Part 6: Predictive Maintenance Using an Unlabelled Dataset

Download the live script

In the previous article we went through a supervised learning approach for Predictive Maintenance. However, with a conservative maintenance schedule this approach is often impractical, due to the absence of failure data. In this case we use an unsupervised learning clustering approach on our data together with domain knowledge from the maintenance team.

In this example, using the same turbofan data, we use different visualisation methods to try and understand the condition of the engines. We start by looking at the first and last operating points recorded for each engine. If engines tend to start in a certain area of the Principal Component Analysis (PCA) but move to a different area just before maintenance is performed, this may give us some indication of what the trend towards failure looks like. In the figure below we can see that the first and last operating points form two groups, with the first points centred closer to the origin and the last points centred further away. However, there is significant overlap between them as most of the engines are still behaving normally at the time maintenance occurs.












The maintenance team know which engines needed maintenance at the time of service and which engines could have run for longer. If we look at an engine which was in need of maintenance (e.g. Engine 39) we could trace its sensor data behaviour for the duration of its run and see that the ending point of a bad engine is far out from the main cluster of functional engines.












This Live Script shows how to estimate the RUL using MATLAB when our equipment maintenance schedule is very conservative, resulting in none of the equipment experiencing failures. The Live Script will take you through:

  1. The typical Predictive Maintenance workflow for an unlabelled dataset
  2. Visualising Big Data using MATLAB
  3. Dimensionality reduction using Principal Component Analysis (PCA)

What Can I Do Next?

Follow us