SGSM: semi-generalist sensing model combining handcrafted and deep learning methods

The significance of intelligent sensing systems is growing in the realm of smart services. These systems extract relevant signal features and generate informative representations for particular tasks. However, building the feature extraction component for such systems requires extensive domain-specific expertise or data. Handcrafted methods focus on general data features rather than specific task characteristics, hence the need for human expertise. Deep learning models place greater emphasis on task-specific labels rather than the underlying data characteristics, necessitating a substantial amount of annotated data. Therefore, it is advantageous to combine the strengths of both approaches by integrating the data handling expertise of existing signal processing methods with the task depicting capabilities of deep learning models. In this paper, we propose SGSM, the first work to link handcrafted and deep learning approaches in the field of sensing. SGSM provides a semi-automated intelligent sensing scheme, which can adaptively and quickly guide the design of particular sensing systems for various given tasks across different sensors. Experimental results on three heterogeneous sensors (acoustic, inertial measurement unit, and Wi-Fi) illustrate that SGSM functions across a wide range of scenarios, thereby establishing its broad applicability. In some cases, SGSM even achieves better performance than sensor-specific specialized solutions.