A cloud-edge collaborative intelligent agent autonomous decision driving method

Abstract

This application provides a cloud-edge collaborative intelligent agent autonomous decision-making driving method. It acquires multi-channel sensing video streams representing the agent's operating environment, performs multi-dimensional feature extraction and real-time behavioral feature capture, and generates edge situational awareness metadata including behavioral anomaly dimensions. The metadata is uploaded to an intelligent agent management service platform, where a local expert knowledge base is retrieved for strategy optimization processing to determine cloud-based enhancement logic with high-dimensional semantic guidance attributes. Cross-level semantic alignment processing is performed on the edge situational awareness metadata and the cloud-based enhancement logic to construct an alignment constraint vector and perform driving load fusion calculation to determine the autonomous decision-making driving instructions for the intelligent agent. This application, through cloud-edge collaborative feature alignment and enhancement, solves the technical defects of traditional solutions, such as the lack of global logical guidance in edge decision-making and the lag in cloud decision-making response, thereby improving the autonomous decision-making adaptability of the intelligent agent in complex operating environments.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and intelligent agent control technology, and more specifically, to a cloud-edge collaborative intelligent agent autonomous decision-making driving method. Background Technology

[0002] With the widespread application of artificial intelligence technology in the operation and smart management of public spaces, the autonomous decision-making capability of intelligent agents (such as inspection robots and intelligent guide terminals) in complex operating environments has become the core of the system. In order to achieve efficient management, it is usually necessary to monitor the operating environment in real time and drive the intelligent agents to perform corresponding tasks.

[0003] In existing agent-driven solutions, an independent decision-making architecture based on edge terminals is typically adopted. This solution first captures single-channel monitoring footage using cameras deployed on-site; then, edge computing nodes perform simple target detection on the footage to identify whether there are any pre-defined violations; finally, the edge nodes directly generate control commands based on locally pre-set fixed logic scripts to drive the agent to move or issue alarms.

[0004] However, this independent decision-making scheme based on edge terminals has significant technical shortcomings. Due to the limited computing resources and storage space of edge terminals, it is difficult to support large-scale expert knowledge bases and complex large-scale model reasoning processes. As a result, the decision logic generated by these terminals can only handle simple, pre-defined local scenarios and lacks a deep understanding of global business objectives. At the same time, the limited perspective of a single monitoring screen makes situational awareness incomplete and prone to blind spots. Furthermore, without high-dimensional semantic guidance from the cloud, edge decision-making has low accuracy in executing actions when faced with complex and sudden behavioral anomalies, making it difficult to achieve true cloud-edge collaborative decision-making and intelligent driving. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a cloud-edge collaborative intelligent agent autonomous decision-making driving method to at least alleviate the aforementioned technical problems.

[0006] A cloud-edge collaborative intelligent agent autonomous decision-making driving method includes: Step 1: Acquire multi-channel sensing video streams representing the agent's operating environment, perform multi-dimensional feature extraction and real-time behavior feature capture on the multi-channel sensing video streams to generate edge situational awareness metadata that represents the local action space and includes the dimension of behavioral anomalies. Step 2: Upload the edge situational awareness metadata to the intelligent agent management service platform, and use the intelligent agent management service platform to retrieve the local expert knowledge base to perform policy optimization processing based on business logic constraints, so as to determine the cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes. Step 3: Perform cross-level semantic alignment processing on edge situational awareness metadata and cloud-based augmentation logic to construct alignment constraint vectors, and perform driving load fusion calculation based on alignment constraint vectors to determine the autonomous decision-making driving instructions that guide the agent's action output and have cloud-edge collaborative decision-making attributes. The multi-channel sensing video stream consists of multiple sub-sensing video streams that perform multi-view coverage of the target physical scene and have consistent time references.

[0007] Optionally, obtaining the multi-channel sensing video stream in step 1 includes: Multiple sensing nodes deployed at different spatial locations within the target physical scene are retrieved to simultaneously capture multiple sub-sensor video streams representing the target physical scene; A unified timestamp operator is used to perform frame-level synchronization calibration on each sub-sensory video stream to generate a multi-channel sensing video stream.

[0008] Optionally, the edge situational awareness metadata generated in step 1 includes: The intelligent sensing edge terminal is used to perform trajectory tracking processing on the motion trajectory of each sub-sensing video stream in the multi-channel sensing video stream in order to determine the motion vector distribution of the target. Based on the set edge recognition operator, edge situational awareness metadata representing the local action space and including the dimension of behavioral anomalies is generated according to the motion vector distribution.

[0009] Optionally, determining the cloud-based enhancement logic in step 2 includes: In the intelligent agent management service platform, edge situational awareness metadata is transformed into semantic retrieval vectors, and the semantic retrieval vectors are used to retrieve related business rule slices from the local expert knowledge base. The logic is refactored by performing business rule slices to determine cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes.

[0010] Optionally, the action of performing cross-level semantic alignment processing in step 3 includes: Local actions matched by edge situational awareness metadata are extracted, and inference processing is performed on the local actions according to the global policy space provided by cloud-enhanced logic to obtain the global semantic inference vector corresponding to the local actions. By calculating the semantic overlap entropy value between the global semantic inference vector and the global policy space, an alignment constraint vector is generated to eliminate execution ambiguity.

[0011] Optionally, the instructions for determining the agent's autonomous decision-making drive in step 3 include: Gain adjustment is performed on edge situational awareness metadata using alignment constraint vectors to generate edge-driven components; The cloud-driven component is extracted from the cloud-enhanced logic, and the edge-driven component and the cloud-driven component are subjected to load-weighted fusion processing to generate an agent autonomous decision-making driving instruction that guides the agent's action output and has cloud-edge collaborative decision-making attributes.

[0012] Optionally, the method also includes edge computing power adaptive adjustment actions: Real-time recording of input sources, model versions, and feature evolution trajectories during semantic analysis of execution intent, matching degree detection, and semantic consistency conflict resolution; The recorded information is encapsulated into an audit payload with logical dependencies to build a traceable generation chain, and the traceable generation chain is used to perform authenticity backtracking verification on the task execution results.

[0013] Optionally, the method also includes communication link optimization processing actions: The task processing engine is used to perform granular decomposition processing on natural language instructions to generate edge preprocessing task sequences and cloud high-dimensional task sequences; Localized routing matching is performed on edge preprocessing task sequences to retrieve locally deployed intelligent agent model entities for low-dimensional business logic processing, and cloud-based high-dimensional task sequences are distributed to cloud-based expert clusters for deep inference through the unified intelligent agent interface service.

[0014] Optionally, the method also includes local logic fallback actions: Real-time detection of heartbeat signals sent down by cloud-enhanced logic; In response to the interruption of the heartbeat signal, the intelligent sensing edge terminal is triggered to execute local decisions based on a preset rule base, and the obtained local decision results are used to temporarily replace the autonomous decision-driving instructions of the intelligent agent.

[0015] Optionally, the method also includes full-chain logical auditing actions: The input and output features during the execution of multidimensional feature extraction and real-time behavioral feature capture, strategy optimization processing, and driving load fusion calculation are encapsulated into a topology audit load that represents the logical dependencies of the decision-making link. The pre-defined traceable generation chain is invoked to perform consistency verification on the topology audit payload, so as to constrain the generation path of the agent's autonomous decision-driven instructions to match the agent's safety behavior specification.

[0016] Technical advantages of the technical solution provided in this application This application's cloud-edge collaborative intelligent agent autonomous decision-making driving method addresses the technical shortcomings of traditional edge independent decision-making schemes, such as limitations in perception and recognition, lack of global logical guidance, and low decision-making accuracy in complex scenarios. It solves the problem of incomplete situational awareness caused by single-view monitoring in traditional schemes by performing multi-dimensional feature extraction on multiple sensing video streams to generate edge situational awareness metadata. Compared to traditional schemes that rely on local recognition from a single video stream, this application, through real-time behavior capture of multiple video streams with viewpoint coverage and temporal consistency, can comprehensively characterize the target's motion vector from multiple spatial dimensions, providing richer feature payloads for generating accurate edge situational awareness.

[0017] Based on the generated edge situational awareness metadata, the cloud-based augmentation logic is determined by retrieving the local expert knowledge base through the intelligent agent management service platform. This solves the problem that traditional edge terminals cannot handle complex business logic due to computing power limitations. In traditional solutions, edge nodes can only execute simple rules, while this application utilizes the high-performance computing capabilities of the cloud platform to perform correlation retrieval and logical reconstruction between edge metadata and massive expert knowledge, endowing local actions with high-dimensional semantic guidance. This enables the decision-making process to take into account both global management strategies and real-time scene changes, significantly improving the logical depth of strategy generation.

[0018] Finally, by performing cross-level semantic alignment and payload fusion calculation on edge situational awareness metadata and cloud-based augmentation logic, the problems of cloud-edge decision misalignment and execution ambiguity are resolved. This application generates alignment constraint vectors by calculating semantic overlap entropy values, and uses these vectors to guide the weighted fusion of driving components, achieving precise gain and correction of edge execution actions by the cloud-based global policy. Compared to the mechanical execution of instructions by edge nodes in traditional solutions, the autonomous decision-driven instructions determined in this application are more adaptable to complex and changing physical environments, enabling the agent to have higher stability and compliance when executing actions. Attached Figure Description

[0019] Figure 1 This application provides an embodiment of a cloud-edge collaborative intelligent agent autonomous decision-making driving method.

