A decision reasoning chain construction method supporting adaptive re-planning and fault-tolerant technology

Abstract

The application relates to a decision reasoning chain construction method supporting adaptive re-planning and fault-tolerant technology and belongs to the technical field of decision reasoning. The application obtains global input information and completes standardization preprocessing, extracts reasoning elements and constructs an element correlation system to generate basic feature information; based on preset reasoning rules and basic feature information, initial reasoning topology structure is generated through traversal matching, initial decision reasoning chain is formed through node reachability verification and conflict detection; in the reasoning chain execution stage, full-link state monitoring is carried out and operation snapshots are maintained, abnormal areas are located and isolation treatment is carried out, and the matched fault-tolerant disposal process is started; based on the snapshots and the running environment, the backtracking basis is extracted, the reasoning chain is incrementally reconfigured, reasoning is executed, and the decision result is output. The application realizes the abnormal fault tolerance and adaptive re-planning of the whole process of decision reasoning, improves the operation stability and continuity of the reasoning chain, and is suitable for automatic decision reasoning tasks in complex scenes.

Description

Technical Field

[0001] This invention belongs to the field of decision reasoning technology, specifically relating to a method for constructing a decision reasoning chain that supports adaptive replanning and fault tolerance technology. Background Technology

[0002] With the widespread application of intelligent decision-making and automated reasoning technologies in complex scenarios, the stable execution and continuous reliable operation of decision reasoning chains have become core technical requirements. Traditional decision reasoning methods often use fixed topologies and preset reasoning paths to achieve logical deduction. When constructing the reasoning chain, they only complete basic path matching and rule execution, lacking a pre-verification mechanism for node connectivity, link validity, and element constraint conflicts. This makes them prone to reasoning interruptions due to initial structural defects.

[0003] During the inference chain operation phase, existing technologies typically lack a full-link status monitoring mechanism, making it impossible to detect abnormal states of execution nodes, transmission links, and data flow in real time, and hindering the rapid location and scope of anomalies. When issues such as node failure, link interruption, timing conflicts, or resource contention occur, only simple restart or termination operations can be performed. The lack of a layered anomaly isolation mechanism and a matched fault-tolerant handling process can easily lead to anomalies spreading throughout the entire inference system, resulting in lost inference results, decision delays, and even task failure.

[0004] Meanwhile, traditional solutions do not support adaptive replanning and incremental topology reconstruction of the inference chain. After an anomaly occurs, the inference chain must be rebuilt from scratch. The backtracking mechanism is imperfect and cannot quickly locate the effective execution segment based on historical snapshots, resulting in high resource overhead and low recovery efficiency. In addition, existing technologies have not built standardized backup link scheduling and local topology adjustment mechanisms, making it difficult to complete local repairs while preserving effective inference logic, resulting in weak overall continuity and robustness.

[0005] The aforementioned problems make it difficult for traditional decision-making reasoning methods to meet the continuous decision-making requirements in scenarios with high reliability, strong real-time requirements, and complex constraints. Therefore, how to achieve pre-verification, in-process monitoring, anomaly isolation, fault-tolerant handling, and adaptive replanning of the reasoning chain has become an urgent technical problem to be solved in this field. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method for constructing a decision reasoning chain that supports adaptive replanning and fault-tolerant technology. The objective of this invention can be achieved through the following technical solution: S1: Obtain the full-domain input information for decision reasoning, perform standardized preprocessing on the full-domain input information, extract reasoning elements and establish an element association system to generate basic feature information; S2: Based on the preset reasoning rules and the basic feature information, perform reasoning path traversal matching to generate an initial reasoning topology; perform node reachability verification and conflict detection on the initial reasoning topology to form an initial decision reasoning chain; S3: Perform full-link status monitoring on the inference chain execution process and synchronously maintain snapshots of the inference chain operation; locate abnormal areas based on monitoring results and perform abnormal isolation processing, and initiate fault tolerance handling procedures that match the abnormality type; S4: Extract reasoning backtracking basis based on the inference chain running snapshot and the current running environment, perform incremental topology reconstruction on the inference chain based on the reasoning backtracking basis, perform reasoning according to the reconstructed inference chain and output decision results.

[0007] Specifically, the establishment of the element association system includes: The reasoning elements extracted from the standardized preprocessed global input information are classified and archived to obtain the category attributes of each reasoning element; By analyzing the semantic tags and attribute fields of each reasoning element, we can identify and distinguish the causal dependencies, temporal constraints, and resource competition relationships among the elements. Construct an element association matrix, add directed association edges between inference elements based on causal dependencies, configure delayed weights for each association edge based on temporal constraints, and mark mutually exclusive node pairs between inference elements based on resource competition relationships to form a structured element association system.

[0008] Specifically, the method for performing inference path traversal matching is as follows: By using a hierarchical and progressive reasoning path traversal mechanism, and combining the node-related characteristics in the basic feature information, the preset reasoning rules are prioritized and divided into core reasoning rules and auxiliary reasoning rules. Based on the principle of prioritizing core reasoning rules and supplementing with auxiliary reasoning rules, the reasoning branches of boundary conditions in the node-related characteristics are traversed. Combined with the directed association edges, delay weights and mutually exclusive node pair markings in the element association system, the feasibility of each reasoning branch is verified. The reasoning path scoring mechanism is used to quantitatively evaluate the valid reasoning branches and generate an initial reasoning topology.

[0009] Specifically, the reasoning path scoring mechanism includes: Multiple scoring dimensions are preset, and corresponding weight coefficients are set for the scoring dimensions. The weight coefficients are dynamically adjusted according to the priority requirements of the current reasoning scenario. For each reasoning branch that passes the verification, the individual score of each scoring dimension is calculated, and the individual score is weighted and summed with the corresponding weight coefficient to obtain the comprehensive score of each reasoning branch. Reasoning branches with a comprehensive score higher than a preset score threshold are considered valid reasoning branches, and reasoning branch combinations are generated.

