How to Use Anomaly Detection in Predictive Maintenance

Posted by yashi ganguly on December 22nd, 2017

Most of the data science so far had been achieved working in comfortable armchairs; a goal is defined, a target class is selected, a model is trained to recognize/predict the target. This has all changed for another vertical called anomaly detection that has wide implications for the predictive maintenance of systems in industrial units. Big data anomaly detection challenges to shift to never-seen-before productive data.

Anomaly detection in a system is often misunderstood in the data parlance. It typically not a programming error or a bug but an event that is not part of the system, in other words, an event that has never taken place in the history of the system. It, however, is seen in a context.  In the case of network data, an anomaly is an intrusion, in healthcare system a sudden pathological status, in online sales a fraudulent payment, and, last but not the least in manufacturing units a mechanical part breakdown. This is indeed relatively a new paradigm for Spark but has become very popular in recent times. In fact, Spark anomaly detection is vertical which is attracting new talent and finding new vistas of operations. 

In the manufacturing sector, it has combined with predictive maintenance which typically is about finding the fault before they occur.  Here the goal is to keep mechanical parts working as long as possible as mechanical pieces are expensive and their breakdown means loss of production and thus revenue. If it can be predicted it means a huge relief for the businesses as by predicting the breaking point before it occurs saves a breakdown which often triggers a series of damages in the other components. Therefore, a high value is usually associated with the early discovery, warning, prediction, and prevention of anomalies. Specifically, the prediction of “unknown” disruptive events in the field of mechanical maintenance takes the name of “anomaly detection.”

The real challenge here is to predict something that one has never seen, an event that is not in the history of the system.  This requires a shift in the analytics perspective! If data describing normal functioning is what we have, then normal functioning we will predict!

For this understanding and analyzing the heat maps in industrial units can be very critical. They are termed Anomaly-heat map which has wide implications for the proper functioning of the system. It is but common knowledge that when a system overheats, it is bound to malfunction. To put it in matrix and algorithm is what anomaly detection data scientists are trained for. The sensor data thus becomes crucial in anomaly detection.