An unmanned inspection vehicle based on multi-agent collaborative decision-making

Abstract

This invention discloses an unmanned inspection vehicle based on multi-agent collaborative decision-making, relating to the field of unmanned aerial vehicle (UAV) inspection. It includes: a data acquisition and management module, a cross-modal fusion module, a situational modeling module, a distributed collaboration module, an analysis and prediction module, a detection optimization module, a knowledge transfer module, a decision allocation module, a joint scheduling module, and a response control module. This invention avoids dependence on a single central node, maintains stable operation even under communication constraints or partial failures, effectively improves perception reliability, reduces inconsistencies caused by multi-source heterogeneous data, avoids the impact of communication congestion on detection results, prevents performance degradation during long-term operation, and enhances the foresight and proactiveness of prevention and control.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) inspection, and specifically to an unmanned inspection vehicle based on multi-agent collaborative decision-making. Background Technology

[0002] With the gradual opening of low-altitude airspace and the rapid development of drone technology, drones have been widely used in logistics, emergency rescue, power, transportation, water conservancy, and urban management. However, the rapid increase in the number of drones has also brought serious challenges to airspace order, security, and privacy protection. Especially in urban core areas, around important infrastructure, and sensitive locations, intrusion incidents by illegal, out-of-control, or malicious drones occur frequently, posing a potential threat to public safety and critical assets. Existing drone detection and control methods mostly rely on fixed monitoring equipment or single mobile platforms, which generally suffer from limited coverage, insufficient environmental adaptability, and weak ability to identify the behavior of new types of drones. At the same time, traditional centralized control architectures face communication bottlenecks and single-point failure risks in large-scale deployment scenarios, making it difficult to meet the requirements for real-time performance, reliability, and scalability in complex dynamic environments.

[0003] Existing unmanned inspection vehicles are highly dependent on a single central node, which makes it impossible to maintain stable operation under conditions of limited communication or partial failure, thus reducing the reliability of perception. The inconsistency caused by multi-source heterogeneous data increases, and communication congestion has a significant impact on the detection effect. Performance degradation still exists in long-term operation, reducing the foresight and initiative of prevention and control. To address this, we propose an unmanned inspection vehicle based on multi-agent collaborative decision-making. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies and provide an unmanned inspection vehicle based on multi-agent collaborative decision-making.

[0005] This invention proposes an unmanned inspection vehicle based on multi-agent collaborative decision-making. The system includes: a data acquisition and management module, a cross-modal fusion module, a situation modeling module, a distributed collaboration module, an analysis and prediction module, a detection optimization module, a knowledge transfer module, a decision allocation module, a joint scheduling module, and a response control module. The data acquisition and management module collects and manages multi-source data from various UAVs in the environment, and dynamically adjusts the working mode and resource allocation of each unmanned inspection vehicle according to environmental complexity and task requirements. The cross-modal fusion module performs time synchronization, spatial alignment, and cross-modal fusion on the collected multi-source heterogeneous data to construct a UAV activity model that includes three-dimensional spatial and temporal dimensions. The situation modeling module is used to create a map of drone activity in a corresponding area for each unmanned inspection vehicle. The distributed collaboration module is used to collaboratively update the drone activity map among various unmanned inspection vehicles across vehicles. The analysis and prediction module analyzes the drone's motion behavior, dwell patterns, and abnormal maneuvers based on real-time drone activity maps, and generates corresponding drone threat levels. The detection optimization module is used to perform adversarial optimization on the drone inspection capabilities of each unmanned inspection vehicle based on historical drone behavior samples and analysis results. The knowledge transfer module is used to share detection experience, model parameters and adversarial knowledge among the unmanned inspection vehicles; The decision allocation module dynamically allocates inspection, tracking, and containment tasks to each unmanned inspection vehicle based on the drone threat level, spatial distribution, inspection vehicle capabilities, and mission status. The joint scheduling module monitors the remaining energy, charging location, and task load of each unmanned inspection vehicle in real time, and dynamically adjusts the inspection range and task intensity. The response control module is used to enable unmanned inspection vehicles to form a defensive array through group collaborative behavior when high-risk drones are detected, and to guide other inspection vehicles to converge. At the same time, the collaborative decision-making results are converted into specific execution control commands.

[0006] As a further aspect of the present invention, the specific steps for the data acquisition and management module to dynamically adjust the working mode and resource allocation of each unmanned inspection vehicle according to environmental complexity and task requirements are as follows: S1.1: Periodically collect the signal-to-noise ratio, real-time error rate, and hardware availability of each sensor source on each unmanned inspection vehicle, and integrate the three sets of information collected into the availability score of the corresponding sensor source through weighted linear scoring. Based on the historical average score, the availability score of each sensor source is exponentially smoothed, and sensors with availability scores below a preset threshold are marked as candidates for decommissioning and will not participate in the next round of sampling. Sensors with availability scores above the preset threshold are recorded in the active sampling pool. S1.2: Read the data from the previous valid time point and the current valid time point of the data streams from different sensors of different unmanned inspection vehicles in the active sampling pool. If the time labels of the two data streams are inconsistent and the difference is less than the preset interpolation window, then the two sets of data streams are aligned to a unified reference timestamp through linear interpolation, the alignment result is recorded, the quality metric of the aligned multi-source sensor data is calculated, and the quality metric is mapped to weights. Then, the generated weights are normalized. S1.3: Calculate the immediate priority of each data stream that needs to be transmitted uplink or between workshops, and select multiple sets of data packets that meet the requirements within the available bandwidth and time window in each scheduling cycle, and cache the remaining data packets locally. When the packet loss rate is detected to be rising, retransmission is triggered and the compression parameters of subsequent cycles are adjusted. Then, the uncertainty of the current area is calculated and the collected uncertainty is mapped to the sampling frequency and sensor activation mode of each unmanned inspection vehicle. S1.4: Decompose the area to be inspected into multiple task units, calculate the detection benefit and expected detection delay of each unit, establish a task allocation vector for the set of available unmanned inspection vehicles, and aim to minimize the weighted expected detection delay of the entire scene. At the same time, constrain the energy, communication and concurrent task limits of each unmanned inspection vehicle, generate multiple allocation schemes using integer programming, and distribute each allocation scheme to the corresponding unmanned inspection vehicle.