[0010] Specifically, the execution node reachability verification and conflict detection include: Based on the node-related characteristics in the basic feature information, node reachability verification is performed through bidirectional traversal verification to judge the connectivity between nodes and the validity of link transmission, and abnormal topology objects are marked. By combining the association constraint information in the element association system, the topological anomaly state in the initial inference topology is judged; The abnormal topology objects based on the labels and the topology anomalies obtained are classified and recorded to generate a verification and detection report. Based on the verification and detection results, the initial inference topology structure is adjusted to form an initial decision inference chain.

[0011] Specifically, the adjustment operation includes: Based on the abnormal topology object markers in the verification and testing report, the abnormal topology objects in the initial inference topology are extracted and isolated to generate an intermediate topology. Based on the topological anomaly state determination results, correction processing is performed on mutual exclusion conflicts and time sequence constraint violations in the intermediate topological structure, and the weight configuration of directed edges is updated. Based on basic feature information and preset reasoning rules, suitable backup links are obtained from the backup associated link set and incorporated into the intermediate topology, the connectivity between nodes is reconstructed, reachability confirmation is performed on the adjusted topology, and the initial decision reasoning chain is output.

[0012] Specifically, the process of performing end-to-end status monitoring is as follows: Real-time sampling is performed on each execution node, directed edge, and data transmission process in the initial decision reasoning chain to obtain a set of running state parameters; The set of operating status parameters is compared with a threshold, abnormal data that exceeds the preset threshold range is marked, and the corresponding node identifier, directed edge identifier and timestamp are associated. The tagged abnormal data and normal data are merged according to time windows to generate a full-link status monitoring log.

[0013] Specifically, the process of maintaining the inference chain runtime snapshot is as follows: Based on a preset time interval, the current inference chain topology, the execution progress of each node, and the unprocessed state vectors in the state buffer are extracted from the end-to-end state monitoring log. The extracted data is serialized into snapshot data blocks, and each snapshot data block is assigned a unique identifier and a collection time label. The snapshot data block is compressed and encoded to generate a compressed snapshot data block. The compressed snapshot data block is written to the storage space, a snapshot index table is established, and the mapping relationship between snapshot identifiers and storage locations is stored.

[0014] Specifically, the process of performing the anomaly isolation procedure is as follows: Based on the anomaly marker information in the end-to-end status monitoring log, the abnormal nodes and abnormal directed edges are located, and the scope of the anomaly is traced through the node dependency relationship, marking the set of affected nodes and the set of directed edges. By performing channel isolation operations, the data transmission channel between the abnormal node and the external node is configured to be interrupted, the abnormal node and the abnormal directed edge are assigned to an independent isolation area, and an area identifier is assigned to the independent isolation area. Abnormal data within the isolation area is redirected to the isolation buffer, and the execution time and area identifier of the isolation operation are recorded and updated to the end-to-end status monitoring log.

[0015] Specifically, the fault-tolerant handling process includes: Extract abnormal feature parameters from the isolation buffer, classify and identify abnormalities based on the abnormal feature parameters, and generate abnormality type labels; Based on the anomaly type label, query the preset fault tolerance strategy mapping table to obtain the matching action sequence; Based on the action sequence, corresponding operations are performed on the abnormal nodes and abnormal directed edges in the isolation area in sequence, and the execution results and action records are written into the end-to-end status monitoring log.

[0016] Specifically, the incremental topology reconstruction process is as follows: Based on the abnormal region boundary and unexecuted node sequence in the reasoning backtracking basis, extract the local subgraph affected by the abnormal from the current reasoning chain topology, and obtain the interface node information between the local subgraph and the unaffected topology part; The local subgraph is subjected to topology adaptation adjustment. Based on the node association relationship and preset configuration parameters, the node execution sequence and directed edge pointing relationship are reconstructed, the adaptation links in the pre-configured backup link library are obtained, and the missing connectivity relationships are adjusted. The adjusted local subgraph is then subjected to connectivity verification, and the local subgraph is integrated with the unaffected topology to generate the reconstructed inference chain topology.

[0017] Specifically, the process of performing reasoning and outputting decision results is as follows: Extract the starting node from the reconstructed inference chain topology, and execute the inference operations corresponding to each node in the topological order. When executing the current node, the output data of the upstream node is obtained as input based on the input data requirements of the node, and the output data of the node is generated after performing inference operations; The output data of the current node is transmitted to the downstream node, the traversal operation of all executable nodes is performed, a decision result sequence is generated, the data integrity of the decision result sequence is verified, and the decision result sequence that passes the verification is output.

[0018] The beneficial effects of this invention are as follows: (1) By constructing a standardized element association system and a hierarchical traversal mechanism for reasoning paths, node reachability verification and conflict detection are completed during the generation stage of the decision reasoning chain. This can identify problems such as topological anomalies, element mutual exclusion, and violation of time sequence constraints in advance, effectively reducing the initial structural defects of the reasoning chain. Through pre-verification and topology adjustment, the anomaly rate during the reasoning execution stage can be reduced, improving the generation quality and structural rationality of the decision reasoning chain, and providing a reliable foundation for subsequent stable execution.

[0019] (2) An operational support mechanism combining full-link status monitoring, anomaly isolation, and matching fault tolerance handling is adopted. This mechanism can perceive the operational status of nodes and links in real time, quickly locate anomalies, and delineate isolation areas to prevent the spread of anomalies from affecting the global reasoning process. By calling the corresponding fault tolerance strategy according to the anomaly type, targeted handling is achieved, which greatly improves the system's ability to cope with faults such as node failure and link interruption, and enhances the stability and continuity of the decision-making and reasoning process.

