Evaluation of process planning in manufacturing by a neural network based on an energy definition of hopfield nets

Abstract

During the planning stages of new factories for the Body-In-White assembly, the processes used per production system need to be defined. Each production system uses a specific combination of processes, with each process belonging to a main process group. The combination of the processes and groups is subject to restrictions. Since the amount of possible combinations is too large to individually check for restrictions, we propose a Neural Network using an energy measurement derived from Hopfield networks. The proposed network memorizes former correct combinations and provides a recommendation score on how likely a new planned configuration is. Since processes can be paired with processes from their own group or with themselves, the Neural Network is modified to allow loops for joining vertices with themselves. This modification is achieved by adjusting the energy function of Hopfield networks to measure the activation of the combinations of clusters, meaning the edges, and not the activation of vertices during the training phase. We implemented the network for the process planning of factories of a leading European automotive manufacturer, and the results using correct, incorrect, and random process combinations indicate a strong capability of detecting anomalous process combinations.