The cross-domain diagnosis method based on transfer learning (TL) provides a solution for fault diagnosis of the aircraft fuel pump. However, due to the fuel pump working in harsh conditions, unknown faults or similar faults that are difficult to distinguish are likely to occur. The diagnosis results determine whether effective maintenance measures can be taken. To overcome these weaknesses, in this article, a class adaptive TL network (CATLN) is proposed for detecting similar and unknown faults of aircraft fuel pumps. First, according to the characteristics of the collected vibration signals, a multibranch feature extractor is constructed to extract multiscale features in time domain and frequency domain. Second, a class adaptive method is designed to reduce the interclass distance of the same class by dividing the core area of the sample, and increase the interclass distance of different classes to distinguish similar faults and unknown faults. The fault data collected in the ground experiment of the fuel pump are used for transfer tasks. The results demonstrate the effectiveness of the CATLN in the cross-domain fault diagnosis task containing similar or unknown fault classes.