Data inversion enhanced multi-swarm crowd wisdom collaborative operation method based on vertical domain large model

Abstract

The application relates to the technical field of multi-agent collaborative control, and discloses a data inversion enhanced multi-machine group wisdom collaborative operation method based on a vertical domain large model. The method comprises the following steps: an agent collects real-time multi-modal perception data of a target operation environment and sends the data to a control server; the control server identifies a hidden interference source by using a pre-trained vertical domain large model in combination with a data inversion mechanism and constructs a global environment state mapping; semantic analysis and sub-task decomposition are performed on multi-machine collaborative operation requirements, a conflict-free task sequence is generated in combination with a space-time dependence relationship, and a grid execution track is planned for each agent; a segmented instruction sequence with a preset time window is generated in combination with communication delay and sensor refresh frequency; inversion verification, environment mapping update and instruction dynamic adjustment are performed according to state feedback data in a task execution process, a collaborative operation execution process is generated in real time through iteration, and the operation requirement execution is completed. The application realizes adaptive closed-loop control of a multi-machine collaborative operation process in a complex environment.

Description

Technical Field

[0001] This application relates to the field of multi-agent cooperative control technology, and in particular to a data inversion-enhanced multi-machine swarm intelligence cooperative operation method based on a vertical domain large model. Background Technology

[0002] With the development of intelligent manufacturing, emergency rescue, complex environment inspection, mine tunnel operations, and unmanned equipment cluster applications, multi-machine collaborative operation technology has gradually become an important research direction in the field of intelligent control. Traditional single-machine operation methods suffer from low efficiency, limited coverage, and insufficient adaptability to sudden environmental changes in complex dynamic environments, making it difficult to meet the application requirements of multi-task parallelism, regional collaborative handling, and high real-time operations. Therefore, a swarm intelligence collaborative operation mode based on multiple intelligent agents working together to complete target tasks has gradually developed.

[0003] The development path of existing multi-machine collaborative operation technology has roughly gone through the following stages: In the early stage, task allocation and path planning methods based on preset rules, fixed logic, and manual modeling were mainly adopted, and multiple work units were scheduled by manually setting constraints or based on heuristic algorithms; Subsequently, with the improvement of sensing, communication and computing capabilities, multi-agent systems gradually introduced collaborative control methods based on graph search, optimized scheduling and local environment modeling to improve the efficiency of task allocation, path planning capabilities and basic obstacle avoidance capabilities in multi-machine operation; In recent years, the development of artificial intelligence technology, especially deep learning and large model technology, has significantly enhanced the ability of multimodal perception and understanding, semantic parsing and task reasoning for complex scenarios, providing a new technical path for the evolution of multi-machine collaborative operation from "rule-driven" to "cognitive-driven".

[0004] However, existing technologies still have significant limitations in multi-machine collaborative operation scenarios facing complex and dynamic environments. On the one hand, traditional methods largely rely on explicit perception results and static modeling to understand the environmental state, making it difficult to effectively infer occlusion areas, hidden interference sources, and potential risk factors. This results in incomplete environmental representation, which in turn affects the accuracy of task decomposition, resource allocation, and path planning. On the other hand, existing collaborative scheduling methods focus primarily on explicit task allocation and local collision avoidance control, failing to adequately consider the spatiotemporal dependencies between multiple tasks, execution conflicts, and global collaborative optimization under dynamic environmental changes. This can easily lead to problems such as task conflicts, uneven resource utilization, and inconsistencies between local decisions and global goals.

[0005] Furthermore, during multi-machine collaborative operations, environmental conditions, communication link quality, agent execution status, and energy consumption levels are typically in constant flux. Existing technologies often employ a one-time planning and distribution mechanism based on initial environmental information, lacking the ability to inversely verify, dynamically update, and iteratively adjust environmental conditions by incorporating feedback from the execution process. This makes it difficult to promptly correct subsequent execution flows in the presence of communication delays, differences in perception refresh rates, and sudden environmental disturbances, thus affecting the synchronization, stability, and overall efficiency of multi-machine collaborative operations. Simultaneously, the limited semantic understanding, task decomposition, and professional knowledge utilization capabilities of general models in specific industry scenarios also hinder the efficient transformation of complex operational requirements into executable collaborative tasks.

[0006] Therefore, how to leverage large vertical domain models to enhance scene understanding and task parsing capabilities of multimodal perception data in complex and dynamic operational environments, and how to combine data inversion mechanisms to effectively infer hidden interference sources and environmental changes, and further construct global environmental state mapping, achieve conflict-free task sequencing, dynamic trajectory planning, and closed-loop iterative adjustments during execution for multi-agent collaborative operations, have become key technical issues that urgently need to be addressed in this field. Summary of the Invention

[0007] To address the problems of incomplete environmental perception, difficulty in identifying hidden interference sources, insufficient accuracy in task decomposition and collaborative scheduling, lack of closed-loop dynamic adjustment capability during execution, and poor synchronization and global coordination among multiple agents in complex dynamic environments, this application provides a data inversion-enhanced multi-agent collaborative operation method based on a vertical domain large model to improve the environmental understanding, task coordination, dynamic adaptability, and overall operation efficiency of multi-agent systems in complex operation scenarios.

[0008] This application provides a data inversion-enhanced multi-machine swarm intelligence collaborative operation method based on a vertical domain large model, applicable to a swarm intelligence collaborative system including a control server and multiple agents. The method includes: S1. Each intelligent agent collects real-time multimodal perception data of the target operating environment, and sends it to the control server after local preprocessing; S2. The control server inputs the multimodal sensing data into the built-in first vertical domain large model for feature extraction, and combines the data inversion mechanism to identify hidden interference sources and construct a global environment state mapping. S3. The control server receives the multi-machine collaborative operation requirements, uses the second vertical domain large model to perform semantic parsing on the multi-machine collaborative operation requirements, decomposes them into various sub-tasks, and determines the task distribution range of each sub-task based on the global environment state mapping. S4. For each subtask within the task distribution range, extract the spatiotemporal dependencies between tasks, and use the collective intelligence collaboration algorithm to sort the anti-conflict priorities to generate a conflict-free task sequence. S5. Assign tasks to each agent according to the task sequence, calculate the penalty cost in combination with the global environment state mapping, and plan a gridded execution trajectory for each agent. S6. The control server extracts the path node parameters of the gridded execution trajectory and evaluates the real-time communication delay with each agent. Then, it performs spatiotemporal matching of the path node parameters, real-time communication delay, and sensor refresh frequency of each agent to generate a segmented instruction sequence with a preset time window and sends it to the corresponding agent. S7. During task execution, each agent collects its state feedback data in real time and sends it back to the control server. If the deviation between the state feedback data and the expected execution state exceeds a preset threshold, the state feedback data is inverted and verified. The global environment state mapping is updated based on the inversion result, and the sequence of unexecuted segmented instructions is adjusted. S8. Based on the adjusted segmented instruction sequence and the updated global environment state mapping, generate the execution flow of multi-machine collaborative operation in real time until the operation requirements are completed.

[0009] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows: 1. By introducing a large vertical domain model into the multi-machine collaborative operation process, it is possible to perform deeper semantic understanding and feature extraction on multimodal perception data in the target operation environment. Combined with the data inversion mechanism, it can identify hidden interference sources that are difficult to detect directly by traditional explicit perception methods, thereby improving the cognitive integrity of complex dynamic environments and the accuracy of environmental modeling.

[0010] 2. By organically combining semantic parsing of job requirements, subtask decomposition, determination of task space scope, and analysis of spatiotemporal dependencies, it is possible to achieve conflict-free priority ranking and generation of conflict-free task sequences among multiple subtasks, thereby improving the rationality of task allocation, job coordination, and resource utilization efficiency of multi-agent systems in complex scenarios.

[0011] 3. By constructing a global environmental state mapping that includes dynamic environmental features and disturbances, and building a gridded cost map based on penalty costs, the execution trajectory of each agent can be globally planned. This can effectively reduce the impact of high-risk areas on the operation process and improve the safety, stability and overall operation efficiency of path planning.

[0012] 4. At the instruction execution level, by comprehensively considering wireless communication delay, path node parameters, and sensor refresh frequency, the control instructions are spatiotemporally matched and time window constrained. This can enhance the synchronous control capability during the execution of multiple agents, reduce the coordination error caused by link delay, refresh difference, and timing offset, and improve the consistency and reliability of multi-machine collaborative operation.

[0013] 5. At the closed-loop control level, by introducing inversion verification, environment mapping update and dynamic adjustment mechanism based on state feedback data during task execution, closed-loop iterative control from environment perception and task planning to execution correction can be realized, improving the system's adaptability to sudden environmental changes and execution deviations, thereby enhancing the robustness and continuous operation capability of multi-machine collaborative intelligence in complex dynamic environments. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the overall control logic for the enhanced multi-machine swarm intelligence collaborative operation based on the data inversion of a vertical domain large model in this application; Figure 2 This is a flowchart of the data inversion-enhanced multi-machine swarm intelligence collaborative operation method based on a vertical domain large model according to this application; Figure 3 This is a schematic diagram of the terminal hardware structure of the intelligent agent in this application. Detailed Implementation

[0016] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] The data inversion-enhanced multi-machine swarm intelligence collaborative operation method based on a vertical domain large model provided in this application is applied to a swarm intelligence collaborative system that includes a control server and multiple agents. The overall control logic diagram of the multi-machine swarm intelligence collaborative operation is shown below. Figure 1 As shown.

[0018] The collaborative operation system of this application includes a control server and multiple intelligent agents. The multiple intelligent agents interact bidirectionally with the control server through a wireless communication link. The intelligent agents receive collaborative operation instructions and execute corresponding sub-tasks, while simultaneously uploading operational environment perception data and operational condition feedback data to the control server. The control server performs task parsing, conflict determination, trajectory planning, feedback comparison, inversion updates, and consistency scheduling on the received data, and then distributes the generated or updated collaborative operation instructions to each intelligent agent.

[0019] Specifically, the control server includes a semantic parsing and task decomposition module, a conflict determination and task orchestration module, a trajectory planning and instruction issuance module, a feedback comparison and inversion update module, and a constraint update and consistency scheduling module. The semantic parsing and task decomposition module receives collaborative task requirements and performs semantic parsing on them, breaking down the overall task objective into multiple sub-tasks that can be executed by different agents. The conflict determination and task orchestration module performs conflict analysis on each sub-task and generates a conflict-free task sequence. The trajectory planning and instruction issuance module plans the collaborative task trajectory of multiple agents based on the conflict-free task sequence and current environmental constraints, and generates corresponding task instructions to issue to each agent. The feedback comparison and inversion update module identifies execution deviations based on the operational feedback data uploaded by the agents and updates the environmental perception results. The constraint update and consistency scheduling module adjusts task allocation and scheduling relationships based on the update results to achieve dynamic collaborative optimization.

[0020] Figure 1 The "inspection," "sweeping," and "spraying" actions performed by multiple intelligent agents are merely illustrative of different types of sub-tasks that different intelligent agents can perform, and do not constitute a limitation on the application scenario of this application. Figure 1 The overall control logic shown establishes a closed-loop control link between the control server and multiple intelligent agents, consisting of "perception uploading—centralized decision-making—instruction issuance—execution feedback—dynamic adjustment," thereby enabling adaptive control of multi-machine collaborative operations in complex environments.