[0020] (3) This invention achieves incremental topology reconstruction and adaptive replanning based on runtime snapshots and backtracking, adjusting only abnormal local subgraphs without requiring global reconstruction of the inference chain. By using backup link scheduling and local topology splicing, the inference path is quickly restored, shortening fault recovery time and reducing system resource consumption. This method can dynamically adapt to changes in the operating environment, enabling autonomous repair and continuous inference of the inference chain, and improving the robustness and real-time performance of the decision-making system in complex scenarios. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating a decision reasoning chain construction method supporting adaptive replanning and fault tolerance technology according to the present invention. Figure 2 This is an architecture diagram of a decision reasoning chain construction method that supports adaptive replanning and fault tolerance technology in this invention. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0024] Please see Figure 1-2 A method for constructing a decision reasoning chain that supports adaptive replanning and fault tolerance techniques, comprising: S1: Collect the full-domain input information for decision-making reasoning, perform standardized preprocessing on the full-domain input information, extract reasoning elements and establish an element association system to generate basic feature information; S2: Based on the preset reasoning rules and the basic feature information, perform reasoning path traversal matching to generate an initial reasoning topology; perform node reachability verification and conflict detection on the initial reasoning topology to form an initial decision reasoning chain; S3: Perform full-link status monitoring on the inference chain execution process and synchronously maintain snapshots of the inference chain operation; locate abnormal areas based on monitoring results and perform abnormal isolation processing, and initiate fault tolerance handling procedures that match the abnormality type; S4: Extract reasoning backtracking basis based on the inference chain running snapshot and the current running environment, perform incremental topology reconstruction on the inference chain based on the reasoning backtracking basis, perform reasoning according to the reconstructed inference chain and output decision results.

[0025] Specifically, the establishment of the element association system includes: The reasoning elements extracted from the standardized preprocessed global input information are classified and archived to clarify the category attributes of each reasoning element. By parsing the semantic tags and attribute fields of each reasoning element, causal dependencies, temporal constraints, and resource competition relationships between elements are identified and distinguished. An element association matrix is ​​constructed, and directed association edges are set for the corresponding reasoning elements in the element association matrix based on the identified causal dependencies. Based on the temporal constraints, corresponding delay weights are configured for each directed association edge, and based on the resource competition relationships, mutually exclusive node pairs are marked for reasoning elements with competing relationships in the element association matrix. The directed association edges, delay weights, and mutually exclusive node pair markings in the element association matrix are integrated to form a structured element association system for subsequent reasoning path traversal and matching.

[0026] Example 1: This embodiment uses the construction of the decision-making reasoning chain of an autonomous vehicle as an application scenario to elaborate in detail the specific process of establishing an element association system based on full-domain input information. In the autonomous driving decision-making system, full-domain input information includes external environment perception data (such as obstacle distance, lane line position, and traffic light status), vehicle own state data (such as current vehicle speed, steering angle, and braking pressure), and reasoning constraint data (such as safe distance threshold, maximum speed limit, and traffic rule priority).

[0027] I. Standardized Preprocessing and Inference Elements After receiving the aforementioned global input information, the extraction system performs standardized preprocessing, including data format standardization, dimension normalization, and noise filtering. After preprocessing, inference elements are extracted, such as: element E1 representing "distance to the obstacle ahead," element E2 representing "current vehicle speed," element E3 representing "safe braking distance," element E4 representing "steering command," and element E5 representing "braking command." These inference elements are then categorized and archived to obtain the category attributes of each element.

[0028] Category attributes are divided into three types: event-related elements (such as obstacle appearance, traffic light change), state-related elements (such as vehicle speed, braking pressure), and constraint-related elements (such as safe distance threshold, speed limit). For example, E1 and E2 are state-related elements, E3 is a constraint-related element, and E4 and E5 are event-related elements (actions performed).

[0029] II. Identifying Relationships Between Elements By analyzing the semantic labels and attribute fields of each inference element, three types of relationships are identified and distinguished: Causal dependency: For example, a decrease in E1 (distance to the obstacle ahead) will cause E2 (current vehicle speed) to need to be reduced, which in turn triggers E5 (braking command). Therefore, E1→E2→E5 constitute a causal chain.

[0030] Temporal constraints: For example, after E1 detects an obstacle, it must issue the E5 braking command within a time interval Δt1 (e.g., 0.5 seconds); after E5 is executed, it needs to reach the target deceleration within Δt2 (e.g., 1 second). These temporal constraints define the delay weights of the associated edges.

[0031] Resource competition: For example, the actuator cannot complete the two actions of emergency braking and steering avoidance at the same time. Therefore, E4 (steering command) and E5 (braking command) constitute a mutually exclusive node pair, that is, only one of them can be selected for execution at the same time.

[0032] III. Constructing an element association matrix Construct an N×N element association matrix M, where N is the total number of inference elements (N=5 in this embodiment). The rows and columns of the matrix correspond to the inference elements (arranged in order from E1 to E5). Each element M(i,j) in the matrix describes the relationship between element i and element j. For causal dependencies, if a directed edge points from E1 to E2, it is marked as 1 in M(1,2), indicating the existence of a direct causal dependency. If there is no direct causal dependency, it is marked as 0.

[0033] In this embodiment, M(1,2)=1 (E1→E2), M(2,5)=1 (E2→E5), and M(1,4)=0 (E1 and E4 have no direct causal relationship). For temporal constraints, a delay weight parameter ω is added to each directed edge. ij For example, ω 12 =0.1 seconds (reaction time from detecting an obstacle to starting deceleration), ω 25 =0.3 seconds (the time from the start of deceleration to full braking). These weights are stored in the additional properties of the association matrix and do not change the 0 / 1 structure of the matrix. For resource competition relationships, a separate list of mutually exclusive node pairs is maintained outside the matrix, with records (E4, E5) as a pair of mutually exclusive nodes, indicating that the two cannot appear in the same reasoning path at the same time.