[0007] As a further aspect of the present invention, the specific steps for the cross-mode fusion module to construct a UAV activity model containing three-dimensional spatial and temporal dimensions are as follows: S2.1: Map the UAV spatial observation results from different unmanned inspection vehicles and different perspectives to the global three-dimensional coordinate system, and construct an initial spatial state containing position and velocity information for each UAV target observed by the unmanned inspection vehicle, and divide the continuous sampling time into equally spaced basic time slices. S2.2: Between adjacent basic time slices, the state of the next time slice is predicted by the state of the previous time slice to form a spatiotemporal continuous trajectory with a short time scale, so as to obtain the high-speed maneuvering and instantaneous change of direction behavior of the UAV. Then, the state sequences of multiple consecutive basic time slices are aggregated into behavior segments, and consistency constraints are applied to each behavior segment based on the known UAV executable operations. S2.3: Between adjacent behavioral segments, the state of the previous moment is used to predict the state of the next moment, forming a long-term behavioral state. If there is a conflict between the short-term spatiotemporal continuous trajectory and the long-term behavioral state, the short-term spatiotemporal continuous trajectory is smoothed and corrected. S2.4: Calculate the deviation between the spatiotemporal continuous trajectory on a short time scale and the behavioral state on a long time scale. If the deviation exceeds a preset threshold, it is considered that there is a spatiotemporal inconsistency. Then, the UAV state sequence is iteratively corrected until the deviation is lower than the preset threshold. After the correction is completed, a UAV activity model that is continuous, smooth and consistent in three-dimensional space and time dimension is generated.

[0008] As a further aspect of the present invention, the specific steps of the situation modeling module in establishing a corresponding UAV activity map for each unmanned inspection vehicle are as follows: S3.1: The unmanned vehicle seeker receives the observation item streams output by each UAV activity model in real time, corrects the corresponding UAV position according to the local coordinate system of the corresponding unmanned inspection vehicle, unifies the timestamps of each observation item stream based on the local reference time, unifies the feature dimensions of each observation item stream, and generates local UAV activity map node candidates for each observation item stream. S3.2: Associate the candidate nodes at the current moment with the corresponding unmanned inspection vehicle to establish a set of active drone trajectories, then calculate the comprehensive matching cost of each node-trajectory pair, and establish a corresponding matching cost matrix. Based on the matching cost matrix, and using the Hungarian algorithm to obtain one-to-one or one-to-many matching relationships, establish new active drone trajectories for unmatched candidate nodes to generate multiple sets of drone activity maps. S3.3: Each newly observed UAV active trajectory is matched with the predicted UAV active trajectory and updated by weighted fusion. Multiple sets of observation samples are sampled using a sliding window to perform uncertainty estimation on the updated UAV active trajectories. For short sequences of multiple sets of behavioral statistics for each trajectory, cumulative statistical tests are used to detect whether the behavioral statistics deviate significantly from the historical baseline. S3.4: If the deviation value is higher than the preset threshold, the corresponding behavior statistics will be recorded as a behavior change event, and the occurrence time and change type will be marked. If the behavior change event is detected multiple times, the priority of the corresponding UAV active trajectory will be increased, and the corresponding uncertainty estimate will be increased. Then, the uncertainty terms of each UAV active trajectory will be counted, and the corresponding trajectory uncertainty scalar and vectorized covariance approximation will be established. S3.5: If the uncertainty exceeds the preset safety threshold, the corresponding UAV active trajectory is marked as high uncertainty, and the trigger time and dominant uncertainty source are recorded. The active trajectory and its nodes are added to the local UAV activity map structure according to the time window. Then, the activity maps of each UAV are pruned. The summary information of the latest UAV activity trajectory is generated into an index record in real time and written into the corresponding UAV activity map, and the map is updated.

[0009] As a further aspect of the present invention, the observation entry stream in S3.1 includes three-dimensional position, timestamp, modality embedding vector, and observation confidence index.

