Explainable neural network for time series-based condition monitoring in sheet metal shearing

Abstract Research indicates the effectiveness of machine learning for condition monitoring in sheet metal shearing. However, existing studies primarily focused on model accuracy while neglecting model explainability. In consequence, potential biases and novel insights captured by the models remained concealed. This work contributes to the state of the art by exploring the intersection of deep learning and causal inference to obtain an explainable condition monitoring model. A causal representation learning framework based on the variational autoencoder architecture is adapted to derive a latent variable model of punch force signals from a fine blanking process. The latent variable model serves two purposes. First, it identifies latent factors explaining variations in observed force signals. Second, it provides a generative model that translates manipulations of latent factors into corresponding force signal changes, thereby, enabling interpretation of the factors. The latent variable model is integrated with a neural network that estimates punch wear. The importance of the latent factors with respect to the network’s wear predictions is analyzed to understand how the model arrives at its predictions. Experimental findings indicate that the latent variable model successfully discovered factors which correspond to real-world mechanisms affecting the punch force. One latent factor isolated a bias from measurement interventions, while another captured force variations which are attributed to punch wear itself. Furthermore, the approach demonstrated effectiveness in detecting biased prediction models, contributing to more reliable condition monitoring systems.