With the wide application of wireless sensor networks (WSN) in smart cities, industrial IoT, and other fields, the security threats it faces are becoming increasingly complex. Traditional detection methods rely too much on manually labeled data and fixed feature engineering, making it difficult to cope with novel attacks in dynamic environments. In this paper, we propose a method, which is based on self-supervised learning and deep learning security detection framework for WSNs, breaks through the data labeling restriction through spatio-temporal feature fusion and comparison learning strategy, which can improve the model's ability to identify abnormal behaviors. In this study, we first constructed a spatio-temporal feature fusion module, which uses gated recurrent unit (GRU) to capture time series dependency, combines with graph neural network (GNN) to model node spatial correlation, and dynamically integrates spatio-temporal features through feature selection and adaptive weighting mechanism. Then a dual self-supervised learning strategy is designed: on the one hand, the original data is reconstructed by an autoencoder to learn the data distribution law; on the other hand, contrast learning is introduced to construct positive and negative sample pairs to enhance the feature discriminative properties. From the results of the experiments, it can be seen that the method achieves an accuracy of 95.3% in the anomaly detection task, with an F1 score of 0.94, which effectively reduces the false alarm rate and enhances the robustness of the method in the case of fewer samples.