Application Of Convolutional Neural Network In Soft Sensor

Lack of real-time measurement of some key product quality is one of the main problems in modern industry.It is possible to produce substandard products.In order to solve this problem,various soft sensor modeling methods have been proposed.Soft sensor is a technique that uses measurable process variables to infer variables that cannot or are difficult to measure online.Convolutional neural network is an important algorithm in deep learning.It can extract features directly from the original information,and its weight sharing characteristics reduce the complexity of the network.This paper uses a convolutional neural network and an existing shallow algorithm to create a dynamic soft sensor model with higher precision and deeper levels.The paper mainly studies the following aspects:(1)Aiming at the soft-sensor modeling of PTA average particle size in PTA refining process,a dynamic soft-measurement model of convolutional neural network and XGBoost hybrid modeling is proposed.The convolutional neural network extracts dynamic features,and the XGBoost algorithm extracts The feature is fitted.The simulation results of the actual PTA average particle size data show that the algorithm has higher prediction accuracy.(2)For the soft-sensor model of 4-CBA content in PX oxidation process,a convolutional neural network using global maximum pooling and linear combination output is proposed,which reduces the complexity of the network and proves based on the convolution through experimental simulation.The dynamic soft-sensor model established by neural networks has good performance(3)Each input sample of the data collected by the PX oxidation process is resampled to generate multiple sets of data,thereby establishing a convolutional neural network with multiple input channels.The content of 4-CBA is predicted in this paper.The performance of the dynamic softmeasurement model of multi-channel convolutional neural networks is analyzed.The simulation results show that the model has better performance than the single-channel convolutional neural network.