Robotic security control system, method and medium based on deep learning

Abstract

This invention relates to the field of artificial intelligence security robot technology, and discloses a deep learning-based robot security control system, method, and medium, comprising the following steps: a security robot collects and preprocesses information within a patrol area to generate a security event flow; a dynamic security scenario graph with security semantic attributes is generated based on the security event flow; the dynamic security scenario graph with security semantic attributes and the security event flow are input into an improved dynamic graph representation network; candidate security events are generated through a risk evolution intensity modeling module; active supplementary data acquisition control commands are generated through an event confusion and disambiguation module; the security robot is controlled to perform supplementary data acquisition actions according to the active supplementary data acquisition control commands to generate supplementary patrol observation data and generate a robot closed-loop control strategy. This invention uses a dynamic graph representation network to achieve closed-loop security control of the robot.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security robot technology, and in particular to a robot security control system, method and medium based on deep learning. Background Technology

[0002] With the increasing application of security patrol robots in parks, factories, buildings and key controlled areas, existing robot security control methods typically acquire on-site information through cameras, radar, infrared sensors and voice acquisition devices, and combine them with target detection, behavior recognition, path planning and alarm linkage to achieve patrol monitoring.

[0003] Existing methods often rely on single identification results or fixed rules to execute alarms, tracking, and route adjustments, lacking dynamic correlation modeling of personnel targets, access control, abnormal sound sources, obstructed areas, fixed camera coverage areas, and robot reachability. In complex scenarios, candidate events can easily become confused between normal waiting, suspicious loitering, tailgating, and unauthorized approach to access control. Robots struggle to determine the source of missing evidence and cannot proactively adjust their observation position, supplement data collection, or update risk assessments based on missing evidence, resulting in insufficient continuity in security incident identification and handling control.

[0004] Therefore, how to provide a robot security control system, method, and medium based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a robot security control system, method, and medium based on deep learning. This invention fully utilizes improved dynamic graph representation networks, security event flow modeling, dynamic security scenario graph construction, event confusion disambiguation, and evidence gap graph generation technologies. It describes in detail the method for security robots to perform dynamic relationship analysis, risk evolution judgment, missing evidence identification, proactive supplementary data collection control, and closed-loop processing of multimodal security data during patrols. It has the advantages of high event recognition accuracy, strong adaptability to complex scenarios, low false alarm and false negative rates, timely robot response, and good security control continuity.

[0006] The robot security control method based on deep learning according to an embodiment of the present invention includes the following steps: Security robots collect information within patrol areas, preprocess the collected information to generate patrol observation datasets, and generate security event streams based on these datasets. A dynamic security scenario graph is constructed based on the security event flow. Node status information is configured for graph nodes, and relationship persistence information and relationship change information are configured for graph edges, generating a dynamic security scenario graph with security semantic attributes. An improved dynamic graph representation network is input with dynamic security context graphs and security event streams that have security semantic attributes. Through a security event stream input module, a security node memory update module, and a security relationship attention aggregation module, security node memory representations and security relationship attention representations are generated. An improved dynamic graph representation module for modeling the risk evolution intensity in the network generates candidate security events based on the security node memory representation, security relationship attention representation, and security event flow. The event confusion and disambiguation module generates event confusion relationships based on the candidate results of security events. The evidence gap diagram generation module generates an evidence gap diagram based on the event confusion relationships and generates an active supplementary sampling control instruction based on the evidence influence edge between the missing evidence node and the candidate event node. The security robot is controlled to perform supplementary sampling actions according to the active supplementary sampling control command, and supplementary sampling patrol observation data is generated. The robot closed-loop control strategy is generated through the supplementary sampling patrol observation data.

[0007] Optionally, the preprocessing includes time synchronization, spatial alignment, and data source marking.

[0008] Optionally, the step of constructing a dynamic security scenario graph based on the security event flow, configuring node state information for graph nodes, configuring relationship persistence information and relationship change information for graph edges, and generating a dynamic security scenario graph with security semantic attributes specifically includes: Read each security event record from the security event stream and generate a graph to construct the source record set; Based on the source record set of the graph, determine the graph nodes of the dynamic security scenario graph and generate a graph node set; Based on the node type and source information of each graph node in the graph node set, generate a graph node set with node status information; Based on the relationship types, target objects, associated regions, and robot poses in the graph source record set, generate a graph edge set with relationship persistence information; Configure relation change information for each graph edge in the graph edge set with relation persistence information, and generate a graph edge set with both relation persistence and relation change information; Based on the graph node set with node state information and the graph edge set with relationship persistence information and relationship change information, a dynamic security scenario graph with security semantic attributes is generated.

[0009] Optionally, the step of inputting a dynamic security context graph with security semantic attributes and a security event flow into an improved dynamic graph representation network, and generating security node memory representations and security relationship attention representations through a security event flow input module, a security node memory update module, and a security relationship attention aggregation module, specifically includes: The network is represented by an improved dynamic graph representation of security scenarios with security semantic attributes and security event flow inputs, generating network input event units. The security event stream input module performs event vectorization on the network input event units to generate fixed-length event input vectors. The security node memory update module reads the historical node memory of the network input event unit and generates an event message vector. The security node memory update module performs gating updates on the node memory based on the event message vector and the historical node memory of the target graph node to obtain the security node memory representation; The security relationship attention aggregation module performs relational semantic encoding on the graph edges connecting the target graph node and its adjacent graph nodes to obtain the relation type embedding vector. Relationship persistence information and relationship change information are encoded to obtain a relationship change vector. The relationship type embedding vector and the relationship change vector are concatenated and mapped to generate a relationship semantic vector. And calculate the relational attention score of the neighboring graph node to the target graph node based on the target graph node memory, the neighboring graph node memory, and the relational semantic vector; The relational attention scores of each adjacent graph node connected to the same target graph node are normalized to generate the relational weights of each adjacent graph node to the target graph node. The security relationship attention aggregation module aggregates neighborhood information of target graph nodes based on relationship weights, adjacent graph node memories, and relationship semantic vectors to obtain a security relationship attention representation.

[0010] Optionally, the risk evolution intensity modeling module in the improved dynamic graph representation network, which generates security event candidate results based on security node memory representation, security relationship attention representation, and security event flow, specifically includes: The target risk evolution input sequence is generated by reading the security node memory representation, security relationship attention representation and security event flow through the risk evolution intensity modeling module. Based on the target risk evolution input sequence, the node memory changes, relation attention changes, and relation persistence state changes of the same target object between the current security event record and the previous security event record are calculated to generate a risk evolution difference representation; The risk evolution intensity modeling module calculates the risk intensity of the risk evolution difference representation and generates the risk evolution intensity value of the target object under different candidate event types. Based on the risk evolution intensity value, the direction of risk change for the same target object in continuous security event records is calculated to generate the risk evolution state; The security event candidate output module calculates the candidate event score for each candidate event type based on the risk evolution status, security node memory representation, and security relationship attention representation. The scores of each candidate event corresponding to the same target object are normalized to form a confidence distribution of candidate events within the same distribution range; Based on the confidence distribution of candidate events, risk evolution status, security node memory representation, and security relationship attention representation, the target object, event occurrence area, event association graph node, and event association graph edge in the security event candidate results are determined. The security event candidate output module combines candidate event types, target objects, event occurrence areas, candidate event confidence distribution, event association graph nodes, and event association graph edges to generate security event candidate results.

[0011] Optionally, the step of generating event confusion relationships through the event confusion and disambiguation module based on the candidate results of security events, generating an evidence gap map through the evidence gap map generation module based on the event confusion relationships, and generating an active supplementary data collection control instruction based on the evidence influence edges between missing evidence nodes and candidate event nodes specifically includes: The event obfuscation and disambiguation module reads the candidate results of security events, and forms candidate event pairs in order of increasing confidence difference of candidate events, generating a set of candidate event pairs. Based on the candidate event pair set, candidate event confidence distribution, event association graph nodes, event association graph edges, risk evolution state, security node memory representation, and security relationship attention representation, calculate the confusion intensity of each candidate event pair and generate a candidate event confusion intensity set. Based on the set of candidate event confusion intensity, an evidence integrity analysis is performed on each candidate event to generate a set of event confusion causes; The set of candidate event pairs, the set of candidate event confusion intensity, and the set of event confusion causes are associated to generate event confusion relationships; The evidence gap graph generation module generates missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary sampling action nodes based on the event confusion relationship. The evidence gap map generation module calculates the evidence influence weight of the evidence influence edge, and connects the missing evidence nodes with the supplementary sampling action nodes according to the evidence influence weight to generate the evidence gap map. Based on the missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary sampling action nodes in the evidence gap diagram, an active supplementary sampling control instruction is generated.

