Structurally-Aware Predictive Maintenance for Autonomous Bot Fleets
A predictive-maintenance system for fleets of autonomous bots navigating structural rails, built to detect track defects before they cause physical grid failures — by mapping physical reality into the data model instead of relying on naive global thresholds.
Patented methodology: WO2024083748A1 →The challenge
Predictive maintenance in Industrial IoT usually fails because algorithms are fed raw sensor thresholds with no physical context. The physical world is not uniform: a rail section bolted next to a heavy structural pillar vibrates completely differently from a rail suspended in the middle of a long span. Applying a single global G-force threshold across such a dynamic environment produces an avalanche of false positives — and an alerting system nobody can trust.
Our approach
Structurally-aware logic
We abandoned global thresholds entirely and categorized every track section by its specific physical properties and location within the grid, so failure detection reasons about context rather than a single number.
Dynamic baselines
For each structural category we established a distinct reference vibration baseline, so "normal" is defined per context — a pillar-mounted rail and a mid-span rail each have their own signature.
The telemetry pipeline
Vast streams of bot telemetry were processed in Google BigQuery, with the data workflows orchestrated end to end by Apache Airflow.
Distributed inference
Predictive models were trained with XGBoost and TensorFlow, containerized with Docker, and deployed as scalable Python microservices on Kubernetes.
The outcome
At inference time, the engine compares a bot’s live telemetry against the reference signature of that exact structural category — not a naive global average. The architecture proved so effective at isolating genuine mechanical defects from normal structural resonance that the core failure-detection methodology was patented.
Technology
IIoT machine learning isn’t just about processing telemetry data — it’s about deeply mapping physical reality into your data architecture.