Physics-Informed LongNet-CNN Prediction Model for Sparse Cross-Device Calibration Transfer in Industrial Sensors

Abstract

Industrial sensor calibration transfer remains challenging because the prediction model was trained on a reference device, but the actual sensor has different sensitivity, an aged state, thermal coupling, or installation history. This paper introduces a physics-informed LongNet-CNN model for cross-device calibration migration with limited target-domain labels. Dilated long-range attention is employed to model slow drift over extended operating windows; convolutional branches capture short transient responses; and a residual physics loss is used to constrain monotonic sensitivity, thermal relaxation and actuator-energy consistency. A controlled benchmark is established by generating several conditions of pressure, flow, vibration and temperature signals from six virtual production lines with realistic bias, lag, noise and hysteresis perturbations. With only 8% of the target calibration samples, the proposed model has an average RMSE of 0.118 engineering units, which is 31.4% lower than the source-only CNN and 18.6% lower than the transfer transformer. The mean absolute calibration slope error is reduced from 0.092 to 0.031, and the ninety-fifth percentile drift overshoot under thermal ramping is limited to 0.44 units. Ablation studies show that the physics residual increases the RMSE by 12.7% when omitted, and removing dilated attention reduces long-window stability by 16.9%. Based on the above results, combine a long-context sequence model with explicit physical constraints to improve the accuracy, robustness and interpretability of calibration transfer for industrial sensing systems. The main contribution is a unified sparse-transfer framework that couples long-context drift memory, local transient extraction, and physically constrained calibration correction. The benchmark is a controlled multi-condition industrial-sensor dataset designed to emulate gain shift, lag, hysteresis, thermal coupling, and vibration disturbance.