A dangerous article high-sensitivity detection method and system based on reinforcement learning

Abstract

The application discloses a dangerous article high-sensitivity detection method based on reinforcement learning, which comprises the following steps: forming a detection evidence set and a current detection state; constructing an enhanced detection action set and an action condition Extremile order surface; forming a historical verification calibration parameter set; calibrating the action condition Extremile order surface and an Extremile regression process, and constructing a reinforcement learning process to obtain a double-track extreme cumulative detection loss; forming an extreme risk closure quantity and a channel residual extreme risk, updating a reinforcement learning detection strategy, and forming a target enhanced detection action; updating the current detection state, the action condition Extremile order surface and the channel residual extreme risk to form a final channel residual extreme risk; determining a result state to form a dangerous article high-sensitivity detection result. The application improves the sensitivity, reliability and detection resource utilization efficiency of dangerous article detection.

Description

Technical Field

[0001] This invention relates to the field of hazardous materials detection technology, and in particular to a highly sensitive method and system for detecting hazardous materials based on reinforcement learning. Background Technology

[0002] Hazardous materials detection is widely used in transportation hubs, logistics inspections, public place security checks, and security protection of key areas. Existing detection equipment typically acquires response data through X-rays, millimeter waves, spectroscopy, magnetic induction, and other detection channels, and then identifies hazardous materials based on fixed thresholds, feature matching models, or classification models. Some solutions improve identification capabilities by fusing response results from different detection channels and adjusting scanning intensity, scanning position, and detection parameters based on the detection results.

[0003] Existing technologies lack effective characterization of the synergistic, attenuation, and redundancy relationships among different detection evidences. They tend to linearly fuse detection evidences according to independent relationships, making it difficult to reflect the combined effects of obstruction, packaging, material overlap, and channel conflict on hazard response.

[0004] Verification data may also result in censored detection trajectories due to insufficient verification time, interruptions in the verification process, or failure to return verification results. Existing training methods directly remove censored detection trajectories or treat them as ordinary samples, which can easily lead to biases in the representativeness of historical samples and inaccuracies in estimating extreme missed detection risks. Furthermore, existing reinforcement learning detection schemes typically select detection actions based on single recognition confidence or immediate gains, lacking a comparison of the extreme cumulative detection losses of the two paths—executing enhanced detection actions and maintaining current detection parameters—and a lack of continuous assessment of the residual extreme risks in the execution action trajectory. This makes it difficult for the detection strategy to achieve closed-loop optimization among risk closure capability, residual missed detection risk, and detection resource consumption, affecting the sensitivity, reliability, and efficiency of hazardous material detection.

[0005] Therefore, how to provide a highly sensitive detection method and system for dangerous items based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a highly sensitive method and system for detecting hazardous materials based on reinforcement learning. This invention improves the ability to identify and suppress extreme missed detection risks of hazardous materials under conditions of occlusion, attenuation, interference, and channel conflict by using Extremile order surface, verification censoring calibration, dual-track extreme cumulative detection loss, and reinforcement learning closed-loop update, thereby improving detection sensitivity, result reliability, and detection resource utilization efficiency.

[0007] A highly sensitive method for detecting hazardous materials based on reinforcement learning according to an embodiment of the present invention includes the following steps: Acquire multi-channel detection response data and detection equipment operating status to form a set of detection evidence and the current detection status; Construct an enhanced set of detection actions and perform a binary additive Choquet integral to form an action-conditional extremile order surface; Identify censored detection trajectories in historical verification trajectories, perform multiplicative active limit censoring correction to obtain verification availability weights, and weight historical verification trajectories with verification availability weights to form a historical verification calibration parameter set; The Extremile order surface and Extremile regression process of the action condition are calibrated using the historical verification calibration parameter set, and an Extremile regression-driven reinforcement learning process is constructed to obtain the dual-track extreme cumulative detection loss. The extreme risk closure quantity is formed based on the extreme detection loss of the dual-track system, and the residual extreme risk of the channel is formed based on the execution action trajectory. The extreme risk closure quantity and the residual extreme risk of the channel are used to update the reinforcement learning detection strategy and form the target-enhanced detection action. Execute target enhancement detection actions and update the current detection status, action conditions, Extremile order surface, and channel residual extreme risk. If the channel residual extreme risk is not closed and the resource boundary is not reached, redetermine the target enhancement detection actions to form the final channel residual extreme risk. The final result status is determined based on the residual extreme risks in the final channel, forming a highly sensitive detection result for hazardous materials.

[0008] Optionally, obtaining the set of detection evidence and the current detection state specifically includes: According to a unified detection cycle, multi-channel detection response data of hazardous materials and the operating status of detection equipment are obtained. Response preprocessing is performed on the multi-channel detection response data to form preprocessed detection response data. Detection evidence extraction is performed on the preprocessed detection response data to form a detection evidence set; The current detection state is formed by performing state construction processing on the collection of detection evidence and the operating status of the detection equipment.

[0009] Optionally, obtaining the action condition extremile order surface specifically includes: Based on the current detection status, the feasibility of the detection control methods that can be executed by the detection equipment is screened, and the detection control methods that meet the equipment operating conditions, detection resource boundaries and hazardous material detection conditions are combined into an enhanced detection action set; For each piece of detected evidence in the detection evidence set and each enhanced detection action in the enhanced detection action set, action evidence association processing is performed. The degree of response and disclosure of the enhanced detection action to the corresponding detected evidence is determined according to the current detection state. The degree of response and disclosure is combined with the evidence strength of the corresponding detected evidence to form the action-related detection evidence quantity. All action-related detection evidence quantities corresponding to the same detection area and the same enhanced detection action are combined according to a fixed evidence order to form the action-related detection evidence vector. Based on the degree of response revelation, the residual evidence vector associated with each action is reconstructed to form the evidence vector before action execution and the residual evidence vector after action execution. The additive fuzzy measure is reconstructed according to the joint revelation relationship between the enhanced detection action and the individual detection evidence and the detection evidence pair. The additive fuzzy measure is then used to perform additive Choquet integration on the evidence vector before action execution and the residual evidence vector after action execution, respectively, to obtain the baseline interaction integral value and the residual interaction integral value. The difference is then calculated to form the action evidence interaction reduction amount. Based on the action evidence interaction reduction and residual interaction integral value corresponding to each enhanced detection action, closed residual competition sorting is performed on enhanced detection actions within the same detection area. The action condition Extremile order is determined according to the closed residual competition sorting result. The action condition Extremile order with consistent evidence interaction structure and the same action identifier in adjacent detection areas is performed to perform order continuity. Order transitions are preserved between detection areas where the evidence interaction relationship is reversed, forming the action condition Extremile order surface.

[0010] Optionally, obtaining the historical verification calibration parameter set specifically includes: Obtain historical verification trajectories, perform trajectory verification status marking and verification availability layering on the historical verification trajectories, and form complete detection trajectories, deleted detection trajectories, and verification availability layers; For each level of verification availability, multiplicative active censoring correction is performed to form a verification availability function; The verification availability function is used to perform weight generation processing to form the verification availability weight; The historical verification trajectory is subjected to historical verification weighted calibration processing based on the verification availability weight, forming a historical verification calibration parameter set.

