Large language model reasoning enhancement framework system for industrial complex decision tasks

Abstract

This invention relates to the field of large language model reasoning technology, and discloses an enhanced framework system for large language model reasoning for complex industrial decision-making tasks. The system includes a collaborative generative automated knowledge construction subsystem and a multi-agent graph structure reasoning subsystem. By decoupling knowledge management and logical reasoning, it achieves a paradigm shift from probabilistic prediction to logical deduction, addressing probabilistic illusions, lack of physical constraints, and multi-objective optimization conflicts in complex industrial decision-making tasks. This invention realizes a paradigm shift from "probabilistic generation" to "controllable logical deduction," significantly suppressing industrial illusions. Through the "constraint injection" module in the multi-agent graph structure reasoning subsystem, unstructured natural language tasks are transformed into controlled structured graph paths; first-order logic rules are introduced as hard constraints to control the boundaries of the LLM generation results, ensuring that process planning and parameter settings meet physical consistency.

Description

Technical Field

[0001] This invention relates to the field of large language model reasoning technology, and discloses a large language model reasoning enhancement framework system for complex industrial decision-making tasks. Background Technology

[0002] With the advancement of industry, leveraging artificial intelligence to solve complex problems in manufacturing scenarios has become a trend. Large Language Models (LLMs), due to their powerful natural language understanding and generation capabilities, have been attempted for application in areas such as industrial knowledge question answering and process assistance. However, industrial manufacturing scenarios are characterized by highly implicit knowledge, strict physical constraints, and strong coupling and conflict among multiple objectives (cost / accuracy / time / equipment capabilities). In complex decision-making tasks (such as manufacturing process planning and part design optimization), the effectiveness of LLMs does not meet the needs of enterprises.

[0003] The current mainstream technical solutions mainly include the following: Industry Model Fine-tuning (SFT) and Reinforcement Learning: Using industry data to train a general model so that it can master domain knowledge; Rule constraints and output validation: Add syntax or logic rule checks to the model output layer to intercept or correct the results; General large-scale model application architecture: includes workflow-based prompt word engineering, basic RAG (retrieval enhancement generation), and a general multi-agent collaborative system; Traditional Neural Networks and Digital Twins: Building high-fidelity simulation models or dedicated neural networks for specific devices or scenarios.

[0004] Although the above solutions have achieved certain results in some general scenarios, they still have significant technical bottlenecks when facing complex industrial decision-making tasks that require rigorous logical deduction, dynamic environmental adaptation and high interpretability, and are difficult to meet the requirements of "high reliability" and "dynamics" in industrial settings.

[0005] The existing technology has the following main drawbacks in practical industrial applications, which are also the technical problems that this invention aims to solve: 1. The fundamental conflict between generation mechanisms and physical constraints ("illusion" and unreliability): LLM is essentially based on predicting the next token based on statistical probability, rather than causally modeling the physical world. It lacks true logical deduction capabilities, and its outputs, while often eloquent, contain serious fallacies (i.e., "illusions") regarding physical facts, causal logic, or engineering constraints (such as tolerances and equipment limits). Simple prompt engineering or post-processing rule validation is insufficient to cover the almost infinite number of corner cases in industrial scenarios, rendering decision recommendations unusable or even posing security risks.

[0006] 2. Lagging knowledge updates and "catastrophic amnesia": The industrial environment is highly dynamic (equipment status and production schedules change in real time).

[0007] Fine-tuning schemes face the dilemma of "training is becoming obsolete," failing to respond to real-time data and easily causing the model to lose its original general reasoning ability (catastrophic forgetting).

[0008] Traditional knowledge graph (KG) construction relies on extensive human intervention from experts, resulting in long update cycles, high costs, and difficulty in handling fragmented, unstructured, and rapidly iterating industrial data.

[0009] 3. "Context dilution" and "error propagation" in long-link inference: When dealing with complex process planning, the attention mechanism of the single-unit model can cause key constraint information to be forgotten or diluted as the context length increases.

[0010] When general multi-agent systems lack unified structured guidance, they are prone to getting caught in a vicious cycle of disagreements or experiencing "error propagation and accumulation" in multiple rounds of dialogue, leading to a final decision that deviates from the goal. Furthermore, the reasoning process is a black box, lacking the interpretability and traceability required by industry.

[0011] 4. The contradiction between generalization ability and deployment cost: Traditional neural network / digital twin solutions are usually customized for specific devices, with poor engineering reusability. Each new scenario requires remodeling, and the expansion cost increases linearly. Moreover, the full simulation computing resources are too large, which cannot meet the needs of rapid iterative inference in industrial fields.

