Early Semiconductor Anomaly Detection Based on Multivariate Time-Series Classification using multilayer Perceptron

This work is focused on the issue of semiconductor anomaly detection during the manufacturing process. It proposes an efficient multivariate time-series fault detection approach aiming to detect wafer anomalies at an early fabrication stage to reduce the yield loss. The raw data consist on eleven (11) multivariate time-series (MTS) measured for 150 seconds and collected during different levels of the fabrication process to describe the wafers status. First of all, the most useful information is extracted from each collected time-series (TS) data to handle the computational complexity of large-scale data processing. For that, three dimensionality reduction techniques, namely: (i) Independent Component Analysis (ICA), (ii) Principal Component Analysis (PCA), and (iii) Factor Analysis (FA) are used for comparison and optimization sake. The aim is to define the better technique allowing to keep only the meaningful information from each time-series. Thereafter, the extracted data is combined to build a new dataset which is used to fit and optimize a multilayer perceptron (MLP) to perform the anomaly detection. The very interesting obtained results show that the proposed approach is promising and could provide a precious decision-making support for abnormal wafer detection in the semiconductor manufacturing process. Copyright (C) 2022 The Authors.