基于RBF神经网络的EMD方法在海平面分析中的应用

SEA LEVEL VARIATION ANALYSIS WITH RBF NEURAL NETWORK BASED EMD METHOD

采用径向基函数神经网络法延拓原始数据序列,有效抑制了EMD分解中出现的端点发散效应,从而实现准确的EMD分解。利用该方法对中国近海验潮站的月平均海平面资料进行处理,分解得到的内在模函数分量代表了海平面各种周期性变化。通过EMD分解得到的总体自适定趋势项为非线性变化,比以往趋势项提取方法更有优势,它反映了在资料长度内海平面的长期升降情况。数据序列越长,该方法所能分解出来的IMF成分越多,可分辨的频率越小。

A new empirical mode decomposition （EMD） method was developed based on radial basic function （RBF） neural network and applied to analyze sea level variability. A temporal signal series is extended with the network before a new round of EMD starts. The result of analysis with the new method agrees better with the result of 5 year running mean than that using traditional EMD. The new method decomposes a series of sea level temporal signal into a set of intrinsic mode functions representing the periodic components of signal and an overall adaptive trend. It was shown that the annual variability was very remarkable in 47 stations along coast of China, and the semi-annual variability was also clear but less remarkable, with the amplitude decreased north to south. In the last 40 years, the sea level fluctuation of the area has been increasing in a non-linear pattern generally but varied geographically. The maximum fluctuation occurs in estuary zones. The results also reveal that the longer the temporal signal series is, the more IMF can be obtained from the data using the new method and a more exact overall adaptive trend can be extracted from signal. A shorter time series could achieve a complete decomposition easily but some of the IMF components may be hidden in overall adaptive trend, resulting in obviously nonlinear sea level variability, which may not be actually true.