Semantic-driven cloud-edge collaborative causal inference method, device, equipment and medium

Abstract

The disclosure provides a semantic-driven cloud-edge collaborative causal reasoning method, device, equipment and medium, relating to the technical field of semantic communication. The scheme is applied to the cloud and includes: receiving multi-source abnormal semantic features from edge nodes; constructing global event correlation nodes according to the multi-source abnormal semantic features, in combination with spatial position correlation information and a pre-defined event knowledge graph; performing causal reasoning on the global event correlation nodes based on a timing constraint mechanism, and constructing a complete event causal evolution path by fusing similar event evolution patterns in a historical memory library; converting the complete event causal evolution path into a structured prompt input into a large language model, and generating an event reasoning result through multi-step thought chain reasoning. The application provides a structured semantic benchmark through edge nodes, and supplements cross-regional correlation and deep reasoning in the cloud, thereby realizing reliable decision-making for complex event causal reasoning under resource-constrained conditions.

Description

Technical Field

[0001] This disclosure relates to the intersection of communication technology and edge computing, specifically to the field of semantic communication technology, and in particular to semantic-driven cloud-edge collaborative causal reasoning methods, devices, equipment, and media. Background Technology

[0002] Semantic communication, as an emerging communication paradigm, demonstrates significant advantages in reducing bandwidth requirements and improving transmission efficiency by extracting and transmitting the core semantics of information rather than the complete bitstream. Unlike traditional communication, which pursues bit-level precise replication, semantic communication focuses on the "meaning" level of information delivery. It achieves collaborative design of source coding and channel transmission through end-to-end joint optimization, providing a new technical path for reliable communication under bandwidth-constrained and poor channel conditions.

[0003] However, applying existing semantic communication methods to cloud-edge collaborative complex event monitoring scenarios still faces significant technical limitations. Complex security events involve behavioral sequences across multiple monitoring areas and time points, exhibiting unique business characteristics such as "fragmented distribution of semantic information and close cross-regional spatiotemporal correlation." This means that a single edge node can only capture abnormal semantic fragments within its local field of view, while the complete causal evolution chain of the event requires the cloud to perform cross-regional correlation and deep inference on multi-source fragmented semantics. However, existing technologies employ a generalized monitoring strategy of uploading the entire video stream, independent analysis of single nodes, and static threshold alarms. This leads to excessive consumption of communication bandwidth, lack of cross-regional semantic correlation, and uninterpretable causal inference. These technical shortcomings limit the inference accuracy and decision reliability of existing methods in cloud-edge collaborative complex event monitoring scenarios. Summary of the Invention

[0004] This disclosure provides a semantic-driven cloud-edge collaborative causal reasoning method, apparatus, device, and medium.

[0005] According to one aspect of this disclosure, a semantic-driven cloud-edge collaborative causal reasoning method is provided, applied in the cloud, the method comprising: Receive multi-source anomaly semantic features from edge nodes; Based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, a global event association node is constructed. Based on the temporal constraint mechanism, causal reasoning is performed on the global event-related nodes, and similar event evolution patterns in the historical memory bank are integrated to construct a complete event causal evolution path; The complete causal evolution path of the event is transformed into a structured prompt input large language model, and the event reasoning result is generated through multi-step thinking chain reasoning.

[0006] According to another aspect of this disclosure, a semantically driven cloud-edge collaborative causal reasoning apparatus is provided, comprising: The receiving module is used to receive multi-source anomaly semantic features from edge nodes; The first construction module is used to construct global event association nodes based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph; The second construction module is used to perform causal reasoning on the global event-related nodes based on the temporal constraint mechanism, and to integrate the evolution patterns of similar events in the historical memory bank to construct a complete event causal evolution path. The reasoning module is used to transform the complete causal evolution path of the event into a structured prompt input large language model, and generate event reasoning results through multi-step thinking chain reasoning.

[0007] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described in any of the above technical solutions.

[0008] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any one of the methods described above.

[0009] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in any one of the above technical solutions.

[0010] This disclosure provides a semantic-driven cloud-edge collaborative causal reasoning method, apparatus, device, and storage medium. By receiving multi-source anomaly semantic features from edge nodes, this application avoids bandwidth waste and computing bottlenecks caused by transmitting the entire original video stream, enabling the cloud to directly perform deep causal analysis based on structured semantics. Next, based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, a global event association node is constructed. This overcomes the limitations of a single node's perspective, integrating fragmented anomaly clues across regions into a unified semantic association network. This solves the technical shortcomings of existing technologies, which only monitor low-level hardware faults of single devices and struggle to extract semantic-level associations of complex cross-regional business events. Furthermore, based on a temporal constraint mechanism, causal reasoning is performed on the global event association node, and similar event evolution patterns from a historical memory bank are integrated. Through temporal logic consistency screening and weighted correction of historical patterns, a complete event causal evolution path with spatiotemporal coherence and empirical support is constructed, overcoming the shortcomings of traditional methods that only perform static association comparisons and cannot reconstruct the dynamic evolution process of events. Furthermore, implicit causal logic is made explicit through multi-step thought chain reasoning, and the chain reasoning capability of the large language model is used to generate interpretable event reasoning results, avoiding the lack of decision-making transparency caused by black-box prediction. Finally, the complete event causal evolution path is transformed into a structured prompt input large language model, ensuring standardized integration of reasoning input with knowledge graphs and historical memory, enabling the receiving end to perform joint reasoning and decision-making based on a unified semantic format. Therefore, this application, by providing a structured semantic benchmark at edge nodes and supplementing cross-regional associations and deep reasoning in the cloud, achieves reliable decision-making for complex event causal reasoning under resource-constrained conditions, effectively overcoming the technical shortcomings of traditional methods such as insufficient utilization of semantic information, blind cross-regional associations, and uninterpretable reasoning results.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a schematic diagram of the steps of a semantically driven cloud-edge collaborative causal reasoning method in one embodiment of this disclosure; Figure 2 This is a schematic diagram of the overall process of the semantic-driven cloud-edge collaborative causal reasoning method in one embodiment of this disclosure; Figure 3 This is a block diagram illustrating the principle of the semantic-driven cloud-edge collaborative causal reasoning device in this embodiment of the present disclosure. Figure 4This is a block diagram of an electronic device used to implement the semantic-driven cloud-edge collaborative causal reasoning method of the embodiments of this disclosure. Detailed Implementation

[0013] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0014] This disclosure provides a semantic-driven cloud-edge collaborative causal reasoning method, which is applied in the cloud. See [link to relevant documentation]. Figure 1 As shown, it includes: Step S101: Receive multi-source anomaly semantic features from edge nodes.

[0015] Specifically, "receiving multi-source anomaly semantic features from edge nodes" refers to the cloud acquiring pre-processed structured semantic information from multiple edge nodes deployed in different monitoring areas, rather than raw video streams or sensor data. Here, "edge nodes" refer to lightweight computing devices deployed in local monitoring areas, responsible for front-end data acquisition and preliminary analysis; "multi-source" indicates that multiple edge nodes report simultaneously, covering monitoring areas in different spatial locations; "anomaly semantic features" are structured information extracted by edge nodes after predictively perceiving raw video or sensor data using pre-inference models, including visual semantics (such as anomaly event type, target attributes), timestamps (time of anomaly occurrence), and spatial location (coordinates of the area where the anomaly occurred). The cloud receives these structured feature nodes through a communication interface as the input basis for subsequent cross-regional semantic aggregation and causal inference, thereby avoiding bandwidth waste and computing power bottlenecks caused by transmitting the full amount of raw data.

[0016] Step S102: Based on the semantic features of multi-source anomalies, combined with spatial location association information and a predefined event knowledge graph, construct global event association nodes.

[0017] Specifically, "multi-source anomaly semantic features" refers to the structured anomaly information filtered out by edge nodes after predictively sensing video streams or sensor data of local monitoring areas through lightweight pre-inference models. This includes the visual semantics of the anomaly events (such as event type and target attributes), timestamps of occurrence, and spatial coordinates. "Spatial location association information" is the association data used to determine the spatial relationships between nodes of each anomaly semantic feature, such as a space-time channel map, which records the topological layout of each monitoring area and the physical distance and accessibility between nodes. "Predefined event knowledge graph" is a structured knowledge base pre-built in the cloud, containing event types, device identifiers, business objects, and causal rules.

