A high-precision very low speed motor control method and system

Abstract

This invention provides a high-precision, ultra-low-speed motor control method and system. The invention relates to the field of motor control technology. The method includes acquiring an image of a seabed sampling area; determining an edge sampling area and multiple protruding sampling areas based on the image; determining multiple candidate sampling point information for the edge sampling area and multiple candidate sampling point information for each protruding sampling area based on the edge sampling area and the multiple protruding sampling areas; determining a second motor control scheme based on the robot sampling video and the multiple candidate sampling point information for each protruding sampling area; and controlling the motor to drive the robot to perform sampling based on the second motor control scheme. This method can accurately determine a suitable seabed sampling motor control scheme to complete seabed sampling efficiently and with low interference.

Description

Technical Field

[0001] This invention relates to the field of motor control technology, specifically to a high-precision ultra-low speed motor control method and system. Background Technology

[0002] With the development of marine resource exploration and deep-sea scientific research, the importance of seabed sampling technology is becoming increasingly prominent. Traditional seabed sampling equipment typically uses fixed-speed or simply speed-adjustable motor control, which is difficult to adapt to the complex and ever-changing seabed topography. Especially under extremely low-speed operating conditions, existing motor control systems generally suffer from insufficient control precision and poor dynamic response performance, leading to damage to fragile seabed samples or incomplete collection during the sampling process. Furthermore, due to the complex seabed topography, current technologies lack intelligent identification and adaptive control capabilities for sampling areas, often requiring manual intervention to adjust sampling strategies. This is not only inefficient but also prone to sampling failure due to human error. In addition, existing systems lack real-time feedback and dynamic adjustment mechanisms for sampling results, making it difficult to achieve high-precision, low-disturbance intelligent sampling operations. These problems severely restrict the efficiency and quality of deep-sea exploration and marine resource development.

[0003] Therefore, accurately determining the appropriate seabed sampling motor control scheme to complete seabed sampling efficiently and with low interference is an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem solved by this invention is how to accurately determine a suitable control scheme for the seabed sampling motor to complete seabed sampling efficiently and with low interference.

[0005] According to a first aspect, the present invention provides a high-precision ultra-low-speed motor control method, comprising: acquiring an image of a seabed sampling area; determining an edge sampling area and multiple protruding sampling areas based on the image of the seabed sampling area; determining multiple candidate sampling point information for the edge sampling area and multiple candidate sampling point information for each protruding sampling area based on the one edge sampling area and the multiple protruding sampling areas; generating a first motor control scheme based on the multiple candidate sampling point information of the edge sampling area; controlling a motor-driven robot to perform sampling based on the first motor control scheme and acquiring a robot sampling video; determining a second motor control scheme based on the robot sampling video and the multiple candidate sampling point information for each protruding sampling area; and controlling a motor-driven robot to perform sampling based on the second motor control scheme.

[0006] In one possible implementation, determining the second motor control scheme based on the robot sampling video and the information of multiple candidate sampling points in each prominent sampling region includes: constructing a sampling map, which includes multiple candidate sampling point nodes and multiple edges between the candidate sampling point nodes, wherein the node feature of each candidate sampling point node is candidate sampling point information, and the edges between nodes are the distances between sampling points; processing the sampling map based on a graph neural network to determine the sampling path for each prominent sampling region; generating sampling videos of the robot at different movement speeds along the sampling path in each prominent sampling region based on the sampling path in each prominent sampling region and the robot sampling video; determining the target movement speed of the robot along the sampling path in each prominent sampling region based on the sampling videos of the robot at different movement speeds along the sampling path in each prominent sampling region; and generating the second motor control scheme based on the target movement speed of the robot along the sampling path in each prominent sampling region.

[0007] In one possible implementation, generating a first motor control scheme based on multiple candidate sampling point information of the edge sampling region includes: clustering multiple candidate sampling point information of the edge sampling region to obtain K clusters; processing the K clusters based on a sampling determination model to obtain a sampling path of the edge sampling region; and generating a first motor control scheme based on the sampling path of the edge sampling region and an initial low moving speed.

[0008] In one possible implementation, determining the target moving speed of the robot's sampling path in each salient sampling area based on the sampling video of the robot's sampling path at different moving speeds in each salient sampling area includes: determining sample destruction information at each moving speed based on the sampling video of the robot's sampling path at different moving speeds in each salient sampling area; and determining the target moving speed of the robot's sampling path in each salient sampling area based on the sample destruction information at each moving speed.

[0009] According to a second aspect, the present invention provides a high-precision ultra-low-speed motor control system, comprising: an acquisition module for acquiring an image of a seabed sampling area; a sampling area determination module for determining an edge sampling area and multiple protruding sampling areas based on the image of the seabed sampling area; a candidate sampling point determination module for determining multiple candidate sampling point information for the edge sampling area and multiple candidate sampling point information for each protruding sampling area based on the one edge sampling area and the multiple protruding sampling areas; a first control scheme generation module for generating a first motor control scheme based on the multiple candidate sampling point information of the edge sampling area; a first sampling control module for controlling a motor-driven robot to perform sampling based on the first motor control scheme and acquiring a robot sampling video; a second control scheme determination module for determining a second motor control scheme based on the robot sampling video and the multiple candidate sampling point information of each protruding sampling area; and a second sampling control module for controlling a motor-driven robot to perform sampling based on the second motor control scheme.

[0010] In one possible implementation, the second control scheme determination module is further configured to: construct a sampling map, the sampling map including multiple candidate sampling point nodes and multiple edges between the multiple candidate sampling point nodes, the node features of each candidate sampling point node being candidate sampling point information, and the edges between nodes being the distance between sampling points; process the sampling map based on a graph neural network to determine the sampling path of each prominent sampling region; generate sampling videos of different moving speeds of the robot on the sampling path of each prominent sampling region based on the sampling path of each prominent sampling region and the robot sampling video; determine the target moving speed of the robot on the sampling path of each prominent sampling region based on the sampling videos of different moving speeds of the robot on the sampling path of each prominent sampling region; and generate a second motor control scheme based on the target moving speed of the robot on the sampling path of each prominent sampling region.

