Identification and Analysis of Non-Stationary Time Series Signals Based on Data Preprocessing and Deep Learning
Author:
Duan Li,Cai Jianxian,Liang Juan,Chen Danqi,Sun Xiaoye
Abstract
Deep learning is not the most accurate way for recognizing time series signals, and it is unable to identify non-stationary time series signals with numerous chaotic classes. Moreover, the signal detection benefits from data preprocessing have gone unnoticed. Therefore, this paper investigates the detection and analysis of non-stationary time series signals using deep learning and data preprocessing. The fitting model of the historical stationarity index is built based on the Gaussian mixture model of single Gaussian models, and the change point of the non-stationary time series signal is detected. To further increase the signal's recognition rate, the non-stationary time series signal is preprocessed using the truncated migration algorithm. The main classification task and the auxiliary classification tasks are constructed to identify non-stationary time series signals characterized by huge chaotic classes through multi-task learning. The efficiency of the suggested method and model is validated by experimental data.
Funder
Langfang Science and Technology Research Self-Funded Project
Fundamental Research Funds for the Central Universities
Publisher
International Information and Engineering Technology Association
Subject
Electrical and Electronic Engineering