[0010] As a further aspect of the present invention, the specific steps of the analysis and prediction module in analyzing the motion behavior, dwelling patterns, and abnormal maneuvers of the UAV are as follows: S4.1: For trajectory observations of the same UAV target, form a continuous observation sequence in chronological order, and divide the sequence into short-time sequence and long-time sequence according to the preset time span. For the continuous observations in the short-time sequence, construct a local dynamics model of the UAV, and for the statistical behavior characteristics in the long-time sequence, establish a low-frequency state model of the UAV. S4.2: Input the continuous observations in the short time series into the local dynamics model, calculate the motion state parameters of the UAV in the current time period by the least squares method, and output the short-term motion response state. Then, extract the statistical behavior features in the long time series and input them into the low-frequency state model to distinguish the macroscopic behavior patterns of the UAV, and output the long-term behavior state probability distribution of the UAV in the current time period. S4.3: Jointly model the probability distribution of short-term motion response state and long-term behavior state, and establish a unified multi-scale hidden state. Then, use historical normal behavior to establish a reference distribution, calculate the degree of deviation between the current joint state and historical normal behavior. If the degree of deviation of multiple rounds of detection continues to exceed the threshold, it is marked as an abnormal trend of the drone. S4.4: Based on the current joint state, recursively predict the state evolution after a preset time interval. At the same time, according to the abnormal trend changes of the UAV, output the behavioral evolution direction within the corresponding time window, assign confidence to each prediction result, map each prediction result to an interpretable behavioral category and trend label, calculate the prediction confidence, and output the behavioral trend, confidence and prediction time point of each UAV.

[0011] As a further aspect of the present invention, the specific steps of the detection optimization module to perform adversarial optimization on the unmanned inspection capabilities of each unmanned inspection vehicle are as follows: S5.1: In the unmanned inspection vehicle cluster, based on computing power, current load and historical detection performance, the unmanned inspection vehicle cluster is divided into behavior generation role and detection and discrimination role, and initial parameters are set for each unmanned inspection vehicle in the behavior generation role and the detection and discrimination role respectively. At the same time, parameter synchronization and state initialization are performed before training begins. S5.2: Each behavior generation role constructs a UAV behavior generation model and generates diverse flight trajectories, maneuvering methods and communication characteristics based on the current unmanned inspection vehicle parameters. At the same time, potential new behavior patterns are generated through parameter perturbation and random constraint combination. The output behavior samples are then encoded into a unified data representation form. S5.3: Perform multidimensional perturbation on each generated behavior sample, and then input the perturbated behavior samples into the detection and discrimination role. The detection and discrimination role builds a UAV behavior discrimination model and inputs the received behavior samples into the UAV behavior discrimination model. At the same time, it outputs the corresponding discrimination category and outputs the probability score of the category as UAV target. S5.4: Calculate the model loss value through the adversarial loss function, and based on the obtained loss value, update the drone behavior generation model and the drone behavior discrimination model respectively through the Adam optimizer. Repeat the alternating update until the loss value converges to the preset range. Then, detect the recognition results of the drone by the unmanned inspection vehicle in real time during the training process and evaluate its detection performance. If the detection performance of the unmanned inspection vehicle decreases in adjacent time periods, retrain it alternately.

[0012] As a further aspect of the present invention, the specific steps of the decision allocation module in dynamically allocating the inspection, tracking, and containment tasks of each unmanned inspection vehicle are as follows: S6.1: Divide the area to be inspected into multiple discrete units, summarize the threat characteristics of each discrete unit from local observation and historical records of the unmanned inspection vehicle, normalize each feature data, assign corresponding weights according to business importance, calculate the initial threat index of each discrete unit, and add corresponding timestamps to each threat index. S6.2: Calculate the current available capabilities of each unmanned inspection vehicle, and calculate the vehicle availability score of the corresponding unmanned inspection vehicle based on the acquired current available capabilities. Calculate the reachability penalty for each unit, and update the vehicle availability score of each unmanned inspection vehicle in real time. Calculate the value gain brought by each unmanned inspection vehicle to the corresponding discrete unit by performing the task according to the matching degree between the unit threat index and the vehicle sensing capability. At the same time, calculate the expected cost of performing the task, and then establish a utility value matrix based on the value gain and the expected cost. S6.3: The allocation results of unmanned inspection vehicles and each discrete unit are used as decision variables. With the goal of maximizing overall utility and satisfying multiple sets of constraints, an integer programming or mixed integer programming model is established. Then, each unmanned inspection vehicle forms the competitive value of the corresponding discrete unit based on the local utility value matrix and broadcasts the bidding. S6.4: Each discrete unit records the current highest bid and the corresponding winning unmanned inspection vehicle. If a higher bid is received, the record is updated and the replaced unmanned inspection vehicle is released. After the bidding ends, a locally consistent allocation scheme is formed. If the threat index of any discrete unit or the state change value of an unmanned inspection vehicle is higher than a preset threshold, a reassessment is triggered. During the reassessment, the replacement gain is calculated and adjusted.

[0013] As a further aspect of the present invention, the constraints described in S6.3 include the upper limit of concurrent tasks per vehicle, energy constraints, minimum coverage requirements for key units, and communication bandwidth limitations.

