As the number of cyber hackers are continues to grow, novel intelligent security measures should be implemented to secure industrial control systems. This paper presents SecDL-Fuse which combines Transformer-Enhanced Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) for developing a hybrid deep learning framework that detects and deals with industrial cyber threats. The proposed framework is designed to achieve three key objectives: (1) Enhance real-time anomaly detection by leveraging spatial and temporal feature extraction through CNN and Transformer-based embeddings, (2) Improve threat classification accuracy by employing an adaptive LSTM-GRU fusion model that captures long-term dependencies in network traffic, and (3) Ensure scalability and robustness in industrial cybersecurity by integrating federated learning to maintain data privacy across distributed environments. An extensive validation of SecDL-Fuse occurred through benchmark experimentation utilizing datasets from NSL-KDD, CICIDS2017 and UNSW-NB15. The proposed approach reaches better accuracy, lower false alarm rate and computational savings when compared to CNN, LSTM, GRU, Bi-LSTM and Transformer-based architectures. The findings present SecDL-Fuse as a superior model which reaches peak accuracy performance of 7.8% higher than competing models while simultaneously decreasing false alarm occurrences leading it to emerge as a powerful security option against current industrial network threats