[0012] Optionally, the step of controlling the security robot to perform supplementary sampling actions according to the active supplementary sampling control command, generating supplementary sampling patrol observation data, and generating a robot closed-loop control strategy based on the supplementary sampling patrol observation data specifically includes: Based on the supplementary mining action type, supplementary mining target object, supplementary mining spatial location, supplementary mining sensor type, supplementary mining duration, and supplementary mining data write-back object in the active supplementary mining control command, a supplementary mining execution task is generated; Based on the task of supplementary mining, the security robot is controlled to perform supplementary mining actions, continuously recording the movement direction, dwelling state, avoidance actions and response status of the target object, and generating supplementary mining patrol observation data; The supplementary patrol observation data is synchronized in time, aligned spatially, and supplemented with data source markings to generate supplementary feedback data units. Based on the supplementary data collection feedback unit, a supplementary data collection feedback event is generated, and the supplementary data collection feedback event is written back to the security event stream to generate an updated security event stream; Update the dynamic security scenario diagram with security semantic attributes based on the supplementary data collection feedback events, and generate the updated dynamic security scenario diagram; The updated security event flow and the updated dynamic security scenario graph are input into the improved dynamic graph representation network to generate an updated security relationship attention representation; Based on the updated security event flow, the updated dynamic security scenario diagram, the updated security node memory representation, the updated security relationship attention representation, the event confusion relationship, and the supplementary feedback events, a risk handling result is generated, and a robot closed-loop control strategy is generated based on the risk handling result.

[0013] Optional, a deep learning-based robot security control system includes the following modules: The patrol data collection module is used to collect information in the patrol area by security robots, preprocess the collected information to generate patrol observation datasets, and generate security event streams based on the patrol observation datasets. The event generation module is used to construct a dynamic security scenario graph based on the security event flow, configure node status information for graph nodes, configure relationship persistence information and relationship change information for graph edges, and generate a dynamic security scenario graph with security semantic attributes. The context graph building module is used to input dynamic security context graphs with security semantic attributes and security event flows into the improved dynamic graph representation network. Through the security event flow input module, security node memory update module, and security relationship attention aggregation module, it generates security node memory representations and security relationship attention representations. The graph representation module is used to model the risk evolution intensity in the network through an improved dynamic graph representation module. It generates candidate results for security events based on the security node memory representation, the security relationship attention representation, and the security event flow. The disambiguation and supplementary acquisition module is used to generate event confusion relationships based on the candidate results of security events through the event confusion disambiguation module, generate an evidence gap map based on the event confusion relationships through the evidence gap map generation module, and generate active supplementary acquisition control instructions based on the evidence influence edges between missing evidence nodes and candidate event nodes. The closed-loop control module is used to control the security robot to perform supplementary sampling actions according to the active supplementary sampling control command, generate supplementary sampling patrol observation data, and generate robot closed-loop control strategy based on the supplementary sampling patrol observation data.

[0014] Optionally, a deep learning-based robot security control medium is provided, wherein the medium stores executable instructions that are read and executed by a computer processor.

[0015] The beneficial effects of this invention are: This invention employs an improved dynamic graph representation network, transforming the traditional handling of ordinary node interaction events in dynamic graph representation networks into a security event flow processing method tailored to robot patrol scenarios. Through security node memory updates and security relationship attention aggregation, it continuously represents the state changes of personnel targets, robots, access control, key control areas, abnormal sound sources, and obstructed areas. This enables the network to generate risk evolution states by combining access control approach, area entry, target dwell, sound source association, obstruction impact, robot reachability, and camera coverage relationships, thereby improving the continuity of risk assessment and scenario adaptability in complex security scenarios.

[0016] This invention utilizes an event confusion and disambiguation module to jointly analyze the confidence distribution of candidate events, nodes in the event association graph, edges in the event association graph, risk evolution status, memory representation of security nodes, and attention representation of security relationships in the candidate results of security events. This analysis identifies the causes of event confusion, such as insufficient target image angle, insufficient target trajectory continuity, insufficient access control side view, insufficient abnormal sound source localization, unclear target identity status, insufficient robot observation position, insufficient fixed camera collaborative information, and insufficient target behavior feedback. This enables the system to form traceable disambiguation evidence among similar events such as normal waiting, suspicious loitering, tailgating preparation, and illegal approach to access control.

[0017] This invention uses an evidence gap map generation module to convert the cause of event confusion into missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary collection action nodes. Based on the evidence influence edges, different missing evidence are mapped to lateral observation, tracking supplementary collection, frontal observation of access control, audio enhancement of sound source direction, identity verification prompts, fixed camera coordinated turning, and target behavior feedback collection. This allows the robot's supplementary collection actions to directly serve the confirmation of security events, reducing misjudgments caused by performing alarms or tracking based solely on a single recognition result. Attached Figure Description

[0018] 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 the deep learning-based robot security control method proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved dynamic graph representation network in this invention; Figure 3 This is a schematic diagram of the structure of the deep learning-based robot security control system proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 and Figure 2 A deep learning-based robot security control method includes the following steps: Security robots collect information within patrol areas, preprocess the collected information to generate patrol observation datasets, and generate security event streams based on these datasets. A dynamic security scenario graph is constructed based on the security event flow. Node status information is configured for graph nodes, and relationship persistence information and relationship change information are configured for graph edges, generating a dynamic security scenario graph with security semantic attributes. An improved dynamic graph representation network is input with dynamic security context graphs and security event streams that have security semantic attributes. Through a security event stream input module, a security node memory update module, and a security relationship attention aggregation module, security node memory representations and security relationship attention representations are generated. An improved dynamic graph representation module for modeling the risk evolution intensity in the network generates candidate security events based on the security node memory representation, security relationship attention representation, and security event flow. The event confusion and disambiguation module generates event confusion relationships based on the candidate results of security events. The evidence gap diagram generation module generates an evidence gap diagram based on the event confusion relationships and generates an active supplementary sampling control instruction based on the evidence influence edge between the missing evidence node and the candidate event node. The security robot is controlled to perform supplementary sampling actions according to the active supplementary sampling control command, and supplementary sampling patrol observation data is generated. The robot closed-loop control strategy is generated through the supplementary sampling patrol observation data.

