Explainable Artificial Intelligence for Time Series Using Attention Mechanism: Application to Wind Turbine Fault Detection

The time series classification problem is highly relevant in many important domains, such as industry, finance, and healthcare. It is a long-known problem that has been addressed by a plethora of mathematical techniques and algorithms, including deep learning models. Despite the rapid growth of deep learning techniques, the black-box nature of these models prevents users from adopting them in real-life scenarios. In this paper, we propose a classification framework for time series using wavelet transformation and transformers. Furthermore, we explain the results using a novel combination of the attention mechanism and Linear Discriminant Analysis. The framework was evaluated on three real-world wind turbine datasets, and it achieved comparable or leading F 1 scores. Moreover, the explanation framework highlights the key frequencies that differentiate between the healthy and faulty states. We show that these frequencies align with the frequencies identified using the mechanical domain knowledge. This compliance increases the trustworthiness of Artificial Intelligence in industrial applications and supports informed decision-making.