應用於希爾伯特黃轉換之多重終止條件經驗模態分析設計與實現

Abstract

在此篇論文中，我們提出一個應用於希爾伯特黃轉換之低計算時間和高精準的多重終止條件經驗模態分析硬體加速器來處理呼吸訊號。使用4級管線及FIFO結構來降低cubic spline的計算延遲。對於精準度方面，本篇採用38位元的浮點運算來近似雙精度正確性。此外，為了處理大量資料EMD計算而使用了外部記憶體方式來設計此應用。在這次研究中，提出了應用於生醫訊號處理之多重終止條件EMD硬體加速器。 相較於軟體執行的方式，由TSMC 90 奈米製程所實作出的硬體加速器。在操作頻率為40 MHz與操作電壓為1.0伏特的情況下，可以提升效能至少43倍。而且對於使用硬體計算結果統計誤差採用正規化均方根誤差來比較以軟體方式執行的結果，其三種終止條件誤差分別是0.03, 0.055 and 0.01。

This thesis presents a low-computation-time and high-accuracy hardware accelerator of multiple stopping criteria empirical mode decomposition (EMD) unit for Hilbert Huang Transform (HHT) with the application to breath signals. The four-stage pipeline and FIFO structures are used to decrease the computation latency for cubic spline unit. From accuracy viewpoint, the 38-bit floating-point format is used to approximate the accuracy of double-precision format. Moreover, we use off-chip memory approach to process the huge input data for EMD calculation. Compared with the EMD in software, the proposed EMD hardware accelerator implemented in TSMC 90nm CMOS process at 40 MHz with core area 4.47mm2 can speed up the performance at least 43 times with respect to that in single core of ARM11 MPcore. From the hardware simulation results, the normalized root mean squared deviation (NRMSD) for breathing signals with three different criteria are at most 0.03, 0.055 and 0.01, respectively.