Rotation-Robust Self-Supervised Learning for IMU-Based HAR With Multi-Scale Convolutional Backbones

Sensor-based human activity recognition (HAR) models are often brittle to sensor orientation changes and placement variability, while heavy backbones for multi-scale temporal modeling hinder on-device deployment; moreover, collecting large labeled corpora is costly, motivating label-efficient representation learning. We address these challenges with a lightweight self-supervised framework that uses discrete canonical views to improve robustness to practical sensor orientation changes without specialized SO(3)-equivariant operators or continuous group modeling, together with InceptionMK, a lightweight backbone that combines Inception-style parallel paths and multi-kernel temporal convolutions for efficient multi-scale modeling. Comprehensive experiments on four public benchmarks show strong performance for the Self-Supervised Learning (SSL) variant, achieving F1 scores of 0.9691 on UCI-HAR, 0.9669 on WISDM, 0.9736 on PAMAP2, and 0.9883 on mHealth. We further validate practical deployability on a Raspberry Pi 3B+, demonstrating compact and real-time on-device inference. Overall, combining a lightweight canonical-view pretext task with an efficient multi-scale backbone provides a practical approach for deployable HAR with improved robustness under the evaluated orientation variability settings and limited-label regimes.