[0021] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 2 The diagram shows a flowchart of the enhanced multi-machine swarm intelligence collaborative operation method based on a vertical domain large model provided by the present invention. The flowchart specifically includes the following steps: S1. Each intelligent agent collects real-time multimodal perception data of the target operating environment, and sends it to the control server after local preprocessing.

[0022] In a specific example, the intelligent agent is configured with a multimodal perception module, a motion state detection module, an energy consumption monitoring module, a communication module, a local control module, and a cooperative operation execution mechanism, wherein: The multimodal perception module includes one or more of the following: a visual acquisition circuit, a ranging acquisition circuit, an attitude detection circuit, and an environmental parameter detection circuit. It is used to acquire real-time multimodal perception data of the target working environment and real-time environmental change information during task execution. The motion state detection module is used to acquire path consumption data, the energy consumption monitoring module is used to acquire real-time energy consumption feedback data, and the local control module is used to generate state feedback data based on real-time environmental change information, path consumption data, and real-time energy consumption feedback data. It also parses the received segmented instruction sequence to generate corresponding underlying control signals. The communication module is used to receive instructions from the control server and send real-time multimodal perception data and state feedback data to the control server. The collaborative operation execution mechanism is used to respond to the underlying control signals and execute corresponding sub-tasks.

[0023] Specifically, in complex and dynamic environments such as mine roadways, underground passages, tunnel construction areas, or disaster sites, each intelligent agent is configured as an independent working unit with a multimodal perception module, motion state detection module, energy consumption monitoring module, communication module, local control module, and collaborative operation execution mechanism. The local control module completes the data aggregation and control forwarding between the modules.

[0024] After the mission starts, the multimodal perception module acquires external environmental information. It can consist of one or more of the following: a visual acquisition circuit, a ranging acquisition circuit, an attitude detection circuit, and an environmental parameter detection circuit. The visual acquisition circuit outputs image or video data; the ranging acquisition circuit outputs obstacle distance, spatial contour, or point cloud data; the attitude detection circuit outputs pose, acceleration, and angular velocity data; and the environmental parameter detection circuit outputs wind speed, dust concentration, temperature, humidity, or harmful gas concentration data, thus forming real-time multimodal perception data. During mission execution, this module continuously outputs real-time environmental change information such as changes in local obstacles, changes in channel occupancy, and changes in wind field. Simultaneously, the motion state detection module acquires the travel distance, trajectory offset, steering change, and actual path length through a wheel speed encoder, odometer, or positioning unit, forming path consumption data. The energy consumption monitoring module acquires instantaneous power, cumulative power consumption, or remaining power through voltage and current sampling circuits, forming real-time energy consumption feedback data.

[0025] The local control module communicates with the above modules through an internal bus or serial interface. After receiving the above data, it aligns and encapsulates the real-time environmental change information, path consumption data and real-time energy consumption feedback data according to a unified timestamp to generate status feedback data.

[0026] The local control module receives segmented command sequences from the control server via the communication module, parses the path nodes, time windows, speed constraints, and action types, and then converts them into underlying control signals corresponding to the start and stop of the drive, motor, steering, braking, or working device. The communication module communicates bidirectionally with the control server via a wireless LAN, private network link, 5G link, or self-organizing network link. On the one hand, it receives segmented command sequences from the control server; on the other hand, it sends real-time multimodal sensing data and status feedback data to the control server, enabling the control server to continuously obtain environmental information and execution status.

[0027] The collaborative operation execution mechanism responds to the underlying control signals to complete the corresponding sub-tasks, which may be obstacle removal, cleaning, inspection, search and rescue, transportation, or local disposal.

[0028] Figure 3 This is a schematic diagram of the hardware structure of the terminal control board of the intelligent agent in this application. Figure 3As shown, the intelligent agent is configured with a multimodal perception module, a motion state detection module, an energy consumption monitoring module, a communication module, a local control module, and a collaborative operation execution mechanism. The multimodal perception module includes one or more of the following: a visual sensor interface circuit, a laser / ultrasonic ranging circuit, an IMU (Inertial Measurement Unit), and an environmental detection circuit. These components together constitute the hardware foundation for the intelligent agent's multimodal perception, used to collect real-time multimodal perception data of the target operating environment and real-time environmental change information during task execution. The motion state detection module includes a motion state and odometer circuit for collecting path consumption data. The energy consumption monitoring module includes a voltage / current energy consumption monitoring circuit for collecting real-time energy consumption feedback data. The local control module includes an MCU (Microcontroller Unit) local control core unit for generating status data based on real-time environmental change information, path consumption data, and real-time energy consumption feedback data. The terminal control board receives state feedback data and parses the received segmented instruction sequences to generate corresponding underlying control signals. The communication module includes a wireless communication module interface, which can connect to wireless LAN, cellular, private network, or self-organizing network communication modules. It receives instructions from the control server and sends real-time multimodal sensing data and state feedback data to the control server, enabling bidirectional communication between the intelligent agent and the control server. The collaborative operation execution mechanism includes operating components connected to the execution mechanism drive circuit, used to respond to underlying control signals and execute corresponding sub-tasks. The power management circuit provides regulated power to each functional unit of the terminal control board and protects against overcurrent, overvoltage, undervoltage, and abnormal load conditions to improve the stability and safety of the intelligent agent in complex dynamic operating environments. Through the above-described terminal control board hardware structure, functions such as environmental perception, state detection, energy consumption monitoring, data communication, local control, and operation execution can be realized, providing hardware support for the multi-machine collaborative operation method of this application.

[0029] The following embodiments illustrate the present invention using a narrow tunnel clearing operation scenario as an example. However, the present invention is not limited to this application scenario. Those skilled in the art will understand that the present invention is also applicable to multi-agent collaborative inspection, obstacle clearing, search and rescue, and operation scheduling scenarios in complex dynamic environments such as mine tunnels, underground passages, tunnel construction areas, and disaster sites.

[0030] Specifically, in the scenario of cleaning narrow alleyways, each intelligent agent, acting as a mobile piece of equipment (such as a cleaning robot, inspection robot, or unmanned transport vehicle), disperses along the alleyway to be cleaned and enters the target area. Each agent's onboard multimodal perception module then collects real-time multimodal perception data of the target working environment. Specifically, the visual acquisition circuit uses cameras to acquire images or video data of the area in front of, on the side walls, and on the roof of the alleyway; the ranging acquisition circuit uses lidar, millimeter-wave radar, or ultrasonic rangefinders to acquire distance information or point cloud information of obstacle accumulations, alleyway boundaries, and passable areas; the attitude detection circuit uses an inertial measurement unit to acquire the agent's current pose, acceleration, and angular velocity information; and the environmental parameter detection circuit acquires information on wind speed, dust concentration, temperature, humidity, or harmful gas concentration inside the alleyway, thus forming the raw multimodal perception data.

[0031] After all types of raw data are acquired, the local control module performs local preprocessing on the raw multimodal sensing data. For example, it performs noise reduction, distortion correction, and brightness compensation on image data; filters and removes outliers from ranging data to reduce abnormal ranging errors caused by dust reflection and irregular surface reflection; performs zero-bias calibration and coordinate unification on attitude data; and filters and smooths environmental parameter data for outliers. Subsequently, the local control module performs time-series alignment of all types of valid data according to a unified timestamp, and spatially registers image data, distance data, and pose data according to the local coordinate system, outputting locally preprocessed multimodal sensing data. This multimodal sensing data includes processed image data, distance data, pose data, and environmental parameter data. Afterward, the local control module compresses and encapsulates the multimodal sensing data, and writes the agent number, acquisition time, and current location information into the data packet, forming an upload data packet. This packet is then sent by the communication module to the control server via a wireless LAN, private network communication link, 5G link, or ad hoc network link, enabling the control server to receive real-time environmental input from different work locations.

[0032] S2. The control server inputs the multimodal sensing data into the built-in first vertical domain large model for feature extraction, and combines the data inversion mechanism to identify hidden interference sources and construct a global environment state mapping.

[0033] In one specific embodiment, the process of performing step S2 may specifically include the following steps: Multimodal sensing data is input into the first vertical domain large model, and features are extracted by combining the pre-built domain knowledge base to generate an environmental feature description that includes dynamic obstacle factors and wind speed change factors. Based on the description of environmental characteristics, the hidden interference sources and their distribution status in the target working environment are determined by reverse deduction through the Bayesian network model in the data inversion mechanism. By fusing environmental feature descriptions with the distribution status of concealed interference sources, a global environmental state mapping that includes dynamic environmental features and interference terms is constructed.

[0034] Specifically, the multimodal perception data includes image data, point cloud or distance data, pose data, and environmental parameter data such as wind speed, dust concentration, temperature and humidity, or harmful gas concentration from multiple agents. Due to severe obstruction, suspended dust, local wind field disturbance, dynamic changes of obstacles, and blind spots in narrow roadways, mine roadways, underground passages, or tunnel construction areas, it is difficult to determine the risk factors in areas not directly observed based solely on explicit perception results. Therefore, the control server inputs the multimodal perception data into the built-in pre-trained first vertical domain large model and performs feature extraction in conjunction with a pre-set domain knowledge base to generate an environmental feature description that can be used for subsequent inversion inference.

[0035] The first large-scale vertical model undergoes pre-training and domain-adaptation training before deployment. It employs a network structure with multi-branch encoding, spatiotemporal alignment fusion, scene feature decoding, and structured semantic output. This structure includes at least an image encoding branch, a point cloud or distance information encoding branch, a pose encoding branch, an environmental parameter encoding branch, a cross-modal fusion layer, a scene feature decoding layer, and a semantic output head. The image encoding branch encodes image data, extracting two-dimensional visual features such as obstacle boundaries, alleyway contours, occluded areas, and dust coverage. The point cloud or distance information encoding branch encodes point cloud or distance data, extracting spatial features such as obstacle spatial distribution, local channel morphology, and passable cross-sectional width. The pose encoding branch encodes pose data and trajectory change information, forming spatial location features corresponding to the acquisition time. The environmental parameter encoding branch encodes environmental parameters such as wind speed, wind direction, temperature, humidity, dust concentration, and harmful gas concentration, forming environmental state features. After each encoded branch outputs, the cross-modal fusion layer first performs spatiotemporal alignment based on a unified timestamp and a unified spatial coordinate system, and then jointly fuses image features, distance features, pose features, and environmental parameter features to form a joint scene feature representation for each spatial unit. Fusion can be achieved using feature concatenation combined with attention weighting, or using a cross-attention approach. The scene feature decoding layer, based on the joint scene feature representation, decodes and extracts scene features such as obstacle contours, channel width variations, local congestion levels, wind speed gradient variations, dust accumulation areas, and passable boundaries. The semantic output head outputs the decoding results as a structured environmental feature description, which includes at least spatial unit identifiers, obstacle categories, obstacle occupancy status, boundary locations, passable widths, local wind speed values, wind speed gradients, occlusion levels, and risk levels. The control server then generates a global environmental state mapping based on the structured environmental feature description. In the pre-training stage, historical data corresponding to the application scenario is used for self-supervised training. The training samples include at least the correspondence between image frames, point cloud frames, pose sequences and environmental parameter sequences to learn the spatiotemporal correlation features between multimodal data. In the domain adaptation training stage, samples with obstacle category labels, wind field disturbance labels, occlusion area labels and risk area labels are used for supervised fine-tuning so that the model output corresponds to the obstacle changes, wind field changes and confined space structure in the scenario of this application.

