A COMPARATIVE STUDY OF AUTOCORRELATED MULTIVARIATE PROCESS MONITORING

抄録

Multivariate process monitoring using sensory data has become increasingly critical for quality control of industrial processes and equipment. The high sampling frequency of sensors generally results in strong autocorrelation within the data. However, extant research on multivariate process monitoring is typically based on the independence assumption. In this study, we present a framework for autocorrelated multivariate process monitoring and study various monitoring algorithms. Specifically, the conventional vector autoregressive (VAR) model, and emerging deep learning models such as the long short-term memory (LSTM) model and the LSTM-based autoencoder model are investigated. The monitoring performance of the different models are compared under different shift patterns based on the metrics of average run length and average computational time. The simulation results demonstrate that the residual-space monitoring scheme of the LSTM autoencoder model is effective in achieving the smallest average run length for mean shifts, while the VAR model and the LSTM model are better for monitoring correlation changes. The average computation time of the LSTM autoencoder-based model is relatively greater than other models, which needs to be further explored to accommodate real-time monitoring scenarios.