Increasing Explainability in Time Series Classification by Functional Decomposition

In this work, we develop a generic methodology for time series classification in order to increase explainability and trustworthiness of the predictions. We achieve this by dividing the whole time series into sub-sequences with a sliding window approach, transforming and classifying each chunk using paradigms from functional decomposition, and then aggregating the results to a final class. Visual as well as dataset-based explanations will be derived, enabling the investigation of errors. We demonstrate our concepts on a case study in the railway domain, where the train type is to be inferred based on the signal of a commonly used axle counting sensor. While outperforming LSTM- and FCN-based end-to-end models regarding the classification accuracy, our method further reveals interesting insights into causes of misclassifications.