[0036] When the control server invokes the first vertical domain large model, it inputs the current moment's image data, point cloud or distance data, pose data, and environmental parameter data into the multimodal coding layer for feature encoding. During scene feature extraction, it incorporates lane cross-section rules, equipment passage boundaries, common obstacle accumulation patterns, dust diffusion patterns, and local wind propagation rules from the domain knowledge base to constrain, correct, and semantically enhance the encoded multimodal features, thereby obtaining an environmental feature description. The environmental feature description includes at least spatial unit numbers, timestamps, dynamic obstacle factors, and wind speed change factors. Both dynamic obstacle factors and wind speed change factors are calculated based on the current moment and adjacent moments' multimodal perception data, and their impact degree is determined and the results corrected using corresponding rules from the domain knowledge base. Dynamic obstacle factors characterize obstacle position changes, movement trends, volume changes, and the degree of passage impact; wind speed change factors characterize local wind speed magnitude, wind direction deflection, and wind speed gradient changes. For example, the intensity of obstacle change can be calculated based on the positional, volumeal, and areaal changes of the obstacle outline center in the same area during two adjacent sampling times. This, combined with the equipment passage width threshold and roadway cross-section rules, determines the degree of its impact on passage capacity, thus forming a dynamic obstacle factor. Alternatively, the wind speed difference, wind direction change, and wind speed gradient between adjacent measuring points can be calculated based on the wind speed, wind direction, and corresponding spatial location of different measuring points in the roadway. The rationality of the change results can then be verified by combining local airflow propagation rules, thus forming a wind speed change factor.

[0037] After obtaining the environmental feature description, the control server invokes the Bayesian network model in the data inversion mechanism to identify hidden interference sources. The Bayesian network model pre-establishes a node and edge structure based on the causal relationships in the tunnel operation scenario. Nodes include at least the states of obstacle accumulation, local collapse, dust obstruction, abnormal wind field, obstructed equipment passage, and hidden interference sources. Edges represent the conditional dependencies between nodes; for example, dust obstruction affects image visibility and ranging confidence, local collapse affects obstacle accumulation and abnormal wind field, and abnormal wind field affects dust diffusion and harmful gas accumulation, thus forming a causal relationship network suitable for collaborative operations in confined spaces. The control server inputs the environmental feature description output from the first vertical domain large model as observation evidence into the Bayesian network model and calculates the posterior probability of each candidate hidden interference source based on the pre-statistically obtained conditional probability table. Hidden interference sources include dust-obscured obstacle accumulations, locally collapsed points that are not directly visible, hazardous gas accumulation areas concealed by airflow disturbances, and dynamic obstruction sources located in blind spots. The training of the Bayesian network model can be based on historical operation records and simulated scenario data. The known obstacle distribution, collapse distribution, gas distribution and wind field distribution are used as ground truth labels, and the corresponding image features, distance features, pose features and environmental parameter features are used as observation inputs. The conditional probability table is updated by the maximum likelihood estimation or expectation maximization method.

[0038] During the inference phase, the control server updates the conditional probabilities of each candidate hidden interference source based on the current input evidence, and selects the candidate results whose posterior probability exceeds a preset probability threshold as the identified hidden interference sources. Simultaneously, it determines their distribution status based on the positional relationship of the corresponding observation areas. For example, when a region simultaneously experiences decreased image visibility, abnormal ranging echoes, sudden changes in wind speed, and a sudden drop in traffic width, the Bayesian network model can deduce the probability distribution of the existence of obstacle accumulation or local collapse points in that region. Similarly, when a region experiences abnormal wind direction deflection and an increase in harmful gas concentration, the Bayesian network model can deduce the probability distribution of the existence of hidden leakage sources or restricted diffusion zones in that region.

[0039] After identifying concealed interference sources and their distribution, the control server, based on the pose information and timestamps in the data uploaded by each agent, transforms the local environmental features observed by different agents into a unified global coordinate system. Using a grid map, topological map, or 3D spatial voxel map as the basic representation, dynamic obstacle factors, wind speed variation factors, and the probability distribution of concealed interference sources are then written into the corresponding spatial units. For cases where multiple agents repeatedly observe the same area, data can be fused according to the principles of observation time priority, sensor confidence weighting, or multi-source consistency to form a unified spatial state result. Based on this, each spatial unit records at least the obstacle state, passage state, wind field state, dust state, and concealed interference source state, thus obtaining a global environmental state mapping for subsequent steps.

[0040] S3. The control server receives multi-machine collaborative operation requests, uses the second vertical domain large model to perform semantic parsing on the multi-machine collaborative operation requests, breaks them down into various sub-tasks, and determines the task distribution range of each sub-task based on the global environment state mapping.

[0041] In one specific embodiment, the process of performing step S3 may specifically include the following steps: The control server inputs the received multi-machine collaborative operation requirements into the second vertical domain large model, extracts the operation intent and operation objectives, and decomposes the semantics of the multi-machine collaborative operation requirements into multiple sub-tasks; Environmental dynamic features and disturbance terms are extracted from the global environmental state map. The environmental dynamic features include the location of dynamic obstacles and wind speed change factors, and the disturbance terms are the identified hidden disturbance sources. The work space requirements corresponding to each sub-task are superimposed with the spatial distribution of environmental dynamic characteristics and interference items, and safety boundary verification is performed. Based on the verification results, areas that conflict with hidden interference sources or exceed the wind speed safety threshold are eliminated, and the actual task distribution range of each sub-task in the workspace is determined.

[0042] Specifically, the control server receives multi-machine collaborative operation requests, which can originate from the scheduling terminal, the host computer system, or preset task templates. These requests must include at least the operation type, target area, operation object, completion conditions, time limits, and constraints on the number of participating agents. In narrow alleyway clearing scenarios, multi-machine collaborative operation requests can include clearing accumulated obstacles in designated alleyway sections, handling localized collapse areas, clearing passable areas, and inspecting and confirming high-risk sections.

[0043] Upon receiving the job request, the control server inputs the multi-machine collaborative job request into the second vertical domain large model. The server performs semantic parsing on the multi-machine collaborative job request, extracting the job intent, job objective, and constraints. The second vertical domain large model adopts a network structure of task requirement encoding, semantic feature fusion, intent recognition and constraint extraction, and task decomposition output. It includes at least a task input adaptation layer, a task semantic encoding layer, a semantic fusion layer, an intent recognition head, an objective constraint extraction head, and a task decomposition output head. The second vertical domain large model can use the same basic model as the first vertical domain large model used for environmental feature extraction in S2, or share some underlying encoding parameters, and complete domain adaptation training for collaborative job task semantics before deployment. The task input adaptation layer is used to uniformly convert text-based, structured tabular, or instruction template-based job requirements into a single format. The task semantic encoding layer is used to encode the converted job requirements and generate task semantic feature vectors. The semantic fusion layer is used to jointly model semantic information such as job actions, job objects, spatial segments, temporal requirements, and safety constraints to form a joint semantic representation of the task. The intent recognition head outputs job intents such as cleaning, inspection, search and rescue, transfer, or hazard removal based on the joint semantic representation of the task. The target constraint extraction head extracts target segments, target objects, completion conditions, priority conditions, and safety constraints. The task decomposition output head outputs a set of sub-tasks and the corresponding job objects, job areas, execution conditions, and completion conditions for each sub-task based on the identified job intents and constraint results. Preferably, the task semantic encoding layer adopts a pre-trained language model encoding structure. Structured tabular requirements are input into the encoding structure after field embedding, and instruction template-based requirements are input into the encoding structure after template slot expansion. The semantic fusion layer uses feature concatenation combined with attention weighting, or cross-attention, to fuse task semantic information from different sources. The control server, based on the results output from the task decomposition output header, forms a set of subtasks in S3 and the operational attributes of each subtask. It then combines this with the global environment state mapping generated in S2 to determine the task distribution range of each subtask. Domain adaptation training samples include historical work orders, manual scheduling instructions, standard tunnel operation procedures, and event handling records. The sample annotations include at least the following labels: operation intent label, operation object label, spatial segment label, priority processing object label, and completion condition label, ensuring that the model output corresponds to the collaborative operation requirements in the application scenario.

[0044] When the control server invokes the second vertical domain large model, it inputs the currently received multi-machine collaborative operation requirements into the task semantic encoding layer and outputs the operation intent and operation objective. The operation intent is used to characterize the operation category to which the current task belongs, and the operation objective is used to characterize the target object, target area, and completion standard to be completed. For example, when the multi-machine collaborative operation requirement is "clear the obstruction in the third section and restore equipment passage", the vertical domain large model can extract the operation intent as passage clearing and the operation objective as clearing the obstruction in the third section and restoring the passable width; when the multi-machine collaborative operation requirement is "inspect the section with abnormal wind speed and investigate hidden risk sources", the vertical domain large model can extract the operation intent as risk inspection and the operation objective as locating the abnormal section and confirming the status of hidden interference sources.

[0045] The semantic decomposition rules are jointly determined by the task structure output by the second vertical domain large model and the operation process template in the domain knowledge base. The domain knowledge base includes at least the lane clearing process rules, equipment operation capability boundaries, hazardous area disposal rules, equipment entry conditions, and operation stage dependencies. Based on the operation object category, target section range, and operation stage relationships, the control server decomposes the overall operation requirements into multiple sub-tasks with independent input and output constraints, and generates a structured task description for each sub-task. The task description includes at least the sub-task type, target section, operation space requirements, resource requirements, and execution constraints. When outputting each sub-task, the control server also records the corresponding operation space requirements, which include at least the target section, minimum operation width, acceptable wind speed range, required approach distance, equipment entry direction constraints, and prohibited entry conditions. For example, in the scenario of clearing narrow alleyways, the overall goal of "restoring passage" can be broken down into the following sub-tasks: obstacle location, main obstacle removal, residual debris sweeping, boundary verification, and passage confirmation. The obstacle location sub-task corresponds to the obstacle range identification requirement, the main obstacle removal sub-task corresponds to the obstacle removal operation requirement, the residual debris sweeping sub-task corresponds to the ground sweeping requirement, the boundary verification sub-task corresponds to the local alleyway boundary confirmation requirement, and the passage confirmation sub-task corresponds to the passage determination requirement.

