Multi-agent collaboration method and system
By deploying signal enhancement points and time window synchronization mechanisms in uninhabited areas, and combining them with distributed ledger technology, the problems of sensor bias correction and collaborative decision-making updates in multi-agent systems in uninhabited areas were solved, thereby improving search and rescue efficiency and decision consistency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NEUSOFT INST GUANGDONG
- Filing Date
- 2025-12-16
- Publication Date
- 2026-05-12
Smart Images
Figure CN122028005A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of multi-agent control technology, and in particular to a multi-agent cooperative method and system. Background Technology
[0002] In recent years, with the rapid development of artificial intelligence and robotics, multi-agent systems have been widely used in complex tasks such as search and rescue, environmental monitoring, and military reconnaissance. Through the collaborative efforts of multiple agents, multi-agent systems can accomplish complex tasks that are difficult for a single agent to handle, offering advantages such as high efficiency, strong robustness, and wide coverage.
[0003] Traditional multi-agent cooperative systems primarily rely on stable communication network infrastructure. In urban environments, agents can communicate in real time via 4G / 5G networks, Wi-Fi, and other methods, enabling data sharing, task allocation, and collaborative decision-making. Global positioning systems such as GPS provide agents with precise location information, supporting navigation and positioning functions. In such network-covered environments, multi-agent systems can achieve efficient collaborative operations.
[0004] However, in uninhabited areas where communication network infrastructure is lacking or there is no signal at all, there are still some problems in multi-agent collaborative task execution. Summary of the Invention
[0005] To address the aforementioned technical problems, or at least partially address them, this application provides a multi-agent collaborative method and system that can effectively update the deviation correction of agent sensor data and collaborative decision-making in environments without communication infrastructure. Through periodic centralized correction, time window synchronization mechanisms, and distributed ledger technology, it ensures efficient collaborative operation of the multi-agent system in complex, uninhabited environments.
[0006] In a first aspect, this application provides a multi-agent cooperative method, the method comprising the following steps: Signal enhancement points are deployed in uninhabited areas. These signal enhancement points include a central base station, deviation correction equipment, and decision update equipment. Multiple heterogeneous intelligent agents are deployed to perform search and rescue missions, with each agent equipped with sensors and communication modules; During the execution of the task, the intelligent agent periodically returns to the signal enhancement point. When it enters the signal coverage area, the communication module establishes a connection with the central base station and uploads sensor data. The central base station is used to communicate with the intelligent agent to obtain the sensor data of the intelligent agent and forward it to the deviation correction device and the decision update device; Deviation correction equipment performs deviation correction processing on the sensor data of the intelligent agent; The decision update device determines whether the collaborative decision needs to be updated based on the corrected data. If so, it generates a decision update package and broadcasts it to all agents within the signal coverage area through the central base station. After completing the calibration and update, the agent leaves the signal coverage area to continue its task.
[0007] Optionally, the method further includes setting a time window mechanism: Calculate the time window based on the task area and the agent's mobility. : in, This represents the distance from 85% of the agents to the signal booster point. The average velocity of all agents; After completing the deviation correction, the agent stays within the signal coverage area until the time window ends before leaving, to ensure that the agent that arrives later can update the important environmental information it carries with the agent that arrives earlier, and to avoid decision-making asynchrony caused by the return time difference.
[0008] Optionally, the time window mechanism further includes a dynamic speed control mechanism, which includes: The central base station counts the number N agents that have arrived and completed processing. arrived And send this statistic to each agent that is about to leave the central base station; Each agent calculates its own departure speed V based on the received data. exit The calculation formula is: The formula for calculating the dynamic weight w is: Among them, R coverage T is the signal coverage radius. window For the time window, T change For the battery swapping time of the intelligent agent, T process N represents the total processing time for data correction and decision updates, k is the adjustment coefficient, and N is the total processing time for data correction and decision updates. arrived N represents the number of agents that have arrived and completed processing. total The total number of intelligent agents; The speed control component calculates the departure speed V of each agent. exit The departure speed of each agent is adjusted to keep the departure time difference of all agents within a preset range, thus preventing frequent decision updates and task execution oscillations caused by excessive departure time differences.
[0009] Optionally, the multi-agent cooperative method further includes: Each agent is also equipped with a distributed ledger; Before leaving the signal coverage area, the intelligent agent records the latest decision content, timestamp, and digital signature generated by the decision update device into the distributed ledger; When performing tasks in areas without signal, if the distance between two agents is less than the preset synchronization distance, the communication module establishes a local communication connection and exchanges their respective ledger information. When an inconsistency is detected, the agent compares the timestamps of each decision in the ledger, automatically selects the decision with the latest timestamp and verified signature as the valid decision, and updates its own distributed ledger. This ensures that agents performing the same task in areas without signal coverage maintain decision consistency and prevents a majority of agents carrying outdated decisions from dominating erroneous decisions in local areas.
