Actuator fault diagnosis in a heat exchanger based on classifiers - A comparative study

Abstract

Five different Fault Detection and Isolation (FDI) approaches are compared using the same experimental system: an industrial shell and tube Heat Exchanger (HE). The FDI approaches are classified into two major groups: the Artificial Neural Networks (ANN), the Naive Bayes and the k-Nearest Neighbor (k-NN) classifier are process historybased methods, while the Fuzzy Logic (FL) system and the Fault Decision Tree (FDT) are considered qualitative model-based methods. The Receiver Operating Characteristic curve and the confusion matrix have been used to compare the detection and classification performance when actuators fail. Experimental results show that the k-NN method reached the lowest total error of fault classification (12.4%) using cross-validation while the FDT method obtained several misclassifications (46.2% of error). In the fault detection stage, k-NN presented the best performance assuming a high probability of correct detection (90 %) with the lowest possible probability of false alarms (13 %); however, the ANN method showed the highest probability of correct detection (93 %) but a poor result in the false alarm rate (31%), while the FL method obtained the minimum false alarm rate (9%). Advantages and disadvantages of each FDI approach are highlighted in a particular context for being implemented in a chemical process.