Prototype scope
What this POC is — and is not
Explicit boundaries so the demo can be judged fairly against Factory Booster Day · Challenge #86.
In scope
- Structured data model for symptoms, components, parameters, causes and solutions
- Guided diagnostic workflow with controlled options (no open LLM answers)
- Probability engine derived from historical occurrences
- Interactive failure graph and knowledge base view
- Human-in-the-loop validation by maintenance specialists
Out of scope
- Live PLC / sensor / IoT integration
- Cybersecurity review required for OT connectivity
- Full population of the FMEA base — depends on client specialists
- Multi-plant rollout and identity federation
- Predictive ML on time-series signals (future layer)
Known limitations
- Mocked dataset — not representative of production failure distributions
- Probability formula is intentionally simple (count / total per symptom)
- No persistence: refreshing the browser resets simulated learning
- AI is scoped as an assistive future layer only, never as authoritative source
Why this shape (not a chatbot)
Maintenance decisions demand traceability. A free-form LLM assistant can hallucinate root causes, confuse similar components, and produce answers that cannot be audited by ISO-9001 / IATF-16949 processes. This prototype instead codifies expert knowledge into structured nodes, uses simple frequency-based probability to rank diagnostic paths, and keeps every step reviewable. AI is proposed as a future assistive layer — for parsing free-text work orders into structured nodes, or clustering similar occurrences — never as the source of the diagnosis itself.
Evolution path
- 1Populate with real FMEA content from client specialists
- 2Ingest historical work orders using LLM-assisted structuring (offline, validated)
- 3Attach sensor / PLC signals as automatic parameter observations
- 4Layer anomaly-detection models on top of the probability graph
- 5Roll out to additional plants with role-based access and audit trails