Knowledge-Retrieval-Augmented LLM Pipeline for Maintenance Guidance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Automating maintenance activities for complex apparatuses like jet turbine engines using large language models can lead to incorrect instructions and inefficient outcomes, especially when augmented with knowledge retrieval systems.

Innovation Solution

A knowledge-retrieval augmented Large Language Model pipeline is developed to automate maintenance recommendations, involving a control circuit that receives a textual description, retrieves relevant knowledge documents, generates a language generation prompt, and outputs candidate recommendations for human review and validation, with the option to retrain the decoder using validated inputs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a large language model is augmented with knowledge retrieval to improve maintenance recommendations, then the accuracy and context-specificity of recommendations are improved, but the complexity of the system increases and may lead to incorrect instructions

Engineering Contradiction:
Improveaccuracy of maintenance recommendationsVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system is divided into distinct functional modules: a retrieval component that searches knowledge bases, a large language model that generates recommendations, and a human reviewer component that validates outputs. This segmentation allows each component to specialize in its function while reducing overall system complexity through modular design.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Human reviewers serve as intermediaries between the automated LLM system and the final maintenance recommendations. This intermediary layer validates and corrects LLM outputs, ensuring reliability while allowing the automated system to handle the bulk of recommendation generation.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If automation is increased to improve maintenance efficiency, then productivity increases, but the reliability of maintenance instructions decreases leading to incorrect recommendations

Engineering Contradiction:
Improvemaintenance efficiencyVSAvoidaccuracy of maintenance instructions
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system applies partial automation where the LLM handles the generation of maintenance recommendations to improve productivity, while human reviewers perform partial validation to ensure reliability. This balanced approach avoids full automation's reliability issues while maintaining efficiency gains.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

Human reviewers provide feedback on LLM-generated recommendations, creating a feedback loop that allows the system to learn from corrections and improve over time while maintaining high productivity through automated initial generation.

Inventive Principle:
Principle #23Feedback

3Reliability

If human review is added to validate candidate recommendations, then the reliability of recommendations is improved, but the time required for maintenance processes increases

Engineering Contradiction:
Improveaccuracy of maintenance recommendationsVSAvoidtime for recommendation validation
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The LLM performs preliminary generation of maintenance recommendations before human review, pre-processing the work to reduce the time burden on human reviewers while ensuring reliable validation occurs.

Inventive Principle:
Principle #10Preliminary action

4Measurement precision

If the task-specific decoder is retrained using human-validated recommendations, then the model's accuracy is improved, but the complexity of the training process and data management increases

Engineering Contradiction:
Improvemodel accuracyVSAvoidtraining pipeline complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system uses its own human-validated recommendations to retrain and improve itself, creating a self-improving loop. The validated data generated during normal operation is directly reused for model enhancement, eliminating the need for separate complex training datasets.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4621633A1Method and apparatus to facilitate building a large language model pipeline
Publication Date: 2025.09.24 GENERAL ELECTRIC CO
  • EP4621633A1 patent drawingFigure 1
  • EP4621633A1 patent drawingFigure 2
  • EP4621633A1 patent drawingFigure 3

AI summary

A textual description of an issue pertaining to at least part of an apparatus is received (201) as input. A data store (102) is accessed (203) and a plurality of knowledge documents (311) that corresponds to the input is retrieved. A language generation prompt (307) is then generated (204), as a function of both the input and the plurality of knowledge documents and output (205) to a task-specific decoder (308) that generates a candidate recommendation (409) to address the aforementioned issue. That candidate recommendation is output to at least one human reviewer (306) who reviews the candidate recommendation as a function of the plurality of knowledge documents and who then provides a corresponding human-validated recommendation (309, 411) to address the issue. The task-specific decoder can be retrained (206) using the human-validation recommendation coupled with the corresponding textual description input.