[0034] IV. Forming a structured system of related elements The constructed feature association matrix, the delayed weight configuration of directed edges, and the labels of mutually exclusive node pairs are integrated to form a structured feature association system. This system can be represented as a triple: G=(V,E,R), where V is the set of inference feature nodes (V={E1,E2,E3,E4,E5}), E is the set of directed edges (E={(E1→E2),(E2→E5)}), and R is the set of mutually exclusive node pairs (R={(E4,E5)}), with a delayed weight assigned to each directed edge. This feature association system, as a core component of basic feature information, is used for subsequent traversal matching of inference paths and generation of topology structures.

[0035] Through the above embodiments, the abstract global input information is transformed into a computable element association system, realizing the structured representation of the decision reasoning chain and providing basic data support for adaptive replanning and fault-tolerant processing.

[0036] Specifically, the method for performing inference path traversal matching based on preset inference rules and the basic feature information is as follows: A hierarchical, progressive reasoning path traversal mechanism is adopted. Combining node-related characteristics from basic feature information, the preset reasoning rules are prioritized, dividing them into core reasoning rules and auxiliary reasoning rules. The node-related characteristics include node attributes, node dependencies, and boundary conditions. Based on the principle of prioritizing core reasoning rules and supplementing with auxiliary reasoning rules, starting from the initial reasoning node, all reasoning branches that meet the boundary conditions in the node-related characteristics are traversed. The feasibility of each reasoning branch is verified by combining directed association edges, delay weights, and mutually exclusive node pair markers in the element association system. A reasoning path scoring mechanism is used to quantitatively evaluate the verified reasoning branches, selecting the optimal combination of reasoning branches to generate the initial reasoning topology. The reasoning path scoring mechanism sets the scoring dimensions and weight ratios based on reasoning efficiency, node reliability, and resource consumption.

[0037] Specifically, the reasoning path scoring mechanism includes: Multiple scoring dimensions are preset, including node reachability of inference branches, data transmission efficiency, resource utilization, and rule matching degree. A corresponding weight coefficient is configured for each scoring dimension, and the weight coefficient is dynamically adjusted according to the priority requirements of the current inference scenario. For each inference branch that passes the verification, the individual score of each scoring dimension is calculated, and the individual score is weighted and summed with the corresponding weight coefficient to obtain the comprehensive score of each inference branch. Inference branches with a comprehensive score higher than a preset scoring threshold are regarded as valid inference branches. They are sorted from high to low comprehensive scores, and the combination of inference branches with the highest comprehensive score is selected as the optimal inference branch combination.

[0038] Specifically, the execution node reachability verification and conflict detection include: Based on the node-related characteristics in the basic feature information, a bidirectional traversal verification method is used to perform node reachability verification. Starting from the initial inference node and the terminating inference node, all execution nodes and associated links in the initial inference topology are traversed. The connectivity between execution nodes and the validity of link transmission are judged, and abnormal topology objects are marked. Combining the association constraint information in the element association system, the topology anomaly state in the initial inference topology is judged. The marked abnormal topology objects and the judged topology anomaly states are classified and recorded, and a verification and detection report is generated. According to the verification and detection results, the initial inference topology is adjusted. After removing abnormal topology objects and resolving topology anomaly states, the initial decision inference chain is formed. The association constraint information includes the constraint requirements of mutually exclusive nodes on marking, delay weights, and node-related characteristics. The abnormal topology objects include unreachable nodes and broken links. The topology anomaly states include node mutual exclusion states, time-series constraint states, and resource contention states.

[0039] Specifically, the adjustment operation includes: Based on the abnormal topology object markers in the verification and testing report, the abnormal topology objects in the initial inference topology are extracted and isolated to generate an intermediate topology. Based on the topological anomaly state determination results, correction processing is performed on mutual exclusion conflicts and time sequence constraint violations in the intermediate topological structure, and the weight configuration of directed edges is updated. Based on basic feature information and preset reasoning rules, suitable backup links are obtained from the backup associated link set and incorporated into the intermediate topology, the connectivity between nodes is reconstructed, reachability confirmation is performed on the adjusted topology, and the initial decision reasoning chain is output.

[0040] Example 2: This example uses a multi-AGV collaborative handling task in a smart factory as a scenario. Basic feature information includes nodes (pickup point A, transport path nodes B / C / D, unloading point E), dependencies (A→B→C→D→E), and boundary conditions (AGV load ≤ 100kg, path width ≥ 1.5m). The element association system defines directed association edges (e.g., A→B, delay weight W). AB =0.3, indicating that the vehicle must depart within 0.3 seconds after pickup; mutually exclusive node pairs (e.g., (B,C) means that B and C cannot be occupied by the same AGV at the same time).

[0041] I. Layered Progressive Reasoning Path Traversal Matching The system's preset inference rule base contains 12 rules, sorted by priority algorithm: each rule is assigned a priority score: P rule =α × urgency + β × historical success rate. Urgency is set from 1 to 5 (5 being the highest), and historical success rate is set from 0 to 1. After calculation, the rules are sorted, and the top 40% are classified into the core rule set (such as "load weight ≤ 100kg" and "path does not conflict"), while the bottom 60% are classified into the auxiliary rule set (such as "lowest energy consumption" and "shortest path").

[0042] Starting from node A, a depth-first search (DFS) is used to traverse all inference branches that satisfy the boundary conditions. Boundary conditions include: node attributes (e.g., A is a pickup point, it can only be dispatched from but not reached), node dependencies (A must be connected to B), and the boundary condition (the AGV's current load capacity is ≤100kg). During traversal, a feasibility check is performed on each branch: Check if directed edges exist: If the branch contains A→C (without direct edges), mark it as infeasible.

[0043] Check the delay weight: record the actual execution time t of edge (A→B). AB If t AB >W AB If +δ (δ is the allowable deviation, preset to 0.05s), then the timing constraint is violated.

[0044] Check for mutually exclusive node pairs: If both B and C appear in the branch, mark them as mutually exclusive.

