Accurate aircraft engine diagnostics is essential for ensuring operational safety and enabling predictive maintenance under heterogeneous operating conditions. Although deep learning models can effectively capture high-dimensional multivariate sensor dynamics, purely data-driven approaches often entangle operating-condition variability with degradation-sensitive patterns, which limits robustness and generalization. This paper introduces PRIME, a physics-guided residual integrated framework for multi-task aircraft engine diagnostics. Rather than embedding explicit thermodynamic equations or physical constraints into the optimization process, PRIME relies on a physically motivated residual decomposition strategy that separates operating-condition-driven nominal behavior from degradation-sensitive sensor deviations. Specifically, nominal responses are estimated from operating-condition representations and subtracted from observed sensor signals to isolate fault-relevant residual patterns. These residual representations are then processed by a hybrid temporal architecture combining temporal convolutional networks and transformer-based self-attention, enabling joint modeling of local degradation signatures and long-range temporal dependencies. Within a unified optimization framework, PRIME simultaneously performs Fault Detection (FD), Fault Type Classification (FTC), and Health State Estimation (HSE). Extensive experiments on NASA C-MAPSS, N-CMAPSS, and the ALFA dataset show consistent and statistically significant improvements over strong baseline models. For FD and FTC, PRIME achieves gains of approximately 2%–4% over the strongest neural baselines evaluated under the same protocol, with larger margins over classical machine learning approaches. For HSE, PRIME yields more faithful degradation trajectories, leading to systematic reductions in estimation error across single- and multi-regime datasets. When Remaining Useful Life (RUL) is projected from the learned health trajectory through a threshold-based mechanism, the resulting estimates also improve substantially, with RMSE reductions of up to about 22% under complex operating conditions. These results show that physics-guided residual disentanglement improves robustness, interpretability, and multi-task diagnostic performance. More broadly, they support the view that HSE provides a useful latent degradation representation for downstream prognostic assessment, even though RUL is not directly optimized by the model.

The rapid emergence of sophisticated, dynamic, and rare or previously unseen attack pattern exposes fundamental limitations of conventional intrusion detection systems (IDS) based on static learning architectures. While deep learning (DL) models have demonstrated strong performance by capturing complex spatial and temporal traffic patterns, existing DL-based IDS largely rely on fixed decision structures, restricting adaptability to evolving threats. Furthermore, current hybrid DL-metaheuristic approaches typically use such metaheuristics as offline or auxiliary optimizers, without interacting with the deep model’s internal latent representations. This paper introduces a novel co-evolutionary IDS that establishes a tight, bidirectional coupling between DL and Particle Swarm Optimization (PSO) through latent-space-guided structural adaptation. A CNN-LSTM (Convolutional Neural Networks-Long Short-Term Memory) encoder learns discriminative spatial–temporal representations of network traffic, which dynamically guide PSO to select and optimize Adaptive Decision Blocks during training. Unlike prior hybrid methods, the proposed framework enables continuous co-evolution of both representation learning and decision structure, allowing the IDS to adapt its internal architecture in response to uncertain, rare, and previously unseen attack patterns. Comprehensive evaluations on UNSW-NB15, CICIDS2017, and ToN-IoT demonstrate statistically significant improvements over state-of-the-art DL and hybrid IDS approaches, achieving over 99.97% accuracy, recall and F₁-score, and low-latency inference suitable for near real-time deployment.