Edge-enabled AI framework for real-time anomaly detection in industrial IoT systems

This paper presents a Federated Learning-based CNN-LSTM Autoencoder model for anomaly detection in Industrial IoT (IIoT) environments. The model integrates CNN for spatial feature extraction and LSTM Autoencoder for capturing temporal patterns, enabling accurate and efficient anomaly detection. Employing federated learning, the method maintains data privacy by locally training models on edge devices. Tested on the WADI dataset, the model performed with 95.6% accuracy and an F1-score of 0.95, beating conventional techniques such as SVM and centralized CNNs. Latency during inference was drastically lowered to 14.7 ms, and the model was trimmed down to merely 3.2 MB in size, rendering it perfect for resource-limited deployments. Visual outputs—learning trends, confusion matrix, ROC curves, and comparative bar charts—validate the strength of the model. The system proposed guarantees real-time response and privacy, solving IIoT security challenges of prime importance. In general, it presents a scalable, decentralized solution to anomaly detection in industrial systems.