Foundation model

HeteroWear (sensor-aware SSL encoder for heterogeneous wearable HAR)

Venue listedCatalog paper date: Nov 13, 2026Added to atlas: Nov 17, 2026Source

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

LODO pretraining + subject-disjoint fine-tune on 5 HAR benchmarks: mean macro-F1 0.724 vs 0.688 scratch; HHAR 0.823 vs 0.741; held-out device +0.066; 50% channel removal keeps 73.8% vs 64.1%. No board row until per-dataset tables are checked.

Signals
PPG, IMU
Form factor
Any
License
MIT (code repository); no released weights
Year
2026
Architecture

Sensor-aware tokenizer (per-stream normalisation, P=8 patches x L=16, modality-specific strided Conv1d) -> within-sensor temporal Transformer -> sensor-context embedding (modality, location, device, log-rate) -> cross-sensor Transformer with padding mask -> latent cross-attention fusion (K=4 queries) (README)

Pretraining data

8 public datasets: PAMAP2, WISDM, HHAR, HARTH, UCI-HAR, PPG-DaLiA, WESAD, MMASH (173 subjects, 358,261 windows)

At a glance
Type
Foundation model
Modalities
PPGIMU
Form factor
Any
Parameters
Not reported
Architecture
Sensor-aware tokenizer (per-stream normalisation, P=8 patches x L=16, modality-specific strided Conv1d) -> within-sensor temporal Transformer -> sensor-context embedding (modality, location, device, log-rate) -> cross-sensor Transformer with padding mask -> latent cross-attention fusion (K=4 queries) (README)
Venue
IEEE BigData 2026
Year
2026
DOI
Not reported
Weights
Not released
Licence
MIT (code repository); no released weights
Caveats
Not reported
Added
Nov 17, 2026
Updated
Nov 17, 2026

Abstract

Self-supervised learning for wearable sensing is usually confined to a single dataset or device family, leaving open whether representations transfer across heterogeneous sensor configurations and whether gains come from pretraining rather than architecture or richer sensor input. We present HeteroWear, a sensor-aware self-supervised framework pretrained on an eight-dataset wearable corpus (173 subjects, 358,261 windows) spanning inertial and physiological signals across body locations, devices, and sampling rates. HeteroWear separates within-sensor temporal modeling from cross-sensor relational modeling using sensor-context embeddings, temporal and cross-sensor Transformers, and latent cross-attention fusion. Training combines structured sensor dropout, cross-configuration consistency, and per-axis masked sensor modeling; missing sensors are omitted rather than reconstructed. Under leave-one-dataset-out pretraining and subject-disjoint evaluation on PAMAP2, WISDM, HHAR, HARTH, and UCI-HAR, HeteroWear improves mean macro-F1 from 0.688 to 0.724 over the identical architecture trained from scratch, with the largest gain on HHAR (0.823 versus 0.741), and achieves the highest macro-F1 on four of five benchmarks relative to single-accelerometer supervised baselines. Held-out-device evaluation yields a mean +0.066 macro-F1 improvement. Under 50% test-time channel removal, HeteroWear retains 73.8% of full-channel performance versus 64.1% for the same architecture trained from scratch without sensor dropout. Ablations show that structured sensor dropout primarily drives robustness, while cross-configuration consistency and sensor-context encoding contribute to transfer.

Verbatim fromarXiv· arXiv metadata (abstracts are released under CC0)

HeteroWear (sensor-aware SSL encoder for heterogeneous wearable HAR): ChatPPG research summary card

Key results

0.724
F1
LODO pretraining + subject-disjoint fine-tune on 5 HAR benchmarks: mean macro-F1 0.724 vs 0.688 scratch

How it works

How HeteroWear (sensor-aware SSL encoder for heterogeneous wearable HAR) works: Input signals → Foundation model → Downstream tasksInput signals: PPG (Photoplethysmogram); IMU (Accelerometer, 3-axis). Datasets: PAMAP2, WISDM AR v1.1, HHAR, HARTH, UCI Human Activity Recognition Using Smartphones (UCI-HAR), PPG-DaLiA, WESAD, MMASH. Foundation model: HeteroWear (sensor-aware SSL encoder for heterogeneous wearable HAR) (Transformer · Masked reconstruction). Downstream tasks: Activity recognition.Input signalsPPG: PhotoplethysmogramPPGPhotoplethysmogramIMU: Accelerometer, 3-axisxyzIMUAccelerometer, 3-axisDatasets: PAMAP2, WISDM AR v1.1, HHAR, HARTH, UCI Human Activity Recognition Using Smartphones (UCI-HAR), PPG-DaLiA, WESAD, MMASHDatasetsPAMAP2, WISDM AR v1.1 +6 moreFoundation model×NHeteroWear (sensor-aware SSL…Transformer · Masked reconstructionDownstream tasksActivity recognitionActivity recognition

Datasets linked in the hub; pretraining vs evaluation roles follow the paper.

Pretraining data
8 public datasets: PAMAP2, WISDM, HHAR, HARTH, UCI-HAR, PPG-DaLiA, WESAD, MMASH (173 subjects, 358,261 windows)
Architecture
Sensor-aware tokenizer (per-stream normalisation, P=8 patches x L=16, modality-specific strided Conv1d) -> within-sensor temporal Transformer -> sensor-context embedding (modality, location, device, log-rate) -> cross-sensor Transformer with padding mask -> latent cross-attention fusion (K=4 queries) (README)
Inputs
PPG, IMU
Approach
foundation model · self supervised · HAR · heterogeneous sensors · sensor dropout · missing modality

Devices· 2

Full photo credits and licences are on each device page. An icon means we have no openly licensed or manufacturer-cleared photo of that exact device yet.

Datasets used· 7

Papers· 1

Cite this model

Cite the publication recorded for this model. Check it against the publisher before submitting.

Mohammadpour, Y., & Shaeri, P. (2026). HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition. IEEE BigData 2026. https://arxiv.org/abs/2611.19027
Mohammadpour, Yasaman and Pouya Shaeri. 2026. "HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition." IEEE BigData 2026. https://arxiv.org/abs/2611.19027.
Mohammadpour, Y. and Shaeri, P. (2026) 'HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition', IEEE BigData 2026. Available at: https://arxiv.org/abs/2611.19027
Mohammadpour, Yasaman, and Pouya Shaeri "HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition." IEEE BigData 2026, 2026. https://arxiv.org/abs/2611.19027.
Y. Mohammadpour and P. Shaeri, "HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition," IEEE BigData 2026, 2026, [Online]. Available: https://arxiv.org/abs/2611.19027.
@article{mohammadpour2026heterowear,
  title         = {{HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition}},
  author        = {Mohammadpour, Yasaman and Shaeri, Pouya},
  year          = {2026},
  journal       = {IEEE BigData 2026},
  eprint        = {2611.19027},
  archivePrefix = {arXiv}
}

Source metadata: 2026

Record updated Nov 17, 2026

History

  1. Added to the ChatPPG research hub