Intelligent fault diagnosis on mechanical transmission system of the train is of great importance to ensure the efficient operation of high-speed trains. However, due to the complex and changeable operation conditions (OCs), it is difficult to ensure the uniformity of the data distribution obtained by real-time acquisition, which sets a higher standard for the design of intelligent algorithms. Empowered by deep transfer learning (DTL) methods, the fuzzy inference system provides an effective solution to the abovementioned problem. In this paper, an Adaptive Pruning Fuzzy Inference System (APFIS) is proposed predicated on the fuzzy theory and the basic structure of fuzzy neural network, along with the hybrid model MSC-APFNN constructed by combining the multi-scale convolutional network with the proposed APFIS. This model can be employed as a backbone for the application to transfer diagnostic detection of rotating systems based on adversarial networks. Experimental results demonstrate that the proposed model achieves excellent diagnostic results under variable operating conditions, which provide a novel insight into the optimization of the traditional convolutional and fully-connected deep network structure and offer a promising avenue for the research of fuzzy decision-making in the context of fault diagnostics for high-speed trains.