Novel Class Detection And Learning Based On Multivariate Time Series

In recent years,with the rapid development of various sensor technology,especially multi-sensor technology.Massive information will be recorded and stored in relative devices,and these data will be processed in form of multivariate time series for downstream tasks.In comparison with uni-variate time series,there are not only time-dependency but also variate-dependency in multivariate time series.How to capture the temporal-spatial information becomes a current research focus.Meanwhile,novelty data will emerge with semantic shift in real complex scene.Novel class can have an adverse effect on recognition tasks.Multivariate time series and temporal evolution are closely related,so that novel class detection based on multivariate time series is important.Although some works are performed in these tasks,there are still some limitations,for instance,a lack of capturing temporal-spatial information in multivariate time series learning.In novel class detection,there are poor flexibility and utilizing label information is still inadequate.Because of such problems,we introduce relative algorithms and innovations are as follows:（1）Aiming at the problem of inadequate learning in temporal-spatial dependency of multivariate time series,we introduce a representation learning model GP-TCN for multivariate time series which combine graph pooling with temporal convolution network.Utilizing graph pooling to learn graph structure information built by multivariate time series,and fuses with temporal dependency learnt by temporal convolution network.Finally,control experiment will show the effectiveness of our model.（2）We introduce a novel class detection model-AE-DMC by combining auto-encoder with dynamic micro clusters,for learning effective low-dimension vector,and building distance metrics in micro clusters space,implement constructing micro clusters in off-line phase and detecting novel class in on-line phase.Our model solve inflexible in novel class detection and difficulties in incremental learning.We carry out experiments on multivariate time series data set and emitter signals from real scene to prove the feasibility of our model in novel class detection.