[0011] Optionally, obtaining the extreme cumulative detection loss of the dual-track system specifically includes: The order-preserving calibration of the action condition Extremile order surface is performed using the historical verification calibration parameter set to form the calibration action condition Extremile order surface; Enhanced detection action execution segments and detection parameter holding segments are extracted from weighted historical trajectory samples to form execution trajectory training samples and holding trajectory training samples, respectively. Based on the immediate detection loss and the discounted extreme cumulative detection loss corresponding to the next detection state, execution trajectory cumulative loss targets and holding trajectory cumulative loss targets are formed. The execution track training samples, hold track training samples, and calibration action condition extremile order are input into the extremile regression process to determine the cumulative position of the execution track cumulative loss target and the hold track cumulative loss target in the historical verification loss weighted distribution. Based on the corresponding calibration action condition extremile order, the cumulative position is converted into execution track extremile weights and hold track extremile weights. The deviation between the cumulative loss target and the extremile action value function prediction result is weighted together using the process verification availability weight and the corresponding extremile weight. The extremile action value function parameters are then updated based on the weighted deviation to form the calibration extremile regression process. The current detection state, the set of enhanced detection actions, and the extremile order surface of the calibration action conditions are input into the calibration extremile regression process to obtain the extreme cumulative detection loss of the execution track and the extreme cumulative detection loss of the hold track for each enhanced detection action. Paired relationships are established according to the same detection region and enhanced detection action to form a dual-track extreme cumulative detection loss. The calibration extremile regression process is used as the action value estimation process of the reinforcement learning detection strategy. The current detection state is used as the reinforcement learning state, the set of enhanced detection actions is used as the reinforcement learning action space, and the dual-track extreme cumulative detection loss is used as the extreme action value basis of the enhanced detection actions to construct an extremile regression-driven reinforcement learning process.

[0012] Optionally, the target enhancement detection action specifically includes: The extreme cumulative detection loss of the dual-track system is processed to generate the extreme risk closure quantity corresponding to each enhanced detection action; The trajectory extension process is performed on the enhanced detection actions and the set of enhanced detection actions to form the execution action trajectory corresponding to each enhanced detection action, and the trajectory residual coefficient is formed based on the degree of response disclosure; Based on the execution action trajectory and trajectory residual coefficient, the residual risk of the execution channel is aggregated to form the extreme residual risk of the channel corresponding to each enhanced detection action; The extreme action value is updated based on the extreme risk closure quantity and the channel residual extreme risk, forming the policy improvement quantity corresponding to each enhanced detection action. The reinforcement learning detection policy is then updated with the policy improvement quantity to form the target enhanced detection action.

[0013] Optionally, obtaining the final channel residual extreme risk specifically includes: The target enhancement detection action is executed, and the status is updated based on the action execution result, forming the updated current detection status and action execution feedback; The action condition extremile order surface is updated based on the action execution feedback, resulting in an updated action condition extremile order surface; Based on the action execution feedback, the calibration Extremile regression process is incrementally updated to form the updated Extremile regression process; The residual risk feedback reassessment will be performed based on the updated current detection status, the updated action condition Extremile order surface, and the updated Extremile regression process to form the updated channel residual extreme risk. Closed-loop control judgment is performed on the updated residual extreme risk of the channel and the cumulative detection resource consumption. When the residual extreme risk of the channel is not closed and the cumulative detection resource consumption has not reached the detection resource boundary, the target enhancement detection action is re-determined. When the residual extreme risk of the channel is closed or the cumulative detection resource consumption reaches the detection resource boundary, the final residual extreme risk of the channel is formed.

[0014] Optionally, obtaining the high-sensitivity detection results for the hazardous materials specifically includes: Risk deviation generation processing is performed on the residual extreme risks of the final channel to form the channel residual risk deviations corresponding to each detection channel; Based on the status determination of the residual risk deviation of the channel, a set of dangerous goods status and high-risk detection channels is formed. The results of hazardous materials detection are correlated with the final channel residual extreme risk and the results of high-risk detection channels to form highly sensitive hazardous materials detection results.

[0015] A highly sensitive detection system for hazardous materials based on reinforcement learning according to an embodiment of the present invention includes: The evidence status construction module is used to acquire multi-channel detection response data and detection equipment operating status to form a set of detection evidence and the current detection status; The order surface construction module is used to construct an enhanced probe action set and perform a binary additivity Choquet integral to form an action condition extremile order surface; The verification and calibration processing module is used to identify the censored detection trajectory in the historical verification trajectory, perform multiplicative active limit censoring correction to obtain the verification availability weight, and weight the historical verification trajectory with the verification availability weight to form a historical verification and calibration parameter set; The extreme regression reinforcement module is used to calibrate the action condition extremile order surface and extremile regression process with the historical verification calibration parameter set, and to construct an extremile regression-driven reinforcement learning process to obtain the dual-track extreme cumulative detection loss. The detection strategy generation module is used to form an extreme risk closure quantity based on the extreme detection loss of the dual-track system, form a channel residual extreme risk based on the execution action trajectory, and update the reinforcement learning detection strategy with the extreme risk closure quantity and the channel residual extreme risk to form target-enhanced detection actions. The closed-loop feedback control module is used to execute target enhancement detection actions and update the current detection state, action condition Extremile order surface and channel residual extreme risk. When the channel residual extreme risk is not closed and the resource boundary has not been reached, the target enhancement detection actions are redefined to form the final channel residual extreme risk. The detection result output module is used to determine the result status based on the final channel residual extreme risk, and to generate highly sensitive detection results for hazardous materials.

[0016] The beneficial effects of this invention are: This invention employs a binary additive Choquet integral to perform non-additive fusion of detection evidence, enabling the synergistic, attenuation, and redundancy relationships among hazard response intensity, response attenuation, background fluctuation, and detection channel consistency to jointly participate in risk assessment. This overcomes the problem that linear fusion cannot reflect the effects of obstruction, packaging, material overlap, and channel conflict. Furthermore, this invention constructs an Extremile order surface for action conditions based on the interaction reduction of action evidence and the residual interaction integral value. This allows different detection areas and different enhanced detection actions to adopt extreme risk concern levels that match the residual interaction risk after the action is executed, thereby improving the high-sensitivity detection capability of weak-response hazardous materials and under complex interference conditions.

[0017] This invention performs multiplicative active limit censoring correction on censored probe trajectories for which historical verification results have not been returned. It corrects the sample missing bias of historical verification trajectories by using verification availability weights and forms a historical verification calibration parameter set. Using multiplicative active limit censoring correction can retain the detection process information that has been observed in the censored probe trajectories, avoiding the decrease in the representativeness of training samples caused by directly removing censored probe trajectories. This ensures that the action condition extremile order surface and extremile regression process are based on historical data corrected for verification availability, thereby improving the accuracy and stability of extreme missed detection risk estimation.

