Condition monitoring and fault diagnosis by principal component analysis and nonlinear PCA

Early detection and diagnosis of process faults while the plant is still operating in a controllable region can help avoid abnormal event progression and reduce productivity loss. Statistical Process Monitoring based on Principal Component Analysis (PCA) is the most effective way for condition monitoring in the process industry.; An early warning system based on linear PCA is developed at first in this dissertation because linear system theory is highly developed, well-defined by the mathematics and computationally efficient.; A major limitation of PCA based monitoring is that the PCA model is linear, while most real processes are nonlinear. An Auto-Associative Neural Network is used for a nonlinear early warning system in this dissertation, although it needs further improvements.; The real-time monitoring of a chemical process or power generation process with multiple operating modes is a challenging problem. One global model can not give satisfactory monitoring performance for a wide operating range. Use of a model family containing adaptive local models based on a moving window data matrix is proposed to cover a wide nonlinear operating range. Dividing the whole operating regime into several sub-ranges, then selecting a local model for each sub-range and updating the local model to current operating conditions are the main features for this proposed method. This method can be used for highly nonlinear processes with multiple operating points.; The PCA model is static, while most real processes are dynamic. Incorporating dynamic information by including lagged information improves the performance of disturbance detection. Nonlinear Dynamic PCA (NDPCA) is proposed here to detect disturbances for nonlinear dynamic systems.; Because the development of nonlinear methods from nonlinear system theory is still crude and incomplete, comparison between the Principal Curve (PC) and the Auto-associate Neural Network (AANN) is conducted and it shows the AANN does not give the optimal projection because of the continuous projection. The half Auto-Associative Neural Network (HAANN) based on nonlinear optimization is proposed to provide a neural network implementation of Principal Curves to avoid the continuous projection.