[0020] Figure 2 This application provides an embodiment of a cloud-edge collaborative intelligent agent autonomous decision-making driving device.

[0021] Figure 3 This is an electronic device according to an embodiment of the present application. Detailed Implementation

[0022] like Figure 1 As shown in the figure, a cloud-edge collaborative intelligent agent autonomous decision-making driving method according to an embodiment of this application includes: Step 1: Acquire multi-channel sensing video streams representing the agent's operating environment, perform multi-dimensional feature extraction and real-time behavior feature capture on the multi-channel sensing video streams to generate edge situational awareness metadata that represents the local action space and includes the dimension of behavioral anomalies. Step 2: Upload the edge situational awareness metadata to the intelligent agent management service platform, and use the intelligent agent management service platform to retrieve the local expert knowledge base to perform policy optimization processing based on business logic constraints, so as to determine the cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes. Step 3: Perform cross-level semantic alignment processing on edge situational awareness metadata and cloud-based augmentation logic to construct alignment constraint vectors, and perform driving load fusion calculation based on alignment constraint vectors to determine the autonomous decision-making driving instructions that guide the agent's action output and have cloud-edge collaborative decision-making attributes. The multi-channel sensing video stream consists of multiple sub-sensing video streams that perform multi-view coverage of the target physical scene and have consistent time references.

[0023] Optionally, obtaining the multi-channel sensing video stream in step 1 includes: Multiple sensing nodes deployed at different spatial locations within the target physical scene are retrieved to simultaneously capture multiple sub-sensor video streams representing the target physical scene; A unified timestamp operator is used to perform frame-level synchronization calibration on each sub-sensory video stream to generate a multi-channel sensing video stream.

[0024] Preferably, in this application, to meet the multi-view, blind-spot-free perception requirements of the target physical scene, multiple sensing nodes are pre-deployed in different spatial orientations within the target physical scene. The deployment location and number of sensing nodes are determined based on the spatial structure of the target physical scene, the operational coverage of the intelligent agent, and the preset perception accuracy requirements. Each sensing node has the ability to continuously acquire video images of the corresponding spatial area, and the perception coverage of adjacent sensing nodes overlaps by a preset proportion to eliminate perception blind spots within the target physical scene and provide a hardware foundation for the synchronous capture of subsequent sub-sensing video streams.

[0025] Preferably, when the agent starts a task, the effective perception area of ​​the target physical scene is defined based on the task boundary of the agent. From all the pre-deployed perception nodes, the perception nodes whose perception coverage falls within the effective perception area are selected as target perception nodes. A synchronous acquisition command is sent to all target perception nodes to trigger all target perception nodes to start video image acquisition operations at the same time. Multiple sub-perception video streams representing the corresponding viewpoint area of ​​the target physical scene are captured synchronously. Each sub-perception video stream is an original video data stream containing a continuous video frame sequence output by the corresponding target perception node.

[0026] Preferably, for each sub-sensing video stream output by a target sensing node, frame-level basic preprocessing operations are performed. These basic preprocessing operations include lens distortion correction, invalid frame removal, and image parameter normalization. Lens distortion correction is used to eliminate image distortion caused by the optical characteristics of the lens at the sensing node. Invalid frame removal is used to remove invalid video frames in the sub-sensing video stream that are completely black or severely blurred. Image parameter normalization is used to adjust the brightness and contrast parameters of the sub-sensing video streams output by different target sensing nodes to a uniform range. The sub-sensing video stream after the basic preprocessing operations will be used as the input object for subsequent timestamp operator processing.

[0027] Preferably, a pre-designed unified timestamp operator is retrieved, and the global clock reference synchronously issued by the intelligent agent management service platform is used as the sole time reference. For each sub-sensing video stream after basic preprocessing, a frame-level timestamp addition operation is performed. Specifically, for each frame of video data in the sub-sensing video stream, the global clock time information corresponding to the acquisition time of the video data, the unique device identifier of the corresponding target sensing node, and the frame sequence number information of the video data in the sub-sensing video stream are added to the frame header of the video data, so that each frame of video data in each sub-sensing video stream has a globally recognizable time reference and source identifier.

[0028] Preferably, based on the timestamp information carried by each frame of video data in each sub-sensing video stream, frame-level synchronization calibration is performed on all sub-sensing video streams. Specifically, taking the continuous time points of the global clock reference as synchronization anchor points, for each synchronization anchor point, video frames from all sub-sensing video streams whose time difference between the frame header timestamp information and the synchronization anchor point is within a preset allowable deviation range (for example, for a sensing node with a frame acquisition frequency of 25 frames per second, the preset allowable deviation range is set to 40 milliseconds) are selected. The selected multi-frame video data from different sub-sensing video streams are combined into a multi-view synchronization frame group corresponding to the synchronization anchor point, thereby eliminating the frame asynchrony problem caused by data transmission delay and acquisition start time difference between different target sensing nodes.

[0029] Preferably, according to the global clock time corresponding to the synchronization anchor point, all multi-view synchronization frame groups are sequentially arranged and encapsulated to generate a multi-channel sensing video stream with time base consistency. Each group of data in the multi-channel sensing video stream corresponds to multi-view video frame data of different spatial orientations of the target physical scene at the same time, completely covering the effective sensing area of ​​the target physical scene, and providing a multi-dimensional, blind-spot-free, and time-synchronized video data foundation for subsequent multi-dimensional feature extraction and real-time behavioral feature capture.

[0030] Preferably, for the encapsulated multi-channel sensing video stream, a synchronization compliance verification process is performed. Specifically, a preset number of synchronization anchors are randomly selected from the multi-channel sensing video stream (for example, the number of selected anchors is not less than 5% of the total number of synchronization anchors). For each selected synchronization anchor, the time difference between the timestamp information of all video frames in the group and the time difference of the corresponding synchronization anchor is verified to be within the preset allowable deviation range. If all selected synchronization anchors pass the verification, the multi-channel sensing video stream is determined to meet the input requirements for subsequent processing. If there are synchronization anchors that fail the verification, the corresponding frame-level synchronization calibration process is re-executed until the multi-channel sensing video stream passes the synchronization compliance verification.

[0031] Optionally, the edge situational awareness metadata generated in step 1 includes: The intelligent sensing edge terminal is used to perform trajectory tracking processing on the motion trajectory of each sub-sensing video stream in the multi-channel sensing video stream in order to determine the motion vector distribution of the target. Based on the set edge recognition operator, edge situational awareness metadata representing the local action space and including the dimension of behavioral anomalies is generated according to the motion vector distribution.

[0032] Preferably, in the specific technical implementation of step 1, the intelligent sensing edge terminal takes the multi-channel sensing video stream that has undergone frame-level synchronization calibration as the input object, and performs frame-level multi-dimensional visual feature extraction processing on all video frames in each multi-view synchronization frame group in the multi-channel sensing video stream to generate a frame-level visual feature map corresponding to each video frame. The frame-level visual feature map contains the contour features, texture features, and spatial position features of the target in the corresponding video frame, providing a unified feature benchmark for subsequent motion trajectory tracking processing of each sub-sensing video stream.

[0033] Preferably, for the continuous frame sequence corresponding to each sub-sensing video stream in the multi-channel sensing video stream, the intelligent sensing edge terminal performs target detection and cross-frame association matching processing based on the corresponding generated continuous frame-level visual feature map to identify all targets present in the continuous frame-level visual feature map, assign a globally unique target identifier to each identified target, and then generate a single-view motion trajectory of each target in the corresponding sub-sensing video stream based on the temporal changes of the spatial position features corresponding to the same target identifier in the continuous frame-level visual feature map. The single-view motion trajectory completely records the position change process and temporal association relationship of the corresponding target in the continuous time series.

[0034] Preferably, the intelligent sensing edge terminal, based on the consistency of the time base of multiple sensing video streams, performs multi-view spatial coordinate fusion processing on the single-view motion trajectories generated by different sub-sensing video streams within the same time interval and bearing the same target identifier. This process maps the single-view motion trajectories from different perspectives to the global three-dimensional spatial coordinate system corresponding to the target's physical scene, thereby generating a continuous and complete motion trajectory for each target in the global three-dimensional spatial coordinate system. Then, based on the position change and corresponding time difference of the continuous and complete motion trajectory within the continuous time series, the motion speed, motion direction, and acceleration parameters corresponding to each target are calculated and combined to form the motion vector distribution of the target.

[0035] Preferably, an edge recognition operator is pre-set in the intelligent sensing edge terminal. The edge recognition operator is constructed based on a lightweight feature inference network adapted to the computing power constraints of the edge terminal. It is pre-trained and converged using a target behavior sample dataset corresponding to the target physical scene. It has the ability to recognize normal patterns and distinguish abnormal patterns of target motion behavior. The input parameters of the edge recognition operator include the target's motion vector distribution and the local action space boundary constraint parameters corresponding to the target physical scene. The output results include the target's behavior attribute labels, behavior anomaly confidence, and a set of executable local action candidates, providing standardized inference output content for the generation of edge situational awareness metadata.

