A physics-informed neural network for dual-parameter decoupling and crosstalk suppression based on dynamic physical weight assignment and dual-branch decoupling

This study proposes a physics-informed neural network based on dynamic physical weight assignment and dual-branch decoupling (DPWA-DualPINN) for dual-parameter decoupling and crosstalk suppression in a fiber Bragg grating-Fabry Perot (FBG-FP) cascaded fiber-optic sensor. The cascaded FBG-FP sensor integrates an FBG sensing unit into the optical fiber link of an extrinsic FP sensor, enabling synchronous temperature and pressure (depth) sensing with parameter-decoupling capability. The dynamic physical weight assignment mechanism automatically adjusts weights based on gradient statistics, thereby enhancing the capability of the network to capture nonlinear relationships. The dual-branch decoupling architecture divides a single network into physically subspace branches with feature disentanglement while introducing a parameter-shared attention mechanism before the output layer to suppress feature crosstalk between branches. The experimental results demonstrate that compared with traditional sensitivity matrix methods, the PINN approach reduces the maximum temperature measurement residuals by 49%, accompanied by a increase of 3.59 times in prediction accuracy, while pressure measurements show a 74.4% reduction in maximum residuals and 5.23 times accuracy enhancement. This method exhibits significant advantages in enhancing the on-site measurement accuracy of marine temperature-depth parameters.