[0011] In one possible implementation, the first control scheme generation module is further configured to: cluster K clusters based on multiple candidate sampling point information of the edge sampling region; process the K clusters based on the sampling determination model to obtain the sampling path of the edge sampling region; and generate a first motor control scheme based on the sampling path of the edge sampling region and the initial low moving speed.

[0012] In one possible implementation, the second control scheme determination module is further configured to: determine sample destruction information at each movement speed based on the sampling video of the robot's sampling path at different movement speeds in each salient sampling area; and determine the target movement speed of the robot's sampling path in each salient sampling area based on the sample destruction information at each movement speed.

[0013] According to a third aspect, embodiments of the present invention provide an electronic device, including: a processor; a memory; and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method as described above, the method including: acquiring an image of a seabed sampling area; determining an edge sampling area and a plurality of protruding sampling areas based on the image of the seabed sampling area; determining multiple candidate sampling point information for the edge sampling area and multiple candidate sampling point information for each protruding sampling area based on the one edge sampling area and the plurality of protruding sampling areas; generating a first motor control scheme based on the multiple candidate sampling point information of the edge sampling area; controlling a motor-driven robot to perform sampling based on the first motor control scheme and acquiring a robot sampling video; determining a second motor control scheme based on the robot sampling video and the multiple candidate sampling point information for each protruding sampling area; and controlling the motor-driven robot to perform sampling based on the second motor control scheme.

[0014] According to a fourth aspect, this embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned high-precision ultra-low-speed motor control method. The method includes: acquiring an image of a seabed sampling area; determining an edge sampling area and multiple protruding sampling areas based on the image of the seabed sampling area; determining multiple candidate sampling point information for the edge sampling area and multiple candidate sampling point information for each protruding sampling area based on the edge sampling area and the multiple protruding sampling areas; generating a first motor control scheme based on the multiple candidate sampling point information of the edge sampling area; controlling a motor-driven robot to perform sampling based on the first motor control scheme and acquiring a robot sampling video; determining a second motor control scheme based on the robot sampling video and the multiple candidate sampling point information for each protruding sampling area; and controlling a motor-driven robot to perform sampling based on the second motor control scheme.

[0015] This invention provides a high-precision, ultra-low-speed motor control method and system. The method includes acquiring an image of a seabed sampling area; determining an edge sampling area and multiple protruding sampling areas based on the image of the seabed sampling area; determining multiple candidate sampling point information for the edge sampling area and multiple candidate sampling point information for each protruding sampling area based on the edge sampling area and the multiple protruding sampling areas; generating a first motor control scheme based on the multiple candidate sampling point information of the edge sampling area; controlling a motor-driven robot to perform sampling based on the first motor control scheme and acquiring a robot sampling video; determining a second motor control scheme based on the robot sampling video and the multiple candidate sampling point information for each protruding sampling area; and controlling a motor-driven robot to perform sampling based on the second motor control scheme. This method can accurately determine a suitable seabed sampling motor control scheme to complete seabed sampling efficiently and with low interference. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a high-precision ultra-low speed motor control method provided in an embodiment of the present invention;

[0017] Figure 2 A flowchart illustrating the generation of a first motor control scheme is provided in an embodiment of the present invention.

[0018] Figure 3 A flowchart illustrating the determination of a second motor control scheme is provided in an embodiment of the present invention.

[0019] Figure 4 A flowchart illustrating the process of determining the target moving speed of a robot's sampling path in each prominent sampling area, as provided in an embodiment of the present invention;

[0020] Figure 5 This is a schematic diagram of a high-precision ultra-low speed motor control system provided in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0022] In this embodiment of the invention, the following are provided: Figure 1 The above describes a high-precision ultra-low speed motor control method, which includes steps S1 to S7:

[0023] Step S1: Obtain an image of the seabed sampling area.

[0024] The images of the seabed sampling area were captured by a high-definition underwater camera mounted on a robot.

[0025] Images of the seabed sampling area can present visual information such as the topographic features, biological distribution, and soil cover type of the seabed in that area.

[0026] Step S2: Based on the image of the seabed sampling area, determine an edge sampling area and multiple protruding sampling areas.

[0027] In some embodiments, a region segmentation model can be used to determine an edge sampling region and multiple salient sampling regions. The region segmentation model is a convolutional neural network model. The input to the region segmentation model is an image of the seabed sampling region, and the output of the region segmentation model is an edge sampling region and multiple salient sampling regions.

[0028] Convolutional Neural Network (CNN) models are deep learning models focused on image processing, comprising convolutional layers, pooling layers, and fully connected layers. Convolutional layers extract local features from images by sliding convolutional kernels. Pooling layers reduce the dimensionality of extracted features, preserving key information while reducing computational cost. Fully connected layers integrate features extracted from multiple layers to achieve classification or segmentation. CNNs can automatically learn hierarchical features of images and can progressively abstract high-level semantic features from low-level visual features.

[0029] An edge sampling region is a region located at the boundary of the overall sampling region, determined by the image analysis of the seabed sampling region using a region segmentation model. It includes information such as the location coordinates, boundary range, and terrain features of the edge sampling region.

[0030] An edge sampling region can be used to perform experimental sampling in order to verify the environmental impact of the sampling operation.

[0031] Multiple prominent sampling areas are multiple regions with significant characteristics identified by the regional division model, and include information such as regional location coordinates, boundary range, topographic features, biological species, and soil type.

[0032] Notable features include dense biological distribution and unique topographical structure.

[0033] Images of the seabed sampling area contain complete visual information about the area to be sampled, such as color differences reflecting soil type, contour shapes reflecting terrain, and pixel distribution reflecting biological aggregation in different areas. These features provide a basis for the convolutional neural network to divide the area.

