AbstractIndustrial Cyber-Physical Systems (ICPSs) are becoming more and more networked and essential to modern infrastructure. This has led to an increase in the complexity of their dynamics and the challenges of protecting them from advanced cyber threats have escalated. Conventional intrusion detection systems (IDS) often struggle to interpret high-dimensional, sequential data efficiently and extract meaningful features. They are characterized by low accuracy and a high rate of false positives. In this article, we adopt the computational design science approach to design an IDS for ICPS, driven by Generative AI and cognitive computing. Initially, we designed a Long Short-Term Memory-based Sparse Variational Autoencoder (LSTM-SVAE) technique to extract relevant features from complex data patterns efficiently. Following this, a Bidirectional Recurrent Neural Network with Hierarchical Attention (BiRNN-HAID) is constructed. This stage focuses on proficiently identifying potential intrusions by processing data with enhanced focus and memory capabilities. Next, a Cognitive Enhancement for Contextual Intrusion Awareness (CE-CIA) is designed to refine the initial predictions by applying cognitive principles. This enhances the system’s reliability by effectively balancing sensitivity and specificity, thereby reducing false positives. The final stage, Interpretive Assurance through Activation Insights in Detection Models (IAA-IDM), involves the visualizations of mean activations of LSTM and GRU layers for providing in-depth insights into the decision-making process for cybersecurity analysts. Our framework undergoes rigorous testing on two publicly accessible industrial datasets, ToN-IoT and Edge-IIoTset, demonstrating its superiority over both baseline methods and recent state-of-the-art approaches.