A confidence propagation and fusion method for multi-source inference results

Abstract

The application discloses a confidence propagation and fusion method for multi-source reasoning results, and comprises the following steps: obtaining reasoning results of at least three heterogeneous reasoning sources and initializing the confidence of the reasoning results to obtain initial confidence; constructing a confidence propagation graph based on the initial confidence, taking reasoning conclusions as nodes and taking the dependency relationship between the reasoning conclusions as edges, wherein the weight of the edge is calculated according to the shared evidence set between the nodes; updating the node confidence based on the confidence of the predecessor node in the confidence propagation graph and the weight of the corresponding edge by using an iterative propagation algorithm, and obtaining a propagated conclusion according to the updated node confidence; detecting a multi-type conflict of the propagated conclusion, triggering conflict resolution according to the detected conflict degree to obtain a resolved conclusion; and performing fusion processing on a plurality of evidences supporting the same conclusion by using evidence theory, obtaining comprehensive confidence according to the fusion result, and outputting a comprehensive conclusion.

Description

Technical Field

[0001] This invention belongs to the field of uncertainty reasoning technology, and particularly relates to a method for the propagation and fusion of confidence of multi-source reasoning results. Background Technology

[0002] With the evolution of artificial intelligence technology, single reasoning models are no longer sufficient to meet the intelligent decision-making needs of complex scenarios. Hybrid intelligent systems, by integrating multiple reasoning mechanisms (such as symbolic rule reasoning, graph algorithm mining, and large language model reasoning) to leverage their strengths and compensate for their weaknesses, have become a current research hotspot. Different reasoning engines possess their own theoretical foundations, knowledge sources, and uncertainty expression methods, enabling them to provide reasoning results for the same problem from different dimensions, thus offering the possibility of improving the overall intelligence level of the system. Currently, simple fusion methods based on voting mechanisms, fusion methods based on probabilistic graphical models, and methods based on evidence theory have been applied to a certain extent, initially achieving the integration of multi-source information and decision support.

[0003] While existing technologies have made some progress in multi-source information fusion, the following prominent problems remain: The outputs of different inference engines lack a unified confidence expression framework, making comparison and fusion difficult on the same scale; when conflicts exist between multi-source results, there is a lack of effective conflict detection and resolution mechanisms, easily leading to decision-making errors; the fusion process lacks interpretability, making it impossible to trace the basis for the final conclusion, and failing to meet the requirements of auditable and interpretable decision-making processes in high-reliability domains; furthermore, most existing methods statically set the weights of inference sources, failing to dynamically adjust them based on their historical performance and contextual environment across different task types, resulting in poor adaptability. Therefore, this invention provides a method for confidence propagation and fusion of multi-source inference results. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for confidence propagation and fusion of multi-source inference results, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides a method for confidence propagation and fusion of multi-source inference results, comprising: Obtain the inference results from at least three heterogeneous inference sources and initialize the confidence of the inference results to obtain the initial confidence. Based on the initial confidence level, a confidence propagation graph is constructed with the reasoning conclusions as nodes and the dependencies between reasoning conclusions as edges, where the weight of the edges is calculated based on the evidence set shared between nodes. An iterative propagation algorithm is used to update the node confidence based on the confidence of the predecessor node and the weight of the corresponding edge in the confidence propagation graph, and the propagation conclusion is obtained based on the updated node confidence. The post-propagation conclusion is subjected to multi-type conflict detection, and conflict resolution is triggered based on the degree of detected conflict to obtain the resolved conclusion; The conclusion after resolution is processed using evidence theory to fuse several pieces of evidence supporting the same conclusion. Based on the fusion result, the overall confidence level is obtained and the overall conclusion is output.

[0006] Optionally, the process of initializing the confidence level of the inference result to obtain the initial confidence level includes: A performance evaluation matrix is ​​constructed based on the historical accuracy of each inference source on different task types; After identifying the task type of the current query, the historical accuracy of each inference source on the identified task type is obtained from the performance evaluation matrix as an adjustment factor, and the original confidence of the inference result is dynamically adjusted to obtain the adjusted confidence. The adjusted confidence level is normalized to obtain the initial confidence level.

[0007] Optionally, the process of calculating the edge weights based on the shared evidence set among nodes includes: The evidence sets supporting each node are obtained separately. The improved Tanimoto coefficient is used to calculate the edge weight from node i to node j based on the intersection modulus of the evidence sets of node i and node j, the modulus of the evidence set of node i, the modulus of the evidence set of node j, and the smoothing factor.

