Inspection robot navigation path optimization method and system based on environment modeling

Abstract

The invention relates to the technical field of robots, in particular to an inspection robot navigation path optimization method and system based on environmental modeling, and the method comprises the steps: obtaining a path optimization instruction containing scene attribute information and real-time perception data; performing dynamic environment feature prediction based on the scene attribute information to generate a multi-scale environment feature flow; performing space-time registration optimization on the multi-source environment modeling database in combination with the real-time sensing data, and establishing a registration environment model library; a multi-path planning learner and a loss analyzer are introduced, a dual-channel navigation path is constructed through loss optimization learning, and the dual-channel navigation path comprises an instant obstacle avoidance path planning channel and a prospective topology path planning channel; and inputting the real-time sensing data and the multi-scale environment characteristic flow into double channels to generate a dynamic navigation path optimization report. According to the method, precise environment modeling and multi-scale feature fusion are realized, and the adaptability, the real-time performance and the global optimization capability of path planning are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robots, in particular to a navigation path optimization method and system for a patrol robot based on environment modeling. BACKGROUND

[0002] With the rapid development of intelligent robot technology, patrol robots are increasingly widely used in complex industrial scenarios such as power, chemical industry, rail transit, and intelligent manufacturing. Traditional patrol robots mostly use preset paths or simple obstacle avoidance strategies based on local perception for navigation, which is difficult to cope with dynamic changes and multiple interference factors coexisting in actual patrol environments, resulting in low path planning efficiency, poor real-time performance, and insufficient adaptability. Especially in scenarios with changing light, frequently moving obstacles, complex structures, or multiple task requirements, traditional methods are prone to path oscillation, local optimal trap, or task interruption.

[0003] In recent years, navigation methods based on environment modeling have gradually become a research hotspot. By constructing a structured or semantic environment model to assist path planning, the understanding ability of robots for the scene is improved. However, existing technologies mostly rely on static maps or low-frequency updated environment information, lack effective prediction of dynamic evolution characteristics of the scene, and there are problems such as time and space asynchronization and coordinate inconsistency between multi-source heterogeneous environment data, which limits the modeling accuracy. In addition, the path planning process often separates obstacle avoidance and global optimization, lacks a collaborative mechanism that takes into account immediate response and long-term topological optimization, and is difficult to achieve efficient, safe, and intelligent dynamic path adjustment.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a navigation path optimization method and system for a patrol robot based on environment modeling, aiming to solve the technical problems that existing patrol robot navigation methods are difficult to effectively integrate dynamic environment perception and multi-source environment modeling, lack a collaborative path planning mechanism that takes into account immediate obstacle avoidance and global topological optimization, resulting in poor path adaptability and low planning efficiency in complex and variable scenarios.

[0006] To achieve the above purpose, the present application provides a navigation path optimization method for a patrol robot based on environment modeling, which comprises: obtaining a navigation path optimization instruction for a patrol robot, wherein the navigation path optimization instruction comprises scene attribute information corresponding to a target patrol scene and real-time perception data; performing dynamic environment feature prediction on the target patrol scene according to the scene attribute information to obtain a multi-scale environment feature flow; According to the scene attribute information and the real-time perception data, a multi-source environment modeling database is spatio-temporally registered and optimized to determine a registered environment model library; A plurality of path planning learners and path optimization loss analyzers are introduced to perform path optimization loss optimization learning on the registered environment model library to build a navigation path double channel of the inspection robot, wherein the navigation path double channel includes an instant obstacle avoidance path planning channel and a forward-looking topological path planning channel. The real-time perception data and the multi-scale environment feature flow are input into the navigation path double channel to generate a dynamic navigation path optimization report of the inspection robot.

[0007] Optionally, the dynamic environment feature prediction of the target inspection scene according to the scene attribute information includes: According to the scene attribute information, the target inspection scene is predicted to determine a key inspection area set; The structure semantic information, dynamic obstacle density information and light time-varying information of the key inspection area set are collected to obtain a target environment feature library; According to the target environment feature library, spatial scale classification is performed to obtain a multi-scale environment initial classification result; According to the multi-scale environment initial classification result, dynamic classification is performed to obtain the multi-scale environment feature flow.

[0008] Optionally, the spatio-temporal registration optimization of the multi-source environment modeling database according to the scene attribute information and the real-time perception data to determine the registered environment model library includes: According to the scene attribute information, the multi-source environment modeling database is scene type associated and filtered to obtain an isomorphic scene environment model library; According to the real-time perception data, the time and space twin degree of each model in the isomorphic scene environment model library is evaluated to obtain a plurality of model registration coefficients; According to the model registration threshold, the plurality of model registration coefficients are optimized and selected to obtain a model registration optimization distribution greater than or equal to the model registration threshold; According to the model registration optimization distribution, the isomorphic scene environment model library is dynamically cropped and fused to generate the registered environment model library.

[0009] Optionally, the introduction of a plurality of path planning learners and path optimization loss analyzers to perform path optimization loss optimization learning on the registered environment model library to build a navigation path double channel of the inspection robot includes: According to the registered environment model library, historical path log retrieval is performed to obtain an instant obstacle avoidance path log library and a forward-looking topological path log library; According to the instant obstacle avoidance path log library, environmental disturbance feature collection is performed to obtain a disturbance feature first record library; According to the prospective topology path log library, long-term task constraint feature collection is performed to obtain a constraint feature second record library; According to the plurality of path planning learners and the path optimization loss resolver, instant obstacle avoidance path optimization loss optimization learning is performed on the registered environment model library, the disturbance feature first record library and the instant obstacle avoidance path log library to generate the instant obstacle avoidance path planning channel; According to the plurality of path planning learners and the path optimization loss resolver, prospective topology path optimization loss optimization learning is performed on the registered environment model library, the constraint feature second record library and the prospective topology path log library to generate the prospective topology path planning channel; The instant obstacle avoidance path planning channel and the prospective topology path planning channel are connected as parallel nodes to generate the double-channel navigation path of the inspection robot.

[0010] Optionally, the instant obstacle avoidance path optimization loss optimization learning performed by the plurality of path planning learners and the path optimization loss resolver on the registered environment model library, the disturbance feature first record library and the instant obstacle avoidance path log library to generate the instant obstacle avoidance path planning channel comprises: The registered environment model library and the disturbance feature first record library are taken as input information, the instant obstacle avoidance path log library is taken as output label, the plurality of path planning learners are respectively supervised and trained, and a plurality of instant path optimization loss coefficients are calculated according to the path optimization loss resolver every predetermined iteration times; If the plurality of instant path optimization loss coefficients are less than an instant path optimization loss threshold, a plurality of instant obstacle avoidance path planning models are generated; The plurality of instant obstacle avoidance path planning models are connected to obtain the instant obstacle avoidance path planning channel.

[0011] Optionally, the path optimization loss resolver comprises a path optimization loss resolution function, and the path optimization loss resolution function is: wherein, represents a path optimization loss coefficient, N represents a predetermined iteration number, n represents the n th iteration, N and n are positive integers, and 1≤n≤N ; represents the number of path collision events in the n th iteration; represents the number of path acceleration overrun segments in the n th iteration; characterize the average deviation distance of the path from the preset topological skeleton line in the nth iteration; λ and μ are non-negative weight coefficients for balancing safety, smoothness and structural constraints.

[0012] Optionally, the real-time perception data and the multi-scale environment feature flow are input into the navigation path double channel to generate an inspection robot dynamic navigation path optimization report, which comprises: The real-time perception data and the multi-scale environment feature flow are input into the instant obstacle avoidance path planning channel to obtain a multi-scale environment-instant obstacle avoidance path set. The real-time perception data and the multi-scale environment feature flow are input into the instant obstacle avoidance path planning channel to obtain a multi-scale environment-instant obstacle avoidance path set. According to the multi-scale environment-instant obstacle avoidance path set and the multi-scale environment-anticipatory topological path set, spatiotemporal consistency fusion and Pareto front screening are performed to output the inspection robot dynamic navigation path optimization report.

[0013] In addition, in order to achieve the above-mentioned purpose, the application also provides an inspection robot navigation path optimization system based on environment modeling, which comprises: An instruction acquisition module is configured to obtain an inspection robot navigation path optimization instruction, wherein the navigation path optimization instruction comprises scene attribute information and real-time perception data corresponding to a target inspection scene. An environment prediction module is configured to perform dynamic environment feature prediction on the target inspection scene according to the scene attribute information to obtain a multi-scale environment feature flow. A registration optimization module is configured to perform spatiotemporal registration optimization on a multi-source environment modeling database according to the scene attribute information and the real-time perception data to determine a registered environment model library. A channel building module is configured to introduce a plurality of path planning learners and path optimization loss analyzers to perform path optimization loss optimization learning on the registered environment model library to build an inspection robot navigation path double channel, wherein the navigation path double channel comprises an instant obstacle avoidance path planning channel and an anticipatory topological path planning channel. A report generation module is configured to input the real-time perception data and the multi-scale environment feature flow into the navigation path double channel to generate an inspection robot dynamic navigation path optimization report.

[0014] Further, in order to achieve the above object, the present application also provides an environment modeling based inspection robot navigation path optimization device, comprising a memory, a processor and an environment modeling based inspection robot navigation path optimization program stored in the memory and executable on the processor, the environment modeling based inspection robot navigation path optimization program being configured to implement the steps of the environment modeling based inspection robot navigation path optimization method according to any one of the above.

[0015] Further, in order to achieve the above object, the present application also provides a medium having an environment modeling based inspection robot navigation path optimization program stored thereon, the environment modeling based inspection robot navigation path optimization program being executable by a processor to implement the steps of the environment modeling based inspection robot navigation path optimization method according to any one of the above.

