Nonlinear Fault Detection Based On KECA

Process monitoring is able to monitor the production process in real time,which plays an in-creasingly irreplaceable role in guaranteeing stable operation,improving product quality and reduc-ing energy consumption,etc.Big data era is coming along with the fast development of information technology and the continuous rising level of hardware in storage and computing.The greatly en-riched data makes the data-driven process monitoring methods become a Popular research field in recent years.Accordingly,these rich data also make the data-driven process monitoring methods face more challenges.In this paper,two kinds of nonlinear fault detection methods based on kernel entropy component analysis(KECA)is proposed for kernel-based method and Manifold Learn-ing owing to the nonlinear characteristic of industrial production process.The specific research includes:Firstly,considering that KECA model with single kernel parameter can not effectively detect the different types of faults in industrial process,an improved KECA fault detection method based on ensemble learning and Bayesian inference is proposed.Because different types of faults often require different size of the kernel parameters to make the model performs a better detection results,we construct off-line KECA models with different kernel parameters in this paper.And all these models are trained using the same training data.After modeling the normal condition process,the on-line detection results of the models are transformed into probabilistic forms by Bayesian inference.The detection results of these probabilistic forms are combined according to the weight to obtain a final decision in which a good detection results for specific faults has a large weight.Therefore,this method can have better fault detection effect for different types of faults.Secondly,due to the merit of KECA's ability to comprehensive select the projection direction of the input data in the dimension reduction process,this paper introduces the idea of information entropy into the maximum variance unfolding(MVU),and proposes a KECA-MVU based nonlin-ear fault detection method.The Renyi entropy is used to measure the validity of the information retained by the kernel matrix obtained by MVU.Then,the feature vectors corresponding to the first few terms of Renyi entropy is taken as the projection direction of the data.And the effective compression of the data is realized through the selected feature vectors.Finally,the linear regres-sion method is adopted to estimate the optimal projection matrix from the input data to the low dimensional structure.The projection matrix is used to achieve the on-line fault detection.Finally,the main research results are summarized,and the difficulties and prospects of the future research work are expounded.