Data augmentation for Human Activity Recognition with Generative Adversarial Networks.

Area of Science:

• Machine Learning
• Signal Processing
• Biomedical Informatics

Background:

• Human Activity Recognition (HAR) requires extensive labeled data for generalization.
• Acquiring labeled data for HAR is often resource-intensive and time-consuming.
• Generative Adversarial Networks (GANs) show promise for synthetic data generation in HAR, outperforming traditional augmentation methods.

Purpose of the Study:

• To introduce conditional Wasserstein Generative Adversarial Networks (cWGANs) as an optimal architecture for generating synthetic HAR accelerometry signals.
• To establish a robust methodology for evaluating the quality and accuracy of GAN-generated synthetic data.
• To investigate the impact of synthetic data on HAR model performance across varying dataset sizes.

Main Methods:

• Implementation of conditional Wasserstein Generative Adversarial Networks (cWGANs) using 1D convolutional layers for accelerometry signal synthesis.
• Calculation of established signal quality and accuracy metrics to evaluate synthetic data.
• Assessment of the influence of incorporating cWGAN-generated data on a large-scale HAR dataset comprising 395 users.

Main Results:

• cWGAN architecture demonstrated superior performance compared to standard Conditional Generative Adversarial Networks (cGANs) for accelerometry signal generation.
• The inclusion of synthetic data led to more significant performance improvements in smaller real datasets.
• An inverse relationship was observed between the quantity of synthetic data needed and the amount of available real data.

Conclusions:

• Conditional Wasserstein Generative Adversarial Networks (cWGANs) provide an effective approach for generating high-quality synthetic accelerometry data for HAR.
• Synthetic data generation is a valuable strategy to augment limited datasets, enhancing HAR model generalization and accuracy.
• The proposed cWGAN methodology and evaluation framework advance the field of synthetic data generation for HAR applications.