<p>This study proposes a method of recognizing knee joint motion intentions based on a flexible sensor array and deep learning for wearable exoskeleton. First, a distributed sensing system is established using a flexible sensor with a three-layer stacked structure, which includes a flexible substrate layer, a functional conductive layer, and a protective encapsulation layer. By capturing skin deformation information, the system obtains changes in joint angles. Secondly, a GCN-Transformer model is proposed, combining the spatial feature extraction capability of the Graph Convolutional Network (GCN) and the temporal modeling advantages of the Transformer encoder to predict exoskeleton joint angles. Finally, experiments are conducted with six participants performing walking, stair climbing, and squatting motions. The GCN-Transformer model achieves an R² value of 0.95 for knee joint angle prediction, demonstrating the experimental model feasibility. The approach combining the flexible sensor with the GCN-Transformer model ensures high accuracy while enhancing wear ability and environmental adaptability, providing a new technical pathway for applying exoskeleton robots in complex operating conditions.</p>