[0021] In this embodiment, a security robot collects information within the patrol area, preprocesses the collected information to generate a patrol observation dataset, and generates a security event stream based on the patrol observation dataset, specifically including: The security robot, equipped with image acquisition devices, audio acquisition devices, infrared detection devices, radar ranging devices, positioning devices, motion control units, and communication interfaces, collects video images, audio signals, infrared heat source information, radar distance information, robot pose information, robot motion status information, access control status information, fixed camera coverage information, and area map information within the patrol area. It also records the acquisition time, acquisition device type, acquisition direction, acquisition location, and data access source for each piece of information, generating an original multimodal acquisition sequence. Based on the acquisition time, pose update time corresponding to robot pose information, motion state update time corresponding to robot motion state information, state update time corresponding to access control status information, and coverage state update time corresponding to fixed camera coverage information in the original multimodal acquisition sequence, the video images, audio signals, infrared heat source information, radar distance information, access control status information, and fixed camera coverage information are time-merged to ensure that the multiple acquisition information within the same patrol period corresponds to the pose changes and motion state changes of the security robot, thus generating a time-synchronized acquisition sequence; Based on the regional map information and robot pose information, the current position, movement direction, observation direction and movable area of ​​the security robot in the patrol area are determined. The video image acquisition field of view, audio signal acquisition direction, infrared heat source position, radar distance measurement position, access control position and fixed camera coverage area in the time synchronization acquisition sequence are uniformly mapped to the spatial coordinate range corresponding to the patrol area to generate a spatially aligned acquisition sequence. Based on the acquisition device type, acquisition location, acquisition direction, acquisition time, data access source, robot pose information, and robot motion state information in the spatial alignment acquisition sequence, data source markers are configured for video images, audio signals, infrared heat source information, radar distance information, access control status information, fixed camera coverage information, and area map information, and a source marker acquisition sequence is generated. Based on the continuity of collection time, spatial location correspondence, data source markers, and robot motion state information in the source marker collection sequence, video images, audio signals, infrared heat source information, radar distance information, access control status information, fixed camera coverage information, and area map information of the same patrol time period, the same spatial range, and the same robot motion state are grouped into patrol observation segments to generate a patrol observation segment set. Based on the acquisition time, spatial coordinate range, data source marker, robot pose information, and robot motion state information corresponding to each patrol observation segment in the patrol observation segment set, the patrol observation segments are organized sequentially, and the correspondence between each patrol observation segment and the extraction process of personnel targets, vehicle targets, abandoned objects, abnormal sound sources, heat source changes, access control status, robot position, robot observation direction, occluded areas, and robot movable areas is preserved to generate a patrol observation dataset. Based on the video images, audio signals, infrared heat source information, radar distance information, robot pose information, robot motion state information, access control state information, fixed camera coverage information, area map information, acquisition time, spatial coordinate range and data source markers corresponding to each patrol observation segment in the patrol observation dataset, and according to the acquisition time order and spatial location correspondence of each patrol observation segment in the patrol area, the patrol observation segment analysis results are generated. Based on the video images, infrared heat source information, and radar distance information in the analysis results of patrol observation segments, target extraction is performed on movable and stationary objects appearing in the patrol area, identifying personnel targets, vehicle targets, and abandoned object targets. Each personnel target, vehicle target, and abandoned object target is associated with the corresponding acquisition time, spatial coordinate range, target appearance area, target heat source response, target distance change, and data source mark to generate security target extraction results. Based on the audio signals, infrared heat source information, access control status information, robot pose information, robot motion status information, fixed camera coverage information, and area map information in the analysis results of patrol observation segments, abnormal sound sources, heat source changes, access control status, robot position, robot observation direction, obstructed areas, and robot movable areas are extracted. The abnormal sound sources, heat source changes, access control status, robot position, robot observation direction, obstructed areas, and robot movable areas are then associated with the corresponding acquisition time, spatial coordinate range, and data source markers to generate security environment extraction results. Based on the security target extraction results, security environment extraction results, and regional map information, the spatial relationships of personnel targets, vehicle targets, and abandoned objects relative to key control areas, ordinary passages, access control, abnormal sound sources, fixed camera coverage areas, obstructed areas, robot positions, and robot movable areas are determined. Based on the changes in the spatial position, movement direction, dwell state, distance from the robot position, and angle with the robot's observation direction of the same target in adjacent patrol observation segments, security relationship change results are generated. Based on the changes in security relationships, newly emerging personnel, vehicle, or abandoned targets are recorded as target emergence relationships; relationships in which personnel, vehicle, or abandoned targets move towards key control areas, ordinary passages, access control areas, or areas with abnormal sound sources are recorded as target approaching area relationships; relationships in which personnel, vehicle, or abandoned targets maintain a continuous positional association within the same spatial range are recorded as target staying relationships; relationships in which different personnel or vehicle targets maintain the same direction of movement and distance association in adjacent patrol observation segments are recorded as target following relationships; and relationships in which personnel or vehicle targets change their movement paths around obstructed areas, access control areas, or the edge of the robot's observation range are recorded as target detour relationships, thus generating target behavior relationship results. Based on the changes in security relationships, the relationships between personnel, vehicles, or abandoned objects entering key control areas, ordinary passages, or areas covered by fixed cameras are recorded as area entry relationships; the changes in spatial distance and orientation between personnel, vehicles, or abandoned objects and access control are recorded as access control proximity relationships; the spatial correspondence between abnormal sound sources and personnel, vehicles, abandoned objects, access control, or key control areas is recorded as sound source association relationships; the relationship between obstructed areas and the robot's observation direction, the area covered by fixed cameras, or the appearance area of ​​the target is recorded as obstruction impact relationships; the connectivity between the robot's position and the robot's movable area is recorded as robot reachability relationships; and the coverage relationship between the area covered by fixed cameras and personnel, vehicles, abandoned objects, access control, or key control areas is recorded as camera coverage relationships, generating security facility relationship results. Based on the target behavior relationship results and the security facility relationship results, security event records are generated for target appearance, target approaching area, target staying, target following, target detouring, area entry, access control approach, sound source association, occlusion impact, robot reachability, and camera coverage, respectively. Each security event record is written with the occurrence time, target object, associated area, data source, robot pose, relationship type, and continuous state, generating a set of security event records. Based on the occurrence time, target object, associated area, relationship type, duration, and data source of each security event record in the security event record set, consecutive security event records belonging to the same target object, the same associated area, and the same relationship type are sequentially linked, and the source relationship between each security event record and the corresponding patrol observation segment in the patrol observation dataset is preserved to generate a security event stream.

[0022] In this embodiment, a dynamic security scenario graph is constructed based on the security event flow. Node state information is configured for graph nodes, and relationship persistence and change information are configured for graph edges. Specifically, generating a dynamic security scenario graph with security semantic attributes includes: Read the occurrence time, target object, associated area, data source, robot pose, relationship type and continuous state of each security event record in the security event stream, and according to the source relationship between each security event record and the corresponding patrol observation segment in the patrol observation dataset, merge security event records belonging to the same patrol time period, the same spatial range and the same target object into graph construction source records, and generate a graph construction source record set; Based on the target objects, associated regions, robot poses, data sources, and relationship types in the graph source record set, identify personnel targets, vehicle targets, abandoned object targets, robots, access control, key control areas, ordinary passages, abnormal sound sources, fixed camera coverage areas, and obstructed areas. Then, determine the identified personnel targets, vehicle targets, abandoned object targets, robots, access control, key control areas, ordinary passages, abnormal sound sources, fixed camera coverage areas, and obstructed areas as graph nodes of the dynamic security scenario graph, and generate a graph node set. Based on the node type and source information corresponding to each graph node in the graph node set, the following configurations are made for personnel targets, vehicle targets, and abandoned objects: target category, target spatial location, target movement direction, target dwelling status, target appearance area, target heat source response, target distance change, and target data source. For robots, the following configurations are made for robot position, robot observation direction, robot movement status, and robot movable area. For access control, the following configurations are made for access control: access control position and access control status. For key control areas and ordinary passages, the following configurations are made for area position and area boundaries. For abnormal sound sources, the following configurations are made for sound source position and sound source data source. For fixed camera coverage areas, the following configurations are made for coverage position and coverage range. For occluded areas, the following configurations are made for occlusion position, occlusion boundary, and occlusion source. The configuration results are then bound to the corresponding occurrence time, spatial coordinate range, data source marker, and patrol observation segment to generate a graph node set with node status information. Based on the relationship type, target object, associated region, and robot pose in the graph source record set, the starting and ending graph nodes corresponding to each relationship are determined in the graph node set with node state information. Target approach, target stay, target follow, target detour, area entry, access control approach, sound source association, occlusion impact, robot reachability, and camera coverage are respectively constructed as graph edges connecting the corresponding graph nodes. Based on the occurrence time, duration, data source, and patrol observation segment source relationship corresponding to each graph edge, the relationship start time, relationship duration, participating graph nodes, relationship data source, and corresponding patrol observation segment are configured for each graph edge, generating a graph edge set with relationship duration information. Based on the spatial location changes, movement direction changes, dwell state changes, distance changes, occlusion coverage changes, camera coverage changes, and robot reachability changes of the same target object, the same associated area, and the same relationship type in continuous security event records, relationship change information is configured for each graph edge in the graph edge set with relationship persistence information. The relationship change information includes the change states corresponding to relationship enhancement, relationship weakening, relationship maintenance, relationship transfer, and relationship termination, and a graph edge set with relationship persistence information and relationship change information is generated. Based on the set of graph nodes with node status information and the set of graph edges with relationship persistence and relationship change information, interconnected graph nodes and edges within the same occurrence time range are organized into security scenario graph units. The security scenario graph units are then associated with patrol observation datasets, security event flows, data source markers, and regional map information. Based on the occurrence time, spatial coordinate range, node status information, relationship persistence information, and relationship change information of each security scenario graph unit, the units are organized in chronological order and connected spatially to generate a dynamic security scenario graph with security semantic attributes.