[0014] The beneficial effects of this invention are: This invention periodically collects the signal-to-noise ratio, error rate, and hardware availability of multi-source sensors from various unmanned inspection vehicles, calculates and smooths sensor availability scores, dynamically selects active sensors, performs cross-vehicle data time alignment and quality weighting, and adaptively schedules data transmission. Based on regional uncertainty, it adjusts the sampling frequency and sensor activation mode. Under constraints of energy, communication, and concurrency, it employs integer programming and a distributed bidding mechanism to complete the allocation of multi-vehicle collaborative tasks. It unifies the UAV observations from different inspection vehicles onto a global three-dimensional spatiotemporal coordinate system, constructs corresponding UAV activity models, and generates local UAV activity maps. Finally, it analyzes UAV trajectories... Short-term dynamics modeling and long-term behavior modeling are performed. By using multi-scale fusion to detect abnormal trends and predict the direction of behavior evolution, an adversarial collaborative learning mechanism is introduced. Some inspection vehicles generate complex UAV behavior samples, while others perform discriminative detection. Through alternating training, the detection capability for new types of UAVs is continuously improved, avoiding dependence on a single central node. Stable operation can still be maintained even under communication constraints or partial failures, effectively improving perception reliability, reducing inconsistencies caused by multi-source heterogeneous data, avoiding the impact of communication congestion on detection results, avoiding performance degradation in long-term operation, and improving the foresight and initiative of prevention and control. Attached Figure Description

[0015] The present invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a framework diagram of an unmanned inspection vehicle based on multi-agent collaborative decision-making. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: This embodiment of the invention provides an unmanned inspection vehicle based on multi-agent cooperative decision-making. See also... Figure 1 , Figure 1 This is a framework diagram of an unmanned inspection vehicle based on multi-agent collaborative decision-making, provided for an embodiment of the present invention. The system includes: a data acquisition and management module, a cross-modal fusion module, a situational modeling module, a distributed collaboration module, an analysis and prediction module, a detection optimization module, a knowledge transfer module, a decision allocation module, a joint scheduling module, and a response control module.

[0020] The data acquisition and management module collects and manages multi-source data from various UAVs in the environment, and dynamically adjusts the working mode and resource allocation of each unmanned inspection vehicle according to the complexity of the environment and the requirements of the mission.

[0021] Specifically, the signal-to-noise ratio, real-time error rate, and hardware availability of each sensor source on each unmanned inspection vehicle are periodically collected. A weighted linear scoring method is used to integrate these three sets of information into an availability score for the corresponding sensor source. Based on the historical average score, the availability score of each sensor source is exponentially smoothed. Sensors with availability scores below a preset threshold are marked as candidates for decommissioning and will not participate in the next round of sampling. Sensors with availability scores above the preset threshold are recorded in the active sampling pool. Data from different sensors on different unmanned inspection vehicles at the previous and current valid time points are read from the active sampling pool. If the time stamps of the two data streams are inconsistent and the difference is less than a preset interpolation window, linear interpolation is used to align the two data streams to a unified reference timestamp. The alignment result is recorded, and the quality metric of the aligned multi-source sensor data is calculated. This quality metric is then mapped to weights and further processed. The generated weights are normalized, and the real-time priority of each data stream that needs to be transmitted uplink or between stations is calculated. Within each scheduling cycle, multiple sets of data packets that meet the requirements are selected and sent within the available bandwidth and time window, while the remaining data packets are locally cached. When an increase in packet loss rate is detected, retransmission is triggered and the compression parameters of subsequent cycles are adjusted. Then, the uncertainty of the current area is calculated, and the collected uncertainty is mapped to the sampling frequency and sensor activation mode of each unmanned inspection vehicle. The area to be inspected is decomposed into multiple task units, and the detection benefit and expected detection latency of each unit are calculated. A task allocation vector is established for the set of available unmanned inspection vehicles, with the goal of minimizing the weighted expected detection latency of the entire scenario. At the same time, the energy, communication and concurrent task limits of each unmanned inspection vehicle are constrained. Integer programming is used to generate multiple allocation schemes, and each allocation scheme is distributed to the corresponding unmanned inspection vehicle.

[0022] The cross-modal fusion module performs time synchronization, spatial alignment, and cross-modal fusion on the collected multi-source heterogeneous data to construct a UAV activity model that includes three-dimensional spatial and temporal dimensions.

[0023] Specifically, the spatial observation results of UAVs from different unmanned inspection vehicles and different perspectives are uniformly mapped to a global three-dimensional coordinate system. For each UAV target observed by the unmanned inspection vehicle, an initial spatial state containing position and velocity information is constructed. The continuous sampling time is divided into equally spaced basic time slices. Between adjacent basic time slices, the state of the next moment is predicted using the state of the previous moment, forming a short-timescale spatiotemporal continuous trajectory to obtain the high-speed maneuvering and instantaneous change-of-direction behavior of the UAV. Then, the state sequences of multiple sets of continuous basic time slices are aggregated into behavior segments, and based on the known executable operations of the UAV, each behavior segment is subjected to... Consistency constraints are applied to predict the state of the next moment between adjacent action segments using the state of the previous moment, forming a long-term action state. If there is a conflict between the spatiotemporal continuous trajectory of the short time scale and the action state of the long time scale, the spatiotemporal continuous trajectory of the short time scale is smoothed and corrected. The deviation between the spatiotemporal continuous trajectory of the short time scale and the action state of the long time scale is calculated. If the deviation exceeds a preset threshold, it is considered that there is a spatiotemporal inconsistency. The UAV state sequence is then iteratively corrected until the deviation is lower than the preset threshold. After the correction is completed, a UAV activity model that is continuous, smooth and consistent in the three-dimensional space and time dimensions is generated.

[0024] The situation modeling module is used to create a map of drone activity in a corresponding area for each unmanned inspection vehicle.

