Study Of Temperature Compensation Methods For MEMS Flow Sensors

Flow sensors based on micro electro mechanical system(MEMS)technology play an important role in automotive electronics,industrial gas control,biological medicine,semiconductor manufacturing,scientific research experiments and so on by virtue of their excellent performance in integration,detection accuracy and sensitivity.Although the differential design is adopted in the structure,it is inevitable that the measurement is affected by the ambient temperature,which reduces the measurement accuracy,resulting in large error of the measurement results and even wrong results.In this paper,a MEMS flow sensor based on thermopile is taken as the research object.Firstly,the influence of ambient temperature changes on it is studied through theoretical formula derivation and multi-physical field finite element simulation.Then,a temperature experimental test platform is designed and built to obtain the relationship between flow fate and output voltage of the MEMS flow sensor at different temperatures.Finally,aiming at the temperature drift problem of the MEMS flow sensor,the following temperature compensation methods are proposed,and the effectiveness of them is verified by the sample data obtained from the temperature experiment:1.A least squares polynomial fitting(LSPF)temperature compensation method is proposed.Flow rate as the function of ambient temperature and output voltage of the flow sensor is obtained by twice polynomial fittings,and the polynomial coefficients are obtained by the ordinary least square method.Inputting test sample data to verify the effect of the temperature compensation model,and the maximum relative error after compensation is 1.65%.2.A Levenberg-Marquard-back propagation neural network(LM-BPNN)temperature compensation method and a genetic algorithm-Levenberg-Marquard-back propagation neural network(GA-LM-BPNN)temperature compensation method are proposed.GA has the advantages of global and parallel,which makes it can be used to optimize the initial weights and biases of BPNN.On this basis,with LM algorithm as train function,GA-LM-BPNN can avoid the problems of slow convergence speed and easy to fall into local minimum during training of BPNN.Compared with LM-BPNN,which is without optimized by genetic algorithm,GA-LM-BPNN has faster convergence speed and higher compensation accuracy.Inputting test sample data to LM-BPNN and GA-LM-BPNN,the maximum relative errors after compensation are 0.4872%、0.3211%respectively.3.A radial basis function neural network(RBFNN)temperature compensation method is proposed.An exact RBFNN is established,which does not need network training and learning,making the algorithm more efficient.The maximum relative error after compensation is 0.4149%.The experimental results above show that the proposed four temperature compensation methods can effectively compensate the temperature drift of the MEMS flow sensor,so as to improve the measurement accuracy and stability.