Abstract: The paper presents a new method to design a predictive filter which suits to analyze non-stationary data.The processed data are approximated by cubic spline function.The predictive filter is constructed with the true value and the observation value of each observation point.The designed filter can perform adaptive analysis to the observation data，and is able to implement multi-resolution analysis as wavelet transform.The paper discusses the derivation of the filter in detail and analyzes its amplitude-frequency characteristics，and then presents the decomposition and composition structure when using the filter to analyze data.Simulations are performed and the experimental results show the entropy of the coefficients produced by the discussed algorithm is less than that by CDF5/3，so it is an effective analysis method for nonstationary data.

摘要： 从三次样条函数入手，提出一种新的适合非平稳数据分析的预测滤波器设计方法。利用三次样条函数对处理后的观测数据进行多项式逼近，由各观测点的真实值和观测值构造出预测滤波器。该滤波器能对观测数据进行自适应分析，且与小波变换一样能实现多分辨率分析。详细讨论了该滤波器的推导和幅频特性分析，并给出了滤波器在分析数据时的分解和重建结构。对文中的分解算法进行了仿真，结果表明该算法所产生的系数的熵小于JPEG2000中CDF5/3小波滤波器产生的系数的熵，是一种有效的非平稳数据有效分析。