Dynamic Principal Component Analysis (DPCA) and Artificial Neural Networks (ANN) are compared in the fault diagnosis task. Both approaches are process history based methods, which do not assume any form of model structure, and rely only on process historical data. Faults in sensors and actuators are implemented to compare the online performance of both approaches in terms of quick detection, isolability capacity and multiple faults identifiability. An industrial heat exchanger was the experimental test-bed system. Multiple faults in sensors can be isolated using an individual control chart generated by the principal components; the error of classification was 15.28% while ANN presented 4.34%. For faults in actuators, ANN showed instantaneous detection and 14.7% lower error classification. However, DPCA required a minor computational effort in the training step. © 2011 Springer-Verlag.
Tudón-Martínez, J. C., Morales-Menendez, R., Garza-Castañón, L., & Ramirez-Mendoza, R. (2011). Comparison of artificial neural networks and dynamic principal component analysis for fault diagnosis. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 6703 LNAI, pp. 10–18). https://doi.org/10.1007/978-3-642-21822-4_2