Hoist vertical domain large model credible thinking chain reasoning method

Abstract

This invention discloses a reliable thought chain reasoning method for a large-scale vertical model of a hoist. It extracts abnormal feature vectors from the hoist's real-time operating data and links them to a domain knowledge graph. Through semantic association evaluation and relevance pruning guided by a large language model, a prior knowledge subgraph is constructed and structurally injected into the large model's prompts. The large model is instantiated as a diagnostic agent, using a structured interaction template of thinking-action observation to force it to logically navigate within the physical boundaries of the knowledge subgraph. A tree-like thinking strategy and a heuristic state evaluator are introduced for multi-path optimization and intelligent pruning. By extracting the logarithmic probability of the large model's decision nodes, the cumulative confidence of the entire path is calculated, outputting a highly reliable diagnostic report and a visualized physical traceability trajectory. This invention effectively suppresses the illusion phenomenon of general large models in industrial scenarios, achieving physical interpretability and high reliability of conclusions in the hoist fault diagnosis process.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and industrial intelligent operation and maintenance technology, specifically involving a reliable thinking chain reasoning method for a vertical domain large model of a hoist. Background Technology

[0002] As the crucial link between the surface and underground production systems, the mine hoist's operational safety directly impacts the normal production order and personnel safety of the coal mine, making it a top priority for the safety of the entire mine's electromechanical system. During long-term operation, the hoist system faces challenges such as high loads and continuous operation, making it highly susceptible to safety hazards such as main bearing wear, braking system failure, or abnormal motor temperature rise. Due to the high coupling of multiple mechanical, electrical, and hydraulic parameters within the hoist system, the fault mechanisms are complex and exhibit strong nonlinear characteristics. Failure to identify faults in a timely manner can lead to serious accidents such as rope breakage and overwinding. Therefore, to ensure the safety and reliability of vertical transportation in mines and effectively avoid production losses due to equipment failure, intelligent fault diagnosis and maintenance decision-making for mine hoists are of significant practical importance.

[0003] Driven by artificial intelligence and large-scale modeling technologies, intelligent operation and maintenance in coal mines has entered a new stage of development. The large-scale model for hoists, through thought chain technology, can decompose complex diagnostic logic, enabling dynamic reasoning and early warning of the hoist's operating status. However, hoist operation and maintenance systems require extremely high decision-making credibility. How to effectively suppress the illusionary phenomena of general large-scale models in complex industrial reasoning processes, and ensure that the conclusions generated by AI conform to the objective laws of the physical world, has become a key issue in coal mine equipment safety management and digital transformation. Traditional neural network-based diagnostic technologies often operate in a black-box mode when handling logical tracing, and their output results lack interpretability, making them difficult for on-site operation and maintenance engineers to fully trust.

[0004] Knowledge graph-based reasoning methods offer a novel solution to the aforementioned problems. Knowledge graphs possess advantages such as high structure, rigorous logic, and strong traceability, exhibiting particularly strong adaptability when handling complex physical topological relationships. Knowledge graphs can not only record the static attribute information of entities but also explicitly express the causal chain of faults through path search. Based on this, deeply integrating the semantic understanding capabilities of large-scale models with the structured constraints of knowledge graphs is gradually becoming one of the core technological directions in mine safety monitoring. In particular, by constructing dynamic knowledge subgraphs, real-time and accurate physical boundary constraints can be provided for the reasoning process of large-scale models. Current technologies lack techniques for effectively integrating dynamic knowledge subgraph constraints with the reasoning chain of large-scale models, failing to achieve step-by-step alignment and quantifiable credibility during the reasoning process. Summary of the Invention

[0005] The purpose of this invention is to provide a reliable reasoning chain method for a large vertical model of a hoist, which can effectively eliminate the illusion phenomenon of general large models in industrial scenarios, realize the physical interpretability of the hoist fault diagnosis process and the high reliability of the conclusions, provide a reliable basis for intelligent operation and maintenance decision-making of large electromechanical equipment in coal mines, and is applicable to long-term online monitoring and precise fault diagnosis of deep well hoist systems.

