A method for generating periodic time series data based on self-encoder

Abstract

The invention provides a method for generating periodic time series data based on a self-encoder. Based on the encoder module, the periodic loop module and the decoder module, the periodic time seriesdata are structured, and the multi-dimensional periodic time series are reduced in dimension by the encoder, so that the data of different dimensions are decorrelated, and the dimension explosion isavoided when the periodic loop is constructed. The whole time series is reconstructed into a ring structure with one cycle as the unit, and the data of each dimension at each time point on the ring isrepresented by an independent Gaussian distribution. Gaussian distribution sample and decoders are use to generate data with strict timing relationships.

Description

technical field [0001] The invention relates to the Internet field, in particular to a method for generating periodic time series data based on an autoencoder. Background technique [0002] The periodic time series data generation technology can obtain a large amount of effective data within a certain period of time, and solve the problem that the actual data collected in the actual scene is usually incomplete. At the same time, the generated data can be used in the fields of big data analysis, machine learning and parallel intelligence. Improve the accuracy of the algorithm. The techniques closest to the present invention are: [0003] (1) Normal cloud technology: it forms a specific structure generator through expectation, entropy and hyper-entropy; then the observed data is represented by Ex (expectation), En (entropy), and He (super-entropy) through the reverse cloud generator To describe the digital characteristics, and on this basis, the cloud droplet is generated by...

AI Technical Summary

Problems solved by technology

The data generated by VAE and the normal cloud model do not have variability, and can only generate data that conforms to the original distribution, so the diversity of generated data is insufficient

Although the text sequence processed by SeqGAN has a certain similarity with the time series, compared with the text sequence, the time series has a stronger front and back dependence. This relationship is more obvious in the periodic time series, and the periodic time series The sequence itself also has obvious periodicity, so SeqGAN's processing effect on the sequential dependence of the time series is not enough

Images