Physically Plausible Data Augmentations in Contrastive Learning for Wearable IMU-based HAR

This repository contains the implementation of Physically Plausible Data Augmentations (PPDA) in contrastive learning for wearable Inertial Measurement Unit (IMU)-based Human Activity Recognition (HAR)

📄 Read our paper (arXiv)

---

Figure: Contrastive learning with Physically Plausible Data Augmentations (PPDAs). PPDAs generate realistic variations (movement amplitude, speed, sensor placement, hardware effects) using physics simulation.

📌 Introduction

This repository provides the official implementation of Physically Plausible Data Augmentations (PPDA) in contrastive learning for wearable IMU-based Human Activity Recognition (HAR).

Recent work has shown that contrastive self-supervised learning (e.g., SimCLR) can learn strong representations from unlabeled sensor data. A key factor in its success is the use of data augmentations to create multiple views of the same signal. However, conventional Signal Transformation-based Data Augmentations (STDA) often rely on arbitrary signal distortions (e.g., jittering, random rotations), which can generate unrealistic samples and potentially break the semantic meaning of activities.

Physically Plausible Data Augmentation (PPDA) [1] overcomes these issues by using a physics-based simulation framework (WIMUSim) [2] to generate realistic variations in:

• Movement Amplitude
• Movement Speed
• Sensor Placement
• Hardware-related Effects (noise and bias)

By pretraining on PPDA-augmented views, our models achieve:

• Higher downstream accuracy across multiple HAR datasets (REALDISP, REALWORLD, and MM-Fit)
• Better label efficiency (reaching supervised performance with <10% labeled data),

⚙️ Installation

Prerequisites

• Python 3.10 or higher
• Git

1. Clone this repository and WIMUSim repository

git clone https://github.com/USERNAME/PPDA_Contrastive.git
git clone https://github.com/STRCWearlab/WIMUSim.git
cd PPDA_Contrastive

2. Install dependencies

pip install -e WIMUSim
pip install -r requirements.txt

3. Configure environment variables Set the WANDB_ENTITY environment variable to your Weights & Biases entity name.

WANDB_ENTITY=`replace_with_your_wandb_entity`

4. Prepare datasets Download datasets (REALDISP, REALWORLD, MM-Fit) as described in data/README.md.

🧪 Running Scripts

Comparing STDA vs PPDA (Section 4.2)

1. Single-augmentation experiment (e.g. MM-Fit dataset):

# PPDA with single augmentation (e.g., magscale only)
python simuclr_pt_ft_ppda.py --wandb \
                             --seeds 1 2 3 \
                             --pt_encoder DeepConvLSTM \
                             --dataset_name mmfit \
                             --pt_num_epochs 200 \
                             --ft_num_epochs 50 \
                             --first_augs magscale \
                             --magscale_sigma 0.2
                             
# STDA with single augmentation (e.g., magscale only)
python simuclr_pt_ft_ppda.py --wandb \
                             --seeds 1 2 3
                             --pt_encoder DeepConvLSTM \
                             --dataset_name mmfit \
                             --pt_num_epochs 200 \
                             --ft_num_epochs 50 \
                             --first_augs magscale \
                             --magscale_sigma 0.2

2. pairwise-augmentation experiment (e.g. MM-Fit dataset):

# PPDA with pairwise-augmentation (e.g., magscale and timewarp)
python simuclr_pt_ft_ppda.py --wandb \
                             --seeds 1 2 3
                             --pt_encoder DeepConvLSTM \
                             --dataset_name mmfit \
                             --pt_num_epochs 200 \
                             --ft_num_epochs 50 \
                             --first_augs magscale timewarp\
                             --magscale_sigma 0.2 \
                             --timewarp_max_speed_ratio 1.5 \
                             --timewarp_knot 4
                             
# STDA with pairwise-augmentation (e.g., magscale and timewarp)
python simuclr_pt_ft_ppda.py --wandb \
                             --seeds 1 2 3
                             --pt_encoder DeepConvLSTM \
                             --dataset_name mmfit \
                             --pt_num_epochs 200 \
                             --ft_num_epochs 50 \
                             --first_augs magscale timewarp \
                             --magscale_sigma 0.2 \
                             --timewarp_max_speed_ratio 1.5 \
                             --timewarp_knot 4

3. Multi-augmentation experiment (e.g. MM-Fit dataset): Similar to the pairwise-augmentation experiment, enable multiple augmentations by listing them in --first_augs. For example, to combine magnitude scaling, time warping, rotation, and noise/bias:

--first_augs magscale timewarp rotation noisebias

Label Efficiency Evaluation (Section 4.3)

To evaluate label efficiency, fine-tune a pretrained model with a fraction of labeled data (e.g., 10%):

python simuclr_ft.py --wandb \
                     --seeds 1 2 3
                     --pt_run_name ppda_magscale_020 \ # pretraining run name (used to find the pretrained model checkpoint)
                     --dataset_name mmfit \
                     --pt_encoder TPN \
                     --ft_num_epochs 50 \
                     --ft_adjust_epochs \
                     --ft_data_sample_type stratified \
                     --ft_data_fraction 0.1 \

📚 Links

• WIMUSim: Paper, GitHub
• PPDA: Paper, GitHub

🙏 Acknowledgements

Parts of this repository are adapted from dl_har_public and soar_2024.
We thank the authors for making their code available.

References

[1] Oishi, N., Birch, P., Roggen, D., & Lago, P. (2025). WIMUSim: simulating realistic variabilities in wearable IMUs for human activity recognition. Frontiers in Computer Science, 7, 1514933.

[2] Oishi, N., Birch, P., Roggen, D., & Lago, P. (2025). Physically Plausible Data Augmentations for Wearable IMU-based Human Activity Recognition Using Physics Simulation. arXiv preprint arXiv:2508.13284