The Choice of AI Matters: Alternative Machine Learning Approaches for CPS Anomalies

We compare the pros and cons of two Artificial Intelligence (AI) solutions, addressing the anomaly detection and identification challenge in industrial Cyber-Physical Systems (CPS). We demonstrate how our current approach, Advanced DL, based on Convolutional Neural Networks (CNN) differs from a previous one, Classic ML. Though both workflows prove to result in highly accurate classification of anomalies, Classic ML is superior in this regard with 99.23% accuracy against 94.85%. This comes at a cost, as Classic ML requires total insight and expertise regarding the system under scrutiny and heavy amounts of feature engineering, while Advanced DL treats the data as a black box, minimising the effort. At the same time, we show that finding the best performing CNN model design is not trivial. We present a quantitative comparison of both workflows in terms of elapsed times for training, validation and preprocessing, alongside discussions on qualitative aspects. Such a comparison, involving analysis of workflows for the given use-case, is of independent interest. We find the choice of AI solution to be use-case dependent.