[0036] Preferably, the intelligent sensing edge terminal inputs the calculated motion vector distribution of the target into a pre-set edge recognition operator. The edge recognition operator performs compliance verification on the target motion behavior corresponding to the motion vector distribution based on the local action space boundary constraint parameters corresponding to the target's physical scene, in order to distinguish whether the target's motion behavior falls within the allowed range of the local action space. At the same time, it performs abnormal pattern matching on the target's motion behavior to generate the target's corresponding behavioral abnormality dimension data. The behavioral abnormality dimension data includes the behavioral abnormality type, behavioral abnormality confidence, spatial location of the behavioral abnormality, and corresponding timestamp information.

[0037] Preferably, the intelligent sensing edge terminal combines the target's behavioral attribute labels, behavioral anomaly dimension data, and executable local action candidate set output by the edge recognition operator with the target's motion vector distribution, the target's corresponding globally unique target identifier, and the corresponding timestamp information, and performs structured encapsulation processing to generate edge situational awareness metadata that represents the local action space and includes behavioral anomaly dimensions. The edge situational awareness metadata adopts a lightweight structured data format, which fully carries the local situational information of the target's physical scene while adapting to the transmission bandwidth constraints of the cloud-edge communication link, providing standardized input data for the strategy optimization processing of the intelligent agent management service platform in the subsequent step 2.

[0038] Preferably, after completing the encapsulation of edge situational awareness metadata, the intelligent sensing edge terminal performs integrity and compliance verification on the edge situational awareness metadata. The verification includes the integrity of the structured fields of the edge situational awareness metadata, the compliance of the confidence values ​​of the abnormal behavior dimension data, and the consistency of the timestamp information with the time base of the multi-channel sensing video stream. Edge situational awareness metadata that passes the verification will be used as valid data in the subsequent processing flow. Edge situational awareness metadata that fails the verification will trigger secondary feature extraction and trajectory tracking processing of the corresponding multi-channel sensing video stream to ensure that the generated edge situational awareness metadata has high availability.

[0039] Optionally, determining the cloud-based enhancement logic in step 2 includes: In the intelligent agent management service platform, edge situational awareness metadata is transformed into semantic retrieval vectors, and the semantic retrieval vectors are used to retrieve related business rule slices from the local expert knowledge base. The logic is refactored by performing business rule slices to determine cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes.

[0040] Preferably, in the specific technical implementation of step 2, the intelligent agent management service platform receives edge situational awareness metadata uploaded by the intelligent sensing edge terminal, performs transmission integrity verification and structured parsing processing on the received edge situational awareness metadata to remove invalid metadata with transmission packet loss or missing fields, and extracts target identifiers, motion vector distributions, abnormal behavior dimension data, local action space constraint information, and corresponding timestamp information contained in the edge situational awareness metadata to generate standardized situational metadata, providing a unified and complete input foundation for the subsequent conversion of semantic retrieval vectors.

[0041] Preferably, the intelligent agent management service platform uses standardized situational metadata as the processing object, retrieves a pre-trained and converged semantic coding network, and performs feature extraction and semantic coding mapping processing on the core semantic fields in the standardized situational metadata to unify unstructured situational description information and structured numerical parameters into a fixed-dimensional semantic feature sequence. Then, it performs normalization processing on the semantic feature sequence to generate a semantic retrieval vector. Each dimension of the semantic retrieval vector corresponds to the semantic representation of event type, behavior anomaly level, spatial location attribute, local action constraint, and business association dimension in the target physical scene, which can completely quantify the core business semantics of the current edge situation.

[0042] Preferably, the intelligent agent management service platform has a pre-built local expert knowledge base. The local expert knowledge base stores the full range of business management rules, standardized handling procedures for abnormal events, global business target constraints, and intelligent agent security behavior specifications corresponding to the target physical scenario. Moreover, all stored business rules are divided into multiple independent business rule slices with exclusive semantic tag vectors according to the dimensions of business execution links, applicable scenarios, and handling objects. Each business rule slice contains the corresponding execution logic, constraints, applicable boundaries, and business priority information, providing standardized basic data support for the retrieval and invocation of business rules.

[0043] Preferably, the intelligent agent management service platform uses the generated semantic retrieval vector as the retrieval basis to perform retrieval processing of associated business rule slices in the local expert knowledge base. Specifically, it calculates the vector similarity between the semantic retrieval vector and the semantic tag vector of each business rule slice in the local expert knowledge base, filters out business rule slices with vector similarity higher than a preset similarity threshold, and then performs scenario adaptability verification on the filtered business rule slices in combination with the current intelligent agent's business logic constraints. It removes business rule slices that do not match the applicable scenarios and business objectives of the current edge situation, and finally obtains the business rule slices associated with the current edge situation.

[0044] Preferably, the intelligent agent management service platform performs compliance verification and priority sorting on the retrieved associated business rule slices. Based on the global business objectives, security behavior specifications, and management requirements of the target physical scenario of the intelligent agent, the platform eliminates business rule slices that have logical conflicts or violate security constraints. Then, according to the behavior anomaly level, business processing priority, and execution timeliness requirements corresponding to the business rule slices, the verified business rule slices are sorted in descending order to generate an ordered sequence of business rule slices. This provides logically coherent and clearly prioritized processing objects for subsequent logical reconstruction processing.

[0045] Preferably, the intelligent agent management service platform takes an ordered sequence of business rule slices as input and performs logic reconstruction processing. Specifically, according to the temporal logic and dependency relationship of business execution, it performs logic splicing and process chaining on each business rule slice in the ordered sequence of business rule slices, performs conflict resolution processing on action conflicts and boundary overlaps between different business rule slices, and supplements the execution boundary constraints and exception fallback handling logic corresponding to the global strategy space, integrating the scattered and independent business rule slices into an initial global strategy logic that covers the complete handling process and matches the global business goals.

[0046] Preferably, the intelligent agent management service platform performs high-dimensional semantic encapsulation and adaptation processing on the initial global policy logic. It supplements the initial global policy logic with semantic guidance information of global business objectives, semantic constraint labels of action execution, and mapping relationship between local actions and global policies to generate cloud-based enhanced logic that represents the global policy space and has high-dimensional semantic guidance attributes. The cloud-based enhanced logic adopts a structured data format adapted to the parsing capabilities of intelligent sensing edge terminals. It not only contains global policy instructions for intelligent agents to execute actions, but also carries high-dimensional semantic guidance information corresponding to global business objectives, which can provide global-level policy guidance and constraints for the execution of local actions at the edge.

[0047] Preferably, the intelligent agent management service platform performs final compliance verification and traceability binding processing on the generated cloud-based enhanced logic. It verifies whether the execution logic of the cloud-based enhanced logic conforms to the security behavior specification of the intelligent agent, whether the semantic information is consistent with the edge situational awareness metadata, and whether the execution instructions are adapted to the hardware execution capabilities of the intelligent agent. After the verification is passed, the cloud-based enhanced logic is bound and stored with the unique identifier of the corresponding edge situational awareness metadata. At the same time, the cloud-based enhanced logic is sent to the corresponding intelligent sensing edge terminal, providing standardized input data for the cross-level semantic alignment processing in the subsequent step 3, and also providing a traceable data source for the subsequent full-chain logic auditing actions.

[0048] Optionally, the action of performing cross-level semantic alignment processing in step 3 includes: Local actions matched by edge situational awareness metadata are extracted, and inference processing is performed on the local actions according to the global policy space provided by cloud-enhanced logic to obtain the global semantic inference vector corresponding to the local actions. By calculating the semantic overlap entropy value between the global semantic inference vector and the global policy space, an alignment constraint vector is generated to eliminate execution ambiguity.

[0049] Preferably, in the specific technical implementation of step 3, the intelligent sensing edge terminal receives the cloud-based enhancement logic issued by the intelligent agent management service platform, and simultaneously retrieves the locally generated edge situational awareness metadata. The edge situational awareness metadata and the cloud-based enhancement logic are subjected to structured parsing and semantic field mapping processing, respectively, to extract the local action candidate set, behavior anomaly dimension data, and local action space constraint information contained in the edge situational awareness metadata. At the same time, the global policy space, high-dimensional semantic guidance information, and global business goal constraints contained in the cloud-based enhancement logic are extracted, providing a unified input foundation with clear semantic boundaries for the subsequent extraction and deduction of local actions.

[0050] Preferably, based on the local action candidate set and local action space constraint information in the parsed edge situational awareness metadata, local action matching and filtering processing is performed to eliminate invalid action candidates that exceed the local action space constraint boundary and are not related to the current behavior anomaly dimension data, thereby obtaining valid local actions that match the current edge situation. Valid local actions fully include the timing, path, triggering conditions, and execution boundary information of the action execution, and can directly correspond to the specific operations that the agent can perform in the current local scene.

