A driving method and system for a quadruped robot based on a neural network model

Abstract

The present invention provides a driving method and system for a quadruped robot based on a neural network model, which relates to the field of robot driving technology. The method includes determining multiple preliminary monitoring points based on a panoramic view of a toxic gas leakage site; determining multiple supplementary monitoring points based on laser radar information at consecutive time points of each preliminary monitoring point and a panoramic view of the toxic gas leakage site; generating a subsequent simulation video of toxic gas diffusion based on the laser radar information at consecutive time points of each preliminary monitoring point, the laser radar information at consecutive time points of each supplementary monitoring point, and a panoramic view of the toxic gas leakage site; determining multiple diffusion danger points and multiple diffusion safety points based on the subsequent simulation video of toxic gas diffusion and the panoramic view of the toxic gas leakage site; and determining a target safety rescue path based on the multiple diffusion safety points and the multiple diffusion danger points. The method can quickly and accurately determine the optimal rescue path for the toxic gas leakage site.

Description

Technical Field

[0001] The present invention relates to the field of robot driving technology, and in particular to a driving method and system for a quadruped robot based on a neural network model. Background Art

[0002] With the acceleration of industrialization and urbanization, toxic gas leaks, such as chemical plant leaks and underground pipeline ruptures, are becoming more frequent, posing a serious threat to human safety and environmental health. At the scene of a toxic gas leak, it is crucial to quickly and accurately assess the spread of the gas and plan safe rescue routes. Traditional toxic gas monitoring methods rely primarily on fixed sensors or manual monitoring, but suffer from slow response speeds, insufficient prediction accuracy, and reliance on manual experience for rescue route planning. Fixed sensors have a limited deployment range and struggle to cover the entire leak area. Manual monitoring is inefficient, unable to obtain dynamic diffusion data in real time, and unable to adapt to complex terrain and dynamically changing airflow conditions, resulting in large prediction errors. Furthermore, rescuers often plan routes based on limited information and may stray into dangerous areas, increasing rescue risks.

[0003] Therefore, how to quickly and accurately determine the optimal rescue path at the gas leakage site is an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem solved by the present invention is how to quickly and accurately determine the optimal rescue path at a toxic gas leakage site.

[0005] According to a first aspect, the present invention provides a driving method for a quadruped robot based on a neural network model, comprising: obtaining a panoramic view of a toxic gas leakage site; determining a plurality of preliminary monitoring points based on the panoramic view of the toxic gas leakage site; controlling the quadruped robot to each preliminary monitoring point, and obtaining laser radar information of continuous time points of each preliminary monitoring point; determining a plurality of supplementary monitoring points based on the laser radar information of continuous time points of each preliminary monitoring point and the panoramic view of the toxic gas leakage site; controlling the quadruped robot to each supplementary monitoring point, and obtaining laser radar information of continuous time points of each supplementary monitoring point; generating a subsequent simulation video of toxic gas diffusion based on the laser radar information of continuous time points of each preliminary monitoring point, the laser radar information of continuous time points of each supplementary monitoring point, and the panoramic view of the toxic gas leakage site; determining a plurality of diffusion danger points and a plurality of diffusion safety points based on the subsequent simulation video of toxic gas diffusion and the panoramic view of the toxic gas leakage site; and determining a target safety rescue path based on the plurality of diffusion safety points and the plurality of diffusion danger points.

[0006] In one possible implementation, the determining of a safe rescue path based on the multiple diffusion safety points and the multiple diffusion danger points includes: generating multiple preliminary safe rescue paths based on the multiple diffusion safety points and the multiple diffusion danger points; determining each preliminary safe rescue path information based on the subsequent simulation video of the toxic gas diffusion and the multiple preliminary safe rescue paths; constructing a graph structure, the graph structure including multiple nodes and multiple edges between the multiple nodes, each node representing a preliminary safe rescue path, the node attribute of each node being the preliminary safe rescue path information, and the edges between the nodes being the similarity between two pieces of preliminary safe rescue path information; processing the graph structure based on a graph neural network to determine a target safe rescue path.

[0007] In a possible implementation, determining a plurality of preliminary monitoring points based on the panoramic view of the toxic gas leakage site includes: determining a plurality of preliminary monitoring points based on the panoramic view of the toxic gas leakage site using a convolutional neural network.

[0008] In one possible implementation, the input of the graph neural network is the graph structure, and the output of the graph neural network is the target safe rescue path.