[0018] This invention employs an Extremile regression process instead of the conventional regression method that primarily focuses on average loss. By altering the contribution of high-loss samples to the update of regression parameters through the Extremile order of the action condition, the reinforcement learning process focuses on learning the missed detection results of hazardous materials with low occurrence probability and high disposal loss. Furthermore, this invention constructs a dual-track joint training mechanism with two paths: executing enhanced detection actions and maintaining current detection parameters. This obtains the extreme cumulative detection loss of both tracks, enabling the reinforcement learning detection strategy to compare the extreme risk reduction effect of executing actions with the extreme missed detection risk of maintaining current detection parameters, thereby enhancing the adaptability of action value estimation to hazardous material detection scenarios.

[0019] This invention forms an extreme risk closure quantity based on the cumulative detection loss of dual-track extremes, and forms the residual extreme risk of the channel by combining the execution action trajectory. The extreme risk closure capability is used as the risk benefit, and the residual extreme risk of the channel is used as the risk cost, driving the update of the reinforcement learning detection strategy. After the target enhancement detection action is executed, this invention continues to update the current detection state, action condition Extremile order surface, Extremile regression process, and residual extreme risk of the channel until the risk is closed or the detection resource boundary is reached. In this way, while reducing the extreme missed detection risk of dangerous goods, it controls the detection time, number of scans, equipment load and energy consumption, and improves the sensitivity, reliability and detection resource utilization efficiency of dangerous goods detection. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a highly sensitive detection method for hazardous materials based on reinforcement learning proposed in this invention; Figure 2 This is a schematic diagram illustrating the closed-loop optimization of the Extremile regression-driven reinforcement learning detection strategy for a high-sensitivity detection method for hazardous materials based on reinforcement learning proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-2 A highly sensitive method for detecting dangerous items based on reinforcement learning includes the following steps: Acquire multi-channel detection response data and detection equipment operating status to form a set of detection evidence and the current detection status; Construct an enhanced set of detection actions and perform a binary additive Choquet integral to form an action-conditional extremile order surface; Identify censored detection trajectories in historical verification trajectories, perform multiplicative active limit censoring correction to obtain verification availability weights, and weight historical verification trajectories with verification availability weights to form a historical verification calibration parameter set; The Extremile order surface and Extremile regression process of the action condition are calibrated using the historical verification calibration parameter set, and an Extremile regression-driven reinforcement learning process is constructed to obtain the dual-track extreme cumulative detection loss. The extreme risk closure quantity is formed based on the extreme detection loss of the dual-track system, and the residual extreme risk of the channel is formed based on the execution action trajectory. The extreme risk closure quantity and the residual extreme risk of the channel are used to update the reinforcement learning detection strategy and form the target-enhanced detection action. Execute target enhancement detection actions and update the current detection status, action conditions, Extremile order surface, and channel residual extreme risk. If the channel residual extreme risk is not closed and the resource boundary is not reached, redetermine the target enhancement detection actions to form the final channel residual extreme risk. The final result status is determined based on the residual extreme risks in the final channel, forming a highly sensitive detection result for hazardous materials.

[0023] In this embodiment, obtaining the set of detection evidence and the current detection state specifically includes: Multi-channel detection response data and detection equipment operating status of hazardous materials are acquired according to a unified detection cycle. Response preprocessing is performed on the multi-channel detection response data to form preprocessed detection response data. The multi-channel detection response data is a data set recording the response location, response intensity, and response distribution of hazardous materials according to the detection channels. The detection equipment operating status is a data set characterizing the actual operating conditions of the detection equipment during the formation of the multi-channel detection response data. The detection equipment operating status includes detection excitation status, response acquisition status, scan execution status, detection channel working status, and equipment load status. The response preprocessing involves establishing the response correspondence between different detection channels according to the acquisition time and detection location, and eliminating time deviation, spatial deviation, and response dimension differences between detection channels based on the response benchmark and response scale of each detection channel. The preprocessed detection response data is the multi-channel detection response data after completing response correspondence and dimension unification. Detection evidence extraction is performed on the preprocessed detection response data to form a set of detection evidence. The detection evidence extraction involves dividing the detection space of hazardous materials into detection units with a unified spatial correspondence, analyzing the intensity of the hazardous response, the degree of response attenuation, the degree of background fluctuation, and the degree of consistency of the detection channel for each detection unit, and forming detection evidence according to a fixed evidence order. The hazard response intensity is used to characterize the response degree of a detection unit that has a hazard response orientation but has not reached the stable identification condition; the response attenuation degree is used to characterize the response loss of the detection unit caused by packaging, obstruction and material overlap; the background fluctuation degree is used to characterize the deviation of the detection unit response from the background response of adjacent detection units; the detection channel consistency degree is used to characterize the consistency of the hazard orientation of the same detection unit in different detection channels; and the detection evidence set is a data set composed of all detection evidence arranged according to the detection position. A state construction process is performed on the set of detected evidence and the operating status of the detection equipment to form the current detection state. The state construction process involves organizing the set of detected evidence according to the detection location and the detection channel identifier, organizing the operating status of the detection equipment according to the acquisition time and the detection channel identifier, establishing a state correspondence between the detected evidence with the same detection channel identifier and the corresponding acquisition time and the operating status of the detection equipment, and combining all state correspondences according to the detection location.

[0024] The detection device is equipped with four detection channels, with a unified detection period of 100ms. The response sampling frequency of each detection channel is set to 2kHz. The time correspondence error between detection channels is controlled within 5ms, and the spatial correspondence error is controlled within 2mm. The hazardous material detection space is divided into detection units with a size of 5mm×5mm. The evidence strength of each detection evidence is normalized to the range of 0 to 1. The hazard response intensity, response attenuation degree, background fluctuation degree, and detection channel consistency degree are written into the detection evidence set in a fixed order and, together with the detection device operating status within the same detection period, form the current detection status.