[0012] The general RAG scheme often loses the deep correlation information between entities (such as the causal chain between "bearing failure" and "motor overheating"), resulting in search results that are relevant but lack guidance for solving complex faults. Summary of the Invention

[0013] To address the aforementioned issues, this invention provides a large language model reasoning enhancement framework system for complex industrial decision-making tasks. The system includes a collaborative generative automated knowledge construction subsystem and a multi-agent graph structure reasoning subsystem. By decoupling knowledge management and logical reasoning, it achieves a paradigm shift from probabilistic prediction to logical deduction, thereby resolving probabilistic illusions, lack of physical constraints, and multi-objective optimization conflicts in complex industrial decision-making tasks.

[0014] Preferably, the generative automated knowledge construction subsystem includes a pre-built industrial ontology library, a document loading and processing module, an unstructured document conversion module, an entity extraction agent, and an entity alignment algorithm module, which can realize the automated closed-loop extraction and storage of industrial knowledge from unstructured multi-source data to structured graphs.

[0015] Preferably, it also includes a knowledge base composed of a graph database and a vector database, wherein the graph database is used to store a structured entity relationship network to support logical reasoning, and the vector database is used to store semantic vectors of knowledge fragments to support fuzzy semantic retrieval.

[0016] Preferably, the multi-agent graph structure reasoning subsystem includes a task planning agent, which is a fine-tuned large language model with tool invocation and industrial task planning capabilities. It can receive user instructions and generate a JSON list containing task dependencies, decompose it into an executable sub-task sequence, and is equipped with a deadlock prevention function, which can detect loops in task dependencies and prompt for replanning.

[0017] Preferably, the multi-agent graph structure reasoning subsystem further includes a graph enhancement retrieval module. This module can locate graph nodes based on the keywords of the current problem and expand outward from 1-hop to 2-hop to mine implicit association information between entities and transform the retrieved subgraph structure into natural language context for the large language model.

[0018] Preferably, the multi-agent graph structure reasoning subsystem includes a graph thinking chain manager. The manager can model the reasoning process as a tree or network structure, supporting the divergence, aggregation, and backtracking of the reasoning process. Each reasoning branch is scored and constrained by an independent evaluation agent. When all current branches are pruned, the system reverts to the parent node to prompt the large language model to regenerate the branches. The manager predefines graphs for four key industrial tasks: process planning, DFM optimization, work instructions, and quality inspection planning.

[0019] Preferably, the multi-agent graph structure reasoning subsystem includes a constraint injection and verification module. This module extracts numerical parameters from the text scheme generated by the large language model through regular expressions, queries the physical rules and parameter library for verification, and if the verification fails or violates the physical constraints, it is rejected and the large language model is required to regenerate.

[0020] Preferably, the multi-agent graph structure reasoning subsystem includes an interpretability output module. The output of this module not only includes decision suggestions, but also displays the reasoning path graph, reference sources, and confidence assessment, meeting the traceability requirements of industrial scenarios.

[0021] Preferably, the unstructured multi-source data includes industrial documents, production logs, and dialogue data. The generative automated knowledge construction subsystem can achieve low-cost, near real-time updates of industrial knowledge, avoiding the high computing costs of repeated retraining.

[0022] Preferably, the complex industrial decision-making tasks include manufacturing process planning, part design optimization, production scheduling, and complex fault diagnosis, and the system can meet the requirements of physical consistency, engineering interpretability, and high reliability in industrial settings.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A paradigm shift from "probabilistic generation" to "controllable logical deduction" was achieved, significantly suppressing industrial illusion. Through the "constraint injection" module in the multi-agent graph structure reasoning subsystem, unstructured natural language tasks were transformed into controlled structured graph paths. First-order logic rules (such as physical laws and equipment limits) were introduced as hard constraints to control the boundaries of the LLM generation results. This ensured that process planning and parameter settings met physical consistency.

[0024] 2. Overcoming the "error propagation" bottleneck of linear reasoning and improving decision-making accuracy for complex, long-chain tasks, this system integrates graph-based reasoning and topological evolutionary reasoning techniques. The reasoning process is no longer a unidirectional straight line, but rather constructed as a branching, merging, and backtracking network structure. The system supports multi-path divergent reasoning and automatically prunes erroneous paths through "weighted result merging" and "constraint feedback iteration" mechanisms. Even when facing extremely long process chains or multi-objective conflict scenarios, it can significantly reduce the error accumulation rate and improve the robustness of the final decision through global optimization rather than local greedy algorithms.