[0051] Preferably, the global policy space in the parsed cloud-enhanced logic is used as the deduction constraint framework to perform full-process deduction processing on the matched effective local actions. Specifically, the execution process of the effective local actions is simulated in the complete business scenario corresponding to the global policy space. The impact of the effective local actions on the global business objectives, the adaptability to other related business rules, and the boundary conflict problems that may occur during the execution process are deduced. Then, the full-link global execution characteristics corresponding to the effective local actions are generated. The full-link global execution characteristics include the global impact dimension of the action execution, the business adaptability level, the conflict risk coefficient, and the matching degree information with the global business objectives.

[0052] Preferably, semantic encoding and normalization are performed on the global execution full-link features corresponding to the effective local actions. A lightweight semantic encoding branch consistent with the semantic encoding network structure in step 2 is retrieved, and the multi-dimensional information in the global execution full-link features is mapped to the same semantic feature space as the global policy space to generate a global semantic inference vector corresponding to the local action. Each dimension of the global semantic inference vector corresponds to the semantic representation of the business target matching degree, scenario adaptability, security compliance, and global influence weight of the action execution, which corresponds one-to-one with the semantic dimensions of the global policy space, providing a dimensionally unified calculation basis for the subsequent semantic overlap entropy value calculation.

[0053] Preferably, the global semantic inference vector and the global policy space provided by the cloud-enhanced logic are used as the computation objects to perform semantic overlap entropy calculation. Specifically, in a unified semantic feature space, the overlap of feature distributions of the global semantic inference vector and the global policy space in each corresponding semantic dimension is calculated. Then, based on the business weights corresponding to each semantic dimension, the overlap of each dimension is weighted and summed and entropy value is transformed to obtain the semantic overlap entropy value between the global semantic inference vector and the global policy space. The semantic overlap entropy value can quantitatively represent the degree of semantic consistency and execution ambiguity of local actions under the global policy framework. The lower the entropy value, the higher the semantic consistency and the smaller the execution ambiguity.

[0054] Preferably, based on the calculated semantic overlap entropy value, the alignment constraint vector is constructed. Specifically, based on the semantic overlap entropy value and combined with the business constraint priority of each semantic dimension in the global policy space, a constraint adjustment coefficient corresponding to each semantic dimension is generated. Then, the constraint adjustment coefficients of each dimension are arranged and combined in the same dimensional order as the global semantic inference vector and the global policy space to generate an alignment constraint vector for eliminating execution ambiguity. Each dimension element of the alignment constraint vector corresponds to the constraint adjustment weight of a semantic dimension in the global policy space, which can perform directional constraints and corrections on dimensions where there is a semantic deviation between local actions and global policies.

[0055] Preferably, after the alignment constraint vector is constructed, a validity verification process is performed on the alignment constraint vector. The verification standard is whether the semantic deviation between the alignment constraint vector and the global semantic inference vector after constraint correction is lower than a preset deviation threshold. The alignment constraint vector that passes the verification will be used as a valid constraint parameter to enter the subsequent driving load fusion calculation stage. The alignment constraint vector that fails the verification will trigger the secondary inference of local actions and the recalculation of semantic overlap entropy value until an alignment constraint vector that meets the validity requirements is generated, so as to ensure that the effect of cross-level semantic alignment processing can meet the needs of subsequent cloud-edge collaborative decision-making.

[0056] Optionally, the instructions for determining the agent's autonomous decision-making drive in step 3 include: Gain adjustment is performed on edge situational awareness metadata using alignment constraint vectors to generate edge-driven components; The cloud-driven component is extracted from the cloud-enhanced logic, and the edge-driven component and the cloud-driven component are subjected to load-weighted fusion processing to generate an agent autonomous decision-making driving instruction that guides the agent's action output and has cloud-edge collaborative decision-making attributes.

[0057] Preferably, in the specific technical implementation of step 3, the intelligent sensing edge terminal takes the alignment constraint vector generated by cross-level semantic alignment processing, the locally generated edge situational awareness metadata, and the cloud-based enhancement logic issued by the intelligent agent management service platform as the basic inputs. It performs structured parsing, temporal synchronization verification, and semantic dimension mapping processing on the three types of input data respectively to remove invalid data with temporal misalignment, missing fields, and semantic dimension mismatch. At the same time, it maps the core semantic fields of the three types of input data to a unified action instruction feature space, providing a standardized input foundation with unified format, temporal synchronization, and semantic dimension matching for the subsequent generation of edge driving components and fusion processing of driving payloads.

[0058] Preferably, the alignment constraint vector and edge situational awareness metadata that have undergone standardization are taken as the processing objects, and the directional gain adjustment processing of the edge situational awareness metadata is performed. Specifically, according to the constraint adjustment weights corresponding to each dimension in the alignment constraint vector, the feature parameters of the corresponding semantic dimensions in the edge situational awareness metadata are weighted and corrected. For dimension features with high semantic consistency with the global policy space, positive gain amplification is performed. For dimension features with semantic deviation or execution ambiguity from the global policy space, reverse constraint correction is performed. At the same time, invalid feature parameters in the edge situational awareness metadata that exceed the execution boundary of the global policy space are removed, so as to generate edge driving components that are adapted to global policy constraints and eliminate execution ambiguity.

[0059] Preferably, the standardized cloud-based enhanced logic undergoes structured parsing and driver feature extraction. Based on the business priority, execution timeliness requirements, and agent hardware execution capability constraints of the global policy space, the core action instructions, action execution timing constraints, action boundary thresholds, and global business goal adaptation parameters for global policy execution are extracted from the cloud-based enhanced logic. At the same time, non-driver-type feature information in the cloud-based enhanced logic that is only used for semantic guidance and does not directly participate in action execution is removed, thereby generating a cloud-based driver component with global policy guidance attributes and adaptable agent execution capabilities.

[0060] Preferably, before performing load-weighted fusion processing, an adaptive determination process for fusion weight coefficients is first performed. The current computing power load status of the intelligent sensing edge terminal, the transmission quality of the cloud-edge communication link, the behavior anomaly level of the current scenario, and the timeliness requirements of business execution are used as the judgment criteria to assign corresponding fusion weight coefficients to the edge-driven component and the cloud-driven component, respectively. The fusion weight coefficient of the edge-driven component is positively correlated with the edge computing power redundancy and the real-time requirements of the scenario, while the fusion weight coefficient of the cloud-driven component is positively correlated with the behavior anomaly level and the global business constraint strength. The sum of the two sets of fusion weight coefficients is a fixed value, providing a quantitative weight basis for subsequent load-weighted fusion processing.

[0061] Preferably, based on the generated fusion weight coefficients, load-weighted fusion processing is performed on the edge-driven component and the cloud-driven component. Specifically, the feature parameters of the same semantic dimension and the same action type in the edge-driven component and the cloud-driven component are weighted and summed according to the corresponding dimension's fusion weight coefficients. Boundary compliance verification is performed on the action parameters of each dimension obtained after fusion. Parameters that exceed the execution capabilities of the intelligent agent's hardware or violate the safety behavior rules are corrected to the compliance range. At the same time, the fused parameters are sorted and encapsulated according to the timing logic of action execution to generate the initial intelligent agent driving instructions.

[0062] Preferably, the generated initial agent-driven instructions undergo full-dimensional compliance verification and scenario adaptability optimization. The verification includes the scenario matching degree between the initial agent-driven instructions and the current edge situational awareness metadata, the consistency with the global strategy of the cloud-enhanced logic, the compliance with the agent's security behavior conventions, and the feasibility of agent hardware execution. For the initial agent-driven instructions that pass the verification, the corresponding abnormal fallback handling logic and state feedback trigger conditions are added to the corresponding action execution. Finally, agent autonomous decision-making driven instructions that guide agent action output and have cloud-edge collaborative decision-making attributes are generated.

[0063] Preferably, after generating the autonomous decision-making driving instructions for the intelligent agent, the autonomous decision-making driving instructions are bound and stored with the unique identifiers of the corresponding edge situational awareness metadata, cloud-based enhanced logic, and alignment constraint vectors. At the same time, the autonomous decision-making driving instructions are sent to the execution control module of the intelligent agent to drive the intelligent agent to perform the corresponding actions. The bound and stored associated data will serve as the core input for subsequent full-chain logic auditing actions, providing a traceable and complete data source for the consistency verification of the decision-making chain.

[0064] Optionally, the method also includes edge computing power adaptive adjustment actions: Real-time monitoring of the video memory usage of intelligent sensing edge terminals, and using the video memory usage to identify the sensing load pressure for each sub-sensing video stream; In response to the sensing load pressure exceeding the preset load threshold, the frame sampling frequency of sensing data interception is reduced, and a computing power offloading request is sent simultaneously to adjust the generation priority of cloud-based enhancement logic.

[0065] Preferably, in the entire process of multi-dimensional feature extraction and real-time behavioral feature capture of multiple sensing video streams in the intelligent sensing edge terminal, the memory occupancy rate of the intelligent sensing edge terminal is continuously monitored in real time according to a preset sampling period. The memory occupancy data of each segment in the video stream decoding stage, feature extraction stage, trajectory tracking stage, and edge recognition operator inference stage of the intelligent sensing edge terminal, as well as the total memory occupancy data of the intelligent sensing edge terminal, are collected. The collected memory occupancy data are processed by moving average filtering to eliminate the monitoring error caused by instantaneous data fluctuations, thereby generating a continuous memory occupancy time series, which provides a stable and accurate quantitative monitoring basis for the subsequent identification of sensing load pressure.