[0034] Convolutional neural networks (CNNs) can process the internal features of seabed sampled areas layer by layer to identify edge sampling regions and salient sampling regions. The convolutional layers of a CNN extract key local features using different types of kernels. Kernels used to locate potential edge sampling regions can capture topographic texture changes and environmental feature differences within the boundary zones of the seabed sampled area, such as the visual transition features of different sediment areas. Kernels used to identify potential salient sampling regions can focus on pixel clustering features corresponding to dense biological distribution or special topographic structures. Pooling layers then perform dimensionality reduction on these local features, eliminating redundant pixel information and reducing computational cost while retaining core information of key features through max pooling, such as feature transition trends in edge zones and the saliency of salient regions. Through iterative cycles of multiple convolutions and pooling, the initially extracted low-level features are gradually integrated into higher-level features with greater semantic value, such as comprehensive features that clearly characterize the attributes of edge sampling regions and feature combinations that reflect the value of salient sampling regions. Fully connected layers can feed these high-level semantic features into a classifier. The classifier can label pixels or small regions in the image according to a preset region determination standard, and finally integrate the samples from the seabed sampling area to form an edge sampling region and multiple salient sampling regions.

[0035] In some embodiments, determining an edge sampling region and multiple salient sampling regions based on the image of the seabed sampling region includes steps S21-S23:

[0036] Step S21: Determine the topographic relief distribution information, biological community distribution information, and sediment type distribution information based on the images of the seabed sampling area.

[0037] In some embodiments, convolutional neural networks can be used to determine information on topographic relief distribution, biological community distribution, and sediment type distribution.

[0038] The topographic relief distribution information is the information on the distribution of elevation changes in different areas of the seabed output by a convolutional neural network, including the location and relative height differences of uplifts, depressions, and flat areas within the region.

[0039] The biological community distribution information is the information on the distribution of seabed organisms output by a convolutional neural network, which includes the location and distribution range of various biological clusters.

[0040] Sediment type distribution information is information on the distribution of seabed sediment types output by a convolutional neural network, which includes the location and coverage of different types of sediments.

[0041] Convolutional neural networks (CNNs) can automatically capture local features related to boundaries, topography, organisms, and sediments in images using convolutional kernels of different sizes. For example, edge detection convolutional kernels can identify grayscale and texture differences between sandy and rocky areas, thereby locating seafloor boundary points. Texture extraction convolutional kernels can distinguish the visual features of sediments. Pooling layers in CNNs can retain key features while compressing redundant information, and simultaneously enhance the differences in topographic relief and the aggregation characteristics of biological communities. Finally, fully connected layers classify and integrate the extracted features to map local features to global distribution information, outputting seafloor boundary points, topographic relief distribution, biological community distribution, and sediment type distribution.

[0042] Step S22: Based on the seabed environment boundary points, the topographic relief distribution, the biological community distribution, and the sediment type distribution, determine multiple preliminary edge area information and multiple preliminary prominent sampling area information.

[0043] In some embodiments, a deep neural network may be used to determine multiple initial edge region information and multiple initial salient sampling region information.

[0044] Deep neural networks (DNNs) are artificial neural networks composed of multiple hidden layers that transmit information by simulating neuron connections. Each neuron in a DNN layer is connected to the next layer via weights, and activation functions introduce nonlinear transformations, enabling them to learn complex nonlinear relationships from input data. DNNs can process high-dimensional, structured data and possess powerful feature fitting and information generation capabilities.

[0045] Multiple preliminary edge region information refers to information on multiple candidate regions with potential edge sampling region characteristics output by a deep neural network. This preliminary edge region information includes the specific coordinate boundaries within the region, boundary transition clarity scores, and terrain stability indices.

[0046] The boundary transition clarity score is a scale of 0-10. The higher the score, the more significant the difference in environmental characteristics between the area and the adjacent area, and the more obvious the boundary transition.

[0047] The terrain stability index is represented by a value from 0 to 100. The lower the value, the more easily the terrain in the area is changed by sampling operations, such as the presence of soft sediments or unstable terrain structures.

[0048] Multiple preliminary salient sampling region information refers to information on multiple candidate regions with potential salient sampling region value output by a deep neural network. This preliminary salient sampling region information includes the specific coordinate boundaries within the region, sample value correlation, and biological density score.

[0049] The sample value correlation is represented by a value between 0 and 100. A higher sample value correlation indicates a greater probability that there are high-value sampling samples, such as special organisms or rare sediments, in the area.

[0050] The biodiversity score is a scale of 0 to 10. A higher score indicates a higher degree of biodiversity in the area and the potential presence of unique biological communities or ecological characteristics.

[0051] Deep neural networks possess the ability to fuse multiple features and perform precise classification and selection. Through hidden layers, deep neural networks can integrate multi-dimensional data such as seabed boundary points and topographic relief distribution, and automatically discover the correlations between different data points, such as the superposition relationship between boundary points and topographic relief, thus avoiding the limitations of single-feature judgment. Deep neural networks can learn the feature patterns of edge regions and prominent sampling regions. For example, they can identify the combination of obvious boundary transitions and low topographic stability as initial edge regions, and the combination of dense biological populations and high sample value correlation as initial prominent sampling regions. Finally, the output layer of the deep neural network can match the integrated features with preset region selection criteria, thereby outputting multiple initial edge region information and multiple initial prominent sampling region information.

[0052] Step S23: Based on the multiple preliminary edge region information and the multiple preliminary salient sampling region information, determine one edge sampling region and multiple salient sampling regions.

[0053] In some embodiments, a deep neural network may be used to determine an edge sampling region and multiple salient sampling regions.

[0054] Deep neural networks can perform weighted analysis on quantitative parameters such as boundary transition clarity scores and terrain stability indices in multiple initial edge region information and sample value correlation and biological density scores in multiple initial salient sampling region information through multi-layer nonlinear transformations. For edge sampling regions, deep neural networks can prioritize the region with the highest boundary transition clarity score and terrain stability index that meets the sampling operation requirements from multiple initial edge region information, ultimately determining one edge sampling region. For salient sampling regions, deep neural networks can select multiple regions from multiple initial salient sampling region information that can both cover high-value sample features and avoid regional feature redundancy based on the rules of sorting sample value correlation from high to low and the rule of non-repetition of regional spatial distribution, ultimately outputting one edge sampling region and multiple salient sampling regions.