[0006] To achieve the above objectives, this invention provides a reliable thought chain reasoning method for a large vertical model of an elevator, comprising the following steps: S1: Dynamic construction and structured injection of prior knowledge subgraphs; Multi-source sensor stream data from the mine hoist monitoring system is acquired. Signal processing algorithms are used to extract high-frequency features of the main bearing vibration spectrum, harmonic components of the motor stator current, and residual oil pressure values ​​of the braking system. These data are then fused and mapped into an abnormal feature vector characterizing the current abnormal operating state of the hoist. ; By utilizing named entity recognition and semantic alignment techniques, the collected abnormal feature vectors are... Mapping to a pre-built domain knowledge graph Specific semantic anchor entities in Establish the connection between physical perception data and logical knowledge space; anchor point entity Execute for the starting node The algorithm employs a hop-based breadth-first search. During each hop of graph expansion, it utilizes a lift-up vertical domain large model as an intelligent semantic pruner to compute candidate neighbor nodes. Anomaly feature vector extracted from the current device Semantic association score between Where M represents the preset graph search depth: ; In the formula, Let j represent the i-th candidate neighbor entity node in the knowledge graph, and j represent the candidate neighbor index. Connect the current node with candidate neighbor nodes in the knowledge graph. edge attributes, This represents the set of candidate neighbor entities in the current search layer. This indicates that the vertical domain large model of the booster accepts textual features, relationships, and entities as input, and outputs a log probability score of the degree of matching. Based on a preset significance threshold The search path is filtered to remove interfering nodes that are weakly related to the current operating conditions, and a prior knowledge subgraph with physical and logical support is constructed. ; then It is converted into a structured text sequence containing entity nodes, relation edges, and topological structure information, and injected as inference constraint hints into the inference context of the lift machine vertical domain large model; S2: Guided thought chain generation based on graph path constraints; The vertical domain large model of the booster is instantiated into a diagnostic agent with interactive capabilities. During the inference process, a hybrid generation strategy combining thought chain (CoT), thought tree (ToT), and thought graph (GoT) is introduced. In the thought graph GoT mode, the operation operators are merged. Multiple intermediate inference chains containing complementary fault information are merged to generate new high-level semantic inference nodes. The new high-level semantic inference nodes inherit the physical constraint information and logical relationships of their corresponding parent nodes. Configure a standardized API instruction space for diagnostic agents Instruction space It includes preset function instructions for interacting with knowledge graphs and sensor databases, and forces the diagnostic agent to follow a closed-loop mechanism of thinking, acting and observing in each round of reasoning iteration; A heuristic state evaluator is introduced to perform real-time quality assessment of each candidate branch and calculate the current inference state. Physical potential score To determine whether the path conforms to the physical operating logic of the hoist: ; In the formula, and These are preset weighting coefficients used to balance the ratio of semantic judgments to physical knowledge constraints in the model. This represents the j-th candidate state in the t-th round of reasoning. This represents the semantic rationality scoring function output by the vertical domain large model of the elevator based on knowledge subgraph constraints. This is a physical consistency indicator function. Refers to the state Jump to The resulting logical reasoning edges, if the reasoning edges generated by the diagnostic agent are... It does not exist in the pre-constructed prior knowledge subgraph If the path is deemed unsuitable, an evaluation penalty will be imposed. A Top-N-based breadth-first search strategy is adopted, in which only the state branch with the highest potential score is retained in each inference depth to continue exploring downwards. By explicitly pruning invalid logic paths, it is ensured that the inference chain converges to a high-confidence diagnostic conclusion. Here, N is the preset search width parameter. S3: Credibility assessment and result verification of inference trajectory; Extract from the final generated reasoning path Log probability distribution of key decision nodes The native decision confidence of the vertical large model of the hoist is calculated in each logical transformation step t using the normalized Softmax function. ; Geometric mean mapping is applied to all steps of the entire path to calculate the cumulative confidence score of the entire path. : ; Where C represents the complete reasoning path; Introducing a logical eccentric function to evaluate the consistency of reasoning trajectories By constructing guiding suffix prompts at the end of the inference process, the large-scale vertical model of the booster is required to determine the rationality and truth value of the entire inference trajectory it generates and output a consistency probability value. This represents the self-examination probability score of the large-scale model in the vertical domain for consistency across the entire inference trajectory; The final credibility score is calculated by combining the original probability with the self-assessment score. : ; Where α is the confidence fusion weight coefficient, and its value ranges from 1 to 10. ; Using path consistency verification function Determine the fault conclusion entity generated by the large vertical model of the hoist. Entity at the end of the reasoning path The degree of topological alignment between them, and the consistency verification of the execution path: ; In the formula, This represents the shortest topological distance between entities in the knowledge graph. This is a physical consistency indicator function, which takes the value 1 when the constraints are met and 0 otherwise. d This is the permissible semantic drift threshold, typically set to {0, 1, 2}, representing the tolerance for semantic drift. The minimum confidence threshold for the diagnostic results output, when the following conditions are met. That is, the generated fault conclusion entity must be located within a reasonable neighborhood of the end of the reasoning chain in the knowledge graph topology. ), and from a physical and logical perspective, exclude the illusory output of the large model, and Exceeding the safety threshold At the same time, the reliability of the diagnostic conclusions is ensured from a probabilistic and statistical perspective, and a diagnostic report is output that includes fault location, evolutionary tracing, and physical evidence chain.

