Knowledge graph enhancement-based industrial large model deterministic reasoning optimization method and system

Abstract

The application discloses a kind of industrial big model certainty inference optimization method and system based on knowledge graph enhancement, belong to industrial artificial intelligence and intelligent control technical field.The method constructs the industrial field multimodal knowledge graph including equipment mechanism, process parameter constraint and fault elimination logic, and collects industrial field multimodal data;After receiving industrial task input, extract task keyword and context feature, based on the task keyword in knowledge graph, carry out path search, extract local subgraph and rule triple related to current task;The rule triple is converted into vector bias or logic mask and injected into industrial big model decoding layer to execute restricted generation inference;Then call physical engine or simulation logic to close loop check on inference result, when not meeting preset safety threshold, trigger feedback correction, to output the certainty inference result meeting industrial physical constraint and process flow requirement.The application can improve the consistency, reliability and explainability of industrial task inference result.

Description

Technical Field

[0001] This invention belongs to the field of industrial artificial intelligence and intelligent control technology, specifically relating to a deterministic reasoning optimization method and system for large industrial models based on knowledge graph enhancement. Background Technology

[0002] With the development of artificial intelligence technology, large language models, represented by pre-trained Transformer architectures, have demonstrated strong capabilities in natural language processing, code generation, and general logical reasoning. However, when these general-purpose large models are directly applied to vertical fields such as industrial production, precision manufacturing, and complex system control, they still face the problem of insufficient determinism. Industrial production environments are characterized by high safety sensitivity and low fault tolerance. Their decision-making processes need to strictly adhere to physical laws and process logic constraints. However, general-purpose large models generate outputs based on probability distributions and lack awareness of the underlying physical mechanisms. Therefore, they are prone to "illusion" phenomena when processing non-standard industrial instructions, making it difficult to meet the deterministic and reliability requirements of industrial scenarios.

[0003] Existing improvement solutions typically employ retrieval-enhanced generation techniques, using relevant fragments from unstructured documents as contextual input models to improve output relevance. However, this approach primarily relies on weak associations between text fragments, making it difficult to impose hard constraints on the structured logic within industrial knowledge. Furthermore, multimodal time-series data from industrial sites, such as vibration, pressure, and flow rates, exhibit high-frequency and nonlinear characteristics, making it difficult for traditional inference architectures to effectively map them into the semantic space of large models. Consequently, it becomes challenging to promptly correct inference results based on changes in on-site operating conditions.

[0004] Furthermore, knowledge in industrial vertical domains typically exists in the form of mechanistic models, expert rules, and structured databases, exhibiting high logical density and rapid knowledge updates. Existing fine-tuning methods are computationally expensive and prone to forgetting issues during incremental knowledge updates. In the inference phase, there is also a lack of an effective mechanism to directly apply structured rule constraints to the large model generation process. Especially in long-chain inference tasks such as cross-device collaboration and fault tracing, existing technologies often only use external knowledge as reference context input, making it difficult to combine current operating conditions to complete local knowledge path search, rule matching, and physical mechanism-based closed-loop verification. This results in insufficient interpretability of the inference path and poor stability of parameter output. Therefore, there is an urgent need to propose a method and system that can maintain the semantic understanding capability of large models while introducing knowledge graph rule constraints and combining physical mechanism feedback for deterministic inference optimization. Summary of the Invention

[0005] To address the problems of logical illusions, mechanism violations, and insufficient stability of reasoning results in existing industrial large-scale models under high-reliability scenarios, this invention provides a deterministic reasoning optimization method and system for industrial large-scale models based on knowledge graph enhancement, so as to improve the determinism, reliability, and interpretability of reasoning results for industrial tasks.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a deterministic reasoning optimization method for large industrial models based on knowledge graph enhancement, comprising the following steps: S1. Construct an industrial multimodal knowledge graph that includes equipment mechanisms, process parameter constraints, and troubleshooting logic, and collect multimodal data from industrial sites. S2. Receive industrial task input and extract task keywords and contextual features; S3. Based on the task keywords, perform path search in the knowledge graph to extract local subgraphs and rule triples related to the current task; S4. Convert the rule triples into vector biases or logical masks and inject them into the industrial large model decoding layer to constrain the prediction probability of candidate tokens. S5. Perform restricted generation inference based on the vector bias or logic mask, and input the inference result into the physics engine or simulation logic for closed-loop verification. S6. When the closed-loop verification result does not meet the preset safety threshold, trigger feedback correction and output a deterministic reasoning result that meets the requirements of industrial physical constraints and process flow.