[0055] Step S3: Based on the one edge sampling region and the multiple salient sampling regions, determine multiple candidate sampling point information for the edge sampling region and multiple candidate sampling point information for each salient sampling region.

[0056] In some embodiments, a candidate information determination model can be used to determine multiple candidate sampling point information for an edge sampling region and multiple candidate sampling point information for each salient sampling region. The candidate information determination model is a deep neural network model. The input to the candidate information determination model is the edge sampling region and the multiple salient sampling regions, and the output of the candidate information determination model is the multiple candidate sampling point information for the edge sampling region and the multiple candidate sampling point information for each salient sampling region.

[0057] The information on multiple candidate sampling points in the edge sampling region is used to determine the information of multiple potential sampling points located in the edge sampling region output by the model. The information on each candidate sampling point in the edge sampling region includes the precise location coordinates of each candidate sampling point, the terrain information of the location, the soil surface condition of the location, the surrounding biological distribution density, and whether the sampling operation space is sufficient.

[0058] Terrain information includes the slope value, flatness, presence and type of obstacles at the location.

[0059] Soil surface conditions include soil particle size, surface moisture, and soil compaction.

[0060] The density of organisms in the surrounding area refers to the number of individuals and the evenness of their distribution within a certain range around the site.

[0061] The information on multiple candidate sampling points for each prominent sampling area is determined by the model output based on the candidate information. The information on each candidate sampling point in each prominent sampling area includes the precise location coordinates of each candidate sampling point, detailed topographic information of the location, surface soil condition, surrounding biological density, and whether there is sufficient space for sampling operations.

[0062] Deep neural networks possess the ability to deeply mine and accurately map regional features. The multiple hidden layers of a deep neural network can abstract the association between regional features and sampling point requirements layer by layer through nonlinear transformations. For experimental sampling needs in edge sampling areas, the model can focus on mining features such as terrain stability and sufficient operational space, and select candidate sampling points with gentle slopes and no obstacles from the edge area. For key sampling needs in prominent sampling areas, the model can associate features such as biological distribution density and soil type specificity, and locate candidate sampling points with minimal environmental interference within the prominent area, obtaining target organisms, special soil samples, and other samples. Simultaneously, the model continuously optimizes weight parameters through backpropagation, constantly improving the accuracy of feature association to ensure that the output candidate sampling point information strictly matches the range and attributes of the corresponding area, and fully covers precise location coordinates, terrain details, soil conditions, biological distribution density, and sampling operational space.

[0063] Step S4: Generate a first motor control scheme based on the information of multiple alternative sampling points in the edge sampling area.

[0064] In some embodiments, Figure 2 This is a flowchart illustrating the generation of a first motor control scheme according to an embodiment of the present invention. The generation of the first motor control scheme includes steps S31 to S33:

[0065] Step S31: Cluster K clusters are obtained based on the information of multiple candidate sampling points in the edge sampling region.

[0066] The clustering algorithm is the K-means clustering algorithm.

[0067] K-means clustering is an unsupervised clustering analysis algorithm that uses iterative computation. It can be used to divide a dataset into K independent clusters based on feature similarity, where data points within each cluster have highly similar features. In some embodiments, the value of K can be pre-set manually.

[0068] The process of clustering multiple candidate sampling points in the edge sampling region using the K-means clustering algorithm is as follows: First, K sampling points are randomly selected from the dataset of multiple candidate sampling points in the edge sampling region as initial cluster centers. Next, for each candidate sampling point in the dataset, the distance between it and the K initial cluster centers is calculated using Euclidean distance, and the candidate sampling point is assigned to the corresponding cluster according to the principle of closest distance. After all candidate sampling points have been partitioned, the feature mean values ​​of all candidate sampling points in each cluster, such as the mean of location coordinates, the mean of slope, and the mean of soil moisture, are recalculated, and the cluster centers of each cluster are updated accordingly. The above partitioning and updating steps are repeated until the change in the features of the cluster centers is less than a preset threshold. At this point, the clustering process converges, and K clusters are finally obtained.

[0069] Each cluster represents a set of candidate sampling points with similar characteristics within the edge sampling area. These characteristics include the location distribution of the sampling points, topographic conditions, soil conditions, etc.

[0070] Clustering can effectively integrate the complex information of candidate sampling points within the edge sampling region. Because the number of candidate sampling points in the edge region is large and their characteristics vary, directly analyzing each candidate sampling point individually in the edge sampling region involves extremely high computational complexity. However, grouping sampling points with similar characteristics into one category through clustering can significantly simplify the data structure, and the clustering results can intuitively present the distribution characteristics and type differences of candidate sampling points within the edge region.

[0071] Step S32: Based on the sampling determination model, process the K clusters to obtain the sampling path of the edge sampling region.

[0072] The sampling determination model is a Transformer model. The input of the sampling determination model is the K clusters, and the output of the sampling determination model is the sampling path of the edge sampling region.

[0073] The Transformer model is a deep learning model based on a self-attention mechanism. It consists of an encoder and a decoder. The encoder uses multi-head self-attention to capture the relationships between elements in the input sequence, transforming the input into a feature vector containing global information. The decoder uses the encoder's output and the self-attention mechanism to generate the target sequence. The Transformer model excels at handling sequential data and data with long-range dependencies, and effectively uncovers global correlations between data points.

[0074] The sampling path in the edge sampling region is the path taken by the robot, determined by sampling, as it sequentially traverses each cluster within the edge sampling region to complete the sampling. The sampling path in the edge sampling region includes the access order of each cluster, the traversal order of sampling points within each cluster, and the specific coordinate trajectory of each path segment.

[0075] The K clusters contain grouping information for candidate sampling points within the edge sampling regions. Features such as the center location of each cluster and the number of sampling points within it reflect the sampling requirements of each region. This information provides the basis for the Transformer model to plan its path, allowing the model to determine the optimal access order by analyzing the spatial relationships between clusters.