[0018] The specific implementation process of this scheme includes: First, extracting visual semantics, timestamps, and spatial location information from multi-source anomaly semantic features to obtain structured anomaly semantic feature nodes. Second, querying spatial location association information based on spatial location information to determine the spatial association relationships between nodes (such as adjacent areas or different perspectives within the same scene). Then, aligning the nodes temporally based on timestamp information to determine the temporal association relationships between nodes (such as concurrency or sequential occurrence). Next, matching the structured anomaly semantic feature nodes with entity nodes in the event knowledge graph to determine the node type mapping relationship (such as mapping "forklift speeding" to the "violation of driving event" type in the knowledge graph). Finally, based on spatial association relationships, temporal association relationships, and node type mapping relationships, a graph neural network is used to aggregate the structured anomaly semantic feature nodes to obtain global event association nodes that integrate cross-regional spatiotemporal associations and semantic types, thereby integrating fragmented local anomaly clues into a unified semantic association network.

[0019] Step S103: Based on the temporal constraint mechanism, perform causal reasoning on the global event-related nodes and integrate the evolution patterns of similar events in the historical memory bank to construct a complete event causal evolution path.

[0020] Specifically, the "temporal constraint mechanism" is a set of rules that ensures the logical consistency of the event sequence. Its core requirement is that the timestamp of the cause node must be earlier than the timestamp of the result node, and the time interval between the two must meet a predefined evolution threshold (e.g., the time difference between "forklift making an illegal U-turn" and "container overturning" should be within a reasonable physical reaction time). The "historical memory bank" is a library of similar event evolution patterns that stores the types of causal nodes, spatial transfer paths, and evolution time distributions in historical events, used to provide empirical references for current causal reasoning. The "complete event causal evolution path" is a causal chain with spatiotemporal coherence formed by piecing together the global event-related nodes in the order of event development.

[0021] The specific implementation process of this scheme includes: First, extracting the timestamp information of each node from the global event association nodes to obtain the event time sequence. Second, determining the temporal relationship and time interval between each node based on the event time sequence. Then, matching the temporal relationship and time interval with the causal rules in the predefined event knowledge graph to filter candidate causal association pairs that meet the temporal logic consistency (e.g., excluding false associations with reversed times or intervals exceeding the threshold). Next, querying the evolution patterns of similar events in the historical memory bank based on the candidate causal association pairs, and weighting and correcting the causal strength of the candidate causal association pairs (e.g., increasing the confidence of the current association if historical data shows that a certain type of event has a high causal probability in a specific scenario). Finally, based on the weighted causal association pairs, concatenating each global event association node according to the chronological order of event development to generate a complete event causal evolution path, thereby overcoming the shortcomings of traditional methods that only perform static association mining and cannot reconstruct the dynamic evolution process of events.

[0022] Step S104: The complete causal evolution path of the event is transformed into a structured prompt input large language model, and the event reasoning result is generated through multi-step thinking chain reasoning.

[0023] Specifically, "structured prompts" refer to input text that organizes key information from the complete causal evolution path of an event into a format understandable by the large language model. This typically includes role settings, background information, reasoning tasks, and output format requirements. "Multi-step reasoning chain" refers to the chain-like reasoning process executed step-by-step by the large language model. By explicitly presenting intermediate reasoning steps, implicit causal logic is transformed into interpretable explicit conclusions, avoiding black-box prediction. "Event reasoning results" are the final decision output generated by comprehensively considering risk level, causal strength, and evolutionary trends, including event characterization, risk scoring, and handling recommendations.

[0024] The specific implementation process of this scheme includes: First, extracting key entities (such as "forklift A" and "hazardous chemical tank B"), entity attributes (such as speed, location, and type of cargo), and causal relationships (such as "speeding → inability to avoid collision") from the complete causal evolution path of the event to obtain structured path information. Second, converting the structured path information into prompt text in the form of natural language description, such as "At time T1, forklift A illegally made a U-turn at speed V in area X, resulting in a collision with container C in area Y at time T2. Please analyze the risk level and evolution trend of this event." Then, the model input prompts are fed into the large language model, triggering a multi-step reasoning chain. This reasoning process specifically includes: first, identifying the entity type and risk level corresponding to the key entities (e.g., determining "forklift speeding" as a high-risk violation); second, assessing the causal strength and evolutionary trend between the key entities (e.g., analyzing whether a collision might lead to a hazardous chemical leak); and third, synthesizing the risk level, causal strength, and evolutionary trend to generate an event reasoning result (e.g., outputting "This event is a high-risk chain accident; it is recommended to immediately activate the emergency response and dispatch drones to verify the leak risk"). Thus, the implicit causal logic is made explicit using the chain reasoning capability of the large language model, outputting interpretable reasoning conclusions.

[0025] This disclosure provides a semantic-driven cloud-edge collaborative causal reasoning method, apparatus, device, and storage medium. By receiving multi-source anomaly semantic features from edge nodes, this application avoids bandwidth waste and computing bottlenecks caused by transmitting the entire original video stream, enabling the cloud to directly perform deep causal analysis based on structured semantics. Next, based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, a global event association node is constructed. This overcomes the limitations of a single node's perspective, integrating fragmented anomaly clues across regions into a unified semantic association network. This solves the technical shortcomings of existing technologies, which only monitor low-level hardware faults of single devices and struggle to extract semantic-level associations of complex cross-regional business events. Furthermore, based on a temporal constraint mechanism, causal reasoning is performed on the global event association node, and similar event evolution patterns from a historical memory bank are integrated. Through temporal logic consistency screening and weighted correction of historical patterns, a complete event causal evolution path with spatiotemporal coherence and empirical support is constructed, overcoming the shortcomings of traditional methods that only perform static association comparisons and cannot reconstruct the dynamic evolution process of events. Furthermore, implicit causal logic is made explicit through multi-step thought chain reasoning, and the chain reasoning capability of the large language model is used to generate interpretable event reasoning results, avoiding the lack of decision-making transparency caused by black-box prediction. Finally, the complete event causal evolution path is transformed into a structured prompt input large language model, ensuring standardized integration of reasoning input with knowledge graphs and historical memory, enabling the receiving end to perform joint reasoning and decision-making based on a unified semantic format. Therefore, this application, by providing a structured semantic benchmark at edge nodes and supplementing cross-regional associations and deep reasoning in the cloud, achieves reliable decision-making for complex event causal reasoning under resource-constrained conditions, effectively overcoming the technical shortcomings of traditional methods such as insufficient utilization of semantic information, blind cross-regional associations, and uninterpretable reasoning results.

[0026] In some optional embodiments, a global event association node is constructed based on multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, including: Visual semantics, timestamps, and spatial location information are extracted from multi-source anomaly semantic features to obtain structured anomaly semantic feature nodes; Based on spatial location information, query spatial location association information to determine the spatial association relationships between each structured anomaly semantic feature node; Based on the timestamp information, the temporal alignment of each structured anomaly semantic feature node is performed to determine the temporal correlation between each structured anomaly semantic feature node; The structured anomaly semantic feature nodes are matched with entity nodes in a predefined event knowledge graph to determine the node type mapping relationship; Based on spatial correlation, temporal correlation, and node type mapping, a graph neural network is used to aggregate structured anomaly semantic feature nodes to obtain global event-related nodes.

[0027] Specifically, "multi-source anomaly semantic features" refers to structured anomaly information from multiple edge nodes. Each edge node uses a lightweight pre-inference model to predictively perceive video streams or sensor data from a local monitoring area, filtering out the visual semantics of anomaly events (such as event type, target attributes), timestamps, and spatial coordinates. "Spatial location association information" is association data used to determine the spatial relationships between nodes, such as a space-time channel map, which records the topological layout of each monitoring area and the physical distance and accessibility between nodes. "Predefined event knowledge graph" is a structured knowledge base pre-built in the cloud, containing event types, device identifiers, business objects, and causal rules. "Graph neural network" is a deep learning model specifically designed to process graph-structured data, achieving high-order relationship modeling by aggregating the feature information of nodes and their neighbors.

