Research On Time Series Classification Via Incorporating Frequency Domain Information

With the rapid development of sensor technology and Internet technology,time series data are continuously generated and accumulated at an unprecedented speed,which contain a wealth of information.Time series analysis can help us understand the essential characteristics and development laws of things,which has important research significance and application value.For example,in the field of network security,abnormal fluctuation of network traffic can be found in time by monitoring network traffic.Time series classification is an important research branch of time series analysis.Compared with traditional classification problems,the time series classification problem is more challenging due to the sequential correlation of time series data.The traditional time series classification methods rely too much on the similarity measurement of two sequence,which are lack of universality.While the machine learning-based methods require complex feature extraction,which is time-consuming.In recent years,deep learning-based methods have been widely used in time series classification due to their good performance and universality.However,most of the existing deep learning-based methods only analyze the time series from the perspective of time domain,which is hard to effectively model the hidden information of time series for the reason that a time series usually contains trend,seasonal,cyclical and irregular components simultaneously in time domain.To tackle the aforementioned challenges,this paper tries to explore effective time series classification methods from the perspective of both time domain and frequency domain.First of all,an Adaptive Multi-level Wavelet Decomposition based neural Network(AMWDNet)is proposed to model the non-stationary and nonlinear time series.For the sake of decomposing a time series into a few components having different frequencies,an Adaptive Multi-level Wavelet Decomposition module(AMWD)is proposed,which approximatively implements the traditional wavelet decomposition algorithm under a deep neural network framework.By this way,the frequency analysis process is seamless embedded into the deep learning frameworks,which enables the fine-turning of wavelet decomposition parameters during the training phrase.Moreover,the Long-term Temporal Patterns Extraction module(LTPE)and Short-term Temporal Patterns Extraction module(STPE)are designed to extract the global trend patterns and local oscillation patterns respectively.Then,a Dual Channel Time-Frequency analysis based neural Network(DCTFNet)is proposed to model the time series from both time domain and time-frequency domain.For the reason that it is hard to detect the frequency information of a time series in a short duration,Time Domain Patterns Extraction module(TDPE)and Time-Frequency Domain Patterns Extraction module(TFDPE)are designed to extract the time-domain features and time-frequency features respectively.Moreover,multi-wavelet analysis is adopted in the TFDPE module to make up for the shortcomings of single wavelet analysis.Considering that time-frequency information at different resolution levels has different importance in many scenarios,information fusion method based on channel-wise attention mechanism is proposed to integrate the time-domain information and timefrequency information.Finally,extensive experiments are conducted on eight datasets from different domains with different lengths and demonstrate that the models proposed in this paper significantly outperform the state-of-the-art time series classification methods.