利用主成分分析與多變量經驗模態針對身體運動數據的特徵萃取

Abstract

近年來，科技在生醫領域不斷有新展品及新技術的發展，無論是微型生物感測器或是搭配應用程式的智慧型手機，都已經可透過這些簡便的裝置來實現個人行動照護的理念，無須像從前須經由醫院裡昂貴的醫療監控儀器或系統獲得生理訊號。而生理訊號有數多種類，如心電訊號、腦電波訊號或運動訊號等，雖然這些非線性及非穩態訊號已經如此方便取得，那如何才能夠有效分析，而本論文將針對人體的運動訊號做詳細分析。 本篇論文提出主成分分析與多變量經驗模態分解的整合使用以及訊號能量選取規則，來幫助達到有效的訊號特徵萃取。透過主成分分析將訊號的重要資訊依序排列，再經由多變量經驗模態分解把訊號拆截成有不同物理意義的本質模態函數。分別算出本質模態函數的能量後，經由設計過的挑選規則，就能找出主要及次要的本質模態函數，就可進行不同的運算，而得到運動時所會產生的特徵值。

There have been more and more new development of products and technologies in the biomedical technology in recent years. For the sake of “personal mobile healthcare”, now it is easier to achieve with the simple mobile devices, such as miniature bio-sensors or smart phones with bio-application, and faster to catch the bio-signals without the expensive medical monitoring equipment or systems from the hospitals like before. And ECG signals, EEG signals and Motion signal are classified as the bio-signals, the same property among them is “non-linear” and “non-stationary”, so how could we efficiently analyze the signals getting so convenient? In this thesis, we choose the body motion data to do the analysis thoroughly. This thesis not only integrate the two kind of analysis methods which are “Principal Components Analysis (PCA)” and “Multivariate Empirical Mode Decomposition (MEMD)”, but design the power selecting rules to assist the signal analysis and effectively extract the features. First of all, by PCA, the information of signals rearrange from the most important to less one. And then the signal would be decomposed to several different physical meaning of intrinsic mode functions (IMFs) by MEMD. Lastly, after calculating the power of IMFs respectively, choosing the primary and secondary IMFs is more accurate with the power selecting rule and we can use the chosen IMFs to do different operations to get the feature extraction from the body motion data.