Dynamic Concentration Estimation of the Ethyl Acetate Produced in Continuous Stirred Tank Reactors Integrated with Adaptive Feature Selection Using Iterative Weighted Support Vector Regression Optimized by Particle Swarm Optimisation

Support vector regression (SVR) is a power ful algorithm for regression and design of soft sensor. The aim ofthis work is a novel soft sensor based on SVR which integretaed with feature selection by adaptivly tuning theinfluencial of each input variables on SVR output using particle swarm optimization (PSO). In this method, during theiterative precedure, the influence of each input variables be evaluated and adaptivley set the more effective inputvariables by concidering their costs. Also the SVR parameters are simuletanusly optimized using PSO. Since the moreeffective input variables needs to be selected among the variables includes of the pressure of CSTR, feed stream flowrate, and temperature of CSTR in current and past times, then in order to have more reliable optimisation procedure,two different coefficient of inertia weight considered which leads to more robust balance between exploration andexploitation. The results indicates that the proposed soft sensor based on this method which called PSO-IWSVR have agood performance in order to construction the more reliable and more robust soft sensor since the effective inputvariables were adaptively selected.

Keywords:

soft sensor , SVR , PSO , feature selection , CSTR