[0025] Specifically, the unmanned vehicle detection system receives observation streams from each UAV activity model in real time, corrects the corresponding UAV position according to the local coordinate system of the UAV inspection vehicle, unifies the timestamps of each observation stream based on the local reference time, and unifies the feature dimensions of each observation stream. It generates local UAV activity map node candidates for each observation stream, associates the current node candidates with the corresponding UAV active trajectory set established by the UAV inspection vehicle, calculates the comprehensive matching cost for each node-trajectory pair, and establishes a corresponding matching cost matrix. Based on the matching cost matrix, and using the Hungarian algorithm to obtain one-to-one or one-to-many matching relationships, it establishes new UAV active trajectories for unmatched candidate nodes to generate multiple sets of UAV activity maps. Each newly observed UAV active trajectory matched with the predicted UAV active trajectory is weighted and fused to update the UAV active trajectories. A sliding window is used to sample multiple sets of observation samples to update the UAV active trajectories. Uncertainty estimation is performed by using a cumulative statistical test to check whether the short sequences of multiple sets of behavioral statistics for each trajectory significantly deviate from the historical baseline. If the deviation exceeds a preset threshold, the corresponding behavioral statistics are recorded as behavioral change events, and the occurrence time and change type are marked. If behavioral change events occur multiple times, the priority of the corresponding UAV active trajectory is increased, and the corresponding uncertainty estimate is improved. Then, the uncertainty terms of each UAV active trajectory are statistically analyzed, and an approximation of the corresponding trajectory uncertainty scalar and vectorized covariance is established. If the uncertainty exceeds a preset safety threshold, the corresponding UAV active trajectory is marked as high uncertainty, and the trigger time and dominant uncertainty source are recorded. The active trajectories and their nodes are added to the local UAV activity map structure according to the time window. Then, the activity maps of each UAV are pruned, and the summary information of the latest UAV activity trajectory is generated into an index record in real time and written into the corresponding UAV activity map, and the map is updated.

[0026] The distributed collaboration module is used to collaboratively update the drone activity map among various unmanned inspection vehicles across vehicles.

[0027] Example 2: This embodiment of the invention provides an unmanned inspection vehicle based on multi-agent collaborative decision-making. See also... Figure 1 , Figure 1 This is a framework diagram of an unmanned inspection vehicle based on multi-agent collaborative decision-making, provided for an embodiment of the present invention. The system includes: a data acquisition and management module, a cross-modal fusion module, a situational modeling module, a distributed collaboration module, an analysis and prediction module, a detection optimization module, a knowledge transfer module, a decision allocation module, a joint scheduling module, and a response control module.

[0028] The analysis and prediction module analyzes the drone's motion behavior, dwell patterns, and abnormal maneuvers based on real-time drone activity maps, and generates corresponding drone threat levels.

[0029] Specifically, for trajectory observations belonging to the same UAV target, a continuous observation sequence is formed in chronological order. This sequence is then divided into short-term and long-term sequences based on a preset time span. For continuous observations in the short-term sequence, a local dynamics model of the UAV is constructed. For statistical behavioral characteristics in the long-term sequence, a low-frequency state model of the UAV is established. Continuous observations from the short-term sequence are input into the local dynamics model, and the motion state parameters of the UAV within the current time period are calculated using a least-squares method. The short-term motion response state is then output. Subsequently, statistical behavioral characteristics from the long-term sequence are extracted and input into the low-frequency state model to distinguish the macroscopic behavioral patterns of the UAV. Simultaneously, the long-term behavioral state profile of the UAV within the current time period is output. The probability distribution model jointly models the probability distributions of short-term motion response states and long-term behavioral states, and establishes a unified multi-scale hidden state. Then, a reference distribution is established using historical normal behavior. The deviation between the current joint state and historical normal behavior is calculated. If the deviation exceeds the threshold in multiple rounds of detection, it is marked as an abnormal trend of the UAV. Based on the current joint state, the state evolution after a preset time interval is recursively predicted. At the same time, according to the abnormal trend changes of the UAV, the behavioral evolution direction within the corresponding time window is output, and a confidence level is assigned to each prediction result. Then, each prediction result is mapped to an interpretable behavioral category and trend label. The prediction confidence is calculated, and the behavioral trend, confidence level, and prediction time point of each UAV are output.

[0030] The detection optimization module is used to perform adversarial optimization on the drone inspection capabilities of each unmanned inspection vehicle based on historical drone behavior samples and analysis results.

[0031] Specifically, within the unmanned inspection vehicle swarm, based on computing power, current load, and historical detection performance, the swarm is divided into behavior generation roles and detection / discrimination roles. Initial parameters are set for each unmanned inspection vehicle in both roles, and parameter synchronization and state initialization are performed before training begins. Each behavior generation role internally constructs a UAV behavior generation model and generates diverse flight trajectories, maneuvering methods, and communication characteristics based on the current unmanned inspection vehicle parameters. Furthermore, potential novel behavior patterns are generated through parameter perturbation and random constraint combinations. The output behavior samples are then encoded into a unified data representation format. Multidimensional perturbations are applied to each generated behavior sample, and the perturbed behavior samples are then... The detection and discrimination role is used to input samples. Internally, a drone behavior discrimination model is built within this role. Each received behavior sample is input into the drone behavior discrimination model, which outputs the corresponding discrimination category and a probability score for the category being a drone target. The model loss value is calculated using an adversarial loss function. Based on the obtained loss value, the Adam optimizer is used to update both the drone behavior generation model and the drone behavior discrimination model. This alternating update process is repeated until the loss value converges to a preset range. During real-time detection training, the unmanned inspection vehicle's recognition results for drones are observed, and its detection performance is evaluated. If the unmanned inspection vehicle's detection performance declines within adjacent time periods, it is retrained using an alternating method.

