A novel fault detection approach based on multilinear sparse PCA: application onthe semiconductor manufacturing processes

Abstract

Batch processes are extremely important to researchers since they are widely used in many fields such as biochemistry, pharmacy, and semiconductors. The powerful batch detection method is critical to increase the performance of the overall equipment and to reduce the use of check wafers. Many techniques have been used in batch process monitoring. Among them, the multivariate statistical process control (MSPC) is very useful in batch process monitoring because of the large number of records data. Therefore, batch processes have certain characteristics, such as multimodal batch nonlinearity trajectories, which were challenged by these MSPCs. In this paper, a novel process monitoring methods based on multilinear sparse PCA (MSPCA) are proposed to overcome these shortcomings. MSPCA handle batch data as a matrix (second order), although most of the other multivariate statistical analysis approaches handles batch data as a vector (first order). Vectorization of batch data tends to ignore parts of the information. Furthermore, The MSPCA can extract more useful data from the batch data with less storage requirements and computational complexity compared to current multivariate statistical analysis approaches. The efficiency of the monitoring technique is implemented in the numerical example and Lam 9600 metal etcher process. The performance of MPSCA is characterized by a fault detection rate of 100%, as well as a false alarm rate higher than 83%. Simulation results show that MSPCA outperforms the traditional techniques.