Since industrial processes become more and more complex, the requirements of the stability, efficiency and safety of production are increasing. In order to get more accurate and timely diagnosis of complex industrial process fault, it is necessary to study more comprehensive theory and method of fault diagnosis.This paper first introduced the development of fault diagnosis. Then, the detail of the data-based fault diagnosis methods are discussed especially independent component analysis (ICA). ICA has been developed in recent years. This algorithm obtains the Independent Components (ICs) by maxizing the non-Gaussinity for fault diagnosis. The advantages are that it doesn’t assume the process variables to meet the Gaussian distribution as Principal Component Analysis (PCA), and the higher-order statistics information can be used. All these make ICA more practical than PCA. The most popolar mehod of ICA is the fixed point algorithm, which is knowed as FastICA. The advantage of FastICA is the volecity of convergence is very fast because of the Newton method. But, Newton method can not keep its convergence ability when the search region is far from the optimum. These have reduced the proformance for fault diagnosis. In order to improve the deficiencies, this paper uses the Particle swarm optimization method (PSO) instead of the Newton method. And a new method for ICs order is proposed. All these construct the PSO-ICA algorithm, which can make the fault diagnosis result more accurate.An extension of ICA for quality preduction is Independent Component Regreeeion (ICR). ICR has the same advantages as ICA. These advantages make ICR more practical than PLS for quality preduction. But, ICR also have his own disadvantages. In ICR, the ICs are extracted using ICA exclusively based on the process measurements. This estimation does not take into account the role of ICs in quality prediction and thus is not optimized for quality prediction. For solving these problems, the quality information between the process variables and qulity variables which was in Partial Least Square (PLS) is introduced into the ICR. All these construct the Modified ICR algorithm. The selection of ICs simultaneously considers both the quality-correlation and the independence, which makes ICs more suitable for quality prediction and improves the proformance of quality prediction.These methods are applied to Tennessee Eastman process and the continuous annealing process for fault diagnosis and quality prediction respectively. Some methods are used for comparision. The results indicate that the proposed methods are effective. |