[0025] In this embodiment, obtaining the Extremile order surface of the action condition specifically includes: Based on the current detection status, the feasibility of the detection control methods that the detection equipment can execute is screened. The detection control methods that meet the equipment operating conditions, detection resource boundaries, and hazardous material detection conditions are combined into an enhanced detection action set. The equipment operating conditions are the execution conditions that the detection equipment meets the hardware working range under the current detection status. The detection resource boundaries are the detection time, number of scans, equipment load, and energy usage limits that the detection equipment is allowed to consume within the current detection cycle. The hazardous material detection conditions are the allowed execution range of the enhanced detection actions determined based on the physical properties, packaging status, spatial dimensions, and safety detection restrictions of the hazardous materials. The enhanced detection action set is the set of actions that the reinforcement learning process can select and control the detection equipment to execute under the current detection status. Each enhanced detection action in the enhanced detection action set has a corresponding action identifier, effective detection area, effective detection channel, control parameter range, and resource consumption. For each piece of detected evidence in the detection evidence set and each enhanced detection action in the enhanced detection action set, action evidence association processing is performed. The degree of response disclosure of the enhanced detection action to the corresponding detected evidence is determined based on the current detection state. This degree of response disclosure is then combined with the evidence strength of the corresponding detected evidence to form the action-related detection evidence quantity. All action-related detection evidence quantities corresponding to the same detection area and the same enhanced detection action are combined according to a fixed evidence order to form an action-related detection evidence vector. The action evidence association processing involves: judging the response enhancement effect, attenuation removal effect, interference suppression effect, and channel conflict reduction effect of the enhanced detection action on the corresponding detected evidence, and determining the corresponding response disclosure degree based on the judgment result. The response disclosure degree is the comprehensive degree of the enhanced detection action's response enhancement, attenuation reduction, interference suppression, and channel conflict reduction effects on the detected evidence, used to characterize the visibility of the detected evidence under the action of the enhanced detection action. The action-related detection evidence quantity is the amount of evidence that the detected evidence affects the extreme risk of missed detection of hazardous materials under the corresponding enhanced detection action. The action-related detection evidence vector is a data set characterizing the interaction relationship between detected evidence in the same detection area and the same enhanced detection action. Based on the degree of response revelation, the residual evidence vector associated with each action is reconstructed to form the evidence vector before action execution and the residual evidence vector after action execution. The additive fuzzy measure is reconstructed according to the joint revelation relationship between the enhanced detection action and the individual detection evidence and the detection evidence pair. The additive fuzzy measure is then used to perform additive Choquet integration on the evidence vector before action execution and the residual evidence vector after action execution, respectively, to obtain the baseline interaction integral value and the residual interaction integral value. The difference is then calculated to form the action evidence interaction reduction amount. The reconstructed action condition residual is as follows: based on the response disclosure level, the expected evidence portion to be revealed by the enhanced detection action is deducted from the amount of detection evidence associated with each action, and the remaining evidence portion is used to form the residual evidence vector after the action is executed; the reconstructed additive fuzzy measure is as follows: the measure value of each individual detection evidence is adjusted according to the response disclosure level corresponding to each individual detection evidence, and the interaction measure value of the detection evidence pair is adjusted according to the joint disclosure relationship corresponding to each detection evidence pair, so that the reconstructed additive fuzzy measure reflects the actual disclosure capability of the enhanced detection action on individual detection evidence and detection evidence pairs in the current detection area; the additive fuzzy measure is a non-additive weight structure that retains only the contribution of individual detection evidence and the interaction contribution of detection evidence pairs, and sets the interaction contribution of three or more detection evidences to zero; the additive Choquet integral is as follows: the evidence contribution of individual detection evidence and the interaction contribution of detection evidence pairs are accumulated respectively, and the interaction of three or more detection evidences is restricted to zero by non-additive integral processing; Based on the action evidence interaction reduction amount and residual interaction integral value corresponding to each enhanced detection action, closed residual competition sorting is performed on the enhanced detection actions in the same detection area. The action condition Extremile order is determined according to the closed residual competition sorting result. The action condition Extremile order with consistent evidence interaction structure and the same action identifier in adjacent detection areas is made into order continuity. Order transitions are preserved between detection areas where the evidence interaction relationship is reversed, forming the action condition Extremile order surface. The closed residual competition sorting is as follows: prioritize increasing the Extremile order of enhanced detection actions with larger residual interaction integral values ​​and smaller action evidence interaction reduction, and decrease the Extremile order of enhanced detection actions with smaller residual interaction integral values ​​and larger action evidence interaction reduction. The Extremile order of the action condition is: an extreme risk concern parameter determined based on the action evidence interaction reduction amount and residual interaction integral value of the corresponding enhanced detection action. The process of determining the Extremile order of the action condition is as follows: the residual interaction integral values ​​corresponding to each enhanced detection action are normalized and sorted within the same detection area, and the normalization result is used as the residual risk order position. Then, the ratio of the action evidence interaction reduction amount to the interaction integral value before the action is executed is calculated to form the interaction reduction ratio. The residual risk order position is deducted in reverse according to the interaction reduction ratio, and the deduction result is limited to the range of zero to one. It is then mapped to the preset minimum Extremile order and the preset maximum Extremile order to form the action condition Extremile order, which is used to control the degree of attention of the Extremile regression process to high-loss missed detection results in the corresponding detection area. The evidence interaction structure is the composition relationship of detection evidence pairs corresponding to positive and negative interaction contributions within the same detection region. The order continuity means limiting the magnitude of Extremile order change corresponding to the same enhanced detection action between adjacent detection regions with consistent evidence interaction structures. The order transition means retaining the discontinuous change of the corresponding Extremile order when the detection evidence pair changes from positive to negative interaction contribution or from negative to positive interaction contribution. The action condition Extremile order surface is used to enable the Extremile regression process to adopt an extreme loss concern level that matches the residual interaction risk after the action for different detection regions and different enhanced detection actions.

[0026] The enhanced detection action set includes nine enhanced detection actions, formed by combining three detection excitation intensity levels (80%, 100%, and 120%) with three scanning angles (-15°, 0°, and +15°). The individual detection evidence measure value and the interaction measure value of detection evidence pairs are normalized to the range of 0 to 1. The preset minimum Extremile order is set to 0.80, and the preset maximum Extremile order is set to 0.99. Between adjacent detection areas with consistent evidence interaction structures, the change in Extremile order corresponding to the same enhanced detection action is limited to 0.03. When the evidence interaction relationship undergoes a positive-to-negative reversal, the change range limitation is removed, but the order transition is retained. The detection resource boundaries for a single detection cycle are set as follows: detection duration 3 seconds, number of scans 6, equipment load rate 85%, and energy consumption 120J.

[0027] This invention characterizes the synergistic, weakening, and redundant relationships between detection evidence through a binary additive Choquet integral, and reconstructs the residual evidence structure after the action is executed by combining the response revelation degree of the enhanced detection action. This allows the Extremile order of the action condition to dynamically change with the detection area, the enhanced detection action, and the residual interaction risk, thereby improving the sensitivity, stability, and accuracy of identifying extreme missed detection risks of dangerous goods under complex occlusion, response attenuation, and channel conflict conditions.

