Modelling Time-series Data Generation with Diffusion Models for triaxial data

Abstract

Recent advancements in diffusion models have demonstrated their capacity to generate high-quality synthetic images. However, their application to time-series data generation remains under-explored. We propose Diff-TSD, a novel method that leverages diffusion models and Modified Recurrence Plots to generate synthetic triaxial time-series data. By transforming raw signals into RGB image representations, we enable the application of image-based generative models for time-series data generation. In this framework, raw triaxial time-series data, such as accelerometer readings, are first transformed into trivariate recurrence plots that preserve the temporal correlation of the signals. These plots are treated as images, where each variate corresponds to an RGB channel, enabling the use of diffusion models traditionally designed for image synthesis. To evaluate the effectiveness of Diff-TSD, we conducted experiments using the WISDM benchmark dataset for human activity classification based on triaxial accelerometer data. The model, trained over 10,000 diffusion epochs, achieved an accuracy of 76.921 %, an F1-Score of 77.281 %, and exhibited stable performance from 5,000 epochs onward. While the accuracy shows a 7.59 % difference compared to real data performance (84.51 %), Diff-TSD proves valuable for augmenting training datasets, particularly under conditions of data scarcity. Additionally, the optimal trade-off between computational cost and performance was identified at 5,000 diffusion epochs, where the Xception classifier demonstrated the most consistent results. The proposed method contributes to the field of soft computing by leveraging diffusion models for synthetic time-series data generation, offering a robust solution for enhancing machine learning tasks in scenarios with limited data availability.