Research On Outlier Detection Methods For Industrial Processes

The problem of outlier detection in industrial processes has been promoted by extenstive applications of data-driven techniques.This problem sufers from various challenges due to the system complexity and harsh working conditions.Among these challenges,the absence of data label should be the most prominent.We have observed that making lables for industrial data is an inherently difficult work,since it requires domain experts to complete it manually.Then,multiple data components should be another big challenge.This issue results from multiple working points in industrial applications.In addition,high-dimensionality of data is also a dominant feature.As modern industrial systems become increasingly complicated,sizes of process variables become larger accordingly.This problem has triggered challenges for most data-driven techniques.Most off-the-peg outlier detection methods can hence hardly be applied directly to industrial processes.Dedicated outlier detection methods for industrial processes are necessary.To this end,this dissertation proposes several such outlier detection models,and we conclude the main contributions as followings:1.A semi-supervised learning strategy,i.e.one-class classification is employed here to alleviate the problem of the absence of data label.When compared with unsupervised learning,semi-supervised learning has less computational complexity at the test phase.This is more appropriate for online applications;2.A robust detection method based on SVDD(support vector data description)is proposed to improve the adaptation of one-class classifiers to complicated data components of industrial data.In this method,a data processing strategy like Boosting is proposed.The bias of the detection model can be reduced by adjusting training samples;3.To reduce the high uncertainty of single model,and alleviate the problems of model selection and curse of dimension,ensemble outlier detection methods are proposed.Under this framework,the initial and core detection models are constructed successively,in order to adapt to different scenarios in industrial processes.Furthermore,a clustering-based ensemble detection is proposed when considering the situation where training samples from multiple working points are included.The objective is to reduce the rate of false negative.For the sake of mining the potential of base learners sufficiently,a dynamic ensemble of outlier detection method is proposed.4.Finally,the problem of outlier detection is extended.The issue of prediction is of great importance in industrial applications.We propose to combine these two problems together,aiming to detect measurements online and obtain data sets with labels.The key idea is to employ a regression model belonging to Bayssian learning.