[0023] In this embodiment, the dynamic security scenario graph with security semantic attributes and the improved dynamic graph representation network for security event flow input are used to generate security node memory representation and security relationship attention representation through a security event flow input module, a security node memory update module, and a security relationship attention aggregation module. Specifically, this includes: The improved dynamic graph representation network inputs a dynamic security scenario graph with security semantic attributes and a security event flow. It reads the graph node set with node state information, the graph edge set with relationship persistence information and relationship change information, the occurrence time, spatial coordinate range, data source marker, and regional map information corresponding to each security scenario graph unit, and matches each security event record in the security event flow with the corresponding graph nodes and graph edges according to the occurrence time, target object, associated region, relationship type, and persistence status to generate network input event units. The network input event units include occurrence time, target object, associated region, data source, robot pose, relationship type, persistence status, participating graph nodes, and participating graph edges. The security event flow input module performs event vectorization processing on network input event units. It encodes the target object, associated region, relationship type, duration, participating graph nodes, and participating graph edges to obtain a field encoding vector. It encodes the occurrence time, relationship start time, and relationship duration to obtain a time position encoding vector. It encodes the data source to obtain a source encoding vector. It encodes the robot position, robot observation direction, and robot motion state in the robot pose to obtain a pose encoding vector. Finally, according to the correspondence of the same security event record, the field encoding vector, time position encoding vector, source encoding vector, and pose encoding vector are sequentially concatenated to generate a fixed-length event input vector. The security node memory update module reads the historical node memories corresponding to personnel targets, robots, access control, key control areas, abnormal sound sources, and obstructed areas involved in the network input event unit. The target graph node memory before the network input event unit is written is used as the target graph node historical node memory, and the adjacent graph node memory that has graph edge connection with the target graph node is used as the adjacent graph node historical node memory. Based on the event input vector, the target graph node historical node memory, the adjacent graph node historical memory, and the relation type embedding vector of the graph edge corresponding to the network input event unit, an event message vector is generated. The generation process of the event message vector includes: performing event feature mapping on the event input vector, performing target node state mapping on the target graph node historical node memory, performing adjacent node state mapping on the adjacent graph node historical memory, performing relation semantic mapping on the relation type embedding vector, superimposing the above mapping results with the message bias term, and generating the event message vector through nonlinear activation processing. The security node memory update module updates the memory of personnel, robots, access control, key control areas, abnormal sound sources, and obstructed areas involved in the current security event by combining and mapping the event message vector and the historical node memory of the target graph nodes, and then performing non-linear activation processing to obtain the node memory update gating vector. The event message vector and the historical node memory of the target graph node are combined and mapped in another way, and then processed by nonlinear activation to obtain the node memory retention gating vector. After multiplying the node memory retention gating vector with the target graph node historical node memory at the corresponding positions, and then performing candidate memory mapping together with the event message vector, the candidate node memory is obtained. Based on the updated gating vector of node memory, the historical node memory and candidate node memory of the target graph node are proportionally fused to generate the target graph node memory after the current security event is written. The updated personnel target node memory, robot node memory, access control node memory, key control area node memory, abnormal sound source node memory and obstruction area node memory are combined to form a security node memory representation. The security relationship attention aggregation module performs relational semantic encoding on the graph edges connecting the target graph node and adjacent graph nodes. It encodes the relationship types corresponding to access control approach, area entry, target stay, sound source association, occlusion effect, robot reachability, and camera coverage to obtain relationship type embedding vectors. It encodes the relationship persistence information and relationship change information corresponding to relationship start time, relationship maintenance time, relationship enhancement, relationship weakening, relationship maintenance, relationship transfer, and relationship dissolution to obtain relationship change vectors. It concatenates and maps the relationship type embedding vectors and relationship change vectors to generate relationship semantic vectors. It calculates the relationship attention score of adjacent graph nodes to the target graph node based on the target graph node memory, adjacent graph node memory, and relationship semantic vector. The process of calculating the relational attention score includes: performing query feature mapping on the target graph node memory, performing key feature mapping on the adjacent graph node memory, performing relational key feature mapping on the relational semantic vector, fusing the key feature mapping results of the adjacent graph node memory with the relational key feature mapping results of the relational semantic vector, calculating the similarity between the fused result and the query feature mapping result of the target graph node memory to obtain the relational attention score, and then normalizing the relational attention scores corresponding to each adjacent graph node connected to the same target graph node to generate the relational weights of each adjacent graph node to the target graph node; The security relationship attention aggregation module aggregates neighborhood information of target graph nodes based on relationship weights, neighboring graph node memories, and relationship semantic vectors. The aggregation process includes: output feature mapping for each neighboring graph node memory, output relationship mapping for the relationship semantic vector between each neighboring graph node and the target graph node, fusing the output feature mapping results and output relationship mapping results corresponding to the same neighboring graph node, and weighting and accumulating the fusion results according to the corresponding relationship weights to obtain the security relationship attention representation of the target graph node.

[0024] In this embodiment, the risk evolution intensity modeling module in the improved dynamic graph representation network generates security event candidate results based on the security node memory representation, security relationship attention representation, and security event flow. Specifically, this includes: The risk evolution intensity modeling module reads the security node memory representation, security relationship attention representation, and security event flow. Based on the occurrence time of each security event record in the security event flow, the security node memory representation, security relationship attention representation, relationship type, continuous state, participating graph nodes, and participating graph edges of the same target object in continuous patrol observation segments are sorted out to generate the target risk evolution input sequence. Based on the target risk evolution input sequence, the node memory changes, relation attention changes, and relation persistence state changes of the same target object between the current security event record and the previous security event record are calculated to generate a risk evolution difference representation; The process of generating the risk evolution difference representation includes: calculating the corresponding positional difference between the security node memory representation corresponding to the current security event record and the security node memory representation corresponding to the previous security event record to obtain the node memory difference representation; calculating the corresponding positional difference between the security relationship attention representation corresponding to the current security event record and the security relationship attention representation corresponding to the previous security event record to obtain the relationship attention difference representation; and sequentially combining the event input vector corresponding to the current security event record, the security node memory representation corresponding to the current security event record, the security relationship attention representation corresponding to the current security event record, the node memory difference representation, the relationship attention difference representation, and the relationship persistence state to generate the risk evolution difference representation. The risk evolution intensity modeling module calculates the risk intensity of the risk evolution difference representation, generating risk evolution intensity values ​​for the target object under different candidate event types. The risk intensity calculation process includes: mapping the risk evolution difference representation to risk features, superimposing the mapping results with risk feature bias terms, and generating a risk implicit representation through nonlinear transformation; for the candidate event types corresponding to normal passage, normal waiting, suspicious loitering, tailgating preparation, illegal approach to access control, abnormal abandoned objects, abnormal sounds, personnel falling, abnormal smoke and fire, and intrusion into key areas, the risk implicit representation is mapped to the candidate event type, and the mapping results corresponding to each candidate event type are converted into non-negative risk evolution intensity values. Based on the risk evolution intensity value, the direction of risk change for the same target object in continuous security event records is calculated to generate the risk evolution state. The direction of risk change includes risk reduction, risk maintenance, risk increase and risk transformation. The process of generating risk evolution states includes: calculating the difference between the risk evolution intensity value of the current security event record under the same candidate event type and the risk evolution intensity value of the previous security event record under the same candidate event type to obtain the risk intensity change; comparing the candidate event type with the largest risk evolution intensity value in the current security event record with the candidate event type with the largest risk evolution intensity value in the previous security event record to obtain candidate event type transformation information; and inputting the risk intensity change, candidate event type transformation information, relationship persistence state, and relationship change information into the state normalization layer to generate risk evolution states including risk reduction state, risk maintenance state, risk increase state, and risk transformation state. The security event candidate output module calculates candidate event scores for each event type based on risk evolution status, security node memory representation, and security relationship attention representation. The calculation process includes: combining the security node memory representation, security relationship attention representation, risk implicit representation, and risk evolution intensity value corresponding to each candidate event type for the current security event record; mapping the combined result to candidate event scores; and superimposing the mapped result with the candidate event score bias term to obtain the candidate event score for each event type; and normalizing the candidate event scores for the same target object to ensure that the confidence levels of each event type form a candidate event confidence distribution within the same range. Based on the confidence distribution of candidate events, risk evolution status, security node memory representation, and security relationship attention representation, the target objects, event occurrence areas, event association graph nodes, and event association graph edges in the security event candidate results are determined. Specifically, personnel, vehicle, or abandoned objects corresponding to the target risk evolution input sequence are identified as target objects. Event occurrence areas are determined based on graph nodes that have relationships with target objects related to area entry, access control proximity, sound source association, occlusion impact, and camera coverage. Event association graph nodes are determined based on graph nodes participating in risk evolution intensity calculation, candidate event score calculation, and relationship attention aggregation. Event association graph edges are determined based on graph edges connecting event association graph nodes and participating in relationship weight allocation, corresponding to target approach, target stay, target follow, target detour, area entry, access control proximity, sound source association, occlusion impact, robot reachability, and camera coverage. The security event candidate output module combines candidate event types, target objects, event occurrence areas, candidate event confidence distribution, event association graph nodes, and event association graph edges to generate security event candidate results. It also establishes a correspondence between the security event candidate results and risk evolution states, security node memory representations, and security relationship attention representations.

