LLM Prompt Framework for Process Information Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training RL agents for control systems in the process industry is challenging due to the unique configurations and complexities of each system, and existing control solutions lack robustness in handling real-life disturbances without prior knowledge, making it difficult to identify and categorize process issues efficiently.

Innovation Solution

A computer-implemented method using a Large Language Model (LLM) with predefined prompts to extract and visualize process information, including system and component events, and generate simulation assistance for identifying and categorizing disturbances.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If control solutions are tested through faults of the corresponding system to ensure robustness, then reliability is improved, but time consumption increases

Engineering Contradiction:
Improverobustness of control systemVSAvoidtime for identifying and categorizing process issues
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the language model on extensive process knowledge and disturbance patterns before actual deployment. The model is pre-loaded with information about various process issues, their symptoms, and categorization rules, enabling it to quickly identify and categorize disturbances without requiring time-consuming manual analysis during operation. This pre-processing of knowledge base allows the system to provide robust control while maintaining rapid response times.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If prior knowledge of the specific process is used to identify and categorize issues, then measurement precision is improved, but device complexity increases

Engineering Contradiction:
Improveaccuracy in identifying process disturbancesVSAvoidcomplexity of control system
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces complex mechanical or manual systems for process monitoring and disturbance identification with a language model-based intelligent system. Instead of using traditional control systems that require manual configuration and expert knowledge input, the invention uses a language model that can naturally process and understand process descriptions, symptoms, and disturbance patterns through language-based interactions. This substitution maintains high measurement precision while reducing device complexity by eliminating the need for manual knowledge base configuration and expert intervention.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4629065A1Framework with large language model for providing process information and method
Publication Date: 2025.10.08 ABB (SCHWEIZ) AG
  • EP4629065A1 patent drawingFigure 1
  • EP4629065A1 patent drawingFigure 2
  • EP4629065A1 patent drawingFigure 3

AI summary

The invention relates to a computer implemented method for providing process information of an industrial process, the computer implemented method being executed on a framework (100) comprising a Large Language Model (102), LLM, a code structure (110) comprising at least one prompt (130, 140, 1050, 160), wherein a prompt contains a configuration as input for the LLM, and an API (104), wherein the API provides the communication at least between the code structure, the computer implemented method comprising the steps: running (502) a first prompt (130) comprising a process description, by the code structure; providing (504), by the code structure, the configuration contained in the first prompt over the API to the LLM; extracting (506), by the LLM, process information according to the configuration contained in the first prompt and providing the extracted process information via the API to the code structure; receiving (508), by the first prompt, the extracted process information; and generating (510), by the first prompt, an output containing the extracted process information.