[0009] According to a second aspect, the present invention provides a driving system for a quadruped robot based on a neural network model, comprising: an acquisition module for acquiring a panoramic view of a toxic gas leak site; a preliminary monitoring point determination module for determining multiple preliminary monitoring points based on the panoramic view of the toxic gas leak site; a first control module for controlling the quadruped robot to each preliminary monitoring point and acquiring laser radar information at consecutive time points of each preliminary monitoring point; a supplementary monitoring point determination module for determining multiple supplementary monitoring points based on the laser radar information at consecutive time points of each preliminary monitoring point and the panoramic view of the toxic gas leak site; a second control module for controlling the quadruped robot to each supplementary monitoring point and acquiring laser radar information at consecutive time points of each supplementary monitoring point; a generation module for generating a subsequent simulation video of toxic gas diffusion based on the laser radar information at consecutive time points of each preliminary monitoring point, the laser radar information at consecutive time points of each supplementary monitoring point, and the panoramic view of the toxic gas leak site; a hazardous area determination module for determining multiple diffusion hazardous points and multiple diffusion safe points based on the subsequent simulation video of toxic gas diffusion and the panoramic view of the toxic gas leak site; and a path planning module for determining a target safe rescue path based on the multiple diffusion safe points and the multiple diffusion hazardous points.

[0010] In one possible implementation, the path planning module is also used to: generate multiple preliminary safe rescue paths based on the multiple diffusion safety points and the multiple diffusion danger points; determine each preliminary safe rescue path information based on the subsequent simulation video of the toxic gas diffusion and the multiple preliminary safe rescue paths; construct a graph structure, the graph structure includes multiple nodes and multiple edges between the multiple nodes, each node represents a preliminary safe rescue path, the node attribute of each node is the preliminary safe rescue path information, and the edge between the nodes is the similarity between the two preliminary safe rescue path information; process the graph structure based on a graph neural network to determine the target safe rescue path.

[0011] In a possible implementation, the preliminary monitoring point determination module is specifically configured to determine a plurality of preliminary monitoring points using a convolutional neural network based on a panoramic view of the toxic gas leakage site.

[0012] In one possible implementation, the input of the graph neural network is the graph structure, and the output of the graph neural network is the target safe rescue path.

[0013] The present invention provides a driving method and system for a quadruped robot based on a neural network model, the method comprising obtaining a panoramic view of a toxic gas leakage site; determining a plurality of preliminary monitoring points based on the panoramic view of the toxic gas leakage site; controlling the quadruped robot to each preliminary monitoring point and obtaining laser radar information at continuous time points of each preliminary monitoring point; determining a plurality of supplementary monitoring points based on the laser radar information at continuous time points of each preliminary monitoring point and the panoramic view of the toxic gas leakage site; controlling the quadruped robot to each supplementary monitoring point and obtaining laser radar information at continuous time points of each supplementary monitoring point; generating a subsequent simulation video of toxic gas diffusion based on the laser radar information at continuous time points of each preliminary monitoring point, the laser radar information at continuous time points of each supplementary monitoring point, and the panoramic view of the toxic gas leakage site; determining a plurality of diffusion danger points and a plurality of diffusion safety points based on the subsequent simulation video of toxic gas diffusion and the panoramic view of the toxic gas leakage site; and determining a target safety rescue path based on the plurality of diffusion safety points and the plurality of diffusion danger points. The method can quickly and accurately determine the optimal rescue path for the toxic gas leakage site. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic flow chart of a driving method for a quadruped robot based on a neural network model provided in an embodiment of the present invention;

[0015] Figure 2 A schematic diagram of a process for determining a target safe rescue path provided by an embodiment of the present invention;

[0016] Figure 3A schematic diagram of a driving system of a quadruped robot based on a neural network model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In an embodiment of the present invention, there is provided Figure 1 A driving method of a quadruped robot based on a neural network model is shown, and the driving method of a quadruped robot based on a neural network model includes steps S1 to S8:

[0018] Step S1: Obtain a panoramic view of the gas leakage site.

[0019] A toxic gas leak site is a physical area where toxic or hazardous gases have leaked, such as chemical plants, warehouses, and other industrial facilities, laboratories, or other enclosed or semi-enclosed environments. Examples of toxic gas leaks include the release of hazardous gases like ammonia and chlorine from ruptured gas storage tanks in chemical plants, and the spread of toxic gases like hydrogen sulfide and carbon monoxide from malfunctioning laboratory ventilation systems.

[0020] The panoramic image of the gas leak site is obtained through a multi-camera array image acquisition device, which is a high-resolution image covering the complete visual information of the gas leak site area.

