融合多变量序列时空信息的事件早期识别方法

Aiming at the problem that existing early event detection algorithms do not fully exploit the spatial-temporal relationships of multivariate time series, an early event detection method based on spatial-temporal fusion is proposed. Time convolutional network（TCN）is adopted for sequence prediction by learning the temporal patterns from the input temporal events described by multivariate time series. Both the original and predicted time series are concatenated along the temporal dimension and fed into a novel multivariate spatial-temporal neural network（MSTNN）for classification. Experimental results for recognizing partial events in multiple scenarios, including action classification, handwritten character recognition, and spoken digit recognition, demonstrate the benefits of the proposed method compared to existing approaches, which achieves an average accuracy of 93.2% when 2/3 of the temporal events have been observed. The performance on the handwritten character recognition dataset is comparable with the case when the complete event is observed.