<p>Wireless Sensor Networks (WSN) are widely employed in various sensitive areas; however, due to their limited computing power, energy, and memory, these networks are exposed to signal-level anomalies as well as cyber-attacks. This study aims to introduce a metaheuristically optimized deep soft-voting ensemble that integrates Deep Neural Networks and CATBoost classifiers to enable secure intrusion detection. To achieve a seamless integration between convergence accuracy and exploration divergence, the authors have employed two complementary metaheuristics, Quadratic Interpolation Optimization (QIO) and Osprey Optimization Algorithm (OOA), for tuning the hyperparameters and decision thresholds. The system derives 23 operational and topological signal features that represent normal traffic and four major attack types, i.e., Blackhole, Grayhole, Flooding, and TDMA Scheduling. Rigorous preprocessing, such as Variance Inflation Factor analysis, supports that the features are independent and stable. The metrics used include accuracy, precision, recall, F1-score, specificity, and AUC, with the QIO-enabled model (DCQI) leading the way with the highest test accuracy of 95.62%.DCQI outperforms the baseline (DNCA) and the OOA-enhanced models, and thus, both class-level detection and balanced performance can be achieved. The feature sensitivity analysis based on the Cosine Amplitude Method and ranking both distinctly and consistently indicates Is CH, Who CH, and Dist to CH as the most influential predictors, which not only strengthens the understanding of WSNs but also shows the computational efficiency of the adopted security.</p>