[0007] As a further aspect of the present invention: In S2, the API instruction space This includes: RetrieveNode for accurately locating entity identifiers, NodeFeature for extracting real-time operating conditions and rated parameters, NeighbourCheck for exploring physical connectivity relationships, and NodeDegree for evaluating the topological importance of nodes.

[0008] As a further aspect of the present invention: In S2, the NodeDegree instruction provides a quantitative basis for evaluating the logical complexity of inference branches by returning the connectivity degree of a given node under a specific edge type; the NodeFeature instruction provides a hard factual basis for physical consistency verification for the diagnostic agent by retrieving the real-time monitoring values ​​and rated design parameters of the entity.

[0009] As a further aspect of the present invention, S2 also includes an automatic graph search and hierarchical pruning mechanism: incremental search of the graph is performed through an alternating approach of language generation and structured retrieval. At each unvisited entity node, the large vertical domain model of the lift mechanism is first used to guide the retrieval and pruning of irrelevant relation types. Subsequently, secondary discovery and filtering of neighboring entities are performed under the selected relation. This hierarchical filtering mechanism ensures that the reasoning trajectory remains computationally processable under the guidance of natural language.

[0010] As a further aspect of the present invention: In S2, the GoT merging operation operator By integrating the contextual information of multiple inference paths, the robustness of the large vertical model of the lift machine in handling multiple concurrent failures of the lift machine is enhanced. The merged inference node is added to the inference graph as a new graph node, and its logical edges are inherited from the corresponding parent node.

[0011] As a further aspect of the present invention: in S3, The logical self-examination mechanism effectively suppresses the illusion of the elevator vertical domain large model by calculating the difference in probability distribution when predicting Yes and No. If the value is below a preset threshold, the diagnostic agent will trigger a backtracking reconstruction of the corresponding inference branch.

