Learning behavioral models by recurrent neural networks with discrete latent representations with application to a flexible industrial conveyor

Abstract Recurrent neural networks (RNN) are being extensively exploited in industry to address complex predictive tasks by leveraging on the increased availability of data from processes. However, the rationale behind model response is encoded in an implicit way, which is difficult to be explained by practitioners. If revealed, such mechanisms could provide deeper insights into RNN execution, enhancing conventional performance evaluations. We propose a new approach based on the introduction of a model-based clustering layer, constraining the network to operate on a discrete latent state representation. By processing context-input conditioned transitions between clusters, a Moore Machine characterizing the RNN computations is extracted. The proposed approach is demonstrated on both synthetic experiments from an open benchmark problem and via the application to a pilot industrial plant, by the behavior cloning of the flexible conveyor of a Remanufacturing process. The finite-state RNN attains the prediction accuracy of RNN with continuous state, providing in addition a more interpretable structure.