[0046] The environmental dynamics include at least the location of dynamic obstacles and the wind speed variation factor. The location of dynamic obstacles characterizes the positional distribution and changing trend of obstacle accumulations, dynamic blockage sources, or movable obstacles in the global coordinate system. The wind speed variation factor characterizes the wind speed magnitude, wind direction changes, and wind speed gradient changes within each spatial unit. The interference term refers to the concealed interference sources identified in S2, including at least obstacle accumulations obscured by dust, locally collapsed points not directly visible, hazardous gas accumulation areas, and dynamic blockage sources within blind spots. The control server maps the operational space requirements corresponding to each sub-task to the corresponding spatial region in the global environmental state mapping, and overlays these operational space requirements with the spatial distribution corresponding to the environmental dynamics and interference terms to form candidate operational regions for each sub-task. The overlay process is performed in a unified global coordinate system and can be implemented using spatial intersection, difference set filtering, neighborhood buffering, and constraint filtering methods. For dynamic obstacle locations, the control server removes the area it occupies and its safety buffer zone from the candidate operation area; for wind speed variation factors, the control server filters spatial units that meet the wind speed requirements based on the upper limit of the allowable wind speed corresponding to the sub-task; for hidden interference sources, the control server marks areas with spatial overlap or adjacent conflicts as prohibited areas based on their spatial distribution and risk neighborhood range.

[0047] After candidate work areas are formed, the control server performs safety boundary checks on each sub-task. Safety boundary checks include at least access boundary checks, wind speed boundary checks, and concealed interference source conflict checks. Access boundary checks determine whether the candidate work area meets the requirements for equipment access width, turning radius, and approach distance; wind speed boundary checks determine whether the wind speed value or wind speed gradient within the candidate work area exceeds the safety upper limit allowed by the corresponding equipment and work type; concealed interference source conflict checks determine whether the candidate work area spatially overlaps with or is adjacent to hazardous gas accumulation areas, local collapse points, dynamic blockage sources, or high-probability concealed obstacle areas. For wind speed safety thresholds, these can be pre-set based on equipment specifications, work modes, and safety rules in the domain knowledge base. The control server compares the wind speed variation factor in the candidate area with the corresponding threshold, eliminating spatial units exceeding the threshold. For concealed interference source conflict areas, the control server determines the conflict boundary based on the probability distribution of the concealed interference source and a preset risk radius, eliminating spatial units overlapping with the conflict boundary.

[0048] Based on the safety boundary verification results, the control server removes areas that conflict with concealed interference sources or exceed the wind speed safety threshold from the candidate work areas. It then performs connectivity analysis and region aggregation on the remaining spatial units to determine the actual task distribution range of each subtask within the work space. Specifically, if a subtask, after verification, forms multiple separate executable regions, the control server marks these executable regions as multiple distribution sub-ranges of that subtask. If a subtask, after verification, has no executable regions that meet the conditions, the control server writes this result into the task constraint information for subsequent task reassignment or execution order adjustment.

[0049] S4. For each subtask within the task distribution range, extract the spatiotemporal dependencies between tasks, and use the collective intelligence collaboration algorithm to sort the anti-conflict priorities to generate a conflict-free task sequence.

[0050] In one specific embodiment, the process of performing step S4 may specifically include the following steps: The local spatial environment details corresponding to each sub-task are reconstructed using a local reconstruction method. By combining the reasoning logic of the task context with the details of the local spatial environment, the spatiotemporal dependencies between various subtasks are analyzed and extracted; Based on spatiotemporal dependencies, a collective intelligence collaboration algorithm is used to determine whether there are temporal or spatial execution conflicts in the execution of each subtask. If a conflict exists, the environmental feature complexity index of the region corresponding to the subtask that caused the conflict is extracted, and the conflict prevention priority of each subtask is ranked according to the index; if there is no conflict, the priority of each subtask is directly determined according to the spatiotemporal dependency. Based on the sorting results, a conflict-free task sequence for multi-machine collaborative operations is generated.

[0051] Specifically, in scenarios with limited space such as narrow alleys, constantly changing obstacle states, and overlapping work surfaces, it is difficult to avoid execution conflicts by simply arranging tasks based on static task types or simple sequence. To address this, the control server first performs local environmental refinement and reconstruction on the corresponding areas of each subtask, then extracts the spatiotemporal dependencies between tasks, and generates a conflict-free task sequence after conflict judgment.

[0052] Local reconstruction is based on the global environmental state mapping obtained in S2 and the multimodal sensing data uploaded in S1. The task distribution range of each subtask is used as a local truncation window to extract corresponding spatial units from the global environmental state mapping. These units are then combined with image features, distance features, pose features, and environmental parameter features for local refinement and restoration. The reconstruction objects include at least: the lane boundary morphology, local obstacle distribution, passable width, location of adjacent work surfaces, local wind disturbance areas, and the neighborhood state of concealed interference sources in the subtask's area. This can be achieved using multi-frame observation fusion for local occupancy raster reconstruction, local point cloud stitching reconstruction, or local voxel map updating. After local reconstruction, detailed local spatial environment data at the subtask level is obtained. This data includes at least local boundary contours, local obstacle occupancy status, passable cross-section width, local wind speed gradient distribution, and the neighborhood range of concealed interference sources.

[0053] The task context reasoning logic includes at least task flow rules, equipment capability boundaries, task completion conditions, and hazardous area handling rules. It is used to determine dependencies between subtasks, such as whether one subtask depends on the completion result of another subtask, or whether there are spatial or equipment constraints preventing parallel execution of the same subtask. The control server establishes pairwise dependency determinations for each subtask based on subtask type, task distribution range, local spatial environment details, and task context rules. For example, if the execution of a subtask requires another subtask to release a passage, remove obstacles, complete boundary confirmation, or reduce risk after completion, a temporal dependency is determined. If the task distribution ranges of two subtasks spatially overlap or share the same passageway within the same time window, a spatial dependency is determined. If the execution of one subtask changes the environmental basis of another subtask, an environmental coupling dependency is determined. The control server constructs a task dependency graph based on this, where nodes correspond to each subtask, edges correspond to dependency types, and edge attributes include at least dependency category, scope of effect, time window of influence, and constraint strength.

[0054] The swarm intelligence collaborative algorithm is used for collaborative conflict determination in multi-agent, multi-subtask, and shared workspace scenarios. Its input data includes at least a set of subtasks, a task dependency graph, the task distribution range of each subtask, local spatial environment details, and capability constraints of the participating agents. The output data includes at least the conflict determination result and a conflict relationship table. Preferably, a constraint graph-based swarm intelligence collaborative algorithm can be used. The control server defines each subtask as a node to be checked and defines the search state during the conflict determination process as: the set of subtasks currently included in the check, the corresponding candidate execution time window, the occupied space range, the occupied resource status, and the remaining set of subtasks to be checked. Starting from an initial empty state, under the premise of satisfying the task dependency graph constraints, candidate subtasks are selected from the remaining subtasks to be checked, and corresponding candidate execution schemes are generated. Each candidate execution scheme includes at least a candidate start time window, a candidate occupied area, a candidate resource call, and a candidate agent identifier. Subsequently, the candidate execution scheme is checked against the constraints of each of the other subtasks currently included in the check to determine whether a conflict occurs. Conflicts include at least the following: timing conflicts, spatial conflicts, resource conflicts, and environmental coupling conflicts. A timing conflict occurs when two subtasks are assigned to overlapping time windows due to a sequential completion relationship. A spatial conflict occurs when two subtasks occupy the same area or passageway that cannot be traversed concurrently within the same time window. A resource conflict occurs when two subtasks call the same type and limited number of execution resources within the same time window, causing resource demand to exceed the available quantity. An environmental coupling conflict occurs when two subtasks execute simultaneously within the same jobspace or adjacent related areas, interfering with sensor stability, occupancy confirmation, or risk release processes. For identified conflicts, the control server generates a conflict relationship table, which records at least the conflicting task pair, conflict type, conflict area, conflict time window, and conflict trigger.

[0055] After obtaining the conflict determination result, different sorting branches are entered based on whether a conflict exists. If no conflict exists, it means that no additional anti-conflict processing is needed between the related subtasks. At this time, the control server directly determines the priority order of each subtask based on the temporal dependency, spatial dependency, result dependency, and environmental coupling dependency in the task dependency graph. Specifically, the execution order of each subtask can be determined according to the principle of "prioritizing the satisfaction of pre-order dependencies, releasing shared channels first, and starting result dependencies last." For subtasks whose task distribution is independent of each other, whose resources do not conflict, and whose environmental coupling effects are not present, they can be arranged into parallel task segments.

[0056] If conflicts exist, priority for conflict prevention is assigned only to the subtasks that caused the conflict. Environmental complexity indicators include at least local obstacle density, fluctuation in passable width, wind speed gradient variation, probability of concealed interference sources, and the number of alternative paths in the local space. Local obstacle density can be characterized by the proportion of obstacles per unit length or area; fluctuation in passable width can be characterized by the width variation of each cross-section within the subtask path range; wind speed gradient variation can be characterized by the wind speed difference between adjacent spatial units; probability of concealed interference sources can be characterized by the risk probability of the corresponding spatial unit in S2; and the number of alternative paths in the local space can be characterized by the number of feasible connecting paths from the current task distribution area to the target area. The control server normalizes the above complexity indicators and performs a weighted sum based on preset weights to obtain the conflict prevention ranking value. The preset weights can be pre-set by scheduling rules in the domain knowledge base, and the weight sum is 1. A higher conflict prevention ranking value indicates a more complex area, higher risk, and fewer alternative paths for the corresponding subtask, and should be prioritized for execution to release bottleneck channels as early as possible and reduce the probability of subsequent conflict propagation.

[0057] Preferably, the control server can further identify conflict-related critical paths by combining the task dependency graph. A critical path refers to at least one dependency link in the task dependency graph, formed by sequentially connecting multiple subtasks with pre- and post-constraint relationships, and significantly constraining the total project duration, bottleneck release order, or the initiation conditions of multiple subsequent subtasks. The control server can determine the criticality of each dependency link based on its cumulative estimated execution time, the bottleneck spatial region involved, and the number of subsequent associated subtasks constrained by it. When a conflicting subtask is located on the critical path, its environmental feature complexity index is extracted first, and a critical path enhancement weight is applied based on the anti-conflict ranking value, giving high-risk conflicting subtasks on the critical path a higher ranking priority. For other conflicting subtasks not located on the critical path, the ranking order is determined based on the environmental feature complexity index, combined with the spatiotemporal dependencies, spatial constraints, and resource constraints.

[0058] After sorting conflict-free branches, the control server merges the two sorting results and arranges the subtasks in descending order of priority. Taking into account spatiotemporal dependencies, task distribution, and resource constraints, it assigns each subtask a permitted execution time window and a space range, ultimately generating a conflict-free task sequence for multi-machine collaborative work. A conflict-free task sequence includes at least the subtask identifier, corresponding execution order, permitted execution time window, permitted space range, and associated preceding task identifier.

[0059] S5. Assign tasks to each agent according to the task sequence, calculate the penalty cost in combination with the global environment state mapping, and plan a gridded execution trajectory for each agent.

[0060] In one specific embodiment, the process of performing step S5 may specifically include the following steps: Assign corresponding subtasks to each agent according to a conflict-free task sequence; By combining the global environment state mapping, the dynamic features and disturbances of the environment within the workspace are extracted, the workspace is divided into grids, and the penalty cost of each grid unit is calculated based on the complexity of the dynamic features of the environment and the danger level of the disturbances, thus constructing a gridded cost map. Based on a gridded cost map, a gridded path search technique is adopted to plan gridded execution trajectories for each agent to avoid high-risk areas, with the goal of minimizing cumulative penalty costs.

