Local Spatiotemporal Feature Representation And Soft Sensors For Industrial Data Based On Convolutional Neural Network

In the modern industrial production,the implementation of intelligent technology,such as process control,optimization and decisionmaking,largely depends on the timely measurement and feedback of key product quality indicators.However,the key product quality indicators cannot be measured online due to the limitations of the harsh environment and measurement technology.Recently,the data-driven soft sensor technology has become the main approach to solve this problem.Because of the cascade characteristics between production units,there are complex local spatiotemporal patterns between process data and quality indicators,which bring a series of problems to the precision soft sensor modeling of industrial process.The convolutional neural network has the natural advantages in local patterns extraction and provides an important modeling method for spatiotemporal features extraction of soft sensors.Hence,based on the convolutional neural network,this paper studies how to extract effective local spatiotemporal patterns from the process data and conduct soft sensor models.The main innovative researches of this paper are as follows:(1)To solve the problem that it is difficult to extract the local spatiotemporal patterns between the variables in the cross-neighborhood structures,a soft sensor modeling method based on the Multi-Channel Convolutional Neural Network(MCNN)is proposed in this paper.The input of MCNN is a 3D multi-channel tensor for each sample that is generated from the 2D dynamic matrix under the natural physical topology by variable reordering.Each channel of the 3D multi-channel tensor corresponds to the spatiotemporal of different variables combinations in the cross-neighborhood topology.Then,the convolution kernel slides on different channels to extract the cross-neighborhood topology local spatiotemporal patterns for soft sensor quality prediction.(2)To solve the problem of local spatiotemporal features of the process data at multiple topological scale,this paper proposes a soft sensor modeling method based on Multi-Scale Attentional Convolutional Neural Network(MSA-CNN).MSA-CNN firstly adopts multi-scale convolution kernels to focus on local spatiotemporal behaviors of different scales,and obtains multi-scale local spatiotemporal features maps.Furthermore,the attention module of feature maps is designed to obtain the attention weight of channel level to distinguish the significance of local behaviors at different topological scales.The multi-scale local spatiotemporal feature maps are weighted by attention for subsequent high-order feature representation and quality prediction.(3)A soft sensor modeling method based on Autoregressive Temporal Convolutional Network(AR-TCN)is proposed to extract the quality-relevant local spatiotemporal feature between process data and quality indicators.The input layer of AR-TCN includes both dynamic process variables and label variable,which fully considers the collection between process variables and quality labels.Moreover,the long-term spatiotemporal trend of quality label is described effectively through the causal dilated convolution to learn the quality-relevant local spatiotemporal feature,which is benefit for improving the predictive performance of the soft sensor model.Figure 32,Table 16,Reference 92...