STAD-FEBTE, a shallow and supervised framework for time series anomaly detection by automatic feature engineering, balancing, and tree-based ensembles: An industrial case study

Modern industrial systems are equipped with multi-sensor units, and building anomaly detection modules to monitor their collected data has become a vital task. Missing such abnormal patterns may cause producing faulty products, unwanted shutdowns in the production line, or even catastrophic damages. Sensor measurements of different natures with different sampling frequencies build a multivariate heterogeneous time series data. Conventional machine learning models fail to capture the temporal characteristics of such data. Deep learning models can address this thanks to their internal network architecture, yet training such models requires large datasets with adequate samples from all anomaly classes. This is not the case in real-world problems where class imbalance is a major issue. Tree-based ensembles are reported to have the dominant performance when dealing with structured tabular data. Inspired by this, we propose a supervised framework that combines an automatic feature engineering pipeline converting the time series dataset into its tabular counterpart with tree-based ensembles. The suggested method tackles class imbalance by generating synthetic anomalies using balancing techniques. Moreover, it allows handling heterogeneous multivariate data and augmenting categorical features with sensor measurements. Two real-world industrial datasets of relatively small size from robotized screwing processes are benchmarked, showing better results for the suggested framework compared to commonly used deep learning architectures.