[0025] In this embodiment, the event obfuscation and disambiguation module generates event obfuscation relationships based on the candidate results of security events, and the evidence gap graph generation module generates an evidence gap graph based on the event obfuscation relationships. Furthermore, based on the evidence influence edges between missing evidence nodes and candidate event nodes, an active supplementary data collection control instruction is generated, specifically including: The event obfuscation and disambiguation module reads the candidate event types, target objects, event occurrence areas, candidate event confidence distribution, event association graph nodes, and event association graph edges from the security event candidate results. It sorts the candidate event types corresponding to the same target object according to the candidate event confidence from high to low. The candidate event type with the highest confidence is paired with the candidate event type whose event occurrence area has spatial overlap and whose candidate event confidence difference is the smallest. For multiple spatially overlapping candidate event types for the same target object, candidate event pairs are formed in order of increasing candidate event confidence difference, generating a set of candidate event pairs. The event confusion and disambiguation module calculates the confusion intensity of each candidate event pair based on the candidate event pair set, candidate event confidence distribution, event association graph nodes, event association graph edges, risk evolution state, security node memory representation, and security relationship attention representation. The calculation process of confusion intensity includes: obtaining the confidence level similarity based on the difference in confidence levels between the two candidate event types in the candidate event pair; and obtaining the graph basis overlap based on the number of event association graph nodes used by the two candidate event types in the candidate event pair, the number of event association graph edges used by the two types, and the total number of event association graph nodes and event association graph edges for each of them. Based on the similarity between the security node memory representation and the security relationship attention representation corresponding to the two candidate event types in the candidate event pair, the representation similarity is obtained; based on whether there is a risk transformation relationship and the direction of risk transformation between the two candidate event types in the risk evolution state, the transformation correlation is obtained; the confidence level, graph basis overlap, representation similarity and transformation correlation are mapped to the same numerical range, and weighted and fused according to the confusion influence coefficient obtained in training to generate a set of candidate event confusion intensity; Based on the set of candidate event confusion intensity, evidence integrity analysis is performed on the corresponding event association graph nodes, event association graph edges, security node memory representation, and security relationship attention representation for each candidate event pair. The target image angle integrity, target trajectory continuity, access control side view coverage, abnormal sound source localization stability, target identity state integrity, robot observation position adaptation, fixed camera collaboration integrity, and target behavior feedback integrity are calculated respectively. Among them, the completeness of the target image angle is determined based on the proportion of the front view visible in the target appearance area and the robot's observation direction; the continuity of the target trajectory is determined based on the position connection relationship of the same target object in the continuous security event records; the access control side view coverage is determined based on the access control position, the robot's current position, the fixed camera coverage information and the occlusion area; the stability of abnormal sound source localization is determined based on the spatial change of the abnormal sound source position in the continuous patrol observation segment; the completeness of the target identity status is determined based on the target object's identity verification information and data source markers; the robot's observation position adaptability is determined based on the robot's current position, the robot's reachable area and the sensor's acquisition range; the completeness of fixed camera collaboration is determined based on the spatial coverage relationship between the fixed camera coverage information and the target object; and the completeness of target behavior feedback is determined based on the record of the target object's behavior changes after the robot approaches, provides voice prompts or changes in perspective. Items with low completeness and high contribution to the confusion intensity of candidate events are identified as the event confusion causes of the corresponding candidate event pairs, generating a set of event confusion causes. The set of candidate event pairs, the set of candidate event confusion intensity, and the set of event confusion causes are associated to generate event confusion relationships. The event confusion relationships include candidate event pairs, confusion intensity, confusion causes, associated target objects, event occurrence areas, association graph nodes, association graph edges, and risk evolution states. The event confusion relationships are then output to the evidence gap graph generation module. The evidence gap map generation module generates missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary sampling action nodes based on event confusion relationships, dynamic security scenario maps with security semantic attributes, robot current position, robot reachable area, sensor acquisition range, and area map information. Among them, missing target frontal image, missing target trajectory continuity, missing access control side view, missing abnormal sound source localization, missing target identity status, missing robot observation position, missing fixed camera collaborative information, and missing target behavior feedback are identified as missing evidence nodes. Candidate event types in event confusion relationships are identified as candidate event nodes. The association between missing evidence nodes and candidate event nodes affected by them is identified as evidence influence edges. Lateral observation supplementary sampling actions, tracking supplementary sampling actions, access control side front observation supplementary sampling actions, sound source direction audio enhancement supplementary sampling actions, identity verification prompt supplementary sampling actions, robot observation position adjustment supplementary sampling actions, fixed camera collaborative turning supplementary sampling actions, and target behavior feedback collection supplementary sampling actions are identified as supplementary sampling action nodes. The evidence gap map generation module calculates the evidence influence weight of the evidence influence edge, and connects the missing evidence nodes with the supplementary sampling action nodes according to the evidence influence weight to generate the evidence gap map. The calculation process for the evidence influence weight includes: determining the confusion contribution value based on the confusion intensity of the candidate event pair corresponding to the missing evidence node; determining the degree of evidence missing based on the evidence completeness corresponding to the missing evidence node; determining the data source reliability based on the number of data sources, data source stability, and data source consistency corresponding to the missing evidence node; determining the robot's suitability for supplementary data collection based on the robot's current location, robot reachable area, sensor acquisition range, and regional map information; mapping the confusion contribution value, degree of evidence missing, data source reliability, and robot suitability for supplementary data collection to the same numerical range, and weighting and fusing them according to the evidence influence coefficients obtained during training to obtain the evidence influence weight of the missing evidence node on the candidate event node. Based on the missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary sampling action nodes in the evidence gap diagram, generate active supplementary sampling control instructions; The process of generating active data acquisition control commands includes: mapping missing frontal images of the target to lateral observation acquisition actions; mapping missing target trajectory continuity to tracking acquisition actions; mapping missing access control side view to access control side front observation acquisition actions; mapping missing abnormal sound source localization to sound source direction audio enhancement acquisition actions; mapping missing target identity status to identity verification prompt acquisition actions; mapping missing robot observation position to robot observation position adjustment acquisition actions; mapping missing fixed camera collaborative information to fixed camera collaborative turning acquisition actions; and mapping missing target behavior feedback to target behavior feedback acquisition acquisition actions. Based on the evidence influence weight of each evidence influence edge, the robot executableness corresponding to the acquisition action node, and the movement direction of the target object, the active data acquisition control command determines the acquisition action type, acquisition target object, acquisition spatial location, acquisition sensor type, acquisition duration, and acquisition data write-back object, and generates the active data acquisition control command.