[0076] The Transformer model's encoder, through a multi-head self-attention mechanism, calculates the association weights between each cluster and other clusters, captures spatial dependencies between clusters, and generates cluster feature vectors containing global location information. The decoder, based on the encoder's output feature vectors and the path planning objective, progressively generates the access order through the self-attention mechanism. The decoder first determines the starting cluster, then selects the next cluster to visit based on the inter-cluster association weights, until all clusters are included in the path. Simultaneously, the model can plan the traversal order of internal sampling points for each cluster, such as expanding outwards from the cluster center, ultimately integrating to form a complete sampling path for the edge sampling region.

[0077] Step S33: Generate a first motor control scheme based on the sampling path of the edge sampling area and the initial low moving speed.

[0078] Initial low moving speed refers to the extremely low baseline moving speed of the robot set in advance when conducting experimental sampling in the edge sampling area. This is done to minimize interference with the environment and ensure the safety of the sampling operation. The goal is to ensure that the robot can reliably complete the sampling test in the edge area at an extremely low speed.

[0079] In some embodiments, a first scheme determination model can be used to generate a first motor control scheme. The first scheme determination model is a deep neural network. The input to the first scheme determination model is a sampling path of an edge sampling region and an initial low moving speed, and the output of the first scheme determination model is the first motor control scheme.

[0080] The first motor control scheme is a specific control scheme generated by the first scheme determination model to control the motor-driven robot to perform sampling tasks in the edge sampling area. The first motor control scheme includes the robot's movement speed adjusted based on the initial low movement speed in each segment of the path, the motor steering angle, the motor start and stop commands, the motor power output during sampling, the inter-region transition commands, and the fault emergency commands.

[0081] Deep neural networks possess the ability to map commands under multiple constraints. The sampling path in the edge sampling region provides the spatial trajectory basis for robot movement, clearly defining key nodes and coordinate ranges such as straight sections, turning sections, and sampling stops. This provides a precise spatial range for the model to divide the control units for speed and steering requirements of different road segments. The initial low movement speed serves as a safe speed benchmark for experimental sampling, limiting the core range of motor speed and preventing the robot from damaging the edge environment due to excessive speed. Through its multi-layered structure, the deep neural network can uncover the correlations in the input data. For straight sections in the path, the model can combine the initial low movement speed with the road slope to generate suitable and smooth speed commands. For turning sections, the model can adjust the steering angle and speed attenuation ratio based on the path curvature. For sampling stops, the model can link speed and start / stop timing to ensure precise motor stopping during sampling. Through iterative optimization, the deep neural network can transform path constraints and speed benchmarks into specific control commands such as motor speed, steering angle, and power output, ultimately forming a first-class motor control scheme that balances safety at extremely low speeds and stable sampling operations.

[0082] Step S5: Based on the first motor control scheme, control the motor to drive the robot to perform sampling and acquire the robot sampling video.

[0083] The robot sampling video is recorded by a camera mounted on the robot during the sampling process, controlled by the first motor control scheme, in an edge sampling area. The video records the robot's movement, the operational status of the sampling device, and the real-time condition of the sampling area, such as whether the sample was damaged due to movement.

[0084] Step S6: Determine the second motor control scheme based on the robot sampling video and the information of multiple alternative sampling points for each prominent sampling area.

[0085] In some embodiments, Figure 3 This is a flowchart illustrating the process of determining a second motor control scheme according to an embodiment of the present invention. The process of determining the second motor control scheme includes steps S41 to S45:

[0086] Step S41: Construct a sampling graph. The sampling graph includes multiple candidate sampling point nodes and multiple edges between the multiple candidate sampling point nodes. The node features of each candidate sampling point node are candidate sampling point information, and the edges between nodes are the distances between sampling points.

[0087] The sampling map is a graphical structure built upon information from multiple candidate sampling points within each salient sampling region. The sampling map includes multiple candidate sampling point nodes and multiple edges connecting these nodes. Each candidate sampling point node corresponds to one of the multiple candidate sampling points within each salient sampling region, and its node feature is that candidate sampling point information. The multiple edges connecting the candidate sampling point nodes are used to connect any two candidate sampling point nodes within the same salient sampling region, and each edge represents the straight-line distance between the corresponding candidate sampling points. By constructing the sampling map, the distribution relationship and spatial distance of candidate sampling points within the same salient sampling region can be visually presented.

[0088] Step S42: Process the sampling map based on the graph neural network to determine the sampling path for each prominent sampling region.

[0089] Graph Neural Networks (GNNs) are deep learning models that can process graph data. GNNs update node representations by aggregating the features of a node itself and its neighbors. GNNs utilize the topological structure of the connections between nodes and edges in the graph to allow each node to learn information from its surrounding nodes, thereby capturing global dependencies within the graph. The input to the GNN is the sampled graph, and the output is the sampling path for each salient sampling region.

[0090] The sampling path for each salient sampling region is the path the robot takes to sequentially visit each candidate sampling point within that region, output by analyzing the sampling map using a graph neural network. Each sampling path for a salient sampling region includes the order in which the sampling points are visited, the coordinate trajectory of each path segment, and the estimated time to reach each sampling point.

[0091] By constructing a sampling map, the spatial distribution relationships between candidate sampling points within a prominent sampling area can be clearly reflected. This spatial distribution information can be used to plan efficient sampling paths, as the spatial location of sampling points directly determines the robot's movement cost and sampling efficiency. Using candidate sampling point information as node features and the distance between sampling points as edge features allows for more comprehensive utilization of the spatial data information of the sampling area. This helps the graph neural network model better understand the spatial relationships and distribution patterns of each sampling point, providing data support for optimizing the sampling path. Processing sampling map data based on a graph neural network model can effectively learn the complex spatial dependencies and feature transfer between candidate sampling points, thereby more accurately planning the optimal path that balances sampling coverage and movement efficiency. Compared to traditional path planning algorithms, graph neural networks have stronger global relationship capture and path optimization capabilities when processing sampling maps with complex spatial relationships.