[0061] Specifically, in scenarios such as narrow alleys and underground passages, there are usually situations where the passage cross-section is limited, the obstacle state changes dynamically, the wind field is disturbed, the distribution of hidden interference sources is uneven, and the passage paths of multiple agents intersect. It is difficult to guarantee the reachability, safety and coordination of the execution trajectory by simply issuing tasks according to the order of sub-tasks. Therefore, it is necessary to construct a gridded cost map based on the global environment state mapping and to plan the execution trajectory with the optimization objective of minimizing the cumulative penalty cost.

[0062] The control server reads the execution order, allowed execution time window, allowed occupation area, resource requirements, and preceding task constraints of each subtask according to the conflict-free task sequence. It also reads the capability description information of each agent, which includes at least type, dimensions, minimum passage width, turning radius, speed range, stable sensor operating range, load capacity, type of working device, and remaining power or fuel. Based on the subtask's operation type, distribution range, resource requirements, and execution constraints, the server matches the subtask with suitable agents. For example, for mechanical obstacle clearing tasks, agents with pushing, grasping, or demolition devices are prioritized; for risk inspection tasks, agents with environmental detection sensors and stable passage capabilities are prioritized. For subtasks requiring multiple agents to collaborate, the server divides the subtask into primary and secondary execution roles based on its distribution range, work area width, and time window constraints, and assigns them to different agents to avoid resource conflicts. The task allocation result includes at least the agent identifier, corresponding subtask identifier, allowed entry time window, starting entry position, and target work area.

[0063] After task allocation, a gridded cost map suitable for path search is constructed. Environmental dynamics include at least dynamic obstacle locations, obstacle density, passage width variations, wind speed, wind speed gradient, and local passability variations. Disturbances include at least collapse risk zones, hazardous gas accumulation zones, high dust concentration zones, blind spots, and high-probability concealed obstacle areas. The control server determines the grid cell side length based on the agent's size and control precision. The task distribution range, allowed occupancy areas, and adjacent passage areas are mapped to a grid space under a unified global coordinate system. Each grid cell is written with a passability marker, obstacle occupancy status, risk level, and time window occupancy information. Grid cells occupied by physical obstacles, located within obstacle safety buffer zones, exceeding the minimum passage width contraction boundary, or overlapping with prohibited areas are marked as impassable cells; the rest are candidate passable cells.

[0064] The penalty cost for each grid cell is calculated based on candidate passage cells. This cost characterizes the passage difficulty, environmental risk, and cooperative avoidance burden of an agent traversing the corresponding grid cell. Preferably, the penalty cost consists of a basic passage cost, an environmental complexity cost, a risk penalty cost, and a cooperative avoidance cost. The basic passage cost characterizes the basic movement cost for an agent traversing the corresponding grid cell and is at least related to the grid cell's side length, ground passage conditions, and the agent's own maneuverability. For straight grid cells without additional restrictions, the basic passage cost can be set to a fixed value. For grid cells with poor ground adhesion conditions or limited turning, the basic passage cost is increased accordingly. The environmental complexity cost characterizes the impact of obstacle density, the degree of fluctuation in passable width, and the degree of wind speed gradient variation on passage stability. The higher the obstacle density, the closer the local passable width is to its limit, and the greater the wind speed gradient, the higher the environmental complexity cost. Risk penalty cost is used to characterize the impact of the hazard level of interference items on safety, and is determined at least based on the risk level or risk probability corresponding to hazardous gas accumulation areas, collapse risk areas, high-probability areas of hidden obstacles, and other hazard sources. When the same grid cell corresponds to multiple interference items, the risk is calculated by superimposing risks based on the type and hazard level of each interference item. Cooperative avoidance cost is used to characterize the impact of the path occupancy status of other agents within the corresponding time window, bottleneck section passing restrictions, and intersection area avoidance requirements on cooperative execution. If a grid cell has been pre-occupied by other agents within the same time window, or is located in a one-way channel, the cooperative avoidance cost is increased.

[0065] When calculating various costs, corresponding environmental dynamic features and disturbance information are extracted from the global environment state mapping, and time window constraints and occupancy constraints are extracted from the task sequence. Parameters such as obstacle density, passable width fluctuation, wind speed gradient, risk probability, and path occupancy are normalized, and then the above four types of costs are weighted and summed according to the preset weights in the domain knowledge base to obtain the comprehensive penalty cost of the corresponding grid. Preferably, for grids completely occupied by physical obstacles, overlapping with prohibited areas, or not meeting the passage conditions, no comprehensive penalty cost is calculated, and they are directly marked as impassable units. For grids located within the obstacle safety buffer zone but not yet meeting the absolute prohibition condition, the environmental complexity cost increases as the distance from the grid unit to the outer contour boundary of the obstacle decreases. For grids where the wind speed value exceeds the current agent's safe working threshold but does not exceed the absolute prohibition threshold, the wind speed-related cost is increased in the environmental complexity cost. To adapt to the passage capabilities and operational requirements of different types of agents, the weight allocation of various costs can be adjusted according to the operation type and agent type. For example, for inspection-type agents, the weights of the components related to wind speed gradient, line-of-sight obstruction, and sensor stability in the environmental complexity cost are increased; for transport-type agents, the weights of the components related to passage width and turning space in the basic passage cost and environmental complexity cost are increased. The comprehensive penalty cost corresponding to all grid cells is written into the cost matrix to form a gridded cost map.

[0066] After obtaining the gridded cost map, gridded execution trajectories are planned for each agent to avoid high-risk areas. The starting point of the execution trajectory is the current grid position of the agent or a preset standby position, and the ending point is the entry point, work point, or exit point of the assigned subtask. If the subtask includes three stages: entry, work, and exit, entry trajectories, work coverage trajectories, and exit trajectories are planned separately and combined in chronological order to form a complete execution trajectory. Path search aims to minimize the cumulative penalty cost. Under the conditions of minimum passage width, maximum turning constraint, prohibited entry area constraint, and task time window constraint, the optimal grid path from the starting point to the ending point is searched. For example, the A* path search algorithm, the Dijkstra path search algorithm, or an incremental grid search algorithm with time window constraints can be used. If the A* path search algorithm is used, the actual cumulative cost consists of the cumulative penalty cost from the starting point to the current grid cell, the heuristic cost consists of the grid distance from the current grid cell to the target grid cell, and the edge expansion cost is the comprehensive penalty cost when expanding adjacent grid nodes, so that the search results prioritize passing through low-risk, low-congestion, and well-accessible areas. For sub-tasks that require coverage (such as cleaning and inspection), after entering the target area, a reciprocating coverage trajectory generation method based on grid coverage can be further adopted to divide the target area into multiple coverage sub-areas to avoid overlapping with obstacle units and high-penalty units.

[0067] To avoid new conflicts arising during multi-agent execution, conflict-free task sequences and time window constraints are also written into the gridded cost map during path planning. For grid cells occupied by other agents within a specific time window, the cooperative avoidance cost within the corresponding time window is increased, or the grid cell is directly marked as a temporarily inaccessible cell within the corresponding time window; narrow passages with insufficient passing capacity only allow one-way traffic, and higher penalty costs are imposed on reverse candidate paths; for grid areas traversed by predicted dynamic obstacle trajectories, time-related dynamic penalty values ​​are set based on the predicted arrival time, so that the planned trajectory avoids the future occupied areas of dynamic obstacles.

[0068] After trajectory planning is completed, the trajectory instruction set for each agent is output, including at least the agent identifier, target subtask identifier, trajectory grid sequence, corresponding time window, expected arrival order, entry point coordinates, exit point coordinates, and velocity constraint information. If no feasible path satisfying the constraints exists during the search process, the result is written to the execution anomaly flag, triggering replanning. This includes expanding the search range of alternative paths, adjusting the entry time window of other agents in the same area, or reverting the corresponding subtask to the task allocation stage to reselect the executing agent.

[0069] S6. The control server extracts the path node parameters of the gridded execution trajectory and evaluates the real-time communication latency with each agent. Then, it performs spatiotemporal matching of the path node parameters, real-time communication latency, and sensor refresh frequency of each agent to generate a segmented instruction sequence with a preset time window and sends it to the corresponding agent.

[0070] In one specific embodiment, the process of performing step S6 may specifically include the following steps: Extract the spatial coordinates and heading status of the gridded execution trajectory as path node parameters; The quality of the wireless communication link between the control server and each agent is detected by heartbeat packets to assess real-time communication latency. The path node parameters, real-time communication delay, and refresh frequency of the sensors carried by each intelligent agent are spatiotemporally matched to analyze the estimated time offset of the command to each path node. The path node parameters are compensated and adjusted in the time dimension based on the estimated time offset so that the arrival time of the adjusted instructions meets the synchronization requirements of multi-machine collaboration, and a segmented instruction sequence with a preset time window is generated. By combining real-time communication latency with the real-time response speed of each agent, the timing of instruction issuance is optimized, and segmented instruction sequences are issued to the corresponding agents for execution.

[0071] Specifically, after obtaining the gridded execution trajectories of each agent, the control server does not directly send out the entire execution trajectory at once. Instead, it first converts the gridded execution trajectory into executable segment-level control data. This is because in complex and confined spaces such as alleyways, there are bottleneck sections, turning sections, intersection sections, and partially obstructed sections. Furthermore, the wireless link latency varies with spatial location. If the entire trajectory is sent out, it is easy for issues such as the arrival time of the instruction to be inconsistent with the refresh time of the on-site perception, overlapping times when multiple agents enter the same confined section, or the subsequent agent entering the channel before the preceding agent has released it, based on the original trajectory, to occur. Therefore, node-level extraction, link latency evaluation, spatiotemporal matching, time compensation, segmented encoding, and pre-send processing are required for the execution trajectory to output a segmented instruction sequence with a preset time window.

[0072] Path node parameters include at least the node's spatial coordinates, heading status, planned passage time, trajectory segment identifier, and speed constraint information. The node's spatial coordinates are determined by the center coordinates of the grid cells the trajectory passes through or the coordinates of control points after path smoothing; the heading status is determined by the direction of the line connecting adjacent nodes or the tangential direction of the smoothed curve; the planned passage time is determined by the S5 time window constraint and the expected arrival order; and the speed constraint information is jointly determined by the passage width of the grid area where the node is located, turning constraints, risk level, and the current agent's maneuverability. Nodes can also be classified as straight-ahead nodes, turning nodes, bottleneck entry / exit nodes, waiting nodes, and operation start / end nodes. For example, a turning node is marked when the directional change between adjacent trajectory segments exceeds a preset turning threshold; a bottleneck entry / exit node is marked when the corresponding grid area in S5 is marked as unidirectional or cannot be occupied by two machines in parallel; and an operation node is marked when it triggers an operation action. After extraction, a node parameter table is formed.

