A local-LLM proof of concept that turns equipment-utilization data into hourly diagnostic reports and a manager review workflow.
Highlights
- Built a data pipeline that reconstructs machining cycles and identifies utilization anomalies and recurring equipment alarms.
- Integrated a local LLM to generate diagnostic summaries and hourly email reports for managers.
- Implemented a review workflow that records managerial feedback and updates the anomaly-assessment baseline.
Purpose
- Give managers actionable context alongside utilization charts through scheduled anomaly summaries and suggested checks.
- Combine deterministic cycle assessment with LLM-assisted interpretation of equipment data.
Implementation
- Built a .NET worker to read existing utilization data, reconstruct machining cycles, and record cycle-time outliers, accumulated idle time, and repeated alarms.
- Integrated ASP.NET Core, Semantic Kernel, read-only DAB/MCP data tools, and a local vLLM inference service.
- Generated structured diagnostic summaries and hourly HTML email reports.
- Added review links and confirmation pages so managerial feedback could update the assessment baseline.
- Deployed the database, worker, API, and inference services using Docker across separate application and LLM virtual machines.
Validation and Future Work
- Completed PoC validation using simulated CNC data rather than a production customer deployment.
- Positioned model output as decision support for managers, with human review of diagnostic suggestions.
- Planned retrieval of manufacturer manuals to supplement data-based interpretation with documented troubleshooting guidance; RAG was not part of the validated PoC.