[0021] Step S2: determining a plurality of preliminary monitoring points based on a panoramic view of the toxic gas leakage site.

[0022] In some embodiments, a convolutional neural network can be used to determine multiple preliminary monitoring points based on a panoramic view of the gas leakage site, where the input of the convolutional neural network is a panoramic view of the gas leakage site, and the output of the convolutional neural network is multiple preliminary monitoring points.

[0023] Convolutional neural networks (CNNs) are deep learning models that excel at processing grid-structured data, such as videos and images. Their core concept is to automatically extract spatial hierarchical features from input data through local receptive fields and weight sharing. Convolutional neural networks consist of convolutional layers, pooling layers, and fully connected layers. Convolutional layers capture local features using sliding filters (convolution kernels), pooling layers reduce data dimensionality and enhance translation invariance, and fully connected layers ultimately complete classification or regression tasks.

[0024] The preliminary monitoring points are key locations in the gas leakage site output by the convolutional neural network. The preliminary monitoring points can be used to deploy quadruped robots to collect environmental data.

[0025] A panoramic image of a toxic gas leak site contains a wealth of environmental information, such as the location of the leak source, the distribution of obstacles, the direction of ventilation openings, and the visual characteristics of toxic gas diffusion. Analysis of the panoramic image can identify key areas near the leak source, critical nodes in the ventilation system, and the distribution of obstacles that may affect the gas diffusion path. These locations are highly valuable for monitoring. For example, the area around the leak source can reflect the initial diffusion state, the area near the ventilation openings can capture the direction of airflow, and open areas can help assess the overall diffusion trend. Selecting initial monitoring points from these key locations ensures that subsequent data collected is representative.

[0026] Convolutional neural networks can extract multi-layered spatial features from panoramic images of gas leak sites. For example, shallow convolution kernels can identify basic geometric features such as obstacle edges and vent shapes, while deeper networks can incorporate contextual information to distinguish semantic units such as open spaces, narrow passages, and enclosed areas. By performing semantic segmentation on the panoramic images, convolutional neural networks can locate key spatial elements such as the physical location of the leak source, the spatial structure of obstacle distribution, and the orientation of vents. These spatial features are highly correlated with potential gas diffusion pathways. For example, the orientation of vents determines the initial diffusion direction of gas driven by airflow, and the distribution of obstacles can form vortices that trap gas. By learning the correlation between spatial layout and diffusion patterns in historical scenarios, convolutional neural networks can automatically identify locations critical for gas diffusion monitoring, such as directly below the leak source, upstream and downstream of vents, and between obstacles. This allows them to prioritize initial monitoring points for deploying quadruped robotic monitoring equipment.

[0027] Step S3: Control the quadruped robot to each preliminary monitoring point and obtain the lidar information of each preliminary monitoring point at consecutive time points.

[0028] LiDAR information is acquired through the following method: When the quadruped robot arrives at a pre-set monitoring point, its onboard gas monitoring LiDAR emits a laser beam of a specific wavelength, then receives the reflected signal and analyzes the three-dimensional spatial distribution of the toxic gas. This LiDAR information includes information such as the gas concentration gradient and the geometric characteristics of the toxic gas diffusion boundary.

[0029] LiDAR data collected at continuous time points by the quadruped robot, which maintains a stable posture and continuously collects data at a fixed sampling rate (e.g., 5 Hz), forms a time series monitoring record. This continuous LiDAR data can capture the dynamic changes in toxic gas concentration and record the temporal evolution of its diffusion direction.

[0030] Step S4: determining a plurality of supplementary monitoring points based on the lidar information of each preliminary monitoring point at consecutive time points and a panoramic view of the toxic gas leakage site.

[0031] In some embodiments, a supplementary point analysis model can be used to determine multiple supplementary monitoring points based on the lidar information of each preliminary monitoring point at consecutive time points and a panoramic view of the gas leakage site. The supplementary point analysis model is a convolutional neural network. The input of the supplementary point analysis model is the lidar information of each preliminary monitoring point at consecutive time points and a panoramic view of the gas leakage site. The output of the supplementary point analysis model is multiple supplementary monitoring points.

[0032] Supplementary monitoring points are key locations optimized and identified based on the initial monitoring points using the supplementary point analysis model, targeting coverage gaps, data-poor areas, or high-risk mutation areas within the monitoring network. These include, but are not limited to, eddy currents behind obstacles, branch nodes in ventilation systems, and areas where the rate of change in concentration gradients exceeds a preset threshold. For example, if the diffusion rate in a particular direction is significantly higher than expected, additional monitoring points are added along the path in that direction to capture mutation trends in real time.

