New Nonlinear Approach for Process Monitoring: Neural Component Analysis

Abstract

Nonlinearity is extremely common in industrial processes. For handling the nonlinearity problem, this paper combines artificial neural networks (ANN) with principal component analysis (PCA) and proposes a new neural component analysis (NCA). NCA has a similar network structure as ANN and adopts the gradient descent method for training, hence it has the same nonlinear fitting ability as ANN. Furthermore, NCA adopts PCA's dimension reduction strategy to extract the uncorrelated components from the process data and constructs statistical indices for process monitoring. The simulation test results show that NCA can successfully extract the uncorrelated components from the nonlinear process data, and it has better performance than other nonlinear approaches.

Survey on recursive statistical process monitoring methods

“…Based on the publications reviewed in this paper, some challenges in the field of RSPM can be drawn as follows: DPM, FOP, and ROM algorithms still have their own shortcomings, and other recursive computation algorithms that have better computation efficiency in feature calculation stage should be considered in future RSPM‐related research. Most of the existing RSPM methods are incremental improvements of RPCA and RPLS, and while some novel RSPM methods applied to other MSPM methods, such as CCA, could be the future research challenge for RSPM. Besides, some process monitoring methods for KPI monitoring and non‐linear processes, such as orthonormal subspace analysis, [ 70 ] neural component analysis, [ 71 ] and artificial neural correlation analysis, [ 72 ] still do not have recursive versions, which will be our future research work. …”

“…Most of the existing RSPM methods are incremental improvements of RPCA and RPLS, and while some novel RSPM methods applied to other MSPM methods, such as CCA, could be the future research challenge for RSPM. Besides, some process monitoring methods for KPI monitoring and non-linear processes, such as orthonormal subspace analysis, [70] neural component analysis, [71] and artificial neural correlation analysis, [72] still do not have recursive versions, which will be our future research work.…”

Process Monitoring Using a Novel Robust PCA Scheme

“…Multivariate statistical-based process monitoring (MSPM) − methods are effective approaches for monitoring modern large-scale industrial processes, which contain large amounts of process and quality data. The most commonly used MSPM method is principal component analysis (PCA), − which transforms high-dimensional process data into a small set of uncorrelated principal components and then monitors these components using several statistical indices. In classical PCA, the data model is obtained by calculating the covariance matrix and then decomposing it using singular value decomposition …”

C-IPLS–IKPLS for Modeling and Detecting Nonlinear Multimode Processes

“…In addition, instead of extracting the nonlinear relations, they can only describe the statistical characteristics of linear processes. Although the corresponding nonlinear regression algorithms − can be used to address nonlinear relations, they also have imprecise separation similar to linear regression algorithms. Moreover, for nonlinear processes containing linear and nonlinear relations, it is difficult to accurately establish the related relation using only one nonlinear model.…”