Fusing Active Sample Labeling Based Semi-supervised Learning Chemical Process Fault Classification Methods With Applications

Supervised learning-based chemical process fault classification methods demand for sufficient labeled fault samples.While semi-supervised learning is able to involve some unlabeled samples into learning so as to reduce the dependence of fault classifiers on the amount of labeled samples.However,semi-supervised learning still relies on the quality of labeled fault samples.In this regard,the thesis proposes an improved active sample labeling method which can significantly improve the classifier performance by comprehensively evaluating unlabeled samples and screening high-quality samples for labeling.Consequently,a semi-supervised learning chemical process fault classification method incorporating active sample labeling is developed.The main contents and results of the research are presented as follows.1.Aiming at improving the quality of labeled fault samples,an improved active sample labeling method is proposed,which helps to select excellent samples for expert labeling considering information amounts and representativeness of samples.Taking account of the existing cosine similarity assessments insensitive to absolute values,a comprehensive approach is established to evaluate samples using a combination of the Euclidean distance and the cosine similarity,selecting samples with high representativeness and rich information for labeling.2.A semi-supervised learning chemical process fault classification method incorporating active sample labeling is developed,which is able to actively screen chemical process fault samples for labeling,constructing training data sets for classifiers.A Tri-training strategy based semi-supervised learning algorithm is established,which maximizes the classifier difference by assigning different weights to the three classifiers according to their different abilities for fault identifications,thereby effectively improving the performance of semi-supervised learning classifications.3.The proposed method is applied to an industrial coal gasification process.Process fault data are collected,and semi-supervised learning fault classification experiments are conducted using actively labeled samples and randomly labeled samples,respectively,achieving satisfactory results and verifying the effectiveness and advantages of the contribution.