Industrial Process Soft Sensor Based On Deep Learning

In order to reduce production costs,monitor production processes and optimize product quality,real-time measurement and accurate prediction of process primary variables are very important.The soft sensor indirectly estimates the primary variable by constructing a mathematical model with the secondary varnable as the input and primary variable as the out:put.It has been developed rapidly and practiced effectively in academia and industry because of its advantages.In this paper,research works are carried out based on the deep model algorithm for complex industrial process soft sensor.Our main research works are described as follows:(1)A stacking recurrent neural network soft sensor method for nonlinear dynamic process is proposed.Stacking with RNN,LSTM and GRU,it can effectively relieve over-fitting while improving prediction effect.The prediction feasibility of sRNN is verified by the actual data simulation of the debutanizer column process.Followed by the effectiveness comparison analysis of nonlinear dynamic processes.(2)A deep forest regression soft sensor method for nonlinear multimode process is proposed.Combining the random forest regression and deep forest structure according to ensemble learning thought,it can adaptively adjust the model structure and enhance the representation learning ability and application scope of the data model.The feasibility of DFR for nonlinear multimode industrial processes soft sensor is verified by the actual data simulation of the primary reformer process.The performance is also compared with other regression methods.(3)A multi-grain cascade recurrent neural network soft sensor method for time-variant process is proposed.Based on the framework of deep forest regression,the long short time memory and gated recurrent neural network are used as the individual learner respectively.It can solve the nonlinear,time-varying,multi-mode,dynamic and other complex characteristics of process data through multi-grained moving window scanning and cascade structure.The method was applied to the debutanizer process and primary reformer process for data simulation to test the accuracy and superiority of this soft sensor method for various complex industrial processes with multiple process characteristics.