Batch Process Monitoring Based On MICA Methods

Batch processes have become more and more important in modern industrial processes. In ensuring the safety and stability of batch processes and high quality final product, on-line monitoring and fault diagnosis in batch processes emerge as an essential and important task. As the development of on-line measurement instruments and computer technology, large amounts of process variables'data can be collected more easily than before. The data can be analyzed to supervise the process behavior, by mining the valuable information and resources.Multi-way principal component analysis (MPCA) and multi-way partial least squares (MPLS), which assume that the variables must subject to the normal distribution condition and only utilize the second-order statistical information, are used most widely multivariate statistical technique in batch processes monitoring. Multi-way independent component analysis (MICA), one type of multivariate statistical method based on ICA technique, is recently developed to apply to the batch processes monitoring. This method can treat with three-way data of batch processes more effectively because it utilizes the high-order statistical information and avoids the assumption of Gaussian distribution. In addition, the extracted latent variables by MICA are statistically independent while principal components generated from MPCA are merely de-correlated. Therefore, the independent variables or components can describe the processes characteristic more intrinsically than MPCA or MPLS. In this work, MICA batch monitoring method is discussed and considering the characteristics of batch processes, two new kinds of monitoring methods are proposed based on MICA.In view of batch-to-batch variation in most industrial batch processes, an adaptive MICA method is proposed to capture the dynamic variation among different batches. This approach first establishes an MICA model based on the historical database, then utilizes the initial monitoring model to detect the next new batch data. If the whole batch behavior is detected normally, we add the new normal batch to model database as the last batch and the oldest one is removed. On the basis of new database the old MICA model is revised by using forgetting factors to adapt to new normal conditions. Otherwise if some fault is detected in the new batch, the MICA model need not updated. The simulation results on a semiconductor etch process show that the proposed approach can construct more exactly statistical monitoring model by grasping new batch data variation information and thereby effectively reduces the false alarms generated by the fixed model.Considering the nonlinearity existing in batch data, a nonlinear monitoring technique is proposed in this paper, which is called multi-way kernel independent component analysis based on feature samples (FS-MKICA). This approach first unfolds three-way data under normal operation condition of batch processes to be two-way. For the unfolded data, the number of samples is too large to quickly obtain a kernel matrix when map the input data into a feature space nonlinearly. In order to reduce the computation complexity of kernel matrix, feature samples can be selected from the large two-way input training samples first, and then kernel ICA (KICA) transformation is performed in the low-dimensional space based on feature samples. Thus the nonlinear independent components are obtained. When the suggested approach is used for batch processes monitoring, I 2and SPE plots are supplied to detect the faults. FS-MKICA approach not only extracts the nonlinear feature of batch processes, but also reduces the computational cost based on whole input samples. The simulation results on penicillin fermentation process clearly demonstrate that the proposed nonlinear method is more sensitive than traditional linear method in fault detection.