[0028] The specific implementation process of this scheme includes: First, extracting visual semantics, timestamps, and spatial location information from the multi-source anomaly semantic features uploaded by each edge node to obtain multiple structured anomaly semantic feature nodes. Each node contains a complete semantic description of "what anomaly, when it occurred, and where it occurred." Second, querying spatial location association information based on the spatial location information of each node to determine the spatial relationship between nodes, such as determining whether two nodes are located in adjacent areas or different perspectives of the same scene. Then, performing temporal alignment based on the timestamp information of each node to determine the temporal relationship between nodes, such as determining whether two anomalies are concurrent events or causal precursors with a sequential order. Next, matching the structured anomaly semantic feature nodes with entity nodes in the event knowledge graph to determine the node type mapping relationship, such as mapping "forklift speeding" to the "illegal driving event" type in the knowledge graph, and mapping "hazardous chemical leak" to the "safety accident event" type. Finally, based on spatial correlation, temporal correlation, and node type mapping, a graph neural network is used to aggregate structured anomaly semantic feature nodes. The graph neural network treats each node as a graph vertex and spatial, temporal, and semantic correlations as edges. Through a multi-layer message passing mechanism, the feature information of neighboring nodes is aggregated to obtain global event correlation nodes that integrate cross-regional spatiotemporal correlation and semantic type, thereby integrating fragmented local anomaly clues into a unified semantic correlation network.

[0029] In this way, by extracting visual semantics, timestamps, and spatial location information from multi-source anomaly semantic features to obtain structured anomaly semantic feature nodes, the unstructured raw data uploaded by edge nodes is transformed into standardized semantic units containing "anomaly type, occurrence time, and spatial coordinates." This avoids the bandwidth waste and computing power bottlenecks caused by directly transmitting the full video stream, enabling the cloud to perform in-depth analysis based on high-dimensional semantic information, significantly reducing communication load and cloud preprocessing overhead. Secondly, by querying spatial location association information based on spatial location information to determine the spatial relationship between nodes, the limitations of the physical field of view of a single monitoring node are overcome. Previously fragmented local anomaly clues are associated according to the actual geographical topology. For example, a spatial adjacency relationship is established between two seemingly independent events, "forklift speeding in area A of the quay crane" and "container displacement in area B of the yard," thereby identifying the potential spatial transmission path of cross-regional chain accidents and solving the technical deficiency of traditional single-point monitoring in capturing the spatial spread of complex events. Furthermore, based on timestamp information, the temporal relationships of each node are determined through time-series alignment. Strict temporal chronology and concurrency analysis eliminate false anomaly combinations with no logical connection in the temporal dimension. For example, it distinguishes between two different event patterns: "multi-regional concurrent anomalies at the same time" and "causal transmission anomalies between different times," providing a reliable temporal constraint basis for subsequent causal inference and avoiding the technical problem of misclassifying time-independent events in static association analysis. Further, structured anomaly semantic feature nodes are matched with entity nodes in a predefined event knowledge graph to determine node type mapping relationships. Using predefined semantic specifications such as event types, device identifiers, and business objects in the knowledge graph, the original semantics extracted from edge nodes are standardized, categorized, and semantically enhanced. For example, the edge node description "forklift driving too fast" is mapped to "violation of driving event—speeding subclass" in the knowledge graph, unifying and resolving semantic differences generated by heterogeneous edge devices in the cloud, improving the consistency and understandability of multi-source data fusion. Finally, based on spatial, temporal, and node type mapping relationships, a graph neural network is used to aggregate structured anomalous semantic feature nodes to obtain global event-related nodes. Utilizing the multi-layer message passing mechanism of the graph neural network, three heterogeneous relationships—spatial adjacency, temporal sequence, and semantic similarity—are uniformly encoded into a high-dimensional embedded representation of the graph structure. This achieves deep fusion of fragmented semantic information across regions, time periods, and types within a unified topological space. Thus, a semantic association network capable of reflecting the global association status of complex events is constructed, overcoming the technical shortcomings of traditional methods that merely perform simple data stacking and cannot uncover high-order implicit associations.

[0030] In some optional embodiments, causal reasoning is performed on global event-related nodes based on a temporal constraint mechanism, and similar event evolution patterns in the historical memory bank are integrated to construct a complete event causal evolution path, including: Extract the timestamp information of each node from the global event association nodes to obtain the event time sequence; Based on the event time sequence, determine the temporal relationship and time interval between the global event-related nodes; Match the temporal sequence and time intervals with the causal rules in the predefined event knowledge graph to filter candidate causal relationship pairs that satisfy the temporal logic consistency; Based on candidate causal association pairs, the evolution patterns of similar events in the historical memory bank are queried, and the causal strength of the candidate causal association pairs is weighted and corrected to obtain weighted causal association pairs. Based on weighted causal pairs, global event-related nodes are pieced together in chronological order of event development to generate a complete causal evolution path.

[0031] Specifically, "Global Event Association Nodes" are fused nodes obtained by aggregating multi-source anomaly semantic features in the cloud using graph neural networks. Each node contains the visual semantics, timestamp, spatial location, and type mapping information of the anomaly event in the knowledge graph. "Event Time Sequence" is a set of global event association nodes arranged chronologically, used to represent the temporal evolution of anomaly events. "Predefined Event Knowledge Graph" is a structured knowledge base pre-built in the cloud, containing event types, device identifiers, business objects, and causal rules. These causal rules specify the allowed causal relationships between specific event types and their corresponding time interval thresholds. "Historical Memory Bank" is a library of similar event evolution patterns storing causal node types, spatial transfer paths, and evolutionary time distributions from historical events, providing empirical references for current causal reasoning. "Complete Event Causal Evolution Path" is a spatiotemporally coherent causal chain formed by piecing together global event association nodes according to the event's development timeline, capable of reconstructing the complete dynamic evolution process of an event from cause to effect.

[0032] The specific implementation process of this scheme includes: First, extracting the timestamp information of each node from the global event association nodes, and sorting them according to time sequence to obtain the event time sequence, for example, obtaining the sequence "forklift illegal U-turn (T1) → Unable to avoid (T2) → Container collision (T3) → Hazardous chemical leak warning (T4)". Second, determining the time sequence and time interval between each node based on the event time sequence, for example, calculating the time difference between T1 and T2. The time difference between T2 and T3 is Then, the temporal sequence and time intervals are matched with the causal rules in the predefined event knowledge graph to filter candidate causal relationships that satisfy temporal logical consistency. For example, the knowledge graph specifies a reasonable time interval threshold of [0, 30 seconds] for "illegal U-turn leading to collision". If the condition is met within seconds, then... Minutes exceeding the threshold are excluded. Next, based on the evolution patterns of similar events in the historical memory database, the causal strength of candidate causal association pairs is weighted and adjusted. For example, if the historical memory database records "in a port operation scenario, the causal probability of a collision after an illegal U-turn is 0.85," then the initial causal strength of the current candidate association is multiplied by this historical weight coefficient to obtain a weighted causal association pair. Finally, based on the weighted causal association pairs, the global event association nodes are concatenated according to the chronological order of event development to generate a complete event causal evolution path, for example, a complete causal chain of "illegal U-turn (cause) → insufficient avoidance (intermediate state) → collision (effect) → leakage warning (secondary risk)." Thus, by using temporal constraints to filter out logically contradictory associations and by using historical memory weighting to improve the confidence of empirical associations, a complete event causal evolution path with spatiotemporal coherence and empirical credibility is ultimately constructed.

