In this work, a facial action unit detector has been proposed. A free-position model to detect emotion and AUs on an input facial image is also provided. The Grad-Cam XAI shows the location of the detected AUs which is the same as the one expected by the FACS system. The proposed method has been validated on the public CK+ dataset and shows very promising results in terms of the explainability of the obtained results and performances, which are competitive in comparison with the state-of-the-art (an average accuracy of 93.4% over all AUs).