通過深度類神經網路從受雜訊汙染的多元時間序列數據中提取信號的直流分量

Abstract

深度學習在去除音頻資料雜訊中的應用引起了很多關注，但對多元時間序列資料的研究相對較少。在本文中，我們致力於透過行動裝置上的深度類神經網路從嘈雜的多元時間序列數據中提取信號的直流分量。為了解決這個問題，我們測試了不同方法對模型性能的影響。從資料增強和雜訊模擬等資料處理方法，到模型結構和特徵工程等模型構建方法，我們比較其性能並分析結果。除了這些實驗之外，我們還提出了一種稱為MPSE的新型損失函數，以幫助模型專注於小振幅信號，以保證模型性能。將我們的最佳設置模型和提出的損失函數與基準進行比較，結果表明它們具有出色的性能和強健性。我們相信這些實驗和分析可以幫助未來需要使用類似資料的研究。

The applications of deep learning to denoising audio data have attracted lots of focus, but relatively little research on multivariate time series data. In the paper, we address extracting the DC component of the signal from noisy multivariate time series data via deep neural networks on mobile devices. To solve the problem, we have tried the effect of different methods on the model performance. From data processing methods like data augmentation and noise simulation to model construction methods like model structure and feature engineering, we compare the performance and analyze the result. Besides these experiments, we also proposed a novel loss function called MPSE to help the model focus on small amplitude signals to guarantee the model's performance. Our optimal setting model and the proposed loss function are compared with their baselines respectively, and the results show they have excellent performance and robustness. We believe these experiments and analyses can help future studies that need to use similar data.