The Challenge
A major global refining company relied on experienced experts to diagnose equipment anomalies detected by their predictive analytics program. But with the number of experienced employees shrinking, response time for these investigations was going up and failures occurring in that dead time between detection, diagnosis, and decision, were becoming more common. When a centrifugal pump at a refining complex began showing elevated vibration in the 2nd stage bearing, all eyes were on the bearing and its lube oil system. But the root cause was going unnoticed.
Multivariate Detection + Full-System Context = Diagnosing Causes, Not Symptoms
UptimeAI Root Cause Agent detected the abnormal vibration when a multivariate model including roughly 30 tags temperatures, pressures, flows, lube oil conditions started to deviate from live vibration values. This was where softwares contribution to alert investigation used to end for the refinery.
Root Cause Agent leveraged context from integration with their CMMS and SharePoint to take an expert like approach to diagnosing the root cause of the sudden increase in vibration. Looking at all available data sources and built in FMEAs, the agent determined the vibration shift was most likely tied to some recent seal work when the unit was returned to service on a temporary cold alignment. Thermal expansion upon return to operation was accelerating bearing wear at a higher than expected rate.
Root Cause Agent issued recommendations to avoid an unplanned bearing failure based on the trajectory of the degradation. By performing a laser alignment with the hot targets, the refinery avoided having to correct a much more serious bearing issue, saving over $500K in maintenance expense and associated unit downtime.
A Repeatable, Confidence-Ranked Root Cause Hypothesis in Minutes
Root Cause Agent assembled a causal chain automatically when the vibration issue was detected. There was no sending the data off to experts to add to their queue to investigate. Instead, the experts were presented with two completely traceable hypotheses. The top carried 90% confidence: thermal growth misalignment following a seal change during the recent turnaround. The causal chain provided full evidence and links to source documentation:
- n SAP work order flagged that the coupling had been broken apart and aligned only while offline — thermal growth after restart drove the misalignment.
- A SharePoint search across tens of thousands of unstructured documents surfaced a prior RCA from a sister pump with identical symptoms, plus an OEM troubleshooting guide on pump-to-driver misalignment.
- The team further refined the hypotheses when they submitted a lube oil lab analysis via “Rooty” AI copilot, which was added as additional evidence that further supported the misalignment hypothesis. The rising iron content across three reports increased diagnostic the maintenance and operations teams confidence further
Results That Scale
After Root Cause Agent successfully diagnosed this misalignment issue, the refiner began leveraging the agent for all their predictive alerts. They saw alert approval rates grew by nearly 20 percentage points, finishing above 70% — against a benchmark in the single digits for their legacy predictive analytics software. The alert approval rating boost indicated a significant increase in alert quality. After many years of generating alert volumes so high only ~5% of alerts could be investigated,
Root Cause Agent was also dramatically lowering the overall number of alerts the team received. Across 15 compressor trains, the platform averaged roughly one alert per asset per month. This high-accuracy, low-volume, completely pre-diagnosed alerting made site-level self-management viable, and kept the people closest to the equipment engaged.
