Wireless sensor networks enable real-time monitoring in critical applications such as industrial automation, environmental sensing, healthcare, and infrastructure management. However, sensor node faults caused by hardware degradation, communication failures, and harsh operating conditions can significantly reduce data reliability and system performance. Existing anomaly detection approaches often depend on large amounts of labelled fault data and suffer from class imbalance, noise sensitivity, computational complexity, and limited generalization in dynamic environments, highlighting the need for adaptive and scalable fault diagnosis frameworks. To address these limitations, this study proposes a self-supervised deep anomaly detection framework for wireless sensor network node fault diagnosis. The proposed method integrates preprocessing and sliding-window segmentation with a Deep Dual Masked Transformer (DDMT) architecture to capture temporal dependencies and learn discriminative features without extensive labelled data. A self-supervised reconstruction objective with a 15% masking strategy is employed to optimize latent representations, while an adaptive anomaly-scoring mechanism generates reliable fault predictions. Experimental results on the Zidi wireless sensor network dataset demonstrate that the proposed method achieves an accuracy of 99.0%, a precision of 98.5%, a recall of 98.3%, an F1-score of 98.4%, and an ROC-AUC of 99.5%, outperforming existing baseline methods. By combining temporal modelling and representation learning within a unified framework, the proposed approach improves anomaly detection accuracy, robustness, scalability, and generalization capability while maintaining low false alarm rates in challenging wireless sensor network environments.