Explainable and interpretable time series forecasting and predictive maintenance

Abstract

Time-series data is ubiquitous in numerous application domains today, including safety-critical settings such as medical and industrial scenarios. However, real-world data can change its characteristics with time, which is referred to as drifts in concept. The phenomenon of concept drifts is particularly troubling for time-series forecasting tasks, where the goal is to predict future time-series values given known, past values. Traditional forecasting methods assume that these characteristics are fixed to achieve a high predicting performance. In this thesis, we will inspect methods how to adapt to changing concepts for time-series forecasting using both complex, deep learning methods and simple, interpretable models. To do that, we utilize the online model selection and online ensemble pruning frameworks for selecting between pools of different models in an online manner. Specifically, we will focus on aspects of explainability and interpretability throughout as these aspects are crucial for applying forecasting methods in practice. Additionally, we investigate a real-world scenario for predictive maintenance, one area of application where explainable and interpretable time-series analysis methods are crucial to gain insights for practitioners and technicians alike.