[0092] When processing a sampling graph, a graph neural network (Graph Neural Network) aggregates the features of its neighbors and updates its representation by combining these features with its own. This process is performed layer by layer, allowing each node to gradually learn the structural information of the global graph, such as which nodes are key hubs. The Graph Neural Network can optimize its parameters using a loss function, ultimately outputting the access priority of each node. Based on this priority, the model selects the node at the center of a salient region as the starting point, and then sequentially selects the highest-priority unvisited node as the next stop, until all candidate sampling points are included in the path, thus forming the sampling path for each salient sampling region.

[0093] Step S43: Based on the sampling path of each salient sampling area and the robot sampling video, generate sampling videos of the robot at different movement speeds along the sampling path of each salient sampling area.

[0094] In some embodiments, a generative adversarial network (GAN) can be used to generate sampled videos of the robot's sampling path at different speeds in each salient sampling region. The input to the GAN is the sampling path of each salient sampling region and the robot's sampled video, and the output of the GAN is the sampled video of the robot's sampling path at different speeds in each salient sampling region.

[0095] Generative Adversarial Networks (GANs) consist of two subnetworks: a generator and a discriminator. The generator produces fake data that conforms to a target distribution, while the discriminator distinguishes the fake data generated by the generator from real data. Both networks continuously optimize through adversarial training. The generator strives to generate fake data that the discriminator cannot distinguish, while the discriminator works to improve its discriminative ability, ultimately enabling the generator to output highly realistic simulated data.

[0096] The sampling videos of the robot at different movement speeds along the sampling path in each salient sampling region are generated by a generative adversarial network, simulating the robot performing sampling at different speeds along the corresponding salient region sampling path. These videos simulate the robot's movement trajectory, sampling process, and anticipated environmental changes caused by sampling at different speeds along the sampling path in each salient sampling region. Each salient region corresponds to multiple simulated videos at different speeds, and each video is based on a real sampling scene and simulates the complete process of the robot moving along the sampling path in that salient region at the corresponding speed.

[0097] The sampling path of each prominent sampling area provides constraints on the robot's movement trajectory. The robot's sampled video provides realistic features of the seabed environment, robot motion characteristics, and environmental interaction features. These features provide the generator with a reference for real data distribution and ensure that the generated simulated video is highly consistent with the real scene in terms of visual effects and motion details. The generator of the Generative Adversarial Network (GAN) can extract the correlation patterns between these spatial constraints and dynamic features through multi-layer neural networks. For example, the matching relationship between the slope of a certain road segment and the robot's speed, the temporal correspondence between the robotic arm's movements and speed at sampling points, etc. When generating new videos, it can adaptively adjust these patterns based on the target speed. For example, while keeping the path spatial coordinates unchanged, it can change the time scale of the motion through interpolation or resampling techniques, while simultaneously adjusting the dynamic performance of environmental feedback and equipment status. The discriminator continuously compares the differences between the generated video and the real video, such as motion coherence, physical rationality, and environmental interaction logic, and then optimizes the generator's parameters in reverse. Ultimately, the sampled videos generated by the generator at different speeds strictly follow the spatial constraints of the sampling path and conform to the physical laws of the real scene in terms of dynamic details.

[0098] Step S44: Determine the target moving speed of the robot's sampling path in each salient sampling area based on the sampling video of the robot's different moving speeds along the sampling path in each salient sampling area.

[0099] In some embodiments, Figure 4 This is a flowchart illustrating a method for determining the target moving speed of a robot's sampling path in each salient sampling area, as provided in an embodiment of the present invention. The determination of the target moving speed of the robot's sampling path in each salient sampling area includes steps S51-S52:

[0100] Step S51: Determine the sample destruction information at each movement speed based on the sampling video of the robot at different movement speeds along the sampling path in each salient sampling area.

[0101] In some embodiments, a destruction information determination model can be used to determine the destruction information of the sampled samples at each movement speed. The destruction information determination model is a Transformer model. The input to the destruction information determination model is the sampled video of the robot's sampling path at different movement speeds in each salient sampling region, and the output of the destruction information determination model is the destruction information of the sampled samples at each movement speed.

[0102] The damage information of the sampled specimens at each movement speed is used to determine the model output. This information provides detailed details of the damage caused to the seabed environment and specimens by the robot at each movement speed. The damage information includes the damage type, quantified damage level, damage location coordinates, damage impact range, and the correlation between damage and movement speed.

[0103] The types of damage include soil sample structure damage, biological sample damage, and seabed topographic disturbance.

[0104] The sampling videos, showing the robot moving at different speeds along its sampling path in each prominent sampling area, simulated complete visual information about the robot's trajectory, sampling actions, and environmental changes at different speeds. For example, excessively high speeds might cause the robot to turn, resulting in soil splashing, while excessively low speeds might lead to low sampling efficiency but minimal environmental interference. This visual information provides the Transformer model with a basis for identifying destructive features.

[0105] The Transformer model can decompose the sampled video at different movement speeds into frames, transforming the video into a continuous sequence of image frames, and labeling each frame with a timestamp and corresponding movement speed. This image frame sequence is then input into the Transformer model's encoder. The encoder's multi-head self-attention mechanism processes the frame sequence, capturing temporal features of damage changes. For example, by comparing image frames of the soil area before and after sampling, it can identify the change in soil structure from intact to loose. Simultaneously, the attention mechanism focuses on the areas where the robot interacts with the environment, extracting damage features from these areas. Through multi-layer processing, the encoder abstracts the visual features of the image frames into high-level semantic features such as "soil damage" and "biological damage," converting them into feature vectors. The decoder receives the feature vectors from the encoder and, combined with the video's speed identifier, decodes them. For soil damage, it calculates the proportion of loose soil area to the sampled area, converting it into a quantified damage value. For biological damage, it counts the number and severity of damaged organisms, converting them into a quantified damage value. For terrain disturbance, it measures changes in terrain height, converting them into a quantified damage value. Simultaneously, the decoder can determine the coordinates of the damage location and the affected area based on the coordinate information of the image frames. The decoder can integrate all information and output the damage information of the sampled samples at each movement speed, clearly defining the damage type, the quantification value of the damage degree, the coordinates of the damage location, the range of damage impact, and the relationship between the damage and the movement speed.

