How AI is changing fault diagnosis in mining maintenance
Maintenance documentation contains thousands of pages of diagnostic procedures. AI can surface the right procedure in seconds — but only if grounded in your actual source material.
The manual problem
A single haul truck can have thousands of pages of technical documentation across electrical, hydraulic, powertrain, and body systems. When a tech encounters a fault code at 2am on a remote site, the last thing they have time for is searching through a PDF library.
So they don't. They rely on experience, ask a colleague, or take their best guess. Sometimes that works. Often it leads to unnecessary part swaps, repeated diagnostic steps, and extended downtime.
The promise of AI — and its pitfall
Large language models can process and summarise technical documentation faster than any human. But general purpose AI models don't know your machines. They'll generate plausible sounding diagnostic steps that may have no basis in your actual maintenance procedures.
This is the critical distinction: AI for maintenance diagnostics must be grounded in verified source material. Every diagnostic step should be traceable to a specific page, section, and part number in your documentation. If it can't cite its source, it shouldn't be trusted.
How Trace works
Trace is FaultPilot's AI diagnostic engine. When your organisation uploads technical documentation, Trace processes it through a multi stage pipeline: extracting content, classifying document types, chunking with context preservation, and building a searchable vector index.
When a tech logs a fault and requests a diagnostic workflow, Trace searches across your documentation to find the relevant procedures. It then generates a structured, step by step workflow — each step citing the source document, page number, and specific measurement or specification.
The pipeline uses multiple AI models, each selected for its strengths. Classification and contextualisation use fast, efficient models. Document retrieval combines vector similarity search with reranking for precision. The final diagnostic generation uses the most capable model available — but always constrained to cite what it finds, never to invent.
Grounding over guessing
Every Trace diagnostic includes a grounding confidence score and source labels. If the documentation doesn't contain enough information to generate a confident diagnosis, Trace says so — rather than filling the gap with assumptions.
This is the difference between AI as a tool and AI as a liability. The tech in the field gets actionable, verifiable steps. The maintenance manager gets an audit trail. And the organisation gets faster resolution times grounded in procedures they already trust.
The best AI diagnostic system isn't the one that sounds the smartest — it's the one that can show you exactly where it found the answer.