[0026] In this embodiment, the security robot is controlled to perform supplementary sampling actions according to the active supplementary sampling control command, generating supplementary sampling patrol observation data. The robot closed-loop control strategy is generated through the supplementary sampling patrol observation data, specifically including: The system reads the supplementary sampling action type, supplementary sampling target object, supplementary sampling spatial location, supplementary sampling sensor type, supplementary sampling duration, and supplementary sampling data write-back object from the active supplementary sampling control command. Based on the supplementary sampling action type, it determines the current supplementary sampling action from the following categories: side observation supplementary sampling action, tracking supplementary sampling action, access control side front observation supplementary sampling action, sound source direction audio enhancement supplementary sampling action, identity verification prompt supplementary sampling action, robot observation position adjustment supplementary sampling action, fixed camera coordinated turning supplementary sampling action, and target behavior feedback acquisition supplementary sampling action. It generates a supplementary sampling movement path based on the supplementary sampling spatial location and area map information. Based on the supplementary sampling sensor type, it determines the acquisition combination of video image acquisition, audio signal acquisition, infrared heat source acquisition, radar distance acquisition, access control status acquisition, and fixed camera coordinated acquisition, and generates a supplementary sampling execution task. The system controls the security robot to perform supplementary sampling actions according to the assigned task. When performing lateral observation supplementary sampling, the robot moves to the lateral observation position of the target object and adjusts its observation direction. When performing tracking supplementary sampling, the robot maintains the tracking distance to the target object and continuously collects the target trajectory. When performing access control side-front observation supplementary sampling, the robot moves to the access control side-front acquisition position and collects access control side view data. When performing sound source direction audio enhancement supplementary sampling, the audio acquisition device is directed towards the abnormal sound source and acquires enhanced audio signals. The system also handles identity verification prompts. During the supplementary sampling action, the security robot is controlled to output identity verification prompts to the target object and collect target response data. When the robot's observation position is adjusted for supplementary sampling, the security robot is controlled to move to an observation position that can simultaneously cover the target object, the area where the event occurred, and the associated access control location. When the fixed camera is coordinated to turn for supplementary sampling, the fixed camera is controlled to adjust its coverage direction and collect coordinated images. When the target behavior feedback collection is performed for supplementary sampling, the security robot is controlled to continuously record the target object's movement direction, dwell state, avoidance actions, and response state during the supplementary sampling duration, generating supplementary sampling patrol observation data. The supplementary patrol observation data is time-synchronized, spatially aligned, and supplemented with data source marking. The video images, audio signals, infrared heat source information, radar distance information, access control status information, fixed camera collaborative images, robot pose information, robot motion status information, and target behavior feedback data obtained from the supplementary data collection are associated with the supplementary target object, supplementary action type, supplementary spatial location, supplementary sensor type, supplementary duration, and supplementary data write-back object in the active supplementary data collection control command to generate supplementary data feedback data units. Based on the supplementary sampling feedback data unit, a supplementary sampling feedback event is generated. The supplementary sampling feedback event includes the supplementary sampling time, supplementary sampling target object, supplementary sampling action type, supplementary sampling spatial location, supplementary sampling data source, supplementary sampling robot pose, supplementary sampling relationship type, supplementary sampling continuity status, supplementary sampling participation graph nodes and supplementary sampling participation graph edges. The supplementary results of the target frontal image, target trajectory, access control side view, abnormal sound source localization, target identity status, robot observation position, fixed camera collaborative information, and target behavior feedback are written into the corresponding supplementary sampling feedback event. Write the supplementary sampling feedback event back to the security event stream. Insert the supplementary sampling feedback event into the corresponding security event record sequence according to the supplementary sampling occurrence time, supplementary sampling target object, supplementary sampling action type and supplementary sampling relationship type. Update the security event records corresponding to target appearance, target approach area, target stay, target follow, target detour, area entry, access control approach, sound source association, occlusion effect, robot reachability and camera coverage according to the supplementary sampling continuity status in the supplementary sampling feedback event, and generate the updated security event stream. The dynamic security scenario graph with security semantic attributes is updated based on the supplementary data collection feedback events. The nodes of the supplementary data collection participation graph in the supplementary data collection feedback events are matched with the set of graph nodes with node status information. The node status information corresponding to personnel targets, vehicle targets, abandoned objects targets, robots, access control, key control areas, ordinary passages, abnormal sound sources, fixed camera coverage areas and obstruction areas is supplemented and written. The edges of the supplementary data collection participation graph in the supplementary data collection feedback events are matched with the set of graph edges with relationship persistence information and relationship change information. The relationship persistence information and relationship change information corresponding to target approach, target stay, target follow, target detour, area entry, access control approach, sound source association, obstruction impact, robot reachability and camera coverage are supplemented and written, generating the updated dynamic security scenario graph. The updated security event flow and the updated dynamic security scenario graph are input into the improved dynamic graph representation network. The security node memory update module updates the security node memory representation based on changes in target spatial location, target movement direction, target dwell state, access control status, sound source location, occlusion coverage, fixed camera coverage, robot position, and robot observation direction in the supplementary feedback events, generating an updated security node memory representation. The security relationship attention aggregation module updates the security relationship attention representation based on the relationship type, relationship persistence information, and relationship change information corresponding to the supplementary feedback events, generating an updated security relationship attention representation. Based on the updated security event flow, updated dynamic security scenario diagram, updated security node memory representation, updated security relationship attention representation, event confusion relationship, and supplementary feedback events, a risk handling result is generated. The process of generating the risk handling result includes: supplementing the missing evidence nodes in the event confusion relationship with status labels based on the supplementary feedback events; adjusting the evidence influence edges corresponding to the candidate event nodes based on the supplementary status labels; regenerating the candidate event confidence distribution based on the adjusted evidence influence edges, updated security node memory representation, and updated security relationship attention representation; and determining the candidate event type with the highest confidence and consistent with the risk evolution state in the candidate event confidence distribution as the confirmed event type. A risk handling result is generated that includes the confirmed event type, target object, event occurrence area, risk level, handling priority, handling basis graph nodes, and handling basis graph edges. Based on the risk mitigation results, a robot closed-loop control strategy is generated. The generation process includes: when the risk mitigation result indicates that the risk has been eliminated, a continued patrol control strategy is generated; when the risk mitigation result indicates that the risk remains, a continuous observation, target tracking, and supplementary recording control strategy is generated; when the risk mitigation result indicates that the risk has increased, a voice alert, background alarm, access control linkage, fixed camera linkage, and sound and light alarm linkage control strategy is generated; when the risk mitigation result indicates that the risk has transformed, a target tracking, detour observation, patrol route adjustment, and patrol task rearrangement control strategy is regenerated based on the confirmed event type, and a correspondence is established between the robot closed-loop control strategy and the updated security event flow, the updated dynamic security scenario map, and the risk mitigation result.

[0027] In this embodiment, the improved dynamic graph representation network includes a security event flow input module, a security node memory update module, a security relationship attention aggregation module, a risk evolution intensity modeling module, a security event candidate output module, an event confusion disambiguation module, and an evidence gap graph generation module. Based on traditional dynamic graph representation networks, this network improves upon the traditional approach of handling only ordinary node interaction events by processing security event flows in robot patrol scenarios. This allows the network to receive information such as occurrence time, target object, associated region, data source, robot pose, relationship type, persistent state, participating graph nodes, and participating graph edges, and to generate security node memory representations, security relationship attention representations, risk evolution states, security event candidate results, event confusion relationships, and evidence gap graphs.

[0028] The training data comes from historical patrol observation datasets, historical security event streams, and historical dynamic security scenario maps collected by security robots in different patrol areas, patrol times, and security scenarios. Training samples are labeled using a combination of machine initial labeling and manual verification. Initial labeling results are generated first through target detection, target tracking, sound source localization, access control status reading, and robot localization. Verification is then performed using video images, audio signals, infrared thermal source information, radar distance information, access control status information, fixed camera collaborative footage, and area map information. Labeling content includes graph node types, graph edge relationship types, relationship change states, real security event types, candidate event types, reasons for event confusion, missing evidence types, evidence influence relationships, and supplementary data collection action types.

[0029] During training, the historical patrol observation dataset is segmented into continuous patrol observation segments, generating historical security event streams and historical dynamic security scenario graphs. The security event stream input module converts each security event record into an event input vector; the security node memory update module generates a security node memory representation based on the event input vector and historical node memory; the security relationship attention aggregation module generates a security relationship attention representation based on graph edge relationship type, relationship persistence information, and relationship change information; the risk evolution intensity modeling module generates risk evolution states; the security event candidate output module generates security event candidate results; the event confusion and disambiguation module generates event confusion relationships; and the evidence gap graph generation module generates an evidence gap graph.

[0030] In this implementation, the event input vector, security node memory representation, and security relationship attention representation all use 256-dimensional vectors, while the relationship type embedding vector and relationship change vector both use 64-dimensional vectors. The security relationship attention aggregation module uses four attention heads to learn the impact of access control approach, area entry, target dwell, sound source association, occlusion effect, robot reachability, and camera coverage on security event judgment. Training employs an adaptive moment estimation optimizer with an initial learning rate of 0.0001, a weight decay coefficient of 0.00001, a batch training sample size of 32 sets of continuous patrol observation segments, a maximum training epoch of 120 epochs, and a random deactivation rate of 0.1.

[0031] The loss function adopts a multi-task joint loss. The total loss consists of node state restoration loss, relation type identification loss, relation change prediction loss, risk evolution state loss, candidate event classification loss, event confusion cause loss, evidence gap graph loss, and evidence influence edge weight loss, which are respectively used to constrain node state restoration, graph edge relation identification, relation change judgment, risk evolution identification, candidate event confidence distribution, event confusion cause identification, evidence gap graph generation, and evidence influence edge weight calculation.

[0032] During training, the security event flow input module, security node memory update module, and security relationship attention aggregation module are trained first. Then, the risk evolution intensity modeling module and security event candidate output module are trained. Next, the event confusion disambiguation module and evidence gap graph generation module are trained. Finally, the entire improved dynamic graph representation network is jointly fine-tuned. Training is stopped and network parameters are saved when the validation set loss does not decrease effectively after 10 consecutive training rounds, and the accuracy rates of candidate event identification, event confusion cause identification, evidence gap graph matching, and risk evolution state identification remain stable.

[0033] Through the above training method, the improvements of the improved dynamic graph representation network compared with the traditional dynamic graph representation network are as follows: the ordinary node interaction events are improved into security event flow inputs, the ordinary node memory is improved into security node memory, the ordinary neighborhood aggregation is improved into security relationship attention aggregation, and risk evolution, event confusion disambiguation and evidence gap graph generation processing are added, enabling the network to identify the cause of candidate event confusion from robot patrol data and provide evidence gap basis for proactively supplementing control commands.