[0010] Optionally, the deviation correction device performs deviation correction processing on the sensor data of the intelligent agent, including the following steps: The agent's sensors include an inertial measurement unit, used to acquire the agent's acceleration data; The acceleration data uploaded by the agent is preprocessed, including format unification and sampling rate standardization, but timestamp alignment is not performed. Use Fast Fourier Transform to convert the acceleration data in the time domain to the frequency domain: Among them, a i (n⋅Δt) is the acceleration sample value of agent i at discrete time n·Δt, where N is the total number of sampling points, Δt is the sampling interval, f is the frequency, j is the imaginary unit, and A i (f) is the complex spectral value corresponding to frequency f; Deviation identification and correction are performed in the frequency domain. After correction, the time domain is transformed back by inverse FFT to avoid time domain alignment interpolation errors caused by the time difference of the agent's return, thus ensuring the accuracy of multi-agent joint correction.
[0011] Optionally, the deviation correction method further includes the following steps: Single agent correction: Analyze the frequency components of the preset ultra-low frequency band. If the amplitude exceeds the first threshold and exists in multiple axes, it is determined to be a sensor hardware deviation and is filtered out using a high-pass filter. Multi-agent joint correction: Calculate the mean μ(f) and standard deviation σ(f) of the frequency domain amplitudes of all agents that have reached them: If μ(f) exceeds the second threshold and σ(f) / μ(f) is less than the preset ratio within the preset low frequency band, it is determined that there is a sharing deviation caused by the environment at that frequency. Among them, the high-frequency part corresponds to the normal maneuvering actions of the intelligent agent and is not subject to deviation correction; Identify the set of shared bias frequencies F shared : Constructing an adaptive filter: Where fs is the identified shared bias frequency, α is the filter strength parameter, and Δf is the filter bandwidth; Application filter: Corrected acceleration data is obtained by inverse FFT transformation to eliminate different types of deviations caused by hardware aging and environmental interference.
[0012] Optionally, the decision update device updates the collaborative decision based on the corrected data, including the following steps: The sensors also include environmental sensing devices for acquiring visual and radar data; The decision update device performs target detection and environmental recognition on the visual and radar data uploaded by the intelligent agent; Task point update judgment: If a suspected trapped person is identified and there is no assigned search and rescue task point within the preset range of the location, the task update is triggered, a cost matrix is constructed and the task allocation algorithm is used to solve the optimal allocation scheme; Path update judgment: If a sudden change in environment is detected, the minimum distance between the obstacle and the planned path is calculated based on the corrected position information output by the deviation correction device. If the distance is less than the safety threshold, the path planning algorithm is used to replan the path. The decision update device generates a decision update package containing update content, timestamp, and digital signature, which is broadcast to all agents within the signal coverage area via the central base station to ensure that agents can update the latest decisions in a timely manner.
[0013] In a second aspect, this application provides a multi-agent cooperative system, the system including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the multi-agent cooperative method as described in any of the first aspects.
[0014] The technical solution provided in this application has the following advantages compared with the prior art: One of its beneficial effects and its working principle is as follows: In traditional urban environments or those with network coverage, intelligent agents can exchange data, correct their location, and coordinate tasks in real time using infrastructure such as 4G / 5G networks and GPS. However, in uninhabited areas (such as deep mountains or primeval forests), where there is no communication base station coverage, intelligent agents cannot communicate in real time. This leads to two serious problems: Sensor bias cannot be corrected: Sensors such as IMUs accumulate biases after long-term operation. In urban environments, this can be corrected in real time via GPS, but in uninhabited areas where there is no GPS signal, the biases will continue to accumulate, leading to problems such as positioning errors. Collaborative decision-making cannot be updated: When an agent discovers important information (such as trapped personnel or path obstacles), it cannot notify other agents, causing the entire team to continue to execute the old plan, which seriously affects the efficiency of search and rescue.
[0015] This application deploys signal enhancement points at the edge of uninhabited areas and designs a periodic return mechanism. The agent returns to the signal enhancement point at regular intervals, completing data upload, deviation correction, and decision updates within the signal coverage area. This centralized correction mechanism solves the communication problem in signal-free environments and fully utilizes the statistical characteristics of multi-agent data. It provides a reliable data foundation for subsequent decision updates and reduces erroneous decisions caused by positioning errors.
[0016] By performing periodic deviation correction at signal enhancement points, the problem of multi-agent systems being unable to perform deviation correction and decision updates due to the lack of communication infrastructure in uninhabited environments is effectively solved.
