Data-efficient explainable modeling: Transfer learning-assisted response prediction and sensor placement

Measuring vibration response data is essential for analyzing aircraft behavior and state under a complex environment. Despite its significance, the high acquisition cost of accurate sensors and the limited amount of response data obtained under specific working conditions hamper its translation into response prediction. Existing methods yield inferior performance due to ignoring two significant challenges: (1) lacking the stringent selection of reliable sensors for response prediction; and (2) ignoring the knowledge transfer from different working conditions, resulting in poor prediction performance when available data is limited. To address these issues, we propose the sparse learning method based on knowledge transfer (SLMKT) to achieve cost-effective and accurate response prediction under specific working conditions. Specifically, our innovative design embodies two main modules: a feature selection module based on sparse learning (FSMSL) and a knowledge transfer module guided by samples (KTMS). The FSMSL module selects the sensor locations most relevant to the target response, yielding interpretable and optimal sensor placement results. The KMTS module then achieves efficient response prediction by transferring knowledge from regular to specific working conditions. These two modules, namely the FSMSL and KMTS, complement each other and together constitute the proposed SLMKT. We validate our method using synthetic datasets, standard aircraft, and large commercial aircraft. The results show that SLMKT can not only predict target responses under small-sample working conditions but also more accurately identify relevant sensor locations that meet industrial requirements. Moreover, SLMKT maintains strong robustness across different scenarios.