[0033] In this way, by extracting the timestamp information of each node from the global event-related nodes to obtain the event time sequence, the nodes that were originally scattered in the global related network are reorganized according to the time dimension, establishing the time main line of the evolution of abnormal events. This allows the cloud to break free from the constraints of static graph structures and examine the development of events from the perspective of time evolution, providing a clear time coordinate benchmark for subsequent causal inference. Secondly, based on the event time sequence, the temporal relationship and time interval between each global event-related node are determined. By quantitatively calculating the time difference between adjacent event nodes, the abstract "sequential occurrence" is transformed into a specific numerical interval. For example, it is precisely determined that there is a 15-second interval between "forklift illegal U-turn" and "container collision," providing a calculable objective basis for the threshold matching of subsequent causal rules, avoiding the arbitrariness of judging the rationality of time based solely on subjective experience. Furthermore, the temporal sequence and time intervals are matched with causal rules in a predefined event knowledge graph to filter candidate causal associations that meet the temporal logic consistency. By using predefined causal rules in the knowledge graph (such as "the timestamp of the cause node must be earlier than the result node and the interval must meet the evolution threshold"), a massive number of candidate associations are rigidly constrained and filtered, effectively eliminating false associations with reversed time or intervals exceeding the threshold. For example, logically contradictory associations such as "the collision occurred before the turn" or "the interval between the two events exceeds the physical reaction limit" are eliminated, significantly reducing noise interference and false alarm rate in causal reasoning. Next, based on the evolution patterns of similar events in the historical memory bank, the causal strength of the candidate causal association pairs is weighted and corrected. The confidence of the current candidate association is dynamically adjusted using causal experience accumulated in the historical memory bank for similar scenarios (e.g., "the historical probability of a collision caused by an illegal U-turn in a port operation scenario is 0.85"). This allows causal reasoning to not only rely on isolated information from a single event but also integrate statistical patterns from long-term historical data, thereby improving the robustness and scenario adaptability of causal judgments and avoiding inference bias due to insufficient samples in rare and abnormal scenarios. Finally, based on the weighted causal association pairs, the global event association nodes are pieced together according to the chronological order of event development to generate a complete event causal evolution path. The high-confidence associations, after being filtered by temporal constraints and corrected by historical weighting, are linked together along the timeline into a complete chain of "cause → transmission → result → secondary risk," for example, a coherent path of "illegal U-turn → failure to avoid → collision → leak warning." This gives the originally fragmented abnormal nodes a dynamically evolving logical framework, outputting a complete picture of the event with spatiotemporal coherence and causal interpretability.

[0034] In some optional embodiments, the complete causal evolution path of the event is transformed into a structured prompt input large language model, and the event reasoning result is generated through multi-step thought chain reasoning, including: Key entities, entity attributes, and causal relationships are extracted from the complete causal evolution path of an event to obtain structured path information; The path structure information is converted into prompt text in the form of natural language description to obtain model input prompts; Inputting model input prompts into the large language model triggers the large language model to execute multi-step thought chain reasoning and obtain the event reasoning result.

[0035] Specifically, the "complete event causal evolution path" is a spatiotemporally coherent causal chain formed by piecing together the relevant nodes of each global event in chronological order, containing complete dynamic evolution information of the abnormal event from cause to effect. "Key entities" refer to the specific objects involved in the path (e.g., forklift A, container C, hazardous chemical tank B), "entity attributes" refer to the characteristic descriptions of the objects (e.g., speed, location, type of cargo), and "causal relationships" refer to the causal transmission logic between entities (e.g., "speeding → inability to avoid collision → collision"). "Path structured information" is an intermediate representation that organizes key entities, entity attributes, and causal relationships in a structured data format, facilitating subsequent conversion into natural language. "Multi-step reasoning chain" refers to the chain-like reasoning process executed step-by-step by the large language model, explicitly presenting intermediate reasoning steps to transform implicit causal logic into interpretable explicit conclusions.

[0036] The specific implementation process of this scheme includes: First, extracting key entities, entity attributes, and causal relationships from the complete causal evolution path of the event to obtain structured path information. For example, extracting "Entity: Forklift A, Attributes: Speed ​​20km / h, Location X region; Entity: Container C, Attributes: Weight 30 tons, Location Y region; Relationship: Forklift A speeding → Unable to avoid collision → Collision with Container C". Second, converting the structured path information into prompt text in the form of natural language description to obtain model input prompts. For example, constructing a prompt statement such as "At time T1, forklift A illegally made a U-turn at a speed of 20km / h in region X, resulting in a collision with container C in region Y at time T2. Please analyze the risk level and evolution trend of this event". Then, the model input prompts are fed into the large language model, triggering it to execute a multi-step reasoning chain. This reasoning process specifically includes: first, identifying the entity type and risk level corresponding to the key entities (e.g., determining "forklift speeding" as a high-risk violation); second, assessing the causal strength and evolutionary trend between the key entities (e.g., analyzing whether a collision might lead to a hazardous chemical leak); and third, synthesizing the risk level, causal strength, and evolutionary trend to generate an event reasoning result (e.g., outputting "This event is a high-risk chain accident; it is recommended to immediately activate the emergency response and dispatch drones to verify the leak risk"). Thus, the implicit causal logic is made explicit using the chain reasoning capability of the large language model, outputting interpretable reasoning conclusions.

[0037] In this way, by extracting key entities, entity attributes, and causal relationships from the complete causal evolution path of an event, structured path information is obtained. Implicit causal knowledge embedded in complex graph structures is transformed into explicit and standardized data representations, enabling the large language model to directly understand the core elements and logical relationships of an event, avoiding semantic ambiguity and information redundancy caused by unstructured raw data input. By converting the structured path information into prompt text in natural language description form, and leveraging the large language model's deep understanding of natural language, technically structured data is transformed into an expression form that conforms to human cognitive habits, lowering the cognitive threshold for model input. Simultaneously, carefully designed prompt templates guide the model to focus on key reasoning dimensions. By inputting prompts into the large language model, multi-step reasoning is triggered, allowing the model to execute a chain-like reasoning process step-by-step: "identifying entity types and risk levels → assessing causal strength and evolutionary trends → comprehensively generating reasoning results." This breaks down traditional black-box prediction into traceable and verifiable intermediate steps. The final event reasoning result not only includes the conclusion itself but also a complete chain of reasoning logic, significantly improving decision-making transparency and interpretability.

[0038] In some optional embodiments, the method further includes: Extract the reasoning confidence of the multi-step thought chain reasoning process from the event reasoning results, calculate the prediction variance of the reasoning confidence, and obtain the degree of uncertainty; Compare the degree of uncertainty with a preset trigger threshold; If the level of uncertainty exceeds a preset trigger threshold, it is determined that the current causal reasoning lacks key evidence, triggering a multi-agent collaborative verification process.

[0039] Specifically, "Inference Confidence" refers to the probability value or confidence score output by the large language model for each intermediate inference step (such as entity type recognition and causal strength assessment) during the multi-step reasoning process, reflecting the model's degree of confidence in the current inference conclusion. "Prediction Variance" is a statistical metric used to quantify the fluctuation of inference confidence across multiple samples or different inference paths; a larger variance indicates lower certainty about the conclusion. "Uncertainty Level" is a numerical indicator obtained by calculating the prediction variance, characterizing the credibility level of the current causal inference result. "Preset Trigger Threshold" is a pre-set upper limit for uncertainty tolerance; when the uncertainty level exceeds this threshold, the inference conclusion is unreliable. "Multi-Agent Collaborative Verification Process" is an active perception mechanism that schedules multiple edge agents in the cloud to perform supplementary evidence collection, used to eliminate inference uncertainty.

[0040] The specific implementation process of this scheme includes: First, extracting the reasoning confidence of the multi-step thought chain reasoning process from the event reasoning results. For example, extracting the confidence of the first step "identifying forklift A as a violation of driving rules" (0.92), the confidence of the second step "assessing the causal strength of the collision leading to leakage as moderate" (0.65), and the confidence of the third step "generating a comprehensive conclusion of a high-risk accident" (0.71). Second, calculating the prediction variance of the reasoning confidence to obtain the degree of uncertainty, for example, calculating the variance value of the confidence of the above three steps. A larger variance indicates more drastic fluctuations in confidence levels at each step and poor overall inference consistency. Then, the degree of uncertainty is compared with a preset trigger threshold, for example, by setting a threshold. If calculated Then determine Finally, in response to an uncertainty level exceeding a preset trigger threshold, the system determines that the current causal reasoning lacks crucial evidence (e.g., the confidence level of the second-step causal strength assessment is low, possibly due to a lack of direct observational data from the leaked area). This triggers a multi-agent collaborative verification process, where the cloud schedules the optimal combination of edge nodes to the target area to perform supplementary data collection, thereby obtaining crucial evidence to eliminate uncertainty. This achieves a transformation from passive monitoring to proactive evidence-based verification, automatically activating agents to collaboratively collect evidence when confidence levels are insufficient, ensuring the reliability of the reasoning results.

