Extracting finite state representations from recurrent models of Industrial Cyber Physical Systems

Neural networks are being broadly explored for theidentification of Industrial Cyber Physical Systems (ICPS) modelsfrom data sequences. However, learned representations typicallylack explainability, representing nowadays a major challenge ofdeep learning. Interpreting the information structured across thesynaptic links is particularly challenging for recurrent neuralnetworks (RNN), encoding input features and observed systemdynamics within a continuous latent space. In this work, weinvestigate the representation built within the RNN while learningbehavioral models of a class of discrete dynamical systems.To this end, we propose a method to extract the symbolicknowledge structured by the continuous state, based on GaussianMixture Model clustering. Experiments are performed on a pilotremanufacturing plant, by learning the model of a conveyorcontroller from process data. We show the capability of the RNNto achieve accurate predictions while providing a Moore Machinerepresentation of the latent activations, consistent with the targetsystem.