Application Of Generative Adversarial Networks In Soft Sensing Of Chemical Process

Soft sensor modeling methods of chemical processes have an important role in guaranteeing the continuous and smooth operation of production equipment,and product quality.It should be further explored to achieve the goal of intelligent production process.In general,the excellent performance of soft sensor modeling methods always depends on a large amount of labeled data.However,the collection of labeled data often causes losses in terms of time,hardware,and labor costs.To ensure that soft sensor method still has better performance when the labeled data is insufficient,it is of great significance to develop advanced soft sensor modeling methods for chemical processes using generative adversarial network.This paper first reviews the current research status of generating confrontation networks.Taking into account the characteristics of chemical process data,a novel soft sensor method using generative adversarial network based on Wasserstein distance is proposed.The main purpose is to improve the performance of the soft sensor model for the quality prediction of key variables in the chemical process based on reducing the cost of data collection.Specifically,the modeling method improves predictive performance by generating labeled data rather than collecting labeled data.Aiming at the problem of low quality generated data,a similarity measurement method is introduced into generative adversarial network,which can efficiently control the quality of generated data to optimize prediction performance of model.Moreover,there are not only process variable data but also image data in chemical processes.Unlike process variable data,image data acquisition has advantages of less hardware requirements and lower complexity.To overcome the modeling problem of image data,a conditional variational autoencoder generative adversarial network is proposed to better capture digital image features in chemical processes.The main work and contributions are as follows:(1)For the case that the traditional soft sensor model can't handle only a small amount of labeled data,the Wasserstein distance-based generation confrontation network is applied to the prediction of key variables in chemical process.Additionally,a similarity measurement-based generative adversarial network is proposed to better avoid the impact of low quality generated data.The experimental results demonstrate the effective of our proposed method,which not only alleviates the problem of long data acquisition cycle and high cost,but also improve prediction performance.(2)Compared with process variable data,digital image data has the characteristic of less hardware types,higher information dimensions,and direct observation.To overcome the difficulty of feature extraction of digital image data,the convolution neural network is introduced.Combining the advantages of the conditional generative adversarial network and the variational autoencoder,a conditional variational autoencoder generative adversarial network is proposed.This not only improves the ability of conditional generative adversarial network in the expression of diversity features,but also solve low efficiency problem of variational autoencoder generative adversarial network on multi-category problems.Consequently,the experimental results show that generative adversarial network can achieve more accurate performance,and successfully applied to digital image data in chemical processes.