A Recurrent Neural Network Architecture for Failure Prediction in Deep Drawing Sensory Time Series Data

Abstract

In the course of the fourth industrial revolution and the coherent phenomenon of industrial big data, production lines and, more specifically, single production processes are complemented by sensor systems to facilitate the acquisition of sensory data on the shop floor level, providing a number of possibilities for data driven optimization applications. One possible application is to support manual quality control at the end of a production line by an automated quality control system based on data driven predictive analysis of the acquired shop floor data. In this paper, we demonstrate the application of such a predictive quality system for a deep drawing manufacturing process of car body parts. As a first step, we formalized the experience and expertise of the domain experts to enable automated a-posteriori quality assessment of the deep drawing process data. As a second step, we trained a combination of two long short-term memory neural networks with the auto-labeled data to predict the occurrence of process failures, i.e. cracks within the manufactured car body part, enabling automated a-priori quality assessment of the deep drawing process. We show that our model correctly predicts the occurrence of more than 94% of the process failures, facilitating the possibility to react to failures pre-emptively before they actually occur. We argue that our approach can be transferred to many other production processes whose quality control can be supported by the acquisition of time series data from the involved production machines.