Composite decoupling of temperature and strain in FBG sensors using physics-informed neural network (PINN)

To address the critical issue of cross-sensitivity between temperature and strain in fiber Bragg grating (FBG) sensors, this paper proposes a multi-parameter composite decoupling method based on a hardware-software collaborative physics-informed neural network (PINN) architecture. Built upon a hardware differential structure, the proposed method embeds the governing physical equations of grating sensing as prior constraints into the loss function of the neural network. This integration guides the model to optimize parameters within a solution space consistent with physical laws. By taking the wavelength shifts of dual gratings from the differential structure as inputs, the model achieves high-precision output of decoupled real-time temperature and flow velocity. Blind test results demonstrate that the PINN algorithm achieves a mean absolute error (MAE) of 0.00458m/s for flow velocity and 0.12 ∘ C for temperature. Bland-Altman consistency analysis confirms that this model not only overcomes the limited generalization of traditional pure data-driven algorithms but also significantly enhances measurement accuracy and stability under extreme operating conditions.