[0012] As a further aspect of the present invention: In S3, path consistency verification is performed using the shortest topological distance. By anchoring generative language descriptions to rigorous industrial knowledge topologies, we ensure that the final fault conclusions are not only semantically coherent but also physically and logically accessible.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: The prior knowledge subgraph constructed in this invention is scored through semantic association. By screening key entities, we can fully utilize the large vertical model of the hoist to analyze feature vectors from multiple sensor sources. x A deeper understanding reveals that this subgraph constraint not only enhances abnormal correlation signals caused by minor equipment defects, but also effectively filters out random noise interference caused by the complex electromagnetic environment of industrial sites, as well as the illusion phenomenon generated by the large vertical model of the hoist in long-range inference. By utilizing structured templates and standardized API instruction spaces for thinking action observation, unreliable factors generated by black-box models can be effectively eliminated, improving the reliability of monitoring and diagnostic data. This guided thinking chain structure ensures diagnostic sensitivity and can quickly capture logical connections across subsystems in the early stages of a fault, thereby significantly improving the detectability and source tracing accuracy of complex fault signals. Using physical consistency verification function topological distance The judgment criteria can more accurately determine the location of the hoist failure and its inducing mechanism; By using a tree-like thinking search algorithm guided by a large vertical model of the hoist to track the reasoning process, and combining it with a heuristic state evaluator for path pruning, we can accurately eliminate misleading branches that are not physically possible, thereby helping to achieve more efficient equipment health management. Adopting based on The real-time credibility classification method of the logic self-examination mechanism can realize real-time logic self-examination of the lifting machine's operation and ensure that it can be quickly corrected when abnormal reasoning deviation occurs, effectively preventing safety hazards caused by AI's erroneous decisions. For matching incomplete sensor feature sequences with mechanism templates in the knowledge base, the thinking chain tracking algorithm guided by the large model of the hoist vertical domain can ensure that logical paths can still be extracted along the limited search zone even when data is missing or signal segmentation is incomplete. This can effectively describe the similarity between actual working conditions and standard mechanism models, thereby accurately obtaining fault evolution characteristics and equipment damage levels. Attached Figure Description

[0014] Figure 1 This is a flowchart of the reliable thought chain reasoning method for the vertical domain large model of the hoisting machine of the present invention; Figure 2 This is a flowchart of the implementation steps of the present invention; Figure 3 This is a schematic diagram of the automatic graph search and inference chain expansion driven by the large vertical model of the elevator in this invention; Figure 4 This is a comparative diagram illustrating the reasoning strategies used in this invention to construct logical chains by combining the vertical domain large model of the elevator with structured knowledge. Detailed Implementation

[0015] The present invention will be further illustrated by the following examples.

[0016] like Figure 1 and Figure 2 As shown, a reliable thought chain reasoning method for a large vertical model of a hoist is applied to a hoist fault diagnosis scenario where a large vertical model of the hoist is deployed. The method includes the following steps: S1: Dynamic construction and structured injection of prior knowledge subgraphs; S11: Establish a low-level probability generation model for hoist operation and maintenance scenarios, and define the probability distribution of the large-scale model for the hoist vertical domain as follows: For any language token sequence of length n generated by the model Its generation probability follows the cumulative distribution of the token sequence: To achieve accurate mapping of the complex operating states of mine hoists, multi-source sensor stream data from the mine hoist monitoring system is accessed in real time via a data gateway. Signal processing algorithms are used to extract high-frequency features of the main bearing vibration spectrum, harmonic components of the motor stator current, and residual oil pressure values ​​of the braking system. These are then fused and mapped into an abnormal feature vector characterizing the current abnormal operating state of the hoist. x Preferably, the characteristic frequencies of the main bearing vibration signal are extracted by fast Fourier transform, and these frequencies, together with the motor stator winding temperature and the brake shoe clearance deviation, constitute an abnormal feature vector. x This allows for a comprehensive characterization of the state of multiple electromechanical-hydraulic systems.

[0017] S12: Pre-construct a heterogeneous knowledge graph for the hoisting machine domain ,in E It represents a collection of entities that includes hoist components, fault symptoms, and maintenance procedures. R It represents a set of relationships that include causality, membership, and spatial connections. h represents the head entity in the knowledge graph triple, r represents the relationship between entities, and t represents the tail entity.

