Research Of Fault Detection Methods Based On Support Vector Data Description And Subspace Indentification

In chemical industry,electricity and other industrial processes,if the fault can not be detected timely,it will cause serious economic losses and even threaten the safety of life.With the increasing complexity of the system and the rapid development of computer and sensor technology,the data-driven based fault detection methods are widely studied.In recent years,the fault detection method based on Support Vector Data Description(SVDD)has attracted much attention because it can deal with non-Gaussian and nonlinear problems.However,in practical application,there are many challenges and problems.This paper mainly handles the following three problems:(1)SVDD has a potential assumption that the process samples in feature space should roughly distribute within a hyper-spherical region.However,it's easily violated in practice,suffering a serious missing detection;(2)Choosing parameters of model is difficult and there is no uniform standard of optimization;(3)SVDD method,including some traditional data-driven fault detection methods,are easily applied.But the detection results provide little effective information for subsequent tasks,i.e.,fault isolation and identification.To address the above three issues,the corresponding methods for each problem are proposed in this paper,and finally a multi-method based fault detection strategy is designed.This paper completes the following innovative work:Firstly,for the problem(1),the influence of the assumption on the detection performance is analyzed,based on which the Hyper-ellipsoid Support Vector Data(ESVDD)method is proposed.By introducing the Mahalanobis distance,the proposed method can reduce the unnecessary area surrounded by decision boundary,which reduces the risk of fault missing detection effectively.Furthermore,the Sequential Minimal Optimization(SMO)is designed to solve the ESVDD detection model.Finally,the simulation results on different data sets show that missing detection rate of the proposed method is significantly reduced.Secondly,for the problem(2),the limitations of the one-class method and the parameter optimization algorithm are analyzed.On this basis,a fault detection method named Harmony Search-based Hyper-ellipsoidal SVDD with Fault samples(HS-FESVDD)is proposed.In one-class method,parameters are only chosen based on normal samples.But such parameters may be unsuitable for fault samples.To guarantee that the selected parameters are suitable for both normal and fault samples,the ESVDD with fault samples(FESVDD)method is proposed.The traditional parameter optimization algorithms,e.g.GA,often suffer the trouble of large computational load,and suffer the difficulty of fine-tuning the parameters of the algorithm itself.To address this issue,a novel method named Harmony Search(HS)algorithm is introduced to acheive the fast convergence.And an adaptive strategy of parameters is designed,which addresses the problem of parameter fine-tuning.Finally,the simulation results on the UCI dataset and the TE process show that the fault detection ability of HS-FES VDD is stronger than that of the similar methods and the detection methods based on principal component analysis.Thirdly,for the problem(3),a fault detection method named Moving Window Subspace Identification(MWSI)is proposed first.This method uses the advantage of the model-based fault detection method to express the fault information.The residual generator is directly constructed from the process observation samples via subspace identification method(SIM).By designing the moving window rules,the method can improve the ability of detecting small faults.Furthermore,a joint fault detection strategy ESVDD-MWSI is designed,in which ESVDD is utilized to meet the requirements of qualities of samples for MWSI.The simulation results show that the designed strategy has a strong ability to detect small faults,and the initial location of the fault can be achieved by test results,which provide the necessary information for fault isolation.