimization feature vectors of the terrain environment to generate the global environment scheduling decision feature map.

2. The UAV system capable of carrying a sampling robot according to claim 1, characterized in that: The terrain feature analysis module includes: a terrain data segmentation unit for spatially dividing the real-time terrain data according to a preset grid size to obtain a set of terrain data units; and a terrain feature encoding unit for inputting the data of each unit in the set of terrain data units into a terrain feature encoder based on a three-dimensional convolutional neural network to obtain a set of local area feature tensors of the terrain environment.

3. The UAV system capable of carrying a sampling robot according to claim 2, characterized in that: The path planning feature enhancement module includes: a feature tensor decomposition unit, which is used to perform 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 calculation unit, which is used to calculate the spatial correlation between any two feature vectors in the set of local area feature vectors of the terrain environment to generate a terrain environment regional correlation matrix; a regional spatial constraint correction unit, which is used to perform distance attenuation correction on the terrain environment regional correlation 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 matrix; and a regional feature optimization unit, which is used to perform multi-scale convolution fusion on the terrain environment regional constraint correlation matrix and the set of local area feature vectors of the terrain environment to obtain the terrain environment local area optimized feature tensor.

4. The UAV system capable of carrying a sampling robot according to claim 3, characterized in that: The regional feature correlation calculation unit includes: a feature space mapping subunit, which is used to map 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 subunit, which is used to calculate 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 area correlation matrix composed of multiple terrain environment area correlation values.

5. The UAV system capable of carrying a sampling robot according to claim 4, characterized in that: The regional feature optimization unit is specifically implemented as follows: performing dilated convolution processing on the terrain environment regional constraint association matrix to obtain a terrain environment regional constraint association feature matrix; inputting the set of terrain environment local area feature vectors and the terrain environment regional constraint association feature matrix into the 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 optimization feature tensor.

6. The UAV system capable of carrying a sampling robot according to claim 1, characterized in that: The regional feature static information entropy calculation unit is specifically implemented as follows: calculating the mean vector and variance vector of the local area optimized feature vector of the terrain environment; performing element-by-element difference calculation between the local area optimized feature vector 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.

7. The UAV system capable of carrying a sampling robot according to claim 6, characterized in that: The regional feature dynamic weight calculation unit is specifically implemented as follows: multiplying the information entropy value of the local area optimization feature vector 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 of the Euclidean distance between the local area optimization feature vector 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.

8. The UAV system capable of carrying a sampling robot according to claim 7, characterized in that: The control instruction generation module is specifically implemented as follows: inputting the global environment scheduling decision feature map into a decision tree-based control signal generator to obtain the robot-mounted control signal, which is used to indicate the deployment strategy of the sampling robot.

9. A method for controlling a drone capable of carrying a sampling robot, characterized in that: include: Environmental perception module, which collects real-time terrain data of the target area through multimodal sensors; A terrain feature analysis module performs 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 performs 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; a task scheduling decision module performs 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; a control instruction generation module generates a robot-mounted control signal based on the global environment scheduling decision feature map; The task scheduling decision module includes: a local feature dimensionality reduction unit, which is used to perform channel-dimensional maximum pooling processing on each tensor in the set of the local area optimization feature tensors of the terrain environment to obtain a set of local area optimization feature vectors of the terrain environment; a regional feature static information entropy calculation unit, which is used to calculate the information entropy value of each feature vector in the set of the local area optimization feature vectors of the terrain environment to generate a terrain environment static information entropy set; a regional clustering center determination unit, which is used to select the local area optimization feature vector of the terrain environment corresponding to the maximum information entropy value in the terrain environment static information entropy set as the initial scheduling center vector; a regional feature dynamic weight calculation unit, which is used to calculate the dynamic scheduling weight of each feature vector in the set of the local area optimization feature vector of the terrain environment and the initial scheduling center vector based on the spatial distance between each feature vector 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, which is used to use the dynamic scheduling weight set to perform weighted summation on the set of the local area optimization feature vectors of the terrain environment to generate the global environment scheduling decision feature map.

Images