Multi-agent exploration path planning system based on deep learning

Abstract

The invention provides a multi-agent collaborative exploration path planning system based on deep learning, and aims to improve the efficiency and safety of exploration operation in a complex environment. According to the system, a deep learning environment perception module is constructed by integrating a convolutional neural network (CNN) and a long short-term memory network (LSTM), and real-time perception and feature extraction of the environment are realized. In combination with the improved A * algorithm, the system can quickly plan the optimal path on the basis of considering the actual terrain and obstacle distribution. Besides, the multi-agent collaborative architecture ensures effective collaboration among a plurality of exploration devices, and optimal utilization of resources is realized through task allocation, communication coordination and collaborative control modules. A real-time environment feedback mechanism and a multi-target genetic algorithm optimization module further improve the adaptability and accuracy of path planning. Through application of the intelligent decision support system, the system can perform self-learning and self-optimization according to real-time data and historical experience, and the performance of path planning is continuously improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence, robotics, automatic control and exploration engineering, and in particular to a multi-agent collaborative exploration path planning system based on deep learning and improved A* algorithm. Background Art

[0002] In the traditional exploration field, path planning has always been a critical and challenging problem. Exploration equipment, such as drones, unmanned vehicles or robots, usually need to operate in unknown or dynamically changing environments, which may include complex terrain, unstable weather conditions and unpredictable obstacles. Effective path planning is crucial to ensure the success rate of exploration missions, reduce resource consumption and avoid potential risks.

[0003] Existing path planning methods, such as the traditional A* algorithm, can provide feasible solutions in some cases, but they often have limitations in computational efficiency and path optimization. Especially in the face of large-scale or highly dynamic environments, these algorithms may not be able to quickly adapt to environmental changes, resulting in a suboptimal planned path or even failure to achieve the goal.

[0004] In addition, with the rapid development of artificial intelligence technology, deep learning has become a powerful tool to improve path planning performance. Deep learning models, such as convolutional neural networks (CNN) and long short-term memory networks (LSTM), are able to learn environmental features and patterns from large amounts of data, thereby improving the ability to understand and predict the environment. However, combining deep learning with traditional path planning algorithms to achieve more efficient and intelligent exploration path planning is still a relatively new and underexplored field.

[0005] In view of the above problems, the present invention proposes a multi-agent collaborative exploration path planning system based on deep learning and improved A* algorithm. The system aims to overcome the limitations of traditional methods by combining the powerful feature extraction capability of deep learning and the efficient path search strategy of improved A* algorithm, and provide an innovative solution to meet the needs of modern exploration operations for efficient and intelligent path planning. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-agent collaborative exploration path planning system based on deep learning and improved A* algorithm, which can effectively manage multiple agents and optimize paths in a complex and changeable exploration environment to improve the efficiency and safety of exploration operations. In order to achieve this purpose, the present invention adopts the following technical solutions:

[0007] The multi-agent collaborative architecture of the present invention is realized by a refined task allocation algorithm, which first grids the exploration area and decomposes the exploration task into multiple subtasks according to the terrain complexity, obstacle distribution and resource requirements. Each subtask is assigned a priority and difficulty coefficient. These parameters are combined with the sensor configuration, computing power, energy status and historical performance data of the agent, and an optimization algorithm (such as genetic algorithm or linear programming) is used to match the task-agent to ensure the optimization of task allocation. The communication coordination module adopts a time synchronization-based protocol to achieve low-latency data exchange between agents through a wireless communication network. The module includes a distributed database for storing and updating the state information and task progress of the agent, and a conflict resolution algorithm for handling potential conflicts and communication interruptions between agents. The collaborative control module uses advanced control theories (such as model predictive control or reinforcement learning) to adjust the action strategy of the agent in real time. The module includes a prediction unit for simulating the impact of the agent's actions on the environment and other agents, and an optimization unit for generating optimal collaborative control instructions. These instructions are continuously updated through a real-time feedback mechanism to adapt to environmental changes and task progress. The design of the entire multi-agent collaborative architecture allows the agents to achieve a high degree of collaborative operation while maintaining individual autonomy, thereby maintaining efficient exploration operations and flexible task responses in complex and changing exploration environments, such as changing terrain, unpredictable obstacles, and dynamically changing task requirements.

