Research On Time Series Classification Based On Spectrum Attention Mechanism And Encoder-Decoder Model

Time series refers to real-valued series data whose attributes have order relationship.It exists widely in our social life and engineering field.The application of traditional time series classifica-tion algorithms often requires expert experience,so the algorithm is cumbersome and not universal.The deep learning model is end-to-end,which can automatically learn the internal pattern of the data,and has the advantages of strong adaptability and good portability.This paper conducts in-depth research to establish deep learning classification algorithms suitable for time series data,and its main contributions are as follows:Firstly,time series data has both global and local features and most of the existing models use stacked networks.As the network deepens,the sensitivity of the network to local modes decreas-es.In addition,the existing models do not take into account the inconsistency of the importance of attributes of time series data in the time dimension.In response to the above problems,an encoder-decoder model is proposed.The model firstly extracts the global semantic information of the original data through a Convolutional Neural Network(CNN)encoder,and then maps it into discriminative features using a Long Short Term Memory(LSTM)decoder.In addition,the atten-tion mechanism is introduced in the decoding process to enhance the model's ability to perceive local patterns.Finally,the Fully Connected layer is used to generate the categorical distribution.Experimental results on 4 representative time series datasets show that the classification perfor-mance of this model exceeds other commonly used algorithms.This paper also demonstrates the rationality of the model by visualizing the decoded output sequence and the attention weight.Secondly,affected by the external environment and processing methods,time series data usu-ally contains a lot of noise,which is not conducive to network training.However,the existing algorithms need to manually design the data preprocessing process when they are applied,and the scheme is complicated and not universal.Considering that the essence of frequency domain filter-ing is to assign an appropriate weight to each frequency component,which is similar to the idea of the attention mechanism.Therefore,a Spectrum Attention Mechanism(SAM)is proposed in the paper.By adding a trainable mask branch to the spectrum of the original data,the network can adaptively learn the importance of each frequency component during the training process,there-by generating a feature representation that is more conducive to network training.In addition,in order to avoid the complete loss of time-domain information,this paper further proposes a Seg-mented Spectrum Attention Mechanism(SSAM)and a corresponding heuristic searching segment number algorithm.Experimental results show that the introduction of this module can improve the classification accuracy of the model,make the network converge faster,and be more robust to noise.Finally,on the basis of the above two modules,this paper studies and establishes accurate and general time series classification algorithms from the perspectives of single-model and multi-model ensembles respectively,and verifies the performance of the algorithm on 42 datasets.For the single model algorithm,the segmented spectrum attention mechanism and the encoder-decoder model are combined to form a strong benchmark model SSAM-CLA.Experimental results show that the classification performance of this model significantly exceeds all other single-model al-gorithms.For the multi-model ensembles algorithm,the deep learning model is used as the base classifier,and the CODLE(Collective Of Deep Learning Ensembles)algorithm is proposed.The voting(CODLE-I),weighted voting(CODLE-?)and learning method(CODLE-?)are used as the combination strategy.The experimental results show that the classification performance of CODLE-III is not statistically different from the state-of-the-art COTE and HIVE-COTE,but the time cost is far reduced,which has more practical significance.