Fault diagnosis of dynamic multi-variate chemical processes using pattern recognition, and smooth representation of trends by a wavelet-based technique.

摘要

Fault detection and diagnosis have become important topics in process industries. The diagnosis of transient upsets in the system behavior can lead to important process or operation modifications that can improve the future behavior of the system, and minimize the impact of costly effects.; A methodology is presented to diagnose patterns generated by a single-variate, dynamic, and continuous system. The patterns are the transient trends of process variables resulting from disturbances in the system. Because large chemical processes lack an accurate dynamic model, a pattern recognition methodology, consisting of feature extractor and feature classifier, is proposed. The Linear Discriminant Basis (LDB) method is modified and adopted as the basis of the proposed feature extraction technique. By using the “Double Wavelet Packet Tree” (DWPT), the input pattern is split amongst a set of most class-discriminant windows. Then Principal Component Analysis (PCA) is used to reduce the number of selected features. The feature classifier is a binary decision tree, based on the Incremental Tree Induction (ITI) technique, coupled with a soft thresholding scheme for recognition of noisy input pattern.; The above mentioned methodology is then extended to extract and classify features of patterns generated by a multivariate, dynamic, and continuous system. PCA technique is applied so that the information space is described by a set of uncorrelated and fictitious data sources. The most class-discriminant features and a binary decision tree are determined for each new data source. Each decision tree outputs not only posteriori probabilities of assigning an input pattern to any one of classes but also confidence limits for the probabilities. The consensus theory and evidence theory is utilized in this work to find the best classes of events for describing system behavior.; Approximating a data stream by a smooth function, which preserves frequency content and duration time of different events is quite useful for certain applications. A methodology is presented to generate a smooth function for a set of discrete and finite data. The objective function is expressed as the weighted summation of some wavelet-based basis functions. The required wavelet filters are provided by a Lifting-Scheme wavelet method. To determine the basis functions and the coefficients, the Wavelet Packet approach is coupled with the Lifting Scheme. An iterative technique is also developed to remove the noise elements from the data set.