Startups share many characteristics with batch processes [2] but exhibit even greater temporal and spatial variability. Consequently, conventional batch monitoring methods, such as multi-way principal component analysis (MPCA), often fall short in monitoring of startups, necessitating more advanced approaches.
Recent studies [3], [4], [5] suggest that autoencoders (AEs) can effectively model nonlinear and dynamic correlations and extract low-dimensional representations [6] that capture intricate process dynamics. In this work, we propose a comprehensive monitoring strategy that uses variable-wise unfolding to preserve intrinsic temporal structures, followed by AE‑based dimensionality reduction. By embedding process knowledge into the model, our approach enhances interpretability for operators and supports intuitive fault diagnosis.
We investigate how activation function choice (from linear to advanced variants) affects both fault detection and diagnostic performance. To facilitate diagnosis, our methodology involves investigating simpler AE architectures to enhance interpretability and pinpoint fault origins, as well as conducting sensitivity analyses on more complex architectures. Comparative assessments against MPCA-based strategies demonstrate that tailoring AE architecture and activation functions to address startup-specific challenges can improve fault detection and diagnosis.
These findings underscore the promise of variable-wise AE strategies, with proper activation function selection, for robust, interpretable, and real-time monitoring and diagnosis during the startup phase of chemical processes.
[1] B. Hassani, Y. Wang, A. Kumar, J. Flores-Cerrillo, J. Wang, and Q. P. He, “Comparison of Machine Learning Techniques for the Modeling and Monitoring of Cryogenic Air Separation Unit Startups,” in 2024 AIChE Annual Meeting, AIChE, 2024.
[2] T. Kourti, “Multivariate dynamic data modeling for analysis and statistical process control of batch processes, start-ups and grade transitions,” in Journal of Chemometrics, John Wiley and Sons Ltd, Jan. 2003, pp. 93–109. doi: 10.1002/cem.778.
[3] J. Qian, Z. Song, Y. Yao, Z. Zhu, and X. Zhang, “A review on autoencoder based representation learning for fault detection and diagnosis in industrial processes,” Chemometrics and Intelligent Laboratory Systems, vol. 231, p. 104711, 2022.
[4] J. Ren and D. Ni, “A batch-wise LSTM-encoder decoder network for batch process monitoring,” Chemical Engineering Research and Design, vol. 164, pp. 102–112, 2020.
[5] P. Agarwal, M. Aghaee, M. Tamer, and H. Budman, “A novel unsupervised approach for batch process monitoring using deep learning,” Comput Chem Eng, vol. 159, p. 107694, 2022.
[6] M. Sakurada and T. Yairi, “Anomaly detection using autoencoders with nonlinear dimensionality reduction,” in Proceedings of the MLSDA 2014 2nd workshop on machine learning for sensory data analysis, 2014, pp. 4–11.