Task-oriented semantic communication towards industrial anomaly detection
Chang Lin, Puning Zhang, Ruyan Wang, Zhigang Yang · Journal of the Franklin Institute · 2025
In industrial production, accurately detecting defective products is crucial for ensuring product quality. Traditional anomaly detection methods utilize embedding from pre-trained models, which overlook the discrepancies between pre-trained and industrial image datasets. Additionally, existing methods fail to consider image transmission distortions caused by complex electromagnetic environments in industrial scenarios, which limits the accuracy of anomaly detection methods. To address these issues, we propose a novel semantic communication architecture for industrial anomaly detection, designed to operate effectively under low signal-to-noise ratio (SNR) conditions. This architecture integrates edge computing concepts and incorporates a domain-adaptive, high-precision anomaly detection network, along with a tailored semantic feature extraction model. Furthermore, based on software-defined radio (SDR), a semantic communication prototype system for industrial anomaly detection is built and compared with traditional methods on authoritative public datasets. The results demonstrate that our network achieves state-of-the-art performance with 98.5 % image-level AUROC and 98.3 % pixel-level AUROC on the MVTec AD benchmark. Additionally, industrial product images are transmitted with a compression ratio of less than 5 %, which enhances robustness under low SNR conditions.