Federated anomaly detection based on collaborative pre-trained generative models in smart manufacturing

To address anomaly detection challenge in smart manufacturing, this paper proposes an edge computing anomaly detection framework, termed the Federated DDPM-GAN framework for Anomaly Detection (FDGAD). This framework is designed to detect anomalies while preserving data privacy by leveraging the strengths of denoising diffusion probabilistic model (DDPM) and generative adversarial network (GAN) to enhance the capabilities of edge devices. The integration of GANs and DDPMs facilitates the generation of realistic synthetic data and also improves the model’s ability to detect subtle and complex anomalies in industrial environments. To further strengthen data privacy, differential privacy techniques are incorporated to ensure the confidentiality of sensitive data. Additionally, we developed a collaborative learning protocol to optimize overall anomaly detection performance. The goal of the protocol is to enable efficient interaction between federated learning processes and combined DDPM-GAN architecture. Extensive case studies conducted on three benchmark datasets demonstrate the effectiveness of the proposed FDGAD framework. while ensuring data privacy. Experimental results on five industrial datasets demonstrate FDGAD achieves 90.7 % F1-score and 94.5 % AUC, outperforming baseline methods by 3.5 % and 2.3 % respectively. The DDPM-based feature extractor reduces class overlap by 41 % compared to autoencoders, while the federated protocol maintains 92.1 % detection accuracy under 50:1 class imbalance. FDGAD proves its effectiveness in handling high-dimensional sensor data and privacy-preserving industrial applications.