[0016] The present application provides an environment modeling based inspection robot navigation path optimization method, which realizes multi-scale dynamic environment feature prediction of a target inspection scene and high-precision space-time registration of a multi-source environment model by fusing scene attribute information and real-time perception data, and constructs a registered environment model library with environment adaptability and modeling accuracy; on this basis, a multi-learner collaboration and loss analysis mechanism is introduced to build a dual-channel navigation path optimization architecture including instant obstacle avoidance and prospective topology planning, realizing the organic unification of local dynamic obstacle avoidance and globally optimal path; finally, a dynamic navigation path optimization report is generated, which significantly improves the real-time performance, safety and task execution efficiency of the inspection robot in a complex and dynamic environment, and enhances the intelligent level and robustness of the system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of an embodiment of the environment modeling based inspection robot navigation path optimization method of the present application.

[0018] The implementation of the object, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0019] It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0020] Reference Figure 1 , Figure 1 is a flowchart of an embodiment of the environment modeling based inspection robot navigation path optimization method of the present application.

[0021] In an embodiment, the environment modeling based inspection robot navigation path optimization method comprises: In one embodiment, a method for optimizing a navigation path of an inspection robot based on environment modeling is provided, and the method comprises: In step S100, an instruction for optimizing a navigation path of an inspection robot is obtained, wherein the instruction for optimizing a navigation path comprises scene attribute information corresponding to a target inspection scene and real-time perception data.

[0022] The scene attribute information can be a static or semi-static data set describing inherent structure and functional characteristics of the target inspection scene, and can be used to provide prior knowledge support for environment modeling and path planning, and to initialize model parameters and constraint conditions. Further, the scene attribute information can be extracted from non-real-time data sources such as pre-recorded engineering drawings, equipment layout diagrams, and regional function division tables, or imported by a manual annotation system. In this embodiment, the scene attribute information can include, but is not limited to, one or more of structure topology information, equipment distribution information, and regional risk level information.

[0023] The real-time perception data can be data reflecting the current environment state collected by sensors in real time during the operation of the inspection robot, and can be used to reflect the state of dynamic change elements in the environment, such as the position of a moving obstacle, the fluctuation of light intensity, and the distribution of temperature abnormal points. Further, the real-time perception data can be periodically collected and transmitted to a central processing system by perception units such as laser radars, visual cameras, infrared sensors, and ultrasonic modules. In this embodiment, the real-time perception data can be laser point cloud data, visual image streams, and thermal imaging data.

[0024] The instruction for optimizing a navigation path of an inspection robot can be obtained by receiving a start signal from an upper-layer scheduling system or a local triggering mechanism, and reading the attached scene attribute information and real-time perception data packet. Further, the instruction for optimizing a navigation path of an inspection robot can be obtained by receiving a remote instruction through an industrial communication protocol, or automatically initiating by a local event trigger of an edge computing node, so as to start the path optimization process and initialize the required input data.

[0025] In step S200, dynamic environment feature prediction is performed on the target inspection scene according to the scene attribute information, and a multi-scale environment feature flow is obtained.

[0026] The multi-scale environment feature flow can be a multi-time granularity and spatial resolution prediction data sequence representing the evolution trend of dynamic elements in the target inspection scene derived based on scene attribute information, and can be used to provide hierarchical prediction results of the future state of the environment to support path decision-making requirements with different response periods. Further, the multi-scale environment feature flow can be obtained by constructing a time series prediction model, combining historical operation rules and current environmental context to recursively calculate short-term mutation features and long-term evolution trends. In this embodiment, the multi-scale environment feature flow can include one or more of instantaneous disturbance features, medium-term evolution features, long-term evolution patterns, etc.

[0027] The dynamic environment feature prediction of the target inspection scene according to the scene attribute information can be to use a pre-trained time series prediction model or a graph evolution network to infer the future environmental change trend combining the structure and function features in the scene attribute information. Further, this operation can be implemented by using an LSTM network for sequence prediction, using a graph attention mechanism to model the evolution law of the correlation between devices, etc., so as to generate multi-granularity prediction results covering short-term disturbance and long-term evolution, and enhance the cognitive ability of the system to the future state.

[0028] In step S300, the multi-source environment modeling database is spatio-temporally registered and optimized according to the scene attribute information and real-time perception data, and a registered environment model library is determined.

[0029] The multi-source environment modeling database can be a data collection storing heterogeneous environment models from multiple modeling methods or data sources, and can be used to provide environment representation under diversified modeling perspectives and to compensate for the limitations of single modeling methods. Further, the multi-source environment modeling database can be obtained by integrating geometric maps generated by SLAM algorithms, semantic maps output by semantic segmentation models, functionally partitioned maps manually annotated, and other modeling results. In this embodiment, the multi-source environment modeling database can include one or more of a geometric modeling sub-database, a semantic modeling sub-database, a functional attribute modeling sub-database, etc.

[0030] The registered environment model library can be a unified coordinate system and time reference fused environment model collection after spatio-temporal consistency correction, and can be used to eliminate the spatio-temporal deviation between multi-source data and form a consistent environment representation that can be used for accurate path planning. Further, the registered environment model library can be obtained by performing spatial coordinate transformation and time stamp alignment operations on each sub-model in the multi-source environment modeling database using a spatio-temporal registration algorithm. In this embodiment, the registered environment model library can include one or more of a spatial registration model, a time synchronization model, a spatio-temporal joint registration model, etc.

[0031] According to the scene attribute information and the real-time perception data, the multi-source environment modeling database is spatiotemporally registered and optimized. A spatiotemporal consistency cost function can be established, combined with the prior constraints provided by the scene attribute and the observation values of the real-time perception data, to solve the optimal transformation parameters to realize model alignment. Further, the operation can be realized by using an iterative closest point algorithm for spatial registration, using a Kalman filter to align multi-source time series data, and the like, so as to solve the inconsistency of multi-source modeling data in spatial coordinate offset and time sampling asynchrony, and improve the fusion model precision.

[0032] In step S400, a plurality of path planning learners and a path optimization loss resolver are introduced to perform path optimization loss optimization learning on the registered environment model library, and a navigation path double channel of the inspection robot is built, wherein the navigation path double channel includes an instant obstacle avoidance path planning channel and a prospective topological path planning channel.

[0033] The path planning learner can be a machine learning model instance for learning path optimization strategies, and can be used to learn optimal behavior strategies from historical path decision samples to improve path generation quality. Further, the path planning learner can be obtained based on a deep reinforcement learning or supervised learning framework, and the parameters can be updated by interacting with the simulation environment to accumulate experience data. In this embodiment, the path planning learner can include but is not limited to one or more of a safety priority learner, an efficiency optimization learner, an energy consumption control learner, and the like.

[0034] The path optimization loss resolver can be a calculation module for decomposing and attributing analysis of the comprehensive loss function in the path planning process, and can be used to guide the plurality of path planning learners to adjust the training direction and realize differentiated adjustment of different optimization objectives. Further, the path optimization loss resolver can be obtained by decomposing the composite loss function into several interpretable sub-items (such as collision penalty, path length, energy consumption), and evaluating the contribution proportion of each sub-item to the total loss. In this embodiment, the path optimization loss resolver can include but is not limited to one or more of a gradient contribution analysis module, an error tracing module, a weight adaptive adjustment module, and the like.

[0035] The navigation path double channel can be a parallel architecture system composed of two functionally complementary path planning channels, and can be used to realize simultaneous processing of local emergency response and global path optimization, avoiding the functional limitations of a single channel architecture. Further, the navigation path double channel can be obtained by concurrent running of the two planning channels through software logic isolation or independent computing threads, and a fusion mechanism is set to coordinate the final output. In this embodiment, the navigation path double channel can include but is not limited to one or more of an instant obstacle avoidance path planning channel and a prospective topological path planning channel.

[0036] The instant obstacle avoidance path planning channel can be a low-delay path adjustment module that focuses on responding to sudden obstacles and local environmental disturbances, and can be used to ensure that the robot can timely avoid unexpected obstacles and maintain operation continuity and physical safety. Further, the instant obstacle avoidance path planning channel can be obtained based on a local planning algorithm such as a dynamic window method or an artificial potential field method, and a bypass trajectory can be quickly generated after receiving the latest perception data. In this embodiment, the instant obstacle avoidance path planning channel can include one or more of a reactive obstacle avoidance channel, a predictive obstacle avoidance channel, a hybrid obstacle avoidance channel, and the like.

[0037] The forward-looking topological path planning channel can be a high-level planning module that performs long-term optimal path exploration based on global environmental awareness, and can be used to avoid falling into a local optimal trap and ensure efficient completion of the overall inspection task. Further, the forward-looking topological path planning channel can be obtained using an improved A*, Dijkstra, or graph neural network algorithm to search for an optimal path that satisfies multiple constraint conditions on a semantic map. In this embodiment, the forward-looking topological path planning channel can include one or more of a heuristic topological channel, a probabilistic graph optimization channel, a semantic guidance path channel, and the like.

[0038] Introducing multiple path planning learners and path optimization loss analyzers for path optimization loss optimization learning on the registered environment model library can be to configure multiple learners with different optimization biases for parallel training, and to decompose the overall performance indicators through the loss analyzers and feed back to each learner to update the strategy. Further, this operation can be achieved by setting a weighted loss function for joint training, designing a gating mechanism to dynamically select the dominant learner output, and the like, so as to realize a collaborative learning mechanism for multi-objective path optimization and promote the balance between local safety and global efficiency.

[0039] In step S500, the real-time perception data and the multi-scale environmental feature stream are input into the navigation path double channel to generate a dynamic navigation path optimization report for the inspection robot.

[0040] The dynamic navigation path optimization report can be a structured output file containing the current recommended path and auxiliary decision information, and can be used to provide complete path execution instructions and emergency plan references for the inspection robot control system. Further, the dynamic navigation path optimization report can be obtained by fusing the double-channel path planning results, and additional metadata such as risk assessment, alternative paths, priority sorting, and the like. In this embodiment, the dynamic navigation path optimization report can include one or more of a main path suggestion report, a multi-path alternative report, a risk warning enhancement report, and the like.

