Predictive maintenance

Predictive maintenance means repairing a machine just before it would fail, by watching sensor readings that show the first signs of wear. It sits between two older ways of maintenance. In reactive maintenance, a machine is repaired only after it breaks. In scheduled maintenance, a part is replaced after a fixed time or fixed use, whether it needs it or not.

Here is how it works in general. Sensors on the machine measure things such as vibration, temperature, sound or current. These readings are stored as time series data and sent to software. The software learns what normal readings look like, using anomaly detection or a model trained on past failures. When the readings start to drift, the software warns that a part will probably fail in a few days or weeks. The maintenance team then plans the repair at a convenient time, with the spare part ready.

The method can fail in two ways. If the software warns too often without a real fault, the team stops believing it. If it warns too late, there is no time to get the part or plan the stoppage. So a warning is useful only when it gives enough lead time. The method suits machines where a stop is costly and where wear gives early signs, such as bearings, motors, pumps and brakes. It suits poorly a part that fails suddenly with no early sign, such as a fuse. For such a part, replacing it on a schedule is cheaper.

Take, for example, a motor in a textile mill. Its vibration slowly rises over three weeks, which suggests that a bearing is wearing. The team replaces it on a Sunday instead of losing a working day to a sudden stop. Consider a railway, where sensors beside the track listen to the wheels and bearings of each passing train and flag a worn one before it damages the track; see rail condition monitoring. Say a hospital has a scanner that is too costly to leave broken. The maker watches its readings from a distance and sends a technician before the fault appears.

The approach needs sensors, connectivity and good data, usually through industrial IoT. It also creates repeat income for makers, because they are paid to watch the machine; see aftermarket services.

How is it different from anomaly detection? Anomaly detection only flags a reading that looks unusual. Predictive maintenance goes further: it estimates which part will fail and by when, and it plans the repair.