Abstract: To solve the problems of low interpretability of deep neural networks and that the existing interpretable networks can not accomplish the cross-domain diagnosis tasks, an interpretable triple feature extractor transfer network (ITFETN) is proposed. To solve the problems of the absence of interpretability, a multi-layer sparse coding model is built and the iterative solving algorithm is deduced for this model. Through unrolling fast iterative soft thresholding algorithm, the equivalent form of network for the algorithm is obtained and regarded as the feature extractor to construct an interpretable algorithm structure equivalent model. In addition, to solve the problems of cross-domain diagnosis, a strategy of triple feature extractor is proposed to extract the shared features from source domain and target domain and their respective private features, and the loss function of transfer learning tasks is designed based on the concept of feature adversary for effective training of ITFETN. The shared features from source domain and target domain with minimum distance are effectively extracted for cross-domain diagnosis to realize interpretable transfer tasks. The results of experiments show that the average accuracy and robustness of ITFETN in two cases of datasets have been improved, compared with the optimal comparing methods and ITFETN can accomplish the tasks of interpretable cross-domain diagnosis effectively.