[0028] In this embodiment, obtaining the historical verification calibration parameter set specifically includes: Historical verification trajectories are acquired, and trajectory verification status marking and verification availability stratification are performed on these trajectories to form complete detection trajectories, censored detection trajectories, and verification availability stratification. The historical verification trajectories are historical data sequences that record the current detection status, enhanced detection actions, and Extremile order of the action conditions according to the detection cycle. The trajectory verification status marking process compares the actual verification completion time of the historical verification trajectory with the observable time allowed for continuous verification. If a historical verification result is obtained within the observable time, it is marked as a complete detection trajectory; if no historical verification result is obtained within the observable time, it is marked as a censored detection trajectory. The verification availability stratification process divides the historical verification trajectories according to the historical enhanced detection action identifier and the order interval to which the historical action condition Extremile order belongs. The historical verification result is the final verification conclusion obtained when the historical verification trajectory ends. Multiplicative active limit censoring correction is performed for each verification availability layer to form a verification availability function. The multiplicative active limit censoring correction is as follows: the censoring time sequence is determined according to the observation and verification duration of the censored detection trajectory. For each censoring time, the number of historical verification trajectories that are still observable and the number of historical verification trajectories that have been censored are counted. The verification retention ratio for the corresponding censoring time is determined based on the two quantities, and the verification retention ratio is continuously multiplied in the order of the censoring time. The verification availability function is used to perform weight generation processing to form verification availability weights. The weight generation processing is as follows: read the verification availability value of the historical verification trajectory before the corresponding observation verification time, and use the reciprocal of the verification availability value to form the process verification availability weight, and use the reciprocal of the verification availability value to form the terminal verification availability weight. The terminal verification availability weight of the censored detection trajectory is set to zero. The process verification availability weight is used to correct the sample missing bias caused by the unavailability of historical verification results during the observation process of the historical verification trajectory. The terminal verification availability weight is used to correct the representativeness of the complete detection trajectory in the statistics of historical verification results. The verification availability weight includes process verification availability weight and terminal verification availability weight. Historical verification weighted calibration is performed on historical verification trajectories using verification availability weights to form a historical verification calibration parameter set. The historical verification weighted calibration process involves weighting the process data in the historical verification trajectory using process verification availability weights to form weighted historical trajectory samples, and weighting the historical verification results corresponding to the complete detection trajectory using terminal verification availability weights to form a historical verification loss weighted distribution. The historical verification calibration parameter set includes a verification availability function, verification availability weights, weighted historical trajectory samples, and a historical verification loss weighted distribution.

[0029] This invention retains effective observation information in censored detection trajectories through multiplicative censoring correction and uses verification availability weights to correct sample representativeness bias caused by missing historical verification results. This makes the weighted distribution of historical trajectory samples and historical verification loss closer to the actual detection risk distribution, thereby improving the accuracy, stability and reliability of action condition extremile order surface calibration, extremile regression training and extreme missed detection risk estimation of dangerous goods.

[0030] In this embodiment, obtaining the extreme cumulative detection loss of the dual-track system specifically includes: The order-preserving calibration of the action condition Extremile order surface is performed using the historical verification calibration parameter set to form the calibration action condition Extremile order surface. The order-preserving calibration is as follows: a monotonic calibration relationship is established based on the historical action condition Extremile order and the cumulative position of the historical verification loss after censoring correction, and the order of each action condition Extremile is limited to a preset order range. Enhanced detection action execution segments and detection parameter holding segments are extracted from weighted historical trajectory samples to form execution track training samples and holding track training samples, respectively. Based on the immediate detection loss and the discounted extreme cumulative detection loss corresponding to the next detection state, execution track cumulative loss targets and holding track cumulative loss targets are formed. The holding track training samples are used to characterize the risk of missing detection of dangerous goods when the current detection parameters remain unchanged. The execution track training samples, hold track training samples, and calibration action condition extremile order are input into the extremile regression process to determine the cumulative position of the execution track cumulative loss target and the hold track cumulative loss target in the historical verification loss weighted distribution. Based on the corresponding calibration action condition extremile order, the cumulative position is converted into execution track extremile weights and hold track extremile weights. The deviation between the cumulative loss target and the extremile action value function prediction result is weighted together using the process verification availability weight and the corresponding extremile weight. The extremile action value function parameters are then updated based on the weighted deviation to form the calibration extremile regression process. The Extremile regression process is a conditional regression process that takes the current detection state, the enhanced detection action, and the Extremile order of the calibration action as inputs, and the extreme cumulative detection loss as output. It changes the contribution of high-loss samples to the update of regression parameters by adjusting the Extremile order of the calibration action. The dual-track joint training is as follows: Under the same detection state and the same Extremile order of the calibration action, the enhanced detection action and the hold-state action are respectively input into the Extremile action value function to obtain the execution track prediction result and the hold-state prediction result. The deviation between the execution track cumulative loss target and the execution track prediction result and the hold-state cumulative loss target and the hold-state prediction result are weighted by the process verification availability weight and the corresponding Extremile weight, respectively. The two weighted deviations are combined to form the dual-track joint training objective. The Extremile action value function is updated according to the dual-track joint training objective, so that the Extremile regression process fits the extreme cumulative detection loss corresponding to the enhanced detection action and the hold-state current detection parameters in a unified parameter space. The current detection state, the set of enhanced detection actions, and the Extremile order surface of the calibration action conditions are input into the calibration Extremile regression process to obtain the extreme cumulative detection loss of the execution track and the extreme cumulative detection loss of the hold track for each enhanced detection action. Pair relationships are established according to the same detection region and enhanced detection action to form a dual-track extreme cumulative detection loss. The calibration Extremile regression process is used as the action value estimation process for the reinforcement learning detection strategy. The current detection state is used as the reinforcement learning state, the set of enhanced detection actions is used as the reinforcement learning action space, and the dual-track extreme cumulative detection loss is used as the extreme action value basis for enhanced detection actions. An Extremile regression-driven reinforcement learning process is constructed. The dual-track extreme cumulative detection loss is used to characterize the high-loss risk of missed detection of hazardous materials corresponding to the two paths of executing enhanced detection actions and maintaining the current detection parameters. The extreme action value basis is used to characterize the change in extreme cumulative detection loss caused by executing enhanced detection actions relative to maintaining the current detection parameters under the current detection state.

[0031] This invention corrects the deviation between the Extremile order of the action condition and the historical extreme loss by order-preserving calibration, and uses dual-track joint training to synchronously fit and execute enhanced detection actions while maintaining the extreme cumulative detection loss corresponding to the current detection parameters. This enables the Extremile regression process to form a stable action value estimate for high-loss missed detection samples, thereby improving the reinforcement learning detection strategy's ability to identify extreme risk differences, the accuracy of action selection, and the ability to suppress the risk of missed detection of dangerous goods.

