A Novel Ordering Method for Independent Component and Its Industrial Application Protocol

Abstract

In this research, a novel ordering method for independent component (IC) and its industrial application protocol are proposed. The goal is developing an IC ordering method that has a more practical background and comparing it with the conventional ordering method. An additional goal is to develop a protocol for industrial application of the proposed method. In Chapter 2, A new IC ordering method has been proposed. The algorithm sorts the ICs by the order that has the greatest effect on the measurements using the mixing matrix. The mixing matrix is computed using the de-mixing matrix computed from the FastICA algorithm. The order of the absolute values of the components of each row of the mixing matrix is given in descending order, and the values of the weight matrix are assigned sequentially from the largest value. Then, each column of the assigned matrix is added to determine the order of the final IC. A simple example and the Tennessee Eastman process (TEP) were used to verify that the order of the important ICs is in order. As a result, it has been verified that the proposed method correctly orders the ICs that are most important and shows the best process fault detection rate in the TEP process. In Chapter 3, A protocol for applying the proposed IC ordering method to the hyper-compressor in LDPE production processes has been proposed. First, the normal operation data is extracted using the interquartile range (IQR). Using the normal operation data, the IC is extracted based on the MICA, and the ICs are sorted using the proposed ordering method. Then the T2 and SPE statistics of normal operation are calculated by extracting the number of ICs that the total sum of eigenvalues reaches 80%. The control limits, which are the criteria for the detection of abnormalities, are calculated in the normal operation data and the control limits of the contribution of each variable for identifying causal variables are also calculated. Then, and SPE of the total operation data are calculated. If each statistic exceeds each control limit, it is assumed that an anomaly in the process has been detected and a contribution plot is calculated at that time. If the contribution of a variable exceeds the control limit, it is considered to be the cause of the process, and the fault propagation pathway is searched using Granger causality. The causal variables may be affected by other variables or affect other variables. Therefore, all the variables related to the anomaly are reported to the operator. As a result, the four out of five shutdowns during the operation were found and the causal variables of each shutdown were extracted. It was confirmed that the variables showed abnormal behavior.