[0106] Step S52: Determine the target moving speed of the robot's sampling path in each salient sampling area based on the sample destruction information of each moving speed.

[0107] In some embodiments, a velocity determination model can be used to determine the target moving speed of the robot's sampling path in each salient sampling region. The velocity determination model is a deep neural network. The input to the velocity determination model is the sample destruction information for each moving speed, and the output of the velocity determination model is the target moving speed of the robot's sampling path in each salient sampling region.

[0108] The target moving speed of the robot on the sampling path in each salient sampling area is determined by the speed determination model after processing the sampling sample destruction information for each moving speed, and is the optimal moving speed applicable to the sampling path of the corresponding salient area.

[0109] The sample destruction information for each moving speed records the specific impact of sampling operations on the samples at different speeds. The model can directly establish a mapping between moving speed and sampling quality using this information. Low speeds may result in minimal destruction but low efficiency, while high speeds may increase efficiency but exacerbate destruction. By comparing destruction data at different speeds, the model can clearly define the speed range corresponding to the acceptable destruction threshold, thus providing a quantitative basis for selecting a target moving speed that balances sampling efficiency and sample integrity.

[0110] Deep neural networks can uncover complex relationships within destructive information through multiple hidden layers, transforming features such as quantified damage levels, speed values, and sampling time into mappings to a high-dimensional feature space. They can also learn latent rules that optimize efficiency within a damage threshold. The model can handle noise and redundancy in the destructive information, such as subtle differences in damage at similar speeds, while focusing on key features through weight adjustments, such as speed thresholds exceeding the damage threshold and speed points where efficiency leaps. Furthermore, the iterative optimization mechanism of deep neural networks can combine preset constraints such as maximum permissible damage and minimum sampling efficiency to precisely locate the optimal solution in the relationship between speed and damage, ultimately outputting a target movement speed that balances sample protection and operational efficiency.

[0111] Step S45: Generate a second motor control scheme based on the target moving speed of the robot's sampling path in each salient sampling area.

[0112] In some embodiments, a second scheme determination model can be used to generate a second motor control scheme. The second scheme determination model is a deep neural network. The input to the second scheme determination model is the target movement speed of the robot along the sampling path in each salient sampling region, and the output of the second scheme determination model is the second motor control scheme.

[0113] The second motor control scheme is generated by processing the target moving speed of the robot's sampling path in each salient sampling area and the sampling path of each salient sampling area using the second scheme determination model. It is a set of control commands used to control the motor-driven robot to perform key sampling in all salient sampling areas. The second motor control scheme includes the moving speed of each segment of the sampling path in each salient area, the motor steering angle, the motor start and stop commands, the motor power output during sampling, the inter-area transition commands, and the fault emergency commands.

[0114] The second motor control scheme can ensure that the robot moves at the target speed in each prominent area, while taking into account sampling efficiency and environmental damage control, and can simultaneously achieve orderly connection of sampling in multiple areas.

[0115] The target movement speed of the robot along the sampling path in each prominent sampling area clearly defines the baseline speed, speed adjustment range, and key nodes to be maintained during the sampling process, serving as the core benchmark for generating motor control commands. These speed parameters are directly related to the motor's rotational speed, torque output, and start-stop timing. The target speed determines the baseline value of the motor's power output, the speed variation range specifies the rate limits for acceleration and deceleration, and the speed requirements at the sampling points define the motor's start-stop timing. Through these parameters, the model can transform abstract speed requirements into specific operational constraints that the motor can execute.

[0116] Deep neural networks possess the ability to accurately model the mapping relationship between target movement speed parameters and motor operating parameters. Through multiple hidden layers, deep neural networks can analyze the multi-dimensional features of target movement speed. The model can transform the target speed of different road segments in each prominent sampling area—such as constant speed on straight sections, limited speed on turning sections, and deceleration thresholds before stopping at sampling points—into specific executable operating parameters for the motor, including the motor's base speed, steering adjustment angle, power output intensity, and duration for the corresponding road segment. Deep neural networks can also uncover the correlation between target speed variation patterns and motor load and path terrain. For example, for target speeds on steep slopes, the model can calculate the additional torque value the motor needs to compensate to prevent speed drops. For deceleration requirements before sampling points, the model can deduce the motor's deceleration rate limit to prevent equipment impact caused by sudden braking. Furthermore, deep neural networks can integrate physical constraints on motor operation, such as maximum speed and torque limits, and process complex scenarios such as speed switching and load fluctuations through nonlinear transformations. Ultimately, the output is a complete second motor control scheme containing motor speed curves for each road segment, steering adjustment timing, power distribution strategies, and emergency adjustment commands, ensuring the robot stably maintains the target movement speed along the sampling path.

[0117] Step S7: Control the motor to drive the robot to perform sampling based on the second motor control scheme.

[0118] Once the second motor control scheme is determined, the robot is driven to enter each prominent sampling area sequentially according to the instructions in the second motor control scheme, moving along the sampling path of the corresponding area, and performing sampling operations at each candidate sampling point, based on the instructions in the second motor control scheme, such as speed, steering, start / stop, and power output.

[0119] Based on the same inventive concept Figure 5 This invention provides a schematic diagram of a high-precision ultra-low speed motor control system, which includes:

[0120] Acquisition module 61 is used to acquire images of the seabed sampling area;

[0121] The sampling area determination module 62 is used to determine an edge sampling area and multiple salient sampling areas based on the image of the seabed sampling area;

[0122] The alternative sampling point determination module 63 is used to determine multiple alternative sampling point information for the edge sampling region and multiple alternative sampling point information for each salient sampling region based on the one edge sampling region and the multiple salient sampling regions.

[0123] The first control scheme generation module 64 is used to generate a first motor control scheme based on the information of multiple alternative sampling points in the edge sampling region.