[0032] In this embodiment, obtaining the target enhancement detection action specifically includes: The extreme cumulative detection loss of the dual-track system is processed to generate an extreme risk closure value, which is then used to generate an extreme risk closure value for each enhanced detection action. The process is as follows: the extreme cumulative detection loss of the hold track corresponding to the same detection area and the same enhanced detection action is subtracted from the extreme cumulative detection loss of the execution track. The positive difference is determined as the regional risk closure value. The regional closure weight is determined based on the corresponding extreme cumulative detection loss of the hold track. The regional risk closure values ​​are then aggregated according to the regional closure weight. The extreme risk closure value is used to characterize the high loss missed detection risk that can be reduced by performing the enhanced detection action relative to maintaining the current detection parameters. The executed enhanced detection actions and the set of enhanced detection actions are subjected to trajectory extension processing to form the execution action trajectory corresponding to each enhanced detection action. The trajectory residual coefficient is formed based on the response disclosure degree. The trajectory extension processing is as follows: the executed enhanced detection actions are arranged according to the execution time, and each enhanced detection action to be evaluated is connected to the end of the arrangement result. The execution action trajectory is an action sequence composed of the executed enhanced detection actions and the enhanced detection actions to be evaluated in the execution order. The trajectory residual coefficient is a value formed by continuously multiplying the response disclosure degree of each enhanced detection action corresponding to each enhanced detection action in the execution action trajectory. It is used to characterize the proportion of detection evidence that is expected to remain undiscovered after the execution action trajectory is completed. Based on the execution action trajectory and trajectory residual coefficient, the residual risk of the execution channel is aggregated to form the residual extreme risk of the channel corresponding to each enhanced detection action. The channel residual risk aggregation is as follows: the cumulative detection loss of the extreme extreme of the execution trajectory corresponding to each detection area is combined with the trajectory residual coefficient to form the regional residual extreme risk, and the regional residual extreme risk is aggregated according to the detection channel affiliation relationship. The channel residual extreme risk is used to characterize the high loss and missed detection risk that the corresponding detection channel is expected to retain after the execution action trajectory is completed. The extreme action value is updated based on the extreme risk closure value and the channel residual extreme risk to form the policy improvement value corresponding to each enhanced detection action. The reinforcement learning detection policy is then updated with the policy improvement value to form the target enhanced detection action. The extreme action value update is as follows: the extreme risk closure value is used as the risk gain of the enhanced detection action, the channel residual extreme risk is used as the risk cost of the enhanced detection action, and the extreme action value represented by the dual-track extreme cumulative detection loss is combined to form the policy improvement value corresponding to each enhanced detection action. The reinforcement learning detection policy update is as follows: the action selection probability corresponding to the enhanced detection action with a larger policy improvement value is increased, the action selection probability corresponding to the enhanced detection action with a smaller policy improvement value is decreased, and the enhanced detection action with the highest updated action selection probability is determined as the target enhanced detection action.

[0033] In this embodiment, obtaining the final channel residual extreme risk specifically includes: The target enhancement detection action is executed, and the action execution result is fed back to update the status, forming an updated current detection status and action execution feedback. The feedback status update is based on the multi-channel detection response data after the action execution and the operating status of the detection equipment to reconstruct the current detection status. The action condition Extremile order surface is updated based on the action execution feedback to form an updated action condition Extremile order surface. The feedback order update is to redetermine the residual interaction integral value, the action evidence interaction reduction amount, and the action condition Extremile order based on the corrected response disclosure level. The calibration Extremile regression process is updated by performing incremental feedback updates based on the action execution feedback, resulting in an updated Extremile regression process. The incremental feedback update is to form a cumulative loss update target based on the actual detection loss and to update the Extremile action value function using feedback Extremile weights. The residual risk feedback reassessment will be performed based on the updated current detection status, the updated action conditions, the Extremile order surface, and the updated Extremile regression process to form the updated channel residual extreme risk. The residual risk feedback reassessment is to re-obtain the dual-track extreme cumulative detection loss and update the trajectory residual coefficient and channel residual extreme risk in combination with the actual response revelation degree. Closed-loop control judgment is performed on the updated channel residual extreme risk and cumulative detection resource consumption. When the channel residual extreme risk is not closed and the cumulative detection resource consumption has not reached the detection resource boundary, the target enhancement detection action is re-determined. When the channel residual extreme risk is closed or the cumulative detection resource consumption reaches the detection resource boundary, the final channel residual extreme risk is formed. The closed-loop control judgment is to determine whether to continue to update the reinforcement learning detection strategy based on the comparison result between the channel residual extreme risk and the preset risk closure threshold.

[0034] In this embodiment, obtaining the highly sensitive detection results of the hazardous materials specifically includes: Risk deviation generation processing is performed on the residual extreme risk of the final channel to form the channel residual risk deviation corresponding to each detection channel. The risk deviation generation processing is to compare the residual extreme risk of the final channel with the corresponding preset risk closure threshold and determine the channel residual risk deviation based on the difference between the two. Based on the status determination of the residual risk deviation of the channel, a result status of dangerous goods and a set of high-risk detection channels are formed. The result status determination is to determine the high-risk pending verification status when the residual risk deviation of any channel is greater than zero, and to classify the corresponding detection channel into the set of high-risk detection channels. When the residual risk deviation of all channels is not greater than zero, the risk closure status is determined. The results of the hazardous material outcome status, the residual extreme risk of the final channel, and the execution results of the high-risk detection channel set are correlated to form a highly sensitive hazardous material detection result. The highly sensitive hazardous material detection result is used to output a risk closure flag when the hazardous material outcome status is in a risk closure state, and to output a high-risk verification flag and a high-risk detection channel set when the hazardous material outcome status is in a high-risk pending verification state. When a risk closure flag is formed, the detection equipment is controlled to end the current detection task. When a high-risk verification flag is formed, hazardous material interception, re-inspection, alarm, and manual verification are triggered, and key re-inspection locations are determined based on the high-risk detection channel set.

[0035] A high-sensitivity detection system for hazardous materials based on reinforcement learning, comprising: The evidence status construction module is used to acquire multi-channel detection response data and detection equipment operating status to form a set of detection evidence and the current detection status; The order surface construction module is used to construct an enhanced probe action set and perform a binary additivity Choquet integral to form an action condition extremile order surface; The verification and calibration processing module is used to identify the censored detection trajectory in the historical verification trajectory, perform multiplicative active limit censoring correction to obtain the verification availability weight, and weight the historical verification trajectory with the verification availability weight to form a historical verification and calibration parameter set; The extreme regression reinforcement module is used to calibrate the action condition extremile order surface and extremile regression process with the historical verification calibration parameter set, and to construct an extremile regression-driven reinforcement learning process to obtain the dual-track extreme cumulative detection loss. The detection strategy generation module is used to form an extreme risk closure quantity based on the extreme detection loss of the dual-track system, form a channel residual extreme risk based on the execution action trajectory, and update the reinforcement learning detection strategy with the extreme risk closure quantity and the channel residual extreme risk to form target-enhanced detection actions. The closed-loop feedback control module is used to execute target enhancement detection actions and update the current detection state, action condition Extremile order surface and channel residual extreme risk. When the channel residual extreme risk is not closed and the resource boundary has not been reached, the target enhancement detection actions are redefined to form the final channel residual extreme risk. The detection result output module is used to determine the result status based on the final channel residual extreme risk, and to generate highly sensitive detection results for hazardous materials.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-channel intelligent security inspection scenario for hazardous materials, performing high-sensitivity detection of hazardous materials under conditions of packaging obstruction, material overlap, response attenuation, and detection channel conflict. The test device was equipped with four detection channels, a uniform detection period of 100ms, a response sampling frequency of 2kHz, and detection units divided into 5mm×5mm units. The maximum detection time for a single detection task was set to 3s, the maximum number of scans was set to 6, the maximum equipment load rate was set to 85%, and the maximum energy consumption was set to 120J. A total of 6000 test samples were collected, including 1800 samples of normal response, 1300 samples of packaging obstruction, 1100 samples of material overlap, 1000 samples of weak response attenuation, and 800 samples of detection channel conflict. Of these, 4800 samples formed complete detection trajectories, while 1200 samples were censored due to the lack of verification results.

