Study On Technique And Application Of Soft Sensor Based On The Neural Networks Optimizing Theory

The thesis is origin from the national natural science funds project "flow soft-measure model and virtual dynamic flowmeter based on flow state complexity measure" project code 60374042.With the fast development of the hydro-transfer and hydro-control, the measurement of flux becomes more and more important. A method of soft sensor is presented to settle the problem.This thesis focuses on the soft sensor modeling based on NN, which solves the problem quite good.First, on the basis of analyzing the reported progress of soft sensor technology and the main trend of relative software development, the modeling methods problems in industrial application of soft sensor technology and developing tendency are discussed. This thesis described the theory of NN(Neural Networks), especially, it analyzed BP algorithm and it's various improved algorithms. As proved, all improved BP algorithm cannot avoid the flaw of "stuck in local minimum". A method of optimize BP with GA(Genetic Algorithm)was proposed.Secondly, the traditional GA algorithm was improved on to overcome its poor local search limitation. A hybrid niche selection operator was proposed. Mutation and crossover operators were all improved on.Then an applied coding scheme of genetic algorithm is proposed to optimize the BPNN topology structures. A new fitness function was established on principle and the algorithm procedure was designed. The simulation results indicate that these optimizing methods are effective.Finally, the whole system for soft sensor of dynamic fluid was designed and realized including variable data collection data disposal and the core of the soft sensor technology-soft sensor modeling based on the optimized BPNN.The soft sensor method proposed in this paper affords a new idea todynamic flow's measurement. The proposed optimize algorithm have great potential in other modeling field.