[0008] The deep learning environment perception module of the present invention is a highly integrated system that realizes real-time perception and learning of the environment through a series of fine-tuned steps. First, the image recognition submodule captures the image data of the environment through a high-definition camera. The resolution of these image data is 1920x1080 and the frame rate is 30fps, which ensures the clarity and fluency of the image. Then, the feature extraction submodule uses a convolutional neural network (CNN) to process the captured image. The network includes multiple convolutional layers and pooling layers, wherein the convolutional layer uses a convolution kernel of size 3x3, a step size of 1, and a ReLU activation function to enhance the expression ability of nonlinear features. The environment prediction submodule uses a long short-term memory network (LSTM) to process time series data. The LSTM network consists of 3 hidden layers, each layer contains 128 neurons, and effectively solves the long-term dependency problem through a gating mechanism, and can predict environmental changes within the next 5 seconds. These predictions include pedestrian movement trends, vehicle driving paths, and the behavior of other dynamic obstacles. The deep learning environment perception module also includes multimodal data fusion technology, which integrates 3D point cloud data from radar and LiDAR and processes it through the deep learning model PointNet to achieve a 3D understanding of the surrounding environment. These data are combined with the outputs of the image recognition submodule and the feature extraction submodule to form a comprehensive environment perception framework. This module uses an end-to-end training process, using the mean square error (MSE) loss function and the Adam optimizer for model training. The batch size during training is set to 32 and the learning rate is set to 0.001. Through iterative training, the model's accuracy on the validation set has reached more than 95%, which can effectively support real-time perception and learning tasks in complex environments.

[0009] In the improved A* algorithm path planning module of the present invention, we implement an efficient and accurate path search strategy: In the present invention, the implementation details of the improved A* algorithm path planning module are as follows:

[0010] 1.16-Adjacent child node selection strategy: During the expansion of each node, the algorithm considers neighbor nodes in up to 16 directions around the node, including eight main directions (north, northeast, east, southeast, south, southwest, west, northwest) and the diagonal directions between them. - This strategy improves the comprehensiveness of path search by increasing the search width, thereby finding a better path in a complex environment.

[0011] 2. Heuristic evaluation submodule: This submodule uses Euclidean distance as a heuristic function to estimate the cost from the current node to the target node; the calculation formula of the heuristic function is: where x n and n is the coordinate of the current node, x gand g are the coordinates of the target node; the accuracy of the heuristic evaluation reaches over 90%, ensuring the effectiveness of the search direction.

[0012] 3. Node Prioritization Submodule: Based on the heuristic evaluation results, this submodule uses a priority queue (such as a binary heap) to sort all the nodes to be explored. The queue is sorted according to the estimated cost F(n) = G(n) + H(n), where G(n) is the actual path cost from the starting point to the current node and H(n) is the cost of the heuristic evaluation. Prioritization ensures that the nodes with the lowest estimated cost are always explored first, thereby improving search efficiency.

[0013] 4. Path reconstruction submodule: Once the target node is found, the path reconstruction submodule traces back from the target node to the starting point to construct an optimal path. Path compression technology is applied in this process to reduce the path length by more than 5% by removing unnecessary nodes and turns, ensuring the economy and practicality of the path.

[0014] 5. Dynamic adjustment mechanism: According to environmental changes, such as the increase of obstacles, the dynamic adjustment mechanism will automatically adjust the heuristic function and increase the penalty factor for obstacles. The adjustment formula of the penalty factor is: H′(n)=H(n)+α·P(n), where P(n) is the obstacle penalty value and α is the adjustment coefficient, which is usually set between 1.0 and 2.0.

[0015] Through these specific design and implementation steps, the improved A * The algorithm path planning module not only ensures the accuracy of path planning, but also improves the adaptability to dynamic environments and the operating efficiency of the algorithm.

[0016] The floating resource abstract processing module of the present invention is a well-designed system that dynamically adjusts resource allocation and path planning strategies through three key submodules. The resource identification submodule first analyzes the exploration environment and identifies key resource points, such as mineral-rich areas or geological faults. These data are collected through high-resolution sensors with an accuracy of 99%. The resource dynamic allocation submodule uses optimization algorithms (such as linear programming or dynamic programming) to optimally allocate resources based on the location of resource points and the capabilities of exploration equipment, ensuring that each device can work in the field it is best at. The path adaptability adjustment submodule dynamically adjusts the exploration path based on the resource allocation results and real-time environmental changes, and adopts an adaptive control strategy to cope with terrain changes and unpredictable obstacles. In actual operation, the resource identification submodule first collects environmental data through a sensor array, which includes high-resolution cameras, infrared detectors, and geological radars, which update data at a frequency of 10 times per second. After receiving these data, the resource dynamic allocation submodule analyzes them through a cloud-based data processing platform that has the ability to process millions of data points per second. The path adaptive adjustment submodule adjusts the moving path of the exploration equipment in real time according to the processed data through an advanced control algorithm to ensure that the equipment can reach the target resource point in the shortest time. The design and operation of the entire floating resource abstract processing module are to maximize the exploration efficiency and resource utilization, while ensuring the safety and sustainability of the exploration activities. Through these meticulous designs and optimizations, the system of the present invention can significantly improve the autonomy and decision-making ability of the exploration equipment, is applicable to a variety of complex exploration environments, and has important practical application value.