[0032] The knowledge transfer module is used to share detection experience, model parameters and adversarial knowledge among the unmanned inspection vehicles; the decision allocation module dynamically allocates inspection, tracking and containment tasks to each unmanned inspection vehicle based on the drone threat level, spatial distribution, inspection vehicle capabilities and mission status.

[0033] Specifically, the area to be inspected is divided into multiple discrete units. Threat characteristics from local observations and historical records of unmanned inspection vehicles are aggregated for each discrete unit. These characteristic data are normalized and weighted according to their importance. An initial threat index is calculated for each discrete unit, and a timestamp is added to each index. The current available capabilities of each unmanned inspection vehicle are calculated, and a vehicle availability score is calculated based on these capabilities. A reachability penalty is calculated for each unit, and the vehicle availability score is updated in real time. Based on the matching degree between the unit threat index and the vehicle's sensing capabilities, the value gain brought by each unmanned inspection vehicle performing a task in the corresponding discrete unit is calculated, along with the expected return on investment for performing the task. The cost is then calculated, and a utility value matrix is ​​established based on the value gain and expected cost. The allocation results of the unmanned inspection vehicle and each discrete unit are used as decision variables. With the goal of maximizing overall utility and satisfying multiple sets of constraints, an integer programming or mixed integer programming model is established. Then, each unmanned inspection vehicle forms the bidding value of the corresponding discrete unit based on the local utility value matrix and broadcasts the bidding. Each discrete unit records the current highest bid and the corresponding winning unmanned inspection vehicle. If a higher bid is received, the system is updated and the resources of the replaced unmanned inspection vehicle are released. After the bidding ends, a locally consistent allocation scheme is formed. If the threat index of any discrete unit or the state change value of any unmanned inspection vehicle is higher than a preset threshold, a reassessment is triggered. During the reassessment, the replacement gain is calculated and adjusted.

[0034] The joint scheduling module monitors the remaining energy, charging location, and task load of each unmanned inspection vehicle in real time, and dynamically adjusts the inspection range and task intensity. The response control module is used to form a defense array through group collaborative behavior when high-risk drones are detected, and guide other inspection vehicles to converge. At the same time, the collaborative decision-making results are converted into specific execution control commands.

[0035] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An unmanned inspection vehicle based on multi-agent collaborative decision-making, characterized in that, include: The module includes a data acquisition and management module, a cross-modal fusion module, a situational modeling module, a distributed collaboration module, an analysis and prediction module, a detection and optimization module, a knowledge transfer module, a decision allocation module, a joint scheduling module, and a response control module. The data acquisition and management module collects and manages multi-source data from various UAVs in the environment, and dynamically adjusts the working mode and resource allocation of each unmanned inspection vehicle according to environmental complexity and task requirements. The cross-modal fusion module performs time synchronization, spatial alignment, and cross-modal fusion on the collected multi-source heterogeneous data to construct a UAV activity model that includes three-dimensional spatial and temporal dimensions. The situation modeling module is used to create a map of drone activity in a corresponding area for each unmanned inspection vehicle. The distributed collaboration module is used to collaboratively update the drone activity map among various unmanned inspection vehicles across vehicles. The analysis and prediction module analyzes the drone's motion behavior, dwell patterns, and abnormal maneuvers based on real-time drone activity maps, and generates corresponding drone threat levels. The detection optimization module is used to perform adversarial optimization on the drone inspection capabilities of each unmanned inspection vehicle based on historical drone behavior samples and analysis results. The knowledge transfer module is used to share detection experience, model parameters and adversarial knowledge among the unmanned inspection vehicles; The decision allocation module dynamically allocates inspection, tracking, and containment tasks to each unmanned inspection vehicle based on the drone threat level, spatial distribution, inspection vehicle capabilities, and mission status. The joint scheduling module monitors the remaining energy, charging location, and task load of each unmanned inspection vehicle in real time, and dynamically adjusts the inspection range and task intensity. The response control module is used to enable unmanned inspection vehicles to form a defensive array through group collaborative behavior when high-risk drones are detected, and to guide other inspection vehicles to converge. At the same time, the collaborative decision-making results are converted into specific execution control commands.

