An Enhanced Kernel Partial Least-Squares Fault Reconstruction Fused With Pattern Classification

Abstract

Partial least squares (PLS) is a well-known multivariate statistical process monitoring (MSPM) method. However, there are two key issues that restrict its application in reconstruction-based fault diagnosis, including weak fault representation ability and uncertainty caused by the overlap among different types of faults. To cope with these two issues, an enhanced kernel PLS (eKPLS) fault reconstruction approach fused with pattern classification is proposed in this study. For the first issue, a fine-grained fault subspace extraction method is developed. These fine-grained fault subspaces exhibit richer fault details, conferring upon the model heightened fault representation ability. For the second issue, the fault magnitude is supplemented into the paradigm of fault reconstruction, and a rediagnosis mechanism is designed to overcome the uncertainty in fault diagnosis. Thus, the ideas of fault reconstruction and pattern classification are fused into a consolidated framework, allowing us to complement their strengths together. Finally, several experiments on a hydraulic system and Tennessee Eastman process (TEP) are performed to validate the effectiveness of the proposed method.