使用最佳化收縮估計法與時頻濾波器進行暫態音誘發性耳聲傳射訊號之降噪

Denoising Click-Evoked Otoacoustic Emissions by Optimal Shrinkage and Time-Frequency Filtering

劉子齊(Tzu-Chi Liu)

指導教授 ： 劉奕汶

國立清華大學/電機資訊學院/電機工程學系所/博士(2021年)

摘要

暫態音誘發性耳聲傳射(click-evoked otoacoustic emission, CEOAE)訊號在臨床上常用於新生兒聽力篩檢，由於CEOAE訊號的強度通常低於背景雜訊，其量測需要以數百、數千次的平均來估測埋藏在雜訊裡的訊號。本論文提出兩個方法來降噪CEOAE訊號：時頻濾波器(time-frequency filtering)與最佳化收縮估計法(optimal shrinkage)，其中最佳估計法有兩種版本：共變異數矩陣為基底與奇異值分解為基底；除此之外本研究也實作了使用適應性步驟找濾波區間的時頻濾波器、最佳化收縮估計法與時頻濾波器的組合、以及兩種針對最佳化收縮估計法的改動：滑動窗口最佳化收縮估計法以及逐塊最佳化收縮估計法。在模擬的CEOAE訊號上，共變異數矩陣的最佳化收縮估計法對比取資料中位數的基本方法，可以穩定增加訊雜比1到2 dB，而奇異值分解的最佳化收縮估計法可以提升3 dB訊雜比、時頻濾波器可以提升6 dB訊雜比。在真實的CEOAE訊號上，奇異值分解的最佳化收縮估計法可以穩定的提升訊雜比，而且原始訊雜比越低提升的幅度就越高；時頻濾波器只對使用非線性量測法的資料有效，而提升的訊雜比優於最佳化收縮估計法。滑動窗口法可以額外增加2 dB訊雜比，而將奇異值分解的最佳化收縮估計法使用在時頻濾波器之前可以額外增加1到2 dB訊雜比。單純考慮CEOAE訊號的降噪，最佳化收縮估計法無法贏過時頻濾波器，但最佳化收縮估計法的優勢是降噪所有訊號而不是只有平均後的訊號，這個特性讓研究長時間CEOAE的動態性成為可能。

關鍵字

耳聲傳射 ； 矩陣降噪 ； 最佳化收縮估計法 ； 時頻濾波器

並列摘要

Click-evoked otoacoustic emissions (CEOAEs) are clinically used as an objective way to infer whether cochlear functions are normal. However, because the sound pressure level of CEOAEs is typically much lower than the background noise, it usually takes hundreds, if not thousands of repetitions to estimate the signal with sufficient accuracy. In this thesis, we propose to improve the signal-to-noise ratio (SNR) of CEOAE signals within limited measurement time by time-frequency filtering (TF) and optimal shrinkage (OS) in two different settings: covariance-based OS (cOS) and singular value decomposition (SVD)-based OS (sOS). We also implemented Wiener filtering (WF), an adaptive procedure for finding the T-F region for the time-frequency filtering, combined the sOS and the TF, and presented two approaches to improve the performance of sOS: window-based sOS (wOS) and block-wise sOS (bOS). By simulation, the WF and the cOS consistently enhanced the SNR by 1 to 2 dB from a baseline method (BM) that is based on calculating the median. The sOS could enhance the SNR by over 3 dB and the TF even achieved 6 dB enhancement. In real data, the sOS improved the SNR consistently, and the level of enhancement increases as the baseline SNR decreases. The TF only worked on the CEOAE signals measured by nonlinear protocol, and it provides greater SNR enhancement than the sOS. The wOS further improved the SNR by 2 dB from the sOS, and applying sOS before TF further improved the SNR by 1 to 2 dB from the TF. For the denosing of CEOAE signals, the performance of TF is better than OS. Nonetheless, an appealing property of OS is that it produces an estimate of all single trials. This property makes it possible to investigate CEOAE dynamics across a longer period of time when the cochlear conditions are not strictly stationary.

並列關鍵字

otoacoustic emissions ； matrix denoising ； optimal shrinkage ； time-frequency filtering

參考文獻

[1] D. L. Donoho, M. Gavish, and I. M. Johnstone, “Optimal shrinkage of eigenvalues in the spiked covariance model,” Ann. Stat., vol. 46, no. 4, p. 1742, 2018.

[2] D. T. Kemp, “Otoacoustic emissions, travelling waves and cochlear mechanisms,”Hearing Research, vol. 22, no. 13, pp. 95–104, 1986.

[3] L. Collet, M. Gartner, A. Moulin, I. Kauffmann, and F. Disant, “Evoked otoacoustic emissions and sensorineural hearing loss,” Archives of Otolaryngology–Head & Neck Surgery, vol. 115, no. 9, pp. 1060–1062, 1989.

[4] A. J. Hall and M. E. Lutman, “Methods for early identification of noiseinduced hearing loss,” Audiology, vol. 38, no. 5, pp. 277–280, 1999.

[5] M. A. Hotz, R. Probst, F. P. Harris, and R. Hauser, “Monitoring the effects of noise exposure using transiently evoked otoacoustic emissions,” Acta Otolaryngologica, vol. 113, no. 4, pp. 478–482, 1993.