[0124] The first sampling control module 65 is used to control the motor to drive the robot to perform sampling based on the first motor control scheme and to acquire the robot sampling video;

[0125] The second control scheme determination module 66 is used to determine the second motor control scheme based on the robot sampling video and the information of multiple alternative sampling points in each salient sampling area.

[0126] The second sampling control module 67 is used to control the motor-driven robot to perform sampling based on the second motor control scheme.

[0127] It should be noted that, in order to simplify the descriptions disclosed herein and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of embodiments of this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. However, this method of disclosure does not imply that the subject matter of this specification requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of a single embodiment disclosed above.

[0128] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A high-precision ultra-low speed motor control method, characterized in that, include: Acquire images of the seabed sampling area; Based on the image of the seabed sampling area, an edge sampling area and multiple prominent sampling areas are determined. An edge sampling area is a region located at the boundary of the overall sampling area, determined by the region segmentation model after analyzing the image of the seabed sampling area. The edge sampling area includes location coordinates, boundary range, and topographic feature information. Multiple prominent sampling areas are multiple regions with significant features, determined by the region segmentation model. The prominent sampling areas include region location coordinates, boundary range, topographic features, biological species, and soil type information. Based on the one edge sampling region and the multiple salient sampling regions, multiple candidate sampling point information for the edge sampling region and multiple candidate sampling point information for each salient sampling region are determined; A first motor control scheme is generated based on multiple candidate sampling point information of the edge sampling region, wherein generating the first motor control scheme based on multiple candidate sampling point information of the edge sampling region includes: K clusters are obtained by clustering multiple candidate sampling points in the edge sampling region; The sampling path of the edge sampling region is obtained by processing the K clusters based on the sampling determination model; A first motor control scheme is generated based on the sampling path and initial low moving speed of the edge sampling area; Based on the first motor control scheme, the motor drives the robot to perform sampling and acquires robot sampling video; A second motor control scheme is determined based on the robot sampling video and multiple candidate sampling point information for each prominent sampling area. This determination includes: Construct a sampling graph, which includes multiple candidate sampling point nodes and multiple edges between the candidate sampling point nodes. The node features of each candidate sampling point node are candidate sampling point information, and the edges between nodes are the distances between sampling points. The sampling path for each prominent sampling region is determined by processing the sampling map using a graph neural network. Based on the sampling path of each salient sampling area and the robot sampling video, a sampling video of the robot moving at different speeds along the sampling path of each salient sampling area is generated; The target moving speed of the robot in each salient sampling area is determined based on the sampling video of the robot's different moving speeds along the sampling path in each salient sampling area. A second motor control scheme is generated based on the target moving speed of the robot's sampling path in each salient sampling area; The second motor control scheme controls the motor to drive the robot to perform sampling.

2. The high-precision ultra-low speed motor control method as described in claim 1, characterized in that, The determination of the target movement speed of the robot's sampling path in each salient sampling area based on the sampling video of the robot's different movement speeds along the sampling path in each salient sampling area includes: Based on the sampling video of the robot at different moving speeds along its sampling path in each salient sampling area, the destruction information of the sampled samples at each moving speed is determined; Based on the sample destruction information at each movement speed, the target movement speed of the robot's sampling path in each salient sampling area is determined.

3. A high-precision ultra-low speed motor control system, characterized in that, include: The acquisition module is used to acquire images of the seabed sampling area; The sampling area determination module is used to determine an edge sampling area and multiple prominent sampling areas based on the image of the seabed sampling area. An edge sampling area is a region located at the boundary of the overall sampling area, determined by analyzing the image of the seabed sampling area using a region segmentation model. The edge sampling area includes location coordinates, boundary range, and topographic feature information. Multiple prominent sampling areas are multiple regions with significant features, determined by a region segmentation model. The prominent sampling areas include region location coordinates, boundary range, topographic features, biological species, and soil type information. The alternative sampling point determination module is used to determine multiple alternative sampling point information for the edge sampling region and multiple alternative sampling point information for each salient sampling region based on the one edge sampling region and the multiple salient sampling regions. The first control scheme generation module is configured to generate a first motor control scheme based on information from multiple candidate sampling points in the edge sampling region. The first control scheme generation module is further configured to: K clusters are obtained by clustering multiple candidate sampling points in the edge sampling region; The sampling path of the edge sampling region is obtained by processing the K clusters based on the sampling determination model; A first motor control scheme is generated based on the sampling path and initial low moving speed of the edge sampling area; The first sampling control module is used to control the motor to drive the robot to perform sampling based on the first motor control scheme and to acquire the robot sampling video. The second control scheme determination module is used to determine a second motor control scheme based on the robot sampling video and information on multiple candidate sampling points for each salient sampling region. The second control scheme determination module is also used for: Construct a sampling graph, which includes multiple candidate sampling point nodes and multiple edges between the candidate sampling point nodes. The node features of each candidate sampling point node are candidate sampling point information, and the edges between nodes are the distances between sampling points. The sampling path for each prominent sampling region is determined by processing the sampling map using a graph neural network. Based on the sampling path of each salient sampling area and the robot sampling video, a sampling video of the robot moving at different speeds along the sampling path of each salient sampling area is generated; The target moving speed of the robot in each salient sampling area is determined based on the sampling video of the robot's different moving speeds along the sampling path in each salient sampling area. A second motor control scheme is generated based on the target moving speed of the robot's sampling path in each salient sampling area; The second sampling control module is used to control the motor-driven robot to perform sampling based on the second motor control scheme.

4. The high-precision ultra-low speed motor control system as described in claim 3, characterized in that, The second control scheme determination module is also used for: Based on the sampling video of the robot at different moving speeds along its sampling path in each salient sampling area, the destruction information of the sampled samples at each moving speed is determined; Based on the sample destruction information at each movement speed, the target movement speed of the robot's sampling path in each salient sampling area is determined.

5. An electronic device, characterized in that, include: processor; Memory; And a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the high-precision ultra-low-speed motor control method as described in any one of claims 1 to 2.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the high-precision ultra-low speed motor control method as described in any one of claims 1 to 2.

Images