[0017] Its second beneficial effect and its working principle are as follows: In a search and rescue area of 80 square kilometers, agents performing tasks may return to signal enhancement points at significantly different times due to their varying locations and task distributions. This time difference can lead to decision-making asynchrony and oscillation effects. For example, suppose five agents are searching collaboratively in a certain area. The first agent returns to the base station at 60 minutes, while the fifth agent returns at 75 minutes. If the fifth agent discovers new trapped individuals and needs to change its search strategy, the first four agents may have already left the signal coverage area of the signal enhancement point, making it impossible to update the latest decision and causing decision-making asynchrony. Alternatively, suppose the agents can update their decisions in areas without signal coverage, but the first four agents have already left and implemented the old strategy for 15 minutes. When they meet in the area without signal coverage and synchronize their decisions, they find that the search path from the previous 15 minutes needs to be replanned. This is equivalent to wasting 15 minutes of work for four agents, causing a search that could have been completed in 90 minutes to take 105 minutes. This 15-minute delay further affects the return time in the next cycle, widening the time difference and creating a vicious cycle that causes oscillation problems.
[0018] The time window ensures that the first arriving agents wait sufficient time within the signal coverage area, allowing later-arriving agents carrying crucial environmental information to trigger decision updates and synchronize them with other agents. The speed control component dynamically adjusts the departure speed based on the number of arriving agents, keeping the time difference between agent departures within a shorter range. This mechanism avoids task execution oscillations caused by frequent decision updates while ensuring that most agents receive the latest decisions, significantly improving the collaborative efficiency of multi-agent systems.
[0019] By employing a time window mechanism and dynamic speed control, the problem of asynchronous decision-making and oscillation caused by the time difference in agent return is effectively solved, ensuring search and rescue efficiency while enabling timely updates of collaborative decisions.
[0020] Its third beneficial effect and its working principle are as follows: Although the time window covers 85% of the agents, 15% are unable to return in time due to distance or special circumstances. This can lead to situations where some agents carry new decisions while others carry old ones when performing tasks in areas without signal. For example, in a valley search area, three agents carrying old decisions are executing the original S-shaped search path, while one agent carrying new decisions knows it needs to switch to a circular search (because new terrain information has been discovered). Due to the lack of real-time communication without signal, their search paths will conflict, reducing search efficiency. The pseudo-majority problem refers to the situation where, although from a global perspective, agents carrying new decisions constitute a majority, from the perspective of a specific area, agents carrying old decisions may constitute a majority. If a simple majority vote is used, the three agents carrying old decisions may reject the one agent carrying new decisions, leading to the erroneous continuation of outdated solutions, or the authenticity of the decisions may be questioned because the number of agents carrying new decisions in the area is extremely small.
[0021] This application equips each agent with a distributed ledger and local area network (LAN) communication capabilities. Each agent updates its latest decision to its distributed ledger when leaving a signal coverage area. In areas without signal coverage, agents exchange ledger information via Wi-Fi Direct. The ledger records the decision content, millisecond-level timestamps, and 2048-bit RSA signatures. When a decision conflict occurs, the decision with the latest timestamp and valid signature is automatically selected, rather than a simple majority vote. Even if only one agent carries a new decision in a local area, as long as its timestamp is updated and its signature is valid, the new decision will be adopted and propagated in that area. The signature mechanism ensures that decisions cannot be forged, and the timestamp ensures the real-time nature of decisions, effectively solving the problems of decision conflicts and spurious majoritys in areas without signal coverage.
[0022] By adding a Wi-Fi communication module and a distributed ledger mechanism to the agent, the problems of decision-making conflicts and false majority in areas without signal are solved, ensuring that the agent can still maintain the final consistency of decisions even in the event of a complete communication interruption.
[0023] Its fourth beneficial effect and its working principle are as follows: Multi-agent joint calibration requires comprehensive analysis of data from multiple agents to identify shared biases caused by the environment (such as wind field effects). However, the return times of agents may differ by 10-15 minutes. Traditional methods require interpolating data from different times to align them to the same time reference. Large-span interpolation introduces errors, and in severe cases, may even misjudge real environmental deviations as interpolation noise, or vice versa, leading to calibration failure or even introducing new errors.
[0024] This application shifts the bias correction to the frequency domain. Environmental disturbances (such as stable wind fields) manifest as energy concentration at specific frequencies in the frequency domain, and this frequency domain characteristic is unaffected by the acquisition time. By using FFT transformation, the amplitude distribution of agent data returned at different times in the 0.01-0.1Hz frequency band is directly compared. When multiple agents exhibit similar frequency domain patterns (standard deviation / mean < 0.3), they can be accurately identified as shared environmental biases. Since the frequency domain characteristics are unaffected by time shift, this method avoids time-domain interpolation errors and ensures correction accuracy under large time difference conditions.