[0073] The control server uses heartbeat packets to probe the quality of the wireless communication links with each agent to assess real-time communication latency. The control server sends probe messages to each agent according to a preset heartbeat cycle. After receiving the probe messages via their local control module, the agents send back acknowledgment messages via their communication modules. The control server calculates the round-trip time (RTD) based on the sending and receiving times of the acknowledgment messages, and takes half of the RTD as an estimate of the current one-way communication latency. One-way latency can be measured directly under clock synchronization conditions; otherwise, the RTD estimate is used. To reduce instantaneous jitter errors, a moving average is applied to the one-way communication latency over the most recent heartbeat cycles, and the latency fluctuation is recorded. If the wireless link is a self-organizing multi-hop link, the control server can also verify whether the one-way communication latency is consistent with the current forwarding topology by considering the number of forwarding nodes corresponding to the current routing path. When the number of forwarding nodes on a certain path increases, the lead time for the corresponding node's trajectory segment increases synchronously.

[0074] Spatiotemporal matching does not uniformly set the refresh frequency of all sensors. Instead, it determines the primary constraint sensor for each node based on the node type and its control objective. For example, the primary constraint sensors for a straight-ahead node might be an attitude sensor and an odometer; for a turning or obstacle avoidance node, they might be a ranging sensor and an attitude sensor; for a work start / end node, the primary constraint sensors might also include a working mechanism status sensor; and for bottleneck entry nodes and meeting-related nodes, the primary constraint sensors might include a vision sensor or an occupancy confirmation sensor. The control server reads the refresh frequency of the primary constraint sensor and calculates the corresponding refresh cycle. Then, combining this with the current one-way communication delay and the local control module's parsing time for the control message, it obtains the estimated time offset when the instruction arrives at the node. The offset can be composed of the one-way communication delay, the local parsing delay, and the waiting time between the instruction arrival time and the next valid sensor refresh time. The local parsing delay can be obtained through pre-task testing or statistical results from recent control cycles; the waiting time is determined based on the position of the expected arrival time of the control message relative to the refresh phase of the primary constraint sensor. For example, if a turning node has a one-way communication latency of 100ms, a local resolution latency of 30ms, a ranging sensor refresh cycle of 50ms, and a message is expected to arrive 20ms after the start of a certain refresh cycle, then it will take another 30ms to wait for the next valid refresh, corresponding to an estimated time offset of 160ms. The control server writes the estimated time offset for each node into the node execution timing table.

[0075] During compensation adjustments, the control server does not alter the global path topology and node sequence already established by S5. Instead, it adjusts the planned passage time, segment start time, and intra-segment speed constraints without compromising spatial reachability and time window occupancy constraints. For ordinary nodes with small estimated time offsets and no bottleneck channel occupancy conflicts, only the planned passage time is corrected. For turning nodes, bottleneck entry nodes, meeting-related nodes, or operation start nodes, in addition to correcting the planned passage time, the speed constraints of the preceding or current segment are adjusted based on the remaining margin of the corresponding time window. For nodes with estimated time offsets exceeding a preset threshold and high-risk obstacles or unreleased occupied areas in front of them, a waiting node is inserted before the node, and the waiting action and subsequent entry action are divided into two adjacent instruction segments to prevent the agent from entering the risk area prematurely when the link latency is large or the perception has not been updated. A preset time window is used to define the allowed start and completion intervals for each segmented instruction. Its starting boundary is at least related to the planned passage time of the first node of the segment, the one-way communication delay, the local resolution delay, and the refresh cycle of the main constraint sensor. Its ending boundary is at least related to the cooperative safety interval, the space occupancy release requirements, and the allowed entry time of adjacent agents. When the segment is located in a bottleneck channel or intersection section, a spatial conflict buffer time is also added. After time compensation is completed, the control server divides the trajectory according to continuous nodes, common velocity constraints, and common time window constraints, forming multiple segmented instruction segments. Each segmented instruction segment includes at least the segment start node, segment end node, segment node coordinate sequence, segment heading sequence, segment velocity constraint sequence, segment corresponding time window, and segment operation action information.

[0076] After forming the segmented instruction sequence, the timing of instruction issuance is optimized. Real-time response speed is the execution response time from when the local control module receives the segmented instruction sequence to when the drive mechanism or operating mechanism begins to execute the corresponding control action. This can be obtained through short message testing before the task starts, no-load action testing, or statistical results from multiple recent control cycles. The control server determines the issuance timing based on a comparison between "the remaining time from the current moment to the planned start time of a certain segment" and "the sum of the one-way communication delay, local execution response time, and safety buffer time corresponding to the first node of that segment." If the remaining time is greater than the sum of the above values, the corresponding segment is written to the pre-issuance queue; if the remaining time is less than or equal to the sum of the above values, the issuance of that segment is executed immediately. To ensure seamless transitions between segments, the next segment is written to the local cache of the corresponding agent before the previous segment has finished executing. If the current link fluctuation exceeds a preset fluctuation threshold, or if there is unreleased space in the area corresponding to the first node of the next segment, a transition segment containing only low-speed hold actions or waiting actions is prioritized. Subsequent segments are only issued after receiving release confirmation or an updated link status. After a segmentation command is issued, the control server records the segment's sending time, confirmation reception time, local cache status, and planned start time for use in subsequent steps.

[0077] S7. During task execution, each agent collects its state feedback data in real time and transmits it back to the control server. If the deviation between the state feedback data and the expected execution state exceeds a preset threshold, the state feedback data is inverted and verified. The global environment state mapping is updated based on the inversion result, and the sequence of unexecuted segmented instructions is adjusted.

[0078] In one specific embodiment, the process of performing step S7 may specifically include the following steps: The control server extracts real-time energy consumption feedback data, path consumption data, and real-time environmental change characteristics from the status feedback data; The real-time energy consumption feedback data, path consumption data, and real-time environmental change characteristics are compared with the preset expected execution state to calculate the execution deviation. If the execution deviation exceeds the preset threshold, the status feedback data is inverted and verified. Based on the inversion verification results, the real-time environmental change characteristics are mapped back to the global coordinates through data correlation analysis, and the global environmental status mapping is updated. Based on the updated global environment state mapping, the segmented instruction sequences that are not currently executed by each agent are dynamically adjusted.

[0079] Specifically, during task execution, various situations will continuously occur, including obstacle displacement, changes in passable width, wind disturbances, dust obstruction, changes in channel occupancy, and load fluctuations. These changes may occur after the command is issued. In this step, the control server does not make static judgments based on the original trajectory. Instead, it first determines whether there is an excessive deviation between the actual execution state and the expected execution state based on the status feedback data. When the triggering conditions are met, it performs inversion verification to determine the cause, location, and scope of the environmental changes that caused the deviation. Subsequently, it updates the global environmental state mapping and dynamically adjusts only the currently unexecuted segmented command sequences accordingly.

[0080] Each agent continuously receives outputs from the motion state detection module, energy consumption monitoring module, and multimodal perception module, and encapsulates them according to a unified timestamp to form state feedback data, which is then uploaded to the control server. The motion state detection module outputs at least the distance traveled, actual path length, current speed, trajectory offset, steering change, and actual node passage time, forming path consumption data. The energy consumption monitoring module outputs at least instantaneous power, cumulative power consumption, remaining power, and energy consumption per unit path, forming real-time energy consumption feedback data. The multimodal perception module outputs at least local obstacle changes, passable boundary changes, channel occupancy changes, local wind speed changes, dust concentration changes, and harmful gas concentration changes, forming real-time environmental change characteristics. The state feedback data also includes agent identifier, current segment identifier, current node identifier, acquisition timestamp, and current pose information.

[0081] After receiving the status feedback data, the control server extracts and aligns the data. It trims the uploaded data into time windows according to the start and end times of the segments corresponding to the segment instruction sequence, and aligns various data types within the same segment to a unified segment-level time base. For high-frequency uploaded data, it takes the average, peak, or last time value within the segment's time window; for low-frequency uploaded data, it takes the most recent valid value and retains its timestamp. The extracted data forms a segment-level feedback table, which includes at least the agent identifier, segment identifier, node range, energy consumption feedback item, path consumption item, and environmental change item.

[0082] The expected execution state is generated jointly by the outputs of S5 and S6. Specifically, the control server uses the corresponding subtask local trajectory node sequence, gridded path length, risk area distribution, and cost constraints generated by S5 as spatial and path benchmarks, and the segment start time, node planned passage time, intra-segment speed constraints, waiting segment settings, and time window compensation results generated by S6 as time and control benchmarks. Combined with the agent type, payload, historical calibration data, and corresponding environmental parameters in the global environment state mapping, it generates a set of target quantities and tolerance quantities that the segment should satisfy under normal environment and normal execution conditions, thus forming the expected execution state. This includes at least the expected unit path energy consumption benchmark value or allowable range, the expected cumulative energy consumption increment benchmark value or allowable range corresponding to energy consumption; the expected node passage time and its time tolerance, the expected intra-segment path length and its length tolerance, the expected intra-segment speed range, the expected trajectory offset tolerance, and the expected turning change tolerance corresponding to the path; and the expected obstacle boundary state, the expected passable width range, the expected channel occupancy state, and the expected environmental disturbance benchmark value or allowable range corresponding to the environment. The expected path length within the segment, the expected node passage time, the expected speed range within the segment, and the expected turning change tolerance are preferably taken directly from the planning and compensation results of S5 and S6; the expected unit path energy consumption benchmark or allowable range is determined based on the agent type, load, speed constraints within the segment, slope within the segment, and historical calibration data; the expected environmental disturbance benchmark or allowable range is determined based on the wind speed change factor, obstacle density, passable width fluctuation, and hidden interference source risk level of the spatial unit where the segment is located in the global environmental state mapping.

[0083] For single-value expected items, the execution deviation is determined by the difference or proportion between the actual value and the expected value. For interval expected items, if the actual value falls within the corresponding allowable interval, the deviation is recorded as zero or as a deviation within the allowable range; otherwise, the deviation is determined by the amount exceeding the interval boundary. Execution deviation includes at least energy consumption deviation, path deviation, and environmental deviation: Energy consumption deviation is determined by the difference between the actual unit path energy consumption and the expected unit path energy consumption benchmark, or by the amount by which the actual unit path energy consumption exceeds the allowable range of the expected unit path energy consumption; Path deviation is determined by a combination of the difference between the actual path length and the expected path length within the segment, the amount by which the actual node passage time exceeds the corresponding time tolerance range, the amount by which the actual trajectory offset exceeds the trajectory offset tolerance, and the amount by which the actual turning change exceeds the turning change tolerance; Environmental deviation is determined by the difference between the actual obstacle boundary changes, passable width changes, local wind speed changes, and channel occupancy changes in the current segment area and the expected environmental disturbance benchmark, or by the amount by which they exceed the corresponding allowable range of the expected environmental disturbance. The control server first calculates the energy consumption deviation, path deviation, and environmental deviation values ​​separately, normalizes these three values, and then weights and sums the normalized values ​​according to preset weights to calculate the overall execution deviation value. Specifically, the path deviation value can be obtained by a normalized weighted sum of path length deviation, node passage time deviation, trajectory offset deviation, and turning change deviation; the environmental deviation value can be obtained by a normalized weighted sum of obstacle boundary change deviation, passable width deviation, wind speed change deviation, and channel occupancy change deviation. Preferably, for bottleneck sections, meeting sections, or high-risk sections, the judgment weights of path deviation and environmental deviation are increased; for long-distance transportation sections or continuous cleaning sections, the joint judgment weight of energy consumption deviation and path deviation is increased.