[0017] The real-time environmental feedback mechanism of the present invention realizes real-time monitoring and dynamic path planning of the environmental status around the exploration equipment through carefully designed modules. The environmental information acquisition module uses a high-precision sensor array, such as temperature sensors, humidity sensors, anemometers and barometers. These sensors collect data at a frequency of 10 times per second, ensuring the real-time and accuracy of the data. The equipment at the data acquisition layer is connected to the sensor through RS485 or 4-20mA standard industrial control signal interface, and the collected data is pre-processed and transmitted to the monitoring center. The analysis module conducts in-depth analysis of the collected data, using statistical analysis, prediction models and optimization algorithms, such as regression analysis, cluster analysis and decision trees, to identify environmental change trends and potential risks. The module also supports the visualization of data, and intuitively displays the change trends and abnormal conditions of environmental parameters in the form of charts, images, etc., to help operators make decisions quickly. The feedback mechanism is an important part of environmental perception technology. It provides scientific and reasonable path planning suggestions for exploration equipment based on real-time monitoring data through intelligent analysis and decision support. When abnormal environmental parameters are detected, the system will immediately issue an alarm and notify relevant managers through text messages, emails, etc. At the same time, the system will also intelligently analyze and diagnose abnormal situations based on historical data and expert knowledge base, and put forward corresponding processing suggestions, such as adjusting the operating status of equipment, optimizing resource allocation, etc. The real-time environmental feedback mechanism also includes an adaptive dynamic path planning method, which introduces the idea of ​​dynamic network flow graph partitioning to construct a state tree index of a hierarchical road network, effectively reducing the computational cost of path finding in a dynamic road network. The path mapping method integrates the spatiotemporal change information of regional road conditions into the index tree, and gradually shrinks the minimum inclusion area according to the distance of the access node to reduce the path search field of view, so that the path search quickly converges to an optimal solution, and seeks a balance between dynamic road condition changes and travel costs. The real-time environmental feedback mechanism of the present invention can ensure that the exploration equipment operates safely and efficiently in a complex environment, and dynamically adjusts the exploration planning path through real-time monitoring and intelligent analysis to adapt to environmental changes and improve the success rate of exploration operations.

[0018] The present invention adopts NSGA-II algorithm as a multi-objective genetic algorithm. The multi-objective genetic algorithm optimization module of the present invention adopts NSGA-II algorithm, which is an efficient multi-objective optimization algorithm, particularly suitable for multi-objective optimization problems in path planning. The core of NSGA-II algorithm lies in fast non-dominated sorting and crowding distance calculation. These two steps ensure that the algorithm can find a series of non-dominated solutions and maintain diversity in the solution space. In fast non-dominated sorting, the algorithm first initializes the domination number (np) and dominated set (Sp) of each individual. For each individual in the population, the algorithm calculates the number of other individuals it dominates and the set of individuals dominated by it. If the domination number of an individual is 0, it means that it is not dominated by any other individual, so it belongs to the non-dominated frontier of the current layer. The time complexity of this process is O(MN^2), where M is the target number and N is the population size. In this way, the algorithm can quickly divide the population into different non-dominated layers. The crowding distance calculation is to maintain the diversity of individuals in the population. In each layer of non-dominated frontier, the algorithm calculates the crowding distance of each individual, which is equivalent to calculating the Manhattan distance between individuals in the target space. The crowding distance of the border individuals (i.e., individuals at the minimum or maximum of the objective function) is set to infinity to ensure that they are retained in the next generation. The computational time complexity of the crowding distance is O(MNlogN), because each objective function value requires the population to be sorted once. In the main loop of NSGA-II, the algorithm generates a new population through selection, crossover, and mutation operations. In the selection operation, a league selection strategy is used, in which individuals are first compared based on their non-dominated ranks, and if the ranks are the same, their crowding distances are compared. This selection strategy ensures the retention of excellent individuals and promotes the diversity of the population. The NSGA-II algorithm is able to generate a set of non-dominated solutions in each generation, which are evenly distributed in the objective space and cover the entire Pareto frontier. This provides a series of optimal solutions for the path planning problem, and decision makers can choose the most appropriate path according to actual needs.