[0008] Optionally, the process of updating the node confidence using an iterative propagation algorithm includes: In each iteration, the confidence of a node is calculated based on the damping factor, the initial confidence of the node, and the weighted sum of the confidence of all predecessor nodes and the corresponding edge weights. This results in the confidence of the node after the current iteration. The iteration is repeated until the change in the confidence of all nodes is less than a preset threshold, at which point the update stops.

[0009] Optionally, the process of performing multi-type conflict detection on the post-propagation conclusion includes: The number of syntax layer conflicts is obtained by processing the post-propagation conclusions based on the syntax layer conflict detection rules. The number of semantic layer conflicts is obtained by processing the post-propagation conclusions based on semantic layer conflict detection rules. The number of pragmatic layer conflicts is obtained by processing the post-propagation conclusions based on pragmatic layer conflict detection rules. The number of time-layer conflicts is obtained by processing the post-propagation conclusions based on the time-layer conflict detection rules. The number of conflicts at each layer is quantified to obtain the degree of conflict at each layer. The total degree of conflict is calculated by weighting and summing the degree of conflict at each layer based on the preset layer weight coefficient.

[0010] Optionally, the process of triggering conflict resolution and obtaining the resolution conclusion based on the detected conflict level includes: When the total conflict level exceeds the preset conflict threshold, the rule chain that triggered the conflict is traced back based on the conclusion node involved in the conflict to perform a rule consistency check and obtain the rule check result. Based on the conflict point information, prompt words are constructed and processed by a large language model to obtain resolution suggestions; An evidence distance matrix is ​​constructed based on the Jousselme distance between each evidence source, and the relative credibility of each evidence source is obtained by processing the evidence distance matrix. The conclusion after resolution is obtained by comprehensively processing the results of the rule check, resolution suggestions, and relative credibility.

[0011] Optionally, the process of fusing several pieces of evidence supporting the same conclusion using evidence theory after resolution includes: The confidence level of each piece of evidence is converted into a basic probability assignment to obtain the basic probability assignment function for each piece of evidence; The Dempster-Shafer evidence theory is used to perform orthogonal sum operations on the basic probability assignment functions of multiple pieces of evidence supporting the same conclusion to obtain the fused basic probability assignment. The fused basic probability allocation is processed to obtain the comprehensive confidence level.

[0012] Compared with the prior art, the present invention has the following advantages and technical effects: This invention addresses the problems in existing technologies, such as the lack of a unified confidence expression framework for multi-source inference results, the absence of conflict detection and resolution mechanisms, poor interpretability of the fusion process, and static setting of inference source weights. It solves the technical problem of incomparable confidence levels of multi-source results by unifying the outputs of heterogeneous inference sources to the same scale through confidence initialization; it solves the technical problem of decision-making errors caused by direct fusion of conflicting conclusions by constructing a four-layer conflict detection system (syntactic, semantic, pragmatic, and temporal) combined with rule backtracking, large language model analysis, and a multi-strategy resolution mechanism based on improved evidence theory; it solves the technical problem of difficulty in tracing and auditing the decision-making process by generating a structured and interpretable report that fully records the entire fusion process; and it solves the technical problem of the inability to adaptively adjust inference source weights based on task type and historical performance by introducing a dynamic weight mechanism based on a performance evaluation matrix. Thus, it achieves efficient, reliable, and interpretable fusion of multi-source inference results. Attached Figure Description

[0013] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1This is a flowchart illustrating the overall process of the multi-source inference result confidence propagation and fusion method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the confidence propagation graph according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the process of the conflict detection and resolution module in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the dynamic weighting mechanism of an embodiment of the present invention. Figure 5 This is an example diagram illustrating the application of the present invention in a space engine fault diagnosis scenario. Detailed Implementation

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0016] Example 1 like Figure 1 As shown, this embodiment provides a method for confidence propagation and fusion of multi-source inference results, including the following steps: Step 1: Obtaining Multi-Source Inference Results and Initializing Confidence Obtain inference results from at least three heterogeneous inference sources, including: (1) Inference source and initial confidence level generation.

