Faults Diagnosis based on Autoencoder and Classifier Techniques

To overcome assumptions made when setting the fault detection index threshold based on reconstruction methods, a classifier is employed to extract hidden patterns in residual data. The faulty data of each fault type used in training the classifier represent only the equivalent of 10% of the fault-free data used in the modeling phase. This considers the fact that, in industrial reality, faulty data are rarer than fault-free data. The proposed methodology is applied to the fault diagnosis of a galvanizing line. The results obtained from this case study show that combining the autoencoder (AE) with a weak classifier such as a Decision Tree (DT) provides a better fault detection rate and a lower false alarm rate than the 95% confidence threshold traditionally associated with AE.