基於多域深度類神經網路之多變量時間序列早期分類

Abstract

多變量時間序列早期預測已在資料探勘領域中成為非常重要的研究議題，並已應用在許多領域中。隨著物聯網(Internet of Things)等技術的快速發展與行動感測設備之日益普及，各種類型的感測資料之快速累積形成了龐大之多變量時間序列資料，如何從中擷取出有意義的特徵資訊並進行有效的分析與處理為一重要課題。以往之時間序列早期預測方法多採用形狀特徵(Shapelet)作為分類器基礎，然而這些方法將無法在任意時間點做出預測。本研究利用結合捲積式類神經網路與長短期記憶網路提出一新穎之多域深度類神經網路方法，於多變數時間序列進行探勘分析，以建構有效的早期分類預測架構。本研究的主軸在於探討如何使用多變量時間序列之前段萃取有用資訊，用於提早預測測試序列之類別，且保證其準確率相當於利用完整資訊的時間序列執行分類之準確率。實驗結果有效證實比起現有之代表性方法，本研究方法在準確性與及早性指標上皆獲得大幅提升。

Early classification on multivariate time series has become an important topic not only in data mining fields but also in areas such as medical research, motion detection, product manufacture, financial market prediction, nature event forecasting and anomaly detection. Most of the existing works use shapelet-based methods for early classification. However, shapelet extraction process is inefficient and the extracted shapelets may not be applicable to every test case at any time point. We present a novel framework named Multi-Domain Deep Neural Network (MD-DNN) that combines the characteristic of convolutional neural network and long-short term memory. Our framework is the first work addressing convolutional neural network and recurrent neural network for multivariate time series early classification. By using a truncation process, it can make predictions at any time point of a multivariate time series. Experimental results on real datasets show that MD-DNN outperforms existing methods on both earliness and accuracy.