2. The unmanned inspection vehicle based on multi-agent collaborative decision-making according to claim 1, characterized in that, The specific steps by which the data acquisition and management module dynamically adjusts the working mode and resource allocation of each unmanned inspection vehicle based on environmental complexity and task requirements are as follows: S1.1: Periodically collect the signal-to-noise ratio, real-time error rate, and hardware availability of each sensor source on each unmanned inspection vehicle, and integrate the three sets of information collected into the availability score of the corresponding sensor source through weighted linear scoring. Based on the historical average score, the availability score of each sensor source is exponentially smoothed, and sensors with availability scores below a preset threshold are marked as candidates for decommissioning and will not participate in the next round of sampling. Sensors with availability scores above the preset threshold are recorded in the active sampling pool. S1.2: Read the data from the previous valid time point and the current valid time point of the data streams from different sensors of different unmanned inspection vehicles in the active sampling pool. If the time labels of the two data streams are inconsistent and the difference is less than the preset interpolation window, then the two sets of data streams are aligned to a unified reference timestamp through linear interpolation, the alignment result is recorded, the quality metric of the aligned multi-source sensor data is calculated, and the quality metric is mapped to weights. Then, the generated weights are normalized. S1.3: Calculate the immediate priority of each data stream that needs to be transmitted uplink or between workshops, and select multiple sets of data packets that meet the requirements within the available bandwidth and time window in each scheduling cycle, and cache the remaining data packets locally. When the packet loss rate is detected to be rising, retransmission is triggered and the compression parameters of subsequent cycles are adjusted. Then, the uncertainty of the current area is calculated and the collected uncertainty is mapped to the sampling frequency and sensor activation mode of each unmanned inspection vehicle. S1.4: Decompose the area to be inspected into multiple task units, calculate the detection benefit and expected detection delay of each unit, establish a task allocation vector for the set of available unmanned inspection vehicles, and aim to minimize the weighted expected detection delay of the entire scene. At the same time, constrain the energy, communication and concurrent task limits of each unmanned inspection vehicle, generate multiple allocation schemes using integer programming, and distribute each allocation scheme to the corresponding unmanned inspection vehicle.

3. The unmanned inspection vehicle based on multi-agent collaborative decision-making according to claim 2, characterized in that, The specific steps for the cross-mode fusion module to construct a UAV activity model that includes three-dimensional spatial and temporal dimensions are as follows: S2.1: Map the UAV spatial observation results from different unmanned inspection vehicles and different perspectives to the global three-dimensional coordinate system, and construct an initial spatial state containing position and velocity information for each UAV target observed by the unmanned inspection vehicle, and divide the continuous sampling time into equally spaced basic time slices. S2.2: Between adjacent basic time slices, the state of the next time slice is predicted by the state of the previous time slice to form a spatiotemporal continuous trajectory with a short time scale, so as to obtain the high-speed maneuvering and instantaneous change of direction behavior of the UAV. Then, the state sequences of multiple consecutive basic time slices are aggregated into behavior segments, and consistency constraints are applied to each behavior segment based on the known UAV executable operations. S2.3: Between adjacent behavioral segments, the state of the previous moment is used to predict the state of the next moment, forming a long-term behavioral state. If there is a conflict between the short-term spatiotemporal continuous trajectory and the long-term behavioral state, the short-term spatiotemporal continuous trajectory is smoothed and corrected. S2.4: Calculate the deviation between the spatiotemporal continuous trajectory on a short time scale and the behavioral state on a long time scale. If the deviation exceeds a preset threshold, it is considered that there is a spatiotemporal inconsistency. Then, the UAV state sequence is iteratively corrected until the deviation is lower than the preset threshold. After the correction is completed, a UAV activity model that is continuous, smooth and consistent in three-dimensional space and time dimension is generated.

4. The unmanned inspection vehicle based on multi-agent collaborative decision-making according to claim 3, characterized in that, The specific steps by which the situation modeling module establishes a corresponding UAV activity map for each unmanned inspection vehicle in the relevant area are as follows: S3.1: The unmanned vehicle seeker receives the observation item streams output by each UAV activity model in real time, corrects the corresponding UAV position according to the local coordinate system of the corresponding unmanned inspection vehicle, unifies the timestamps of each observation item stream based on the local reference time, unifies the feature dimensions of each observation item stream, and generates local UAV activity map node candidates for each observation item stream. S3.2: Associate the candidate nodes at the current moment with the corresponding unmanned inspection vehicle to establish a set of active drone trajectories, then calculate the comprehensive matching cost of each node-trajectory pair, and establish a corresponding matching cost matrix. Based on the matching cost matrix, and using the Hungarian algorithm to obtain one-to-one or one-to-many matching relationships, establish new active drone trajectories for unmatched candidate nodes to generate multiple sets of drone activity maps. S3.3: Each newly observed UAV active trajectory is matched with the predicted UAV active trajectory and updated by weighted fusion. Multiple sets of observation samples are sampled using a sliding window to perform uncertainty estimation on the updated UAV active trajectories. For short sequences of multiple sets of behavioral statistics for each trajectory, cumulative statistical tests are used to detect whether the behavioral statistics deviate significantly from the historical baseline. S3.4: If the deviation value is higher than the preset threshold, the corresponding behavior statistics will be recorded as a behavior change event, and the occurrence time and change type will be marked. If the behavior change event is detected multiple times, the priority of the corresponding UAV active trajectory will be increased, and the corresponding uncertainty estimate will be increased. Then, the uncertainty terms of each UAV active trajectory will be counted, and the corresponding trajectory uncertainty scalar and vectorized covariance approximation will be established. S3.5: If the uncertainty exceeds the preset safety threshold, the corresponding UAV active trajectory is marked as high uncertainty, and the trigger time and dominant uncertainty source are recorded. The active trajectory and its nodes are added to the local UAV activity map structure according to the time window. Then, the activity maps of each UAV are pruned. The summary information of the latest UAV activity trajectory is generated into an index record in real time and written into the corresponding UAV activity map, and the map is updated.

