State Space Model Based Approches For Complex Dynamic Process Monitoring

Process monitoring technologies are important in the process automation systems. Since process monitoring technologies make the operation of processes safe and stable, and improve the quality of products in modern industrial processes, therefore, the research work on process monitoring has great industrial value. In modern industry, due to the expansion of process scales and the process complexity, it is difficult to obtain the mechanism model. Meanwhile, the popular applications of distributed control systems in industrial processes and the development of computer science lead to huge amounts of process data accumulated in the industrial enterprises. Therefore, multivariate statistical process monitoring has become a popular research subject. In the past two decades, there are lots of progresses in both academic research and industrial applications.Due to the intrinsic characteristics and the wide applications of closed-loop control technology, dynamic behavior is common in industrial processes. In recent years, researchers have made efforts in the area of dynamic process monitoring and proposed lots of approaches, such as dynamic multivariate statistical process monitoring methods, time series models, state space model based methods. These methods assume that the dynamic process operates in a single operating condition, process variables are linearly correlated and Gaussian distributed. However, in industrial applications, the dynamic processes tend to operate in a more complex environment, making the process variables nonlinear, non-Gaussian, multimode, and affected by the process noises. Based on the state space models, this thesis proposed several novel methods and solutions to model and monitor complex dynamic processes.(1) A dynamic modeling and monitoring method based on sparse partial least squares (SPLS) was proposed in the case of small sample issues. Via sparsity constraints on the projection coefficients, SPLS is able to obtain sparse projection vectors and select the most relevant variables in subspace identification. In the case with small number of training samples, the SPLS based model can also work well, reducing the over-fitting problem greatly and enhancing the generalization performance.(2) For dynamic processes in noisy environment, a process monitoring method based on linear Gaussian state space models (LGSSM) was proposed. Similar to probabilistic principal component analysis (PPCA) and factor analysis (FA), LGSSM models the statistical characteristics of process noise as well as the latent states. The model parameters are estimated via the expectation-maximization (EM) algorithm. In LGSSM, the estimation of the states and the prediction of the observations are based on Kalman filter. Monitoring statistics are set up in the state space and residual space, seperately. To identify fault related variables, a fault identification algorithm was proposed, based on the contributions to the monitoring statistics in the residual space. Compared with the other fault identification methods in principal component analysis, the proposed algorithm makes use of dynamic information and avoids fault smearing problem. Therefore, the proposed method achieves better performance in both fault detection and fault identification.(3) A multimode dynamic process monitoring method based on mixture canonical variate analysis (MCVA) was proposed. In MCVA, it is assumed that the process is constructed with different modes, each of which follows Gaussian distribution. In each mode, a model based on canonical variate analysis (CVA) is used to model the dynamic behavior of the process. In process monitoring step, the posterior probabilities of each sample corresponding to different Gaussian modes are calculated via Bayes rule and marginal distributions. The local monitoring statistics are calculated in each Gaussian mode and integrated into a global monitoring index according to the posterior probabilities. Compared with the traditional multimode process monitoring methods, e.g., Gaussian mixture models (GMM), and the tradional dynamic process monitoring methods, e.g., CVA, the proposed approach shows its superior in monitoring multimode dynamic processes.(4) A kernel partial least squares (KPLS) based process monitoring approach was proposed to model and monitor nonlinear non-Gaussian dynamic processes. Different with the linear subspace identification methods, KPLS maps the input model variables to a feature space and capture the realtionship between the mapped inputs and the outputs. Therefore, both the process dynamics and nonlinearity are modeled simultaneously by KPLS. By introducing the kernel trick, the nonlinear mapping between the original input space and the feature space is implicit and the complex nonlinear calculation is avoided. In process monitoring, since the scores (states) of KPLS are non-Gaussian, the traditional monitoring statistics based on Gaussian distributed variables is no longer appropriate. In order to monitor the non-Gaussian states, support vector data description (SVDD) was developed as the monitoring method. SVDD does not depend on the distribution form of the monitored variables. Therefore, the proposed approach achieves good performance in monitoring nonlinear and non-Gaussian processes.