Simple Heuristics as a Viable Alternative to Machine Learning-Based Anomaly Detection in Industrial IoT

Abstract

This article evaluates the efficacy of simple heuristic approaches compared to sophisticated machine learning by quantifying the accuracy and timeliness of selected multivariate anomaly detectors on industrial time series. It specifically examines the efficacy of two probabilistic detectors, a statistical detector and a deep learning anomaly detector. The presented work stems from the observation that the application of machine learning methods may be unfounded in a variety of use cases. The findings made in this study imply that there is no reason to over-engineer a solution by applying sophisticated methods without genuine grounds. The conventional autoregressive heuristic model outperforms the autoencoder by up to 7.2 percent. Furthermore, the autoencoder also underperforms in terms of execution time. Compared to the simpler approaches, its computational time complexity is up to 47 percent higher. Simple methods thus emerge as viable alternatives to sophisticated multivariate time-series anomaly detection on the evaluated application domain. Our conclusions remained valid through examining datasets originating from other domains. We infer that the performance of more elaborated methods requires verification to justify their usage.