| Title |
Time Series Data Augmentation Algorithm Using Wavelet Packet Conversion And Generation Neural Network Model |
| Authors |
이상훈(Sang-Hun Lee) ; 김경호(Kyung-Ho Kimark) |
| DOI |
https://doi.org/10.5370/KIEE.2025.74.11.1998 |
| Keywords |
Acceleration?Plethysmogram; Wavelet?Packet?Transform; Time?GAN; GAN; VAE; Data?Augmentation; 1D?CNN; LSTM; LightGBM |
| Abstract |
This study tackles the data scarcity and motion artifact problems of wearable APG signals by pairing wavelet?based quality screening with generative augmentation. Clean 66?sample APG cycles were extracted, Time?GAN, GAN, and VAE. A db4, 4?level Wavelet Packet Transform filtered out low?quality samples using LF ratio?≥?0.60, HF diff?≤?0.05, phase shift?≤?10°. Four training sets (Real?Only plus three augmented variants) were tested on 1D?CNN, LSTM, and LightGBM with 5?fold cross?validation. Time?GAN data pushed LightGBM Accuracy/F1/AUC to?1.00 and showed near?perfect overlap with real signals in PCA space, whereas VAE was slightly lower and GAN lagged due to mode collapse. Results confirm that wavelet?screened Time?GAN augmentation markedly improves APG?based cardiovascular classification. |