[0045] The branch that passes the verification will proceed to the scoring stage.

[0046] II. Reasoning Path Scoring Mechanism Four scoring dimensions are preset, and the scoring method for each dimension is as follows: Node reachability score R1: number of actually connected nodes in the branch / total number of nodes.

[0047] Data transmission efficiency score R2: 1 - (total path delay / maximum allowable delay), where total path delay = ΣW ij .

[0048] Resource utilization score R3: 1 - (peak CPU utilization / 100%).

[0049] Rule matching score R4: Number of core rules for branch matching / Total number of core rules.

[0050] The weighting coefficients are dynamically adjusted based on the current task scenario (e.g., urgent orders). w1=0.2, w2=0.4, w3=0.1, w4=0.3 (weighted sum is 1). For each branch, the overall score is: S=Σ(wi·Ri).

[0051] For example, a branch score of (0.8, 0.9, 0.7, 0.6) → S = 0.2 × 0.8 + 0.4 × 0.9 + 0.1 × 0.7 + 0.3 × 0.6 = 0.16 + 0.36 + 0.07 + 0.18 = 0.77. A preset scoring threshold θ = 0.7 is used; branches scoring above 0.7 are considered valid. These branches are sorted in descending order of S, and the highest-scoring branches are combined to generate the initial inference topology.

[0052] III. Node Reachability Verification and Conflict Detection Starting from the initial node A, perform a forward Depth-First Search (DFS) and record the set V of all reachable nodes. fwd ={A,B,C,D,E}. Starting from the terminal node E, perform a reverse DFS (along the reverse edges) and record the set V of all nodes that can be reached in the reverse direction. rev ={E,D,C,B,A}. Calculate V. fwd With V rev The intersection of V yields the truly connected set of nodes. fwd Nodes not in the intersection of V are marked as "unreachable nodes" (e.g., if there is an isolated node F, then F is marked); nodes whose starting point is in V are marked as "unreachable nodes"; fwd But the destination is not V. fwd The edge is marked as "broken link".

[0053] Detecting topological anomalies by combining associated constraint information: Node mutual exclusion state: Traverse all mutually exclusive node pairs (B, C) and check if the same path contains both B and C. If so, record the mutual exclusion conflict.

[0054] Temporal constraint state: For a directed edge (A→B), if the actual delay t AB =0.25s, W AB =0.3s, deviation -0.05s (not exceeding δ=0.05s) then passes; if t AB =0.4s, deviation +0.1s>δ, then the timing of the marking is violated.

[0055] Resource contention status: Check if the same resource (such as a communication channel) is requested by more than one node at the same time, and mark it as a resource contention.

[0056] Classify and record abnormal topology objects (unreachable node F, broken link F→G) and abnormal topology states (mutual exclusion conflict, timing violation), and generate a verification and detection report.

[0057] IV. Adjustment Operation Based on the report, abnormal topology objects F and F→G were extracted and isolated: node F and edge F→G were temporarily disabled, and an intermediate topology (excluding F and F→G) was generated. For mutual exclusion conflicts (both B and C appear simultaneously), the directed edge with the smallest weight in the conflicting path was deleted (assuming B→C has a weight of 0.2 and C→D has a weight of 0.3, then B→C is deleted). For temporal violations (t... AB =0.4s>0.35s), adjust the delay weight W_AB to 0.4s (update to the actual value), and notify the upper-level monitoring.

[0058] Select an appropriate link from the set of backup links: The backup link library pre-stores alternative paths to bypass the faulty node. For example, after F is isolated, the original path A→B→F→G→E can be replaced by A→B→H→E (where H is the backup node). Selection criteria: The end node of the backup link matches the interface node of the current topology (e.g., B and E), and the total link delay does not exceed 1.2 times the delay of the original path. Incorporate the selected backup link into the intermediate topology structure to reconstruct the connectivity between nodes.

[0059] Finally, re-perform bidirectional reachability verification on the adjusted topology (repeat step three), and output the initial decision reasoning chain after confirming that there are no anomalies.

[0060] Specifically, the process of performing end-to-end status monitoring is as follows: Real-time sampling is performed on each execution node, directed edge, and data transmission process in the initial decision reasoning chain to collect a set of running status parameters, including the execution status parameters of nodes, the transmission status parameters of directed edges, and data flow parameters. The collected parameters are compared with thresholds, and abnormal data exceeding the preset threshold range is marked and associated with the corresponding node identifier, directed edge identifier, and timestamp. The marked abnormal data and normal data are merged according to time windows to generate a full-link status monitoring log.

[0061] Specifically, the process of maintaining the inference chain runtime snapshot is as follows: Extract the current inference chain topology, execution progress of each node, and unprocessed state vectors in the state buffer from the end-to-end status monitoring log at preset time intervals; serialize the extracted data into snapshot data blocks, and assign a unique identifier and collection time label to each snapshot data block; perform compression encoding on the snapshot data blocks to generate compressed snapshot data blocks; write the compressed snapshot data blocks into the storage space, establish a snapshot index table, and record the mapping relationship between snapshot identifiers and storage locations.

[0062] Specifically, the process of performing the anomaly isolation procedure is as follows: Based on the anomaly marker information in the end-to-end status monitoring log, the abnormal nodes and abnormal directed edges are located. The scope of the anomaly is traced through node dependencies, and the set of affected nodes and the set of directed edges are marked. By performing channel isolation operations, the data transmission channel between the abnormal node and external nodes is interrupted, and a region identifier is assigned to the isolation region. The abnormal data in the isolation region is redirected to the isolation buffer, the execution time of the isolation operation and the region identifier are recorded, and the end-to-end status monitoring log is updated.