[0033] Convolutional neural networks possess unique capabilities for analyzing spatiotemporal features and recognizing dynamic patterns. Through their three-dimensional convolutional structure, convolutional neural networks can simultaneously process static panoramic spatial features and dynamic time-series lidar information, thereby extracting information such as sudden changes in diffusion direction and concentration gradients during the gas diffusion process. Furthermore, the convolutional neural network's deep semantic understanding enables it to identify coverage gaps or potential high-risk areas in the initial monitoring point distribution. This is then combined with the effects of the ventilation system on airflow to generate multiple supplementary monitoring points.

[0034] In some embodiments, the supplementary point analysis model includes a diffusion analysis layer, an anomaly monitoring layer, and a point planning layer. The diffusion analysis layer, the anomaly monitoring layer, and the point planning layer all include a convolutional neural network structure. The input of the diffusion analysis layer is the lidar information of each of the preliminary monitoring points at consecutive time points. The output of the diffusion analysis layer is the time series concentration field sequence of the toxic gas and the regional diffusion trend of the toxic gas. The input of the anomaly monitoring layer is the time series concentration field sequence of the toxic gas, the regional diffusion trend of the toxic gas, and a panoramic view of the toxic gas leakage site. The output of the anomaly monitoring layer is the abnormally high concentration area mark and the diffusion path mutation point. The input of the point planning layer is the abnormally high concentration area mark, the diffusion path mutation point, and a panoramic view of the toxic gas leakage site. The output of the point planning layer is multiple supplementary monitoring points.

[0035] The toxic gas concentration field sequence is a sequence of toxic gas concentration distribution images indexed by time. Each frame of the image corresponds to a moment and can fully display the toxic gas concentration distribution in the entire toxic gas leakage site area and its changes over time. The concentration values ​​are visualized through color gradients and numerical annotations.

[0036] The regional diffusion trend of toxic gas includes the diffusion direction and speed distribution of toxic gas in the area of ​​toxic gas leakage site.

[0037] Abnormally high concentration area markers are spatial areas where the concentration of toxic gas exceeds the safety threshold and the duration reaches or exceeds the set standard.

[0038] The mutation point of the diffusion path is the spatial location where the diffusion direction or speed of the poison gas changes significantly in a short period of time.

[0039] Different layers can be responsible for processing information in different dimensions. For example, the diffusion analysis layer focuses on analyzing the concentration distribution and time-series diffusion characteristics in lidar data. The anomaly monitoring layer uses panoramic images to locate high-risk areas and diffusion mutation points. The point planning layer can generate specific supplementary monitoring point plans based on the results of the first two. This layered processing allows each module to perform its own duties, avoiding the inefficiencies caused by complex data mixing. This not only improves the accuracy of analyzing toxic gas diffusion characteristics, but also enables the precise deployment of supplementary monitoring points, thereby enhancing the integrity and effectiveness of monitoring.

[0040] Step S5: Control the quadruped robot to each supplementary monitoring point and obtain the laser radar information of each supplementary monitoring point at consecutive time points.

[0041] For step S5, please refer to step S3, which will not be repeated here.

[0042] Step S6, generating a subsequent simulation video of poison gas diffusion based on the laser radar information of each preliminary monitoring point at consecutive time points, the laser radar information of each supplementary monitoring point at consecutive time points, and the panoramic view of the poison gas leakage site.

[0043] In some embodiments, a simulation generation model can be used to generate a subsequent simulation video of the toxic gas diffusion based on the lidar information of each preliminary monitoring point at continuous time points, the lidar information of each supplementary monitoring point at continuous time points, and a panoramic view of the toxic gas leakage site. The simulation generation model is a generative adversarial network. The input of the simulation generation model is the lidar information of each preliminary monitoring point at continuous time points, the lidar information of each supplementary monitoring point at continuous time points, and a panoramic view of the toxic gas leakage site. The output of the simulation generation model is a subsequent simulation video of the toxic gas diffusion.

[0044] A Generative Adversarial Network (GAN) is a deep learning model composed of a generator and a discriminator. The generator learns the distribution patterns of input data and generates realistic new data, while the discriminator attempts to distinguish generated data from real data. Both are continuously optimized through adversarial training. GANs can generate high-fidelity and diverse outputs and are suitable for modeling complex dynamic processes such as the spatiotemporal evolution of toxic gas diffusion.

