A Dual-Branch Incremental Learning Framework for Industrial Production Anomaly Detection
Xue Xu, Yucan Qiu, Qiaodong Jia, Qiang Zhu, Pan Zhu, Danpei Zhao, Xiaoying Shang, Nan Sheng · 2025
Efficient anomaly detection is crucial to achieve intelligent monitoring and quality control of industrial processes. Industrial anomaly detection remains challenging due to complex temporal dependencies, high dimensionality, and limited labeled anomalies. Existing mainstream approaches often fail to model non-linear dynamics or adapt to evolving anomaly conditions. To address these issues, we propose a dual-branch parallel architecture that integrates Transformer-based global attention with multi-scale convolutional feature extraction, effectively capturing both long-range dependencies and local variations. To enhance adaptability and reduce reliance on labeled data, we incorporate Generative Adversarial Networks (GANs) for data augmentation and adopt a teacher-student self-training mechanism to support incremental learning. Extensive experiments on benchmark datasets SMD, SWaT, MSL, SMAP and real-world cigarette production data demonstrate that our framework significantly outperforms mainstream methods with notable gains in precision, recall, and F1-score. These results validate the proposed approach's robustness and practical applicability for high-precision quality control in complex industrial environments.