[0025] 3. It solves the challenges of updating and the associated costs associated with static knowledge bases in handling highly dynamic industrial environments. Utilizing the semantic understanding capabilities of generative agents, it automatically completes the closed-loop extraction and storage of unstructured multi-source data (documents, logs, dialogues) into a structured knowledge graph. This achieves low-cost, near real-time updates of industrial knowledge, retaining the generalization capabilities of LLM while possessing real-time awareness of dynamic production environments, thus avoiding the high computational costs of repeated retraining.

[0026] 4. Enhanced retrieval capabilities for "deep implicit relationships," improving the comprehensiveness of fault attribution and planning, by employing AgenticGraphRAG (Graph-Enhanced Retrieval) technology. The system not only retrieves text fragments but also performs multi-hop traversal on the knowledge graph to uncover implicit causal and relational paths between entities. In complex fault diagnosis or production scheduling, it can provide complete context containing causal chains, rather than fragmented information, significantly improving the model's depth of understanding of complex industrial problems.

[0027] 5. It endows the large model's decision-making process with "white-box" characteristics, meeting the needs of industrial auditing and traceability. Each reasoning step corresponds to a node transition or the triggering of a logical rule on the graph, and the reasoning results support interpretable traceability. Engineers can clearly view the logical path and source of the decision (which KG relationship or document was referenced), making AI decision-making possess the auditability necessary for industrial sites. Attached Figure Description

[0028] Figure 1 This is a diagram illustrating the technical solution architecture of the present invention. Figure 2 This is an example flowchart (process planning) of the present invention; Figure 3 This is an example flowchart of the present invention. Detailed Implementation

[0029] To overcome the endogeneity defects commonly found in existing industrial large-scale model applications, such as probabilistic illusions, lack of physical constraints, and multi-objective optimization conflicts, please refer to... Figure 1 This invention proposes a physical perception and knowledge-enhanced reasoning system for complex industrial scenarios. This system achieves a paradigm shift from "probabilistic prediction" to "logical deduction" by decoupling knowledge management and logical reasoning. Its core features include a collaborative, generative, automated knowledge construction subsystem and a multi-agent graph structure reasoning subsystem. 1. Generative Automated Knowledge Construction Subsystem: It realizes the automated closed loop of industrial knowledge from "text black box" to "structured graph", and solves the timeliness problem of traditional knowledge graph construction relying on a large number of human experts and being unable to cope with real-time changes in the production environment (such as equipment parameter adjustment and production schedule change).

[0030] 2. Multi-agent graph structure reasoning subsystem: By restricting "statistic-based generation" to the framework of "graph-based logic", it effectively suppresses model illusion and ensures that the results of complex decisions such as process planning and production scheduling meet physical consistency and engineering interpretability.

[0031] Example: Please refer to Figures 2 to 3 This embodiment provides an enhanced reasoning framework for large language model (LLM) systems designed for complex industrial decision-making tasks. By decoupling knowledge management from logical reasoning and utilizing graph thinking chains (GoT) and physical constraint injection, this framework addresses issues such as illusion, lack of physical common sense, and untraceable reasoning inherent in traditional large industrial models.

[0032] 1. Generative Automated Knowledge Construction Subsystem: This subsystem consists of a core layer, comprising pre-built industrial ontology library, document loading and processing, unstructured document conversion, entity extraction agent, entity alignment algorithm, and other core modules.

[0033] 2. Knowledge Base: This knowledge base consists of a graph database and a vector database. The graph database stores the structured network of entity relationships, supporting logical reasoning. The vector database stores the semantic vectors of knowledge fragments, supporting fuzzy semantic retrieval.

[0034] 3. Multi-agent graph structure reasoning subsystem: This subsystem is the core brain of decision-making, and mainly includes the following collaborative agents and modules: Task planning agent: As the system's "scheduler," this is a finely tuned LLM with "tool invocation" and "industrial task planning" capabilities. Upon receiving user instructions, it generates a JSON list containing task dependencies. It then breaks this down into multiple executable sub-task sequences (DAG, Directed Acyclic Graph).

[0035] For example, the task of "diagnosing equipment A's fault and providing a solution" can be broken down into: 1. Retrieving the status of equipment A; 2. Matching the fault mode; 3. Consulting the maintenance manual; 4. Generating a solution.

[0036] This agent is equipped with a deadlock prevention check function: it detects the existence of loops. If a loop is found (A depends on B, and B depends on A), the system reports an error and requires the agent to replan.

[0037] AgenticGraphRAG (Graph Enhanced Retrieval) module: Locate graph nodes based on keywords of the current problem and expand outward (1-hop, 2-hop) to find implicit correlation information (such as common sources of failure). Transform the retrieved subgraph structure into natural language context and feed it into LLM.