[0045] The subsequent simulation video of poison gas diffusion is a dynamic simulation video generated by a simulation generation model to simulate the diffusion path, concentration changes and evolution of dangerous areas of poison gas in the future.

[0046] The generator of a generative adversarial network extracts features from the input data distribution and synthesizes new samples based on these features. First, the continuous time-point LiDAR information from the initial and supplementary monitoring points is fused with a panoramic image of the gas leak site and input into the generator. The generator learns the gas diffusion pattern through a spatiotemporal attention mechanism and incorporates the Fick diffusion equation constraint "∇²C=∂C / ∂t" and a mass conservation regularization term into its loss function to ensure that the concentration field evolution conforms to physical laws. The discriminator compares the statistical distribution differences between the generated video frames and the real monitoring data, including the concentration field distribution in a single frame and the gradient of change between adjacent frames. This adversarial game improves the generator's ability to predict diffusion behavior in unmonitored areas, ensuring that the model ultimately generates physically accurate subsequent simulation videos of gas diffusion.

[0047] Step S7: determining a plurality of diffusion danger points and a plurality of diffusion safety points based on the subsequent simulation video of the poisonous gas diffusion and the panoramic view of the poisonous gas leakage site.

[0048] In some embodiments, a gated loop unit can be used to determine multiple diffusion danger points and multiple diffusion safety points based on the subsequent simulation video of the toxic gas diffusion and the panoramic view of the toxic gas leakage site. The input of the gated loop unit is the subsequent simulation video of the toxic gas diffusion and the panoramic view of the toxic gas leakage site, and the output of the gated loop unit is multiple diffusion danger points and multiple diffusion safety points.

[0049] The Gated Recurrent Unit (GRU) is a variant of a recurrent neural network that can process sequential data and temporal information and excels at capturing dependencies in long sequences. A GRU consists of three components: a memory cell, an update gate, and a reset gate.

[0050] Multiple diffusion hazard points are spatial locations where the concentration of toxic gas output by the gated circulation unit exceeds the safety threshold and lasts for a long time, such as the center of a high-concentration air mass, the low wind speed retention area on the diffusion path, etc.

[0051] The diffusion safety point is a location where the concentration of toxic gas output by the gated circulation unit is always lower than the safety threshold and has a certain degree of environmental safety, such as a high place against the wind, a closed and airtight area, or the upstream area of ​​the ventilation system.

[0052] The Gated Recurrent Unit (GRU) is capable of analyzing temporal dynamics and integrating spatial semantics. The GRU uses a gating mechanism to capture the temporal dependencies of concentration field changes in subsequent simulation videos of toxic gas diffusion, thereby identifying areas where concentrations consistently exceed limits. Panoramic images can extract spatial features such as obstacles and vents. The GRU combines dynamic temporal features with static spatial features to analyze the environmental impact of toxic gas diffusion, thereby locating high-risk and low-risk areas and ultimately identifying multiple diffusion risk points and safe diffusion points.

[0053] Step S8: determining a target safety rescue path based on the multiple diffusion safety points and the multiple diffusion danger points.

[0054] In some embodiments, the Figure 2 The process is used to determine the target safe rescue path, Figure 2 A schematic diagram of a process for determining a target safe rescue path provided by an embodiment of the present invention. The process for determining a target safe rescue path includes steps S21 to S24:

[0055] Step S21: generating a plurality of preliminary safety rescue paths based on the plurality of diffusion safety points and the plurality of diffusion danger points.

[0056] In some embodiments, a plurality of preliminary safety rescue paths may be generated based on the plurality of diffuse safety points and the plurality of diffuse danger points using a variational autoencoder. The variational autoencoder inputs the plurality of diffuse safety points and the plurality of diffuse danger points, and outputs the plurality of preliminary safety rescue paths.

[0057] The Variational Autoencoder (VAE) is a generative deep learning model. The VAE compresses the input data into probability distribution parameters in a low-dimensional latent space through the encoder. The decoder can generate new data by sampling from this distribution. Its core is to use variational inference to balance the reconstruction accuracy and the regularization of the latent distribution to achieve data generation and feature decoupling.

[0058] The initial safe rescue path is a candidate route generated by a variational autoencoder based on safe and dangerous points. It consists of a series of continuous coordinate points. The initial safe rescue path satisfies connectivity from the starting point to multiple diffuse safe points in the middle and then to the end point, while avoiding diffuse dangerous points along the way.