[0084] When the execution deviation does not exceed the preset threshold, the current global environment state mapping and instructions remain unchanged, and the current execution result is written to the historical cache. When the deviation exceeds the preset threshold, inversion verification is triggered. If S2 uses a Bayesian network model to identify hidden interference sources, S7 will not add an independent training process. Instead, it will input the increased energy consumption per unit path, increased node throughput delay, increased trajectory offset, sudden increase in local wind speed, forward movement of obstacle boundaries, and changes in channel occupancy status corresponding to the current segment as new observation evidence into the Bayesian network model. It will update the posterior probabilities of candidate environmental causes such as increased local obstacle accumulation, increased risk of local collapse, abnormal wind field enhancement, and dynamic blockage of blind spots in the corresponding area. If the posterior probability of a candidate environmental cause exceeds the preset probability threshold, it will be regarded as the main environmental cause corresponding to the current execution deviation. If the posterior probabilities of all candidate environmental causes do not exceed the threshold, the deviation will be regarded as a local execution deviation, and only the time window, speed constraint, or waiting period of the unexecuted segment will be adjusted, without immediately rewriting the global environment state mapping.

[0085] When the inversion verification results support that the environmental state has changed, an environmental state update is performed. The control server reads the pose information, segment identifier, node range, and timestamp of the agent that triggered the deviation at the corresponding moment in that segment. Combined with the installation extrinsic parameters and detection range of the multimodal perception module, it transforms local observation results such as changes in local obstacles, traversable boundaries, wind speed, and passage occupancy from the local coordinate system to the global coordinate system. Then, it performs data association analysis based on the grid cell range corresponding to the current segment to determine the spatial location, temporal location, and impact range of the aforementioned changes. Preferably, the data association analysis employs dual constraints of time threshold and spatial neighborhood threshold. That is, only when change information from one or more agents falls within the same time window and the corresponding spatial location is located in adjacent grids or the same region is it determined to be the same environmental change event. If multiple agents provide consistent observations of the same spatial cell, the confidence level of the change event is increased.

[0086] After completing the data association analysis, the global environmental state mapping is updated. For changes in obstacle boundaries, the obstacle occupancy status, passable boundaries, and safety buffer range of the corresponding spatial unit are updated; for changes in passage occupancy, the time window occupancy status and cooperative avoidance constraints of the corresponding spatial unit are updated; for changes in wind speed, the wind speed magnitude, wind direction, and wind speed gradient of the corresponding spatial unit are updated; for environmental changes such as increased dust obstruction or increased concentration of harmful gases, the risk level and disturbance item status of the corresponding spatial unit are updated; if the Bayesian network model determines that the posterior probability of a candidate hidden disturbance source exceeds a preset probability threshold after online inversion, the candidate result is written into the hidden disturbance source status field of the corresponding spatial unit. To avoid mapping jitter caused by a single false detection, confidence-weighted fusion or time-decay fusion is performed on the old and new states of the same spatial unit. That is, for changes observed consistently by multiple agents and with high confidence, the writing weight is increased; for changes observed only once and with low confidence, only a limited magnitude update is performed.

[0087] Dynamic adjustment targets are limited to segments that are not currently being executed. For segments that are being executed but allow for safe interruption, it is preferable to first generate waiting segments, low-speed holding segments, or safe exit segments, and then connect the adjusted subsequent segments. During dynamic adjustment, segments affected by the updated region are first identified. Affected segments include those whose path nodes are located in the updated region, whose time windows conflict with the updated occupancy state, whose speed constraints do not match the updated environmental risks, or whose action types do not match the updated operating conditions. At least one of the following adjustments is performed on the affected segments: adjusting the segment start time and node planned passage time, adjusting the speed constraints within the segment, inserting waiting nodes or yielding nodes, shortening the original segment and re-splitting subsequent segments, changing the local path node sequence of the affected segment, or, if necessary, reverting the corresponding subtask to S5 to replan the trajectory. If the updated global environment state mapping shows that a certain area is inaccessible, the unexecuted entry segment targeting that area is canceled, and a waiting segment or detour segment is generated for the corresponding agent; if it shows that the original waiting conditions have been eliminated, the entry segment after the corresponding waiting segment is released.

[0088] S8. Based on the adjusted segmented instruction sequence and the updated global environment state mapping, generate the execution flow of multi-machine collaborative operation in real time until the operation requirements are completed.

[0089] In one specific embodiment, the process of executing step S8 may specifically include the following steps: Integrate the currently adjusted segmented instruction sequences of each agent; By combining the real-time update frequency of the state feedback data and based on the updated global environment state mapping, the wind speed change factor and global energy consumption constraints in the environment are dynamically adjusted. Based on the integrated segmented instruction sequence, dynamically adjusted wind speed change factor, and global energy consumption constraints, a refined execution process for multi-machine collaborative intelligent operation is generated in real time through iterative generation. Based on the refined execution process, the overall coordination arrangement of the operation is determined, each intelligent agent is scheduled to execute continuously, and the inversion verification and instruction adjustment steps are returned in a loop until all sub-tasks covered by the multi-machine collaborative operation requirements are completed.

[0090] Specifically, the control server merges the unexecuted segmented instruction sequences of each agent according to a unified global time base, spatial coordinate system, and task constraints to form a global segmented scheduling table. This table includes at least the agent identifier, segment identifier, subtask identifier, segment start / end node, spatial unit range of the segment, segment start / completion time window, velocity constraint, action type, preconditions, and completion conditions. If a segment has not yet been executed but its preconditions are not met, it is marked as a segment to be activated; if multiple segments simultaneously meet their preconditions, they are retained as candidate executable segments.

[0091] S8 operates not on individual feedback variables, but rather on global scheduling constraints for multiple agents whose segments are currently not executed. Therefore, the real-time update frequency of state feedback data adopts a component-based update method. State feedback data includes at least real-time energy consumption feedback data, path consumption data, and real-time environmental change characteristics. Therefore, instead of updating all constraints at a single fixed period, the update intervals related to wind speed change factor updates—including local wind speed changes, obstacle boundary changes, passable width changes, and channel occupancy changes—are calculated separately. The update intervals related to global energy consumption constraint updates—including unit path energy consumption changes, cumulative energy consumption increment changes, remaining power changes, and current load state changes—are also calculated. The environmental constraint refresh cycle and energy consumption constraint refresh cycle are then determined based on the update intervals of different feedback items. If the update interval for wind speed-related feedback items shortens, it indicates that the wind field state in the candidate segment area is changing faster, and the control server correspondingly shortens the refresh cycle of the wind speed change factor. If the update interval for energy consumption-related feedback items shortens, it indicates that the energy consumption state of at least one agent is changing faster, and the control server correspondingly shortens the refresh cycle of the global energy consumption constraint.

[0092] The dynamic adjustment of the wind speed change factor is based on the local wind speed change results mapped back to the global coordinate system in S7, combined with the updated global environmental state mapping, to re-correct the wind speed magnitude, wind direction change, and wind speed gradient of the spatial unit where the candidate segment is located and its adjacent spatial units. If multiple agents report consistent local wind speed change results for the same spatial area within adjacent time windows, the confidence of the wind speed update for that area is increased, and the constraint weight of the wind speed change factor in the current round of execution is correspondingly increased. If there is only a single observation by a single agent, the wind speed change factor for that area is corrected to a limited extent based on the confidence information formed in S7, without immediate and significant rewriting. After completing the spatial state correction, the corrected wind speed change factor is reread according to the spatial unit where the current candidate segment is located, and the actual effective wind speed impact value for that segment is calculated by combining the travel direction, operation direction, and equipment wind resistance capability of each segment. If the direction of travel of a candidate segment is highly sensitive to the local crosswind direction, the allowable speed range of the segment will be tightened, the time window margin of the segment will be extended, or the start priority of the segment will be reduced in the current round of refined execution process; if a candidate segment is located in an area with small wind speed changes and stable wind direction, its original speed constraints and time window will be maintained.

[0093] The dynamic adjustment of global energy consumption constraints involves reading the unit path energy consumption, cumulative power consumption, remaining power, current load status, and path length of the corresponding segment extracted from S7. This data, combined with the wind speed variation factor, slope conditions, obstacle complexity, and operational mechanism action type of the spatial unit where the candidate segment is located in the current round, is used to recalculate the estimated energy consumption of each agent's subsequent segments. The estimated segment energy consumption includes the movement energy consumption corresponding to the estimated path length of the segment and the mechanism energy consumption corresponding to the segment's operational actions. If a candidate segment is located in a high-wind-speed disturbance zone, a slope-varying zone, or a high-obstacle-complexity zone, the estimated energy consumption of that segment is increased accordingly. Based on the recalculation results, a remaining energy consumption budget table is established. This table includes at least the agent identifier, current remaining power, the total estimated energy consumption of allocated but unexecuted segments, the minimum energy consumption reserve, and the upper limit of allowed new segment energy consumption. If the total estimated energy consumption of an agent's allocated but unexecuted segments is close to its allowed energy consumption limit for adding new segments, the startup priority of its subsequent high-energy-consuming segments will be reduced, or some unstarted and transferable subtasks will be allocated to agents with more remaining energy. If an agent's remaining power is decreasing rapidly but its current segment is not yet completed, its subsequent unexecuted segments will be adjusted first, without directly interrupting its current executing segment, provided that the S6 time window synchronization relationship and S7 map update results are not disrupted. For example, for any agent, the minimum energy reserve is determined based on the minimum energy consumption required for the agent to safely exit the current work area from its current location and complete the minimum safe stay. This includes at least: the estimated energy consumption for moving from the current location along the shortest safe exit path to the safe zone, the energy consumption of the mechanism performing necessary braking / steering / staying actions, and a preset safety redundancy. The preset safety redundancy can be set according to a fixed proportion or a fixed absolute value of the aforementioned minimum energy consumption. For any agent, the maximum allowable energy consumption E1 for a new segment can be determined by the following formula: E1 = E2 − E3 − E4, where E2 is the available energy consumption corresponding to the current remaining power, E3 is the total estimated energy consumption of the segments already allocated to the agent but not yet executed, and E4 is the minimum energy reserve. When E1 ≤ 0, it is determined that the agent does not currently have the energy consumption conditions to continue receiving new segments; when the estimated energy consumption of a new candidate segment is greater than E1, the new candidate segment will not be allocated to the agent.

[0094] Preferably, task reallocation is triggered when any of the following conditions are met: First, the maximum allowable energy consumption of a new segment for an agent is less than the preset minimum executable segment energy consumption threshold; second, the total estimated energy consumption of subsequent segments to be executed by an agent exceeds the upper limit after deducting the minimum energy reserve from the current remaining available energy consumption; third, the updated wind speed change factor, space occupancy status, or risk level renders the original segment assigned to the agent unexecutable in the current round and can be replaced by another agent for execution. When task reallocation is triggered, only subtasks that have not yet been started and meet the transfer conditions are reallocated.

