110th Anniversary: Real-Time End Point Detection of Fluidized Bed Drying Process Based on a Switching Model of Near-Infrared Spectroscopy

For timely detection of the drying end point of a fluidized bed drying (FBD) process, a switching model based monitoring method is proposed based on in situ measurement of granule moisture content via near-infrared (NIR) spectroscopy. The least-squares support vector classification (LSSVC) method is adopted to build a global model for monitoring the initial underdrying phase with relatively higher granule moisture content. Subsequently, the instance based learning (IBL) strategy is used to select similar samples from historical batches for building up a local model to check on each query sample in the current process, in order to detect whether the real drying end point is reached. To solve the problem of selecting similar samples in high-dimensional NIR spectral space, the t-distributed stochastic neighbor embedding (t-SNE) strategy is introduced into the IBL model building method to ensure efficiency of dimension reduction. For online monitoring of an FBD process, a model switch strategy is proposed between the above established global model and local models, such that good prediction performance can be obtained with significantly reduced computational effort. Experimental results on the FBD process of silica gel granules demonstrate well the effectiveness and merit of the proposed method in comparison with the existing global model or local model building methods.