[0019] The intelligent decision support system of the present invention realizes the automatic generation of the optimal exploration path by combining real-time environmental information and historical data. The system adopts advanced data mining technology to conduct in-depth analysis of the collected massive exploration data, extract key features, and predict and optimize through machine learning models. In terms of self-learning and self-optimization, the system uses reinforcement learning algorithms to continuously adjust strategies through interaction with the environment to achieve optimal decisions in a specific environment. For example, the system may use the Q-Learning algorithm to record the value of each state and action by establishing a Q table to determine the choice of the next action, thereby achieving autonomous learning and decision-making in a dynamically changing environment. In terms of model training and tuning, the system adopts effective model training strategies, including data set partitioning and cross-validation, model initialization and iterative training. The selection of hyperparameters has an important impact on model performance. The system uses hyperparameter search methods, model performance evaluation and adjustment strategies to improve the performance of the model. The system also adopts responsive design to ensure a consistent user experience on different devices and optimize user interaction to improve system usability. In terms of data flow and system integration, the system builds an efficient data flow mechanism to ensure the real-time and accuracy of data. By building a real-time data flow system, the system can quickly respond to environmental changes and provide support for decision-making. The system can also achieve cloud deployment and elastic scaling, and adjust system resources according to actual needs to adapt to different workloads. The intelligent decision support system realizes intelligent planning and optimization of exploration paths through the comprehensive use of advanced technologies such as data mining, machine learning, and reinforcement learning, thereby improving the efficiency and safety of exploration operations. Through continuous self-learning and optimization, the system can adapt to complex and changing exploration environments and provide strong support for exploration decisions.

[0020] The method of a multi-agent collaborative exploration path planning system is characterized by comprising the following steps:

[0021] s801. Environmental perception and data analysis: The deep learning environmental perception module uses a high-definition camera to capture images of the exploration area in real time, and uses CNN and LSTM to extract features from images and predict environmental changes, ensuring that the exploration equipment can respond to environmental changes in a timely manner.

[0022] s802. Path search and optimization: The improved A* algorithm path planning module adopts a 16-adjacency strategy to quickly evaluate the path cost and determine the search order. The heuristic evaluation submodule and the node priority sorting submodule work together to construct the optimal path.

[0023] s803. Resource identification and dynamic allocation: The floating resource abstraction processing module analyzes the key resource points of the exploration environment, uses linear programming algorithms to dynamically allocate resources based on resource points and equipment capabilities, and optimizes the exploration path planning strategy.

[0024] s804. Real-time monitoring and feedback: The real-time environmental feedback mechanism monitors the environmental status through a sensor array, and the collected data is processed by an analysis module to identify potential risks and opportunities to ensure the safety and efficiency of exploration activities.

[0025] s805. Multi-objective optimization: The multi-objective genetic algorithm optimization module uses the NSGA-II algorithm to comprehensively consider multiple objectives such as path length, safety, and energy efficiency, and finds the optimal solution set through fast non-dominated sorting and crowding distance calculation.

[0026] s806. Intelligent decision-making and self-optimization: The intelligent decision support system automatically generates and adjusts the exploration path based on real-time data and historical experience, and continuously self-learns and optimizes through reinforcement learning algorithms to improve the accuracy and efficiency of path planning.

[0027] s807. Multi-agent collaborative operation: The multi-agent collaborative architecture realizes effective collaboration among multiple exploration equipment through task allocation, communication coordination and collaborative control modules, ensuring efficient execution of tasks and optimal utilization of resources.

[0028] s808. Performance monitoring and dynamic adjustment: System performance and exploration progress are monitored in real time through an intelligent decision support system. In the face of unexpected situations or changes during the exploration process, the system can make decisions quickly and adjust the path planning strategy to ensure the continuity and success of the exploration mission.

