Time Series Modeling Method Based On Deep Learning

Time series has a wide range of applications in real life.How to effectively model time series data has become one of the focuses of extensive research.Time series classification and time series anomaly detection tasks have important application values in time series information systems.In recent years,many deep learning methods have made great progress in these time series modeling tasks.Traditional Echo State Network(ESN)has achieved excellent perfor-mance in chaotic time series tasks due to its ability to model short-term dependence.In time series classification tasks,the time series classification methods based on echo state networks have achieved excellent performance in recent years.However,these methods have shortcom-ings in the modeling of long-term dependencies of time series.In time series anomaly detection tasks,deep learning methods based on reconstruction framework have achieved good results in recent years.However,it is difficult to capture the multi-scale information patterns of normal sample fragments in time series,which hampers these methods in downstream anomaly detec-tion tasks.Based on the existing research works and their shortcomings,this paper proposes two deep learning-based time series modeling methods as follows:(1)We proposed a multi-time series classification method called Multihead Echo Self-Attention Memory Network(MESAMN).Specifically,MESAMN consists of a multi-head self-attention encoder and a convolutional memory learner.In the multi-head self-attention en-coder,multiple differently initialized ESNs are utilized for high-dimensional projection,which is then followed by a self-attention mechanism to capture the long-term dependent features.Subsequently,a convolutional memory learner learns features extracted by the memory encoder and applies classification.Experimental results show that,compared with the existing models,MESAMN exhibits superior performance on 18 multivariate time series classification tasks and three 3D skeletal-based action recognition tasks.In addition,we verify the MESAMN’s ability to capture long-term dependent features through experiments.(2)We proposed a time series anomaly detection method named Self-Supervised Discrimi-native Network(S~2DN).Specifically,our S~2DN includes a multiscale down-sampling module,a feature extraction module,and a surrogate supervision module.The multiscale down-sampling module downsamples the original time series by using different downsampling rates.Thus,it creates different subsequences and pseudo-labels,which can capture the multiscale behavior in time series.The feature extraction module,based on the convolutional network,learns to capture multiscale feature information of the time series under different downsampling rates.Finally,the surrogate supervision module uses the self-supervised loss function to perform optimization training to separate abnormal and normal samples.Experimental results show that the proposed S~2DN achieves great performance on 18 time series anomaly detection tasks.