Process Fault Detection And Diagnosis Based On Independent Component Analysis

Efficient fault detection and diagnosis play key roles in ensuring the plant safety and improving the product quality.Massive process data collected by the ubiquitous sensors in modern industrial processes have facilitated the development of data-driven based fault detection and diagnosis.Currently,independent component analysis(ICA)is the reearch hotspot of data-driven methods and attracts much attention from researchers.However,process data usually contain outliers or measurement noise,or have nonlinear structure,which is not effectively taken into consideration by the conventional fault detecton and diagnosis methods based on ICA.This dissertation considers the above three cases which process data frequently encounter,conducts the researches on robust ICA algorithm,noisy ICA algorithm and nonlinear ICA algorithm around the ICA theory,and further establishes monitoring statistics.From this,a set of ICA-based fault detection and diagnosis methods are obtained.The main research work and results are as follows.(1)To deal with the problem that outliers in process data may lead to the deteriorated fault detection performance,a robust ICA based process fault detection method is proposed.Firstly,through the robust whitening and the robust search of maximum non-Gaussian directions,a robust ICA algorithm which can reduce the effects of outliers is developed to estimate the mixing matrix and the independent components.Then,for each estimated independent component,a robust calculation method of monitoring statistic is constructed.The simulation studies on the mixing matrix estimation and the fault detection in the continuous stirred tank reactor(CSTR)system demonstrate that the proposed method can estimate the mixing matrix more accurately and can improve the fault detection performance effectively.(2)Taking into consideration the problem that Gaussian measurement noises contaminate process data and make the sensitivity of fault detection methods decline,a NoisyICAn based process fault detection method is proposed.Firstly,a noisy ICA algorithm referred to as NoisyICAn in this dissertation,which can attenuate the effects of Gaussian measurement noises effectively,is developed based on the fourth-order cumulant to conduct the mixing matrix estimation.The kurtosis relationships of independent components and process variables are subsequently obtained based on the estimated mixing matrix.Then,the kurtosis of each independent component is recursively estimated and two monitoring statistics which can further reduce the effects of Gaussian measurement noises are built.The simulation studies on the fault detection in a three-variable numerical system and the CSTR system demonstrate that the proposed method can improve the sensitivity for detecting process faults obviously and can give the timely fault indication.(3)Consideirng the problem that not all measurement noises are Gaussian distributed,a NRJDICA based process fault detection method is proposed.Firstly,A NRJDICA algorithm which can explicitly take measurement noises into consideration and is not subject to the effect of measurement noises' distributions is firstly developed to estimate the mixing matrix by whitening of process variables and performing the joint diagonalization of whitened variables' time-delayed covariance matrices.Subsequently,the relationships between the kurtosis statistics of independent components and the fourth-order cross cumulant statistics of process variables are derived based on the estimated mixing matrix to help sorting the estimated independent components and selecting the dominant independent components.The serial correlation information of each dominant independent component is then estimated by using the moving window technique,on the basis of which a noise-resistant monitoring statistic is constructed.The simulation studies on the fault detection in a three-variable numerial system and the CSTR system illustrate that the proposed method requires no assumption of measurement noises' distributiones and achieves the superior performance in terms of the sensitivity to process faults.(4)Considering the nonlinearity and the time correlation of process data,nonlinear fault detection and fault identification methods based on kernel time structure ICA are proposed.Firstly,a kernel time structure ICA algorithm which can tackle the nonlinear structure of process data effectively is proposed to estimate independent components from nonlinear process data without any strict assumption about the distributions of independent components for calculating monitoring statistics.Then,a nonlinear contribution plot method is developed to identify fault variables based on the idea of sensitivity analysis.The simulation studies on the fault detection and fault identification in the Tennessee Eastman(TE)benchmark industrial process illustrate that,the proposed method achieves satisfactory fault detection performance and can identify fault variables accurately.(5)In order to make full use of independent components extracted by kernel time structure ICA for fault detection,a nonlinear fault detection method based on kernel time structure ICA and weighted independent components is proposed.Firstly,the kernel time structure ICA is adopted to estimate independent components from nonlinear process data.Then,Gaussian mixture model is utilized to estimate the probability density of each estimated independent component.On this basis,the emerging probabilities of independent components' samples are measured to guide the weight allocation for independent components' samples,so that independent components associated with fault can be highlighted whereas independent components unassociated with fault can be suppressed at different sample time and the fault information can be captured duly and effectively.The simulation studies on the fault detection in a four-variable numerial system and the TE benchmark industrial process,demonstrate the superiority of the proposed method in fast fault detection.(6)Finally,both the level control system of double-holding water tank in the laboratory and the fluid catalytic cracking unit in the refinery oil plant of a petrochemical corporation are taken as the objects for verifying the effectiveness of the proposed methods.The results show that,in accordance with the demonstration of simulation studies on the numerial systems,the CSTR system and the TE benchmark industrial process,the proposed methods can also effectively detect faults for the devices both in the laboratory and in the industrial field.In this dissertation,a set of fault detection and diagnosis methods based on ICA are proposed and their effectiveness is validated by the theoretical analysis,the simulation studies,the level control system of double-holding water tank in the laboratory and the realistic fluid catalytic cracking unit.Moreover,the parameter selection issues in the proposed methods which have effects on the performance of fault detection and diagnosis are discussed in detail,and a series of the effectual strategies for parameter selection are provided.The proposed methods can provide new solutions to the problem of fault detection and diagnosis in realistic industrial processes and have the promising application prospect.