Human activity recognition (HAR) models are increasingly deployed in mobile and wearable systems for healthcare, smart environments, and ubiquitous computing. Although deep learning approaches achieve high accuracy, their limited interpretability restricts adoption in safety-critical applications. Existing post-hoc explainable artificial intelligence (XAI) methods are difficult to apply to HAR due to the high-dimensional, multi-sensor, and temporally complex nature of IMU data, and most prior studies focus on supervised learning settings, limiting understanding of self-supervised representations. In this paper, we revisit interpretability for self-supervised inertial measurement unit (IMU) sensing models and propose a unified explanation framework to improve the reliability of post-hoc analysis. The framework enables representative XAI methods to identify discriminative motion patterns and sensor dependencies within IMU sequences. Experiments on multiple HAR datasets demonstrate consistent and informative explanations across activities, architectures, and sensor modalities, providing a practical foundation for benchmarking interpretability in self-supervised IMU sensing systems.