[0018] Due to strong electromagnetic interference and unstable operating conditions at the mine site, in order to suppress misjudgments caused by noise, named entity recognition and semantic alignment techniques are used to analyze the collected abnormal feature vectors. x Mapping to knowledge graph G Specific semantic anchor entities in Establish the connection between physical perception data and logical knowledge space; S13: Anchor point entity An M-hop breadth-first search (BFS) is performed on the starting node to explore possible evolutionary paths and related factors of the failure. During the graph expansion process at each hop, the elevator vertical domain large model is invoked as an intelligent semantic pruner to calculate candidate neighbor nodes by constructing specific evaluation prompts. Compared with the current device anomaly feature vector x Semantic association score between Where M represents the preset graph search depth: ; In the formula, Let j represent the i-th candidate neighbor entity node in the knowledge graph, and j represent the candidate neighbor index. Connect the current node with candidate neighbor nodes in the knowledge graph. edge attributes, This represents the set of candidate neighbor entities in the current search layer. This indicates that the vertical domain large model of the booster accepts textual features, relationships, and entities as input, and outputs a log probability score of the degree of matching. S14: Based on the set significance threshold The search path is filtered to remove interfering nodes that are weakly related to the current operating conditions. For the preferred nodes, a significance threshold is set. Only retain Using the nodes and edges, construct a highly compressed subgraph of prior knowledge that is supported by physical logic. ; then It is transformed into a structured text sequence containing topological structure, and injected as a hard inference constraint into the inference context prompts of the vertical domain large model of the booster; S2: Guided thought chain generation based on graph path constraints; S21: The large-scale vertical model of the booster is instantiated into a diagnostic agent with interactive capabilities. To ensure that the reasoning process strictly follows industrial mechanisms, a hybrid generation strategy of thought chain (CoT), thought tree (ToT), and thought graph (GoT) is introduced during the reasoning process to form a structured template for thinking action observation. This forces the agent to use a standardized API instruction space during the generation of thought chains. Interacting with a physics knowledge base; such as Figure 4 The diagram shows a comparison of reasoning strategies for constructing logical chains using a large vertical model of the elevator and structured knowledge.

[0019] In the GoT mind mapping model, the reasoning process is modeled as a directed graph structure. This allows the diagnostic agent to merge operation operators. The two intermediate reasoning chains with complementary information are integrated into a higher-level semantic node, and the merging formula is expressed as: ; in, and These represent intermediate semantic nodes in different branches of reasoning. This represents the node fusion function, resulting in a new semantic node after merging. Inheriting the physical constraint information of the parent nodes enables dynamic refinement of the reasoning space; GoT merge operation operator By integrating the contextual information of multiple inference paths, the robustness of the large vertical model of the lift machine in handling multiple concurrent failures of the lift machine is enhanced. The merged inference node is added to the inference graph as a new graph node, and its logical edges are inherited from the corresponding parent node.

[0020] S22: Configuring a standardized API instruction space for diagnostic agents It mandates that the diagnostic agent follow a closed-loop mechanism of thinking, acting, and observing in each round of inference iteration; instruction space This includes: RetrieveNode for accurately locating entity identifiers, NodeFeature for extracting real-time operating conditions and rated parameters, NeighbourCheck for exploring physical connectivity relationships, and NodeDegree for evaluating the topological importance of nodes. Furthermore, the NodeDegree instruction provides a quantitative basis for evaluating the logical complexity of inference branches by returning the connectivity degree of a given node under a specific edge type; while the NodeFeature instruction provides a hard factual basis for physical consistency verification for the diagnostic agent by retrieving the real-time monitoring values ​​and rated design parameters of the entity.

[0021] S23: To evaluate reasoning quality in real time, a heuristic state evaluator is introduced to perform real-time quality evaluation on each candidate branch of the mind tree or mind map and calculate the current reasoning state. Physical potential score To determine whether the path conforms to the physical operating logic of the hoist: ; In the formula, and These are preset weighting coefficients used to balance the ratio of semantic judgments to physical knowledge constraints in the model. This represents the j-th candidate state in the t-th round of reasoning. This represents the semantic rationality scoring function output by the vertical domain large model of the elevator based on knowledge subgraph constraints. This is a physical consistency indicator function. Refers to the state Jump to The resulting logical reasoning edges, if the reasoning edges generated by the diagnostic agent are... It does not exist in the pre-constructed prior knowledge subgraph If the path is not specified, an evaluation penalty will be imposed.

[0022] S24: Employs a Top-N based breadth-first search strategy, retaining only the top performers with the highest potential scores at each inference depth. Each state branch continues to explore downwards, ensuring that the inference chain converges to a high-confidence diagnostic conclusion with limited computational resources by explicitly pruning invalid logical paths (paths subject to evaluation penalties). Here, N is a preset search width parameter. Furthermore, S2 also includes automatic graph search and hierarchical pruning mechanisms: such as Figure 3 As shown, incremental search of the graph is performed by alternating language generation and structured retrieval. At each unvisited entity node, the large vertical model of the lift mechanism is first used to guide the retrieval and prune irrelevant relation types. Then, secondary discovery and filtering of neighboring entities are performed under the selected relation. This hierarchical filtering mechanism ensures that the reasoning trajectory remains computationally processable under the guidance of natural language.

