Wireless Sensor Networks (WSNs) are increasingly vital across domains, yet remain highly susceptible to malicious node attacks which can compromise network reliability and security. This paper presents an interpretable, energy-efficient dual-stage detection framework combining adaptive trust-aware decision trees at the cluster head with server-side hybrid deep and ensemble models. Suspicious cases are effectively elevated for advanced classification using CNN+ RF and AE+ LightGBM, while SHAP-based interpretability provides actionable transparency. A thorough analysis of the SensorNetGuard data indicates that our framework is superior to the previous ML/DL network robotics baselines in terms of performance and stability, imposing a smaller cost on resources, demonstrating an effective learning mechanism suitable for deployment in real-world WSN setups.