[0017] Three heterogeneous inference sources generate inference conclusions and their initial confidence levels: 1) OWL rule inference: Based on ontology and SWRL rules, new assertions are derived through an inference engine (such as Pellet or HermiT). The initial confidence level is usually determined by the confidence level of the rule (if the rule itself has a confidence level) or the deterministic factor of the inference engine. If there is no explicit confidence level, it can be set to a fixed value (such as 0.9). 2) Graph algorithm mining: Algorithms such as PageRank, community detection, and similarity propagation in graph databases are used to mine implicit relationships. The initial confidence level can be obtained by mapping the weights output by the algorithm (such as similarity scores). 3) Large Language Model Inference: LLM is guided to perform thought chain inference through prompting engineering, and the conclusion and its confidence level are output (LLM may output probability or confidence level statements). The LLM output needs to be parsed to extract the confidence level value.

[0018] (2) Dynamic initial confidence adjustment.

[0019] A dynamic weighting mechanism is adopted to adjust the initial confidence level based on the historical performance of the inference source on different task types: 1) Construct a performance evaluation matrix: Suppose there are (m) types of task categories (such as fault diagnosis, entity classification, relation prediction), and record the historical accuracy of each inference source (s) on each type of task (t). 2) Task type identification: For the current query, determine its task category through keyword matching or classification models. 3) Adjustment factor calculation: Use historical accuracy as a multiplicative factor to adjust the initial confidence level: ;in Let be the average accuracy of all sources on task (t). The adjustment factor (e.g., 0.2) is used to ensure that the adjustment range is reasonable and the confidence level remains in the [0,1] range.

[0020] (3) Normalization The initial confidence levels adjusted from each source are standardized to the same scale, and Min-Max linear normalization is applied: in and These are the minimum and maximum values ​​for all initial confidence levels, respectively. If probabilistic characteristics need to be preserved, softmax normalization can also be used. The temperature parameter (T) controls the smoothness of the distribution. After normalization, all confidence levels are located in ([0,1]) and the sum is not necessarily 1, but the confidence levels of each conclusion can be compared.

[0021] Step 2: Construct a confidence propagation graph, where: (1) Node definition: 1) Atomic conclusion: such as "the engine has bearing wear", which is a single fact. 2) Composite conclusion: composed of multiple atomic conclusions, such as "the engine has bearing wear and lubricating oil contamination", or connected by logical operations (AND / OR).

[0022] (2) Edge definition, edge represents the support or dependency relationship between conclusions, including three types: 1) Rule triggering relationship: if OWL rule If triggered, a directed edge is established from (A) to (B). 2) Graph algorithm similarity relationship: If the graph algorithm finds that the entities or concepts corresponding to two conclusions are highly similar (e.g., based on Jaccard similarity > threshold), then an undirected edge or a bidirectional edge is established. 3) LLM thought chain dependency: If the LLM declares in the inference that (A) supports (B), then a directed edge is established from (A) to (B).

[0023] (3) Algorithm construction. 1) Collect all conclusion nodes and establish a node set (V). 2) Traverse the OWL rule base, for each rule If all conclusions in the Antecedent node exist in (V), then add an edge from each Antecedent node to the Consequent node. 3) The similarity results output by the graph traversal algorithm: for each pair of similar conclusions ((u,v)), if the similarity... 4) Parse the LLM thought chain output, extract the supporting relationships in the reasoning steps, and add corresponding directed edges. 5) Remove duplicates and assign weights to each edge.

[0024] (4) Edge weight calculation. The improved Tanimoto coefficient is used to calculate the support between nodes: in: This is the set of evidence supporting node (i), including: the initial confidence source of node (i) itself (which can be considered as a virtual piece of evidence); the predecessor nodes pointing to (i) and their edge weights (which need to be accumulated). In practice, iterative calculation can be used: initially, each node's evidence set only contains its own initial confidence (considered as a piece of evidence), and as propagation progresses, the contributions of predecessor nodes are added as new evidence. For simplification, the weighted sum of predecessor node confidences can also be used to approximate the number of pieces of evidence. To smooth out the fraction and avoid a denominator of zero, it is usually taken as a smoothing factor. The coefficient ranges from [0,1], with a larger value indicating more shared evidence between the two nodes and higher support. For undirected edges, the weights are symmetric; for directed edges, asymmetric weights can be calculated based on the direction.

[0025] Step 3: Confidence Propagation Calculation. Based on the confidence propagation graph, an iterative propagation algorithm is used to update the confidence of each node, similar to a variant of belief propagation or PageRank. Specifically, it includes: (1) Initialization. Initial confidence level of each node (v). Set it to the value after normalization in step 1.

