Monitoring autocorrelated processes using the latent vector in LSTM autoencoder

This paper proposes a novel process monitoring procedure using latent vectors from a long short-term memory (LSTM) autoencoder to effectively detect anomalies in autocorrelated processes.While the traditional reconstruction error-based monitoring procedure is effective for detecting abnormal states in process data, it has limitations in capturing subtle or gradual variations.To address this issue, we introduce a latent vector-based procedure and analyze how the dimensionality of latent vectors affects detection performance using AR(1), AR(2), and ARMA(1,1) models.The two procedures are evaluated using average run length (ARL) as a performance metric.The results indicate that the latent vector-based procedure is more effective in detecting small shifts, such as process level changes, with lower-dimensional latent vectors enabling faster detection.In contrast, the reconstruction error-based procedure performs better in identifying large shifts, such as process variance changes, with high-dimensional latent vectors capturing complex patterns more precisely.These findings suggest that selecting an appropriate monitoring procedure based on the type of process anomaly is crucial, with the latent vector-based procedures being more suitable for detecting subtle changes and the reconstruction error-based procedures being more effective for identifying large variations.