[0066] Preferably, the generated video memory usage time sequence is used as input, and combined with the resource usage mapping relationship of the perception processing tasks corresponding to each sub-perception video stream, the perception load pressure identification processing for each sub-perception video stream is performed. First, based on the frame resolution, frame sampling frequency, and computational complexity of the corresponding perception processing task of each sub-perception video stream, the video memory usage weight coefficient corresponding to each sub-perception video stream is determined. Then, combined with the sub-item video memory usage data in the video memory usage time sequence, the single-channel perception load value corresponding to each sub-perception video stream and the comprehensive perception load pressure corresponding to all sub-perception video streams are calculated respectively. The comprehensive perception load pressure can quantitatively characterize the degree of occupation and load pressure of the current multi-channel perception video stream processing task on the video memory resources of the intelligent perception edge terminal.

[0067] Preferably, based on the hardware memory specifications of the intelligent sensing edge terminal, the minimum operating resource requirements of the edge recognition operator, and the minimum performance improvement requirements for multi-channel sensing video stream processing, corresponding preset load thresholds are set in advance. The preset load thresholds include a warning load threshold and an overload load threshold. The identified comprehensive sensing load pressure is compared with the warning load threshold and the overload load threshold respectively. When the comprehensive sensing load pressure exceeds the warning load threshold, a load warning notification is generated and the single-channel sensing load value of the corresponding sub-sensing video stream is recorded. When the comprehensive sensing load pressure exceeds the overload load threshold, subsequent frame extraction frequency reduction actions and computing power unloading request sending actions are triggered, thereby realizing the hierarchical judgment and precise triggering of the edge computing power load status.

[0068] Preferably, in response to the trigger condition that the overall sensing load pressure exceeds the preset load threshold, the frame extraction frequency reduction processing of the sensing data is performed. First, based on the spatial coverage area of ​​the sensing node corresponding to each sub-sensing video stream, the correlation with the intelligent agent's operation path, and the probability of historical abnormal behavior in the corresponding area, the sensing priority of each sub-sensing video stream is divided. Then, the frame extraction frequency of the corresponding sub-sensing video stream is reduced in a stepwise manner according to the sensing priority from low to high. After each reduction, the memory occupancy rate and overall sensing load pressure of the intelligent sensing edge terminal are re-monitored until the overall sensing load pressure falls back below the preset load threshold. At the same time, after the frame extraction frequency is reduced, a unified timestamp operator is used to re-perform frame-level synchronization calibration for each sub-sensing video stream to maintain the consistency of the time base of the multiple sensing video streams.

[0069] Preferably, while performing the frame rate reduction processing, a corresponding computing power offload request is generated. The computing power offload request encapsulates the unique device identifier of the intelligent sensing edge terminal, the current video memory occupancy rate data, the comprehensive sensing load pressure level, the executed frame rate reduction parameters, the behavior anomaly level corresponding to the current edge situational awareness metadata, and the timeliness requirements of business processing. Through a dedicated communication link between the intelligent sensing edge terminal and the intelligent agent management service platform, the computing power offload request is synchronously sent to the intelligent agent management service platform, providing a complete edge load status basis for the intelligent agent management service platform to adjust the generation strategy of cloud-based enhancement logic.

[0070] Preferably, after receiving a computing power offloading request, the intelligent agent management service platform parses the encapsulated data in the computing power offloading request to obtain the current computing power limitation, business priority, and timeliness requirements of the edge. Based on this, it performs priority adjustment processing on the cloud-based enhanced logic generation tasks, increasing the priority of cloud-based enhanced logic generation tasks corresponding to edge situational awareness metadata with high behavioral anomaly levels and high timeliness requirements, and decreasing the priority of cloud-based enhanced logic generation tasks corresponding to non-urgent and non-critical scenarios. At the same time, it performs lightweight optimization on high-priority cloud-based enhanced logic, eliminating unnecessary redundant reasoning content, and reducing the data volume of cloud-based enhanced logic and edge parsing computing power overhead while improving high-dimensional semantic guidance attributes.

[0071] Preferably, after reducing the frame extraction frequency and adjusting the priority of cloud-based enhancement logic generation, the memory usage and overall perception load pressure of the intelligent sensing edge terminal are continuously monitored. When the overall perception load pressure is found to be continuously lower than the preset recovery load threshold for a preset duration, the frame extraction frequency of each sub-sensing video stream is gradually restored to the initial set value in descending order of perception priority. At the same time, a computing power recovery notification is sent to the intelligent agent management service platform to restore the priority of cloud-based enhancement logic generation tasks. During the frame extraction frequency recovery process, the frame-level synchronization calibration of each sub-sensing video stream is maintained to ensure that the time reference consistency of multiple sensing video streams is not disrupted, thus forming a complete processing flow for adaptive adjustment of edge computing power.

[0072] Optionally, the method also includes communication link optimization processing actions: The edge situational awareness metadata is subjected to feature dimensionality reduction and compression processing using the unified interface service of intelligent agents to generate a bit situational payload containing key situational features. On the agent management service platform side, feature restoration processing is performed on the bit situation payload to reduce cloud-edge communication bandwidth consumption while maintaining the accuracy of cross-level semantic alignment processing.

[0073] Preferably, after the intelligent sensing edge terminal completes the generation of edge situational awareness metadata, the unified interface service of the intelligent agent deployed between the intelligent sensing edge terminal and the intelligent agent management service platform is invoked to perform structured parsing and compliance preprocessing on the generated edge situational awareness metadata. First, the completeness of the fields and the compliance of the format of the edge situational awareness metadata are verified, and invalid metadata with missing fields or disordered timing is removed. Then, the edge situational awareness metadata that has passed the verification is decomposed and classified. According to the data dimension and business purpose, the feature fields in the metadata are divided into situational core fields, auxiliary description fields and redundant filling fields, thereby generating a standardized and classified metadata feature set, which provides a clear-bounded and well-classified processing object for subsequent feature dimensionality reduction and compression processing.

[0074] Preferably, using the standardized and categorized metadata feature set as input, and combining the accuracy requirements of subsequent cross-level semantic alignment processing with the business needs of cloud strategy optimization processing, an importance weight quantification assessment is performed on each feature field in the metadata feature set. The core evaluation dimensions are the degree of influence of the feature field on the cross-level semantic alignment processing results and its business support value for cloud strategy optimization processing. A corresponding importance weight coefficient is assigned to each feature field, and then the feature fields in the metadata feature set are sorted in descending order of importance weight coefficient. Feature fields with importance weight coefficients higher than a preset weight threshold are selected as key situational features, and the remaining fields are classified as non-critical redundant features. This clarifies the core retention objects and the objects that can be removed in feature dimensionality reduction and compression.

[0075] Preferably, the unified interface service for intelligent agents takes the key situation features obtained from screening as the core of processing, performs feature dimensionality reduction and compression processing, first removes non-critical redundant features in the metadata feature set, eliminates invalid data volume that does not support the subsequent core processing flow, and then performs lightweight lossless encoding processing on the retained key situation features, mapping the high-dimensional structured feature data to a low-dimensional compact encoding space. Lossless compression encoding is used for the core situation features with the highest importance weight coefficient, and near-lossless compression encoding is used for the auxiliary key situation features with the second highest weight coefficient. Under the premise of completely preserving the semantic information of the key situation features, the volume of feature data is further compressed, and finally a compressed key situation feature encoding sequence is generated.

[0076] Preferably, the unified interface service of the intelligent agent performs standardized encapsulation processing on the compressed key situation feature encoding sequence. At the beginning of the encoding sequence, a globally unique identifier of the corresponding edge situation awareness metadata, a dimension mapping dictionary of key situation features, a data check code, and corresponding timestamp information are added. At the end of the encoding sequence, a transmission end identifier and an integrity check field are added to generate a bit situation payload containing key situation features. The bit situation payload adopts a binary bit stream format adapted to the cloud-edge communication protocol. Its data volume is significantly reduced compared to the original edge situation awareness metadata. At the same time, the encapsulated dimension mapping dictionary can completely represent the mapping relationship between key situation features and original metadata features, providing a complete decoding basis for subsequent feature restoration processing on the cloud side.

[0077] Preferably, the unified interface service for intelligent agents transmits the generated bit situation payload to the intelligent agent management service platform through a dedicated communication link between the intelligent sensing edge terminal and the intelligent agent management service platform. After receiving the bit situation payload, the intelligent agent management service platform first performs a transmission integrity check based on the verification field in the payload to verify whether the bit situation payload has experienced packet loss, code errors, or data tampering during transmission. For bit situation payloads that fail the verification, a retransmission request is sent to the intelligent sensing edge terminal. For bit situation payloads that pass the verification, decapsulation processing is performed to extract the key situation feature encoding sequence, dimension mapping dictionary, metadata globally unique identifier, and timestamp information, providing standardized input for subsequent feature restoration processing.

