Development Of Soft Sensor Using Local Modeling Method

Concomitant with developments in production technology and complexities in technological process,the process dataset is characterized with time-variant,non-Gaussianity,multi-stage,and uncertainties.Knowledge-based modeling has become more and more difficult.In this case,the local model through observations resolves uncertainty and the complexity of the system,achieving the real-time key variables estimation.Under this framework,many mathematical tools such as machine learning,data mining,and statistical techniques are used.The approach simplifies the complex system and benefits from the low development cost,the high real-time performance,and the facilitation of maintenance.The major contribution of this study is to combine local information with modeling algorithms to address soft sensor methods of complex systems.The main works are as follows:With insufficient modeling data in nonlinear system,a hyperparameter-varying recursive Gaussian process regression model is presented to calibrate the online model.The model calculates the hyperparameter of the accumulated data in the functional space through the reset model parameters of the conjugate gradient algorithm.Then,the recursive Gaussian process regression modeling under the moving window mechanism is implemented.Applications on a numerical case and a penicillin fermentation process evaluate the performance of proposed method.With non-Gaussianity of plant-wide processes,an adaptive multi-state partial least squares regression algorithm using local neighborhood standardization is proposed.The algorithm uses the maximum difference information of auxiliary information to locally normalize the original data,which makes data follow one single distribution.Sample-wise updating and block-wise updating strategies improve adaptive qualities by using partial least squares method.Applications on CSTR and IPB concentration prediction evaluate the performance of proposed method.To the overlapped clustering problem of different subset distributions in batch processes,a recurrent neural network using steady state identification(SSID RNN)algorithm and a temporal feature embedded Gaussian mixture model(TFGMM)are presented.In SSID RNN,the steady state describes different stages,which is identified through variances.The corresponding local model is contented by the memory unit in recurrent neural network.In TFGMM,the posterior probability of overlapped samples in Gaussian mixture model is modified by referring temporal features to improve the flexibility of phase division and the estimation of unknown parameters.Applications on a numerical case and a penicillin fermentation process evaluate the performance of the proposed method.For dealing with the problem of disturbance caused by input random noise,a modified subspace identification method with limited iterative expectation-maximization algorithm is proposed.The data-driven model is constructed with an unknown state-space model by using the locally optimized estimate in Kalman filter,which reduces the influence form input noise.The biased estimation obtained by subspace identification under least squares framework is calibrated by using the expectation-maximization algorithm.And a limited iterative criterion is given for improving predictive performance.Applications on a numerical simulation and the TE process evaluate performance of the proposed method.