[0026] (2) Topological order propagation. Since the graph may contain cycles, an iterative approach is used until convergence. In each iteration, each node receives contributions from all its predecessor nodes: ;in is the damping factor (e.g., 0.2), preserving the effect of the initial confidence level; (\text{pre}(v)) is the set of predecessor nodes pointing to (v).

[0027] (3) Propagation function. The above formula is a linear weighted sum. More complex nonlinear functions (such as product, T-norm) can also be used, but the linear form is simple and effective.

[0028] (4) Iteration and convergence. Repeat the update until the confidence change of all nodes is less than the threshold. (like Or reach the maximum number of iterations (e.g., 100 times). If there are cycles in the graph, convergence must be ensured (this can be guaranteed by the damping factor).

[0029] Step 4: Conflict Detection and Quantification. Conflict detection is performed on the post-propagation conclusions, including: (1) Syntax layer conflicts. Specifically, this includes: 1) Detection objects: naming conflicts (the same entity is assigned different identifiers), type conflicts (the same entity is asserted as a mutually exclusive type, such as owl:disjointWith). 2) Algorithm: Construct an entity-assertion mapping and check if there are two assertions. and and and Declare them as disjoint in the ontology. This can be achieved by querying the classifier of the ontology inference engine.

[0030] (2) Semantic layer conflicts. Specifically, this includes: 1) Detection objects: violations of OWL ontology logical consistency, such as contradictions derived through rules (e.g., simultaneous existence of (P(a,b)) and 2) Algorithm: Load all current conclusions as facts into the OWL inference engine and check if the ontology is consistent. If inconsistent, locate the set of assertions that cause the contradiction.

[0031] (3) Pragmatic layer conflict. Specifically, this includes: 1) Detection object: The conclusion is inconsistent with the context, for example, some conclusions are unreasonable in a specific scenario (such as "air conditioning cooling in winter" contradicting the low outdoor temperature). 2) Algorithm: Detection is performed based on a predefined context rule base (such as common sense constraints). Each rule is in the form of: If condition (C) is true, then conclusions (A) and (B) cannot be true at the same time. This is checked through pattern matching.

[0032] (4) Time-level conflicts. Specifically, this includes: 1) Detection objects: Inconsistencies with the time series, such as reversed event order or contradictory durations. 2) Algorithm: If the conclusion has a timestamp, check the consistency of the time series. For example, if the rule requires (A) to occur before (B), but the timestamp shows that (B) is earlier than (A), then there is a conflict. Allen's interval algebra can be used for reasoning.

[0033] (5) Conflict level quantification. A weighting coefficient is assigned to each level of conflict. These correspond to the grammatical, semantic, pragmatic, and temporal layers, respectively. The setting principle is: 1) Semantic layer conflicts have the highest weight (e.g., ...). ), because logical inconsistencies directly undermine the foundation of the knowledge base. 2) Syntactic level ( Naming and type conflicts affect entity recognition. 3) Pragmatic layer ( ) and time layer ( The weights are relatively low because they depend on the external context, which may allow for some flexibility.

[0034] Overall conflict level Normalized to ([0,1]). If If the result is 0.3, proceed to step 5; otherwise, proceed directly to step 6.

[0035] Step 5: Conflict Resolution When a conflict is detected, the conflict resolution process is executed: (1) Backtrack to check the consistency of the rule base that triggered the conflict; When conflicts involve OWL rule reasoning, extract the rule chain that caused the conflict. Construct a conflict dependency graph, where nodes represent conclusions and edges represent rule triggering relationships. Traverse backwards from the conflicting nodes, marking all rules involved in the derivation. Check if these rules themselves are contradictory (e.g., conflicting rule premises). The consistency of the rule set can be tested through static rule analysis (e.g., checking if rule heads are mutually exclusive) or using an inference engine.

[0036] (2) Call the large language model to perform interpretive analysis on the conflict points and generate conflict resolution suggestions; 1) Package the points of conflict, related conclusions, and evidence into a hint and request the LLM to generate possible resolution suggestions, such as pointing out which piece of evidence may be unreliable or providing additional knowledge to explain the contradiction.

[0037] 2) After the LLM outputs suggestions, they are parsed and transformed into weight adjustments or evidence corrections.

[0038] (3) Improved integration of evidence theories.

