|
The traditional system identification methods are mostly based on the basis of the model structure known , and require too much of a priori knowledge . Exist in reality a large number of non-linear time-varying systems, the lack of prior knowledge , the model structure is difficult to determine , to the identification work has brought great difficulties . In this paper, difficult to exact description of nonlinear time- varying systems , the use of black-box characteristics of the neural network and the ability of the non-linear model , a more in-depth analysis and discussion of the neural network identification method to arbitrary precision approximation . Article first introduced the development and research status quo of the artificial neural network system identification , analysis commonly used BP network , RBF network and GRNN network learning methods and the simulation examples . More in-depth analysis on the basis of the comparison of simulation results , focusing on the GRNN network , summed up the web presence of two major problems : First, the the GRNN network mode layer node number of training samples is proportional to the number of training samples increase in volume , the pattern layer node corresponding increase ; smoothing factor values ??have a major impact on network performance , smoothing factor if we take a single value , the calculation is easy and accurate insufficient , for different values ??of , the result is more accurate , but the computational complexity high . Then , based on the analysis of the causes of the above problems , and took two corresponding solutions . Response to the first question , the FCM clustering method for processing , reduction of network structure . And the similarity index of the input data is defined by the specifications given threshold comparison to decide whether to carry out the iterative clustering operation, in order to solve the iterative operation is cumbersome , inefficient operation . For the second problem , the size of the contribution rate of each feature vector based on the network output , to select a corresponding smoothing factor IMPROVED method . Finally summarized earlier , summed up the the GRNN network based on improved identification strategy , and in the actual gas hydrate resistivity measurement system analysis . Experimental results show that the network identification model processing speed, high recognition accuracy and generalization ability , has good practical value .
|