[0078] Preferably, the intelligent agent management service platform uses the key situation feature encoding sequence and dimension mapping dictionary obtained from decapsulation as processing objects, performs feature restoration processing, and performs inverse mapping decoding processing on the low-dimensional key situation feature encoding sequence according to the feature mapping relationship recorded in the dimension mapping dictionary to restore the restored situation feature data with the same dimension and complete semantic information as the original edge situation awareness metadata. Then, semantic consistency verification is performed on the restored situation feature data, and the semantic deviation value between the restored situation feature data and the key situation features in the original edge situation awareness metadata is calculated to ensure that the semantic deviation value is within the error range allowed by cross-level semantic alignment processing. In this way, while completely restoring the situation information on the edge side, the accuracy of subsequent cross-level semantic alignment processing is improved without being affected by the feature dimensionality reduction and compression process.

[0079] Preferably, the intelligent agent management service platform binds and stores the verified restored situational feature data with the globally unique identifier of the corresponding edge situational awareness metadata. At the same time, it uses the restored situational feature data as an effective input for cloud-based policy optimization processing, retrieves the local expert knowledge base to perform policy optimization processing based on business logic constraints, and generates corresponding cloud-based enhancement logic. The entire communication link optimization process, through feature dimensionality reduction compression and standardized encapsulation, significantly reduces the data transmission volume in the cloud-edge communication process while maintaining the accuracy of cross-level semantic alignment processing, and reduces the occupation of communication bandwidth. At the same time, through the standardized processing of the unified interface service of intelligent agents, the compatibility, stability and security of data interaction between the edge side and the cloud side are improved.

[0080] Optionally, the method also includes local logic fallback actions: Real-time detection of heartbeat signals sent down by cloud-enhanced logic; In response to the interruption of the heartbeat signal, the intelligent sensing edge terminal is triggered to execute local decisions based on a preset rule base, and the obtained local decision results are used to temporarily replace the autonomous decision-driving instructions of the intelligent agent.

[0081] Preferably, after establishing a cloud-edge communication link between the intelligent sensing edge terminal and the intelligent agent management service platform, continuous real-time heartbeat signal detection processing is performed on the downlink of the cloud-enhanced logic. The intelligent agent management service platform sends a dedicated heartbeat signal bound to the cloud-enhanced logic to the corresponding intelligent sensing edge terminal according to a preset sending cycle. The heartbeat signal encapsulates a globally unique identifier of the corresponding cloud-enhanced logic, a link status verification code, a downlink timing stamp, and link health parameters. The intelligent sensing edge terminal continuously listens to the dedicated heartbeat signal within a preset receiving window period and performs link status verification and timing continuity verification on each received heartbeat signal. This enables real-time monitoring of the downlink status of the cloud-enhanced logic throughout the entire time period, providing accurate status basis for subsequent link interruption determination.

[0082] Preferably, the intelligent sensing edge terminal, based on the real-time detected heartbeat signal reception status, executes the cloud-enhanced logic's heartbeat signal transmission interruption judgment processing. It pre-sets the maximum allowed timeout number of heartbeat signals and a continuous monitoring sliding window. When no valid heartbeat signal is received within a single reception window, it is marked as a single timeout event, and a timeout retransmission request is initiated to request heartbeat signal retransmission from the intelligent agent management service platform. When the number of single timeout events reaches the preset maximum allowed timeout number within the continuous monitoring sliding window, and the retransmission request does not receive a valid response, it is determined that the cloud-enhanced logic's heartbeat signal transmission is interrupted, and a link interruption alarm message is generated. This avoids misjudgment caused by instantaneous network fluctuations and improves the accuracy of triggering local logic fallback actions.

[0083] Preferably, a pre-built and stored rule library adapted to the target physical scene and the agent's operation type is constructed in advance within the intelligent sensing edge terminal. The pre-built rule library is constructed in a hierarchical and classified manner according to the type of abnormal behavior, the level of urgency of the scene, and the hardware execution capability of the agent. It includes standardized handling rules for all normal events and abnormal events in the target physical scene, security boundary constraints for agent action execution, autonomous decision-making logic in local scenes, and anomaly fallback handling process. Each handling rule has a clear triggering condition, execution action sequence, timing constraints, and security verification standard. Moreover, all handling rules in the pre-built rule library are fully matched with the agent's security behavior specification, providing a complete and compliant execution basis for local decision-making after link interruption.

[0084] Preferably, in response to the determination result of the interruption of the heartbeat signal sent by the cloud-enhanced logic, the local decision-making execution process of the intelligent sensing edge terminal is immediately triggered, the cloud-edge collaborative decision-making link based on the cloud-enhanced logic is suspended, and the edge situational awareness metadata generated in real time is retrieved. The edge situational awareness metadata is subjected to structured parsing to extract the behavioral anomaly dimension data, target motion vector distribution, and local action space constraint information. The parsed feature data is used as the matching basis to perform precise matching retrieval of the corresponding handling rules in the preset rule base, and the target handling rules that completely match the current scene state and behavioral anomaly type are selected. Based on this, the local decision result is generated.

[0085] Preferably, after generating the local decision result, compliance verification and executability verification are first performed on the local decision result. The verification includes whether the local decision result conforms to the agent's safe behavior specification, whether it is within the allowed boundaries of the local action space, and whether it is compatible with the agent's hardware execution capabilities. For the local decision result that passes the verification, it is encapsulated according to the standardized format of the agent's autonomous decision-driven instructions, and the timing constraints, state feedback trigger conditions, and exception fallback logic for action execution are supplemented to generate standardized local execution instructions. The generated local execution instructions temporarily replace the original agent autonomous decision-driven instructions and are sent to the agent's execution control module to drive the agent to continuously execute compliant actions that meet the current scenario requirements during the interruption of the cloud-edge communication link.

[0086] Preferably, during the period between local decision execution and local execution instruction substitution, the intelligent sensing edge terminal continuously sends link detection requests to the intelligent agent management service platform and continuously listens for the heartbeat signals sent by the cloud-based augmentation logic. When a valid and sequentially continuous heartbeat signal sent by the cloud-based augmentation logic is received again, it is determined that the cloud-edge communication link has returned to normal. Then, the link recovery verification and the integrity verification of the cloud-based augmentation logic are performed. After the verification is passed, the local decision-making process based on the preset rule base is terminated, the temporary substitution of the local execution instruction for the intelligent agent's autonomous decision-driven instruction is revoked, and the execution link of the intelligent agent's autonomous decision-driven instruction based on cloud-edge collaboration is smoothly switched back. At the same time, the local decision execution records, intelligent agent action execution status, and scene situation change data during the link interruption are encapsulated, stored, and uploaded to the intelligent agent management service platform.

[0087] Preferably, throughout the entire execution process of the local logic rollback action, the intelligent sensing edge terminal performs time-series recording of the heartbeat signal detection and recording, the link interruption judgment process, the matching and retrieval process of the preset rule base, the generation and verification data of local decision results, and the issuance and execution status data of local execution instructions. All recorded data is encapsulated into a link interruption event audit payload. The link interruption event audit payload contains the complete event triggering cause, decision execution link, action execution result and time sequence correlation. This provides a complete traceable data source for consistency verification in subsequent full-chain logic audit actions, ensuring that the decision execution process of the intelligent agent throughout its entire life cycle meets the security requirements of traceability and verifiability.

[0088] Optionally, the method also includes full-chain logical auditing actions: The input and output features during the execution of multidimensional feature extraction and real-time behavioral feature capture, strategy optimization processing, and driving load fusion calculation are encapsulated into a topology audit load that represents the logical dependencies of the decision-making link. The pre-defined traceable generation chain is invoked to perform consistency verification on the topology audit payload, so as to constrain the generation path of the agent's autonomous decision-driven instructions to match the agent's safety behavior specification.

[0089] Preferably, in the entire generation process of the intelligent agent's autonomous decision-making driving instructions, for the three core decision-making stages of multi-dimensional feature extraction and real-time behavior feature capture, policy optimization processing, and driving load fusion calculation, the input and output features are synchronously collected and processed throughout the entire time period. Based on a unified timestamp operator, the input multi-channel sensing video stream and output edge situational awareness metadata of the multi-dimensional feature extraction and real-time behavior feature capture stage are collected respectively; the input edge situational awareness metadata and output cloud-based enhancement logic of the policy optimization processing stage are collected respectively; and the input edge situational awareness metadata, cloud-based enhancement logic, alignment constraint vector, and output intelligent agent autonomous decision-making driving instructions of the driving load fusion calculation stage are collected respectively. At the same time, the model version, calculation parameters, timestamp information, and device unique identifier are collected during the execution of each stage. The format standardization and time alignment processing of all collected feature data are performed to eliminate the format differences and time misalignment problems of data in different stages, providing a complete, synchronous, and standardized basic data source for the subsequent encapsulation of topology audit payloads.