[0041] Inputting real-time sensing data and multi-scale environmental feature streams into the dual-channel navigation path allows for the simultaneous feeding of the current sensing state and predicted features into the dual-channel architecture. Each channel generates a local response and a global planning result, which are then integrated through fusion logic to output a structured report. Furthermore, this operation can be achieved by employing a voting mechanism to fuse the dual-channel outputs and using a Bayesian inference framework to evaluate and rank the confidence of each path. This enables a comprehensive path decision output based on real-time state and future predictions, improving path adaptability and robustness.

[0042] Taking the inspection of rail transit tunnels as an example, the navigation path optimization method for inspection robots based on environmental modeling in this embodiment can be as follows: deploy inspection robots in subway tunnels to perform track and power supply system inspection tasks. The tunnel interior contains dynamic obstacles such as temporary construction debris and moving maintenance vehicles, and lighting conditions vary significantly depending on the section. The system first loads tunnel structure maps and equipment distribution information as scene attribute information. Combined with real-time perception data collected by LiDAR and visible light cameras, it predicts the airflow disturbance cycle caused by trains entering and leaving the station and the hot spots of personnel activity, forming a multi-scale environmental feature flow. Then, it performs spatiotemporal registration of the geometric map constructed by SLAM and the semantic map of manually annotated high-risk areas to generate a registration environment model library under a unified coordinate system. Multiple path planning learners are trained with different focuses on obstacle avoidance success rate and path smoothness, respectively. The loss analyzer identifies that safety weight should be higher than efficiency in the current environment and dynamically adjusts the learning strategy. The real-time obstacle avoidance channel quickly generates detour trajectories based on the latest point cloud data, while the forward-looking topology channel re-plans the inspection sequence according to the importance of all equipment. The final generated dynamic navigation path optimization report not only recommends the main path, but also includes three alternative detour routes and collision risk scores for each section, ensuring that the robot can stably complete the inspection task in a complex dynamic environment.

[0043] This embodiment provides a navigation path optimization method for inspection robots based on environmental modeling. It obtains navigation path optimization instructions containing scene attribute information and real-time perception data, driving the generation of multi-scale environmental feature flows to achieve dual prediction of short-term environmental disturbances and long-term evolution. A spatiotemporal registration optimization mechanism integrates heterogeneous models from multi-source modeling databases to construct a spatially consistent and temporally synchronized registered environment model library. Multiple path planning learners, guided by a path optimization loss parser, conduct collaborative training to achieve dynamic balance in multi-objective path optimization. A dual-channel navigation path architecture processes local obstacle avoidance and global path planning in parallel, ultimately generating a dynamic navigation path optimization report containing main path suggestions, alternative solutions, and risk assessment information. Through the coordinated operation of the above steps, this method can improve the real-time performance of path planning, enhance operational safety, and ensure the stability of task execution.

[0044] In one of the embodiments, the target inspection scene is dynamically predicted according to the scene attribute information to obtain a multi-scale environment feature flow, including: The target inspection scene is spatially functionally zoned according to the scene attribute information to determine a key inspection region set.

[0045] The key inspection region set can be a set of spatial regions with high task priority or potential risk, and can be used to narrow the range of environment feature collection and prediction calculation. Further, the key inspection region set can be combined with prior knowledge such as device importance, historical failure frequency, operation intensity, etc. in the scene attribute information to obtain spatial functional zoning prediction through a rule engine or a classification model. In this embodiment, the key inspection region set includes one or more of high-risk device area, personnel activity frequent area, unstable transition area of light, etc. Illustratively, the determination of the key inspection region set can be combined with risk score weighting through K-means clustering, or a random forest classifier is constructed to identify key functional blocks, so as to realize intelligent filtering from the whole scene to the key region and reduce invalid calculation overhead.

[0046] The structural semantic information, dynamic obstacle density information and light time-varying information of the key inspection region set are collected to obtain a target environment feature library.

[0047] The structural semantic information can be data describing the physical structure and its functional meaning in the key inspection region, and can be used to enhance the understanding ability of the robot to the environment function. Further, the structural semantic information can be obtained by extracting labeled spatial information such as device type, installation position, passage purpose, etc. from a pre-built semantic map or BIM model. The dynamic obstacle density information can be a statistical quantity reflecting the appearance frequency and aggregation degree of moving entities in the key inspection region, and can be used to quantify the space-time interference level of the environment. Further, the dynamic obstacle density information can be obtained by performing motion target detection and trajectory tracking on the laser radar point cloud or video stream to count the number of targets in a specific region per unit time. The light time-varying information can be a data sequence representing the variation law of the light intensity in the key inspection region with time, and can be used to identify the visual perception degradation area caused by day and night alternation or obstruction. Further, the light time-varying information can be obtained by periodically collecting brightness values through a light-sensitive sensor or a vision system, and recording the fluctuation period and amplitude characteristics. In this embodiment, the target environment feature library can include a structural topology sub-library, a dynamic distribution sub-library, and an optical environment sub-library, etc. By calling the perception system interface and semantic map service, the above three types of information are periodically obtained and integrated to construct a comprehensive environment description and improve the completeness of feature expression.

[0048] The spatial scale is classified according to the target environment feature library to obtain a multi-scale environment initial classification result.

[0049] The multi-scale environment initial classification result can be a result of classifying environment features according to a spatial range hierarchy, and can be used to establish a spatial granularity hierarchy of environment features. Further, the multi-scale environment initial classification result can be obtained by setting a spatial scale division standard (such as a meter level, a ten-meter level, and a hundred-meter level), and automatically matching a corresponding scale in combination with the size and scope of a geographical entity to which the feature belongs. In this embodiment, the multi-scale environment initial classification result includes, but is not limited to, a local device-level feature group, a channel connection-level feature group, a region function-level feature group, and the like. Illustratively, the spatial scale classification can be performed by using an ontology model to define a scale hierarchy relationship, or by designing a rule engine to match features and scale templates, thereby forming a hierarchical environment feature organization method and supporting hierarchical path planning strategies.

[0050] The multi-scale environment initial classification result is dynamically graded to obtain a multi-scale environment feature stream.

[0051] The dynamic grading can be a process of grading the degree of change of environment features over time, and can be used to distinguish stable parts from variable parts in the environment. Further, the dynamic grading can be obtained by analyzing the time series fluctuation characteristics (such as variance, frequency, and mutation times) of each feature, in combination with a weight model to output a dynamic intensity grade. In this embodiment, the dynamic grading includes, but is not limited to, a high-frequency disturbance grade, a trend evolution grade, an occasional mutation grade, and the like. Illustratively, the dynamic grading can be performed by calculating a feature change rate based on a sliding window, or by introducing a Markov model to predict state transition probabilities to determine a dynamic grade, thereby generating an output stream with a dynamic grade label, so that the path planning can distinguish between stable and rapidly changing areas.

[0052] Taking large chemical industrial park pipeline inspection as an example, the navigation path optimization method of the inspection robot based on environment modeling in the embodiment can be: deploying an inspection robot in a chemical industrial park containing a reaction device area, a storage tank area and a pipe gallery passage. The system first performs spatial functional area prediction based on process flow diagrams and safety zoning drawings, identifies the high-temperature high-pressure reactor periphery, the valve concentration area and the personnel cross-operation passage as a key inspection area set; then collects the structural semantic information (such as pipeline direction, maintenance platform position) of these areas, the dynamic obstacle density information (through radar monitoring of the moving frequency of workers during the inspection period) and the time-varying illumination information (recording the change of picture noise caused by insufficient night light); after the above data is summarized into a target environment feature library, it is divided into three categories according to the spatial scale: device level (single reaction kettle), passage level (passage between adjacent devices), and area level (entire device unit); then, combined with the obstacle occurrence frequency and illumination fluctuation amplitude at each scale, dynamic classification is performed, and it is found that although a certain maintenance passage belongs to the passage level, it has the highest dynamic level; the finally generated multi-scale environment feature flow guide focuses on the sudden lane occupation of this passage for immediate obstacle avoidance, while the forward-looking topological passage avoids the high dynamic area to re-plan the main path, effectively avoiding the path interruption problem caused by temporary construction.

[0053] The navigation path optimization method of the inspection robot based on environment modeling provided in the embodiment, by performing spatial functional area prediction on the target inspection scene according to the scene attribute information, determining the key inspection area set, narrowing the environment feature collection and prediction calculation range; by collecting the structural semantic information, dynamic obstacle density information and illumination time-varying information of the key area, constructing the target environment feature library, providing the data basis representing the comprehensive state; by spatial scale classification to form a multi-scale environment initial classification result, establishing the spatial granularity level of the environment feature; by dynamic classification to generate a multi-scale environment feature flow, giving the environment feature change speed discrimination attribute. Through the above steps, the method cooperates to make the path planning pay attention to the importance of the key area and judge the dynamic change degree of the area, so as to realize more fine and adaptive path decision in complex industrial scenes, and significantly improve the safety and execution efficiency of the inspection process.

[0054] In one of the embodiments, the multi-source environment modeling database is spatio-temporally registered and optimized according to the scene attribute information and real-time perception data to determine a registered environment model library, including: According to the scene attribute information, the multi-source environment modeling database is associated and filtered by scene type to obtain a homogeneous scene environment model library.

[0055] The isomorphic scene environment model library can be a subset model set that has high similarity in structural type, functional layout, and operation characteristics with the current target inspection scene, and can be used to narrow down the modeling data search range. Further, the isomorphic scene environment model library can combine the context to describe its acquisition method, that is, based on the semantic labels such as structural topology information and device distribution information in the scene attribute information, the models in the database are classified and filtered through matching rules or vector similarity calculation. In this embodiment, the isomorphic scene environment model library can be one or more of a power equipment area model sub-library, a pipeline dense area model sub-library, and an open channel area model sub-library. Further, the isomorphic scene environment model library can be an operation object for spatio-temporal twin degree evaluation, providing a candidate set for subsequent generation of model registration coefficients.

