Research On Time Series Classification Based On Deep Learning

Time series are ubiquitous in life,and sequence-related data can be treated as time series,so there are many application scenarios of time series processing.Time series classification is a main research direction of time series processing,which plays an important role in the fields of medical signal analysis and human activity recognition.Most of the current time series classification models are constructed based on traditional methods,which are heavily dependent on domain knowledge,and their performance is mainly determined by feature engineering,resulting in a long model construction period and a lack of generalization ability.Time series classification models based on deep learning can learn features from training,and has stronger versatility.However,compared with traditional methods,the models’ advantage of accuracy is not obvious,and require higher computing resources.Meanwhile,the problems of long-term dependencies and multi-dimensional feature extraction are tricky.Combining the advantages of various deep learning networks,it proposes two lightweight and robust time series classification models,which improve the classification accuracy and simplify the model structure,respectively solving the problems of long-term dependencies in univariate time series classification and multi-dimensional feature extraction in multivariate time series classification.(1)For univariate time series classification,it proposes Legendre Fully Convolutional Networks.The model consists of a fully convolutional module and a Legendre module.The fully convolutional module can extract the features of the data,while the Legendre module can learn the temporal information from the data.These two modules are able to process the input data in parallel,then the model combines the outputs of the two modules for classification.By Gramian Angular Field transformation,the model supports the conversion of one-dimensional signals to two-dimensional images for classification,which further improves the performance and scope of the model.The experimental results on 85 UCR univariate time series classification datasets show that the model has high classification accuracy,can effectively deal with time series classification problems of various lengths,and has strong generality and scalability.(2)For multivariate time series classification,it proposes a Quaternion Time Series Transformer.The model converts real number information into quaternion data for storage and calculation,and classifies the data through the quaternion transformer module,which can capture the correlation between multi-dimensional input sequences while effectively reducing the complexity of the networks.With unsupervised pre-training,the model converges quickly with fine-tuning and has stronger classification ability.The experimental results on 11 UEA multivariate time series classification datasets show that the model has obvious advantages in classifying high-dimensional information and has excellent classification performance and generalization ability.