Automobiles are increasingly transforming into user-centered intelligent service platforms, and driver identification models capable of perceiving driver identity in real time are essential for delivering personalized services. The identification performance of these models can be affected by variations in fatigue states. However, most existing models are trained using data collected under alert states, without considering the effects of fatigue. Therefore, this study proposes a driver identification method that adapts to fatigue evolution. First, multimodal data, including driving operation behaviors, facial behaviors, and fatigue ratings, were collected. Then, the Kruskal-Wallis test was employed to select discriminative driver identification features. Finally, by integrating driving data from different driving states, a driver identification model based on principal component analysis and a one-dimensional convolutional neural network was developed. The results show that the proposed model achieved average accuracy, precision, and recall of 88.72%, 89.97%, and 88.80%, respectively, for driver groups of 3, 9, 15, and 21 participants, representing increases of 23.34%, 22.53%, and 23.24%, respectively, over the conventional model. The proposed method effectively improves the driver identification accuracy and the adaptability to fatigue evolution, providing support for accurate driver identification and the development of personalized driving systems.