论文部分内容阅读
介绍了小波神经网络对于一维数据进行有损压缩和特征提取的数学基础,重点讨论了BP小波神经网络收敛速度的改进方法.新的改进措施包括:1.优化选取初始网络参数和信号标度变换.2.适当改进常用的自适应调节学习率方法.3.利用差值逐步逼近技术对原始数据进行多次压缩.为了尽可能提高压缩比,在保证相对误差不变的情况下,还讨论了变结构的小波神经网络,能够自动删去贡献很小的隐层神经元.仿真计算表明,在满足相对误差要求的情况下,这些新的改进方法能够获得较快的收敛速度和较高的压缩比.
This paper introduces the mathematical foundation of wavelet neural network for lossy compression and feature extraction of one-dimensional data, and focuses on the improvement of convergence speed of BP wavelet neural network.The new improvement measures include: 1. Optimize the selection of initial network parameters and signal scaling Transform.2.Appropriate to improve the commonly used method of adaptive adjustment of learning rate.3.The use of difference stepwise approximation of the original data for multiple compression.In order to improve the compression ratio as much as possible, to ensure that the relative error remains unchanged, also discussed The wavelet neural network with variable structure can automatically delete hidden neurons with small contribution.The simulation results show that these new improved methods can obtain faster convergence rate and higher Compression ratio.