Every equipment show this year features AI-powered predictive maintenance, and the marketing is running well ahead of the plumbing. A decade into connected pump monitoring, here is the honest picture of what works.
What actually detects faults today
Current, pressure and runtime, trended over time, still catch the overwhelming majority of real failures. A motor current drifting upward is wear; pressure sagging under load is fouling; start counts rising is a dead vessel. This detection is rules-based, cheap, and works on a spreadsheet. Whether the alert is raised by a threshold or a neural network matters less than whether somebody receives it.
Where AI genuinely helps
Pattern recognition across fleets. Comparing a hundred similar installations and flagging the statistical outlier is something thresholds do poorly and machine learning does well. Vibration-signature analysis on large rotating equipment is the other credible application, where the physics is rich enough to reward the modelling.
Where it remains mostly slide
Claims of predicting failures weeks out on a single small pump with minimal instrumentation. Good predictions need good data, and a domestic-size pump with one sensor does not generate enough signal for the model to learn from. On those machines, the sensor investment matters more than the algorithm.
The hierarchy has not changed: instrument first, trend second, automate alerts third, and let the models layer on top where the fleet is big enough to teach them anything.