Accurate prediction of disk failure is of great significance to the stability of data center. But in the real production environment, the number of failed disks and normal disks in data centers is extremely imbalanced. Therefore, when artificial intelligence is used to predict disk failures, this paper improves the focal loss function by introducing the sample weight w, which reduces the weight of normal samples far away from the center of failed samples in the loss function, thus improving the performance of the prediction model under extremely unbalanced datasets. Finally, based on real datasets, experiments are carried out to verify the effectiveness of the proposed method in extremely imbalanced datasets.