Visual saliency distribution maps for explaining time-series AI models used in continuous production of textile fibers

• Visual Saliency Distribution Maps provide aggregated results of individual saliency maps • Visual Saliency is strongly affected by model architecture • Out of Distribution Data not only affects regression outcomes but also visual saliency maps With the prevalence of AI solutions in industries and the adaptation of deep learning techniques, it is crucial to understand these often opaque or black-box models. For image-based classification tasks, visual saliency maps offer an excellent solution for explaining model behavior. In the scope of this work, we adopt visual saliency maps for time-series regression tasks. We investigate the interaction between time-series architectures, distribution shifts, and saliency techniques on real-world continuous production data and explore how domain experts and developers can benefit from the results by proposing to make use of distribution plots of saliency maps.