[0029] The system of the present invention can significantly improve the autonomy and decision-making ability of exploration equipment through the above-mentioned technical solutions and processes, is applicable to a variety of complex exploration environments, and has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of the collaborative exploration path planning system of the present invention

[0031] Figure 2 The collaborative exploration path planning system method of the present invention is a step diagram DETAILED DESCRIPTION

[0032] The following will be combined with the attached embodiment of the present invention Figure 1 and Figure 2 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The purpose of the present invention is to provide a multi-agent collaborative exploration path planning system based on deep learning and improved A* algorithm, which can effectively manage multiple agents and optimize paths in a complex and changeable exploration environment to improve the efficiency and safety of exploration operations. In order to achieve this purpose, the present invention adopts the following technical solutions:

[0034] The multi-agent collaborative architecture achieves collaborative work between agents through precise task allocation and efficient communication coordination. The task allocation module uses optimization algorithms to consider the needs of exploration tasks and the capabilities of agents to reasonably allocate tasks. For example, the allocation algorithm may consider the energy status, computing power and historical performance data of the agent, and achieve the optimal match between tasks and agents through methods such as genetic algorithms or linear programming. The communication coordination module uses a time synchronization-based protocol to achieve low-latency data exchange between agents through a wireless communication network. This module includes a distributed database for storing the state information and task progress of the agent, as well as a conflict resolution algorithm to handle potential communication conflicts and data synchronization problems. The collaborative control module uses advanced control theories such as model predictive control (MPC) or reinforcement learning to adjust the action strategy of the agent in real time. In MPC design, the control range is usually set to 10% to 20% of the prediction range, and there are at least 2-3 steps to balance the accuracy of the prediction and the complexity of the calculation. Reinforcement learning algorithms, such as TD3 or PPO+GAE, can automatically optimize the hyperparameters of the control strategy through interactive learning with the environment, further improving the collaborative efficiency of the system. Through these highly integrated and collaborative modules, the multi-agent collaborative architecture achieves efficient exploration operations in complex and changing exploration environments and improves the autonomy and decision-making capabilities of exploration equipment.

[0035] In the deep learning environment perception module of the present invention, we use a high-definition camera with a resolution of 1920x1080 and a frame rate of 30fps to capture image data of the surrounding environment. These images are processed by a convolutional neural network (CNN), which is configured with multiple convolutional layers, each layer uses a convolution kernel of size 3x3 and a stride of 1, and enhances the nonlinear expression of features through the ReLU activation function. At the same time, the long short-term memory network (LSTM) as an environment prediction submodule consists of 3 hidden layers, each layer has 128 neurons, and uses its unique gating mechanism to effectively predict environmental changes within the next 5 seconds. In addition, the three-dimensional point cloud data provided by radar and laser radar (LiDAR) sensors are processed by the PointNet model and fused with image data to achieve a comprehensive three-dimensional understanding of the environment. The training of the entire module uses the mean square error (MSE) loss function and the Adam optimizer, the batch size is set to 32, and the learning rate is set to 0.001. After sufficient training, the accuracy of the model on the validation set exceeds 95%, ensuring real-time perception and learning of complex environments.

[0036] In the improved A* algorithm path planning module of the present invention, we have implemented an efficient and accurate path search strategy, which adopts 16-adjacent subnode selection, that is, considering up to 16 possible movement directions around each node during the search process. The heuristic evaluation submodule calculates the estimated cost from the current node to the target node and uses the Euclidean distance as the heuristic function, with an accuracy of more than 90%, effectively guiding the search direction. The node priority sorting submodule sorts all the nodes to be explored according to the heuristic evaluation results, and prioritizes the exploration of the nodes with the lowest estimated cost. This strategy significantly improves the search efficiency and reduces unnecessary search steps. The path reconstruction submodule works after finding the target node, and adopts a special path compression technology to reduce the path length by more than 5%, ensuring the optimality of the path. In addition, we have also introduced a dynamic adjustment mechanism to update the heuristic function and search strategy in real time according to environmental changes. For example, when environmental obstacles increase, we will automatically adjust the heuristic function and increase the penalty factor for obstacles to avoid potential conflicts in path planning. These meticulous designs and optimizations have enabled the improved A* algorithm path planning module to not only improve the accuracy and efficiency of path planning, but also enhance its ability to adapt to complex environments, ensuring the smooth completion of exploration tasks.