[0041] In this way, by extracting the reasoning confidence of the multi-step reasoning process from the event reasoning results, it is possible to penetrate the final output surface of the large language model and deeply track the degree of certainty of each intermediate reasoning step. For example, it can identify situations where the confidence of the "causal strength assessment" step is low while the confidence of the "entity recognition" step is high, thereby accurately locating the weak links in the reasoning chain and avoiding the one-sided judgment that relies solely on the final conclusion score while ignoring the reliability of the process. By calculating the prediction variance of the reasoning confidence, the degree of uncertainty is obtained, quantifying the fluctuation of the confidence of multiple steps into a single numerical indicator, objectively reflecting the consistency and stability of the model's overall reasoning. The larger the variance, the higher the dispersion of the confidence of each step and the worse the self-consistency of the reasoning logic, providing the system with a calculable and comparable decision-making basis. By comparing the degree of uncertainty with a preset trigger threshold, a clear automated decision boundary is established, enabling the system to determine the reliability of reasoning based on quantitative indicators rather than subjective experience, and realizing the standardization and configurability of trigger conditions. By responding to uncertainty levels exceeding a preset threshold, the system determines that the current causal reasoning lacks key evidence and triggers a multi-agent collaborative verification process. When the confidence level of the reasoning is insufficient, the system automatically activates a supplementary evidence collection mechanism. For example, when the confidence level of the "leakage risk" assessment fluctuates too much due to a lack of direct on-site observation, the system immediately dispatches drones or nearby edge nodes to conduct verification, rather than blindly outputting high-risk warnings that lead to false alarms or conservatively ignoring the real risks that lead to missed alarms.

[0042] In some optional embodiments, triggering a multi-agent collaborative verification process includes: The active perception task is modeled as a Markov decision process. The state of the Markov decision process includes the field of view of each edge node, the computing load status, the communication channel bandwidth and the queue of tasks to be verified. The action includes the task allocation scheme for assigning evidence collection tasks to each edge node. A comprehensive reward function is constructed, which is a weighted combination of the estimated benefit of task completion quality, communication transmission cost, and repeated collection penalty. The estimated benefit of task completion quality is determined based on the multi-perspective complementarity of each edge node. Based on the comprehensive reward function, the task allocation scheme is evaluated using a reinforcement learning algorithm to select the target task allocation scheme with the greatest long-term cumulative benefit. The target task allocation scheme is analyzed to obtain the optimal combination of edge nodes and the corresponding scheduling instructions.

[0043] Specifically, a "Markov Decision Process" is a mathematical model of sequential decision-making, consisting of a state space, an action space, state transition probabilities, and a reward function. Its core characteristic is that the next state depends only on the current state and action, and is independent of historical states. "Field of view" refers to the physical coverage area of ​​the monitoring equipment at each edge node, determining the range of spatial information it can collect. "Computing load status" refers to the resource occupancy of the currently assigned computing tasks at each edge node, reflecting its remaining processing capacity. "Communication channel bandwidth" refers to the available data transmission rate between each edge node and the cloud, affecting the timeliness of evidence feedback. "Pending verification task queue" refers to the set of multiple evidence collection tasks currently awaiting assignment in the system; "Task allocation scheme" refers to the combination strategy for assigning specific evidence collection tasks to each edge node; "Multi-perspective complementarity" refers to the set of non-repetitive, mutually corroborating observation perspectives that different edge nodes can provide due to spatial location differences; the greater the difference in perspectives, the higher the evidence complementarity. The "comprehensive reward function" is a multi-objective utility function used to evaluate the merits of task allocation schemes; the "long-term cumulative return" refers to the cumulative discounted value considering the returns of multiple future decisions, reflecting the global optimality of the task allocation scheme.

[0044] The specific implementation process of this scheme includes: First, the active perception task is modeled as a Markov decision process. The state space is defined as a joint representation of the field of view, computing power load, communication channel bandwidth, and queue of tasks to be verified for each edge node. The action space is defined as all possible task allocation schemes for assigning evidence collection tasks to each edge node. Second, a comprehensive reward function is constructed. This function is a weighted combination of the estimated benefit of task completion quality, communication transmission cost, and duplicate collection penalty. The estimated benefit of task completion quality is determined based on the multi-view complementarity of each edge node. For example, if node A covers the front of the leak area and node B covers the side, the combination of the two can obtain three-dimensional evidence. Higher complementarity results in greater benefits. The communication transmission cost is calculated based on the distance between the node and the cloud and the channel bandwidth. The duplicate collection penalty provides negative incentives for redundant situations where multiple nodes cover the same view. Then, based on the comprehensive reward function, the value of the task allocation scheme is evaluated using a reinforcement learning algorithm (such as multi-agent reinforcement learning). The state-action value function is iteratively updated to select the target task allocation scheme with the maximum long-term cumulative benefit. Finally, the target task allocation scheme is analyzed to obtain the optimal combination of edge nodes and corresponding scheduling instructions, such as outputting specific instructions like "Schedule node A to perform frontal leak collection, node B to perform lateral corroboration, and node C to stand by." Thus, a dynamic balance scheduling that achieves optimal evidence complementarity and lowest cost is realized under conditions of limited communication resources.

[0045] Specifically, reinforcement learning algorithms can include the following optional implementation methods: Method 1: Multi-Agent Proximal Policy Optimization Algorithm: This algorithm adopts an actor-critic architecture, which is directly adapted to the multi-agent collaborative scheduling scenario of this application. The policy network takes the joint state representation of the field of view, computing power load, communication channel bandwidth, and queue of tasks to be verified of each edge node as input, and outputs the action probability distribution for allocating evidence collection tasks to each edge node. For example, it selects the scheme of "node A performs frontal collection and node B performs lateral corroboration" with a probability of 0.7. The value network takes the same state representation as input and outputs the estimated long-term cumulative benefit that the system expects to obtain under this scheme. The policy network limits the update range through a proximal pruning mechanism to prevent multiple nodes from repeatedly collecting or missing key perspectives due to drastic changes in task allocation. The value network estimates the state value through a multi-step bootstrapping method and performs gradient updates by combining the estimated benefit of task completion quality based on multi-perspective complementarity, communication transmission cost, and repeated collection penalty in the comprehensive reward function. Through alternating iterations until the policy converges, the task allocation scheme with the optimal long-term cumulative benefit that "maximizes evidence complementarity and minimizes communication cost and repeated penalty" is selected.

[0046] Method 2: Multi-Agent Q-Learning Algorithm: This algorithm maintains a deep Q-network for a cloud-based global scheduler. It takes the joint state representation of each edge node as input and outputs the Q-value of each candidate task allocation scheme. The current state, the selected task allocation scheme, the comprehensive reward obtained after execution, and the next state after transition are combined into an experience sample and stored in a replay buffer. The comprehensive reward is calculated based on the quality of the verification results fed back by the nodes after actual execution, the actual communication overhead, and whether there is view overlap. The Q-value is updated using a temporal difference method, enabling the network to gradually learn to accurately evaluate the true long-term value of different allocation schemes in complex state combinations such as "node A has ample bandwidth but its view overlaps with node C" and "node B has limited bandwidth but provides a unique view." Finally, the task allocation scheme with the largest Q-value is selected, and the optimal edge node combination and scheduling instructions are obtained through analysis.

[0047] Method 3: Multi-Agent Actor-Critic Algorithm: This algorithm's policy network takes the joint state representation as input and directly outputs a deterministic task allocation scheme, such as "scheduling node A to collect data from region X, node B to provide lateral evidence from region Y, and node C on standby." The value network takes the state and deterministic actions as joint input and outputs the expected long-term cumulative benefit under this allocation scheme. The policy network is updated through a deterministic policy gradient method, adjusting the task allocation strategy along the gradient direction indicated by the value network: "improving the predicted benefit of task completion quality, reducing communication and transmission costs, and reducing penalties for repeated collection." The value network is updated by minimizing the error between the predicted benefit and the actual cumulative discounted benefit. The dual-network collaborative iteration enables the cloud scheduler to gradually master how to dynamically allocate multi-agent collection tasks to achieve global optimum under different network loads and task urgency levels.

[0048] All three methods mentioned above use the estimated benefits of task completion quality based on multi-perspective complementarity, communication transmission costs, and penalties for repeated collection as real-time feedback signals to drive continuous optimization of task allocation strategies. Ultimately, a dynamic balance is achieved between evidence collection quality, communication resource consumption, and collection redundancy, ensuring efficient execution of multi-agent collaborative verification under resource-constrained conditions.

