Revolutionizing heavy asset maintenance: The power of LLM-based virtual assistants

Part of a maintenance engineer’s job is to diagnose equipment issues and perform root-cause analysis to take the correct action and prevent the issue from happening again. Naturally, this work is often performed where the equipment is found, and has to be done quickly to avoid prolonged downtime.

It can both be cumbersome and take too long to sift through old maintenance logs and equipment manuals, which could help pinpoint the root cause. A “virtual maintenance assistant” based on a large language model (LLM) could make this knowledge much more accessible, both in the field and in the back office.

In this whitepaer, we give a brief overview of how such a virtual assistant could become a reality in the maintenance domain. We will start by exploring two essential tasks: Question Answering and Information Retrieval. We give concrete examples for both tasks using maintenance logs contributed by an operator of commercial tanker vessels, and share some thoughts on the practical implementation of these approaches. Lastly, we also discuss various methods to improve the model outputs beyond prompting a base LLM: Retrieval Augmented Generation (RAG) allows the model to search through external data sources for answers, while fine-tuning changes the model parameters to “teach” the model itself domain knowledge.

‍

Contents:

• Summary
• Question Answering
• Information Retrieval
• Improving the results: Augmenting base models with domain-specific data
• Retrieval Augmented Generation
• Fine-tuning
• Conclusion: Empowering maintenance engineers with “virtual maintenance assistants” is easier than expected
• Appendix