Research On Soft Sensor Modeling Method Based On Spatial-temporal Feature Fusion Deep Neural Network

Key quality variables in large-scale production processes such as metallurgy and chemical industry are closely related to product quality and production safety.Because of the limitations of measurement technology and the high cost of developing corresponding instruments,many key quality variables are difficult to measure directly.As a method to predict key quality variables,soft sensor modeling realizes the estimation of key quality variables by establishing the mathematical relationship between related variables and target variables.In recent years,a variety of modeling methods have been applied,but most of these methods rely on a single feature in time or space,and are difficult to apply to production processes with complex working conditions.At the same time,there is a relatively lack of soft-sensor modeling research on spatio-temporal multi-dimensional feature fusion and multi-model ensemble learning.Based on the above considerations,this paper takes the deep neural network as the core technology,develops a soft sensor model integrating time and space features,and applies it to the prediction of key quality variables in the actual industrial production process to verify its feasibility and effectiveness.The specific research work and research results of the paper are as follows:(1)A twin-stream deep neural network soft-sensor model that fuses spatiotemporal features of data is proposed.Among them,the historical value of the quality variable is used as the input of the gated recurrent neural network to extract temporal features,and the variable data related to the quality variable is used as the input of the convolutional neural network to extract the spatial features.Fusion is performed to achieve accurate estimates of quality variables.Compared with traditional methods,the method based on spatial-temporal feature fusion of deep neural network proposed in this paper can take the features of data space dimension and time dimension into consideration.This method can better realize real-time estimation of process quality variables.During the training process,the feature fusion layer of the neural network is defined,and the process of multi-feature fusion is optimized.The application results in the prediction case of silicon content in blast furnace hot metal show that the prediction accuracy of this model is significantly improved compared with other methods.(2)Based on the twin-stream spatial-temporal fusion deep neural network model,a method based on iterative deep aggregation(IDA)is proposed.This method also uses gated neural network and convolutional neural network to consider the features of time and space dimensions,but different from the twin-stream spatialtemporal fusion deep neural network model,the method based on iterative deep aggregation draws on the idea of residual neural network.The nodes aggregates the spatiotemporal features layer by layer,and iteratively transmits the output of the nodes to the output layer to realize the effective transfer of information from shallow to deep layers.The neural network parameters are further updated and optimized to improve the model prediction accuracy.Under the same conditions,the water wall pipe temperature of thermal power plant is predicted,and the results show that the model has better generalization ability and higher prediction accuracy.The twin-stream spatial-temporal fusion deep neural network model and the IDA-based model proposed in this paper can effectively solve the problem of industrial soft measurement modeling difficulties of complex processes with multiple variables.