[0063] Specifically, the fault-tolerant handling process includes: Anomaly feature parameters are extracted from the isolation buffer, and anomalies are classified and identified based on these parameters to generate anomaly type labels. A preset fault tolerance strategy mapping table is queried based on the anomaly type labels to obtain a matching sequence of handling actions, which includes node retry operations, node skip operations, and function degradation operations. Based on the sequence of handling actions, corresponding operations are executed sequentially on the abnormal nodes and abnormal directed edges in the isolation area, and the execution results and action records are written to the end-to-end status monitoring log.

[0064] Example 3: This example uses the decision-making reasoning chain of an autonomous vehicle as a scenario. The initial decision-making reasoning chain includes a perception node (detecting obstacles), a planning node (planning avoidance paths), a control node (outputting steering / braking commands), and directed edges and data transmission channels between the nodes.

[0065] I. End-to-end status monitoring The system samples each node, directed edge, and data transmission process in the inference chain in real time at a sampling frequency of 10Hz, collecting a set of runtime state parameters. For example: Node execution status parameters: node processing time (in milliseconds), output data validity flag; Directed edge transmission state parameters: data transmission delay (ms), packet loss rate; Data flow parameters: data packet size, checksum.

[0066] The system presets threshold ranges for each parameter: node processing time ≤ 50ms, transmission latency ≤ 10ms, and packet loss rate ≤ 5%. After sampling, the measured values ​​are compared with the thresholds. Assuming that at a certain moment, the processing time of the planning node reaches 120ms, exceeding the 50ms threshold, the system marks this data as abnormal, associating it with the node ID (planning node), directed edge ID (perception → planning), and timestamp T0. The system merges abnormal and normal data into a 1-second time window, generating a full-link status monitoring log. Each log entry includes the start and end times of the time window, node / edge identifier, parameter value, and whether it is abnormal.

[0067] II. Maintaining Inference Chain Snapshots The system triggers a snapshot operation every 30 seconds. It extracts the current inference chain topology (node ​​list, directed edge list), the execution progress of each node (percentage of completed nodes), and unprocessed state vectors in the state buffer (such as the current sensing data frame sequence number) from the state monitoring log. This data is serialized into JSON-formatted snapshot data blocks, each assigned a unique snapshot ID (e.g., Snap_20250417_093000) and a collection time tag. The snapshot data blocks are then gzip-compressed to generate compressed snapshot data blocks. These compressed blocks are written to SSD storage and recorded in the snapshot index table (snapshot ID, storage path, compressed size, checksum). The index table is stored in memory for fast lookup.

[0068] III. Implement abnormal isolation procedures Based on the anomaly marker information (planning node processing time exceeded the limit) in the end-to-end status monitoring log, the system identifies the abnormal node as the "planning node," with the abnormal directed edge being "perception → planning." The impact range of the anomaly is traced through node dependencies: the abnormal output of the planning node causes its downstream "control node" to be unable to obtain valid planning results; therefore, the affected set includes both the "planning node" and the "control node," as well as the directed edge "planning → control." By performing a channel isolation operation, the data transmission channel between the abnormal node (planning node) and external nodes (perception node, control node) is interrupted, i.e., the input and output queues are temporarily blocked. A zone ID "Zone_01" is assigned to the isolation zone. Data to be processed within the isolation zone (such as incomplete planning requests) is redirected to the isolation buffer (a dedicated queue in memory), and the execution time T1 and zone ID of the isolation operation are recorded, updating the end-to-end status monitoring log.

[0069] IV. Fault Handling Procedures Extract anomaly characteristic parameters from the isolation buffer, including: anomaly type (processing time exceeded limit), anomaly node ID, input data volume before the anomaly occurred, node CPU utilization, etc. Classify and identify anomalies based on these characteristic parameters: planning node timeouts are usually due to excessively high algorithm complexity or a surge in input data volume; in this example, it is classified as a "performance degradation anomaly," generating the anomaly type label "PERF_DEGRADE."

[0070] The system pre-defines a fault tolerance policy mapping table, as shown below:

[0071] Based on the "PERF_DEGRADE" label, the mapping table is queried to obtain the sequence of actions: [Switch to degrade mode, skip the current planning cycle, record the exception]. The system then performs the above operations sequentially on the exception nodes (planning nodes) within the isolation area: Switch the planning algorithm from high-precision mode to fast mode (sacrificing a small amount of accuracy for speed).

[0072] After executing the fast mode, monitor the processing time again. If it still exceeds 50ms, skip the current cycle planning, reuse the valid planning results from the previous cycle, and send a "planning downgrade" flag to the control node.

[0073] Increment the exception counter by 1.

[0074] After each step is executed, the execution result (success / failure, actual time taken, degradation flag) and action record are written to the end-to-end status monitoring log to form a closed loop.

[0075] If the exception type is NODE_CRASH (node ​​crash), the handling sequence will include a node restart operation: sending a restart command to the node management process, waiting for the node to recover and reloading the most recent snapshot, and retrying unfinished tasks. If the retry fails after 3 attempts, a skip operation will be performed, setting the node output to a safe default value (e.g., braking to a stop), and triggering an advanced alarm.

[0076] Through the aforementioned end-to-end status monitoring, snapshot maintenance, anomaly isolation, and fault-tolerant handling based on policy tables, the system can automatically recover or degrade when node performance declines or crashes, ensuring the continuity and safety of autonomous driving decisions.

[0077] Specifically, the incremental topology reconstruction of the inference chain is performed as follows: Based on the abnormal region boundaries and unexecuted node sequences in the inference backtracking basis, the local subgraph affected by the anomaly is extracted from the current inference chain topology, and the interface node information between the local subgraph and the unaffected topology is recorded. Topology adaptation adjustment is performed on the local subgraph: a node reordering operation is performed, rearranging the node execution order according to the dependencies and backup priorities between nodes; a directed edge redirection operation is performed, redirecting directed edges that bypass abnormal nodes to backup nodes, establishing a correspondence between directed edges and alternative paths; a backup link injection operation is performed, selecting suitable backup links from a pre-configured backup link library, replacing failed directed edges, and incorporating missing connectivity relationships; connectivity verification is performed on the adjusted local subgraph, and after confirming that the interface nodes and the unaffected topology are correctly connected, the adjusted local subgraph is spliced ​​with the unaffected topology to generate the reconstructed inference chain topology.