[0039] The Dempster-Shafer evidence theory, which incorporates a conflict weighting factor, is adopted: 1) Calculate the Jousselme distance: for two pieces of evidence and (That is, the confidence assignment of two sources of reasoning to the same proposition), the distance is: in This is the focal element similarity matrix (such as the Jaccard index).

[0040] 2) Construct an evidence distance matrix, calculate the average distance of each piece of evidence to the other pieces of evidence, and obtain the relative credibility of the evidence: Normalization makes the sum of confidence levels equal to 1.

[0041] 3) Murphy's average preprocessing: For highly conflicting evidence (distance greater than a threshold), a weighted average is first performed to obtain the average evidence. Then, it is combined with low-conflict evidence using Dempster's algorithm. Specifically, the evidence is sorted by credibility, the top k pieces of evidence with the highest credibility are averaged, and then combined with other evidence.

[0042] (4) Expert intervention. If the above steps still cannot resolve the conflict, submit the conflict information (including relevant conclusions, propagation paths, and attempted resolution methods) to the experts through a visual interface. The experts can manually select the conclusions to trust or adjust the weights.

[0043] Step 6: Result Fusion and Overall Confidence Calculation. The conclusions after propagation and resolution are fused together: (1) Integration of Evidence Theories. For multiple pieces of evidence supporting the same conclusion, the Dempster-Shafer evidence theory is used to calculate the overall confidence level: in Assign basic probabilities to the conclusion for each piece of evidence (which can be derived from confidence levels, for example, by using confidence levels as support for a single proposition and allocating the remainder to the entire set). The resulting aggregate confidence level is then obtained.

[0044] (2) Hierarchical fusion. For conclusions with hierarchical relationships (such as "fault" and "bearing fault"), bottom-up fusion is adopted: first fuse the subclass conclusions, and then propagate upwards. For example, if the confidence of "bearing fault" is high, it will contribute to the confidence of the parent class "fault", but the inheritance relationship in the ontology needs to be considered.

[0045] (3) Output. Output the integrated conclusion after fusion and its confidence level, along with the fusion path (e.g., which evidences were involved in the fusion and the weights of each step).

[0046] Step 7: Generate an interpretability report. Record the entire confidence propagation and fusion process to generate an interpretability report.

[0047] The report includes the original conclusions and their confidence levels of each inference source; the structure and propagation path of the confidence propagation graph; conflict detection results and resolution process; and the basis for the formation of the final fusion conclusion.

[0048] (1) Data Structure Design: The report's data structure adopts a hierarchical JSON format (or a compatible document model) to facilitate serialization, parsing, and interaction with the front-end display. The core data structure is shown in Table 1.

[0049] Table 1 (2) Data storage design.

[0050] Based on the data volume, query requirements, and usage scenarios, the following storage solutions or combinations can be adopted under the above data structure: 1) Document database (Recommended: MongoDB) Applicability: The JSON structure of the reports is naturally suited for document databases. Each report is stored as a document, fields are indexable, and nested queries are supported.

[0051] Storage example: { "_id": "report_001", "metadata": {...}, "sources": [...], "propagation_graph": {...}, } Index design: Create indexes for metadata.timestamp, metadata.task_type, and final_results.conclusion.

[0052] 2) Graph databases (such as Neo4j) Applicability: If the confidence propagation graph is very large (number of nodes > 100,000) and frequent graph traversal queries are required (such as finding all predecessor paths of a certain conclusion), the graph structure can be stored separately in the native graph database.

[0053] Storage method: Node label: Conclusion, with attributes including label, initial_conf, final_conf, etc.

[0054] Relation type: SUPPORTS, attributes include weight, relation_type, details.

[0055] Each report can be a separate subgraph or distinguished by the report ID attribute.

[0056] Integration with a document repository: The graph database stores the propagation graph, while the document repository stores other structured information (metadata, conflict records, etc.), linked by report IDs.

[0057] 3) Object storage (such as AWS S3, MinIO) Applicability: Storing large raw files, such as full text of LLM thought processes, image evidence, and log files.

[0058] Storage method: Each report is packaged as a ZIP file or stored as a JSON file, with the file path containing the report ID.

[0059] 4) Hybrid storage strategy In practical systems, hybrid storage is often used to balance performance and cost: Hot data (recent reports, frequently accessed data): Stored in MongoDB for fast retrieval and API response.

[0060] Graph structure: For complex graph queries, synchronize to Neo4j.