[0023] S3: Credibility assessment and result verification of inference trajectory; S31: To provide a high-confidence basis for coal mine safety decisions, multi-level verification is performed on the final generated reasoning chain, and extracting data from the final generated reasoning path. Log probability distribution of key decision nodes The native decision confidence of the vertical large model of the hoist is calculated in each logical transformation step using the normalized Softmax function. This characterizes the internal determinism of the vertical domain large model of the lifting machine for the causal relationship of a specific fault, where Lt specifically represents the log probability distribution of the target output Token corresponding to the t-th logical transformation step, and K specifically represents the number of key decision nodes in the inference path. S32: Apply a geometric mean mapping to all steps of the entire path and calculate the cumulative confidence score of the entire path. : ; Where C represents the complete inference path, the geometric mean confidence score of the entire path is obtained by using the arithmetic mean in logarithmic space followed by exponential mapping. This is to reduce the confidence compression effect caused by long paths.

[0024] S33: Introduction The logical self-examination mechanism constructs guiding suffix prompts at the end of the inference process for self-reflection, requiring the lifter's vertical domain large model to determine the rationality and truth value of the entire inference trajectory it generates; it extracts the logarithmic probability of the Yes token output by the lifter's vertical domain large model and converts it into a consistency probability value. ,in, The self-examination probability score of the elevator vertical domain large model represents the consistency of the entire inference trajectory, and is used to correct the illusion problem that may exist in the elevator vertical domain large model; The logical self-examination mechanism effectively suppresses the illusion of the elevator vertical domain large model by calculating the difference in probability distribution when predicting Yes and No. If the value is below a preset threshold, the diagnostic agent will trigger a backtracking reconstruction of the corresponding inference branch.

[0025] S34: Combine the original probability with the self-assessment score to calculate the corrected final credibility score. : ; Where α is the confidence fusion weight coefficient, and its value ranges from (0,1); S35: Utilizing Path Consistency Verification Functions Determine the fault conclusion entity generated by the large vertical model of the hoist. Entity at the end of the reasoning path The degree of topological alignment between them is evaluated, path consistency verification is performed, and the final credibility score is combined. Make comprehensive decisions: ; In the formula, This represents the shortest topological distance between entities in the knowledge graph. This is a physical consistency indicator function, which takes the value 1 when the constraints are met and 0 otherwise. d This is the permissible semantic drift threshold, typically set to {0, 1, 2}, representing the tolerance for semantic drift. The minimum confidence threshold for the diagnostic results output, when the following conditions are met. That is, the generated fault conclusion entity must be located within a reasonable neighborhood of the end of the reasoning chain in the knowledge graph topology. ), and from a physical and logical perspective, exclude the illusory output of the large model, and Exceeding the safety threshold At the same time, the reliability of the diagnostic conclusions is ensured from a probabilistic and statistical perspective, and a diagnostic report is output that includes fault location, evolutionary tracing, and physical evidence chain.

[0026] Furthermore, path consistency verification is performed using the shortest topological distance. By anchoring generative language descriptions to rigorous industrial knowledge topologies, we ensure that the final fault conclusions are not only semantically coherent but also physically and logically accessible.

[0027] Due to the large number of parameters in the vertical domain model of mining, conventional reasoning is prone to logical chain breaks when faced with complex multi-hop logic. However, the search method based on knowledge subgraph constraints is highly robust to logical divergence caused by the increase in the number of reasoning layers, and therefore can handle large-scale multi-point linkage monitoring tasks across systems such as cables, brakes, and main drives.

