A new process monitoring method based on noisy time structure independent component analysis

Conventional process monitoring method based on fast independent component analysis (FastICA) cannot take the ubiquitous measurement noises into account and may exhibit degraded monitoring performance under the adverse effects of the measurement noises. In this paper, a new process monitoring approach based on noisy time structure ICA (NoisyTSICA) is proposed to solve such problem. A NoisyTSICA algorithm which can consider the measurement noises explicitly is firstly developed to estimate the mixing matrix and extract the independent components (ICs). Subsequently, a monitoring statistic is built to detect process faults on the basis of the recursive kurtosis estimations of the dominant ICs. Lastly, a contribution plot for the monitoring statistic is constructed to identify the fault variables based on the sensitivity analysis. Simulation studies on the continuous stirred tank reactor system demonstrate that the proposed NoisyTSICA-based monitoring method outperforms the conventional FastICA-based monitoring method.

A NoisyTSICA algorithm which can consider the measurement noises explicitly is developed to estimate the mixing matrix and extract the independent components (ICs). The commonly used PE index is adopted for quantitative comparison. A smaller PE suggests that the corresponding algorithm has a higher accuracy for estimating the mixing matrix and extracting the ICs. It can be easily observed that most of the PE values of the NoisyTSICA algorithm are much lower than the PE values of the FastICA algorithm.D2 monitoring statistic is built to detect process faults on the basis of the recursive kurtosis estimations of the dominant ICs. A contribution plot for the D2 monitoring statistic is constructed to identify the fault variables based on the sensitivity analysis.Figure optionsDownload as PowerPoint slide