Plunger pumps play a pivotal role in numerous large-scale equipment and engineering machinery. However, the working environment faced by plunger pumps is exceptionally complex, which not only increases the operational risk under certain conditions but also makes data collection extremely difficult. Currently, most fault diagnosis methods for plunger pumps are based on data collected under a single working condition, with relatively less research on fault diagnosis under varying conditions. This paper proposes hybrid transfer learning (HTL) for fault diagnosis of plunger pumps based on features and models, aimed at enhancing the generalizability of fault diagnosis models. Initially, a high-performance fault diagnosis model was constructed using Convolutional Neural Networks (CNNs). Subsequently, the Maximum Mean Discrepancy (MMD) function was introduced to achieve feature-based transfer learning. Finally, the paper undertakes model transfer, implementing state transfer learning that integrates both feature transfer and model transfer, thereby further improving the model’s generalizability. This allows the model to maintain robust performance when dealing with data from plunger pumps under different operational conditions. Compared to existing transfer learning methods, the approach presented in this paper demonstrates a greater advantage in enhancing model generalizability.