[0028] This invention provides a reliable thought chain reasoning method for a large-scale vertical model of a hoist. It deeply embeds structured knowledge into the thought chain generation process of the large model, and through a physical consistency indicator function, ensures that the diagnostic path generated by the model not only conforms to semantic logic but also strictly follows the physical operating laws of the hoist's electromechanical-hydraulic system. This ensures synchronization between digital spatial reasoning and physical entity logic, and significantly improves the accuracy of fault diagnosis for deep-well hoists. Furthermore, it can be deeply integrated with existing mine monitoring systems to achieve reliable monitoring of the hoist's operating status and precise maintenance decisions, resulting in significant social safety benefits and economic application value.

Claims

1. A reliable reasoning chain method for a large-scale vertical model of a hoisting machine, characterized in that, Includes the following steps: S1: Dynamic construction and structured injection of prior knowledge subgraphs; Multi-source sensor stream data from the mine hoist monitoring system is acquired. Signal processing algorithms are used to extract high-frequency features of the main bearing vibration spectrum, harmonic components of the motor stator current, and residual oil pressure values ​​of the braking system. These data are then fused and mapped into an abnormal feature vector characterizing the current abnormal operating state of the hoist. ; By utilizing named entity recognition and semantic alignment techniques, the collected abnormal feature vectors are... Mapping to a pre-built domain knowledge graph Specific semantic anchor entities in Establish the connection between physical perception data and logical knowledge space; anchor point entity Execute for the starting node The algorithm employs a hop-based breadth-first search. During each hop of graph expansion, it utilizes a lift-up vertical domain large model as an intelligent semantic pruner to compute candidate neighbor nodes. Anomaly feature vector extracted from the current device Semantic association score between Where M represents the preset graph search depth: ； In the formula, Let j represent the i-th candidate neighbor entity node in the knowledge graph, and j represent the candidate neighbor index. Connect the current node with candidate neighbor nodes in the knowledge graph. edge attributes, This represents the set of candidate neighbor entities in the current search layer. This indicates that the vertical domain large model of the booster accepts textual features, relationships, and entities as input, and outputs a log probability score of the degree of matching. Based on a preset significance threshold The search path is filtered to remove interfering nodes that are weakly related to the current operating conditions, and a prior knowledge subgraph with physical and logical support is constructed. ; then It is converted into a structured text sequence containing entity nodes, relation edges, and topological structure information, and injected as inference constraint hints into the inference context of the lift machine vertical domain large model; S2: Guided thought chain generation based on graph path constraints; The vertical domain large model of the booster is instantiated into a diagnostic agent with interactive capabilities. During the inference process, a hybrid generation strategy combining thought chain (CoT), thought tree (ToT), and thought graph (GoT) is introduced. In the thought graph GoT mode, the operation operators are merged. Multiple intermediate inference chains containing complementary fault information are merged to generate new high-level semantic inference nodes. The new high-level semantic inference nodes inherit the physical constraint information and logical relationships of their corresponding parent nodes. Configure a standardized API instruction space for diagnostic agents Instruction space It includes preset function instructions for interacting with knowledge graphs and sensor databases, and mandates that the diagnostic agent follow a closed-loop mechanism of thinking, acting, and observing in each round of reasoning iteration; A heuristic state evaluator is introduced to perform real-time quality assessment of each candidate branch and calculate the current inference state. Physical potential score To determine whether the path conforms to the physical operating logic of the hoist: ； In the formula, and These are preset weighting coefficients used to balance the ratio of semantic judgments to physical knowledge constraints in the model. This represents the j-th candidate state in the t-th round of reasoning. This represents the semantic rationality scoring function output by the vertical domain large model of the elevator based on knowledge subgraph constraints. This is a physical consistency indicator function. Refers to the state Jump to The resulting logical reasoning edges, if the reasoning edges generated by the diagnostic agent are... It does not exist in the pre-constructed prior knowledge subgraph If the path is deemed unsuitable, an evaluation penalty will be imposed. A Top-N-based breadth-first search strategy is adopted, in which only the state branch with the highest potential score is retained in each inference depth to continue exploring downwards. By explicitly pruning invalid logic paths, it is ensured that the inference chain converges to a high-confidence diagnostic conclusion. Here, N is the preset search width parameter. S3: Credibility assessment and result verification of inference trajectory; Extract from the final generated reasoning path Log probability distribution of key decision nodes The native decision confidence of the vertical large model of the hoist is calculated in each logical transformation step t using the normalized Softmax function. ; Geometric mean mapping is applied to all steps of the entire path to calculate the cumulative confidence score of the entire path. : ； Where C represents the complete reasoning path; Introducing a logical eccentric function to evaluate the consistency of reasoning trajectories By constructing guiding suffix prompts at the end of the inference process, the large-scale vertical model of the booster is required to determine the rationality and truth value of the entire inference trajectory it generates and output a consistency probability value. This represents the self-examination probability score of the large-scale model in the vertical domain for consistency across the entire inference trajectory; The final credibility score is calculated by combining the original probability with the self-assessment score. : ； Where α is the confidence fusion weight coefficient, and its value ranges from 1 to 10. ; Using path consistency verification function Determine the fault conclusion entity generated by the large vertical model of the hoist. Entity at the end of the reasoning path The degree of topological alignment between them, and the consistency verification of the execution path: ； In the formula, This represents the shortest topological distance between entities in the knowledge graph. This is a physical consistency indicator function, which takes the value 1 when the constraints are met and 0 otherwise. d This is the permissible semantic drift threshold, typically set to {0, 1, 2}, representing the tolerance for semantic drift. The minimum confidence threshold for the diagnostic results output, when the following conditions are met. That is, the generated fault conclusion entity must be located within a reasonable neighborhood of the end of the reasoning chain in the knowledge graph topology. ), and from a physical and logical perspective, exclude the illusory output of the large model, and Exceeding the safety threshold At the same time, the reliability of the diagnostic conclusions is ensured from a probabilistic and statistical perspective, and a diagnostic report is output that includes fault location, evolutionary tracing, and physical evidence chain.