[0025] By cleverly circumventing the complexity of time-domain data alignment through frequency domain processing, it can still accurately perform multi-agent joint bias correction even when there is a large time difference in the agent's return. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating the multi-intelligent collaborative method provided in the embodiments of this application. Detailed Implementation
[0027] The technical solutions in this application will now be described with reference to the accompanying drawings.
[0028] Many specific details are set forth in the following description to provide a thorough understanding of this application, but this application may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of this application, and not all embodiments. It should be noted that, unless otherwise specified, the embodiments of this application and the features in the embodiments can be combined with each other.
[0029] Firstly, such as Figure 1 As shown, this application provides a method for deviation correction and decision update in multi-agent collaborative search and rescue missions in uninhabited areas. The implementation process of this method is described in detail below using a mountainous forest search and rescue scenario as an example.
[0030] S101: Agent deployment and time window calculation.
[0031] The search and rescue mission area covers approximately 80 square kilometers, with complex terrain including valleys, dense forests, and steep slopes, and lacks communication base station coverage. A temporary signal booster point will be deployed in the search and rescue area as a coordination center.
[0032] The central base station at each signal enhancement point is equipped with a deviation correction device and a decision update device. The deviation correction device is responsible for handling sensor data deviations of the intelligent agent, while the decision update device is responsible for determining whether the collaborative decision needs to be updated based on the corrected data. The signal enhancement point provides Wi-Fi signal coverage through a high-gain antenna array, with a coverage radius of approximately 1.5 kilometers. Wi-Fi is used instead of 5G because of the lack of infrastructure support in uninhabited areas, and Wi-Fi devices are easier to deploy temporarily and have lower power consumption.
[0033] The search and rescue operation involved 15 heterogeneous intelligent agents: 8 quadcopter drones and 7 tracked unmanned vehicles. Each agent was equipped with an inertial measurement unit (IMU), a current technology, for navigation and positioning, including a three-axis accelerometer and a three-axis gyroscope to record the agent's motion. In addition, the drones carried lidar and thermal imaging cameras, while the unmanned vehicles carried millimeter-wave radar and visible-light cameras for environmental perception and target identification. Each agent also had a Wi-Fi communication module supporting high-speed data transmission with signal enhancement points, as well as Wi-Fi Direct functionality, enabling local area network connections with other agents in areas without signal coverage.
[0034] Before the task begins, a reasonable time window T needs to be calculated. window A time window defines the length of time from when an agent completes deviation correction after arriving at a signal enhancement point to when it leaves the signal coverage area. The core purpose of setting a time window is to solve the problem of time difference in agent return to the base station: due to the vast task area and the agents being distributed in different locations, their return times to the signal enhancement point will inevitably differ. Without a time window mechanism, the agents that arrive first will leave immediately after completing correction, while the agents that arrive later may carry important environmental information (such as the location of newly discovered trapped personnel or fallen trees blocking the original path). This information needs to be updated in the collaborative decision-making of all agents, but agents that have already left will not receive the updates.
[0035] The time window is calculated based on the scope of the mission area and the mobility of the agents. Considering the 80 square kilometer search and rescue area is approximately circular with a diameter of about 10 kilometers, and signal enhancement points are generally located at the edges (where terrain is typically flatter than in the center), the furthest distance from an agent to a signal enhancement point is approximately 10 kilometers. The average cruising speed of the drone is 10 meters per second, and the average speed of the unmanned vehicle in complex terrain is 4 meters per second. To ensure that most agents (approximately 85%) can return within the time window, while avoiding a prolonged time window that could decrease search and rescue efficiency, the calculation formula is as follows: Where D 85percentile This indicates that 85% of the agents are approximately 7 kilometers away from the signal booster point; V average The average velocity of all agents is approximately 6 meters per second. T is calculated. window = 7000 / 6 ≈ 1200 seconds (20 minutes).
[0036] This 20-minute time window ensures that most agents have a chance to return and share information without excessively prolonging the mission cycle and affecting search and rescue efficiency. It's important to note that approximately 15% of agents may still fail to return within the time window, which is one of the reasons for introducing the S106 distributed ledger mechanism.
[0037] S102: Data upload and preprocessing.
