Automated vehicle fault diagnosis and report generation using hybrid machine learning with multi-step RAG approach

Accurate vehicle diagnostics is challenging in the automotive industry due to complex failure patterns and fragmented information. While AI-based predictive maintenance systems can forecast failures using historical data, they often lack interpretability and overlook valuable technical documentation. This study proposes a holistic diagnostic framework that utilizes generative AI to improve fault prediction and explainability. A novel Retrieval-Augmented Generation (RAG) architecture is introduced to extract knowledge from unstructured manuals and structured fault code databases, enabling the generation of comprehensive vehicle diagnostic reports. The core innovation lies in a hybrid diagnostic pipeline that (i) generates realistic synthetic failure cases using a conditional tabular generative adversarial network (CTGAN) to mitigate class imbalance, and (ii) employs a multi-source RAG architecture to unify structured fault-code databases with unstructured technical manuals for context-aware diagnostics. CTGAN achieved a 96% correlation with real data. GPT-4o is employed for report generation, utilizing a hybrid prompting technique to enhance accuracy and coherence. Report quality was evaluated using multiple LLMs, yielding average scores of 85% for readability and contextual relevance. However, factual accuracy scores were comparatively lower, reflecting current limitations of LLMs in maintaining consistency across technical domains. This highlights an important research direction–improving domain alignment and verification mechanisms–while demonstrating that the proposed framework remains a strong step toward integrating explainable AI in automotive diagnostics and fleet management.