Research On Model And Information Transform Mechanism Of Process Support Vector Machines

Support vector machines (SVM) is a novel machine learning method developed on statistical learning theory; it has virtue of self-contained theory, global optimization, good generalization ability, and independence of algorithm complexity on feature space dimension etc and has well applied effect in pattern recognition and fitting of time-invariant system. Aiming at the inelasticity of existing SVM method for solving dynamic pattern classification and nonlinear regression and so on in time-varying system, a process support vector machines (PSVM) model with time-varying functions as its inputs was proposed in the paper. PSVM with procedural inputs is a more generalized support vector machine model, it broadens traditional SVM's synchronous instantaneous restrict to inputs and widens application field of SVM, many practical applications can come down to this problem. Thus, research on topology structure, kernel function information transform mechanism and learning algorithm etc of PSVM has significant meaning in aspect of process signal pattern recognition, dynamic system modeling and function fitting and so on of time-varying system.The paper introduced theory background, basic thought and learning algorithm etc of traditional SVM synoptically, on this basis analyzed mapping mechanism and property of kernel function of PSVM, proposed PSVM based on orthogonal basis expansion and on kernel transformation and also studied on relative problems such as model and learning algorithm etc. PSVM based on orthogonal basis expansion converts the signal classification problem of time-varying function space into pattern recognition problem in high dimensional real-vector space via isomorphic principle of function space and high dimensional vector space and solves the pattern recognition problem of time-varying function effectively by dint of traditional SVM method. Because of adopting new process kernel functions, PSVM based on kernel transformation can availably deal with process signal and time-varying function of system input, and it is of equivalence to a three-layer feedforward process neural networks. Identification of oil-water bearing reservoir in oilfield development proves that PSVM based on kernel transformation has better dynamic pattern classification ability than PSVM based on orthogonal basis expansion. In the end, the paper extended PSVM used to settle pattern recognition assignment to regression analysis problem and put forward process support vector regression (PSVR) based onε-insensitive loss function, constructed parameter optimization method based on GASA mixed strategy and applied it to time series prediction. From the point of functional analysis, time series prediction problem is transformed to functional approximation, so PSVR is fitter for settling simulation, modeling and problem solving of complicated nonlinear system with time-varying factors.