The Affective Behavior Analysis in-the-wild (ABAW) 2022 Competition gives Affective Computing a large promotion. In this paper, we present our method of AU challenge in this Competition. We use improved IResnet100 as backbone. Then we train AU dataset in Aff-Wild2 on three pertained models pretrained by our private au and expression dataset, and Glint360K respectively. Finally, we ensemble the results of our models. We achieved F1 score (macro) 0.731 on AU validation set.