[0056] The scene type associated filtering of the multi-source environment modeling database according to the scene attribute information can be to extract the type identification field in the scene attribute information, and match and search the metadata labels of each model in the multi-source environment modeling database to filter out the consistent subset. Further, this operation can be implemented by using a keyword-based label matching method, using semantic encoding vector to calculate similarity and sort filtering, etc., so as to reduce the data scale of subsequent processing and improve the pertinence and efficiency of the modeling process.

[0057] The spatio-temporal twin degree evaluation of each model in the isomorphic scene environment model library according to the real-time perception data obtains a plurality of model registration coefficients.

[0058] The model registration coefficient can be a numerical index quantifying the consistency degree of a certain environment model and the current real scene in time evolution and space form, and can be used to evaluate the credibility of its participation in the final modeling. Further, the model registration coefficient can combine the context to describe its acquisition method, that is, by comparing the difference between the real-time perception data and the model prediction state, and combining the dynamic weight to fuse the time synchronization error and the space geometric deviation to calculate. In this embodiment, the model registration coefficient can include one or more of a time consistency coefficient, a space similarity coefficient, and a spatio-temporal joint matching coefficient. Further, the model registration coefficient can be an output result of the spatio-temporal twin degree evaluation, which is a basic input for constructing the model registration optimization distribution.

[0059] The spatio-temporal twin degree evaluation can be a comprehensive analysis process of evaluating the consistency of the environment model and the real scene in time dynamic change and space geometric structure, and can be used to support model optimization decision. Further, this operation can be implemented by using real-time perception data to drive simulation playback, comparing the time sequence evolution track and space topological relationship difference between the model output and the actual observation, etc., so as to establish a quantifiable corresponding relationship between the model and the reality, and support fine model optimization.

[0060] The model registration threshold is used to optimize and select the plurality of model registration coefficients, to obtain a model registration optimization distribution greater than or equal to the model registration threshold.

[0061] The model registration threshold can be a critical numerical standard used to determine whether the environment model has sufficient matching accuracy to participate in the final fusion, and can be used to exclude modeling results of low quality or severe deviation. Further, the model registration threshold can be combined with the context to describe its acquisition method, that is, dynamically determined according to the system set safety margin, historical data statistical distribution or online learning adjustment mechanism. In this embodiment, the model registration threshold can include one or more of a static preset threshold, a dynamic adaptive threshold, and a differentiated threshold for different regions. Further, the model registration threshold can be applied to the model registration coefficient, guiding the formation of the model registration optimization distribution.

[0062] The model registration optimization distribution can be a set of high matching degree environment models and the distribution state of the corresponding registration coefficients after threshold screening, and can be used to identify high-quality model resources that can be used for fusion. Further, the model registration optimization distribution can be combined with the context to describe its acquisition method, that is, comparing the registration coefficients of each model with the model registration threshold, and selecting the models greater than or equal to the threshold to form an effective distribution set. In this embodiment, the model registration optimization distribution can include one or more of a unimodal centralized distribution, a multimodal parallel distribution, and a sparse discrete distribution. Further, the model registration optimization distribution can be used as the basis for dynamic pruning and fusion, to determine which models are retained and participate in fusion.

[0063] Optimizing and selecting the plurality of model registration coefficients according to the model registration threshold can be comparing the registration coefficient corresponding to each model with a preset or adaptive model registration threshold, and retaining the models and their coefficients that meet the conditions. Further, this operation can be implemented by setting a fixed threshold for hard screening, introducing a probability threshold for soft selection, and the like, so as to ensure that only high matching degree models enter the subsequent fusion stage, and improve the reliability of the final model.

[0064] The model registration optimization distribution is used to dynamically prune and fuse the isomorphic scene environment model library, to generate a registration environment model library.

[0065] The dynamic pruning and fusion can be a combination of selective removal and information integration of the candidate model set according to the model registration optimization distribution, and can be used to generate a compact and more accurate fusion model. Further, the dynamic pruning and fusion can combine the context to obtain the way, that is, the models below the registration threshold are removed, and the models above the threshold are combined into a unified expression by weighted average, graph structure splicing or feature embedding fusion. In this embodiment, the dynamic pruning and fusion can include one or more of linear weighted fusion, graph topology splicing fusion, attention mechanism guided fusion, etc. Further, the dynamic pruning and fusion can be applied to the isomorphic scene environment model library, and the registration environment model library is output.

[0066] The dynamic pruning and fusion of the isomorphic scene environment model library according to the model registration optimization distribution can be based on the effective model set marked in the model registration optimization distribution, deleting the unselected models, and performing unified coordinate transformation and data fusion on the spatial structure and semantic attributes of the selected models. Further, the operation can be implemented by using a grid map superposition fusion method, using a graph neural network to aggregate the topological structure, so as to eliminate redundant modeling information, integrate high-credibility segments, and generate a registration environment model library with clear structure and improved precision.

[0067] Taking the inspection of a complex device area in a chemical park as an example, the navigation path optimization method of the inspection robot based on environment modeling in this embodiment can be that in a chemical device area containing a large number of reaction kettles, pipeline groups and valve groups, the environment structure is dense and there are dynamic interferences such as high-temperature steam shielding and mobile maintenance platforms. The system first selects historical three-dimensional reconstruction models under the same working condition from a multi-source environment modeling database according to the "high-density pipeline area" and "high-pressure equipment aggregation" labels in the scene attribute information, and constructs an isomorphic scene environment model library; then, laser radar and thermal imaging real-time data are connected, and the space-time twin degree of each model in the library is evaluated, the consistency of the current pipeline surface temperature distribution trend and the model prediction is compared, and the spatial matching degree of the steam cloud diffusion path is verified, to generate a model registration coefficient of each model; a dynamically adjusted model registration threshold is set, and only the models with high matching degree are retained to form a model registration optimization distribution; finally, dynamic pruning and fusion are performed on the high-matching models, the old pipeline parts that have been removed are removed, and the new sensor positions are integrated, to generate a registration environment model library that accurately reflects the current real state, provide a high-quality environment base map for double-channel path planning, and ensure that the robot can safely and efficiently pass through in a limited vision and complex structure environment.

[0068] The embodiment provides a navigation path optimization method for a patrol robot based on environment modeling. The method performs scene type correlation screening on a multi-source environment modeling database according to scene attribute information, constructs an isomorphic scene environment model library, and realizes accurate positioning of the modeling category. The method uses real-time perception data to evaluate the space-time twin degree of the model in the library, generates a model registration coefficient, and changes the model selection from experience dependence to data driving. The method optimizes and selects the coefficient through a model registration threshold, forms a model registration optimization distribution, and ensures that only high-confidence models participate in subsequent modeling. Finally, the method performs dynamic cropping and fusion operations based on the distribution, removes low-matching-degree model segments, integrates high-quality modeling results, and generates a registration environment model library with clear structure and improved precision. The process significantly optimizes the integration process of multi-source heterogeneous environment data, solves the modeling distortion problem caused by model mixing and update lag in traditional methods, improves the space-time consistency and reality fit of environment expression, and provides a more accurate and reliable environment cognition basis for subsequent path planning.

[0069] In one of the embodiments, a plurality of path planning learners and path optimization loss analyzers are introduced to perform path optimization loss optimization learning on the registration environment model library, and a navigation path double channel for the patrol robot is built, including: According to the registration environment model library, historical path log retrieval is performed to obtain an instant obstacle avoidance path log library and a forward-looking topological path log library.

[0070] The instant obstacle avoidance path log library can be a structured data set recording local obstacle avoidance behaviors and related environment states of the patrol robot in historical operation, and can be used to provide behavior sample support in typical disturbance scenes for training of the instant obstacle avoidance path planning channel. Further, the instant obstacle avoidance path log library can include but is not limited to one or more of static obstacle bypass logs, dynamic obstacle prediction logs, and multi-obstacle cooperative avoidance logs. The forward-looking topological path log library can be a data set storing long-term path decisions and context information of the patrol robot formed in the global path exploration process, and can be used as a basic data source for the forward-looking topological path planning channel to learn long-term task logic and structured path patterns. Further, the forward-looking topological path log library can include but is not limited to one or more of multi-site sequential planning logs, cross-region transfer path logs, and periodic patrol optimization logs.

[0071] In the embodiment, the operation can be performed by taking a spatial node or a semantic region in the registration environment model library as an index key, matching and classifying and extracting corresponding behavior records in the historical path log database. Further, the operation can be implemented by matching log entries based on a geographic fence range or using path semantic labels for automatic classification retrieval, so that the function separation of path knowledge can be realized, and learning data bases for local response and global optimization are respectively constructed.

[0072] According to the instant obstacle avoidance path log library, environmental disturbance feature collection is performed to obtain a first disturbance feature record library.

[0073] The first disturbance feature record library can be a feature set representing environmental instantaneous uncertainty, and can be used to model common dynamic interference types in the local environment, thereby improving the generalization ability of the obstacle avoidance strategy to unknown disturbances. Further, the first disturbance feature record library can include, but is not limited to, one or more of sensor noise patterns, obstacle emergence patterns, and environmental vibration interference patterns.

[0074] In this embodiment, the operation can be statistical analysis of trajectory deviation, speed fluctuation, control frequency, and other parameters in the instant obstacle avoidance path log library, identification of high-frequency disturbance patterns, and formation of feature templates. Further, the operation can be achieved by using principal component analysis to extract key disturbance factors or using an autoencoder network for nonlinear feature compression, thereby converting the original obstacle avoidance behavior into quantifiable disturbance feature representation and enhancing the model's understanding of the dynamic environment.

[0075] According to the long-term task constraint feature collection of the prospective topological path log library, a second constraint feature record library is obtained.

[0076] The second constraint feature record library can be a high-level semantic feature set reflecting long-term task requirements and system limitation conditions, and can be used to provide task-level optimization guidance for global path planning and avoid short-sighted decisions based on geometric distance. Further, the second constraint feature record library can include, but is not limited to, one or more of timing dependency constraints, energy budget constraints, and device access permission constraints.

