Predictive Maintenance Dashboard using Python

Description

We operate twelve large water pumps on a continuous duty cycle. When one fails unexpectedly it costs us about nine hours of production, and it has happened three times in two years. We have vibration sensors on all twelve and eighteen months of data, including two of those three failures. That is a small number of positive examples and we want to be realistic about it: we are not expecting a model that predicts failure a week out from two labelled events. What we want is anomaly detection against each pump's own baseline, surfaced to the maintenance team with enough context to decide whether to inspect. If the honest answer after looking at the data is that eighteen months is not enough, we would rather hear that in week two than in month three.

• Assess the eighteen months of vibration data and report honestly on what it can and cannot support
• Build per-pump anomaly detection against each unit's own baseline rather than a fleet average
• Surface anomalies with the underlying signal so an engineer can judge, not just a score
• Backtest against the two labelled failures and state the lead time achieved
• Deliver a dashboard the maintenance team will actually open, not a notebook

• Data assessment report with a clear statement of limitations
• Per-pump anomaly detection model
• Maintenance dashboard with signal context
• Backtest results against the known failures