Time Series Feature Representation With Deep Learning

As a form of time-series data has great value out of a widespread, its space exists throughout all areas of life, multimedia data, financial data, weather data, census data, system log data are the existing form of time series. Accordingly, the time-series data mining research has achieved successful application in the field of meteorology, transport, finance, industry, medicine, computer networks, and agriculture. However, due to the timing correlation, high dimensions, massive and multi-feature time-series data of characteristics such as how to effectively characterize time series data features, in order to more effectively carry out data mining faces enormous challenges. Time series data contains a huge valuable information, including intuitive visible and potential implicit information, how to effectively mine the implicit potential of knowledge and information from these vast amounts of high-dimensional time series data has important theoretical and practical significance.Characteristics of time series is expressed in the time-series data to ensure that critical information condition, how to effectively represent data, reduce data dimension and remove noise effectively select the dimension low, a small number can reflect the original time series data of key information feature subset in order to achieve the dimension of simple time-series data. Traditional time series feature representation algorithm is mainly based on a hypothetical model formula for time series check out the location of sampling, this method is not intuitive, and the need for time-series data useful research experts to build mathematical models feature when the projector is easy to lose out important information on the time domain, so the traditional time-series data feature representation algorithms can’t deal effectively with today’s era of big data mass time-series data.Paper studies the characteristics of time series data mining problem represented by the time series analysis of existing strengths and deficiencies feature representation algorithm. Reference depth study of computer vision/image recognition in the field of groundbreaking research, the latest research results applied to deep learning time series data feature representation work through in-depth study of a variety of deep learning models, eventually discover the depth of belief network model for the time series feature indicates good performance. And the model on the 19 kinds of UCR reference time series data sets and four traditional characteristics indicate algorithm performance comparison.