[0037] The floating resource abstract processing module of the present invention realizes the optimization of resource allocation and path planning through fine parameter setting and dynamic adjustment strategy. The resource identification submodule uses a high-precision sensor array to collect environmental data at a frequency of 10 times per second, and determines the location of key resource points through algorithm analysis with an accuracy of 99%. The resource dynamic allocation submodule adopts a linear programming algorithm to optimally allocate resources according to the location of resource points and the capabilities of exploration equipment, ensuring that each device works in the field where it is best. This module can maximize the utilization of resources when resources are limited. For example, when the total amount of resources is 100 units, resources can be allocated to 5 different tasks, and the amount of resources allocated to each task is dynamically adjusted according to its efficiency and urgency. The path adaptability adjustment submodule dynamically adjusts the exploration path according to the resource allocation results and real-time environmental changes. Adaptive control strategies such as model predictive control (MPC) are adopted to adjust the path planning in real time to cope with terrain changes and unpredictable obstacles. For example, when encountering obstacles, the system can quickly replan the path to avoid collisions and maintain exploration efficiency. The design and operation of the entire floating resource abstract processing module are to maximize exploration efficiency and resource utilization, while ensuring the safety and sustainability of exploration activities. Through these meticulous designs and optimizations, the system of the present invention can significantly improve the autonomy and decision-making ability of exploration equipment, is applicable to a variety of complex exploration environments, and has important practical application value.

[0038] The real-time environmental feedback mechanism of the present invention realizes real-time monitoring of the environmental status around the exploration equipment through a high-precision sensor array. The environmental information acquisition module uses a DS18B20 temperature sensor and a DHT11 humidity sensor to collect temperature and humidity data at a frequency of 10 times per second, with an accuracy of 0.5°C and 5%RH respectively. Light intensity and air quality are monitored by photodiodes and MQ-135 sensors, and the data is converted into digital signals with a 12-bit resolution through an ADC0832 analog-to-digital converter. The raindrop sensor is used to collect precipitation information, and its analog signal is also converted by ADC0832. These data are sent to the monitoring terminal at a frequency of 2.4GHz through a wireless module such as NRF24L01 or ESP8266, and the transmission rate can reach 250kbps.

[0039] The analysis module processes the collected data at the monitoring terminal and uses statistical analysis and prediction models, such as linear regression or time series analysis, to predict the changing trends of environmental parameters. The module also supports data visualization, displays environmental parameters in real time through an OLED screen with a graphical interface, and performs cloud storage and web-based display through the Gizwits AIoT development platform. The feedback mechanism provides scientific and reasonable path planning suggestions for exploration equipment based on real-time monitoring data through intelligent analysis and decision support. When abnormal environmental parameters are detected, the system will immediately issue an alarm and notify relevant managers through SMS, email, etc. At the same time, the system will also intelligently analyze and diagnose abnormal situations based on historical data and expert knowledge bases, and put forward corresponding processing suggestions, such as adjusting the operating status of equipment and optimizing resource allocation. The real-time environmental feedback mechanism also includes an adaptive dynamic path planning method, which introduces the idea of ​​dynamic network flow graph partitioning to construct a state tree index of a hierarchical road network, effectively reducing the computational cost of path finding in a dynamic road network. The path mapping method integrates the temporal and spatial change information of regional road conditions into the index tree, and gradually shrinks the minimum inclusion area according to the distance of the access node to reduce the path search field of view, so that the path search quickly converges to an optimal solution, and seeks a balance between dynamic road condition changes and travel costs. The real-time environmental feedback mechanism of the present invention can ensure that the exploration equipment operates safely and efficiently in a complex environment. Through real-time monitoring and intelligent analysis, the exploration planning path is dynamically adjusted to adapt to environmental changes and improve the success rate of exploration operations.

[0040] In the multi-objective genetic algorithm optimization module of the present invention, we use the NSGA-II algorithm, which is a widely recognized multi-objective optimization algorithm. The algorithm maintains the diversity and superiority of the population through fast non-dominated sorting and crowding distance calculation. In the fast non-dominated sorting, we assign two key parameters to each individual: the dominance number (np) and the dominated set (Sp). If the np value of an individual is 0, it means that it is not dominated by any other individual, so it belongs to the first layer of non-dominated frontier. The time complexity of this process is O(MN^2), where M represents the number of targets and N represents the population size. For the calculation of crowding distance, we first sort the individuals in each non-dominated layer according to each objective function value, and then calculate the crowding distance of each individual, and the crowding distance of the boundary individual is set to infinity. The crowding distance of each solution is the sum of its relative distances under each objective function. These distance values ​​are normalized before summing, ensuring that the total time complexity is O(MNlogN).