5. The unmanned inspection vehicle based on multi-agent collaborative decision-making according to claim 1, characterized in that, The specific steps of the analysis and prediction module in analyzing the drone's motion behavior, hovering patterns, and abnormal maneuvers are as follows: S4.1: For trajectory observations of the same UAV target, form a continuous observation sequence in chronological order, and divide the sequence into short-time sequence and long-time sequence according to the preset time span. For the continuous observations in the short-time sequence, construct a local dynamics model of the UAV, and for the statistical behavior characteristics in the long-time sequence, establish a low-frequency state model of the UAV. S4.2: Input the continuous observations in the short time series into the local dynamics model, calculate the motion state parameters of the UAV in the current time period by the least squares method, and output the short-term motion response state. Then, extract the statistical behavior features in the long time series and input them into the low-frequency state model to distinguish the macroscopic behavior patterns of the UAV, and output the long-term behavior state probability distribution of the UAV in the current time period. S4.3: Jointly model the probability distribution of short-term motion response state and long-term behavior state, and establish a unified multi-scale hidden state. Then, use historical normal behavior to establish a reference distribution, calculate the degree of deviation between the current joint state and historical normal behavior. If the degree of deviation of multiple rounds of detection continues to exceed the threshold, it is marked as an abnormal trend of the drone. S4.4: Based on the current joint state, recursively predict the state evolution after a preset time interval. At the same time, according to the abnormal trend changes of the UAV, output the behavioral evolution direction within the corresponding time window, assign confidence to each prediction result, map each prediction result to an interpretable behavioral category and trend label, calculate the prediction confidence, and output the behavioral trend, confidence and prediction time point of each UAV.

6. The unmanned inspection vehicle based on multi-agent cooperative decision-making according to claim 1, characterized in that, The specific steps of the detection optimization module to perform adversarial optimization on the drone inspection capabilities of each unmanned inspection vehicle are as follows: S5.1: In the unmanned inspection vehicle cluster, based on computing power, current load and historical detection performance, the unmanned inspection vehicle cluster is divided into behavior generation role and detection and discrimination role, and initial parameters are set for each unmanned inspection vehicle in the behavior generation role and the detection and discrimination role respectively. At the same time, parameter synchronization and state initialization are performed before training begins. S5.2: Each behavior generation role constructs a UAV behavior generation model and generates diverse flight trajectories, maneuvering methods and communication characteristics based on the current unmanned inspection vehicle parameters. At the same time, potential new behavior patterns are generated through parameter perturbation and random constraint combination. The output behavior samples are then encoded into a unified data representation form. S5.3: Perform multidimensional perturbation on each generated behavior sample, and then input the perturbated behavior samples into the detection and discrimination role. The detection and discrimination role builds a UAV behavior discrimination model and inputs the received behavior samples into the UAV behavior discrimination model. At the same time, it outputs the corresponding discrimination category and outputs the probability score of the category as UAV target. S5.4: Calculate the model loss value through the adversarial loss function, and based on the obtained loss value, update the drone behavior generation model and the drone behavior discrimination model respectively through the Adam optimizer. Repeat the alternating update until the loss value converges to the preset range. Then, detect the recognition results of the drone by the unmanned inspection vehicle in real time during the training process and evaluate its detection performance. If the detection performance of the unmanned inspection vehicle decreases in adjacent time periods, retrain it alternately.

7. The unmanned inspection vehicle based on multi-agent cooperative decision-making according to claim 1, characterized in that, The specific steps of the decision allocation module in dynamically allocating inspection, tracking, and containment tasks for each unmanned inspection vehicle are as follows: S6.1: Divide the area to be inspected into multiple discrete units, summarize the threat characteristics of each discrete unit from local observation and historical records of the unmanned inspection vehicle, normalize each feature data, assign corresponding weights according to business importance, calculate the initial threat index of each discrete unit, and add corresponding timestamps to each threat index. S6.2: Calculate the current available capabilities of each unmanned inspection vehicle, and calculate the vehicle availability score of the corresponding unmanned inspection vehicle based on the acquired current available capabilities. Calculate the reachability penalty for each unit, and update the vehicle availability score of each unmanned inspection vehicle in real time. Calculate the value gain brought by each unmanned inspection vehicle to the corresponding discrete unit by performing the task according to the matching degree between the unit threat index and the vehicle sensing capability. At the same time, calculate the expected cost of performing the task, and then establish a utility value matrix based on the value gain and the expected cost. S6.3: The allocation results of unmanned inspection vehicles and each discrete unit are used as decision variables. With the goal of maximizing overall utility and satisfying multiple sets of constraints, an integer programming or mixed integer programming model is established. Then, each unmanned inspection vehicle forms the competitive value of the corresponding discrete unit based on the local utility value matrix and broadcasts the bidding. S6.4: Each discrete unit records the current highest bid and the corresponding winning unmanned inspection vehicle. If a higher bid is received, the record is updated and the replaced unmanned inspection vehicle is released. After the bidding ends, a locally consistent allocation scheme is formed. If the threat index of any discrete unit or the state change value of an unmanned inspection vehicle is higher than a preset threshold, a reassessment is triggered. During the reassessment, the replacement gain is calculated and adjusted.

Images