[0038] Graph Thinking Chain (GoT) Manager: Breaking the linear thought chain of LLM (CoT), the reasoning process is modeled as a tree or network structure: supporting divergence (proposing multiple hypothetical causes or based on known task strategies), convergence (integrating multiple pieces of evidence to reach a conclusion), and backtracking (returning to the previous node to re-examine the reasoning when a path is negated). This greatly improves the success rate of diagnosing complex faults.

[0039] Each branch is scored and constrained by an independent "evaluation agent". If all current branches are pruned (all scores are low), the system reverts to the parent node, prompting LLM: "The previous assumptions are all invalid. Please try a completely new approach", and regenerates the branches.

[0040] This method predefines graphs for four key industrial tasks: process planning, DFM optimization, work instructions, and quality inspection planning.

[0041] Constraint Injection and Verification Module: When LLM generates decision parameters (such as "recommended pressure setting is 50MPa"), the numerical parameters are extracted from the text scheme generated by LLM using regular expressions and then queried in the "Physical Rules and Parameters Library" for verification. If the verification fails or violates physical constraints, the decision is immediately rejected and LLM is required to regenerate the scheme, thus eliminating the safety risks caused by "illusions".

[0042] Interpretable output module: The final output not only includes suggestions, but also displays the reasoning path diagram (which graph nodes were passed through), the source of the reference (which manual and which page it is based on), and the confidence assessment, meeting the stringent traceability requirements of industrial scenarios.

Claims

1. A large language model reasoning enhancement framework system for complex industrial decision-making tasks, characterized in that: It includes a generative automated knowledge construction subsystem that works collaboratively and a multi-agent graph structure reasoning subsystem. By decoupling knowledge management and logical reasoning, it realizes a paradigm shift from probabilistic prediction to logical deduction, which is used to solve probabilistic illusions, lack of physical constraints and multi-objective optimization conflicts in complex industrial decision-making tasks.

2. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, The generative automated knowledge construction subsystem includes a pre-built industrial ontology library, a document loading and processing module, an unstructured document conversion module, an entity extraction agent, and an entity alignment algorithm module, which can realize the automated closed-loop extraction and storage of industrial knowledge from unstructured multi-source data to structured graphs.

3. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, It also includes a knowledge base consisting of a graph database and a vector database. The graph database is used to store structured entity relationship networks to support logical reasoning, and the vector database is used to store semantic vectors of knowledge fragments to support fuzzy semantic retrieval.

4. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, The multi-agent graph structure reasoning subsystem includes a task planning agent, which is a fine-tuned large language model with tool invocation and industrial task planning capabilities. It can receive user instructions and generate a JSON list containing task dependencies, decompose it into an executable sub-task sequence, and is equipped with a deadlock prevention function. It can detect loops in task dependencies and prompt for replanning.

5. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, The multi-agent graph structure reasoning subsystem also includes a graph enhancement retrieval module. This module can locate graph nodes based on the keywords of the current question and expand outward from 1-hop to 2-hop to mine implicit association information between entities and transform the retrieved subgraph structure into natural language context for the large language model.

6. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, The multi-agent graph structure reasoning subsystem includes a graph thinking chain manager. The manager can model the reasoning process as a tree or network structure, supporting the divergence, aggregation, and backtracking of the reasoning process. Each reasoning branch is scored and constrained by an independent evaluation agent. When all current branches are pruned, the system reverts to the parent node to prompt the large language model to regenerate branches. The manager predefines graphs for four key industrial tasks: process planning, DFM optimization, work instructions, and quality inspection planning.

7. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, The multi-agent graph structure reasoning subsystem includes a constraint injection and verification module. This module extracts numerical parameters from the text scheme generated by the large language model through regular expressions, queries the physical rules and parameter library for verification, and if the verification fails or violates the physical constraints, it is rejected and the large language model is required to regenerate.

8. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 1, characterized in that, The multi-agent graph structure reasoning subsystem includes an interpretability output module. The output of this module not only includes decision suggestions, but also displays the reasoning path graph, reference sources, and confidence assessment, meeting the traceability requirements of industrial scenarios.

9. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to claim 2, characterized in that, The unstructured multi-source data includes industrial documents, production logs, and dialogue data. The generative automated knowledge construction subsystem can achieve low-cost, near real-time updates of industrial knowledge, avoiding the high computing power cost of repeated retraining.

10. The large language model reasoning enhancement framework system for complex industrial decision-making tasks according to any one of claims 1-9, characterized in that, The complex industrial decision-making tasks include manufacturing process planning, part design optimization, production scheduling, and complex fault diagnosis. The system can meet the requirements of physical consistency, engineering interpretability, and high reliability in industrial settings.

Images