[0061] Large text / binary data: Stored in object storage; MongoDB only stores the access link (URL).

[0062] Full-text search: Synchronize report text (such as conflict descriptions and conclusions) to Elasticsearch, supporting keyword search.

[0063] (3) Interactive follow-up query design To enable users to click on any node to view detailed evidence, the storage supports the following interactive follow-up query design: Node details query: Find the corresponding node in the MongoDB report by node_id, or directly match the node in Neo4j and return its attributes and relationships.

[0064] Edge details query: Similarly, it returns the details field of the edge.

[0065] Path tracing: Input the starting node and find all predecessor / successor nodes and edges. This can be quickly achieved using graph database traversal algorithms.

[0066] Evidence fusion process: Given the final conclusion, backtrack its fusion_path and aggregate the contributions of all participating nodes.

[0067] (4) Multi-granularity interpretation generation The above data structures and storage methods enable the generation of multi-granularity interpretations: Macro level: Generated from report metadata and statistical information from various sources, the macro_summary field can be read directly or calculated in real time.

[0068] At the meso level: extract the key paths (such as the edge with the highest weight and the nodes involved in the conflict) in the propagation_graph, and generate a subgraph after filtering by graph algorithms (such as shortest path, PageRank).

[0069] Micro level: Directly return the details field of the node / edge, or load the full text of the associated LLM mind chain from object storage.

[0070] Through the above data structure and storage scheme, the entire process of multi-source reasoning fusion can be systematically recorded, providing end users with transparent, traceable, and interactive explanatory capabilities.

[0071] Example 2 This embodiment provides a method for confidence propagation and fusion of multi-source inference results. The overall process of the confidence propagation and fusion method is as follows: Figure 1 As shown, the specific steps include: Multi-source inference result acquisition: Inference results and their initial confidence scores are obtained from the OWL rule inference engine (such as Pellet), graph algorithm mining module (such as Neo4j GDS), and large language model (such as ChatGLM-6B).

[0072] Confidence initialization: The initial confidence of the three inference sources is normalized and dynamically adjusted according to the task type and historical performance.

[0073] Confidence propagation graph construction: Construct a confidence propagation graph with reasoning conclusions as nodes and reasoning dependencies as edges.

[0074] Confidence propagation calculation: The confidence of nodes is updated along the propagation graph using an iterative algorithm.

[0075] Conflict detection: Detects whether there are logical, numerical, or categorical conflicts in the conclusions after propagation.

[0076] Conflict resolution: If a conflict is detected, the conflict resolution process is executed.

[0077] Results fusion: The conclusions after resolution are fused, and the overall confidence level is calculated.

[0078] Explainable report generation: Record the entire process and generate an explainable report.

[0079] The construction and calculation of confidence propagation graphs are as follows: Figure 2 As shown, it specifically includes: Source nodes S1, S2, and S3: represent the original conclusions of rule-based reasoning, graph algorithms, and LLM, respectively.

[0080] Intermediate nodes I1 and I2: represent intermediate conclusions after preliminary fusion.

[0081] Target node T: Represents the final conclusion.

[0082] Edge weight w_ij: represents the degree of support of node i for node j, with a value range of [0,1].

[0083] The confidence propagation calculation formula is: C(j) = C0(j) + Σ[i∈pre(j)] C(i) × w_ij × φ(i,j) where C(j) is the updated confidence of node j, C0(j) is the initial confidence, pre(j) is the set of predecessor nodes of j, and φ(i,j) is the propagation attenuation factor.

[0084] In this embodiment, the Tanimoto coefficient is used to calculate the edge weights, with α=0.5. Iterative propagation continues until the confidence change of all nodes is less than the threshold ε=0.01.

[0085] The detailed process of conflict detection and resolution is as follows: Figure 3 As shown, the workflow includes: (1) Obtain the set of conclusions after propagation.

[0086] (2) Perform four-layer conflict detection sequentially: Syntax layer: Checks whether entity names are consistent and type definitions match; Semantic layer: Loads the OWL ontology and calls the inference engine to check logical consistency; Pragmatic layer: Based on contextual rules, check whether the conclusion contradicts the current scenario; Time layer: Check whether the conclusions are consistent with the time series data.

[0087] (3) Calculate the conflict score = α1×syntax_conflict + α2×semantic_conflict + α3×pragmatic_conflict + α4×temporal_conflict, where α1-α4 are weighting coefficients.