[0078] Specifically, the process of performing reasoning and outputting decision results is as follows: Extract the starting node from the reconstructed inference chain topology, and execute the inference operations corresponding to each node in the topological order. When executing the current node, obtain the output data of the upstream node as input according to the input data requirements of the node, and generate the output data of the node after performing the inference operation. Pass the output data of the current node to the downstream node as the input basis of the downstream node, until all executable nodes are traversed. Assemble the output data generated by each node during the traversal into a decision result sequence according to the execution order. After verifying the data integrity of the decision result sequence, output the decision result sequence that has passed the verification.

[0079] Example 4: This example uses the decision-making reasoning chain of an autonomous vehicle under complex road conditions as a scenario. The initial reasoning chain includes the following node sequence: perception node (N1) → localization node (N2) → planning node (N3) → control node (N4). Due to the presence of a construction area ahead, the planning node (N3) timed out and has been isolated. The system performs incremental topology reconstruction.

[0080] 1. Extract the local subgraph and interface nodes affected by the anomaly. Based on the boundary of the abnormal region (abnormal node N3 and its upstream and downstream directed edges) in the reasoning backtracking basis, extract the local subgraph affected by the abnormality from the current reasoning chain topology. This subgraph contains nodes N2, N3, and N4, and directed edges N2→N3 and N3→N4. The unaffected part is node N1 and the starting segment of the directed edges. Record the interface nodes: the upstream interface node is N2 (subgraph input), and the downstream interface node is N4 (subgraph output).

[0081] II. Topology Adaptation and Adjustment The system performs three types of adjustment operations: Node reordering: Since N3 is unavailable, the system selects a backup planning node N3' (with the same input / output interface but a more robust algorithm) from the backup node library. Based on the dependency (the output of N2 must be used as planning input), the node execution order is rearranged to N2→N3'→N4.

[0082] Directed edge redirection: The original directed edge (N2→N3) that bypasses N3 is redirected to N2→N3', and N3→N4 is redirected to N3'→N4. At the same time, an alternative path mapping table is established to record the correspondence between the original edge (N2,N3) and the alternative edge (N2,N3').

[0083] Alternate Link Injection: Select an appropriate link from the pre-configured alternative link library. The library stores alternative paths for various failure scenarios. For example, when a planned node fails, an "emergency braking link" N2→N4 (skipping planning and directly braking) can be injected. In this scenario, since the vehicle is still in motion, a "degraded planning link": N2→N3'→N4, is injected, replacing the failed N3 node. Simultaneously, missing connectivity is supplemented: an input data buffer queue and output data routing are configured for N3'.

[0084] III. Connectivity Verification and Segmentation Perform connectivity checks on the adjusted local subgraph: Starting from N2, can N3' and N4 be reached via directed edges? The check passes. Confirm that the output of interface node N2 matches the output of the unaffected part (N1), and that N4 matches the input of the downstream (actuator). Concatenate the adjusted local subgraph with the unaffected part (N1→N2) to generate the reconstructed inference chain topology: N1→N2→N3'→N4.

[0085] IV. Perform reasoning and output decision results Extract the starting node N1 from the reconstructed topology and execute the following steps sequentially: Execute N1 (perception), obtain upstream (no) output as input, and generate perception data D1 (obstacle position, lane line).

[0086] D1 is passed to N2 (location), and N2 generates the location result D2 based on D1 and GPS data.

[0087] D2 is passed to N3' (degraded planning), and N3' quickly generates the avoidance trajectory D3 based on D2 (the output is a sequence of steering angles).

[0088] D3 is passed to N4 (control), and N4 generates braking / steering command D4 based on D3.

[0089] After traversing all nodes, D1, D2, D3, and D4 are assembled into a decision result sequence [D1, D2, D3, D4] in execution order. The system performs data integrity verification on this sequence: checking whether each data packet contains the required fields, whether the timestamps are consecutive, and whether the checksum is correct. After the verification passes, D4 is output as the final decision result (instruction set) and sent to the vehicle actuator.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for constructing a decision reasoning chain that supports adaptive replanning and fault-tolerant techniques, characterized in that, include: S1: Obtain the full-domain input information for decision reasoning, perform standardized preprocessing on the full-domain input information, extract reasoning elements and establish an element association system to generate basic feature information; S2: Based on the preset reasoning rules and the basic feature information, perform reasoning path traversal matching to generate an initial reasoning topology; perform node reachability verification and conflict detection on the initial reasoning topology to form an initial decision reasoning chain; S3: Perform full-link status monitoring on the inference chain execution process and synchronously maintain snapshots of the inference chain operation; locate abnormal areas based on monitoring results and perform abnormal isolation processing, and initiate fault tolerance handling procedures that match the abnormality type; S4: Extract reasoning backtracking basis based on the inference chain running snapshot and the current running environment, perform incremental topology reconstruction on the inference chain based on the reasoning backtracking basis, perform reasoning according to the reconstructed inference chain and output decision results.

2. The method according to claim 1, characterized in that, The establishment of the element association system specifically includes: The reasoning elements extracted from the standardized preprocessed global input information are classified and archived to obtain the category attributes of each reasoning element; By analyzing the semantic tags and attribute fields of each reasoning element, we can identify and distinguish the causal dependencies, temporal constraints, and resource competition relationships among the elements. Construct an element association matrix, add directed association edges between inference elements based on causal dependencies, configure delayed weights for each association edge based on temporal constraints, and mark mutually exclusive node pairs between inference elements based on resource competition relationships to form a structured element association system.