[0077] In this embodiment, the operation can be to extract task sequences, dwell time, path selection preferences, and other information from the prospective topological path log library, and to derive implicit constraint rules in combination with external task management system data. Further, the operation can be achieved by applying association rule mining to discover task dependency relationships or using a graph neural network to model task topology, thereby mining the high-level task logic behind historical decisions and supporting more strategic path generation.

[0078] According to the plurality of path planning learners and the path optimization loss resolver, instant obstacle avoidance path optimization loss optimization learning is performed on the registered environment model library, the first disturbance feature record library, and the instant obstacle avoidance path log library to generate an instant obstacle avoidance path planning channel.

[0079] In this embodiment, the operation can be to fuse three types of input data into joint training samples, to learn optimal response strategies under different disturbance scenarios in parallel by multiple path planning learners, and to decompose and guide gradient updates for collision risk, reaction delay and other sub-item losses by the path optimization loss resolver. Further, the operation can be implemented by setting an adversarial training mechanism to enhance noise resistance or introducing an attention module to focus on key disturbance features, so as to construct a local obstacle avoidance channel with strong robustness and fast response capability, which can effectively cope with complex dynamic disturbances.

[0080] According to the multiple path planning learners and the path optimization loss resolver, the registered environment model library, the constraint feature second record library and the prospective topological path log library are subjected to prospective topological path optimization loss optimization learning, to generate a prospective topological path planning channel.

[0081] In this embodiment, the operation can be to train the learning ability of the learning learner to search for an optimal path under multiple restriction conditions based on the global semantic map, in combination with task constraint features and historical successful path samples, and the path optimization loss resolver dynamically adjusts the weight influence of each constraint term. Further, the operation can be implemented by using a hierarchical reinforcement learning architecture to process multi-level decision or designing a soft constraint mechanism to allow limited violations in exchange for overall optimization, so as to generate a long-term path scheme that takes into account task integrity and resource efficiency, and reduces repeated exploration and invalid movement.

[0082] The instant obstacle avoidance path planning channel and the prospective topological path planning channel are connected as parallel nodes to generate a dual-channel navigation path for the inspection robot.

[0083] In this embodiment, the operation can be to keep the two channels running independently on the computing architecture, to receive the outputs of both sides through the intermediate coordination module and to select the dominant channel or the fusion result according to the current environment state. Further, the operation can be implemented by using a dynamic weight fusion strategy or setting a conflict detection and arbitration mechanism to handle the contradiction between the outputs of the two channels, so as to realize functional specialization and structural decoupling, and to ensure that local safety response does not interfere with the stability of the global path.

[0084] Taking the night inspection of a chemical industrial park as an example, the navigation path optimization method of the inspection robot based on environment modeling in the embodiment can be as follows: a chemical plant inspection robot needs to complete a gas leakage monitoring task in a high-risk area at night. The system calls the layout and pipeline distribution information of the plant in the registered environment model library, retrieves historical path logs, separates the logs that have dealt with the passing of transport vehicles in the past to form an instant obstacle avoidance path log library, and includes records of high-temperature area detouring and multi-point sampling sequence optimization to form a forward-looking topological path log library; further, common disturbance patterns (such as radar echo abnormalities and visual blurring) are extracted to form a first record library of disturbance features, and rules such as “high-pressure storage tanks must be detected first” and “single task power consumption should not exceed 80%” are summarized to form a second record library of constraint features; multiple path planning learners are trained for emergency avoidance success rate and task completion rate respectively, and a loss analyzer identifies that safety weight is higher in the current task, and strengthens the obstacle avoidance channel output; finally, in the dual-channel architecture built, the instant obstacle avoidance channel responds to sensor false alarms caused by steam clusters in real time, while the forward-looking topological channel continuously evaluates whether the remaining power is sufficient to complete all sampling point coverage and timely suggests returning to charge in advance; the two channels are connected in parallel nodes, and the final action instruction is determined by the central coordinator to ensure that the robot can stably perform key inspection tasks in a highly disturbed environment.

[0085] The navigation path optimization method of the inspection robot based on environment modeling provided in the embodiment realizes the structural decoupling of path experience in the functional dimension by indexing the registered environment model library to carry out historical path log retrieval, dividing the original path data into an instant obstacle avoidance path log library and a forward-looking topological path log library; further, by extracting disturbance features and constraint features from the two types of log libraries, a first record library of disturbance features and a second record library of constraint features are constructed, so that the system can distinguish between short-term dynamic interference and long-term optimization goals; using multiple path planning learners combined with a path optimization loss analyzer, the two feature libraries and the corresponding log libraries are trained respectively, the former focuses on improving the robustness and response speed of local obstacle avoidance, and the latter is committed to optimizing the strategic rationality of the global path; finally, the dual channels are connected through parallel nodes to form a unified architecture, which not only avoids internal target conflicts in a single channel, but also retains the respective professional advantages, and realizes decision fusion through a coordination mechanism. Through the above steps, the local safety and global efficiency can be balanced in complex industrial scenarios.

[0086] In one of the embodiments, according to the multiple path planning learners and the path optimization loss analyzer, the registered environment model library, the first record library of disturbance features, and the instant obstacle avoidance path log library are subjected to instant obstacle avoidance path optimization loss optimization learning to generate an instant obstacle avoidance path planning channel, which includes: The registered environment model library and the disturbance feature first record library are taken as input information, the real-time obstacle avoidance path log library is taken as output label, and a plurality of path planning learners are supervised and trained respectively, and when a predetermined number of iterations is trained, a plurality of real-time path optimization loss coefficients are calculated according to a path optimization loss resolver; If the plurality of real-time path optimization loss coefficients are less than a real-time path optimization loss threshold, a plurality of real-time obstacle avoidance path planning models are generated; The plurality of real-time obstacle avoidance path planning models are connected to obtain a real-time obstacle avoidance path planning channel.

[0087] The real-time path optimization loss coefficient can be a quantitative index representing the deviation degree of a specific optimization target obtained by the path optimization loss resolver after multidimensional decomposition of the output result of each path planning learner in the supervised training process, and can be used to evaluate the compliance degree of the model in different obstacle avoidance performance dimensions to support training termination judgment and model selection. Further, the real-time path optimization loss coefficient can be combined with the context to describe its acquisition method, that is, by comparing the learner predicted path with the real path in the real-time obstacle avoidance path log library, the error values of collision risk, trajectory jitter, response delay and other sub-items are calculated respectively, and are normalized to comparable coefficients. In the embodiment, the real-time path optimization loss coefficient can include but is not limited to one or more of safety constraint loss coefficient, dynamic response loss coefficient, motion smoothing loss coefficient, etc.

[0088] The predetermined number of iterations can be a model parameter update period unit set in the supervised training process, which can be used to control the loss evaluation frequency and balance the training efficiency and convergence stability. Further, the predetermined number of iterations can be combined with the context to describe its acquisition method, that is, in the training loop, the training stage is divided into a fixed batch size unit, and the loss resolver is triggered once for each completed parameter update. In one specific embodiment, the predetermined number of iterations can use one or more of fixed step iteration, adaptive interval iteration, event-driven iteration, etc.

[0089] The registered environment model library and the disturbance feature first record library are taken as input information, the real-time obstacle avoidance path log library is taken as output label, and a plurality of path planning learners are supervised and trained respectively, which can be to splice the spatial semantic state of the registered environment model library and the dynamic interference mode of the disturbance feature first record library into a joint input vector, match the obstacle avoidance action sequence or trajectory point set at the corresponding time in the real-time obstacle avoidance path log library as a supervision signal, and carry out end-to-end supervised learning. Further, the operation can be achieved by predicting the trajectory point sequence using the sequence-to-sequence architecture and learning the expert-level obstacle avoidance action strategy using the behavior cloning method, so that the mapping relationship from the high-precision environment state + typical disturbance feature to the historical successful obstacle avoidance strategy can be established by each path planning learner, and the authenticity and reliability of the local path generation can be improved.

[0090] The plurality of instant path optimization loss coefficients can be calculated according to the path optimization loss resolver at each predetermined iteration of the training, that is, after each predetermined iteration is completed, the path optimization loss resolver is called to compare the current output path of each learner with the label path in multiple granularities, and output the normalization coefficient of each sub-item error. Further, this operation can be realized by evaluating the trajectory distribution similarity based on the Jensen-Shannon divergence and calculating the path topology consistency score using the spatio-temporal graph convolution network, so as to realize the phased quality monitoring and direction calibration of the training process, and avoid overfitting of the model on a single target at the expense of other key performance.

[0091] If the plurality of instant path optimization loss coefficients is less than the instant path optimization loss threshold, a plurality of instant obstacle avoidance path planning models are generated. The instant path optimization loss threshold can be a preset upper limit value of the loss coefficient for determining whether the path planning learner meets the instant obstacle avoidance performance requirement, and can be used as a decision criterion for model solidification and channel construction to ensure that the generated instant obstacle avoidance path planning model has engineering usability. Further, the instant path optimization loss threshold can be combined with the context to describe its acquisition method, that is, it is set according to the safety level specification and real-time performance requirement of the industrial inspection scene, supporting differential configuration according to task type or regional risk level. In the embodiment, the instant path optimization loss threshold can include one or more of a hard safety threshold, a soft response timeliness threshold, a fault tolerance smoothness threshold, and the like.

[0092] The plurality of instant obstacle avoidance path planning models can be a plurality of functionally complementary local path planning model instances formed by supervised training and meeting the loss threshold condition, which can be used to constitute a redundant and diversified model set to improve the generalization ability and anti-single-point failure ability of the channel to unobserved disturbance patterns. Further, the plurality of instant obstacle avoidance path planning models can be combined with the context to describe their acquisition method, that is, each model is obtained by differentiating training of an independently initialized path planning learner, and there are design differences in structure or training target. In one exemplary embodiment, the plurality of instant obstacle avoidance path planning models can include one or more of an emergent obstacle response model, a sensor noise robust model, and a multi-obstacle collaborative avoidance model.

