Due to the large-scale growth of data, the storage scale of data centers is getting larger and larger. Hard disk is the main storage medium, once a failure occurs, it will bring huge losses to users and enterprises. In order to improve the reliability of storage systems, many machine learning methods have been widely employed to predict hard disk failure in the past few decades. However, due to the large number of different models of hard disks in the heterogeneous disk system, traditional machine learning methods cannot build a general model. Inspired by a DANN based unsupervised domain adaptation approach for image classification, in this paper, we propose the DFPTL (Disk Failure Prediction via Transfer Learning) approach, which introduce the DANN approach to predict failure in heterogeneous disk systems by reducing the distribution differences between different models of disk datasets. This approach only needs unlabeled data (the target domain) of a specific disk model and the labeled data (the source domain) collected from a different disk model from the same manufacturer. Experimental results on real-world datasets demonstrate that DFPTL can achieve adaptation effect in the presence of domain shifts and outperform traditional machine learning algorithms.