[0038] The agent continuously collects sensor data during task execution and returns to the signal enhancement point after the cycle time to perform data correction and decision updates. When an agent returns and enters the Wi-Fi coverage area within a 1.5 km radius of the signal enhancement point, it immediately establishes a connection with the base station and begins uploading data. In this example, the uploaded data mainly includes: IMU data sequence: contains acceleration and angular velocity measurements from past mission cycles, with a sampling rate of 50Hz; Environmental perception data: obstacle information detected by lidar or millimeter-wave radar, and suspected targets discovered by image recognition; Current location information: Location estimate obtained by fusing IMU integration and visual odometry; Because the data formats of heterogeneous intelligent agents (such as drones and unmanned vehicles in this example) differ, data preprocessing is a necessary step for subsequent deviation correction. Unlike traditional timestamp alignment, this method avoids the deviation caused by time-domain alignment because there is a time difference (potentially 10-15 minutes) between the intelligent agents' returns to the base station. Direct alignment of time-domain data would introduce additional interpolation errors. Therefore, this invention uses frequency-domain processing for deviation correction, thus avoiding the deviations caused by time-domain alignment. The preprocessing mainly performs the following tasks: Standardized format: Different IMU models may use different data formats and units. They should be uniformly converted to the standard format (acceleration unit: m / s², angular velocity unit: rad / s). Sampling rate standardization: Data with different sampling rates are standardized to 50Hz through interpolation to ensure consistency in subsequent Fast Fourier Transform (FFT) processing; Data integrity check: Identify and mark outliers or missing segments in the data to avoid affecting subsequent processing; The preprocessed data is sent to the deviation correction device, while the original environmental perception data is sent to the decision update device along with the corrected data.
[0039] S103: Agent bias correction.
[0040] Bias correction is a crucial prerequisite for ensuring the accuracy of decision updates. In mountainous forest environments, the accelerometer sensors of intelligent agents are susceptible to biases caused by various factors, particularly changes in wind fields within valleys. If these wind-induced acceleration biases are not corrected, they will accumulate errors during inertial navigation integration, leading to inaccurate position estimations. The accuracy of the position directly impacts the decision update judgment—for example, if the agent's position calculation has a 2-3 meter bias, it may incorrectly determine whether fallen trees are obstructing the planned path, resulting in an erroneous decision update.
[0041] Bias correction consists of two steps: single-agent correction and multi-agent joint correction, which are performed sequentially. Single-agent calibration primarily targets hardware deviations within the sensor itself, such as those caused by the aging of individual components. These deviations typically manifest as low-frequency components because changes in the sensor's physical characteristics are slow processes. The calibration process uses a Fast Fourier Transform (FFT) to convert the time-domain acceleration data to the frequency domain: Where a i (n·Δt) is the acceleration sample value of agent i at discrete time n·Δt, where N is the total number of sampling points, Δt is the sampling interval, f is the frequency, j is the imaginary unit, and A i (f) is the complex spectral value corresponding to frequency f.
[0042] In the frequency domain, the frequency components of a preset ultra-low frequency band are analyzed. In this embodiment, the preset ultra-low frequency band is below 0.01 Hz. If the amplitude of these components exceeds a first threshold and exists in all three axes, it is determined to be a sensor hardware deviation. In this embodiment, the first threshold is 0.05 m / s². These low-frequency deviations are filtered out using a high-pass filter: After filtering, the data is transformed back to the time domain by inverse FFT to obtain the corrected acceleration data.
[0043] Multi-agent joint correction is used to correct biases caused by environmental factors. This example focuses on acceleration bias caused by wind fields in a valley, where the wind field simultaneously affects multiple agents within the area. This shared bias cannot be determined using historical data or calibration values of a single agent; that is, looking at the frequency domain data of a single agent alone makes it difficult to determine whether the bias is caused by environmental interference, such as continuous acceleration due to the agent performing a specific task. The key to multi-agent joint analysis's ability to identify shared bias lies in utilizing the spatial correlation of environmental interference. In valleys or specific terrains, although wind fields are dynamic, they exhibit similarity within a certain spatial range. When multiple agents pass through similar areas within similar time periods, they are affected by similar wind fields, resulting in similar acceleration biases. By statistically analyzing the spectral data of multiple agents, if it is found that multiple agents share components with similar amplitudes in a certain low-frequency band (e.g., the ratio of standard deviation to mean is less than 0.3), it can be determined with high confidence that this is a shared bias caused by environmental factors, rather than an independent hardware problem or task behavior.
[0044] Joint correction is also performed in the frequency domain. For each frequency component, the mean and standard deviation of the amplitudes arriving at the agent are calculated: If, within a preset low-frequency band, μ(f) exceeds a second threshold and σ(f) / μ(f) is less than a preset ratio (indicating that most agents have similar deviations), then it is determined that there is a shared deviation caused by the environment at that frequency. In this embodiment, the preset low-frequency band is 0.01-0.1 Hz, the second threshold is 0.1 m / s², and the preset ratio is 0.3.
[0045] The reason for choosing to use the low-frequency band for identification is that high-frequency acceleration changes are often normal maneuvers of the agent (such as turning or climbing), and it is difficult to distinguish whether they are task execution or environmental interference; while the continuous wind field influence is mainly reflected in the low-frequency band, which is easier to identify.
[0046] An adaptive filter is designed to remove the identified shared bias frequencies, thus completing the joint correction. The corrected acceleration data is then used to update the agent's position estimate.