[0059] The variational autoencoder's encoder encodes data such as safe point coordinates and dangerous point masks into low-dimensional latent vectors. These latent vectors implicitly represent path characteristics, such as detour propensity. The decoder then samples the latent space to generate a diverse sequence of paths that meet safety rules. The probabilistic generation mechanism of the variational autoencoder effectively addresses the uncertainty in rescue scenarios, outputting multiple preliminary safe rescue paths covering different risk preferences.

[0060] Step S22: determining information of each preliminary safe rescue path based on the subsequent simulation video of poison gas diffusion and the multiple preliminary safe rescue paths.

[0061] In some embodiments, each preliminary safe rescue path information can be determined based on the subsequent simulation video of the toxic gas diffusion and the multiple preliminary safe rescue paths through a path analysis model. The path analysis model is a Transformer model. The input of the path analysis model is the subsequent simulation video of the toxic gas diffusion and the multiple preliminary safe rescue paths. The output of the path analysis model is each preliminary safe rescue path information.

[0062] The Transformer model consists of two parts: an encoder and a decoder. In the Transformer model, the encoder is responsible for performing representation learning on the input sequence, integrating a self-attention mechanism and a feed-forward neural network. Unlike the encoder, the decoder inherits the encoder architecture and incorporates a multi-head attention mechanism, specifically designed to parse the encoder output and generate the corresponding target sequence. This architectural design enables the Transformer model to accurately identify each preliminary safe rescue path.

[0063] Each preliminary safe rescue path information is a quantitative description of the risk and feasibility of each preliminary safe rescue path in a dynamic toxic gas environment output by the path analysis model. The preliminary safe rescue path information includes risk indicators, time and space indicators, environmental adaptability indicators, etc.

[0064] The risk index is used to measure the risk of poisoning for rescuers along their route. Factors influencing the risk index include the cumulative exposure concentration and the proportion of high-risk periods.

[0065] The cumulative exposure concentration is the integral value of the concentration at each point along the entire route over time. For example, if the cumulative exposure along a route reaches 500 ppm·min and far exceeds the safety threshold, it means that the poisoning risk along this route is high.

[0066] The proportion of high-risk periods reflects the proportion of time that the route passes through areas with excessive toxic gases. The higher the proportion, the longer the route is exposed to the dangerous environment during operation, and the greater the risk.

[0067] Time and space indicators include estimated travel time and path length to reflect travel efficiency.

[0068] Environmental adaptability indicators include the number of alternative safety points along the path and the risk fluctuation range of the path under different wind speed and wind direction scenarios. They are used to evaluate the reliability of the path in responding to dynamic environmental changes.

[0069] The Transformer model can process subsequent simulation videos of toxic gas diffusion that contain time series information. Leveraging a self-attention mechanism, it effectively captures the correlation between video frames at different time points, as well as the spatial correspondence between preliminary safe rescue paths and risk areas in the video. The Transformer model comprehensively considers the impact of toxic gas diffusion conditions on the paths at each moment, then calculates indicators such as the risk level and feasibility of each path at different times, ultimately determining the initial safe rescue path information for each path.

[0070] The Transformer model's encoder uses a self-attention mechanism to extract features from the input video of the subsequent simulation of the gas diffusion and the initial safe rescue path. This allows it to capture spatiotemporal relationships such as gas concentration changes and diffusion direction at different moments in the video, as well as the spatial connections between points along the path. A feedforward neural network further transforms and enhances these features. Based on the encoder output, the decoder uses a multi-head attention mechanism to focus on key information, deeply integrating video features with path features to ultimately generate path information.

[0071] Step S23, constructing a graph structure, which includes multiple nodes and multiple edges between the multiple nodes. Each node represents a preliminary safe rescue path, the node attribute of each node is the preliminary safe rescue path information, and the edges between the nodes are the similarities between the two preliminary safe rescue path information.

[0072] The graph structure consists of multiple nodes and multiple edges connecting these nodes. It can abstractly express complex relationships in the form of a topological network. Each node and edge carries specific attributes and meanings and can be used to describe objects and their relationships.

[0073] A node corresponds to each preliminary safe rescue path, and its node attributes are the detailed information of the path.

[0074] The edge between nodes indicates the degree of association between two paths, and the edge is the similarity between the information of two preliminary safety rescue paths to reflect the similarity or difference of the paths in terms of risk characteristics, spatial direction, etc.

[0075] In some embodiments, the similarity between two pieces of preliminary safe rescue path information can be calculated using a deep neural network model.

[0076] Step S24: Process the graph structure based on a graph neural network to determine a target safe rescue path.