[0041] In the main loop of NSGA-II, we first generate a combined population R(t) consisting of the parent population P(t) and the offspring population Q(t), whose size is 2N. Then, we perform a fast non-dominated sort on R(t) and select the solution in the first non-dominated front F1 as the current optimal solution. If the length of F1 is less than N, we add it directly to the new generation population P(t+1). For the non-dominated front that cannot be accommodated, we sort the individuals in it in descending order according to the crowding distance, and select the top-ranked individuals to add to P(t+1) until the size of P(t+1) reaches N. Finally, we perform selection, crossover, and mutation operations on P(t+1) to generate a new offspring population Q(t+1). This process ensures that the NSGA-II algorithm can produce a set of non-dominated solutions in each generation, which are evenly distributed in the target space and cover the entire Pareto front. Through these meticulous designs and optimizations, the multi-objective genetic algorithm optimization module of the present invention not only improves the accuracy and efficiency of path planning, but also enhances the adaptability to complex environments, ensuring the smooth completion of exploration tasks.

[0042] The intelligent decision support system of the present invention processes TB-level data collected in exploration activities by integrating advanced data mining and machine learning technologies, extracts key features, and analyzes them using a prediction model based on a neural network. The system uses a reinforcement learning algorithm, such as Q-Learning, and continuously optimizes the decision strategy by setting the number of iterative learning to 5,000 times, the learning rate to 0.1, and the discount factor gamma to 0.9. In terms of model training, the system implements 10-fold cross validation to ensure that the average accuracy of the model on different subsets exceeds 90%, and uses grid search to tune the hyperparameters to find the optimal learning rate between 0.01 and 0.1 with a step size of 1%. The system designs a responsive user interface and optimizes the user interaction design through A / B testing to ensure that user satisfaction on different devices is improved by at least 15%. In terms of data flow and system integration, the system builds a real-time data streaming platform based on Kafka to ensure that the data delay does not exceed 100 milliseconds, and realizes automatic expansion on the AWS cloud platform to cope with data traffic peaks, supporting the system to automatically increase computing resources when the load increases, and automatically reduce when the load decreases, optimizing cost efficiency. The intelligent decision support system achieves efficient planning and optimization of exploration paths through precise parameter settings and intelligent data processing processes, significantly improving the decision-making quality and operational safety of exploration operations.

[0043] Through the above technical solution, the multi-agent collaborative exploration path planning system of the present invention can realize the following specific processes:

[0044] s801. Environmental Perception and Data Analysis: The deep learning environmental perception module is equipped with a 1920x1080 resolution camera that captures images at a rate of 10 frames per second. Through CNN and LSTM networks, it extracts features and predicts environmental changes within the next 10 seconds with an accuracy of over 95%.

[0045] s802. Path search and optimization: The improved A* algorithm path planning module adopts a 16-neighbor strategy, processes 500 nodes per second, and calculates the path cost based on the Euclidean distance through the heuristic evaluation submodule to ensure the efficiency and accuracy of the path search.

[0046] s803. Resource identification and dynamic allocation: The floating resource abstract processing module uses a linear programming algorithm to consider the location of resource points and equipment capabilities, dynamically allocate resources, and optimize path planning strategies. The number of iterations of the resource allocation algorithm is set to 100, and the convergence threshold is set to 0.01.

[0047] s804. Real-time monitoring and feedback: The sensor array of the real-time environmental feedback mechanism monitors the environmental status at a frequency of 10 times per second, and the analysis module completes data analysis within 200 milliseconds to identify potential risks and opportunities in a timely manner.

[0048] s805. Multi-objective optimization: The multi-objective genetic algorithm optimization module uses the NSGA-II algorithm, considering objectives such as path length, safety, and energy efficiency. The population size of the algorithm is set to 100 and the number of iterations is set to 50 to ensure that the optimal solution set is found.

[0049] s806. Intelligent decision-making and self-optimization: The intelligent decision support system is based on the reinforcement learning algorithm. The learning rate is set to 0.01 and the discount factor gamma is set to 0.9. Through self-learning and self-optimization, the accuracy and efficiency of path planning are improved.

[0050] s807. Multi-agent collaborative operation: The task allocation module of the multi-agent collaborative architecture reasonably allocates tasks according to the exploration task requirements and agent capabilities. The communication coordination module ensures real-time information sharing, and the collaborative control module ensures effective collaboration between multiple exploration equipment.

[0051] s808. Performance monitoring and dynamic adjustment: System performance and exploration progress are monitored in real time through an intelligent decision support system at a frequency of once per second. When an abnormal situation is detected, the system can make a decision within 50 milliseconds and adjust the path planning strategy to ensure the continuity and success of the exploration mission.

