A new incipient fault monitoring method based on modified principal component analysis

Abstract

A novel multivariate statistical process monitoring (MSPM) method based on modified principal component analysis (PCA) is proposed to solve the low detection rate problem of incipient fault. In this modified PCA, on the basis of normal PCA model, the columns of loading matrix are reordered by mutual information between different statistic component matrices and training data. Then, instead of cumulative percent variance criterion, the principal component subspace is selected according to the largest mutual information. The selected PC subspace can maximally reflect fault characteristics into a new statistical index. Besides, a detection index based on the sliding average control chart statistic is proposed, which eliminates the effect of noise by proper averaging of the most recent samples, greatly improving the ability and accuracy of fault detection. The case study of a numerical simulation and Tennessee Eastman process shows that the proposed method can effectively detect incipient fault.

Just‐in‐time latent autoregressive residual generation for dynamic process monitoring

“…Particularly, multivariate statistical process monitoring (MSPM) with a goal of extracting normal statistical signatures from a dataset given from the normal operating condition (NOC) has become a quite popular methodology for detecting possible faults or anomalies in industrial plants 4–6 . On the basis of two fundamental analytical algorithms, that is, principal component analysis (PCA) and independent component analysis (ICA), 7–9 plenty of unsupervised modeling methods for representing the normal variation in the given dataset have been motivated 4–11 …”

A Novel Process Monitoring Method Based on Dynamic Related ReliefF-SFA Method

“…As the important researching branch of the data-driven process monitoring methods, multivariate statistical process monitoring (MSPM) has received extensive attention [1]- [4], which can extract feature information representing complex industrial data. Some traditional methods such as Principal Component Analysis (PCA) [5], [6], Partial Least Squares (PLS) [7], [8], Fisher Discriminant Analysis (FDA) and Canonical Variate Analysis (CVA) [9], [10] have been widely considered. As one of data-based fault diagnosis methods, Slow feature analysis [11] (SFA), as a new feature extraction and dimensionality reduction method, has been researched widely in the fields of process monitoring and fault diagnosis.…”