| dc.description.abstract | With the advent of Industry 4.0, variable relations in modern industrial processes are increasingly complex due to their high dimensions and complex auto-correlations and cross-correlations. Multivariate statistical methods such as principal component analysis (PCA), partial least squares (PLS), slow feature analysis, and canonical correlation analysis (CCA) are widely employed to exploit variable relations for process data analytics, process monitoring, and fault diagnosis. However, it is not appropriate to directly apply these multivariate statistical methods to cross-correlated data from a dynamic process, since these methods are only able to model static components of the data. Subsequently, numerous adaptations of multivariate statistical methods have been proposed to address the intricate interdependencies inherent in dynamic process data. Among these, dynamic inner methods have emerged as the most advanced and efficient approaches in recent years. However, it is pertinent to acknowledge that these dynamic inner methods still exhibit certain limitations owing to the intricate properties of time series data.
In this research, several novel algorithms are proposed to address the aforementioned issues.
First, basis functions are adapted into dynamic inner PLS, referred to as dynamic weighted PLS (DWPLS), to deal with redundant information in neighboring samples caused by high sampling rates in industrial processes. Nowadays, however, because of the improvement of sensory technologies, redundant information exists among adjacent samples with high sampling rates. The existing dynamic algorithms ignore the redundant information, causing suboptimal modeling and monitoring schemes. In this paper, a dynamic weighted PLS (DWPLS) method is proposed to deal with the aforementioned issue. DWPLS maximizes the covariance between the quality latent scores and a weighted representation of lagged process scores. The relations of lagged process scores are learned through a weighted summation of basis functions. The monitoring scheme of DWPLS is also constructed. Case studies with two processes, the Tennessee Eastman process, and a distillation process, are conducted to show the advantages of DWPLS over existing methods in terms of regression and fault detection performance.
Further, to fully exploit quality information, a novel dynamic auto-regressive latent variable model (DALVM) is proposed to capture both auto- and cross-correlations simultaneously from high-dimensional time series data. Some dynamic supervised learning algorithms are designed to extract their dynamic cross-correlations, but auto-correlations among quality variables are rarely considered, which, however, can provide additional valuable information. This article proposes a novel dynamic auto-regressive latent variable model (DALVM) to capture both auto and cross-correlations from high-dimensional time series data. DALVM is designed to maximize the covariance between the current quality score and the weighted sum of past quality and process scores, and an auto-regressive exogenous inner model is developed for consistency purposes. Further, a concurrent anomaly detection system is developed based on DALVM, referred to as ConDALVM, which conducts subsequent decompositions in the extracted latent spaces. ConDALVM realizes comprehensive monitoring for both static and dynamic anomalies in process and quality spaces. The superiority of the methods is demonstrated through a numerical simulation and
two industrial processes.
Then to continue to improve the flexibility and performance of DALVM, an efficient dynamic auto-regressive CCA (EDACCA) is also developed, with an emphasis on missing data imputation. an efficient dynamic auto-regressive canonical correlation analysis (EDACCA) method is proposed with a modified auto-regressive exogenous model to extract dynamics in both auto-correlations and cross-correlations. The flexibility and efficiency of EDACCA are improved with the design of weighting parameters and the economic singular value decomposition. EDACCA is further adapted for multi-step ahead (MS) prediction and missing data imputation. Two industrial processes are employed to evaluate the prediction performance and imputation performance of EDACCA.
After that, the exponentially weighted moving average model is adapted into dynamic SFA to capture dynamic relations among data with great interpretability. Slow variations of process variables retain normal operating conditions and important patterns of manufacturing processes, and slow feature analysis (SFA) is widely adopted to study slowly varying signals. However, SFA only extracts first-order dynamics, and most of its multi-step dynamic extensions focus on predictability instead of slowness. Then, a novel exponentially weighted SFA (EWSFA) is proposed to extract slow features by following the decaying nature of temporal signals. Exponentially decaying weights are learned in EWSFA to model dynamic relations between current and historical samples, and its computation efficiency is further improved with the aid of singular value decomposition. An EWSFA-based monitoring scheme is developed to detect anomalies in both slow and fast variations, and a novel dynamic index is designed to identify the occurrence of operating condition changes. The effectiveness of EWSFA is demonstrated through two industrial case studies.
Finally, for future works, the power of neural networks will be effectively incorporated into the design of dynamic algorithms to deal with complex nonlinearities. | en |
