The underground mine environment is inherently complex and hazardous, which necessitates ensuring mining workers’ safety. With the fast development of AI techniques, human activity recognition has become a crucial task to enhance daily life, especially in such challenging settings. This enables automatic monitoring and detection of the mine workers’ activities and operations on a large scale, which leads to better support of workers’ safety and welfare. Existing traditional uni-modal approaches, which use data of one single modality, suffer from low data quality and noise, so they struggle to generalise effectively in such extreme conditions (Li et al, 2022). As underground mines have extreme and complicated environments, such as dust, low light, signal interference, and rugged terrain, uni-modal approaches could not resolve such obstacles and are not suitable for real-world deployment. Multi-modal human activity recognition approaches could use data from multi-modal sources, thus capturing complementary information and producing robust and effective performance in recognising and monitoring mining activities. In this case, designing suitable multi-sensor hardware and AI activity recognition models has become a pivotal task. Producing efficient multi-modal approaches for edge deployment is another problem worth exploring.