Data Science in Heavy Industry and the Internet of Things

The math behind predictive maintenance is referred to in some literature as Prognostics.  Leading up to a machine failure, signatures of the impending failure - for example, an increasing temperature or a dropping pressure - can sometimes be captured by sensor data. The prognostic/failure prediction model focuses on detecting these signatures as soon as possible. If a problem is caught early enough, repairs may be as minor as tightening a bolt. The longer a potential failure goes undetected, often the more expensive it is to repair.

Figure 1 is adapted from Blann and shows this rough phenomenon: a failing component's performance and condition degrades as it reaches a total failure point. Along the way it hits a couple important points in relation to predictive maintenance.  Point S is the start of a failing component. Point P is the point where it is observed in the data. And point F indicates a completely failed component.  Notably P is different from S, indicating the actual onset of a failing component may occur significantly earlier than it is actually detectable in what data is available.  The period between when a problem is detected (P) and when a component completely fails (F) is what is labeled the PF interval. It is in this period that attempted corrective actions can be taken to reduce overall costs. Different components will have different curves and thus different PF intervals. For example, a single failing bolt on some machines may not be detectable before failure. In this case, point P is on top of point F. Predictive maintenance is concerned with moving point P to as early in time as possible.

Figure 1. Adapted from Blann, the theoretical PF curve describes important points in a predictive maintenance problem. Point S indicates the start of a failing component.  Point P indicates the point at which a failing component is observed given existing data streams, and point F indicates the point of functional failure.  The condition of a component and timing around these points may vary from component to component.