Research On Nonlinear Process Fault Diagnosis Based On Data-driven

With the development of digital technology and network solution, time-series data that is col-lected in the domain of large-scale chemical industrial exhibits some remarkable characteristics, including high dimensionality, nonlinearity, uncertainty and incompleteness, etc. These will cause a lot of problems in the procedure of data analysis, for example, inefficient computation, degen-erated performance and curse of dimensionality. These are great challenges for the traditional technology in data mining. Statistical learning theory has been attracted a lot of attentions in the academia and engineering by its intelligibility and tolerance for uncertainty. In considerations of process monitoring, this dissertation focuses on the discussions which are the appearance of uncer-tainty, high dimensionality and nonlinearity in the time series with combining soft computing and data mining. The more effective methods have been proposed for the problems, and the validity and superiority of the proposed methods are demonstrated on the numerical example and benchmark simulation. The contributions of this dissertation are listed as follows.Ⅰ. A novel multiple kernel learning based on inexact projection method is proposed, named GCISP-MKL(Generalized Convexity-based Inexact Projection for Multiple Kernel Learn-ing). In the condition of smoothness, this method is able to converge to the stationary point accurately and effectively in the limited memory space. A theoretical research on the gener-alized convexity shows that the objective function in the model is strictly pseudo-convex, and the constraint is pseudo-linear. According to the theorem, we can get the fact that the primal and dual problem is strong duality. In addition, a nonmonotone gradient method is proposed to optimize the kernel weights and the Hessian matrix is approximated with a variation of Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) approach.Ⅱ. To overcome the monotonicity of existing feature selection methods, a new monotonic fea-ture selection algorithm based on multiple kernel scheme is proposed, named PrimMKL-FS (Primal method for Multiple Kernel Learning-based Feature Selection). The importance of feature in the large-scale process can be indicated by the weights of kernel matrices which is obtained by the corresponding optimization. The weights in the model are updated by the Nesterov’s normal projected gradient descent method in the primal. The optimal assign-ment can be performed on the basis of quantity solved by the optimization problem.Ⅲ. To overcome the drawbacks of correlations among the variables not satisfying the funda-mental assumption in the practical process, that is the monitored variables follow a Gaussian distribution, a scheme based on ICA and deep architecture is proposed, named ICA-Sparse Autoencoder. The latent variables in the non-Gaussian distribution can be extracted by ICA, then a deep architecture is used for the residual sequence to handle the nonlinear problems. Meanwhile, a pre-train process based on a restricted Boltzmann machine is proposed to over-come the problem of local minima in the optimization of sparse autoencoder. The parameters in the deep architecture is optimized by the L-BFGS algorithm.Ⅳ. A robust Bayesian model based on Gaussian mixture model, named BRR (Bayesian Robust Regression), is proposed to overcome a drawback in the process monitoring which is difficult to have a probabilistic output. One of the major limitations in GMR is its lack of robustness to outliers, since the estimates of the means and the precisions can be severely affected by atypical observations. Another issue is the infinite trouble, which is using maximum likelihood to fit a Gaussian mixture model (GMM).Specifically, a prior is placed on the mixture component in order to identify outliers in the process. Moreover, a separate precision weight is used for the inverse covariance matrix to address the infinite trouble. The prior in the mixture weight is designed by a Dirichlet distribution in which the parameters is a probability of the mixing proportions.Ⅴ. A key issue in the GMM is that the performance is often prone to overfitting with respect to the number of mixture models. A nonparametric Bayesian fault detection method based on Dirichlet process mixture model is developed. The model addresses this problem by as-suming that there is an infinite number of latent clusters, the posterior provides a distribution over the number of clusters, the assignment of data to clusters, and the parameters associated with each cluster. In this method, a prior is placed on the mixed weight which is builded as a Dirichlet process mixture by the stick-breaking construction. The truncation is used during the progress of variational approximate inference. Meanwhile, the truncated model which is adjusted according to the variational free energy makes the model selection worked together with the variational inference.Finally, further emphases and directions in the research area are given by the fact which is based on a summary and discussion of this thesis.