[0088] (4) If conflict_score < θ (threshold set to 0.3), proceed directly to the fusion step.

[0089] (5) If conflict_score ≥ θ, initiate conflict resolution: Use LLM to analyze the causes of conflicts and generate explanations and suggestions; Calculate the Jousselme distance between each source of evidence and construct an evidence distance matrix; The relative credibility of each piece of evidence is calculated based on the evidence distance. Highly conflicting evidence was preprocessed using the Murphy averaging method. The improved Dempster combination rule is applied for fusion; If the conflict cannot be resolved, the conflict information will be packaged and sent to the expert system.

[0090] (6) Record the conflict detection and resolution process for subsequent interpretation report generation.

[0091] Dynamic weighting mechanism, such as Figure 4 As shown, the dynamic weighting mechanism includes: Offline training phase: Construct an inference source performance evaluation matrix E[i][j], where i represents the inference source and j represents the task type (e.g., "fault diagnosis", "situation prediction", "entity recognition"). Each matrix element records the historical accuracy of the inference source on that type of task.

[0092] Online reasoning stage: a) Classify the current query by task type; b. Retrieve the accuracy of each inference source on this task type from the evaluation matrix; c uses accuracy as an adjustment factor to update the initial confidence level: C_adjusted[i] = C0[i] × (0.7 + 0.3 × accuracy[i]); d. Update the evaluation matrix regularly based on user feedback.

[0093] In this embodiment, the task type classification uses a lightweight BERT model, which is finely tuned on aerospace datasets, achieving a classification accuracy of 92%.

[0094] Applications in aerospace engine fault diagnosis, such as Figure 5 As shown, the test run of a certain type of liquid rocket engine includes: (1) The sensor detected excessive vibration of the turbo pump (16.2 mm / s) and a decrease in speed (27,500 rpm).

[0095] (2) All three inference engines start simultaneously: The rule reasoning engine triggers the SWRL rule "Excessive vibration ∧ decreased speed → bearing wear", and outputs the conclusion "bearing wear" with an initial confidence level of 0.95. The graph algorithm retrieved similar cases and found three historical cases (F012, F035, and F078) that were all bearing wear. Based on the case similarity, the conclusion "bearing wear" was calculated with a confidence level of 0.88. LLM analysis of the test report text, combined with domain knowledge, deduced that "it may be due to bearing wear or rotor imbalance", and output two possible conclusions and their confidence levels: "bearing wear" (0.82) and "rotor imbalance" (0.45).

[0096] (3) Construct a confidence propagation graph, with nodes including the conclusions of the three engines and intermediate nodes.

[0097] (4) After propagation calculation, the confidence of the "bearing wear" node increased to 0.94, and the confidence of the "rotor imbalance" node decreased to 0.23.

[0098] (5) Conflict detection found that "rotor imbalance" conflicted with most conclusions, so conflict resolution was initiated: LLM analysis suggests that the vibration spectrum is dominated by the first harmonic, which is more consistent with the characteristics of bearing wear. After improving the integration of evidence theories, the overall confidence level of the final conclusion "bearing wear" is 0.96.

[0099] (6) Generating an interpretable report includes: Rule reasoning path: Triggering rule R12 (confidence level 0.95); Graph algorithm retrieval results: Similar cases F012, F035, F078 (similarity 0.85, 0.82, 0.79); LLM analysis process: thought chain "vibration characteristics → spectrum analysis → probability ranking"; Fusion process: confidence propagation graph, conflict detection results, final fusion formula; Conclusion: Bearing wear (confidence level 0.96), it is recommended to check the bearing lubrication system.

[0100] This system supports interactive user exploration and interpretable reports: Macro view: Shows the main contributions of the three modules "rule reasoning", "graph algorithm" and "LLM", as well as the confidence score of 0.96 after fusion.

[0101] Mid-view: Click the “Rule Reasoning” module to expand the specific triggered rule R12 and its preconditions, and show the degree to which each precondition is satisfied.

[0102] Microscopic view: Click "Vibration characteristics" to expand the original sensor data waveform and mark the time and amplitude of exceeding the standard.

[0103] Conflict View: Click "Conflict Resolution" to see the reasons why "rotor imbalance" was ruled out, including spectrum analysis charts, historical case comparisons, etc.

