Study On Signal Reconstruction Method Of Multifunctional Sensor And Its Application

With the development of technology and the growing demand for measurement, microsensor, smart sensor, network sensor, wireless sensor and multifunctional sensor have become the main evolution directions of Sensing Technology. Multifunctional sensor can simultaneously detect several different signals; accurately and objectively assess measured object and environment and greatly reduce the size and consumption of the measurement system, therefore it have promising prospects in many fields, such as aeronautics, astronautics, environmental perception and industry measurement.The accurate measurement of multifunctional sensor depends on the reasonable signal reconstruction method. However, the uncertainty of measurement conditions, increase of measurand dimensions and the complex sensitive mechanism of multifunctional sensor usually cause many unfavorable factors for existing signal processing method such as poor model Predictive ability, low signal prediction accuracy, over-fitting, local minimum and curse of dimensionality. Therefore, this thesis studies several practical multifunctional signal reconstruction methods to overcome the existing related problem, improve the signal processing accuracy and promote the development and utilization of multifunctional sensor. The main contents of this dissertation are as follows:Multiple regression models with linearization parameters are often used in multi-functional sensor signal reconstruction, which has the quality of identification model simplification and easy to perform the signal reconstruction process. However, the least squares parameter estimation method only consider the error of the observe vector, unfortunately, in practice, the data matrix and observe vector are consist of measurement signals, which are inevitably contaminated by noise. Therefore, the total least squares method is much more reasonable in the case of noise data, which considers the bias of both data matrix and observe vector. The experimental results show that the total least squares have higher precision than the least squares in the multifunctional sensor signal reconstruction.Support vector machines are the new machine learning algorithm that is suited for small sample size problem. It use the structural risk minimization principle to replace the empirical risk minimization principle and can avoid the disadvantages of over-fitting problem, local minimum problem and curse of dimensionality in traditional learning methods. The least squares support vector machines involve equality constraints instead of inequality constraints and adopts a least squares cost function. Therefore it expresses the training by solving a set of linear equations instead of a quadratic programming problem in standard support vector machines, which greatly reduces computational cost and suit for the multifunctional sensor signal reconstruction. However, the set of linear equations consist of measurement signals, which are rarely known exactly. Thus the set of linear equations is often nearly ill-conditioned in practice. Though the inverse of a badly conditioned matrix is possible, the solution often exhibits numerical instabilities. Therefore, the total least squares method was used to solve the linear equations problem, which can restrain the matrix perturbation and improve the stabilities of solution. The experimental results suggest that the proposed solution is suitable for the multifunctional sensor signal reconstruction and can achieve better generalization performance and stability of signal reconstruction than the standard least squares support vector machines and support vector machines.The weight function used in the standard robust least squares support vector machines lack of elimination processes for outlier data, therefore the outlier data still influence the final solution process and reduce the robustness of algorithm. Therefore, a improve method was discussed in this paper, which used IGGIII weight function to evaluate the regression error and total least squares method to calculate the parameters of regression model. IGGIII weight function can reduce the effects of outlier data and enhance the robustness of reconstruction model by assigning zero weight to outliers, while the total least squares method can improve the stability of solution and calculation accuracy of prediction model parameters. The experiment results suggest that the improved method can obtain more accuracy and robust prediction than standard robust least squares support vector machines for different volume of outliers in training set of multifunctional sensor.The disadvantages of global model for a complex reconstruction problem are poor prediction efficiency, low generalization ability and high identification model complexity. The local model method can improve the generalization ability and prediction efficiency and reduce the complexity of identification model, which build the reconstruction model only in a local domain. In this paper, the local model method for the signal reconstruction of multifunctional sensor is discussed based on the vicinal risk minimization principle and least squares support vector machines, and the solution of the local model is calculated by total least squares algorithm. The experiment results suggest that local least squares support vector machines method can get more accurate and stability estimation than the global least squares support vector machines.To verify the feasibility of the proposed methods in practical application, it has been used to implement the signal reconstruction of a novel design multifunctional multicomponent concentration sensor. The relationship of input signals and output signals can be identified based on calibration sample set and the least squares support vector machines with total least squares solution. Therefore we can get the related concentration of sodium chloride and sucrose from different groups of the multifunctional sensor output signals. The experimental results prove that the proposed algorithms are suitable for practical application.