Ensemble Based Semi-supervised Learning For Fault Classification

In modern industry,timely identification of malfunctions and corresponding types plays important role in satisfying the demands of production safety and stability of production quality.Traditional process monitoring methods mainly rely on mechanism-based model,knowledge-based model and so on.Recently,with the development of modern industrial processes and the generation of industrial process data,data-driven process monitoring algorithm has been emerged.However,in actual industrial processes,labeled data is often very scarce.This has restricted the application of traditional and matured supervised algorithms in industrial processes.At the same time,semi-supervised learning algorithm has gradually drawn a lot of scientific attention.However,semi-supervised learning algorithms often have the drawback of instability.Ensemble learning algorithm can integrate,weak learners based on certain rules to generate a more generalized and powerful learner.Here,in this paper,based on the fact of lacking labeled data and strong dependence of data distribution of semi-supervised learning algorithms,semi-supervised learning algorithms and ensemble learning have been combined together to perform a series of semi-supervised ensemble learning.Specific works include:(1)In order to avoid the accumulation of false label errors commonly found in ensemble semi-supervised learning algorithms and to use as much information as possibility on unlabeled samples,based on Fisher Discriminant Analysis and Principal Component Analysis,an Ensemble Semi-supervised Fisher Discriminant Analysis classification method(ESFDA)combined with metric-level fusion algorithm has been emerged.This algorithm invokes principal component analysis to introduce global data distribution information and uses the metric layer fusion algorithm to improve the generalization performance of base classifiers.Algorithm simulation was completed based on Tennessee Eastman platform,and the effectiveness of algorithm was proved.(2)To further improve the diversity among base classifiers,an adaptive method based ensemble semi-supervised learning(Ada-ESFDA)malfunction classification method is proposed.This method improves the accuracy of the model by adaptively adjusting the weights of labeled data during the iteration process to improve the diversity among the classifiers.At the same time,by adjusting the weight coefficients of each classifier,the accuracy of the model is further improved.The simulation results confirmed the improvement of the algorithm classification effect by the adaptive improvement.(3)As extremely lacking of labeled samples,a malfunction classification method based on active-learning ensemble semi-supervised learning(Active-ESFDA)has been provided.This method introduces the active learning method to provide the samples that have the largest information for model training,achieving better classification performance with less labor cost.The effectiveness of the algorithm was verified by platform simulation.At the end of the essay,a series of research contents of this paper are systematically summarized,and future research work is carried out.