A Research On Representation Methods Of High-level Semantic Features Of Time Series

High-level semantic is the result of the gradual accumulation of non-verbal information,which is embodied in the knowledge of the physical and mechanical world that is difficult to express in language.High-level semantic information is the information of data itself,which is helpful to improve the network performance.At present,the mainstream deep neural network is a data-driven,end-to-end network structure,and the exploration of network characteristics is a kind of afterwards and explanatory research.In the industry production,which is lack of label data and dependent on human cognition,the existing neural network application is difficult.The study of high-level semantic feature representation method is helpful to build an interpretable neural network and alleviate the problem of neural network application in the industry caused by incomplete training set.Using high-level semantic features to guide neural network belongs to a new field.How to effectively represent the high-level semantic information in data and how to use the core of these information problems.In this thesis,theoretical analysis,method research and simulation verification are carried out.The main contributions are as follows:(1)Aiming at the problem that cognition is difficult to apply to neural network models,a prototype system combining high-level semantic features and neural networks is proposed,which is compatible with mainstream network models.This system combines human cognition and data features,constructs specific high-level semantic tags,uses neural networks to achieve high-level semantic feature representation,and combines transfer learning to complete feature migration,transforming human cognition into prior knowledge of neural networks Improve network performance.(2)In the time series regression problem,aiming at the problem that the network model lacks data feature guidance and poor performance,combined with the time series model in the traditional method,it is proposed to extract different frequency components in the time series to obtain high-level semantic data,and use experience in the field of signal processing Modal decomposition performs frequency division extraction on time series to realize the construction of high-level semantic labels.Aiming at the problem that the receptive field of the convolutional neural network is limited and cannot effectively extract time series features,frequency-frequency convolution is used to complete the effective characterization of high-level semantic features.(3)Aiming at the problem that neural networks are difficult to apply in industrial applications,and high-level semantic structures are difficult to model effectively,it is proposed to use dynamic time warping to cluster industry data to construct basic labels.In order to parse high-level semantics,a decoupled neural network is used to ensure that the network corresponds to high-level semantic features;by visualizing the intermediate results of the neural network,high-level semantic labels are adjusted to achieve effective representation of high-level semantic labels,while improving the results.Interpretable.The time-varying convolution of the receptive field can be changed by adapting to the characteristics of the data to improve the migration ability.The validity of the method is proved by comparison with traditional models on the data in the exploration field.