[0093] If the plurality of instant path optimization loss coefficients are less than the instant path optimization loss threshold, a plurality of instant obstacle avoidance path planning models are generated. That is, all instant path optimization loss coefficients output by each path planning learner are compared item by item, and when all coefficients are lower than the corresponding threshold, the current weight parameter is fixed and encapsulated as an independent model. Further, the operation can be realized by setting a relaxation judgment mechanism to allow a single item to slightly exceed the threshold but the overall weighted sum to meet the standard, and introducing a confidence gating mechanism to filter low-confidence samples for re-evaluation, so that each generated instant obstacle avoidance path planning model can meet the engineering usable standard in terms of safety, response, smoothness and the like.

[0094] The plurality of instant obstacle avoidance path planning models are connected to obtain an instant obstacle avoidance path planning channel. That is, the plurality of fixed instant obstacle avoidance path planning models are combined into a unified inference unit through an integrated architecture to support parallel forward propagation and result fusion. Further, the operation can be realized by using a weighted average to fuse the output trajectories of each model, and using a gating network to dynamically select the optimal model output, so that a path generation channel with model diversity and decision redundancy can be constructed, and the adaptability to unknown disturbance types and system robustness can be enhanced.

[0095] Taking the intelligent factory AGV cooperative inspection as an example, the navigation path optimization method of the inspection robot based on environment modeling in the embodiment can be as follows: In a clean room of a semiconductor wafer factory, multiple inspection robots need to pass through among a high-density device array while avoiding frequently scheduled logistics AGVs. The system calls the device contour and channel semantic map in the registered environment model library with an accuracy of 0.5 cm, combines the typical interference modes recorded in the disturbance feature first record library such as AGV emergency stop trajectory mutation and laser point cloud jitter caused by clean room airflow, and uses the 237 instant obstacle avoidance path logs of historical successful AGV avoidance as labels to carry out supervised training on five path planning learners initialized differently. After every 100 iterations, the path optimization loss resolver calculates the collision risk coefficient (based on the minimum safety distance), the response delay coefficient (based on the time difference between the issuance of an instruction and the effectiveness of the trajectory), and the trajectory jitter coefficient (based on the curvature change rate variance) of each model. When the three coefficients of a certain learner are continuously lower than the preset threshold (0.02, 150 ms, and 0.85) for two times, the learner is fixed as an instant obstacle avoidance path planning model. Finally, the five qualified models are connected through a gating fusion mechanism to form an instant obstacle avoidance path planning channel. When an AGV suddenly cuts into the inspection channel, the channel can generate a collision-free detour path within 68 ms, and the trajectory curvature is continuous and the execution is smooth, which is significantly better than the single model scheme.

[0096] The embodiment provides a navigation path optimization method for a patrol robot based on environment modeling. The method jointly takes a registration environment model library and a disturbance feature first record library as inputs, takes an instant obstacle avoidance path log library as a supervision label, performs supervision training on multiple path planning learners, tightly couples a real environment state and historical effective decisions in a model learning process, and improves physical rationality and task adaptability of local path generation. The method introduces a path optimization loss resolver to calculate multiple instant path optimization loss coefficients at a predetermined iteration number, realizes multi-angle quantitative evaluation and dynamic guidance on model performance, and avoids optimization deviation caused by only relying on a total loss in traditional training. The method solidifies the instant obstacle avoidance path planning model when all loss coefficients output by the learners are lower than a preset threshold, ensures that each model meets industrial-level reliability requirements in key performance indicators, and finally connects multiple qualified models into an instant obstacle avoidance path planning channel, improves coverage ability for diversified disturbances, and enhances system fault tolerance. The mechanism enables the channel to have high precision, strong robustness and fast response characteristics, and provides structural guarantee for stable output of the local obstacle avoidance function in the navigation path double channel.

[0097] In one embodiment, the path optimization loss resolver includes a path optimization loss analysis function, and the path optimization loss analysis function is: wherein, represents a path optimization loss coefficient, N represents a predetermined iteration number, n represents the nth iteration, N and n are positive integers, and 1≤n≤N; represents the number of path collision events in the nth iteration; represents the number of path acceleration over-limit segments in the nth iteration; represents the average deviation distance of the path from the preset topological skeleton line in the nth iteration; and λ and μ are non-negative weight coefficients for balancing safety, smoothness and structural constraints.

[0098] The path optimization loss analysis function can be a mathematical expression for quantifying path planning quality and supporting multi-objective trade-off, and can be used to provide an interpretable path quality evaluation mechanism. Further, the path optimization loss analysis function can be a differentiable function in the form of a weighted sum based on the preset three indicators: the number of collision events, the number of acceleration over-limit segments, and the average path deviation distance, and is used as a loss feedback signal in the learner training process. In the embodiment, the path optimization loss analysis function can include but is not limited to one or more of a safety-dominant loss function, a smoothness-enhanced loss function, and a structure-constrained enhanced loss function.

[0099] may be a scalar coefficient representing the optimization performance of path synthesis in a single iteration, which can be used to drive the path planning learner to optimize safety, smoothness, and structural consistency in iterations. Further, may be obtained by , multiplying JERK, DEV, and JERK by their corresponding weights and adding them together, as the input basis for the backpropagation calculation of the path planning learner. In the present embodiment, may include but not limited to one or more of the following: high safety penalty type , high smoothness constraint type , balanced adjustment type , etc.

[0100] N can be the maximum number of iteration calculations preset in the path optimization process, which can be used to limit the total number of repeated evaluations of the loss function. Further, N can be determined by system initialization configuration, which is used to control the time boundary and convergence judgment condition of the optimization process. In the present embodiment, N can include but not limited to one or more of the following: short period iteration configuration, medium period iteration configuration, long period fine optimization configuration, etc.

[0101] n can be the iteration number currently in, which can be used to mark the time node of loss calculation. Further, n can be incremented from the initial value 1 round by round until it reaches the predetermined iteration number N and terminates. In the present embodiment, n can include but not limited to one or more of the following: initial exploration stage (n is small), medium convergence stage (n is medium), and final stable stage (n is close to N).

[0102] may be the statistical number of path and obstacle intersection events occurring in the nth path optimization iteration, which can be used to reflect the safety property of the path. Further, The number of times the candidate path segment overlaps with the known obstacle region can be determined by a spatial geometry detection algorithm, and the number of times the candidate path segment overlaps with the known obstacle region can be accumulated. In the present embodiment, may include but not limited to one or more of the following: point-level collision count, line segment-level conflict count, and potential collision warning count within a time-space window.

[0103] may be the total number of continuous path segments in the path generated in the nth iteration, whose motion acceleration change rate exceeds the allowed threshold, which can be used to measure the dynamic execution stability of the path. Further, The instantaneous jerk value can be derived based on the path curvature and speed profile, and the segments that exceed the device's bearing limit can be identified and counted. In the present embodiment, may include but not limited to one or more of the following: linear acceleration mutation segment, angular acceleration jitter segment, and compound motion impact segment.

[0104] The average spatial deviation distance of the generated path relative to the preset topology skeleton line in the nth iteration can be used to maintain the structure of the inspection task logic. Further, The Euclidean distance mean value can be calculated as the deviation measure by projecting each sampling point on the actual path to the reference skeleton line. In this embodiment, It can include but is not limited to one or more of the following: main road deviation, key node proximity deviation, functional area coverage offset, etc.

[0105] Lambda can be a non-negative numerical coefficient for adjusting the relative importance of the collision term ( ) in the total loss, which can be used to control the influence strength of the safety goal in path optimization. Further, lambda can be set by a task priority strategy or an online parameter adjustment mechanism, participating The weighted operation of the formula. In this embodiment, lambda can include but is not limited to one or more of the following: static preset weight, dynamic adaptive weight, scene-aware switching weight, etc.

[0106] Mu can be a non-negative numerical coefficient for adjusting the relative importance of the path deviation term ( ) in the total loss, which can be used to affect the degree of adherence of the path to the preset inspection logic. Further, mu can be similar to lambda, configured by a strategy module or automatically updated by a learner feedback, used to balance the structural compliance requirements. In this embodiment, mu can include but is not limited to one or more of the following: global structure preservation weight, local node anchoring weight, segmented differentiated constraint weight, etc.

[0107] The non-negative weight coefficient can be an adjustment parameter with a value greater than or equal to zero, which can be used to achieve flexible trade-off between different optimization goals. Further, the non-negative weight coefficient can be generated by manual configuration, rule engine or meta-learning mechanism, which can remain constant or dynamically change in each iteration. In this embodiment, the non-negative weight coefficient can include but is not limited to one or more of the following: safety-efficiency trade-off coefficient, smoothing-structure trade-off coefficient, dynamic-static constraint switching coefficient, etc.

[0108] Defining a path optimization loss analysis function to calculate a path optimization loss coefficient It can be to establish a weighted sum formula containing , , Three indicators, which can be used to build a decomposable and adjustable loss evaluation system. Further, this operation can be batch trained by using fixed weight combination, or designed online weight adjustment module to dynamically modify lambda and mu according to environmental changes, so as to achieve the technical effect of supporting multi-objective collaborative optimization.

[0109] Count the number of path collision events in the nth iteration It can be a spatial intersection detection of the currently generated path with the obstacle region in the registered environment model library, which can be used to quantify the path obstacle avoidance performance. Further, this operation can be achieved by using bounding box fast culling non-collision area, or using precise grid scanning to detect subtle intrusion, so as to form a clear punishment signal for insufficient safety.

[0110] Count the number of path acceleration over-limit sections in the nth iteration It can be based on the jerk value of each section of motion derived from the path curvature and speed planning model, which can be used to suppress the sharp changes in the path that are not conducive to smooth operation. Further, this operation can be achieved by pre-setting the jerk threshold for hard truncation judgment, or introducing a soft punishment mechanism to impose progressive punishment on adjacent over-limit sections, so as to improve the execution reliability.

