Gearboxes with complex structures are key components in rotating machinery, and their failures exhibit both sporadic and coupling characteristics. However, in engineering practice, gearbox fault data are extremely scarce, and the sample size for composite faults is effectively zero. Therefore, it is of great significance to investigate fault diagnosis modeling methods under scarce fault samples and zero composite fault samples. This paper designs a new loss function for single single-fault samples and zero composite fault samples based on the prior knowledge of similarity between time series monitoring signals and the difference between single faults, as well as the prior knowledge that composite faults are composed of single faults. This enables the effective optimization of the feature encoder. On this basis, a diagnostic algorithm for single faults and composite faults of gearboxes is constructed, achieving effective diagnosis of various states of gearboxes. This paper validates the proposed method using two gearbox fault experimental platforms. The results show that the proposed method can construct and obtain the gearbox fault diagnosis model under single single-fault samples and zero composite fault samples.