Anomaly detection and prediction in discrete manufacturing based on cooperative LSTM networks
B. Lindemann, Nasser Jazdi, Michael Weyrich
Abstract
Manufacturing processes are characterized by their temporal and spatial distributed nonlinear physics. Analytical models are not available and numerical models do not incorporate abnormal process effects that are not known to the engineer. These unknown anomalies cause reduced process stability and fluctuant product quality. To tackle the problem, numerous approaches for anomaly detection based on neural networks have been developed over the years. Long short-term memory (LSTM) networks have also been investigated intensively for prediction purposes. Current approaches lack in the capability of constructing prediction models for both process and anomaly behavior. Furthermore, they do not deliver a solution for short-term as well as long-term anomalies. Hence, the current paper presents a novel detection and prediction procedure based on a LSTM architecture to cooperatively predict process outputs and anomalies by using two separate but interacting models. The anomaly detector is realized as stacked LSTM auto-encoder and the cooperative prediction models are based on sequence-to-sequence networks with gated recurrent units for short-term and LSTM for long-term effects. The approach is evaluated within a real industrial environment by means of a production plant for hot forging at a German automotive supplier for metal components.