[0111] Calculate the average deviation distance of the path from the preset topological skeleton line in the nth iteration It can be to project multiple sampling points on the current path onto the preset topological skeleton line, which can be used to maintain the consistency of the path with the original inspection logic. Further, this operation can be achieved by using KD tree to accelerate the nearest point search, or giving higher calculation weight to the area near the key node, so as to prevent structural deviation of the task.

[0112] By adjusting the non-negative weight coefficients λ and μ to balance safety, smoothness and structural constraints, it can be to change the numerical ratio of λ ( weight) and μ ( weight) according to the current task demand or environmental risk level, which can be used to achieve dynamic trade-off between multiple optimization goals. Further, this operation can be achieved by automatically increasing the value of λ to cope with high-risk areas according to the risk map, and temporarily increasing μ in the structure-sensitive section to lock the inspection route, so as to adapt to different priority scenarios.

[0113] Taking the night inspection of a chemical plant area as an example, the inspection robot navigation path optimization method based on environment modeling of the embodiment can be: when performing night equipment temperature monitoring task in a chemical plant area, the system loads the three-dimensional layout map of the plant area as scene attribute information, combines laser radar and thermal imaging real-time sensing data to predict personnel activity hotspots and temporary occupation situation, and generates multi-scale environment feature flow; the SLAM map and semantic annotation map in the multi-source modeling database are spatiotemporally registered to form a consistent registered environment model library; the path planning enters the iterative optimization stage, and the path optimization loss analysis function respectively counts 、 、 : when the path passes through the temporary stacking area increase, triggering high lambda value reinforcement obstacle avoidance; in the dense area of the curve increase, prompting the learner to reduce the curvature change rate; if the path deviates from the main inspection line due to serious detour, the DEV increases, and the path is pulled back to the vicinity of the topology skeleton by increasing the mu value; the final generated dynamic navigation path optimization report not only contains the main path, but also attaches the risk score and backup route suggestion, ensuring that the robot can still safely, smoothly and accurately complete the inspection task in a complex environment.

[0114] The embodiment provides a navigation path optimization method for an inspection robot based on environment modeling, which inputs the real-time perception data and the multi-scale environment feature flow into the path optimization loss analysis function , , The three indicators are weighted and fused with non-negative weight coefficients lambda and mu to construct a clear multi-objective optimization evaluation system; the path quality is evaluated through round-by-round iteration, wherein reflects the spatial conflict frequency with obstacles, captures the acceleration mutation segment, measures the average distance from the topology skeleton line; the dynamic balance of safety, smoothness and structural constraint is achieved by adjusting lambda and mu; finally, the path planning learner continuously optimizes the strategy in the training, so that the local response and global decision are coordinated in a unified framework. This mechanism significantly improves the path quality, running stability and task completion reliability of the inspection robot in a dynamic and complex environment.

[0115] In one of the embodiments, the real-time perception data and the multi-scale environment feature flow are input into the navigation path double channel to generate a dynamic navigation path optimization report for the inspection robot, including: The real-time perception data and the multi-scale environment feature flow are input into the instant obstacle avoidance path planning channel to obtain a multi-scale environment-instant obstacle avoidance path set.

[0116] The multi-scale environment-instant obstacle avoidance path set can be a set containing various local obstacle avoidance paths, which can be used to provide diversified response schemes for sudden obstacles. Further, the multi-scale environment-instant obstacle avoidance path set can combine the real-time perception data and short-term disturbance information to evaluate the feasibility of the detour trajectory in a dynamic window in parallel, and output a candidate path group that meets the safety constraint. Exemplarily, the multi-scale environment-instant obstacle avoidance path set can include one or more of an instantaneous reaction type path subset, a short-time buffer type path subset, and a middle-range adjustment type path subset. It can be understood that the multi-scale environment-instant obstacle avoidance path set, as one of the inputs of the spatiotemporal consistency fusion, needs to be coordinate-aligned and time sequence-matched with the prospective topology path set.

[0117] Further, inputting the real-time perception data and the multi-scale environmental feature stream into the instant obstacle avoidance path planning channel can be activating multiple parallel obstacle avoidance strategy calculation units through the local planning module to generate differentiated detour schemes. For example, this operation can generate candidate paths through a multi-hypothesis tracking framework or run a dynamic window method with different parameters in parallel to produce path variants, thereby improving the diversity and adaptability of local path generation and enhancing the ability to cope with complex dynamic obstacles.

[0118] Inputting the real-time perception data and the multi-scale environmental feature stream into the prospective topology path planning channel obtains a multi-scale environmental-prospective topology path set.

[0119] The multi-scale environmental-prospective topology path set can be a set containing multiple global candidate paths, which can be used to provide long-term inspection route options covering different optimization objectives. Further, the multi-scale environmental-prospective topology path set can combine the semantic information of the registered environment model library and the long-term evolution mode to generate path combinations using a multi-objective search algorithm. For example, the multi-scale environmental-prospective topology path set can include one or more of the shortest path subset, the low-energy path subset, the high-safety path subset, etc. It can be understood that the multi-scale environmental-prospective topology path set participates in the Pareto front screening process and needs to be combined with the instant obstacle avoidance path set to form a multi-objective decision input.

[0120] Further, inputting the real-time perception data and the multi-scale environmental feature stream into the prospective topology path planning channel can be starting a multi-round graph search process through the global planner, focusing on different optimization objectives to generate path subsets in each round. For example, this operation can traverse different weight combinations through a weighted A* algorithm or introduce Monte Carlo tree search to sample potential long-period inspection sequences, thereby expanding the breadth of global path exploration and improving the probability of discovering high-quality paths.

[0121] According to the multi-scale environmental-instant obstacle avoidance path set and the multi-scale environmental-prospective topology path set, spatiotemporal consistency fusion and Pareto front screening are performed to output a dynamic navigation path optimization report for the inspection robot.

[0122] The spatio-temporal consistency fusion can be an operation mechanism of aligning and integrating path results under a unified spatio-temporal reference, and can be used to eliminate spatio-temporal conflicts of local and global paths. Further, the spatio-temporal consistency fusion can synchronize path segments through timestamp interpolation and spatial coordinate transformation techniques. For example, the spatio-temporal consistency fusion can use one or more of path point-level fusion, time interval fusion, and event-triggered fusion. It can be understood that the spatio-temporal consistency fusion is applied to the multi-scale environment-immediate obstacle avoidance path set and the multi-scale environment-prospective topological path set, and the output result is used as the basis input for Pareto frontier screening.

[0123] The Pareto frontier screening can be a multi-objective optimization decision method, and can be used to select a non-dominated solution set from candidate paths. Further, the Pareto frontier screening can construct a target function space including safety, path length, energy consumption, and turning frequency, and compare the dominance relationship between paths. For example, the Pareto frontier screening can use one or more of dominance ordering-based screening, crowding degree calculation-based screening, and reference point guide-based screening. It can be understood that the Pareto frontier screening takes the path results after spatio-temporal consistency fusion as input, and the output result is used to generate a final path combination.

[0124] Further, the spatio-temporal consistency fusion and the Pareto frontier screening according to the multi-scale environment-immediate obstacle avoidance path set and the multi-scale environment-prospective topological path set can be to first align the path sets in time and space, and then map them to a multi-dimensional target space for dominance relationship analysis. For example, the operation can be implemented by a sliding time window mechanism for path segment fusion or a hierarchical screening process, so as to realize the coordination and unification of local and global paths, and complete intelligent optimization under multi-objective trade-off, and output a comprehensive path suggestion with decision transparency.

[0125] Taking the night inspection of a chemical plant as an example, the navigation path optimization method for the inspection robot based on environment modeling can be as follows: the system loads a plant layout map as scene attribute information, combines real-time perception data collected by a laser radar and a thermal imaging camera, predicts a personnel work cycle and a high-temperature area diffusion trend, and forms a multi-scale environment feature flow; an instant obstacle avoidance path planning channel generates multiple emergency avoidance paths based on the latest point cloud and instantaneous disturbance features, and forms a multi-scale environment-instant obstacle avoidance path set; a prospective topology path planning channel re-plans an inspection sequence according to equipment importance and historical failure rates, and outputs a multi-scale environment-prospective topology path set containing characteristics such as low energy consumption and few stops; then, a spatiotemporal consistency fusion mechanism is used to align the time nodes and spatial positions of the two types of paths, and eliminate the path jumping problem caused by planning delay; then, a Pareto front screening method is used to select non-dominated paths that meet the safety distance requirements and take into account the overall time consumption and power consumption from the fusion results; and finally, a dynamic navigation path optimization report is generated, which not only recommends an optimal execution path, but also labels high-risk intersection areas and provides two backup routes, significantly improving the autonomous decision-making ability and task reliability of the robot in a low-visibility and high-dynamic environment.

[0126] The navigation path optimization method for the inspection robot based on environment modeling provided in this embodiment can be as follows: real-time perception data and a multi-scale environment feature flow are respectively input into a double-channel structure, so that the instant obstacle avoidance path planning channel can generate a multi-scale environment-instant obstacle avoidance path set containing multiple response strategies based on current dynamic disturbances, ensuring that the robot has diversified avoidance capabilities in the face of sudden obstacles; at the same time, the prospective topology path planning channel generates a multi-scale environment-prospective topology path set covering different optimization objectives in combination with long-term evolution prediction, supporting continuous optimization of global inspection efficiency; on this basis, a spatiotemporal consistency fusion mechanism is used to align the time sequence and unify the spatial coordinates of the two path sets, solving the problem of path segment misplacement caused by planning period differences or perception delay, and ensuring that local actions and global plans are coordinated and consistent; further, a Pareto front screening method is used to perform multi-dimensional non-dominated sorting on candidate paths without presetting weights, and automatically identify the path combination with the best overall performance; finally, the output dynamic navigation path optimization report contains not only the main path, but also alternative solutions and risk prompts, realizing a closed-loop decision-making process from local response to global planning to intelligent optimization, and enhancing the adaptability, robustness and decision-making transparency of the system in complex industrial scenarios.

