Anomaly detection for quality control based on sequence-to-sequence LSTM networks

Unforeseeable process events and anomalies are drivers of increased inefficiencies in terms of fluctuating product quality. This paper presents a data-driven approach for quality optimization that is used to characterize anomalies being unknown at the time the system was designed. A network architecture based on a sequence-to-sequence network with Long Short-Term Memory (LSTM) cells is presented. Thus, it can be predicted which adaptation of the actuating variables has to be carried out in order to compensate expected anomalies. This keeps the quality result within tolerance. The approach is prototypically implemented and evaluated on the basis of two process chains of discrete manufacturing.