[0052] The system of the present invention can significantly improve the autonomy and decision-making ability of exploration equipment through the above-mentioned technical solutions and processes, is applicable to a variety of complex exploration environments, and has important practical application value.

[0053] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A multi-agent exploration path planning system based on deep learning, characterized in that: include: Multi-agent collaborative architecture, used to achieve collaborative operation of multiple exploration equipment, including task allocation module, communication coordination module and collaborative control module; Deep learning environment perception module, integrating convolutional neural network (CNN) and long short-term memory network (LSTM), is used to improve the recognition ability of complex environmental features, including image recognition submodule, feature extraction submodule and environmental prediction submodule; Improved A* algorithm path planning module, adopts 16-adjacent subnode selection strategy, optimizes the path search process, and improves the efficiency of path planning, including heuristic evaluation submodule, node priority sorting submodule and path reconstruction submodule; Floating resource abstract processing module, used to abstract exploration path planning data to adapt to different types of resources and environmental changes, including resource identification submodule, resource dynamic allocation submodule and path adaptive adjustment submodule; The real-time environmental feedback mechanism includes an environmental information collection module and an analysis module, which are used to monitor the environmental status around the exploration equipment in real time and dynamically adjust the exploration planning path; the multi-objective genetic algorithm optimization module uses the NSGA-II algorithm to simultaneously consider multiple optimization goals such as path length, safety, and energy efficiency in path planning; the intelligent decision support system is used to automatically generate the optimal exploration path based on real-time environmental information and historical data, and has the ability of self-learning and self-optimization.

2. The system according to claim 1, characterized in that The deep learning environment perception module acquires environmental images through the image recognition submodule, extracts key features through the feature extraction submodule, and predicts environmental changes through the environment prediction submodule.

3. The system according to claim 1, characterized in that The improved A* algorithm path planning module evaluates the path cost through the heuristic evaluation submodule, determines the search order through the node priority sorting submodule, and constructs the optimal path through the path reconstruction submodule.

4. The system according to claim 1, characterized in that The floating resource abstract processing module identifies resource requirements in the exploration path through the resource identification submodule, dynamically allocates resources through the resource dynamic allocation submodule, and adjusts the path planning strategy according to resource changes.

5. The system according to claim 1, characterized in that The real-time environmental feedback mechanism collects environmental data through the environmental information acquisition module and analyzes the data through the analysis module to identify potential risks and opportunities.

6. The system according to claim 1, characterized in that The multi-objective genetic algorithm optimization module finds the optimal solution set through the NSGA-II algorithm, while considering multiple objectives such as path length, safety, and energy efficiency.

7. The system according to claim 1, characterized in that The intelligent decision support system continuously improves the accuracy and efficiency of path planning through self-learning and self-optimization based on real-time data and historical experience.

8. A method for using the multi-agent collaborative exploration path planning system according to claim 1, characterized in that: The following steps are involved: Step S801: Use the deep learning environment perception module to perceive the exploration environment in real time, obtain the environment image through the image recognition submodule, extract key features through the feature extraction submodule, and predict environmental changes through the environment prediction submodule; Step S802: perform path search through the improved A* algorithm path planning module, evaluate the path cost using the heuristic evaluation submodule, determine the search order through the node priority sorting submodule, and construct the optimal path through the path reconstruction submodule; Step S803: According to the output of the floating resource abstraction processing module, identify the resource requirements in the exploration path, dynamically allocate resources, and adjust the path planning strategy according to resource changes; Step S804: using the real-time environmental feedback mechanism to monitor the environmental status, collecting environmental data through the environmental information collection module, and analyzing the data through the analysis module to identify potential risks and opportunities; Step S805: Optimize the path planning result through the multi-objective genetic algorithm optimization module, consider multiple objectives such as path length, safety, energy efficiency, etc., and find the optimal solution set through the NSGA-II algorithm; Step S806: Automatically generate and adjust the exploration path using the intelligent decision support system. The system continuously improves the accuracy and efficiency of path planning through self-learning and self-optimization based on real-time data and historical experience. Step S807: During the path planning process, task allocation, communication coordination and collaborative control are continuously performed through the multi-agent collaborative architecture to ensure effective collaboration among multiple exploration devices and optimal utilization of resources; Step S808: During the execution of the system, the system performance and exploration progress are monitored in real time, and dynamic adjustments are made through the intelligent decision support system to cope with any unexpected situations or changes that occur during the exploration process.

Images