Fault Diagnosis For Industrial Processes Based On Machine Learning

Fault diagnosis is one of the important parts of the industrial process automation system,and it aims to detect and eliminate the fault timely and effectively.Therefore,it is of great significance to ensure production safety and product quality.On the one hand,with the increasingly fierce market competition,modern industries tend to be more and more complex and large-scale.On the other hand,process data can naturally reflects the process state,and the fast development of information technologies such as the distributed control system,the industrial internet of things and the intelligent instruments enables the easy acquisition and storage of process data.Hence,in this era of intelligent manufacturing,data-based fault diagnosis has been one of the hotspots in the area of process control.The detection,location and classification of faults are the main subjects of fault diagnosis.For the fault detection and location of the large-scale process,distributed modeling is often utilized.The distributed principal component analysis(DPCA)is a classical and pioneering distributed modeling method,since it can conduct block division purely based on data and without any prior knowledge about the process.However,there are still some important and ignored defects for DPCA.For the fault classification,in the traditional fault classification models,training samples are usually assumed to be labeled ideally,and this restricts the practicability of the fault classification model.Therefore,the problem of fault detection and location for the large scale process,as well as the problem of fault classification in the case of imperfect labeling for training samples,are studied in this thesis,in which the following new methods are proposed(1)As to the problem of fault detection and location for the large scale process,since the classical DPCA method has some important and ignored defects in the selection of sub-block variables,the utilization of dynamic information of process data,and the reliability of fault location,a performance-driven distributed canonical variate analysis(DCVA)method is proposed.Firstly,with historical fault information,the genetic algorithm is utilized to select appropriate variables for each sub-block.Secondly,canonical variate analysis is introduced to capture the dynamic information of process data for performance improvement,and to establish the fault detection model for each sub-block.In this way,the fault detection for the large scale is performed based on the DCVA model.Finally,a novel fault location method based on the contribution plot is developed for the DCVA model,and the reliability of fault location is improved.(2)As to the problem of fault classification in the case of imperfect labeling for training samples,due to the label-noise in the real labeling,a manifold-preserving sparse graph based ensemble Fisher discriminant analysis(FDA)model is initially proposed.Firstly,a manifold-preserving sparse graph is utilized to filter the training samples,and this can not only eliminate the obviously label-noise,but also retain the training samples with stronger information representation ability.In this way,the information quality of training samples is improved.Secondly,based on Bagging and FDA,ensemble learning is employed to further upgrade the robustness of the fault classification model(3)As to the problem of fault classification in the case of imperfect labeling for training samples,considering the fact that online samples may come from new fault classes,as well as the limited information representation ability of available historical training samples,an online learning based FDA model is proposed.This method can perform online model updating based on incremental learning.For online updating,a novel online sample selection criterion is developed.And this criterion can recognize samples from new fault classes through a max-min deviation rule with online updating capability,as well as ensure the reliability and the information content of the chosen online sample by the weighting sum of its maximum posteriori probability and information entropy(4)As to the problem of fault classification in the case of imperfect labeling for training samples,since only a minority of the training samples are labeled in real industrial processes,and the limited and costly expert labeling resources should be utilized as effectively as possible,an active learning based semi-supervised exponential discriminant analysis(EDA)is proposed.Firstly,to make EDA applicable to the semi-supervised industrial scenario,scatter matrices are transformed into their regularization variants through combining unlabeled training samples.Secondly,based on active learning,the best versus second-best rule is employed to select more informative training samples to upgrade the model classification performance in an active way and to utilize the expert labeling resources in an effective way.