LINEAR AND NONLINEAR INTERVAL MULTISCALE PCA BASED FAULT DETECTION METHODS

Abstract

Statistical process control is an integral set of techniques in chemical engineering that can help guarantee the effective operation of various engineering systems. A multitude of techniques have been and are still being developed in order to detect and monitor faults in processes using process data. Deviations from the normal operating conditions can cause a myriad of abnormalities in the data, such as temperature and pressure readings, which can then be used to detect faults in the process. Processes variables can be linearly or non-linearly correlated, and the aim of this work is to start off with linearly correlated data, and then move on to non-linearly correlated data, which is more complex and requires more advanced methods to deal with. The basis of fault detection when the data is linearly correlated is principal component analysis, which is widely used. Two extensions of this method, interval PCA (IPCA) and multiscale PCA (MSPCA), have been developed to help increase the efficiency of fault detection. In MSPCA, the data are decomposed using wavelets at multiple scales, and PCA is applied at each scale before reconstructing the data back to the time domain, where PCA is applied again to detect faults. MSPCA helps deal with auto-correlated noise and reduces its effect on the accuracy of fault detection. Interval PCA, on the other hand, is a technique that helps deal with uncertainty in the data by converting the single valued data interval by aggregating the measured samples over a time horizon, and PCA is applied on the generated intervals. These two methods will be combined, and new fault detection method, called interval multiscale PCA (IMSPCA), will use the advantages of both to have more efficient fault detection method. For the nonlinear case, a neural network-based modification of the algorithm will be used to developed neural network IMSPCA (NNIMSPCA). Neural networks are a group of techniques, modeled after neurons in the human brain that are taught to recognize complex patterns in data.