[0049] In this way, by modeling the proactive perception task as a Markov decision process and defining the state as a joint representation of the field of view, computing load, communication channel bandwidth, and queue of tasks to be verified of each edge node, the system can perceive the global resource situation in real time. For example, it can identify heterogeneous states such as node A having sufficient computing power but limited bandwidth, and node B having sufficient bandwidth but not covering the target area. This provides comprehensive and dynamic state input for subsequent task allocation, avoiding the shortcomings of traditional fixed assignment strategies that cannot adapt to dynamic changes in the network environment. Secondly, by defining the action as a task allocation scheme for assigning evidence collection tasks to each edge node, and constructing a comprehensive reward function that is a weighted combination of the estimated benefit of task completion quality, communication transmission cost, and duplicate collection penalty, a quantitative evaluation framework for multi-objective trade-offs is realized. The estimated benefit of task completion quality is determined based on the multi-view complementarity of each edge node. For example, scheduling two nodes covering the front and side of the leak area can obtain three-dimensional evidence. The higher the complementarity, the greater the benefit. At the same time, the communication transmission cost constrains the blind scheduling of long-distance high bandwidth consumption, and the duplicate collection penalty suppresses the resource waste of multiple nodes covering the same view. The weighted combination of the three factors allows the system to pursue evidence quality while taking into account economy and efficiency. Furthermore, by using a reinforcement learning algorithm based on a comprehensive reward function to evaluate the value of task allocation schemes and select the target task allocation scheme with the maximum long-term cumulative benefit, the system not only considers the immediate benefit of the current single-step decision, but also evaluates the global return of future multi-step decisions through a discount accumulation mechanism. For example, the current scheduling node A may have slightly higher short-term costs, but it can create favorable conditions for the collaborative collection of subsequent node B, thereby obtaining higher long-term cumulative benefits, realizing the decision upgrade from local optimum to global optimum. By analyzing the target task allocation scheme to obtain the optimal combination of edge nodes and the corresponding scheduling instructions, the abstract mathematical optimization results are transformed into executable scheduling strategies, such as outputting specific instructions such as "schedule node A to perform positive leakage collection, node B to perform lateral evidence collection, and node C not to be assigned due to full computing power," ensuring that the optimized scheme is accurately implemented in the physical system. Thus, the above scheme effectively overcomes the technical defects of traditional scheduling methods that allocate tasks based on only a single indicator (such as bandwidth or distance), leading to repeated collection from multiple perspectives or omission of key evidence. Under resource-constrained conditions, it achieves a dynamic balance of optimal evidence complementarity, lowest communication cost, and minimal collection redundancy, significantly improving the scheduling efficiency and evidence quality of multi-agent collaborative perception in complex scenarios.

[0050] In some optional embodiments, the method further includes: According to the scheduling instructions, control the optimal edge node combination to execute the evidence collection task and receive the verification results fed back by the optimal edge node combination; The complete causal evolution path of the event is updated based on the verification results, and the multi-step thought chain reasoning is re-executed based on the updated complete causal evolution path of the event until the uncertainty level is lower than the preset trigger threshold.

[0051] Specifically, the "scheduling instruction" is a specific execution command generated by the cloud after analyzing the target task allocation scheme through reinforcement learning algorithms. It includes the identity identifier of the optimal edge node combination, the evidence collection task content to be executed by each node (such as collection area, collection angle, sensor type), and feedback time limit requirements. The "optimal edge node combination" is the set of edge nodes with the greatest long-term cumulative benefits selected through value assessment. Each node forms an optimal configuration in terms of field of view, computing power load, communication bandwidth, and multi-view complementarity. The "verification result" is supplementary data fed back to the cloud after the optimal edge node combination executes the evidence collection task, including newly collected visual semantics, timestamps, spatial location, and on-site environmental parameters. The "complete event causal evolution path" is a causal chain with spatiotemporal coherence formed by piecing together the global event-related nodes according to the event development time sequence. The "multi-step thought chain reasoning" is a chain-like reasoning process executed step by step by the large language model, including identifying entity types and risk levels, assessing causal strength and evolution trend, and comprehensively generating event reasoning results. "Uncertainty level" is a quantitative indicator obtained by calculating the variance of the reasoning confidence, which represents the credibility level of the current causal reasoning result; "Preset trigger threshold" is the upper limit of uncertainty tolerance set in advance by the system.

[0052] The specific implementation process of this scheme includes: First, according to the scheduling instructions, the optimal edge node combination is controlled to execute the evidence collection task. The cloud sends the parsed scheduling instructions to each optimal edge node. After receiving the instructions, the nodes adjust their sensor parameters, move to the target area to perform secondary focusing collection and multi-dimensional corroboration. For example, node A goes to the front of the leak area to collect high-definition images, and node B collects thermal imaging data from the side. After the collection is completed, the verification results are sent back to the cloud. Second, the verification results fed back by the optimal edge node combination are received, and the cloud integrates the supplementary data from multiple nodes to form a complete evidence set. Then, the complete event causal evolution path is updated according to the verification results. For example, the newly collected visual semantics of "liquid seepage exists in the leak area" and its corresponding timestamp and spatial location are inserted into the original path to correct or strengthen the causal relationship strength of "collision → leakage". Next, based on the updated complete event causal evolution path, multi-step thought chain reasoning is re-executed. The large language model uses the updated path as input to re-perform entity recognition, causal evaluation, and comprehensive reasoning. Finally, it is determined whether the uncertainty level after re-reasoning is lower than the preset trigger threshold. If it is still higher than the threshold, the collaborative verification process is triggered again. If it is lower than the threshold, the final causal reasoning result with an interpretable logical chain and an early warning instruction are output. Thus, through the closed-loop iterative mechanism of "collection-feedback-update-re-reasoning", reasoning uncertainty is gradually eliminated until the confidence level requirement is met.

[0053] In this way, by controlling the optimal combination of edge nodes to execute evidence collection tasks according to scheduling instructions, the abstract optimization scheme output by the reinforcement learning algorithm is transformed into precise execution in the physical world. For example, node A is instructed to collect evidence directly in the leak area, while node B provides corroboration from the side. This ensures the spatial coverage and complementary perspectives of supplementary evidence collection, avoiding the technical defect of the scheduling scheme being disconnected from the execution implementation. By receiving the verification results fed back by the optimal combination of edge nodes, the cloud obtains first-hand on-site data from the optimal node configuration, rather than relying on historical data or indirect inference, ensuring the timeliness and relevance of supplementary evidence. Secondly, by updating the complete causal evolution path of the event according to the verification results, the newly collected visual semantics, timestamps, and spatial locations are integrated into the original causal chain in real time. For example, new evidence of "liquid seepage" is inserted into the "collision → leakage" correlation link, correcting or strengthening the strength of the corresponding causal edges. This allows the causal path to evolve dynamically as the evidence becomes richer, overcoming the rigidity defect of traditional methods where the path is fixed once generated and cannot absorb new information. Furthermore, by re-executing multi-step reasoning based on the updated complete causal evolution path of the event, the large language model uses the evidence-enhanced path as input to re-identify entities, assess causality, and perform comprehensive reasoning. For example, the confidence level of the second step, "leakage risk assessment," increased from 0.65 to 0.91 due to the addition of direct observational evidence, and the reasoning conclusion gradually converged as the quality of evidence improved. Finally, by cyclically executing the above update and re-reasoning process until the uncertainty level falls below a preset trigger threshold, an automated closed loop of "uncertainty quantification → proactive evidence collection → evidence fusion → reasoning optimization" was established. The system only triggers verification when the confidence level is insufficient and terminates iteration when the confidence level meets the threshold, avoiding the waste of resources caused by indiscriminate continuous collection and eliminating the decision-making risk of blindly outputting conclusions when evidence is insufficient. In summary, the above solution effectively overcomes the technical defects of traditional monitoring systems that involve one-time reasoning and the inability to perform closed-loop correction, realizing a paradigm shift from "single prediction" to "evidence-based iteration." While ensuring the real-time nature of monitoring, it continuously improves the reliability of decision-making through dynamic evidence supplementation and reasoning iteration, significantly enhancing the robustness of causal reasoning and the accuracy of early warning in complex scenarios.

