Remaining useful life prediction for tunneling magnetoresistance sensors: a physics-guided few-shot learning approach

Accurate remaining useful life prediction is essential to ensure the stable, long-term operation of high-reliability sensors in critical infrastructure. Emerging magnetic sensors operate under coupled multi-physics fields, making their degradation mechanisms inherently complex. Conventional approaches are typically limited to predicting aging trends under single variables and lack the capability to estimate remaining useful life across multiple physical quantities. Moreover, data-driven methods such as neural networks require extensive test data to analyze multi-parameter degradation, which poses a significant challenge for engineering practice where available failure samples are extremely scarce. To address this limitation, we propose a physics-guided temporal attention network. Our framework systematically integrates domain-specific physical degradation principles across all stages of the machine learning pipeline. It employs a physics-guided data generator to augment the training dataset and a dual-path adaptive fusion architecture to jointly model degradation dynamics. The global path captures long-term degradation trends using self-attention, while the local path captures short-term fluctuations using convolutional operations. Physical prior knowledge is encoded as explicit input features to guide the network's attention toward critical degradation stages. By unifying physics-guided constraints with data-driven learning. Our approach stabilizes optimization during virtual pretraining under data scarcity and enables accurate prediction of real degradation trajectories. Experimental results on measured sensor data demonstrate high prediction accuracy, attaining MAE of 0.036 and coefficient of determination R² of 96.9%. Our research provides a robust solution for sensor predictive maintenance that ensures efficient use of limited data while maintaining analytical interpretability.