Research On Compensation Technology Of Pressure Sensor Based On Machine Learning And Intelligent Optimization Algorithm

As Micro Electro Mechanical System takes many advantages as low power consummation,high sensitivity,small volume,standardized manufacturing process and high cost performance ratio,piezo-resistive pressure sensor based on which has been extensively applied to the pressure measurement in almost all the industrial areas such as automobile,aviation,petrochemical and consumer electronics,etc.The requirement of the synthetic performance pressure sensor in industrial process increases rapidly with the improvement of the social industrialization level.However,the environmental temperature and the static pressure effect are two key factors which may bring some unsatisfied bias to measurement characteristics of high precision pressure sensor,which have become the technical bottleneck to provide more measurement accuracy.To deal with these two issues,the research work in this paper is based on piezo-resistive effect and MEMS craft.The absolute pressure sensor and the relative pressure sensor are taken as the research objects.Some methods including Machine learning models as support vector regression(SVR),least squares support vector machine(LSSVM),kernel extreme learning machine(KELM)and intelligent optimization algorithms as adaptive mutation particle swarm optimization(AMPSO),Chaotic ions motion algorithm,coupled simulated annealing with simplex(CSA-Simplex)are combined to solve critical problems as nonlinear output characteristic caused by environmental temperature variation and static pressure output error derived from static pressure.The main research works are summarized as follows:(1)By analyzing the reason for temperature drift,it indicates that the nonlinear output of the pressure sensor is not only rely on the response relationship of piezo-resistive coefficient of the semiconductor and the environmental temperature variation but also influenced by some other unsatisfied factors in the pressure sensor manufacturing process.Both hardware compensation methods and software compensation methods are reviewed to give the details of advantages and disadvantages of these existed temperature compensation schemes.(2)The improved AdaBoost.RT ensemble adaptive mutation particle swarm optimization optimized support vector regression(AMPSO-SVR)is proposed to compensate the temperature error of 250kPa piezo-resistive differential pressure sensor.Taking into consideration the nonlinear effect of the sensor output characteristic rise from the temperature variation,the temperature compensation problem could be deemed as a nonlinear function regression problem.The frame of structure risk minimum guarantees SVR good nonlinear approximation ability and generalization ability.Because the model parameters are important to the SVR the classical intelligent optimization algorithm-particle swarm optimization(PSO)is introduced in this research work to select appropriate model parameters.Nevertheless,classical PSO algorithm suffers from premature phenomena and relative limited searching ability,Levy flight is applied to adaptively adjust the particle swarm trajectory to avoid premature.To accelerate the convergence process of PSO,opposition-based initialization is used to uniformly distribute the swarm particles over the solution space to learn as much as possible about the solution space structure.A dynamic calibration experiment was designed and implemented to reduce the calibration time,several compensation models including Cuckoo Search Optimized Support Vector Regression(CS-SVR),Firefly Algorithm Optimized Support Vector Machine(FA-SVR),Shuffled Leap Frog Algorithm Optimized Support Vector Regression(SLFA-SVR),Particle Swarm Optimization Optimized Support Vector Regression(PSO-SVR)and Particle Swarm Optimization with Levy Flight Optimized Particle Swarm Optimization Optimized Support Vector Regression(Levy-PSO-SVR)are compared on the experiment data with Adaptive Mutation article Swarm Optimization Optimized Support Vector Regression(AMPSO-SVR),the results indicate the presented AMPSO is superior to other optimization scheme for SVR.Furthermore,the proposed improved AdaBoost.RT takes the AMPSO-SVR as base learning machine.The result derived from the comparison between back-propagation neural network,radial basis function neural network,AMPSO-SVR and the proposed compensation scheme shows that the improved AdaBoost.RT ensemble AMPSO-SVR can provide more satisfied compensation performance than other compensation methods and more feasible for engineering realization.(3)The chaotic ions motion algorithm(CIMA)optimized least squares support vector machine(LSSVM)is proposed to deal with the temperature compensation problem of 40kPa piezo-resistive differential pressure sensor.LSSVM converts the quadratic optimization to a linear equation system,in which the regularization theory is able to achieve the balance between modeling ability and model complexity.Constructing hybrid kernel function on the basis of analyzing the model approximation ability using different kernel functions.A new intelligent optimization method-ions motion algorithm(IMA)is introduced,moreover,the CIMA is presented to select appropriate hybrid kernel parameters.The chaotic mapping is embedded in the CIMA to promise the ergodicity and stochasticity of the searching process.Compared with SVM merely has RBF kernel(RBF-SVM),Improved Particle Swarm Optimization Optimized Support Vector Machine(IPSO-RBF-SVM),Improved Particle Swarm Optimization Optimized Least Square Support Vector Machine(IPSO-RBF-LSSVM),Improved Particle Swarm Optimization Optimized Least Square Support Vector Machine(IPSO-Hybrid-LSSVM),Ions Motion Algorithm Optimized Least Square Support Vector Machine(IMA-Hybrid-LSSVM)and Chaotic Ions Motion Algorithm Optimized Least Square Support Vector Machine(CIMA-Hybrid-LSSVM)on static experiment data from viewpoints of static partition and random partition.Three conclusion can be drawn from the simulation results:LSSVM framework is more efficient than SVM;hybrid kernel has better generalization ability than single RBF kernel;CIMA is more suitable than other referred optimization algorithms for searching global optima.(4)A new sparse scheme named quantum particle swarm sparse least squares support machine(QPSO-sparse-LSSVM)is presented to deal with relative large sample dataset.QPSO-sparse-LSSVM can not only obtain the optimal model parameters of the sparse LSSVM model with high probability but also avoid the early stop caused by the order stop criterion of the classical sparse LSSVM algorithm.Compared with the LSSVM without sparseness and classical sparse LSSVM on different dataset,the proposed sparse LSSVM shows the most satisfied performance for the temperature compensation problem of 1MPa absolute pressure sensor.(5)The coupled simulated annealing and simplex optimized kernel extreme learning machine(CSA-simplex-KELM)framework is proposed to solve the compensation problem result from synthetic influence of the static effect and environmental temperature.Extreme Learning Machine(ELM)is a prediction model with fast modeling speed and decent generalization ability which owes to the application of the kernel trick and regularization theory.The parameter optimization task is accomplished by a two steps searching strategy consists of coupled simulated annealing(CSA)and simplex.The CSA is employed to find a local optimal position followed by the simplex which would perform a more precise local search.Compared with Back Propagation Neural Network(BP),Radial Basis Function Neural Network(RBF),Particle Swarm Optimization Optimized Support Vector Machine(PSO-SVM),Particle Swarm Optimization Optimized Least Square Support Vector Machine(PSO-LSSVM)and Extreme Learning Machine(ELM),the proposed compensation strategy not only improves the compensation performance about the temperature compensation and synthetic compensation but also exhibits strong learning ability about the characteristic of pressure sensor.Theoretic analysis,intelligent optimization algorithm,machine learning theory and experiment research are combined in this research.Some key methods are developed along with related work to eliminate the temperature error and synthetic error of the MEMS piezo-resistive pressure sensor which have important theoretical significance and application value for the diversification of the research methods of the pressure measurement area and exploring the development direction of this discipline.