Process Support Vector Machine And Its Application To Satellite Thermal Equilibrium Temperature Prediction

The change of the thermal equilibrium temperature in the satellite thermal equilibrium test is actually a process of the accumulation with the time. Process neural network (PNN) can well embody this time accumulation effect, but its generalization ability is poor. Traditional support vector machine (SVM) has a good generalization ability. However, it is hard for the SVM to embody the time accumulation in the real satellite thermal equilibrium test effectively. According to the actual demand of the satellite thermal equilibrium test, a novel learning machine named process support vector machine (PSVM) is proposed in this dissertation based on the analysis of the advantages and disadvantages of the PNN and the SVM. The mechanism of the PSVM is studied. Finally, the proposed PSVM is utilized to predict the thermal equilibrium temperature of some satellite.Firstly, the generalization ability of the PNN is studied in this dissertation, and some methods are developed to improve the generalization ability of the PNN. Based on the study of the support vector machines for regression, a PSVM model is proposed. The inputs of the PSVM are time-varying functions. The inner product of the kernel function of the PSVM is a product of two time-varying functions. To simplify the computation complexity, the inner product can be resolved by an integral operator based on the expansion of the orthogonal basis functions. Thus, the PSVM can be used to simulate various time-varying systems in the real world. An experiment is presented to prove the performance of the PSVM, and the results indicate that the PSVM can well embody the time accumulation effect with a higher prediction accuracy.The kernel function is the core of the PSVM. It can transform the inputs of the PSVM into a linear hyperplane by mapping the inputs to a high dimension feature space. The properties of the PSVM kernel function and its construction methods are studied in this dissertation, and the concrete formulations of the different kernel functions and its selection methods are given under the condition of the time-varying inputs. The generalization ability of the PSVM is influenced by the kernel parameter, the regularization parameter and the insensitive coefficient. In order to find a PSVM with good generalization ability, the particle swarm optimization (PSO) algorithm is proposed to optimize the above parameters. Meanwhile, the nonlinear decreasing strategy of the maximum velocity is proposed to improve the precision of the parameter optimization.The performance and the life of the satellite are mainly influenced by the space environment and the corresponding accumulation effects. In order to guarantee the reliability of the operation of the satellite, thermal equilibrium test should be implemented on ground before the launching of the satellite. In order to shorten the test time and cut down the cost of the satellite development, the unstable thermal equilibrium test method should be energetically developed, that is to say, the thermal equilibrium temperature prediction should be energetically studied. Based on the theory research in this dissertation, a thermal equilibrium temperature prediction method by the PSVM is proposed, and the results indicate that the PSVM has a higher prediction accuracy than the traditional SVM, because that the PSVM can better embody the time accumulation in the change of the thermal equilibrium temperature. The results also indicate that the PSVM has a better generalization ability than the PNN. Thus, the PSVM is more suitable for practical engineering problems. Therefore, the study in this dissertation has some theoretical meaning and higher practical value.