[0054] In some optional embodiments, the method further includes: Based on the event reasoning results, the prior information of the scene is adaptively updated to obtain the updated prior model parameters; The updated prior model parameters are distributed to the edge nodes to drive the dynamic evolution of the edge nodes' predictive perception capabilities.

[0055] Specifically, "scene prior information" refers to the experiential knowledge maintained in the cloud that describes the distribution patterns of abnormal events and prediction models under specific monitoring scenarios. This includes prior probabilities of event types, spatiotemporal correlation patterns, and prediction model parameters. "Prior model parameters" are a quantitative representation of the scene prior information, such as the frequency weights of various abnormal events in port operation scenarios and the probability matrix of abnormal transmission between different areas. "Adaptive update" refers to the process of automatically adjusting prior model parameters to match the dynamic changes of the real scene based on feedback from actual inference results and prediction deviations. "Predictive perception" refers to the ability of edge nodes to perform forward-looking anomaly screening of local monitoring areas by combining scene prior information. Prior guidance enables nodes to prioritize high-probability anomaly types, improving perception efficiency.

[0056] The specific implementation process of this scheme includes: First, adaptively updating the prior information of the scene based on the event inference results. The cloud compares the current event inference result (such as the complete causal chain of "illegal U-turn leading to collision and leakage") with the prior information of the scene before the update, and calculates the prediction deviation (for example, the prior model predicts a leakage probability of 0.3 in this scene, while the actual inference confirmation probability is 0.8). Based on this deviation, the prior model parameters are adjusted through adaptive algorithms such as Bayesian update or gradient descent to obtain the updated prior model parameters. For example, the prior weight of the "illegal U-turn" event type in this scene is increased, and the probability of abnormal transmission from region X to region Y is enhanced. Second, the updated prior model parameters are distributed to the edge nodes. The cloud transmits the parameter file to each edge node through a wireless channel. Then, the edge nodes load the updated prior model parameters and replace the local old parameters to drive the dynamic evolution of the edge nodes' predictive perception capabilities. For example, in subsequent perception, the nodes improve the detection sensitivity of "illegal U-turn" behavior and prioritize associating anomalies in region X with region Y for early warning, so that the edge prediction model is continuously optimized as the scene evolves. Thus, through a closed-loop mechanism of cloud-based inference feedback and edge parameter updates, continuous accumulation of scene knowledge and dynamic evolution of edge perception capabilities are achieved.

[0057] In this way, by adaptively updating the prior information of the scene based on the reasoning results of the event, the cloud compares the causal chain confirmed by the current reasoning with the original prior model, and automatically adjusts the parameters of the prior model based on the prediction deviation. For example, if the probability of "illegal U-turn leading to collision and leakage" confirmed by the actual reasoning is higher than the prior prediction, the prior weight of this causal path is increased. This transforms the prior information of the scene from a static experience base into a dynamic knowledge body that continuously learns with real events, overcoming the technical defects of traditional fixed prior models that cannot adapt to scene evolution and gradually become ineffective. By distributing the updated prior model parameters to edge nodes, a knowledge feedback channel is established between deep reasoning in the cloud and front-end perception at the edge, enabling edge nodes to obtain the latest scene cognition extracted from the cloud from a global perspective in a timely manner, rather than relying on outdated local experience in the long term. By driving the dynamic evolution of the predictive perception capabilities of edge nodes with updated prior model parameters, edge nodes improve their sensitivity to detecting high-probability anomalies and optimize their ability to predict the transmission of anomalies across regions in subsequent perception. For example, after loading updated parameters, nodes prioritize establishing an association warning between "forklift speeding" and "hazardous chemical area", realizing a perception paradigm upgrade from "passively discovering anomalies" to "actively predicting risks", significantly improving the accuracy and timeliness of initial anomaly screening and reducing the uploading of invalid data.

[0058] For a better overall understanding of the technical solution of this application, please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic flowchart illustrating the overall process of a semantic-driven cloud-edge collaborative causal reasoning method according to an embodiment of this disclosure. The flowchart demonstrates the complete workflow of the semantic-driven cloud-edge collaborative causal reasoning method, and its logical flow is as follows: At the edge, monitoring devices such as intelligent agents A / B / C are deployed. Each intelligent agent inputs the collected video stream or sensor data into a lightweight pre-inference module. This module combines the scene prior information sent from the cloud to perform predictive perception, and after filtering out abnormal semantic features, it uploads them to the cloud.

[0059] After receiving abnormal semantic features in the cloud, the system first integrates fragmented semantics from multiple sources by combining a cross-regional semantic aggregation module with a spatiotemporal channel map and an event knowledge graph. The aggregation results are then input into a deep causal reasoning and temporal constraint module, which integrates historical event evolution patterns from memory backtracking to uncover complete causal chains. The causal reasoning results are then input into a large language model thinking chain module for multi-step reasoning to generate event reasoning results and risk scores.

[0060] Whether the inference result enters the uncertainty quantification judgment node satisfies ( If the confidence level is low (Yes), the reinforcement learning collaborative scheduling module is triggered. This module generates the optimal scheduling scheme based on a comprehensive evaluation of complementarity, communication cost, and repetition penalty, and issues active perception / secondary collection instructions to the edge agent. After the agent performs supplementary collection, it sends back new evidence. The cloud updates the causal reasoning path and re-executes the thought chain reasoning until the uncertainty is lower than the threshold. If the confidence level is high (No), the final risk score and warning are directly output.

[0061] Meanwhile, the deep causal reasoning and temporal constraint module adaptively updates the prior information of the scene based on the reasoning results, and sends the updated parameters to the lightweight pre-reasoning module at the edge, driving the dynamic evolution of edge perception capabilities and forming a closed-loop collaborative system of "edge perception → cloud reasoning → knowledge evolution → edge optimization".

[0062] The following describes an apparatus embodiment of this application, which can be used to execute the semantic-driven cloud-edge collaborative causal reasoning method in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the semantic-driven cloud-edge collaborative causal reasoning method described above.

[0063] This disclosure also provides a semantically driven cloud-edge collaborative causal reasoning device 300, such as... Figure 3 As shown, it includes: The receiving module 301 is used to receive multi-source anomaly semantic features from edge nodes; The first construction module 302 is used to construct global event association nodes based on multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph; The second construction module 303 is used to perform causal reasoning on global event-related nodes based on the temporal constraint mechanism, and to integrate similar event evolution patterns in the historical memory bank to construct a complete event causal evolution path. The reasoning module 304 is used to transform the complete causal evolution path of an event into a structured prompt input large language model, and generate event reasoning results through multi-step thinking chain reasoning.

[0064] In some optional embodiments, the first construction module 302 constructs global event association nodes based on multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, including: Visual semantics, timestamps, and spatial location information are extracted from multi-source anomaly semantic features to obtain structured anomaly semantic feature nodes; Based on spatial location information, query spatial location association information to determine the spatial association relationships between each structured anomaly semantic feature node; Based on the timestamp information, the temporal alignment of each structured anomaly semantic feature node is performed to determine the temporal correlation between each structured anomaly semantic feature node; The structured anomaly semantic feature nodes are matched with entity nodes in a predefined event knowledge graph to determine the node type mapping relationship; Based on spatial correlation, temporal correlation, and node type mapping, a graph neural network is used to aggregate structured anomaly semantic feature nodes to obtain global event-related nodes.

[0065] In some optional embodiments, the second construction module 303 performs causal reasoning on global event-related nodes based on a temporal constraint mechanism, and integrates similar event evolution patterns in the historical memory bank to construct a complete event causal evolution path, including: Extract the timestamp information of each node from the global event association nodes to obtain the event time sequence; Based on the event time sequence, determine the temporal relationship and time interval between the global event-related nodes; The temporal sequence and time intervals are matched with the causal rules in the predefined event knowledge graph to filter candidate causal relationship pairs that satisfy the temporal logic consistency. Based on candidate causal association pairs, the evolution patterns of similar events in the historical memory bank are queried, and the causal strength of the candidate causal association pairs is weighted and corrected to obtain weighted causal association pairs. Based on weighted causal pairs, global event-related nodes are pieced together in chronological order of event development to generate a complete causal evolution path.

