Our customer is monitoring critical equipment on their vessel with 64 sensors. Based on the data acquired, it applied conditional monitoring with thresholds specified by experts in the field. The time to figure out how to process the data and specify the decision rules takes too long. Weak signals that lead to failures are not found, and this process is not scalable to other equipment.
Our customer wanted a failure prediction model, produced without any knowledge of the underlying physical system, no knowledge of the nature of the sensors, and could identify the sensors causing the failure.
Using anonymised health data acquired over a year, DiagFit was used to create a unique model within hours that captured correlations between sensor data and produced a health status for each sensor.
The predictive model produced by DiagFit performed much better than the system developed by our customer. It took only a few hours to create instead of several months, and the fact that it was created without any knowledge of the underlying physical system means that the process could be replicated quickly on other equipment.
The illustration below shows 3 curves
From top to bottom, the first curve is a snapshot of the data acquired by a sensor.The second curve is the ground truth of the failure, which means that when the curve dropped to -1, the system was in an abnormal state. When it returned to 1, the system failed and a maintenance operation solved the problem to return to a healthy state.
The last curve is our health indicator calculated from our model. It demonstrates the ability to predict failure almost a day in advance.