[0095] The refined execution process is a multi-agent collaborative scheduling result formed for the current execution round. It includes at least the set of segments allowed to be started in the current round, the start order, waiting order, yielding order, occupation order, release order, segment switching conditions, and cross-agent coordination relationships for each segment. Based on the global segment scheduling table, the control server filters currently unexecuted segments by round and sequentially performs space occupancy constraint judgment, wind speed safety constraint judgment, energy consumption executable constraint judgment, prerequisite task completion constraint judgment, and collaborative consistency constraint judgment. Segments that simultaneously meet the above constraints are added to the executable set for the current round; segments that do not meet the constraints are retained in the pending execution set, and a blocking reason field is written to them. The collaborative consistency constraint determination includes at least the following criteria: 1. Whether the occupancy order of shared channels or shared work surfaces across agents is consistent with the published entry, waiting, yielding, and release order; 2. Whether the start time of a segment is consistent with the completion and release conditions of its associated preceding segment; 3. Whether multiple parallel candidate segments within the same time window will cause subsequent critical path segments to fail to start as planned; 4. Whether the switching conditions, action types, and speed constraints between adjacent segments of the same agent are compatible. A candidate segment is considered to satisfy the collaborative consistency constraint only when it simultaneously meets the above consistency criteria. Subsequently, a hierarchical sorting is performed on the executable set of the current round. The sorting criteria include at least the subtask priority, the degree of critical path constraint, the risk level corresponding to the wind speed change factor, and the degree of matching between the expected energy consumption and the remaining power. The degree of critical path constraint is used to characterize the impact of a candidate segment on the advancement of the critical path. The higher the degree of critical path constraint of a candidate segment, the more it is located at the forefront of the task link corresponding to the critical path, the more it undertakes the release function of bottleneck space units or shared channels, and the more likely its delay is to cause the delay of the start of subsequent segments on the critical path or the extension of the overall project duration. Preferably, the degree of critical path constraint is determined based on at least one of the following: whether the candidate segment is a component segment of the subtask corresponding to the critical path; whether the completion of the candidate segment releases critical bottleneck space units or shared channels; whether the delay of the candidate segment causes the expected start time of subsequent segments on the critical path to be postponed; and whether the candidate segment occupies the shared resources required by the critical path segments. The more of the above factors a candidate segment satisfies, or the greater its impact on the advancement of the critical path, the higher its degree of critical path constraint is determined, and its ranking priority is increased accordingly.

[0096] If multiple segments are spatially independent, they are allowed to be executed in parallel. If multiple segments are spatially adjacent but do not conflict, their order is determined based on their impact on the release of subsequent critical paths. After sorting, an execution orchestration table for the current round is formed, which includes at least the list of segments to be started in the current round, the start time of each segment, the triggering conditions of each segment, the corresponding agent for each segment, the release conditions after the segment is completed, and the set of candidate segments for the next round.

[0097] The overall coordination of the operation is determined based on the execution schedule, which includes at least the order of entry, passage, waiting, yielding, occupying, and exit for each agent in the current round. For scenarios with shared work surfaces or shared channels, shared resource occupation windows and release windows are also written. The control server issues execution permits or waiting permits to each agent according to the overall operation coordination schedule. The current round starts the segment, using the segment instruction sequence formed in S6 and adjusted in S7 as the underlying execution content; segments that are pending execution but have not yet met the conditions are maintained as waiting segments, low-speed holding segments, or safe dwell segments. After receiving the corresponding execution permission, each agent continues to execute according to the current segment instruction sequence and continues to send status feedback data back to the control server through S7. When the status feedback data arrives, the local environmental state changes, the wind speed change factor is updated, the global energy consumption constraint changes, or a subtask is completed early or delayed, the agent rereads the updated global environmental state mapping and the set of currently unexecuted segments from S7, and then performs the integration, constraint update, process iteration, and coordination arrangement in this step to generate the next round of refined execution process until all subtasks are completed and all segments have been executed or legally revoked.

[0098] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A data inversion enhanced multi-swarm collaborative work method based on vertical domain large model, applied to a swarm collaborative system comprising a control server and multiple agents, characterized in that, The methods include: S1. Each intelligent agent collects real-time multimodal perception data of the target operating environment, and sends it to the control server after local preprocessing; S2. The control server inputs the multimodal sensing data into the built-in first vertical domain large model for feature extraction, and combines the data inversion mechanism to identify hidden interference sources and construct a global environment state mapping. S3. The control server receives the multi-machine collaborative operation requirements, uses the second vertical domain large model to perform semantic parsing on the multi-machine collaborative operation requirements, decomposes them into various sub-tasks, and determines the task distribution range of each sub-task based on the global environment state mapping. S4. For each subtask within the task distribution range, extract the spatiotemporal dependencies between tasks, perform anti-conflict priority sorting through the collective intelligence collaboration algorithm, and generate a conflict-free task sequence. S5. Assign tasks to each agent according to the task sequence, calculate the penalty cost in conjunction with the global environment state mapping, and plan a gridded execution trajectory for each agent. S6. The control server extracts the path node parameters of the gridded execution trajectory and evaluates the real-time communication delay with each agent. Then, it performs spatiotemporal matching of the path node parameters, the real-time communication delay, and the sensor refresh frequency of each agent to generate a segmented instruction sequence with a preset time window and sends it to the corresponding agent. S7. During task execution, each intelligent agent collects its status feedback data in real time and transmits it back to the control server. If the deviation between the status feedback data and the expected execution state exceeds a preset threshold, the status feedback data is inverted and verified. The global environment state mapping is updated according to the inversion result, and the unexecuted segmented instruction sequence is adjusted. S8. Based on the adjusted segmented instruction sequence and the updated global environment state mapping, generate the execution flow of multi-machine collaborative operation in real time until the operation requirements are completed.

2. The method of claim 1, wherein S2 include: The multimodal sensing data is input into the first vertical domain large model, and features are extracted by combining the preset domain knowledge base to generate an environmental feature description that includes dynamic obstacle factors and wind speed change factors. Based on the environmental feature description, the hidden interference sources and their distribution status in the target working environment are determined by reverse deduction using the Bayesian network model in the data inversion mechanism. The environmental feature description and the distribution status of the concealed interference source are fused to construct the global environmental state mapping that includes environmental dynamic features and interference items.

3. The method of claim 1, wherein S3 include: The control server inputs the received multi-machine collaborative operation requirements into the second vertical domain large model, extracts the operation intention and operation goal, and semantically decomposes the multi-machine collaborative operation requirements into multiple sub-tasks; Environmental dynamic features and interference terms are extracted from the global environmental state mapping, wherein the environmental dynamic features include dynamic obstacle locations and wind speed change factors, and the interference terms are identified hidden interference sources. The workspace requirements corresponding to each sub-task are superimposed with the dynamic characteristics of the environment and the spatial distribution of the interference items, and a safety boundary verification is performed. Based on the verification results, areas that conflict with the hidden interference sources and exceed the wind speed safety threshold are eliminated, and the actual task distribution range of each sub-task in the workspace is determined.

4. The method according to claim 1, characterized by S4 include: The local spatial environment details corresponding to each sub-task are reconstructed using a local reconstruction method. By combining the reasoning logic of the task context with the details of the local spatial environment, the spatiotemporal dependencies between the various subtasks are analyzed and extracted; Based on the aforementioned spatiotemporal dependency, a collective intelligence collaboration algorithm is used to determine whether there are temporal or spatial execution conflicts in the execution of each subtask. If a conflict exists, the environmental feature complexity index of the region corresponding to the subtask that caused the conflict is extracted, and the conflict prevention priority of each subtask is ranked according to the index. If there is no conflict, the priority order of each subtask is directly determined based on the spatiotemporal dependency relationship. Based on the sorting results, a conflict-free task sequence for multi-machine collaborative operation is generated.

5. The method according to claim 1, characterized by S5 include: Assign corresponding subtasks to each agent according to the conflict-free task sequence; Based on the global environment state mapping, the environmental dynamic features and disturbances within the workspace are extracted, the workspace is divided into grids, and the penalty cost of each grid unit is calculated according to the complexity of the environmental dynamic features and the danger level of the disturbances, thus constructing a gridded cost map. Based on the gridded cost map, a gridded path search technique is used to plan gridded execution trajectories for each agent to avoid high-risk areas, with the goal of minimizing cumulative penalty costs.

6. The method of claim 1, wherein S6 include: Extract the spatial coordinates and heading status of the gridded execution trajectory as the path node parameters; The quality of the wireless communication link between the control server and each intelligent agent is detected by heartbeat packets to assess real-time communication latency. The path node parameters, the real-time communication delay, and the refresh frequency of the sensors carried by each intelligent agent are spatiotemporally matched to analyze the estimated time offset of the command reaching each path node. The path node parameters are compensated and adjusted in the time dimension based on the estimated time offset so that the arrival time of the adjusted instructions meets the synchronization requirements of multi-machine collaboration, and the segmented instruction sequence with a preset time window is generated. By combining the real-time communication latency with the real-time response speed of each agent, the timing of instruction issuance is optimized, and the segmented instruction sequence is issued to the corresponding agent for execution.

7. The method of claim 1, wherein, S7 includes: The control server extracts real-time energy consumption feedback data, path consumption data, and real-time environmental change characteristics from the status feedback data; The real-time energy consumption feedback data, the path consumption data, and the real-time environmental change characteristics are compared with the preset expected execution state to calculate the execution deviation. If the execution deviation exceeds the preset threshold, the state feedback data is inverted and verified. Based on the inversion verification result, the real-time environmental change characteristics are mapped back to global coordinates through data correlation analysis, and the global environmental state mapping is updated. Based on the updated global environment state mapping, the segmented instruction sequences that are not currently being executed by each agent are dynamically adjusted.

8. The method of claim 1, wherein, S8 includes: Integrate the currently adjusted segmented instruction sequences of each agent; Based on the real-time update frequency of the state feedback data and the updated global environment state mapping, the wind speed change factor and global energy consumption constraints in the environment are dynamically adjusted. Based on the integrated segmented instruction sequence, dynamically adjusted wind speed change factor, and global energy consumption constraints, a refined execution process for multi-machine collaborative intelligent operation is generated in real time through iterative generation. Based on the refined execution process, the overall coordination arrangement of the operation is determined, each intelligent agent is scheduled to execute continuously, and the inversion verification and instruction adjustment steps are returned in a loop until all sub-tasks covered by the multi-machine collaborative operation requirements are completed.

9. The method of claim 1, wherein, The intelligent agent is equipped with a multimodal perception module, a motion state detection module, an energy consumption monitoring module, a communication module, a local control module, and a cooperative operation execution mechanism, among which: The multimodal perception module includes one or more of a visual acquisition circuit, a ranging acquisition circuit, an attitude detection circuit, and an environmental parameter detection circuit, used to acquire real-time multimodal perception data of the target working environment and real-time environmental change information during task execution; the motion state detection module is used to acquire path consumption data, the energy consumption monitoring module is used to acquire real-time energy consumption feedback data, the local control module is used to generate state feedback data based on the real-time environmental change information, the path consumption data, and the real-time energy consumption feedback data, and to parse the received segmented instruction sequence to generate corresponding underlying control signals; the communication module is used to receive instructions issued by the control server and send the real-time multimodal perception data and the state feedback data to the control server; the collaborative operation execution mechanism is used to respond to the underlying control signals and execute corresponding sub-tasks.

Images