Study On Online Kernel Learning For Process Modeling

There is such a phenomena in many industrial processes:some critical variables closely related to the product quality cannot be measured online by sensors and the delayed information obtained from offline assay may have difficulty to meet the requirement of the following real-time monitoring, optimization and control. Due to the extremely complex mechanism, strong nonlinear and time-varying characteristics of those industrial processes, data-based soft-sensing approaches gained more and more attention. A large number of studies indicate that kernel learning (KL) methods, such as support vector regression (SVR) and least squares SVR (LSSVR) can achieve a preferable performance in dealing with nonlinear modeling within limited-sample circumstance. In addition, the online KL (OKL) algorithms, including recursive KL (RKL) methods in global learning framework and local learning approaches, show a better ability to trace the time-varying characteristics and reduce computation load.However, there are still some problems in these OKL methods:(1) In the updating procedure, sub-inverse technique is always executed only once for each new sample, which leads to a failure when OKL is used to identify some complex systems'model.(2) In OKL approaches, most models are established and updated with parameters fixed. The research on the online parameters optimization, especially for weighted kernel learning methods is still rare.(3) In local learning, the similarity measurement between different samples is estimated based on input information only, which leads to a waste of output information.(4) All the proposed OKL algorithms are supervised learning techniques, no semi-supervised approach has been proposed yet. In this dissertation, we will start with a brief review of online kernel learning. Then a few modification of RKL are proposed to meet the requirement of different industrial applications. The main contributions of our work are as follows:(1) A RKL based adaptive system identification algorithm is proposed to identify a special nonlinear structural Hammerstein system:(a) By utilizing a procedure of data-preprocessing, the form of the offline identification method is modified. Thus, it is possible to solve the system identification problem in a sparse and recursive way. (b) Instead of calculating the inverse matrix directly, a twice sub-inverse technique is adopted to reduce the computation load.(c) A prediction error based strategy is employed for forward key node selection and adaptive switch between recursive updating and offline re-initialization, which helps the proposed algorithm be able to achieve a sparse way and overcome certain problems, e.g., error accumulation, poor stability. The proposed adaptive identification method is applied on a benchmark numerical example. The results show that the proposed method is more suitable for Hammerstein model online identification as it can improve the accuracy, stability and computation efficiency.(2) To deal with the modeling problem in batch processes, such as fed-batch fermentation processes with characteristics of batch-to-batch variation and few labeled samples within the same batch, two algorithms, "local weighted" kernel learning and semi-supervised "weighted kernel" learning, are presented to handle the batch-to-batch modeling and within-batch estimation problems.(a) Inspired by global and local learning, an adaptive local weighted kernel regression (ALW-KR) is proposed. Once a global model is available, ALW-KR adjusts the model with suitable weights according to the similarity between the query sample and key nodes in the training dataset. Compared to the local modeling method, only one matrix inverse calculation is needed for the proposed method, which makes the algorithm more efficient. Also, the proposed method can acquire a better accuracy than global modeling as a slight adjustment is utilized. ALW-KR is applied on a benchmark penicillin process "Pensim" to online model the batch-to-batch biomass concentration. The experimental results demonstrate the effectiveness of ALW-KR. (b) A semi-supervised "weighted kernel" learning algorithm is presented to handle the situation where there are few labeled nodes within one batch. By introducing a time-index based kernel function as a weight for traditional kernels, the proposed method is able to predict the output more accurate than traditional semi-supervised method. The simulation on penicillin concentration online prediction shows that the proposed semi-supervised approach can improve the prediction accuracy.(3) A multi-kernel LSSVR (MK-LSSVR) is proposed for multi-mode systems to establish a universal model. A new multi-kernel is designed by combining samples' mode and secondary variable information, which is also used to estimate the node's prior mode probability. By employing the procedures including posterior probability re-estimation, online sparsification and updating, the proposed approach is able to guarantee its sparsification, accuracy and generalization ability. Simulations on the penicillin processes indicate that the proposed MK-LSSVR is able to predict the biomass concentration more accurately and stabile, compared to traditional single-kernel based kernel learning methods.4) A new adaptive local kernel learning (ALKL) approach is proposed. Both the input/output information are adopted to construct the similarity criterion, and a new supervised locality preserving projection technique (SLPP) for regression problems is proposed to obtain a more comprehensive relevant sample set. Meanwhile, an adaptive weighted LSSVR (AW-LSSVR) is presented to establish the local model with a low-rank recursive updating trick to optimize the weight of relevant samples. The experimental results on the gas dry point prediction in fluidized catalytic cracking unit (FCCU) show that ALKL is more accurate and suitable, compared to traditional global and local learning methods.Finally, conclusions and further studies are given.