Abstract: Modern large-scale building structures are becoming increasingly complex, which imposes higher requirements on the optimal layout of sensors in their structural health monitoring systems. However, when neural network models are used for damage identification, they cannot effectively capture the spatially coupling relationship between multiple sensors and easily generate recognition mappings that deviate from physical laws. To address the aforementioned issues, an intrinsic reinforcement network damage identification method based on J-divergence physical constraint is proposed. This study first separates the intrinsic modes of time-domain signals via variational mode decomposition, extracts spatial dimension information from multiple sensors using self-attention mechanism, and extracts temporal dimension information using Gated Recurrent Unit (GRU) to construct a seed-transfer intrinsic reinforcement neural network. Then, the J-Divergence Constraint Intrinsic Reinforcement Network (JDC-IRN) damage identification method is established by using the J-divergence and Auto Regressive Moving Average Vector (ARMAV) models to train the neural network and constructing a loss function physics module that integrates physical prior knowledge. Finally, the effectiveness of the proposed method is verified through numerical simulation and experimental research. The results show that this method is superior to traditional long short-term memory network and convolutional neural network in damage identification, and has stronger robustness under noisy environments and small-sample conditions.