[0034] refer to Figure 3 The deep learning-based robot security control system includes the following modules: The patrol data collection module is used to collect information in the patrol area by security robots, preprocess the collected information to generate patrol observation datasets, and generate security event streams based on the patrol observation datasets. The event generation module is used to construct a dynamic security scenario graph based on the security event flow, configure node status information for graph nodes, configure relationship persistence information and relationship change information for graph edges, and generate a dynamic security scenario graph with security semantic attributes. The context graph building module is used to input dynamic security context graphs with security semantic attributes and security event flows into the improved dynamic graph representation network. Through the security event flow input module, security node memory update module, and security relationship attention aggregation module, it generates security node memory representations and security relationship attention representations. The graph representation module is used to model the risk evolution intensity in the network through an improved dynamic graph representation module. It generates candidate results for security events based on the security node memory representation, the security relationship attention representation, and the security event flow. The disambiguation and supplementary acquisition module is used to generate event confusion relationships based on the candidate results of security events through the event confusion disambiguation module, generate an evidence gap map based on the event confusion relationships through the evidence gap map generation module, and generate active supplementary acquisition control instructions based on the evidence influence edges between missing evidence nodes and candidate event nodes. The closed-loop control module is used to control the security robot to perform supplementary sampling actions according to the active supplementary sampling control command, generate supplementary sampling patrol observation data, and generate robot closed-loop control strategy based on the supplementary sampling patrol observation data.

[0035] In this embodiment, a robot security control medium based on deep learning is provided. The medium stores executable instructions that can be read by a computer processor.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to a nighttime robotic security patrol scenario in a closed industrial park. The park includes office building entrances, warehouse side doors, equipment room passageways, parking areas, and fenced entrances / exits. The patrol robot performs its patrol tasks in low-light conditions at night, with limited personnel movement, partial obstructions, and incomplete coverage by fixed cameras. Existing patrol methods primarily rely on robot cameras to identify personnel targets and directly trigger an alarm when personnel linger for an extended period or approach the access control. This can easily confuse normal waiting, overtime passage, equipment maintenance approach, and suspicious loitering, especially when there are shelves obstructing the access control, the target's back is to the robot, or the source of abnormal noise is unclear, leading to numerous false alarms and the need for secondary manual verification.

[0037] In this scenario, the security robot collects video images, audio signals, infrared thermal source information, radar distance information, robot pose information, access control status information, fixed camera coverage information, and area map information along its patrol route. The system synchronizes and aligns the collected information temporally and spatially to form a patrol observation dataset. Based on this dataset, the system generates a security event stream, recording events such as personnel approaching access control, target lingering, target detours, sound source association, occlusion effects, and camera coverage relationships as continuous security events. It also constructs a dynamic security context graph with security semantic attributes. An improved dynamic graph representation network generates candidate security events based on security node memory representations and security relationship attention representations. When the system detects that the same target simultaneously exhibits candidate events such as "normal waiting," "suspicious loitering," and "illegal approach to access control," the event confusion and disambiguation module further analyzes the target image angle, target trajectory continuity, access control side view, abnormal sound source localization, and fixed camera collaborative information to generate event confusion relationships. Subsequently, the evidence gap map generation module transforms the missing side view of the access control system, the lack of continuity in the target trajectory, and the lack of a frontal image of the target into supplementary data acquisition action nodes. This controls the robot to move to the observation position in front of the access control system, maintaining a safe distance to track the target, and requests the fixed camera to coordinate its turning. After completing the supplementary data acquisition, the robot writes the supplementary patrol observation data back to the security event stream and dynamic security scenario map. The system then regenerates the risk handling results and generates corresponding closed-loop control strategies for the robot based on whether the risk is eliminated, maintained, escalated, or transformed.

[0038] This embodiment was validated over 30 consecutive night patrol shifts, completing a total of 1248 patrol cycles, collecting 18640 valid patrol observation segments, identifying 1782 personnel targets, 436 vehicle targets, and 57 abandoned objects, and generating 9236 security incident records. Among these, 126 abnormal events were manually verified, including 41 suspicious loitering incidents, 28 unauthorized approaching of access control points, 17 tailgating attempts, 13 abnormal abandoned objects, 18 abnormal sounds, and 9 intrusions into key areas. During testing, compared to the traditional patrol method of "target recognition plus fixed rule alarm," this invention can proactively collect key evidence when there is confusion in the candidate event results. This increases the accuracy of candidate event confirmation from 82.4% to 93.6%, reduces the false alarm rate from 14.8% to 5.1%, the missed alarm rate from 8.7% to 3.2%, shortens the average event confirmation time from 43.5 seconds to 27.8 seconds, and reduces the number of manual reviews per shift from an average of 9.6 to 3.4. These results demonstrate that this invention can improve the stability of event recognition in complex security scenarios and enhance the targeted nature of robot handling and control.

[0039] Table 1 Comparison of Robot Security Control Effects As can be seen from the table above, the present invention significantly improves upon traditional patrol control methods in several key performance indicators of robot security control. With consistent evaluation benchmarks, both methods were tested based on 30 patrol shifts, 18,640 valid patrol observation segments, 9,236 security event records, and 126 verified abnormal events; therefore, the comparison results are highly comparable. Regarding the accuracy rate of candidate event confirmation, the traditional patrol control method achieves 82.4%, while the present invention reaches 93.6%, an improvement of 11.2 percentage points. This demonstrates that the present invention, through security event flow, dynamic security scenario diagrams, and an improved dynamic graph representation network, can more fully utilize the relationships between personnel targets, access control, abnormal sound sources, obstructed areas, and fixed camera coverage to more accurately distinguish easily confused events such as normal waiting, suspicious loitering, tailgating preparation, and unauthorized approach to access control.

[0040] Regarding false alarm rate and false negative rate, the false alarm rate of traditional patrol control method is 14.8% and the false negative rate is 8.7%, while the method of this invention reduces them to 5.1% and 3.2% respectively. This indicates that the present invention, through event confusion disambiguation and evidence gap map generation mechanism, can first determine the source of missing evidence when the recognition result is unstable, and then control the robot to perform supplementary collection actions such as lateral observation, tracking supplementary collection, front observation of the access control side, audio enhancement of the sound source direction, and fixed camera coordinated turning, thereby reducing false alarms and abnormal omissions caused by insufficient recognition results in a single time.

[0041] In terms of response efficiency, the average event confirmation time of the traditional patrol control method is 43.5 seconds, while the method of this invention shortens it to 27.8 seconds, a reduction of 15.7 seconds. The number of manual reviews per shift is reduced from 9.6 to 3.4, indicating that this invention can transform the cause of confusion in candidate events into executable proactive supplementary data collection control instructions, enabling the robot to autonomously supplement key evidence on-site and reducing the workload of repeated manual verification in the background. Regarding the event confirmation stability rate after supplementary data collection, the method of this invention reaches 92.8%, higher than the 76.9% of the traditional method, indicating that after the supplementary data collection feedback is written back to the security event stream, dynamic security scenario diagram, security node memory representation, and security relationship attention representation, the continuity of risk assessment can be effectively improved. The robot closed-loop control strategy matching rate increases from 79.5% to 91.7%, indicating that the control strategies generated by this invention, such as continued patrol, target tracking, voice alerts, background alarms, access control linkage, and patrol task rescheduling, have a higher degree of matching with the actual risk state. Overall, this invention can improve the accuracy of abnormal event identification in robot security patrols, reduce false alarms and missed alarms, and enhance the stability and timeliness of closed-loop handling.