[0104] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for confidence propagation and fusion of multi-source inference results, characterized in that, Includes the following steps: Obtain the inference results from at least three heterogeneous inference sources and initialize the confidence of the inference results to obtain the initial confidence. Based on the initial confidence level, a confidence propagation graph is constructed with the reasoning conclusions as nodes and the dependencies between reasoning conclusions as edges, where the weight of the edges is calculated based on the evidence set shared between nodes. An iterative propagation algorithm is used to update the node confidence based on the confidence of the predecessor node and the weight of the corresponding edge in the confidence propagation graph, and the propagation conclusion is obtained based on the updated node confidence. The post-propagation conclusion is subjected to multi-type conflict detection, and conflict resolution is triggered based on the degree of detected conflict to obtain the resolved conclusion; The conclusion after resolution is processed using evidence theory to fuse several pieces of evidence supporting the same conclusion. Based on the fusion result, the overall confidence level is obtained and the overall conclusion is output.

2. The method for confidence propagation and fusion of multi-source inference results according to claim 1, characterized in that, The process of initializing the confidence level of the inference results to obtain the initial confidence level includes: A performance evaluation matrix is ​​constructed based on the historical accuracy of each inference source on different task types; After identifying the task type of the current query, the historical accuracy of each inference source on the identified task type is obtained from the performance evaluation matrix as an adjustment factor, and the original confidence of the inference result is dynamically adjusted to obtain the adjusted confidence. The adjusted confidence level is normalized to obtain the initial confidence level.

3. The method for confidence propagation and fusion of multi-source inference results according to claim 1, characterized in that, The process of calculating the edge weights based on the evidence set shared between nodes includes: The evidence sets supporting each node are obtained separately. The improved Tanimoto coefficient is used to calculate the edge weight from node i to node j based on the intersection modulus of the evidence sets of node i and node j, the modulus of the evidence set of node i, the modulus of the evidence set of node j, and the smoothing factor.

4. The method for confidence propagation and fusion of multi-source inference results according to claim 1, characterized in that, The process of updating node confidence using the iterative propagation algorithm includes: In each iteration, the confidence of a node is calculated based on the damping factor, the initial confidence of the node, and the weighted sum of the confidence of all predecessor nodes and the corresponding edge weights. This results in the confidence of the node after the current iteration. The iteration is repeated until the change in the confidence of all nodes is less than a preset threshold, at which point the update stops.

5. The method for confidence propagation and fusion of multi-source inference results according to claim 1, characterized in that, The process of performing multi-type conflict detection on the post-propagation conclusions includes: The number of syntax layer conflicts is obtained by processing the post-propagation conclusions based on the syntax layer conflict detection rules. The number of semantic layer conflicts is obtained by processing the post-propagation conclusions based on semantic layer conflict detection rules. The number of pragmatic layer conflicts is obtained by processing the post-propagation conclusions based on pragmatic layer conflict detection rules. The number of time-layer conflicts is obtained by processing the post-propagation conclusions based on the time-layer conflict detection rules. The number of conflicts at each layer is quantified to obtain the degree of conflict at each layer. The total degree of conflict is calculated by weighting and summing the degree of conflict at each layer based on the preset layer weight coefficient.

6. The method for confidence propagation and fusion of multi-source inference results according to claim 1, characterized in that, The process of triggering conflict resolution based on the detected level of conflict to obtain the resolution conclusion includes: When the total conflict level exceeds the preset conflict threshold, the rule chain that triggered the conflict is traced back based on the conclusion node involved in the conflict to perform a rule consistency check and obtain the rule check result. Based on the conflict point information, prompt words are constructed and processed by a large language model to obtain resolution suggestions; An evidence distance matrix is ​​constructed based on the Jousselme distance between each evidence source, and the relative credibility of each evidence source is obtained by processing the evidence distance matrix. The conclusion after resolution is obtained by comprehensively processing the results of the rule check, resolution suggestions, and relative credibility.

7. The method for confidence propagation and fusion of multi-source inference results according to claim 6, characterized in that, The process of fusing several pieces of evidence supporting the same conclusion using evidence theory after the resolution includes: The confidence level of each piece of evidence is converted into a basic probability assignment to obtain the basic probability assignment function for each piece of evidence; The Dempster-Shafer evidence theory is used to perform orthogonal sum operations on the basic probability assignment functions of multiple pieces of evidence supporting the same conclusion to obtain the fused basic probability assignment. The fused basic probability allocation is processed to obtain the comprehensive confidence level.

Images