Research On Data-driven Modeling Method For Complex Industrial Soft Sensor

Complex industrial processes have increasingly higher requirements for control systems,and it is the general trend to build a data-based,safe,efficient,green,and low-carbon intelligent manufacturing system.This background promotes the development of soft sensor technology.The data-driven soft sensor modeling method has become an effective means to solve difficult measurement problems of key variables in complex industrial processes.It has also become the focus of intelligent manufacturing and complex industrial control research.This work,combined with a data-driven intelligent learning algorithm,deeply studies the nonlinear and noise problems of process data in complex industrial processes.The main research work is as follows:(1)According to the problem of nonlinear data and insufficient feature extraction from a single soft sensor model,the method based on SCAE-ACNN was proposed.Two homogeneous base learners are constructed using the same data to extract different features from two aspects.The first homogenous base learner uses a deep stacked and pre-trained convolution autoencoder to extract high-valued features from the original auxiliary variables.The second homogeneous base learner uses convolutional neural networks of channel and spatial attention for feature extraction.The BP algorithm trains the parameter learning of the two base learners.Convolutional autoencoders have a good low-dimensional representation of nonlinear and high-dimensional data of industrial processes.At the same time,the attentional mechanism can suppress irrelevant features.Experimental results show that the proposed soft sensor model based on ensemble learning SCAE-ACNN is superior to traditional modeling methods in terms of prediction performance.(2)A soft sensor modeling method based on the soft threshold function is proposed due to the problem of process data inclusion noise in complex industrial processes.The data is filtered by the soft threshold automatically fitted by the neural network,and the soft threshold autoencoder model is constructed.The weight factor is obtained through the nonlinear mapping of perception.Then the average value of input data is multiplied by the corresponding weight factor to obtain the appropriate threshold.In addition,a numerical simulation experiment is set up to prove this method’s effectiveness.After a batch of noisy data verification,the method has a good effect in denoising.The experimental results show that the prediction performance of the regressor can be improved by denoising the data.(3)According to the problem that the soft threshold function will turn part of the original data into "0" and cause data information to be lost,a soft sensor model of multi-attention fusion is proposed.This model is based on the residual network.The main road cleans the data in each residual module through the soft threshold function.The branch road extracts high-valued features from the original data by constructing an attention fusion module.Then,it converges the high-valued features with the cleaned data and transmits it to the next residual module.The network uses the Bayesian strategy to find the optimal hyperparameter.The experimental results show that the proposed method performs well when the data noise is included,especially for predicting maximum and minimum values in the data.