[0090] Preferably, the standardized input and output features of the entire process are used as the processing object. Topology mapping and topology audit payload encapsulation of the logical dependencies of the decision-making link are performed. First, according to the execution sequence of the agent's autonomous decision-making, the link topology relationship of the three core decision-making links is constructed. The upstream and downstream logical dependencies of the multi-dimensional feature extraction and real-time behavior feature capture link as the pre-input of the strategy optimization processing link, the strategy optimization processing link as the global strategy input of the driving payload fusion calculation link, and the driving payload fusion calculation link as the final decision output link are clarified. Then, the input and output features of each link are bound with the corresponding unique identifier of the link node and the upstream and downstream association identifier. According to the constructed link topology relationship, all feature data, node identifiers, association relationships, timing information and check codes are structurally encapsulated to generate a topology audit payload that represents the logical dependencies of the decision-making link. The topology audit payload can completely restore the entire link data flow path and logical association relationship from environmental perception to decision output, without any missing or truncated link information.

[0091] Preferably, a traceable generation chain is pre-constructed and stored based on the agent's security behavior specifications, the business management rules of the target physical scenario, and the standard execution process of cloud-edge collaborative decision-making. The traceable generation chain includes the standard link topology of the entire process of the agent's autonomous decision-making, the compliance input and output feature boundaries of each decision-making link, the logical dependency verification rules of upstream and downstream nodes, the security behavior constraint thresholds, and the violation judgment criteria. Each standard link node in the traceable generation chain is set with a unique compliance verification identifier and corresponding feature verification rules, and all rules and constraints are fully matched with the agent's security behavior specifications. The traceable generation chain is stored in the trusted storage area of ​​the agent management service platform and the local secure storage area of ​​the intelligent sensing edge terminal, respectively, providing a unified and compliant standard basis for the consistency verification of the subsequent topology audit payload.

[0092] Preferably, after the topology audit payload is encapsulated, a preset traceable generation chain is retrieved, and a full-dimensional consistency verification process is performed on the topology audit payload. First, the topology audit payload is decapsulated and structured parsed to extract the actual decision-making link topology, the input and output feature data of each link node, the upstream and downstream logical dependencies, and the full-link timing information. Then, based on the traceable generation chain as the standard, the link topology consistency verification is performed according to the execution sequence. The actual decision-making link topology is compared with the standard link topology in the traceable generation chain to confirm that there are no missing, reversed, or additional abnormal nodes in the flow order and dependencies of the link nodes. Next, the node feature compliance verification is performed to compare whether the actual input and output feature data of each link node is within the compliance boundaries specified by the traceable generation chain. Finally, the logical dependency consistency verification is performed to confirm that the input and output mapping relationship and logical flow between upstream and downstream nodes comply with the verification rules of the traceable generation chain. Finally, a full-link consistency verification report containing the verification results of each link and anomaly location information is generated.

[0093] Preferably, based on the results of the end-to-end consistency verification report, constraint processing is performed on the generation path of the agent's autonomous decision-driven instructions to ensure that the generation path fully matches the agent's safety behavior specification. For topology audit payloads that pass the end-to-end consistency verification, it is confirmed that the generation path of the corresponding agent's autonomous decision-driven instructions complies with the compliance requirements of the traceable generation chain and the agent's safety behavior specification. A compliance audit mark is added to the corresponding agent's autonomous decision-driven instructions, allowing the instructions to be normally sent to the agent's execution control module. For topology audit payloads that fail the end-to-end consistency verification, based on the anomaly location information in the verification report, the nodes in the decision-making link that have violations or anomalies are accurately located. Link anomaly alarm information is immediately generated, and the sending channel of the corresponding agent's autonomous decision-driven instructions is blocked to prevent the issuance of violations to the agent and avoid the execution of actions that do not comply with the safety behavior specification.

[0094] Preferably, for abnormal decision-making links that fail consistency verification, a safety fallback and generation path correction process is executed. While intercepting the autonomous decision-making driving instructions of the non-compliant intelligent agent, the safety fallback decision rules for the corresponding scenario are retrieved from the preset rule base in the intelligent sensing edge terminal. A fallback driving instruction that conforms to the intelligent agent's safe behavior specification is generated to temporarily replace the non-compliant autonomous decision-making driving instruction, thereby improving the safety of the intelligent agent's actions in abnormal states. At the same time, based on the location results of the abnormal nodes, the cause of the abnormality is traced back. For abnormalities in the feature extraction stage, the running status of the edge recognition operator and the feature extraction parameters are verified. For abnormalities in the strategy optimization stage, the compliance of the business rule slice and the correctness of the logical reconstruction process are verified. For abnormalities in the driving load fusion stage, the generation process of the alignment constraint vector and the compliance of the weighted fusion calculation parameters are verified. After completing the investigation of the cause of the abnormality, the parameters, models or rules of the corresponding stage are corrected to ensure that the decision generation path returns to the compliance range specified by the traceable generation chain.

[0095] Preferably, the entire process of full-chain logical auditing is performed with trusted recording and continuous optimization. This involves binding the topology audit payload, the retrieval records of the traceable generation chain, the full-chain consistency verification report, anomaly handling actions, and decision path correction records to the unique identifiers of the corresponding intelligent agent's autonomous decision-driven instructions. This is then subjected to time-sequential trusted encrypted storage, generating a complete full-chain audit archive for decision-making. The audit archive enables full-process backtracking of the generation path, logical dependencies, and compliance status of any intelligent agent's autonomous decision-driven instruction. Simultaneously, based on continuously accumulated audit archive data, the verification rules, compliance boundaries, and constraint thresholds in the traceable generation chain are iteratively optimized. This includes supplementing security behavior specifications for new scenarios, continuously improving the coverage and accuracy of consistency verification, and continuously strengthening the compliance constraint capabilities on the generation path of intelligent agent's autonomous decision-driven instructions.

[0096] Preferably, considering the architectural characteristics of cloud-edge collaboration, distributed audit collaborative processing is performed between the edge and the cloud. On the intelligent sensing edge terminal side, local audit fragments are generated in real time for the local execution processes of multi-dimensional feature extraction and real-time behavioral feature capture, cross-level semantic alignment processing, and driving payload fusion calculation, and are synchronized to the intelligent agent management service platform. At the same time, based on the lightweight and traceable generation chain of local storage, real-time compliance pre-verification is performed on the local decision-making process. Even when the cloud-edge communication link is interrupted or a local logic rollback action is triggered, consistency verification can still be performed on the local decision generation path, improving decision compliance in offline mode. On the intelligent agent management service platform side, the local audit fragments synchronized from the edge side are received, and combined with the audit data of the cloud policy optimization processing, a complete topology audit payload is generated, and consistency verification of the entire link is performed, realizing seamless connection and full-link audit coverage of audit data at both ends of the cloud and edge, adapting to various operating scenarios under the cloud-edge collaborative architecture.

[0097] like Figure 2 As shown, this is a cloud-edge collaborative intelligent agent autonomous decision-making driving device, which includes: The edge situation awareness module is used to acquire multi-channel sensing video streams that characterize the working environment of the intelligent agent, and to perform multi-dimensional feature extraction and real-time behavior feature capture on the multi-channel sensing video streams to generate edge situation awareness metadata that characterizes the local action space and includes the dimension of behavioral anomalies. The cloud-based enhanced logic module is used to upload the edge situational awareness metadata to the intelligent agent management service platform, and use the intelligent agent management service platform to retrieve the local expert knowledge base to perform policy optimization processing based on business logic constraints, so as to determine the cloud-based enhanced logic that represents the global policy space and has high-dimensional semantic guidance attributes. The decision-driven instruction module is used to perform cross-level semantic alignment processing on the edge situational awareness metadata and the cloud-based enhancement logic to construct an alignment constraint vector, and to perform driving load fusion calculation based on the alignment constraint vector to determine an agent autonomous decision-driven instruction that guides the agent's action output and has cloud-edge collaborative decision-making attributes.

[0098] Furthermore, in an optional implementation, the edge situational awareness module, when acquiring multiple sensing video streams: Multiple sensing nodes are positioned at different spatial locations within the target physical scene to synchronously capture multiple sub-sensing video streams representing the target physical scene; a unified timestamp operator is used to perform frame-level synchronization calibration on each sub-sensing video stream to generate the multi-channel sensing video stream.

[0099] Furthermore, in an optional implementation, the edge situation awareness module, when generating the edge situation awareness metadata: The intelligent sensing edge terminal performs trajectory tracking processing on the motion trajectories of each of the sub-sensing video streams in the multi-channel sensing video stream to determine the motion vector distribution of the target; based on the set edge recognition operator, edge situational awareness metadata representing the local action space and including the behavioral anomaly dimension is generated according to the motion vector distribution.

[0100] Furthermore, in an optional implementation, when determining the cloud enhancement logic, the cloud enhancement logic module: In the intelligent agent management service platform, the edge situational awareness metadata is converted into a semantic retrieval vector, and the semantic retrieval vector is used to retrieve the associated business rule slices from the local expert knowledge base; the business rule slices are subjected to logical reconstruction processing to determine cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes.

[0101] Furthermore, in an optional implementation, the decision-driven instruction module, when performing cross-level semantic alignment processing: Local actions matched by the edge situational awareness metadata are extracted, and inference processing is performed on the local actions according to the global policy space provided by the cloud enhancement logic to obtain the global semantic inference vector corresponding to the local actions; by calculating the semantic overlap entropy value between the global semantic inference vector and the global policy space, an alignment constraint vector for eliminating execution ambiguity is generated.

