Research On Time Series Classification Algorithms Based On Multimodal Neural Network

Time series is an important form of big data,one of whose application is mining the category information through clustering or classification.Time Series Classification(TSC)task,a universal and significant topic,is often seen in fields of industry,finance and medicine.However,time series data is deficiency in overhigh dimensionality and misalignment in time steps,so it is difficult to extract useful discriminative information for classification task from them.Deep learning model has certain tolerance for input data and is able to extract features automatically,so that it performs well in TSC problems.And different types of neural networks are variant in structures and mechanisms,so there are also differences in the deep features they can learn.Integrating multiple feature information by ensemble learning or the fusion of multi models can further improve the performance.Therefore,this paper studies on TSC methods based on the multimodal neural network,aiming to learn high-dimensional time series from multiple perspectives and fully explore the features implied in them.Specifically,two TSC algorithms are proposed and validated in a relevant field:The multi-granularity curve-tendencies and time-dependencies of time series are both the important information which benefit for TSC task.A multimodal neural network composed of multiscale FCN and LSTM is proposed in this paper.This model can focus on both the multi-scale geometric spatial features of time series curves and the time-dependencies features of its values simultaneously,so that time series instances can be better distinguished by virtue of the comprehensive grasp of characteristics of them.The large-scale receptive field in this model is realized by dilated convolution to ensure that training pressure will not increase significantly.The effectiveness and superiority of the proposed model are verified by a series of experiments on UCR datasets.Multimodal networks are generally fused at feature-level,at present.Under this manner,parameters will be of large amounts and components will be coupled to each other,leading the model hard to be adequately-trained.A TSC solution which integrates the decision information of CNN and BPNN at decision level with Dempster-Shafer evidence theory is designed in this paper.In the two deep learning modules,one focuses on features extraction,and the other on the fitting of complex function relationships,and they are both trained to good states independently.The complementary information obtained by fusion helps our method to make more accurate judgments on basis of the high accuracies of single modules.Besides conventional validation experiments,this method is also tried to solve the problem of aero-engine fault diagnosis which belongs to traditional sequence classification filed.And good performances are achieved in both experiments.