The tightening process has become increasingly automated, but the current quality diagnosis methods are insufficient to meet the necessary standards. To address the lack of efficient tightening diagnosis, this paper proposes a knowledge-embedded GAN. To reduce the need for data labels through the use of semi-supervised learning and improves the model’s generalization ability through knowledge embedding. Finally, the method’s effectiveness in improving fault diagnosis is verified using actual production line data. The results indicate that the method with knowledge embedding can diagnose the quality of the tightening process more accurately than the method without it.