An IPNN-Based Parameter Identification Method for a Vibration Sensor Sensitivity Model

Vibration sensors, as critical components in motion control and measurement systems, have dynamic characteristics that directly affect measurement accuracy. However, existing sensitivity models, due to structural simplifications and parameter uncertainties, hinder conventional vibration and shock calibration methods from fully characterizing their dynamic performance. In addition, traditional parameter identification approaches are often noise-sensitive and lack interpretability, making them inadequate for high-precision applications. To address these challenges, this study proposes an Algorithm-Unrolled Interpretable Physics-Informed Neural Network (IPNN) for parameter identification of a vibration sensor sensitivity model. By integrating the physical characteristics of the sensors with vibration calibration data, the method enables high-precision parameter identification and interpretable dynamic modeling. Comparative experimental results show that the proposed IPNN reduces the RMSE of sensor voltage predictions by over 60% compared with GRU and LSTM and decreases the average full-frequency relative deviation from laser interferometry calibration results by approximately 65% relative to LSM.