A novel methodology based on long short-term memory stacked autoencoders for unsupervised detection of abnormal working conditions in semiconductor manufacturing systems

The detection of the occurrence of working conditions in semiconductor manufacturing systems is fundamental to minimize yield losses and enhance production quality. The main challenges are the lack of data labeled with the machine state (normal/abnormal) and the nonlinear time evolution of the monitored signals. This work develops a novel methodology for the detection of abnormal conditions that, differently from the existing approaches, do not require the availability of labeled data. It consists of: (a) an approach based on k -fold cross-validation for automatically building a training set which does not contain data collected during abnormal conditions; (b) a signal reconstruction model based on stacked autoencoders with Long Short-Term Memory (LSTM) cells for reproducing the machine expected behavior in normal conditions; and (c) a decision module based on an abnormality indicator computed using the Mahalanobis distance. The proposed methodology is shown to outperform other state-of-the-art approaches in two case studies based on data taken from plasma etching machines.