[0042] The above are merely preferred embodiments 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 robot security control method based on deep learning, characterized by, Includes the following steps: Security robots collect information within patrol areas, preprocess the collected information to generate patrol observation datasets, and generate security event streams based on these datasets. A dynamic security scenario graph is constructed based on the security event flow. Node status information is configured for graph nodes, and relationship persistence and change information are configured for graph edges to generate a dynamic security scenario graph with security semantic attributes. An improved dynamic graph representation network is input with dynamic security context graphs and security event streams that have security semantic attributes. Through a security event stream input module, a security node memory update module, and a security relationship attention aggregation module, security node memory representations and security relationship attention representations are generated. An improved dynamic graph representation module for modeling the risk evolution intensity in the network generates candidate security events based on the security node memory representation, security relationship attention representation, and security event flow. The event confusion and disambiguation module generates event confusion relationships based on the candidate results of security events. The evidence gap diagram generation module generates an evidence gap diagram based on the event confusion relationships and generates an active supplementary sampling control instruction based on the evidence influence edge between the missing evidence node and the candidate event node. The security robot is controlled to perform supplementary sampling actions according to the active supplementary sampling control command, and supplementary sampling patrol observation data is generated. The robot closed-loop control strategy is generated through the supplementary sampling patrol observation data. 2.The deep learning-based robot security control method of claim 1, wherein, The preprocessing includes time synchronization, spatial alignment, and data source marking. 3.The deep learning-based robot security control method of claim 1, wherein, The process of constructing a dynamic security scenario graph based on the security event flow, configuring node state information for graph nodes, configuring relationship persistence information and relationship change information for graph edges, and generating a dynamic security scenario graph with security semantic attributes specifically includes: Read each security event record from the security event stream and generate a graph to construct the source record set; Based on the source record set of the graph, determine the graph nodes of the dynamic security scenario graph and generate a graph node set; Based on the node type and source information of each graph node in the graph node set, generate a graph node set with node status information; Based on the relationship types, target objects, associated regions, and robot poses in the graph source record set, generate a graph edge set with relationship persistence information; Configure relation change information for each graph edge in the graph edge set with relation persistence information, and generate a graph edge set with both relation persistence and relation change information; Based on the graph node set with node state information and the graph edge set with relationship persistence information and relationship change information, a dynamic security scenario graph with security semantic attributes is generated. 4.The deep learning-based robot security control method of claim 1, wherein The improved dynamic graph representation network, which inputs dynamic security scenario graphs and security event streams with security semantic attributes, generates security node memory representations and security relationship attention representations through a security event stream input module, a security node memory update module, and a security relationship attention aggregation module. Specifically, this includes: The network is represented by an improved dynamic graph representation of security scenarios with security semantic attributes and security event flow inputs, generating network input event units. The security event stream input module performs event vectorization on the network input event units to generate fixed-length event input vectors. The security node memory update module reads the historical node memory of the network input event unit and generates an event message vector. The security node memory update module performs gating updates on the node memory based on the event message vector and the historical node memory of the target graph node to obtain the security node memory representation; The security relationship attention aggregation module performs relational semantic encoding on the graph edges connecting the target graph node and its adjacent graph nodes to obtain the relation type embedding vector. Relationship persistence information and relationship change information are encoded to obtain a relationship change vector. The relationship type embedding vector and the relationship change vector are concatenated and mapped to generate a relationship semantic vector. And calculate the relational attention score of the neighboring graph node to the target graph node based on the target graph node memory, the neighboring graph node memory, and the relational semantic vector; The relational attention scores of each adjacent graph node connected to the same target graph node are normalized to generate the relational weights of each adjacent graph node to the target graph node. The security relationship attention aggregation module aggregates neighborhood information of target graph nodes based on relationship weights, adjacent graph node memories, and relationship semantic vectors to obtain a security relationship attention representation.

5. The robot security control method based on deep learning according to claim 1, characterized in that, The risk evolution intensity modeling module in the improved dynamic graph representation network generates candidate security event results based on the security node memory representation, security relationship attention representation, and security event flow. Specifically, this includes: The target risk evolution input sequence is generated by reading the security node memory representation, security relationship attention representation and security event flow through the risk evolution intensity modeling module. Based on the target risk evolution input sequence, the node memory changes, relational attention changes, and relational persistence state changes of the same target object between the current security event record and the previous security event record are calculated to generate a risk evolution difference representation. The risk evolution intensity modeling module calculates the risk intensity of the risk evolution difference representation and generates the risk evolution intensity value of the target object under different candidate event types. Based on the risk evolution intensity value, the direction of risk change for the same target object in continuous security event records is calculated to generate the risk evolution state; The security event candidate output module calculates the candidate event score for each candidate event type based on the risk evolution status, security node memory representation, and security relationship attention representation. The scores of each candidate event corresponding to the same target object are normalized to form a confidence distribution of candidate events within the same distribution range; Based on the confidence distribution of candidate events, risk evolution status, security node memory representation, and security relationship attention representation, the target object, event occurrence area, event association graph node, and event association graph edge in the security event candidate results are determined. The security event candidate output module combines candidate event types, target objects, event occurrence areas, candidate event confidence distribution, event association graph nodes, and event association graph edges to generate security event candidate results.

6. The robot security control method based on deep learning according to claim 1, characterized in that, The process of generating event confusion relationships through the event confusion and disambiguation module based on the candidate results of security events, generating an evidence gap map based on the event confusion relationships through the evidence gap map generation module, and generating active supplementary data collection control instructions based on the evidence influence edges between missing evidence nodes and candidate event nodes specifically includes: The event obfuscation and disambiguation module reads the candidate results of security events, and forms candidate event pairs in order of increasing confidence difference of candidate events, generating a set of candidate event pairs. Based on the candidate event pair set, candidate event confidence distribution, event association graph nodes, event association graph edges, risk evolution state, security node memory representation, and security relationship attention representation, calculate the confusion intensity of each candidate event pair and generate a candidate event confusion intensity set. Based on the set of candidate event confusion intensity, an evidence integrity analysis is performed on each candidate event to generate a set of event confusion causes; The set of candidate event pairs, the set of candidate event confusion intensity, and the set of event confusion causes are associated to generate event confusion relationships; The evidence gap graph generation module generates missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary sampling action nodes based on the event confusion relationship. The evidence gap map generation module calculates the evidence influence weight of the evidence influence edge, and connects the missing evidence nodes with the supplementary sampling action nodes according to the evidence influence weight to generate the evidence gap map. Based on the missing evidence nodes, candidate event nodes, evidence influence edges, and supplementary sampling action nodes in the evidence gap diagram, an active supplementary sampling control instruction is generated.

7. The robot security control method based on deep learning according to claim 1, characterized in that, The process of controlling the security robot to perform supplementary sampling actions according to the active supplementary sampling control command, generating supplementary sampling patrol observation data, and generating a robot closed-loop control strategy based on the supplementary sampling patrol observation data specifically includes: Based on the supplementary mining action type, supplementary mining target object, supplementary mining spatial location, supplementary mining sensor type, supplementary mining duration, and supplementary mining data write-back object in the active supplementary mining control command, a supplementary mining execution task is generated; Based on the task of supplementary mining, the security robot is controlled to perform supplementary mining actions, continuously recording the movement direction, dwelling state, avoidance actions and response status of the target object, and generating supplementary mining patrol observation data; The supplementary patrol observation data is synchronized in time, aligned spatially, and supplemented with data source markings to generate supplementary feedback data units. Based on the supplementary data collection feedback unit, a supplementary data collection feedback event is generated, and the supplementary data collection feedback event is written back to the security event stream to generate an updated security event stream; Update the dynamic security scenario diagram with security semantic attributes based on the supplementary data collection feedback events, and generate the updated dynamic security scenario diagram; The updated security event flow and the updated dynamic security scenario graph are input into the improved dynamic graph representation network to generate an updated security relationship attention representation; Based on the updated security event flow, the updated dynamic security scenario diagram, the updated security node memory representation, the updated security relationship attention representation, the event confusion relationship, and the supplementary feedback events, a risk handling result is generated, and a robot closed-loop control strategy is generated based on the risk handling result.

8. A deep learning-based robot security control system, executing the deep learning-based robot security control method according to any one of claims 1 to 7, characterized in that, Includes the following modules: The patrol data collection module is used to collect information in the patrol area by security robots, preprocess the collected information to generate patrol observation datasets, and generate security event streams based on the patrol observation datasets. The event generation module is used to construct a dynamic security scenario graph based on the security event flow, configure node status information for graph nodes, configure relationship persistence information and relationship change information for graph edges, and generate a dynamic security scenario graph with security semantic attributes. The context graph building module is used to input dynamic security context graphs with security semantic attributes and security event flows into the improved dynamic graph representation network. Through the security event flow input module, security node memory update module, and security relationship attention aggregation module, it generates security node memory representations and security relationship attention representations. The graph representation module is used to model the risk evolution intensity in the network through an improved dynamic graph representation module. It generates candidate results for security events based on the security node memory representation, the security relationship attention representation, and the security event flow. The disambiguation and supplementary acquisition module is used to generate event confusion relationships based on the candidate results of security events through the event confusion disambiguation module, generate an evidence gap map based on the event confusion relationships through the evidence gap map generation module, and generate active supplementary acquisition control instructions based on the evidence influence edges between missing evidence nodes and candidate event nodes. The closed-loop control module is used to control the security robot to perform supplementary sampling actions according to the active supplementary sampling control command, generate supplementary sampling patrol observation data, and generate robot closed-loop control strategy based on the supplementary sampling patrol observation data.

9. A storage medium, characterized in that, The medium stores executable instructions for being read by a computer's processor to perform the method of any one of claims 1 to 7.

Images