2. The reliable thought chain reasoning method for a large vertical model of a hoist as described in claim 1, characterized in that, In S2, the API instruction space This includes: RetrieveNode for accurately locating entity identifiers, NodeFeature for extracting real-time operating conditions and rated parameters, NeighbourCheck for exploring physical connectivity relationships, and NodeDegree for evaluating the topological importance of nodes.

3. The reliable thought chain reasoning method for a large vertical model of a hoist as described in claim 2, characterized in that, In S2, the NodeDegree instruction provides a quantitative basis for evaluating the logical complexity of inference branches by returning the connectivity degree of a given node under a specific edge type; the NodeFeature instruction provides a hard factual basis for physical consistency verification for the diagnostic agent by retrieving the real-time monitoring values ​​and rated design parameters of the entity.

4. A reliable thought chain reasoning method for a large vertical model of a hoisting machine according to any one of claims 1-3, characterized in that, S2 also includes an automatic graph search and hierarchical pruning mechanism: incremental search of the graph is performed through an alternation of language generation and structured retrieval. At each unvisited entity node, the lifter vertical domain large model is first used to guide the retrieval and prune irrelevant relation types. Then, secondary discovery and filtering of neighboring entities are performed under the selected relation. This hierarchical filtering mechanism ensures that the reasoning trajectory remains computationally tractable under the guidance of natural language.

5. The reliable thought chain reasoning method for a large vertical model of a hoist as described in claim 1, characterized in that, In S2, the GoT merge operation operator By integrating the contextual information of multiple inference paths, the robustness of the large vertical model of the lift machine in handling multiple concurrent failures of the lift machine is enhanced. The merged inference nodes are added to the inference graph as new graph nodes, and their logical edges are inherited from their corresponding parent nodes.

6. The reliable thought chain reasoning method for a large vertical model of a hoist as described in claim 1, characterized in that, In S3 The logical self-examination mechanism effectively suppresses the illusion of the elevator vertical domain large model by calculating the difference in probability distribution when predicting Yes and No. If the value is below a preset threshold, the diagnostic agent will trigger a backtracking reconstruction of the corresponding inference branch.

7. The reliable thought chain reasoning method for a large vertical model of a hoist as described in claim 1, characterized in that, In S3, path consistency verification is performed using the shortest topological distance. By anchoring generative language descriptions to rigorous industrial knowledge topologies, we ensure that the final fault conclusions are not only semantically coherent but also physically and logically accessible.

Images