[0047] The adaptive filter design in multi-agent joint correction is as follows: First, identify the set of shared bias frequencies: Then construct the adaptive filter: Where f s The identified shared bias frequency is α, which is the filter strength parameter with a value of 5, and Δf is the filter bandwidth with a value of 0.02Hz. This Gaussian filter is close to 0 at the shared bias frequency and close to 1 at other frequencies.
[0048] Next, apply the filter: Finally, the corrected acceleration data is obtained by inverse FFT transformation back to the time domain.
[0049] S104: Decision Update.
[0050] The decision update device determines whether the collaborative decision needs to be updated based on the bias-corrected data and the environmental perception information uploaded by the agent. The determination process is described as follows: The decision update device first performs target detection and recognition on the visual and radar data uploaded by the agent: Personnel target identification: Deep learning models (such as the YOLO series of target detection algorithms) are used to analyze thermal imaging and visible light images to identify potential trapped personnel. If a humanoid target is detected with a confidence score exceeding 0.8, the target's location coordinates are extracted.
[0051] Environmental mutation identification: By using the change detection algorithm of LiDAR point cloud data, new obstacles in the environment are identified, such as fallen trees, soil and rocks caused by landslides, etc.
[0052] Based on the identification results, determine whether the decision needs to be updated: Task point update judgment: If a suspected trapped person is identified, their location is first compared with the existing task list. If there is no assigned search and rescue task point near the location (within 50 meters), it is determined that a new task point needs to be added, triggering a task update decision.
[0053] Path update decision: If a sudden environmental change is detected (such as a fallen tree), the minimum distance between the obstacle and the currently planned paths of all agents needs to be calculated based on the corrected and accurate location information. This highlights the necessity of bias correction—if the location estimation is biased, it may incorrectly determine whether a fallen tree with a diameter of 1 meter truly blocks the path. Path update decisions are only triggered when the minimum distance between the obstacle and the path is less than a safety threshold (3 meters for drones, 2 meters for unmanned vehicles).
[0054] Once a decision update is deemed necessary, the decision update device first uses a path planning algorithm to replan the path of the affected agent, avoiding newly discovered obstacles, or adds the newly discovered rescue target to the task list. Then, based on urgency and distance, it assigns the target to the most suitable agent using a task allocation algorithm (corresponding to the processing methods for the two decision scenarios). In this embodiment, the path planning algorithm is A. The algorithm, namely the task allocation algorithm, includes a cost matrix and the Hungarian algorithm. The specific implementation is as follows: When an obstacle is detected blocking the original path, use A. Algorithm (existing path planning algorithm) redesign: Where g(n) is the actual cost from the starting point to node n, h(n) is the heuristic function, and P n P is the position coordinate of node n. goal The target location is indicated by the terrain factor, such as 1.0 for flat ground, 1.5 for slopes, and 2.0 for dense forest.
[0055] During path planning, obstacle locations are set as impassable nodes, and the search space is divided into 2m x 2m grids.
[0056] When a new rescue target is discovered, it is necessary to decide which agent to assign it to. A cost matrix method is used: Construct the cost matrix C: Where C[i,j] represents the comprehensive cost of agent i performing task j, D[i,j] is the distance (meters) from agent i to task j, V[i] is the average speed (m / s) of agent i, S[i,j] is the suitability of agent i for task j (0-1), such as the high suitability of drones for thermal imaging tasks, E[i] is the energy consumption factor of agent i, which is related to the remaining power, and w1, w2, w3 are weighting coefficients.
[0057] The Hungarian algorithm is used to find the optimal allocation (an existing task allocation algorithm) to minimize the total cost.
[0058] The decision update device then generates a decision update package containing the updated content, a timestamp, and a digital signature. This package is in JSON format and is broadcast via Wi-Fi to all agents within signal coverage. Upon receiving the package, each agent verifies the signature and updates its local decision. Simultaneously, the agent records the entire update package, including the original JSON content and signature, in a distributed ledger, ensuring that other agents can verify and adopt the update in areas without signal coverage.
[0059] S105: Speed adjustment and oscillation suppression.
[0060] After the agent completes deviation correction and possible decision updates, it needs to leave the signal coverage area to continue its task (calculation begins from leaving the central base station). At this point, the time window mechanism comes into play.
[0061] Each agent dynamically adjusts its departure speed based on the number of agents that have arrived. The basic principle is: agents that arrive earlier should slow down their departure speed appropriately to allow time for later agents to update; once most agents have completed their updates, the remaining agents should accelerate their departure speed to reduce the overall departure time difference.
[0062] The speed adjustment formula is: Where R coverage The signal coverage radius is 1.5 kilometers, T window It's a 20-minute time window, T change The battery swapping time for the intelligent agent is approximately 5 minutes. process The total processing time for data correction and decision updates is approximately 1 minute, k=1.2 is the adjustment coefficient, and N arrived This represents the number of agents that have arrived and completed processing.
