<p>Human activity recognition (HAR) has become a fundamental research area in applications, such as healthcare monitoring and security surveillance. Although significant progress has been made, learning feature representations from time-series data remains challenging, especially when labeled data is limited. This paper proposes a novel frequency-domain feature learning framework named FIRE, which enhances domain generalization to improve HAR performance. The framework consists of three core components: frequency-domain data augmentation, multi-scale feature extraction, and a frequency-domain multilayer perceptron (MLP). FIRE leverages existing labeled data for training while ensuring generalization to unseen domains. Frequency-domain data augmentation improves robustness to intra-domain variations, the multi-scale module captures richer temporal representations, and the frequency-domain MLP learns discriminative features by modeling global spectral dependencies. Extensive experiments have been conducted based on three public datasets—USC-HAD, DSADS, and PAMAP2—under a cross-subject domain generalization setting demonstrate the effectiveness of FIRE, achieving classification accuracies of 81.48%, 91.18%, and 85.56%, respectively. The evaluation results indicate that FIRE achieves superior performance compared with other existing methods, validating its robustness against domain shifts.</p>