[0007] Furthermore, in step S4, the predicted probability of candidate tokens in the decoding layer of the industrial big model is dynamically adjusted according to the confidence level of the constraints corresponding to the rule triples, so as to suppress candidate outputs that do not conform to industrial physical laws and process flow, and enhance candidate outputs that conform to logical constraints.

[0008] Furthermore, in steps S5 to S6, the closed-loop verification includes automatically verifying the output parameters corresponding to the inference result to determine whether the output parameters meet the energy conservation, stress limit, or safety production red line, and triggering feedback correction when the verification result exceeds the preset range.

[0009] This invention also provides a deterministic reasoning optimization system for large industrial models based on knowledge graph enhancement, including a knowledge graph construction module, a multimodal data acquisition module, a task keyword and context feature extraction module, a path search and rule triple extraction module, a decoding layer constraint injection module, and a loop closure verification module. The path search and rule triple extraction module is used to extract local subgraphs and rule triples related to the current task from the knowledge graph based on task keywords. The decoding layer constraint injection module is used to convert the rule triples into vector biases or logical masks and inject them into the decoding layer of the large industrial model. The loop closure verification module is used to call a physics engine or simulation logic to perform loop closure verification on the reasoning results.

[0010] The beneficial effects of this invention are as follows: by introducing structured knowledge from the industrial field into the decoding process of large models in the form of rule triples, the model's ability to constrain industrial mechanisms and process logic can be enhanced; by combining multimodal data from the industrial site for path search and rule matching, the model's ability to perceive the current working conditions can be improved; and by introducing a physics engine or simulation logic for closed-loop verification and feedback correction, the determinism, reliability, and interpretability of the output results can be improved. Attached Figure Description

[0011] Figure 1 This is a flowchart of a deterministic reasoning optimization method for large industrial models based on knowledge graph enhancement, as proposed in this invention.

[0012] Figure 2 This is a schematic diagram of a deterministic reasoning optimization system for large industrial models based on knowledge graph enhancement, as described in this invention. Figure 1 .

[0013] Figure 3 This is a schematic diagram of a deterministic reasoning optimization system for large industrial models based on knowledge graph enhancement, as described in this invention. Figure 2 .

Claims

1. A deterministic reasoning optimization method for large industrial models based on knowledge graph enhancement, characterized in that, The process includes the following steps: constructing a multimodal knowledge graph for the industrial field, which includes equipment mechanisms, process parameter constraints, and troubleshooting logic; collecting multimodal data from the industrial site; receiving industrial task input; extracting task keywords and contextual features; performing path search in the knowledge graph based on the task keywords; extracting local subgraphs and rule triples related to the current task; converting the rule triples into vector biases or logical masks through an enhanced reasoning module and injecting them into the decoding layer of the large industrial model; performing restricted generative reasoning; and calling a physics engine or simulation logic to perform closed-loop verification of the output results.

2. The deterministic reasoning optimization method for large industrial models based on knowledge graph enhancement according to claim 1, characterized in that, The process of converting rule triples into vector biases or logical masks and injecting them into the industrial big model decoding layer is based on the confidence level of the constraints corresponding to the rule triples, which in real time corrects the predicted probability of candidate tokens in the industrial big model decoding layer.

3. The deterministic reasoning optimization method for large industrial models based on knowledge graph enhancement according to claim 1, characterized in that, The closed-loop verification includes automated verification of whether the output parameters meet the requirements of energy conservation, stress limits, or safety production red lines.

4. A deterministic reasoning optimization system for large industrial models based on knowledge graph enhancement, characterized in that, The system includes a knowledge graph construction module, a multimodal data acquisition module, a task keyword and context feature extraction module, a path search and rule triple extraction module, a decoding layer constraint injection module, and a closed-loop verification module. The knowledge graph construction module constructs an industrial multimodal knowledge graph containing equipment mechanisms, process parameter constraints, and troubleshooting logic. The multimodal data acquisition module collects multimodal data from the industrial field. The path search and rule triple extraction module extracts local subgraphs and rule triples related to the current task from the knowledge graph based on task keywords. The decoding layer constraint injection module converts the rule triples into vector biases or logical masks and injects them into the decoding layer of the large industrial model. The closed-loop verification module calls a physics engine or simulation logic to perform closed-loop verification on the output results.

Images