التحدي
A national energy operator managing gas compression assets across multiple sites experienced frequent unplanned shutdowns, each outage costing six figures in lost production. Despite extensive PI historian instrumentation capturing pressure, temperature, vibration and other metrics, engineers relied on manual dashboard inspection and spreadsheet tracking, unable to detect the slow degradation patterns that preceded failures. Previous attempts with generic anomaly-detection software produced excessive false positives and weren't tuned to the operator's specific compressor operating conditions.
ما الذي قمنا به
Built an automated predictive-maintenance pipeline on Databricks with three layers: Delta Live Tables standardising PI system streams across sites into a Bronze, Silver, Gold data flow; a rules engine encoding the operator's own engineering knowledge across operational stages (enclosure, pre-lube, yard valve, ignition, on-load); and an ML layer of MLflow-tracked XGBoost classifiers and autoencoders detecting subtle degradation patterns. LangGraph orchestration unified both layers, generating automated alerts and plain-language daily performance summaries explaining the reasoning behind each prediction.
النتيجة
Detected faults up to seven days in advance across more than 10,000 daily sensor readings per asset, letting engineers move from reactive emergency response to proactive scheduling. Reduced unplanned downtime and delivered real-time anomaly dashboards for ongoing operational awareness.
Turned six-figure unplanned outages into seven-day advance warnings, with plain-language daily summaries explaining every prediction.
Delta Live Tables normalised multi-site PI tag names and units into a governed Bronze, Silver, Gold flow. MLflow-tracked XGBoost classifiers and autoencoders scored degradation risk against engineering rules for each operational stage, with LangGraph orchestration turning combined rule and model outputs into alerts and daily narrative summaries rather than raw scores.