[0102] Furthermore, in an optional implementation, when the decision-driven instruction module determines the autonomous decision-driven instruction for the intelligent agent: The edge situation awareness metadata is subjected to gain adjustment using the alignment constraint vector to generate an edge driving component; the cloud driving component is extracted from the cloud enhancement logic, and the edge driving component and the cloud driving component are subjected to load-weighted fusion processing to generate an agent autonomous decision-making driving instruction that guides the agent's action output and has cloud-edge collaborative decision-making attributes.

[0103] Furthermore, in an optional implementation, the device further includes a computing power adjustment module, used for: The memory usage rate of the intelligent sensing edge terminal is monitored in real time, and the memory usage rate is used to identify the sensing load pressure for each of the sub-sensing video streams; in response to the sensing load pressure exceeding the preset load threshold, the frame extraction frequency of sensing data is reduced, and a computing power offloading request is sent simultaneously to adjust the generation priority of the cloud enhancement logic.

[0104] Furthermore, in an optional implementation, the apparatus further includes a link optimization module for: The edge situation awareness metadata is subjected to feature dimensionality reduction and compression processing using the unified interface service of the intelligent agent to generate a bit situation payload containing key situation features; the bit situation payload is subjected to feature restoration processing on the intelligent agent management service platform to reduce cloud-edge communication bandwidth consumption while maintaining the accuracy of the cross-level semantic alignment processing.

[0105] Furthermore, in an optional implementation, the device further includes a logic fallback module for: The system detects the heartbeat signals sent by the cloud-enhanced logic in real time; in response to the interruption of the heartbeat signals, it triggers the intelligent sensing edge terminal to execute local decisions based on a preset rule base, and uses the obtained local decision results to temporarily replace the autonomous decision-making driving instructions of the intelligent agent.

[0106] Furthermore, in an optional implementation, the apparatus further includes a decision audit module, used for: The input and output features during the execution of the multidimensional feature extraction and real-time behavior feature capture, the strategy optimization processing, and the driving load fusion calculation are encapsulated into a topology audit load that represents the logical dependency relationship of the decision link; a preset traceable generation chain is invoked to perform consistency verification on the topology audit load, so as to constrain the generation path of the agent's autonomous decision-making driving instructions to match the agent's safety behavior specification.

[0107] like Figure 3 The image shows an electronic device that includes a processor and a memory. The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the functions of each module of the cloud-edge collaborative intelligent agent autonomous decision-making driving device as described in claim 1, or implements the steps of the cloud-edge collaborative intelligent agent autonomous decision-making driving method, wherein the method includes: Step 1: Acquire multi-channel sensing video streams representing the agent's operating environment, perform multi-dimensional feature extraction and real-time behavior feature capture on the multi-channel sensing video streams to generate edge situational awareness metadata that represents the local action space and includes the dimension of behavioral anomalies. Step 2: Upload the edge situational awareness metadata to the intelligent agent management service platform, and use the intelligent agent management service platform to retrieve the local expert knowledge base to perform policy optimization processing based on business logic constraints, so as to determine the cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes. Step 3: Perform cross-level semantic alignment processing on edge situational awareness metadata and cloud-based augmentation logic to construct alignment constraint vectors, and perform driving load fusion calculations based on alignment constraint vectors to determine the autonomous decision-making driving instructions that guide the agent's action output and possess cloud-edge collaborative decision-making attributes.

[0108] A computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the aforementioned cloud-edge collaborative intelligent agent autonomous decision-making driving methods.

Claims

1. A cloud-edge collaborative intelligent agent autonomous decision-making driving method, characterized in that, include: Step 1: Acquire multi-channel sensing video streams that represent the working environment of the intelligent agent, and perform multi-dimensional feature extraction and real-time behavior feature capture on the multi-channel sensing video streams to generate edge situational awareness metadata that represents the local action space and includes the dimension of behavioral anomalies. Step 2: Upload the edge situational awareness metadata to the intelligent agent management service platform, and use the intelligent agent management service platform to retrieve the local expert knowledge base to perform policy optimization processing based on business logic constraints, so as to determine the cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes. Step 3: Perform cross-level semantic alignment processing on the edge situation awareness metadata and the cloud enhancement logic to construct an alignment constraint vector, and perform driving load fusion calculation based on the alignment constraint vector to determine the agent autonomous decision-making driving instruction that guides the agent's action output and has cloud-edge collaborative decision-making attributes. The multi-channel sensing video stream consists of multiple sub-sensing video streams that perform multi-view coverage of the target physical scene and have consistent time references. 2.The cloud-edge collaborative intelligent agent autonomous decision driving method of claim 1, wherein, Step 1, which involves acquiring the multi-channel sensing video stream, includes: Multiple sensing nodes arranged in different spatial locations within the target physical scene are retrieved to simultaneously capture multiple sub-sensing video streams representing the target physical scene; A unified timestamp operator is used to perform frame-level synchronization calibration on each of the sub-sensory video streams to generate the multi-channel sensing video stream. 3.The cloud-edge collaborative intelligent agent autonomous decision driving method of claim 1, wherein, The generation of the edge situation awareness metadata in step 1 includes: The intelligent sensing edge terminal is used to perform trajectory tracking processing on the motion trajectories of each of the sub-sensing video streams in the multi-channel sensing video stream to determine the motion vector distribution of the target. Based on the set edge recognition operator, edge situational awareness metadata representing the local action space and including the dimension of behavioral anomalies is generated according to the motion vector distribution. 4.The cloud-edge collaborative intelligent agent autonomous decision driving method of claim 1, wherein, Determining the cloud enhancement logic in step 2 includes: In the intelligent agent management service platform, the edge situational awareness metadata is converted into a semantic retrieval vector, and the semantic retrieval vector is used to retrieve the associated business rule slices from the local expert knowledge base; The business rule slices are subjected to logical reconstruction processing to determine cloud-based enhancement logic that represents the global policy space and has high-dimensional semantic guidance attributes.

5. The cloud-edge collaborative intelligent agent autonomous decision-making driving method according to claim 1, characterized in that, The action of performing cross-level semantic alignment processing in step 3 includes: Local actions matched by the edge situational awareness metadata are extracted, and inference processing is performed on the local actions according to the global policy space provided by the cloud enhancement logic to obtain the global semantic inference vector corresponding to the local actions; by calculating the semantic overlap entropy value between the global semantic inference vector and the global policy space, an alignment constraint vector for eliminating execution ambiguity is generated.

6. The cloud-edge collaborative intelligent agent autonomous decision-making driving method according to claim 5, characterized in that, Step 3, in which the autonomous decision-making driving instructions for the intelligent agent are determined, includes: Gain adjustment is performed on the edge situational awareness metadata using the alignment constraint vector to generate edge-driven components; The cloud-driven component is extracted from the cloud-enhanced logic, and the edge-driven component and the cloud-driven component are subjected to load-weighted fusion processing to generate an agent autonomous decision-making driving instruction that guides the agent's action output and has cloud-edge collaborative decision-making attributes.

7. The cloud-edge collaborative intelligent agent autonomous decision-making driving method according to claim 1, characterized in that, The method also includes edge computing power adaptive adjustment actions: The memory usage rate of the intelligent sensing edge terminal is monitored in real time, and the memory usage rate is used to identify the sensing load pressure for each of the sub-sensing video streams. In response to the sensing load pressure exceeding a preset load threshold, the frame extraction frequency of sensing data is reduced, and a computing power offloading request is sent simultaneously to adjust the generation priority of the cloud enhancement logic. 8.The cloud-edge collaborative intelligent agent autonomous decision driving method of claim 1, wherein, The method also includes communication link optimization processing actions: The edge situation awareness metadata is subjected to feature dimensionality reduction and compression processing using the unified interface service of the intelligent agent to generate a bit situation payload containing key situation features. On the intelligent agent management service platform side, feature restoration processing is performed on the bit situation payload to reduce cloud-edge communication bandwidth consumption while maintaining the accuracy of the cross-level semantic alignment processing. 9.The cloud-edge collaborative intelligent agent autonomous decision driving method of claim 1, wherein, The method also includes local logical fallback actions: Real-time detection of the heartbeat signals sent down by the cloud-enhanced logic; In response to the interruption of the heartbeat signal, the intelligent sensing edge terminal is triggered to execute a local decision based on a preset rule base, and the obtained local decision result is used to temporarily replace the autonomous decision-making driving instruction of the intelligent agent.

10. The cloud-edge collaborative intelligent agent autonomous decision driving method according to claim 1, characterized in that, The method also includes full-chain logical auditing actions: The input and output features during the execution of the multidimensional feature extraction and real-time behavioral feature capture, the strategy optimization process, and the driving load fusion calculation are encapsulated into a topology audit load that represents the logical dependencies of the decision link. A preset traceable generation chain is invoked to perform consistency verification on the topology audit payload, so as to constrain the generation path of the agent's autonomous decision-driven instructions to match the agent's security behavior specification.

Images