[0077] Graph neural networks (GNNs) are a type of deep learning model designed for graph-structured data. Their core principle is to enable nodes to exchange information through message passing and gradually update their feature representations. In this way, GNNs can effectively capture high-order relationships between nodes in a graph structure, thereby uncovering global features and underlying patterns in the data. The input of a GNN is the graph structure, and its output is a target safe rescue path.

[0078] The target safe rescue path is an optimal rescue route generated by processing the graph structure through a graph neural network. This path is composed of continuous spatial coordinate points and must comprehensively meet the requirements of safety, efficiency, and environmental adaptability. Specifically, the target safe rescue path must avoid dangerous areas where toxic gas concentrations exceed the standard throughout the entire process to control the cumulative exposure concentration of personnel within the safe threshold, while also shortening the path length or travel time as much as possible to improve rescue efficiency. In addition, the target safe rescue path must be robust to the dynamic changes in toxic gas diffusion, such as having low risk fluctuations under different meteorological conditions and being close to alternative safe points to ensure emergency evacuation capabilities.

[0079] The information of each preliminary safe rescue path covers multi-dimensional attributes such as risk, time, and space. The node attribute design of the graph structure can integrate this complex information, providing rich contextual data for the graph neural network, enabling the model to comprehensively consider various factors for path evaluation.

[0080] The graph structure intuitively presents the associations between preliminary safety rescue paths in the form of edges, making it easier for graph neural networks to capture potential connections between paths.

[0081] The path information attributes of each node form the core basis for the graph neural network's decision-making. For example, the cumulative exposure concentration is used to assess the health threat posed by the path to rescuers; higher values ​​indicate greater risk. The estimated travel time directly affects rescue efficiency, with shorter travel time paths being more advantageous in emergency scenarios. The environmental adaptability index reflects the path's stability in a dynamic toxic gas diffusion environment; higher indexes indicate a path is less susceptible to environmental changes. This data helps the graph neural network accurately calculate the overall score for each path, thereby selecting the most suitable target safe rescue path.

[0082] Based on the same inventive concept, Figure 3 A schematic diagram of a driving system of a quadruped robot based on a neural network model provided in an embodiment of the present invention, wherein the driving system of the quadruped robot based on the neural network model includes:

[0083] An acquisition module 31 is used to obtain a panoramic image of the toxic gas leakage site;

[0084] A preliminary monitoring point determination module 32 is configured to determine a plurality of preliminary monitoring points based on a panoramic view of the toxic gas leakage site;

[0085] The first control module 33 is used to control the quadruped robot to each preliminary monitoring point and obtain the lidar information of each preliminary monitoring point at consecutive time points;

[0086] A supplementary monitoring point determination module 34 is configured to determine a plurality of supplementary monitoring points based on the lidar information of each preliminary monitoring point at consecutive time points and a panoramic view of the gas leakage site;

[0087] The second control module 35 is used to control the quadruped robot to each supplementary monitoring point and obtain the laser radar information of each supplementary monitoring point at consecutive time points;

[0088] A generating module 36 is configured to generate a subsequent simulation video of poison gas diffusion based on the laser radar information of each of the preliminary monitoring points at consecutive time points, the laser radar information of each of the supplementary monitoring points at consecutive time points, and the panoramic image of the poison gas leakage site;

[0089] a danger zone determination module 37 for determining a plurality of diffusion danger points and a plurality of diffusion safety points based on the subsequent simulation video of the poisonous gas diffusion and the panoramic view of the poisonous gas leakage site;

[0090] The path planning module 38 is configured to determine a target safety rescue path based on the multiple diffusion safety points and the multiple diffusion danger points.

[0091] Similarly, it should be noted that, in order to simplify the presentation of this specification and thus facilitate understanding of one or more embodiments of the invention, the foregoing descriptions of the embodiments of this specification sometimes combine multiple features into a single embodiment, figure, or description thereof. However, this disclosure method does not imply that the subject matter of this specification requires more features than those recited in the claims. In fact, an embodiment may have fewer features than all of the features of a single disclosed embodiment.

[0092] Finally, it should be understood that the embodiments described in this specification are intended only to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, by way of example and not limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the implementations explicitly described and illustrated in this specification.

