What is unsupervised learning? Unsupervised Learning (see wikipedia) is used when no historical data is available that involves past faulty behaviour and/or ageing progression. Therefore,
Engines test bench
During testing phase of catalyst converters, customer used a pressure measurement to decide when the catalyst converter should be replaced. However, this pressure measurement came from an additional sensor added exceptionnaly for the test phase. This sensor was not normally available, and objective was to replace it by a software sensor.
Detect/Anticipate failure from other measurements than the pressure one used, in order to see if there was redundancy among sensors. Objective was to know if it was possible to detect this failure from sensors already present in the engine.
DiagFit offered two ways to deal with this problem. First method was the creation of a virtual sensor calculated from available sensors to simulate the pressure. Second method was the creation of a normality space, derived from the nominal operation of the catalyst converter, in order to measure a deviation as one was getting close to failures.
All kinds of failures (6) with zero false positive rate have been anticipated whatever solution used.