Time Dimension Expanded Local Weighted ELM On Industrial Soft Sensor Modelling

In the process of modern industrial activities,in order to ensure the safety of production processes and the quality of products,it is very important to have direct access to key process parameters.However,due to technical and cost constraints,it is difficult to directly measure critical quality variables in complex industrial processes.In many cases it is inevitable to use expensive measuring instruments or to cause some interference to the production process.To indirectly measure the quality variables,Soft Sensor Modelling method establish a mathematical model from some easily measurable auxiliary variables to difficult-to-measure quality variables.In the development of Soft Sensor Modeling algorithm,the traditional multivariate statistical regression algorithms are simple and fast,and the neural network algorithms and the depth learning have high precision to adapt to the nonlinear process,but the calculation speed is often slow.This paper focuses on Extreme Learning Machine(ELM)which is a single hidden layer neural network algorithm.This algorithm does not require repeated iterations of neuron parameters.It is fast,accurate and does not fall into the local optimal solution having great research potential.This paper based on the ELM algorithm take the industrial process soft sensor modeling as the application background.The main research results are summarized as follows:1)Aiming at the nonlinear problem of industrial process,a Soft Sensor Modelling method based on Local-weighted Extreme Learning Machine(LWELM)is proposed.The local weighted method is used to establish the local model for each test sample point.Compared with the global model,the process complexity is reduced,which help improving the generalization performance of ELM in complex industrial nonlinear process.Simulation results show that this improves Soft Sensor Modelling prediction precision effectively.2)Aiming at the dynamic problem of industrial process,a Soft Sensor Modelling method based on Time-dimension-expanded Extreme Learning Machine(TELM)is proposed.This method merges the sample points before and after the current time into new sample points,and obtains an expansion matrix with high dimension and high linear correlation containing the process dynamic information.The Extreme Learning Machine can overcome the high dimension and linear correlation problems.And use the increased data to improve the insufficient generalization performance of ELM when apply on limited samples of industrial process.Simulation results show that this improves Soft Sensor Modelling prediction precision effectively.3)Aiming at the problem of dynamic,nonlinear and noise error of industrial process,a Soft Sensor Modelling method based on Time-dimension-expanded Local-weighted Extreme Learning Machine(TLWELM)is proposed.This method considers the dynamic correlation between the samples close in the time dimension,and the local model similarity between the neighboring sample points in the spatial dimension improving ELM with both methods.The Local-weighted method eliminates the outliers and singular values reducing the influence of noise and errors.At the same time,it improves the generalization performance of ELM and reduces the expansion dimensions improving the generalization performance of ELM.As the expansion matrix shrinks,the noise and error interference is further reduced.Simulation results show that this improves Soft Sensor Modelling prediction precision effectively.