A new neural network soft sensing method based on the multi-resolution data fusion

Abstract

For the sampling frequency is different between the quality indexes and the auxiliary variables in the soft sensing training patterns, a new neural network soft sensing method based on the multi-resolution data fusion is presented in the paper. In the presented method, the neural network is trained based on the quality indexes and auxiliary variables with low sampling frequency first and the high resolution quality indexes is estimated using the trained network. Then the high resolution quality indexes are multi-scale decomposed using the biorthogonal wavelets and are fused with the low resolution quality indexes. Then the fused high resolution quality indexes are got and are used for the next training of the neural network to improve the accuracy of the soft sensing model. The method is used in the soft sensing modelling for the dry point of the crude gasoline of the Fluidized Catalytic Cracking Unit (FCCU) fractionator, results show that the neural network soft sensing model has higher accuracy after several iterative fusion and training steps.