(588a) Optimal Hierarchical Stationarity Feature Extraction for Monitoring Start-up Intermittent Manufacturing

Intermittent manufacturing processes play an important role in providing low-volume and high-value-added products [1-3]. Unknown process disturbances, misoperations, and failures in sensing instruments/actuators will affect the operation performance of the processes and may even deteriorate the operating safety. As an effective solution to this type of issue, process monitoring has been widely applied in intermittent manufacturing industries. The development of an effective process monitoring scheme requires a high-fidelity model. However, it has been challenging and expensive to develop accurate first-principles models as the basis of process monitoring of intermittent manufacturing processes due to their complex structures, high nonlinearity, and non-stationarity [4]. As data resources become increasingly rich, and new sensing and data storage technologies are made available, data-driven process monitoring methods have attracted great research attention [5,6].

In the existing literature, there have been some results on data-driven monitoring of intermittent manufacturing processes. For example, multi-way principal component analysis (MPCA) was proposed [5,6]. Multi-phase modeling and monitoring methods were proposed in [7-9]. However, in addition to the process operations at steady-state levels, appropriate start-up, which involves starting the machine and warming up the machine material is also critical. Since process variations in the start-up phase are non-stationary in both time-wise and batch-wise directions, the above methods can be hardly applied during the start-up phase. While there have been some approaches on modeling and monitoring of the start-up phase of process operation [10-12], they indeed suffer from a few critical limitations: 1) there have been few methods that can address identification and fault detection in one framework; 2) online monitoring is yet to be enabled; 3) the application of the existing method is challenging when there is a relatively small number of batches for modeling.

In this work, we propose a two-layer hierarchical stationarity feature extraction framework to address the identification and monitoring of intermittent manufacturing processes simultaneously. The first layer focuses on obtaining a consistently stationary subspace for each start-up batch. The projection directions are optimized via finding latent variables using Kullback-Leibler divergence. This way, process variations in each batch are decomposed into two parts. The first part contains consistently stationary information that is almost the same as overall start-up batches, and the other part represents non-stationary process variations in each batch. The second layer of the hierarchical structure of the proposed framework finds a long-term equilibrium trend by treating all start-up batches using cointegration analysis. With consistent stationary space, the abnormal process can be timely detected at each time interval with the explicit projection directions. For the long-term equilibrium trend, this part of information contributes to distinguishing the start-up stage from the steady production stage. By integrating the above stationary information, identification and monitoring of the batch start-up stage can be achieved every time the process is restarted. The contribution of the proposed method is summarized below:

1. Start-up stage simultaneous identification and online monitoring is achieved for the first time.

2. Instantaneous modeling is realized for each restart to avoid potential model mismatch;

3. Two types of stationary process variations are treated and decomposed with considerations of both inter-batch and intra-batch directions.

References

[1] Z. Zhou, C. Wen, C. Yang. Fault detection using random projections and k-Nearest neighbor rule for semiconductor manufacturing processes. IEEE Transactions on Semiconductor Manufacturing, 28(1):70–79, 2015.

[2] G. Wang, S. Yin, O. Kaynak. An LWPR-Based data-driven fault detection approach for nonlinear process monitoring. IEEE Transactions on Industrial Informatics, 10(4): 2016-2023, 2014.

[3] K. Peng, K. Zhang, B. You, J. Dong, Z. Wang. A quality-based nonlinear fault diagnosis framework focusing on industrial multimode batch processes. IEEE Transactions on Industrial Electronics, 63(4):2615–2624, 2016.

[4] M. Kheradmandi, P. Mhaskar. Adaptive model predictive batch process monitoring and control. Industrial & Engineering Chemistry Research, 57(43):14628–14636, 2018.

[5] P. Nomikos, J. F. MacGregor. Monitoring batch processes using multi-way principal component analysis. AIChE Journal, 40:1361–1375, 1994.

[6] P. Nomikos, J. F. MacGregor. Multivariate SPC charts for monitoring batch process. Technometrics, 37:41–59, 1995.

[7] C. Undey, A. Cinar. Statistical monitoring of multistage, multiphase batch processes. CIEEE Control Systems Magazine, 22:40–52, 2002.

[8] N. Lu, F. Gao, F. Wang. Sub-PCA modeling and on-line monitoring strategy for batch processes AIChE Journal, 50(1):255–259, 2004.

[9] C. Zhao, F. Wang, N. Lu, M. Jia. Stage-based soft-transition multiple PCA modeling and on-line monitoring strategy for batch processes. Journal of Process Control, 17(9):728–741, 2007.

[10] B. L. Huang, Y. Yao. Automatic steady state identification for batch processes by nonparametric signal decomposition and statistical hypothesis test. Chemometrics and Intelligent Laboratory Systems, 138:84–96, 2014.

[11] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-C. Yen, C. C. Tung, H. H. Liu. The empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis. Proceedings of the Royal Society of London, 454:903–995,1998.

[12] Z. Yan, B.-L. Huang, Y. Yao. Multivariate statistical process monitoring of batch-to-batch startups. AIChE Journal, 61(11):3719-3727, 2015.