Failure prediction on an anonymized system of a ship


Our customer monitors a critical equipment on his ship with 64 sensors. Based on acquired data, he used to apply condition-based monitoring with thresholds specified by domain experts. Time to find how to process the data and to specify the decision rules takes too much time, it does not find weak signals leading to failures, and this process is not scalable to other equipment.


Our customer wanted to get a failure prediction model, produced without any knowledge of the underlying physical system, any knowledge of the nature of sensors, and could identify which sensors caused the failure.


From anonymised healthy data acquired during one year, DiagFit was used to create in a couple of hours a single model which captured correlations between sensors data and produced a state of health per sensor.


The predictive model produced by DiagFit gave much better results than the system developed by our customer, it only took few hours to create it instead of many months, and the fact that is has been created without any knowledge of the underlying physical system means that this process can be reproduced to other equipment rapidly.

The illustration above showcases 3 curves. 

From top to bottom, the first curve is a snapshot of data acquired by one sensor.
The second curve is the failure ground truth meaning that when the curve fell to -1, the system was in an abnormal state. When it returned to 1, the system failed and a maintenance operation fixed the problem to return to a healthy state
The last curve is our health indicator computed from our model. It demonstrates the capacity to predict the failure almost one day in advance.