<p>Wireless sensor networks, or WSNs, are essential for communication in cyber-physical systems. WSNs have several applications, such as data transmission, identification of objects, and environmental monitoring. However, their Internet of Things (IoT) connectivity makes them highly vulnerable to cyberattacks. This research offers a new, end-to-end framework for preventing and detecting intrusions in WSNs. The framework introduces three key innovations: a newly proposed Multilayer Edge-Point Steerable Convolutional Attention Network (MEPSCAN) for accurately classifying multiple intrusion types by extracting fine-grained spatial features; a newly enhanced metaheuristic optimizer called the Enhancing Sparrow Search Algorithm (ESSA) for tuning hyperparameters and accelerating convergence; and a custom two-step data normalization method that improves robustness to heterogeneous and imbalanced WSN data. In addition, a Multi-axis Vision Transformer (MAViT)—adapted from advanced vision models—is employed for effective feature extraction by capturing both global and local contextual relationships. For intrusion prevention, the system includes targeted mitigation strategies such as packet verification, rate limiting, and scheduling reinforcement based on detected threats, thereby maintaining WSN integrity and resilience. On the dataset of WSN-DS, the proposed framework shows improved measures of accuracy (99.99%), specificity (99.99%), precision (99.99%), F1-score (99.99%), and recall (99.99%) in different attack scenarios compared to STLGBM-DDS, DL, Decision Tree, WOGRU-IDS, and MLSTL-WS. The WSN-DS dataset, comprising over 370,000 records with five classes, including blackhole, grayhole, flooding, scheduling, and normal traffic, is used as a benchmark due to its realistic simulation of diverse and complex WSN intrusion scenarios. This multi-disciplinary approach serves as an efficient solution to the problem of WSN security, offering a highly accurate and computationally efficient IDS solution.</p>