[0127] In addition, to achieve the above object, the present application also provides a navigation path optimization system for an inspection robot based on environment modeling, which comprises: An instruction obtaining module is configured to obtain a navigation path optimization instruction for the inspection robot, wherein the navigation path optimization instruction comprises scene attribute information corresponding to a target inspection scene and real-time perception data. An environment prediction module is configured to perform dynamic environment feature prediction on the target inspection scene according to the scene attribute information, and obtain a multi-scale environment feature flow. A registration optimization module is configured to perform spatio-temporal registration optimization on a multi-source environment modeling database according to the scene attribute information and the real-time perception data, and determine a registered environment model library. A channel building module is configured to introduce a plurality of path planning learners and a path optimization loss parser to perform path optimization loss optimization learning on the registered environment model library, and build a navigation path double channel for the inspection robot, wherein the navigation path double channel comprises an instant obstacle avoidance path planning channel and a forward-looking topological path planning channel. A report generation module is configured to input the real-time perception data and the multi-scale environment feature flow into the navigation path double channel, and generate a dynamic navigation path optimization report for the inspection robot.

[0128] Other embodiments or specific implementations of the inspection robot navigation path optimization system based on environment modeling according to the present application can refer to the above-mentioned method embodiments, which will not be described here.

[0129] In addition, in order to achieve the above-mentioned purpose, the present application also provides an inspection robot navigation path optimization device based on environment modeling, which comprises a memory, a processor and an inspection robot navigation path optimization program based on environment modeling stored on the memory and executable on the processor, wherein the inspection robot navigation path optimization program based on environment modeling is configured to implement the steps of the inspection robot navigation path optimization method based on environment modeling according to any one of the above embodiments.

[0130] In addition, in order to achieve the above-mentioned purpose, the present application also provides a medium having an inspection robot navigation path optimization program based on environment modeling stored thereon, wherein the inspection robot navigation path optimization program based on environment modeling is executed by a processor to implement the steps of the inspection robot navigation path optimization method based on environment modeling according to any one of the above embodiments.

[0131] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent flow transformation according to the content of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for optimizing the navigation path of an inspection robot based on environmental modeling, characterized in that, The method includes: Obtain navigation path optimization instructions for the inspection robot, wherein the navigation path optimization instructions include scene attribute information and real-time perception data corresponding to the target inspection scene; Based on the scene attribute information, dynamic environmental feature prediction is performed on the target inspection scene to obtain a multi-scale environmental feature stream. Based on the scene attribute information and the real-time perception data, spatiotemporal registration optimization is performed on the multi-source environment modeling database to determine the registration environment model library; Multiple path planning learners and path optimization loss parsers are introduced to perform path optimization loss optimization learning on the registered environment model library, and a dual-channel navigation path is built for the inspection robot. The dual-channel navigation path includes an instantaneous obstacle avoidance path planning channel and a forward-looking topology path planning channel. The real-time sensing data and the multi-scale environmental feature stream are input into the dual channels of the navigation path to generate a dynamic navigation path optimization report for the inspection robot.

2. The inspection robot navigation path optimization method based on environment modeling as described in claim 1, characterized in that, The step of dynamically predicting environmental features of the target inspection scene based on the scene attribute information to obtain a multi-scale environmental feature stream includes: Based on the scene attribute information, spatial functional zoning prediction is performed on the target inspection scene to determine the key inspection area set; Collect structural semantic information, dynamic obstacle density information, and time-varying illumination information of the key inspection area set to obtain a target environment feature library; Based on the target environment feature library, spatial scale classification is performed to obtain initial multi-scale environment classification results; Dynamic grading is performed based on the initial classification results of the multi-scale environment to obtain the feature flow of the multi-scale environment.

3. The method for optimizing the navigation path of an inspection robot based on environmental modeling as described in claim 1, characterized in that, The step of performing spatiotemporal registration optimization on the multi-source environment modeling database based on the scene attribute information and the real-time perception data to determine the registration environment model library includes: Based on the scene attribute information, the multi-source environment modeling database is filtered by scene type association to obtain a homogeneous scene environment model library; Based on the real-time perception data, the spatiotemporal twinning degree of each model in the isomorphic scene environment model library is evaluated to obtain multiple model registration coefficients. The multiple model registration coefficients are optimized and selected based on the model registration threshold to obtain a model registration optimization distribution that is greater than or equal to the model registration threshold. Based on the model registration optimization distribution, the isomorphic scene environment model library is dynamically trimmed and merged to generate the registration environment model library.

4. The method for optimizing the navigation path of an inspection robot based on environmental modeling as described in claim 1, characterized in that, The process involves introducing multiple path planning learners and path optimization loss parsers to perform path optimization loss optimization learning on the registered environment model library, thereby establishing a dual-channel navigation path for the inspection robot, including: Historical path logs are retrieved based on the registered environment model library to obtain an instant obstacle avoidance path log library and a forward-looking topology path log library. Environmental disturbance features are collected based on the real-time obstacle avoidance path log library to obtain a first disturbance feature record library; Based on the aforementioned prospective topology path log library, long-term task constraint features are collected to obtain a second constraint feature record library. Based on the multiple path planning learners and the path optimization loss parser, the real-time obstacle avoidance path optimization loss optimization learning is performed on the registration environment model library, the first perturbation feature record library and the real-time obstacle avoidance path log library to generate the real-time obstacle avoidance path planning channel. Based on the multiple path planning learners and the path optimization loss parser, the prospective topology path optimization loss optimization learning is performed on the registration environment model library, the second constraint feature record library and the prospective topology path log library to generate the prospective topology path planning channel. The real-time obstacle avoidance path planning channel and the forward-looking topology path planning channel are connected as parallel nodes to generate the dual-channel navigation path for the inspection robot.

5. The inspection robot navigation path optimization method based on environmental modeling as described in claim 4, characterized in that, The step of performing real-time obstacle avoidance path optimization loss optimization learning on the registration environment model library, the first perturbation feature record library, and the real-time obstacle avoidance path log library based on the multiple path planning learners and the path optimization loss parser to generate the real-time obstacle avoidance path planning channel includes: Using the registration environment model library and the first perturbation feature record library as input information, and the real-time obstacle avoidance path log library as output label, supervised training is performed on the multiple path planning learners respectively. During each predetermined number of training iterations, multiple real-time path optimization loss coefficients are calculated based on the path optimization loss parser. If the multiple real-time path optimization loss coefficients are less than the real-time path optimization loss threshold, multiple real-time obstacle avoidance path planning models are generated. By connecting the multiple real-time obstacle avoidance path planning models, the real-time obstacle avoidance path planning channel is obtained.

6. The method for optimizing the navigation path of an inspection robot based on environmental modeling as described in claim 1, characterized in that, The path optimization loss parser includes a path optimization loss parsing function, which is: in, The path optimization loss coefficient is represented by N, where N represents the predetermined number of iterations and n represents the nth iteration. Both N and n are positive integers, and 1 ≤ n ≤ N. This represents the number of path collision events that occur in the nth iteration. Characterizes the number of path acceleration exceeding limits in the nth iteration; The average deviation distance of the path from the preset topology skeleton line in the nth iteration is represented; λ and μ are non-negative weight coefficients used to balance safety, smoothness and structural constraints.

7. The method for optimizing the navigation path of an inspection robot based on environmental modeling as described in claim 1, characterized in that, The step of inputting the real-time sensing data and the multi-scale environmental feature stream into the dual channels of the navigation path to generate a dynamic navigation path optimization report for the inspection robot includes: The real-time sensing data and the multi-scale environmental feature stream are input into the real-time obstacle avoidance path planning channel to obtain a multi-scale environment-real-time obstacle avoidance path set; The real-time sensing data and the multi-scale environmental feature stream are input into the forward-looking topology path planning channel to obtain a multi-scale environment-forward-looking topology path set. Based on the multi-scale environment-instantaneous obstacle avoidance path set and the multi-scale environment-forward-looking topological path set, spatiotemporal consistency fusion and Pareto front screening are performed to output the dynamic navigation path optimization report of the inspection robot.

8. A navigation path optimization system for inspection robots based on environmental modeling, characterized in that, The system includes: The instruction acquisition module is used to obtain navigation path optimization instructions for the inspection robot, wherein the navigation path optimization instructions include scene attribute information and real-time perception data corresponding to the target inspection scene; The environment prediction module is used to perform dynamic environmental feature prediction on the target inspection scene based on the scene attribute information to obtain a multi-scale environmental feature stream. The registration optimization module is used to perform spatiotemporal registration optimization on the multi-source environment modeling database based on the scene attribute information and the real-time perception data, and to determine the registration environment model library. A channel module is built to introduce multiple path planning learners and path optimization loss parsers to perform path optimization loss optimization learning on the registration environment model library, and to build a dual-channel navigation path for the inspection robot, wherein the dual-channel navigation path includes an instantaneous obstacle avoidance path planning channel and a forward-looking topology path planning channel. The report generation module is used to input the real-time perception data and the multi-scale environmental feature stream into the dual channels of the navigation path to generate a dynamic navigation path optimization report for the inspection robot.

9. A navigation path optimization device for inspection robots based on environmental modeling, characterized in that, The device includes: a memory, a processor, and an environment-modeling-based inspection robot navigation path optimization program stored in the memory and executable on the processor, the environment-modeling-based inspection robot navigation path optimization program being configured to implement the steps of the environment-modeling-based inspection robot navigation path optimization method as described in any one of claims 1 to 8.

10. A medium, characterized in that, The medium stores an inspection robot navigation path optimization program based on environment modeling. When the inspection robot navigation path optimization program based on environment modeling is executed by the processor, it implements the steps of the inspection robot navigation path optimization method based on environment modeling as described in any one of claims 1 to 8.

Images