Claims

1. A driving method for a quadruped robot based on a neural network model, characterized in that: include: Obtain a panoramic view of the gas leak site; Determining a plurality of preliminary monitoring points based on a panoramic view of the toxic gas leakage site, wherein determining a plurality of preliminary monitoring points based on the panoramic view of the toxic gas leakage site comprises: A convolutional neural network is used to determine multiple preliminary monitoring points based on a panoramic image of the gas leakage site; Control the quadruped robot to each preliminary monitoring point and obtain the lidar information of each preliminary monitoring point at consecutive time points. The lidar information includes the gas concentration gradient distribution and the geometric characteristics of the gas diffusion boundary; Determine multiple supplementary monitoring points based on the lidar information of each preliminary monitoring point at consecutive time points and a panoramic view of the toxic gas leakage site; Control the quadruped robot to each supplementary monitoring point and obtain the lidar information of each supplementary monitoring point at consecutive time points; generating a subsequent simulation video of the poisonous gas diffusion based on the laser radar information at consecutive time points of each preliminary monitoring point, the laser radar information at consecutive time points of each supplementary monitoring point, and the panoramic view of the poisonous gas leakage site; Determining multiple diffusion danger points and multiple diffusion safety points based on the subsequent simulation video of the poisonous gas diffusion and the panoramic view of the poisonous gas leakage site; A target safety rescue path is determined based on the multiple diffusion safety points and the multiple diffusion danger points.

2. The driving method of a quadruped robot based on a neural network model according to claim 1, characterized in that: The determining of a safe rescue path based on the multiple diffusion safety points and the multiple diffusion danger points includes: generating a plurality of preliminary safety rescue paths based on the plurality of diffusion safety points and the plurality of diffusion danger points; Determining each preliminary safe rescue path information based on the subsequent simulation video of the poisonous gas diffusion and the multiple preliminary safe rescue paths; Constructing a graph structure, which includes multiple nodes and multiple edges between the multiple nodes. Each node represents a preliminary safe rescue path. The node attribute of each node is the preliminary safe rescue path information. The edges between the nodes are the similarities between the two preliminary safe rescue path information. The graph structure is processed based on a graph neural network to determine a target safe rescue path.

3. The driving method of a quadruped robot based on a neural network model as claimed in claim 2, characterized in that: The input of the graph neural network is the graph structure, and the output of the graph neural network is the target safe rescue path.

4. A driving system for a quadruped robot based on a neural network model, characterized in that: include: An acquisition module is used to obtain a panoramic view of the gas leakage site; A preliminary monitoring point determination module is used to determine multiple preliminary monitoring points based on the panoramic view of the toxic gas leakage site. The preliminary monitoring point determination module is specifically used to: A convolutional neural network is used to determine multiple preliminary monitoring points based on a panoramic image of the gas leakage site; The first control module is used to control the quadruped robot to each preliminary monitoring point and obtain the lidar information of each preliminary monitoring point at consecutive time points; a supplementary monitoring point determination module, configured to determine a plurality of supplementary monitoring points based on the lidar information of the consecutive time points of each preliminary monitoring point and a panoramic view of the toxic gas leakage site; The second control module is used to control the quadruped robot to each supplementary monitoring point and obtain the laser radar information of each supplementary monitoring point at consecutive time points. The laser radar information includes the gas concentration gradient distribution and the geometric characteristics of the gas diffusion boundary; a generation module for generating a subsequent simulation video of poison gas diffusion based on the laser radar information of each of the preliminary monitoring points at consecutive time points, the laser radar information of each of the supplementary monitoring points at consecutive time points, and the panoramic image of the poison gas leakage site; a hazardous area determination module, configured to determine a plurality of diffusion hazardous points and a plurality of diffusion safe points based on the subsequent simulation video of the poisonous gas diffusion and a panoramic view of the poisonous gas leakage site; A path planning module is used to determine a target safety rescue path based on the multiple diffusion safety points and the multiple diffusion danger points.

5. The driving system of the quadruped robot based on the neural network model as claimed in claim 4, characterized in that: The path planning module is also used to: generating a plurality of preliminary safety rescue paths based on the plurality of diffusion safety points and the plurality of diffusion danger points; Determining each preliminary safe rescue path information based on the subsequent simulation video of the poisonous gas diffusion and the multiple preliminary safe rescue paths; Constructing a graph structure, which includes multiple nodes and multiple edges between the multiple nodes. Each node represents a preliminary safe rescue path. The node attribute of each node is the preliminary safe rescue path information. The edges between the nodes are the similarities between the two preliminary safe rescue path information. The graph structure is processed based on a graph neural network to determine a target safe rescue path.

6. The driving system of the quadruped robot based on the neural network model according to claim 5, characterized in that: The input of the graph neural network is the graph structure, and the output of the graph neural network is the target safe rescue path.

Images