[0063] This speed control mechanism effectively solves the oscillation problem. Oscillation refers to a chain reaction caused by excessively large time differences between agents leaving, leading to frequent updates of decisions. For example, if another agent arrives 10 minutes after the first agent leaves to join a task in a certain area, the arriving agent may carry new collaborative decisions that need to be updated again (via Wi-Fi Direct and distributed ledger in areas without signal). This renders the work done in the first 10 minutes ineffective and further widens the time difference, creating a vicious cycle. By controlling the speed, the departure time difference of all agents is kept within a short period, significantly reducing the probability of oscillation.
[0064] S106: Decision synchronization in areas without signal.
[0065] Although the time window mechanism can solve most decision synchronization problems, about 15% of agents may still be unable to return within the time window due to excessive distance or special circumstances. This may cause some agents that left earlier to still be using old decisions when performing tasks in areas without signal, resulting in incoordination with the actions of other agents.
[0066] Before leaving the signal coverage area, each agent records its latest decision in a local distributed ledger. Each record in the ledger contains: Decision content (path coordinate sequence or new task point information); Timestamps accurate to milliseconds; 2048-bit RSA digital signature; When performing regional tasks in areas without signal, if the distance between two agents is less than a preset synchronization distance, they will establish a local area network connection via Wi-Fi Direct to exchange their respective ledger information. In this embodiment, the preset synchronization distance is set to 300 meters. If an inconsistency is detected, the decision with the latest timestamp and verified signature will be automatically selected as the valid decision.
[0067] This mechanism effectively prevents the spurious majority problem. The distributed ledger application proposed in this invention ensures that even if a majority of agents carry the old decision in a certain local area, the new decision will be adopted and propagated in that area as long as at least one agent carries the new decision and the signature verification is valid. Signature verification ensures the authenticity of the decision. Combined with the time window mechanism, in actual operation, it can almost guarantee that at least one agent in each task area carries the latest decision.
[0068] Secondly, embodiments of this application provide a multi-agent cooperative task system for unmanned areas, including: The signal enhancement point is deployed at the edge of the uninhabited area, with built-in deviation correction and decision update equipment, and multiple heterogeneous intelligent agents. Each intelligent agent is equipped with a sensor module, a communication module, a speed control component, and a distributed ledger storage module.
[0069] The devices in the system work together to execute the deviation correction and decision update method as described in the above embodiments.
[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Moreover, in the description of the embodiments of this application, unless otherwise stated, " / " signifies "or," for example, A / B can mean A or B; "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist, for example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, in the description of the embodiments of this application, "multiple" refers to two or more.
[0071] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
[0072] It should be emphasized that the various technical features of the technical solution in this application complement each other to form an organic whole. Those skilled in the art should realize that the evaluation must adhere to the overall evaluation criteria, strictly follow the requirements of the Patent Law, and disregard any negative or subjective requirements. The inventiveness of this application should be evaluated objectively and realistically, without considering the individual technical features in isolation. Furthermore, the evaluation process should first determine whether the relevant prior art documents have the motivation to introduce common general knowledge or combine them with other materials, thus prudently incorporating common general knowledge and commonly used methods into the evaluation.
Claims
1. A multi-agent cooperative method, characterized in that, The method includes the following steps: Signal enhancement points are deployed in uninhabited areas. These signal enhancement points include a central base station, deviation correction equipment, and decision update equipment. Multiple heterogeneous intelligent agents are deployed to perform search and rescue missions, with each agent equipped with sensors and communication modules; During the execution of the task, the intelligent agent periodically returns to the signal enhancement point. When it enters the signal coverage area, the communication module establishes a connection with the central base station and uploads sensor data. The central base station is used to communicate with the intelligent agent to obtain the sensor data of the intelligent agent and forward it to the deviation correction device and the decision update device; Deviation correction equipment performs deviation correction processing on the sensor data of the intelligent agent; The decision update device determines whether the collaborative decision needs to be updated based on the corrected data. If so, it generates a decision update package and broadcasts it to all agents within the signal coverage area through the central base station. After completing the calibration and update, the agent leaves the signal coverage area to continue its task.
2. The multi-agent cooperative method according to claim 1, characterized in that, The method also includes setting a time window mechanism: Calculate the time window based on the task area and the agent's mobility. : Among them, D 85percentile V represents the distance from 85% of the agents to the signal booster point. average The average velocity of all agents; After completing the deviation correction, the agent stays within the signal coverage area until the time window ends before leaving, to ensure that the agent that arrives later can update the important environmental information it carries with the agent that arrives earlier, and to avoid decision-making asynchrony caused by the return time difference.