[0037] After detection begins, the system acquires multi-channel detection response data and the operating status of the detection equipment. It performs time-based correspondence, spatial correspondence, and response scale unification on the response data from different detection channels. Furthermore, it extracts the hazard response intensity, response attenuation degree, background fluctuation degree, and detection channel consistency from each detection unit, forming a set of detection evidence and the current detection status. The enhanced detection action set is formed by combining three detection excitation intensity levels (80%, 100%, and 120%) with three scanning angles (-15°, 0°, and +15°), comprising a total of nine enhanced detection actions.

[0038] The system generates action-related detection evidence vectors based on the degree to which enhanced detection actions reveal the responses of each piece of detection evidence. It then performs additive Choquet integration on the evidence vectors before and after the action, obtaining the baseline interaction integral value, the residual interaction integral value, and the action evidence interaction reduction. The action condition extremile order is set between 0.80 and 0.99. Increasing the residual interaction integral value and decreasing the action evidence interaction reduction increases the corresponding action condition extremile order, thus strengthening the extremile regression process's focus on high-loss missed detection samples.

[0039] For 1200 sets of censored detection trajectories, the system performs multiplicative censoring correction with a 30-second observable duration, generating process verification availability weights and terminal verification availability weights, and forming a weighted historical trajectory sample and a weighted distribution of historical verification loss. The extremile regression training batch is set to 128 samples, with 64 training samples for the execution track and 64 training samples for the hold track. The cumulative loss discount factor is set to 0.95, and the parameter update step size is set to 0.001. Under the same current detection state and the same calibration action conditions at the extremile order, the system predicts the extreme cumulative detection loss corresponding to executing the enhanced detection action and holding the current detection parameters, forming a dual-track extreme cumulative detection loss.

[0040] The reinforcement learning detection strategy uses the extreme risk closure value as the risk reward and the residual extreme risk of the channel as the risk cost. It also combines the cumulative detection loss from both extreme risks to form the strategy improvement value, determining the target enhancement detection action from the set of enhancement detection actions. After the target enhancement detection action is executed, the system updates the current detection state, the action condition Extremile order surface, the Extremile regression process, and the residual extreme risk of the channel based on the actual detection response. The risk closure threshold for each detection channel is set to 0.08. Detection stops when the residual extreme risk of all channels is no higher than 0.08. If the risk closure condition is not met and the detection resource boundary is not reached, the target enhancement detection action is redefined.

[0041] The same test samples were processed using a fixed-threshold fusion detection method, a conventional reinforcement learning detection method, and the method of this invention. The fixed-threshold fusion detection method used linear weighted fusion and a fixed scanning order, while the conventional reinforcement learning method used the average cumulative loss as the action value and did not perform multiplicative active limit censoring correction, additive Choquet integration, and dual-track Extremile regression. The comparison results are shown in Table 1. Table 1. Comparison of the overall performance of different hazardous material detection methods

[0042] As shown in Table 1, the method of the present invention achieves a dangerous goods identification accuracy of 98.2%, which is 6.5 percentage points higher than that of the fixed threshold fusion detection method and the ordinary reinforcement learning detection method, respectively. The false negative rate of dangerous goods is reduced to 1.1%, and the recall rate of high-loss false negative samples is increased to 97.6%. Under the conditions of packaging obstruction, weak response attenuation, and detection channel conflict, the identification rate of the method of the present invention reaches 97.4%, 96.1%, and 95.8%, respectively. The extreme cumulative detection loss estimation bias is reduced to 4.3%, and the average detection time is shortened to 2.08s. This indicates that the present invention can improve the dangerous goods identification capability, the accuracy of extreme false negative risk estimation, and the detection efficiency under complex detection conditions.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A highly sensitive method for detecting hazardous materials based on reinforcement learning, characterized in that, Includes the following steps: Acquire multi-channel detection response data and detection equipment operating status to form a set of detection evidence and the current detection status; Construct an enhanced set of detection actions and perform a binary additive Choquet integral to form an action-conditional extremile order surface; Identify censored detection trajectories in historical verification trajectories, perform multiplicative active limit censoring correction to obtain verification availability weights, and weight historical verification trajectories with verification availability weights to form a historical verification calibration parameter set; The Extremile order surface and Extremile regression process of the action condition are calibrated using the historical verification calibration parameter set, and an Extremile regression-driven reinforcement learning process is constructed to obtain the dual-track extreme cumulative detection loss. The extreme risk closure quantity is formed based on the extreme detection loss of the dual-track system, and the residual extreme risk of the channel is formed based on the execution action trajectory. The extreme risk closure quantity and the residual extreme risk of the channel are used to update the reinforcement learning detection strategy and form the target-enhanced detection action. Execute target enhancement detection actions and update the current detection status, action conditions, Extremile order surface, and channel residual extreme risk. If the channel residual extreme risk is not closed and the resource boundary is not reached, redetermine the target enhancement detection actions to form the final channel residual extreme risk. The final result status is determined based on the residual extreme risks in the final channel, forming a highly sensitive detection result for hazardous materials.

2. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The acquisition of the detection evidence set and the current detection state specifically includes: According to a unified detection cycle, multi-channel detection response data of hazardous materials and the operating status of detection equipment are obtained. Response preprocessing is performed on the multi-channel detection response data to form preprocessed detection response data. Detection evidence extraction is performed on the preprocessed detection response data to form a detection evidence set; The current detection state is formed by performing state construction processing on the collection of detection evidence and the operating status of the detection equipment.

3. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The specific process of obtaining the Extremile order surface of the action condition includes: Based on the current detection status, the feasibility of the detection control methods that can be executed by the detection equipment is screened, and the detection control methods that meet the equipment operating conditions, detection resource boundaries and hazardous material detection conditions are combined into an enhanced detection action set; For each piece of detected evidence in the detection evidence set and each enhanced detection action in the enhanced detection action set, action evidence association processing is performed. The degree of response and disclosure of the enhanced detection action to the corresponding detected evidence is determined according to the current detection state. The degree of response and disclosure is combined with the evidence strength of the corresponding detected evidence to form the action-related detection evidence quantity. All action-related detection evidence quantities corresponding to the same detection area and the same enhanced detection action are combined according to a fixed evidence order to form the action-related detection evidence vector. Based on the degree of response revelation, the residual evidence vector associated with each action is reconstructed to form the evidence vector before action execution and the residual evidence vector after action execution. The additive fuzzy measure is reconstructed according to the joint revelation relationship between the enhanced detection action and the individual detection evidence and the detection evidence pair. The additive fuzzy measure is then used to perform additive Choquet integration on the evidence vector before action execution and the residual evidence vector after action execution, respectively, to obtain the baseline interaction integral value and the residual interaction integral value. The difference is then calculated to form the action evidence interaction reduction amount. Based on the action evidence interaction reduction and residual interaction integral value corresponding to each enhanced detection action, closed residual competition sorting is performed on enhanced detection actions within the same detection area. The action condition Extremile order is determined according to the closed residual competition sorting result. The action condition Extremile order with consistent evidence interaction structure and the same action identifier in adjacent detection areas is performed to perform order continuity. Order transitions are preserved between detection areas where the evidence interaction relationship is reversed, forming the action condition Extremile order surface.

4. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The acquisition of the historical verification and calibration parameter set specifically includes: Obtain historical verification trajectories, perform trajectory verification status marking and verification availability layering on the historical verification trajectories, and form complete detection trajectories, deleted detection trajectories, and verification availability layers; For each level of verification availability, multiplicative active censoring correction is performed to form a verification availability function; The verification availability function is used to perform weight generation processing to form the verification availability weight; The historical verification trajectory is subjected to historical verification weighted calibration processing based on the verification availability weight, forming a historical verification calibration parameter set.

5. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The specific methods for obtaining the extreme cumulative detection loss of the dual-track system include: The order-preserving calibration of the action condition Extremile order surface is performed using the historical verification calibration parameter set to form the calibration action condition Extremile order surface; Enhanced detection action execution segments and detection parameter holding segments are extracted from weighted historical trajectory samples to form execution trajectory training samples and holding trajectory training samples, respectively. Based on the immediate detection loss and the discounted extreme cumulative detection loss corresponding to the next detection state, execution trajectory cumulative loss targets and holding trajectory cumulative loss targets are formed. The execution track training samples, hold track training samples, and calibration action condition extremile order are input into the extremile regression process to determine the cumulative position of the execution track cumulative loss target and the hold track cumulative loss target in the historical verification loss weighted distribution. Based on the corresponding calibration action condition extremile order, the cumulative position is converted into execution track extremile weights and hold track extremile weights. The deviation between the cumulative loss target and the extremile action value function prediction result is weighted together using the process verification availability weight and the corresponding extremile weight. The extremile action value function parameters are then updated based on the weighted deviation to form the calibration extremile regression process. The current detection state, the set of enhanced detection actions, and the extremile order surface of the calibration action conditions are input into the calibration extremile regression process to obtain the extreme cumulative detection loss of the execution track and the extreme cumulative detection loss of the hold track for each enhanced detection action. Paired relationships are established according to the same detection region and enhanced detection action to form a dual-track extreme cumulative detection loss. The calibration extremile regression process is used as the action value estimation process of the reinforcement learning detection strategy. The current detection state is used as the reinforcement learning state, the set of enhanced detection actions is used as the reinforcement learning action space, and the dual-track extreme cumulative detection loss is used as the extreme action value basis of the enhanced detection actions to construct an extremile regression-driven reinforcement learning process.

6. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The specific steps for obtaining the target enhancement detection action include: The extreme cumulative detection loss of the dual-track system is processed to generate the extreme risk closure quantity corresponding to each enhanced detection action; The trajectory extension process is performed on the enhanced detection actions and the set of enhanced detection actions to form the execution action trajectory corresponding to each enhanced detection action, and the trajectory residual coefficient is formed based on the degree of response disclosure; Based on the execution action trajectory and trajectory residual coefficient, the residual risk of the execution channel is aggregated to form the extreme residual risk of the channel corresponding to each enhanced detection action; The extreme action value is updated based on the extreme risk closure quantity and the channel residual extreme risk, forming the policy improvement quantity corresponding to each enhanced detection action. The reinforcement learning detection policy is then updated with the policy improvement quantity to form the target enhanced detection action.

7. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The determination of the final channel residual extreme risk specifically includes: The target enhancement detection action is executed, and the status is updated based on the action execution result, forming the updated current detection status and action execution feedback; The action condition extremile order surface is updated based on the action execution feedback, resulting in an updated action condition extremile order surface; Based on the action execution feedback, the calibration Extremile regression process is incrementally updated to form the updated Extremile regression process; The residual risk feedback reassessment will be performed based on the updated current detection status, the updated action condition Extremile order surface, and the updated Extremile regression process to form the updated channel residual extreme risk. Closed-loop control judgment is performed on the updated residual extreme risk of the channel and the cumulative detection resource consumption. When the residual extreme risk of the channel is not closed and the cumulative detection resource consumption has not reached the detection resource boundary, the target enhancement detection action is re-determined. When the residual extreme risk of the channel is closed or the cumulative detection resource consumption reaches the detection resource boundary, the final residual extreme risk of the channel is formed.

8. The highly sensitive detection method for hazardous materials based on reinforcement learning according to claim 1, characterized in that, The high-sensitivity detection results for hazardous materials are obtained specifically in the following ways: Risk deviation generation processing is performed on the residual extreme risks of the final channel to form the channel residual risk deviations corresponding to each detection channel; Based on the status determination of the residual risk deviation of the channel, a set of dangerous goods status and high-risk detection channels is formed. The results of hazardous materials detection are correlated with the final channel residual extreme risk and the results of high-risk detection channels to form highly sensitive hazardous materials detection results.

9. A high-sensitivity detection system for hazardous materials based on reinforcement learning according to claim 1, comprising executing a high-sensitivity detection method for hazardous materials based on reinforcement learning according to any one of claims 1 to 8, characterized in that, include: The evidence status construction module is used to acquire multi-channel detection response data and detection equipment operating status to form a set of detection evidence and the current detection status; The order surface construction module is used to construct an enhanced probe action set and perform a binary additivity Choquet integral to form an action condition extremile order surface; The verification and calibration processing module is used to identify the censored detection trajectory in the historical verification trajectory, perform multiplicative active limit censoring correction to obtain the verification availability weight, and weight the historical verification trajectory with the verification availability weight to form a historical verification and calibration parameter set; The extreme regression reinforcement module is used to calibrate the action condition extremile order surface and extremile regression process with the historical verification calibration parameter set, and to construct an extremile regression-driven reinforcement learning process to obtain the dual-track extreme cumulative detection loss. The detection strategy generation module is used to form an extreme risk closure quantity based on the extreme detection loss of the dual-track system, form a channel residual extreme risk based on the execution action trajectory, and update the reinforcement learning detection strategy with the extreme risk closure quantity and the channel residual extreme risk to form target-enhanced detection actions. The closed-loop feedback control module is used to execute target enhancement detection actions and update the current detection state, action condition Extremile order surface and channel residual extreme risk. When the channel residual extreme risk is not closed and the resource boundary has not been reached, the target enhancement detection actions are redefined to form the final channel residual extreme risk. The detection result output module is used to determine the result status based on the final channel residual extreme risk, and to generate highly sensitive detection results for hazardous materials.

Images