GenAI-Augmented Diagnostic Reasoning for Diesel Engine Fault Triage: A Large Language Model Framework for Technician DecisionSupport at Scale
DOI:
https://doi.org/10.47363/JMSMR/2025(6)226Keywords:
GenAI, Diagnostic Reasoning, Fault Triage, Diesel Engine, LLM Framework, Decision Support, Data Integration, Evaluation, Large Language Models (LLMs) for Diesel Diagnostics, GenAI-Augmented Fault Triage Systems, AIPowered Technician Decision Support, Predictive Maintenance for Diesel Engines, Intelligent Fault Reasoning Frameworks, Scalable Diagnostic Assistance for Heavy Machinery, Explainable AI for Engine Troubleshooting, Automated Diesel Engine Failure Analysis, Industrial Generative AI for Maintenance OperationsAbstract
Diesel engine maintenance is critical for on-road safety and operational efficiency, yet technician capability is a current bottleneck. Working with a major manufacturer, data on Diesel Engine fault triage are collected and curated to address a key maintenance issue: Given observed symptoms, what faults are likely candidates? Strong performance in Diagnostics is therefore a logical path, but development in this area is complicated by four factors: heterogeneity and incompleteness of the data; the need for results in diagnostic timeframes; the requirement to minimize risk to life and limb; and the need for results that can be understood and trusted by staff. A Generalized-Gen AI approach is adopted to Evidence-Based Decision Support across the full breadth of Diesel Engine Maintenance. The diagnostic triage task is just one of the areas where Gen AI plays a role; other roles range from ingestion of multiple data sources to sensor data interpretation (e.g., audio analysis).
A large language model architecture, Google’s T5, is used for the first time to provide Evidence-Based Triage Support for Diesel Engine Faults. A Gen AI approach is essential to support the effective combination of disparate datasets into a reliable triage support system. Results on a held-out test set are highly promising, and evaluations in the required real-world setting (effectively an audit of triage decisions) show diagnostic completions that are correct when present in the recorded training data. The combination of strong diagnostic performance across the test set and a high-confidence low-latency real-world validation indicates that Engine Fault Triage can be assisted by this system. Further confidence in the overall approach arises from earlier, well-founded GenAI Applications across virtually all Diesel Maintenance Areas. The triage task’s solid performance, therefore, lends further indirect support to these other tasks.
