With the rapid growth of signal data in industrial internet, the issue of missing labels has become increasingly prominent, making self-supervised learning a critical solution. To address the problems of coarse feature granularity, unstable representation, and weak transferability in existing contrastive learning methods for signal recognition tasks, a window-based signal reconstruction contrastive learning approach was proposed. The method divided feature maps into multiple fixed windows and introduces local similarity constraints to construct a fine-grained contrastive structure. It also incorporated a signal reconstruction module to enhance the stability and semantic consistency of feature representations. Furthermore, a loss function integrating reconstruction error was designed to improve the feature fitting capability to the original signals. Experiments on three signal recognition datasets—RML, ADS-B, and CSI—show that the proposed method achieves up to 28.32% higher performance compared to other contrastive approaches. Cross-dataset transfer accuracies reach 65.36%, 66.17%, and 66.96%, respectively, significantly outperforming existing methods and demonstrating strong transfer and generalization capabilities.