3. The multi-agent cooperative method according to claim 2, characterized in that, The time window mechanism also includes a dynamic speed control mechanism, which includes: The central base station counts the number N agents that have arrived and completed processing. arrived And send this statistic to each agent that is about to leave the central base station; Each agent calculates its own departure speed V based on the received data. exit The calculation formula is: The formula for calculating the dynamic weight w is: Among them, R coverage T is the signal coverage radius. window For the time window, T change For the battery swapping time of the intelligent agent, T process N represents the total processing time for data correction and decision updates, k is the adjustment coefficient, and N is the total processing time for data correction and decision updates. arrived N represents the number of agents that have arrived and completed processing. total The total number of intelligent agents; The speed control component calculates the departure speed V of each agent. exit The departure speed of each agent is adjusted to keep the departure time difference of all agents within a preset range, thus preventing frequent decision updates and task execution oscillations caused by excessive departure time differences.
4. The multi-agent cooperative method according to claim 1, characterized in that, The multi-agent cooperative method also includes: Each agent is also equipped with a distributed ledger; Before leaving the signal coverage area, the intelligent agent records the latest decision content, timestamp, and digital signature generated by the decision update device into the distributed ledger; When performing tasks in areas without signal, if the distance between two agents is less than the preset synchronization distance, the communication module establishes a local communication connection and exchanges their respective ledger information. When an inconsistency is detected, the agent compares the timestamps of each decision in the ledger, automatically selects the decision with the latest timestamp and verified signature as the valid decision, and updates its own distributed ledger. This ensures that agents performing the same task in areas without signal coverage maintain decision consistency and prevents a majority of agents carrying outdated decisions from dominating erroneous decisions in local areas.
5. The multi-agent cooperative method according to claim 1, characterized in that, The deviation correction device performs deviation correction processing on the sensor data of the intelligent agent, including the following steps: The agent's sensors include an inertial measurement unit, used to acquire the agent's acceleration data; The acceleration data uploaded by the agent is preprocessed, including format unification and sampling rate standardization, but timestamp alignment is not performed. Use Fast Fourier Transform to convert the acceleration data in the time domain to the frequency domain: Among them, a i (n⋅Δt) is the acceleration sample value of agent i at discrete time n·Δt, where N is the total number of sampling points, Δt is the sampling interval, f is the frequency, j is the imaginary unit, and A i (f) is the complex spectral value corresponding to frequency f; Deviation identification and correction are performed in the frequency domain. After correction, the time domain is transformed back by inverse FFT to avoid time domain alignment interpolation errors caused by the time difference of the agent's return, thus ensuring the accuracy of multi-agent joint correction.
6. The multi-agent cooperative method according to claim 5, characterized in that, The deviation correction method further includes the following steps: Single agent correction: Analyze the frequency components of the preset ultra-low frequency band. If the amplitude exceeds the first threshold and exists in multiple axes, it is determined to be a sensor hardware deviation and is filtered out using a high-pass filter. Multi-agent joint correction: Calculate the mean μ(f) and standard deviation σ(f) of the frequency domain amplitudes of all agents that have reached them: If μ(f) exceeds the second threshold and σ(f) / μ(f) is less than the preset ratio within the preset low frequency band, it is determined that there is a sharing deviation caused by the environment at that frequency. Among them, the high-frequency part corresponds to the normal maneuvering actions of the intelligent agent and is not subject to deviation correction; Identify the set of shared bias frequencies F shared : Constructing an adaptive filter: Among them, f s It is the identified shared bias frequency, α is the filter strength parameter, and Δf is the filter bandwidth; Application filter: Corrected acceleration data is obtained by inverse FFT transformation to eliminate different types of deviations caused by hardware aging and environmental interference.
7. The multi-agent cooperative method according to claim 1, characterized in that, The decision update device updates collaborative decision-making based on the corrected data, including the following steps: The sensors also include environmental sensing devices for acquiring visual and radar data; The decision update device performs target detection and environmental recognition on the visual and radar data uploaded by the intelligent agent; Task point update judgment: If a suspected trapped person is identified and there is no assigned search and rescue task point within the preset range of the location, the task update is triggered, a cost matrix is constructed and the task allocation algorithm is used to solve the optimal allocation scheme; Path update judgment: If a sudden change in environment is detected, the minimum distance between the obstacle and the planned path is calculated based on the corrected position information output by the deviation correction device. If the distance is less than the safety threshold, the path planning algorithm is used to replan the path. The decision update device generates a decision update package containing update content, timestamp, and digital signature, which is broadcast to all agents within the signal coverage area via the central base station to ensure that agents can update the latest decisions in a timely manner.
8. A multi-agent cooperative system, characterized in that, The system includes a processor and a memory, the memory storing at least one instruction, at least one program, a code set, or an instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by the processor to implement the multi-agent cooperative method as described in any one of claims 1-7.