3. The method according to claim 1, characterized in that, The method for performing inference path traversal matching is as follows: By using a hierarchical and progressive reasoning path traversal mechanism, and combining the node-related characteristics in the basic feature information, the preset reasoning rules are prioritized and divided into core reasoning rules and auxiliary reasoning rules. Based on the principle of prioritizing core reasoning rules and supplementing with auxiliary reasoning rules, the reasoning branches of boundary conditions in the node-related characteristics are traversed. Combined with the directed association edges, delay weights and mutually exclusive node pair markings in the element association system, the feasibility of each reasoning branch is verified. The reasoning path scoring mechanism is used to quantitatively evaluate the valid reasoning branches and generate an initial reasoning topology.

4. The method according to claim 3, characterized in that, The reasoning path scoring mechanism specifically includes: Multiple scoring dimensions are preset, and corresponding weight coefficients are set for the scoring dimensions. The weight coefficients are dynamically adjusted according to the priority requirements of the current reasoning scenario. For each reasoning branch that passes the verification, the individual score of each scoring dimension is calculated, and the individual score is weighted and summed with the corresponding weight coefficient to obtain the comprehensive score of each reasoning branch. Reasoning branches with a comprehensive score higher than a preset score threshold are considered valid reasoning branches, and reasoning branch combinations are generated.

5. The method according to claim 1, characterized in that, The execution node reachability verification and conflict detection specifically include: Based on the node-related characteristics in the basic feature information, node reachability verification is performed through bidirectional traversal verification to judge the connectivity between nodes and the validity of link transmission, and abnormal topology objects are marked. By combining the association constraint information in the element association system, the topological anomaly state in the initial inference topology is judged; The abnormal topology objects based on the labels and the topology anomalies obtained are classified and recorded to generate a verification and detection report. Based on the verification and detection results, the initial inference topology structure is adjusted to form an initial decision inference chain.

6. The method according to claim 5, characterized in that, The adjustment operation specifically includes: Based on the abnormal topology object markers in the verification and testing report, the abnormal topology objects in the initial inference topology are extracted and isolated to generate an intermediate topology. Based on the topological anomaly state determination results, correction processing is performed on mutual exclusion conflicts and time sequence constraint violations in the intermediate topological structure, and the weight configuration of directed edges is updated. Based on basic feature information and preset reasoning rules, suitable backup links are obtained from the backup associated link set and incorporated into the intermediate topology, the connectivity between nodes is reconstructed, reachability confirmation is performed on the adjusted topology, and the initial decision reasoning chain is output.

7. The method according to claim 1, characterized in that, The specific process for performing end-to-end status monitoring is as follows: Real-time sampling is performed on each execution node, directed edge, and data transmission process in the initial decision reasoning chain to obtain a set of running state parameters; The set of operating status parameters is compared with a threshold, abnormal data that exceeds the preset threshold range is marked, and the corresponding node identifier, directed edge identifier and timestamp are associated. The tagged abnormal data and normal data are merged according to time windows to generate a full-link status monitoring log.

8. The method according to claim 1, characterized in that, The specific process for maintaining the inference chain runtime snapshot is as follows: Based on a preset time interval, the current inference chain topology, the execution progress of each node, and the unprocessed state vectors in the state buffer are extracted from the end-to-end state monitoring log. The extracted data is serialized into snapshot data blocks, and each snapshot data block is assigned a unique identifier and a collection time label. The snapshot data block is compressed and encoded to generate a compressed snapshot data block. The compressed snapshot data block is written to the storage space, a snapshot index table is established, and the mapping relationship between snapshot identifiers and storage locations is stored.

9. The method according to claim 1, characterized in that, The specific process for performing anomaly isolation is as follows: Based on the anomaly marker information in the end-to-end status monitoring log, the abnormal nodes and abnormal directed edges are located, and the scope of the anomaly is traced through the node dependency relationship, marking the set of affected nodes and the set of directed edges. By performing channel isolation operations, the data transmission channel between the abnormal node and the external node is configured to be interrupted, the abnormal node and the abnormal directed edge are assigned to an independent isolation area, and an area identifier is assigned to the independent isolation area. Abnormal data within the isolation area is redirected to the isolation buffer, and the execution time and area identifier of the isolation operation are recorded and updated to the end-to-end status monitoring log.

10. The method according to claim 1, characterized in that, The fault-tolerant handling process specifically includes: Extract abnormal feature parameters from the isolation buffer, classify and identify abnormalities based on the abnormal feature parameters, and generate abnormality type labels; Based on the anomaly type label, query the preset fault tolerance strategy mapping table to obtain the matching action sequence; Based on the action sequence, corresponding operations are performed on the abnormal nodes and abnormal directed edges in the isolation area in sequence, and the execution results and action records are written into the end-to-end status monitoring log.

11. The method according to claim 1, characterized in that, The specific process of performing incremental topology reconstruction is as follows: Based on the abnormal region boundary and unexecuted node sequence in the reasoning backtracking basis, extract the local subgraph affected by the abnormal from the current reasoning chain topology, and obtain the interface node information between the local subgraph and the unaffected topology part; The local subgraph is subjected to topology adaptation adjustment. Based on the node association relationship and preset configuration parameters, the node execution sequence and directed edge pointing relationship are reconstructed, the adaptation links in the pre-configured backup link library are obtained, and the missing connectivity relationships are adjusted. The adjusted local subgraph is then subjected to connectivity verification, and the local subgraph is integrated with the unaffected topology to generate the reconstructed inference chain topology.

12. The method according to claim 1, characterized in that, The specific process of performing reasoning and outputting decision results is as follows: Extract the starting node from the reconstructed inference chain topology, and execute the inference operations corresponding to each node in the topological order. When executing the current node, the output data of the upstream node is obtained as input based on the input data requirements of the node, and the output data of the node is generated after performing inference operations; The output data of the current node is transmitted to the downstream node, the traversal operation of all executable nodes is performed, a decision result sequence is generated and the data integrity is verified, and the decision result sequence that passes the verification is output.

Images