[0066] In some optional embodiments, the reasoning module 304 transforms the complete causal evolution path of the event into a structured prompt input large language model, and generates event reasoning results through multi-step thought chain reasoning, including: Key entities, entity attributes, and causal relationships are extracted from the complete causal evolution path of an event to obtain structured path information; The path structure information is converted into prompt text in the form of natural language description to obtain model input prompts; Inputting model input prompts into the large language model triggers the large language model to execute multi-step thought chain reasoning and obtain the event reasoning result.

[0067] In some optional embodiments, the inference module 304 is further configured to: Extract the reasoning confidence of the multi-step thought chain reasoning process from the event reasoning results, calculate the prediction variance of the reasoning confidence, and obtain the degree of uncertainty; Compare the degree of uncertainty with a preset trigger threshold; If the level of uncertainty exceeds a preset trigger threshold, it is determined that the current causal reasoning lacks key evidence, triggering a multi-agent collaborative verification process.

[0068] In some optional embodiments, the inference module 304 triggers a multi-agent collaborative verification process, including: The active perception task is modeled as a Markov decision process. The state of the Markov decision process includes the field of view of each edge node, the computing load status, the communication channel bandwidth and the queue of tasks to be verified. The action includes the task allocation scheme for assigning evidence collection tasks to each edge node. A comprehensive reward function is constructed, which is a weighted combination of the estimated benefit of task completion quality, communication transmission cost, and repeated collection penalty. The estimated benefit of task completion quality is determined based on the multi-perspective complementarity of each edge node. Based on the comprehensive reward function, the task allocation scheme is evaluated using a reinforcement learning algorithm to select the target task allocation scheme with the greatest long-term cumulative benefit. The target task allocation scheme is analyzed to obtain the optimal combination of edge nodes and the corresponding scheduling instructions.

[0069] In some optional embodiments, the inference module 304 is further configured to: According to the scheduling instructions, control the optimal edge node combination to execute the evidence collection task and receive the verification results fed back by the optimal edge node combination; The complete causal evolution path of the event is updated based on the verification results, and the multi-step thought chain reasoning is re-executed based on the updated complete causal evolution path of the event until the uncertainty level is lower than the preset trigger threshold.

[0070] In some optional embodiments, the inference module 304 is further configured to: Based on the event reasoning results, the prior information of the scene is adaptively updated to obtain the updated prior model parameters; The updated prior model parameters are distributed to the edge nodes to drive the dynamic evolution of the edge nodes' predictive perception capabilities.

[0071] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0072] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0073] Figure 4A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0074] like Figure 4 As shown, the electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0075] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0076] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as a semantically driven cloud-edge collaborative causal reasoning method. For example, in some embodiments, the semantically driven cloud-edge collaborative causal reasoning method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps described above can be performed. Alternatively, in other embodiments, computing unit 401 may be configured to perform a semantically driven cloud-edge collaborative causal reasoning method by any other suitable means (e.g., by means of firmware).

[0077] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0078] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0079] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0080] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0081] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0082] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0083] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0084] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A semantically driven cloud-edge collaborative causal reasoning method, wherein, Applied to the cloud, the method includes: Receive multi-source anomaly semantic features from edge nodes; Based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, a global event association node is constructed. Based on the temporal constraint mechanism, causal reasoning is performed on the global event-related nodes, and similar event evolution patterns in the historical memory bank are integrated to construct a complete event causal evolution path; The complete causal evolution path of the event is transformed into a structured prompt input large language model, and the event reasoning result is generated through multi-step thinking chain reasoning.

2. The method according to claim 1, wherein, The step of constructing global event association nodes based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph, includes: Visual semantics, timestamps, and spatial location information are extracted from the multi-source anomaly semantic features to obtain structured anomaly semantic feature nodes; Based on the spatial location information, query the spatial location association information to determine the spatial association relationship between each structured anomaly semantic feature node; Based on the timestamp information, the structured anomaly semantic feature nodes are time-series aligned to determine the temporal correlation between them. The structured anomaly semantic feature nodes are matched with entity nodes in the predefined event knowledge graph to determine the node type mapping relationship; Based on the spatial association, the temporal association, and the node type mapping, the structured anomaly semantic feature nodes are aggregated using a graph neural network to obtain the global event association nodes.

3. The method according to claim 1, wherein, The causal reasoning of the global event-related nodes based on the temporal constraint mechanism, and the integration of similar event evolution patterns in the historical memory bank, construct a complete event causal evolution path, including: Extract the timestamp information of each node from the global event association nodes to obtain the event time sequence; Based on the event time sequence, determine the temporal relationship and time interval between each global event-related node; The temporal sequence and time interval are matched with the causal rules in the predefined event knowledge graph to filter candidate causal association pairs that satisfy temporal logic consistency. Based on the candidate causal association pairs, the evolution patterns of similar events in the historical memory bank are queried, and the causal strength of the candidate causal association pairs is weighted and corrected to obtain weighted causal association pairs. Based on the weighted causal relationship pairs, the global event association nodes are spliced ​​together in the chronological order of event development to generate the complete event causal evolution path.

4. The method according to claim 1, wherein, The process of transforming the complete causal evolution path of the event into a structured prompt input large language model, and generating event reasoning results through multi-step thought chain reasoning, includes: Key entities, entity attributes, and causal relationships are extracted from the complete causal evolution path of the event to obtain structured path information. The path structure information is converted into prompt text in the form of natural language description to obtain model input prompts; The input prompts from the model are fed into the large language model, triggering the large language model to perform multi-step thought chain reasoning and obtain the event reasoning result.

5. The method according to any one of claims 1 to 4, wherein, The method further includes: Extract the reasoning confidence of the multi-step thinking chain reasoning process from the event reasoning results, calculate the prediction variance of the reasoning confidence, and obtain the degree of uncertainty; The degree of uncertainty is compared with a preset trigger threshold; In response to the uncertainty level being higher than the preset trigger threshold, it is determined that the current causal reasoning lacks key evidence, and a multi-agent collaborative verification process is triggered.

6. The method according to claim 5, wherein, The process of triggering multi-agent collaborative verification includes: The active perception task is modeled as a Markov decision process. The state of the Markov decision process includes the field of view of each edge node, the computing load state, the communication channel bandwidth and the queue of tasks to be verified. The actions include a task allocation scheme for assigning evidence collection tasks to each edge node. A comprehensive reward function is constructed, which is a weighted combination of the estimated benefit of task completion quality, communication transmission cost, and repeated collection penalty. The estimated benefit of task completion quality is determined based on the multi-view complementarity of each edge node. Based on the comprehensive reward function, the task allocation scheme is evaluated using a reinforcement learning algorithm to select the target task allocation scheme with the maximum long-term cumulative benefit. The target task allocation scheme is analyzed to obtain the optimal combination of edge nodes and the corresponding scheduling instructions.

7. The method according to claim 6, wherein, The method further includes: The optimal edge node combination is controlled to perform evidence collection tasks according to the scheduling instructions, and the verification results fed back by the optimal edge node combination are received. The complete event causal evolution path is updated based on the verification results, and the multi-step thought chain reasoning is re-executed based on the updated complete event causal evolution path until the uncertainty level is lower than the preset trigger threshold.

8. The method according to any one of claims 1 to 4, wherein, The method further includes: Based on the event reasoning results, the prior information of the scene is adaptively updated to obtain the updated prior model parameters; The updated prior model parameters are sent to the edge nodes to drive the dynamic evolution of the edge nodes' predictive perception capabilities.

9. A semantically driven cloud-edge collaborative causal reasoning device, wherein, The device includes: The receiving module is used to receive multi-source anomaly semantic features from edge nodes; The first construction module is used to construct global event association nodes based on the multi-source anomaly semantic features, combined with spatial location association information and a predefined event knowledge graph; The second construction module is used to perform causal reasoning on the global event-related nodes based on the temporal constraint mechanism, and to integrate the evolution patterns of similar events in the historical memory bank to construct a complete event causal evolution path. The reasoning module is used to transform the complete causal evolution path of the event into a structured prompt input large language model, and generate event reasoning results through multi-step thinking chain reasoning.

10. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

11. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-8.

12. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.

Images