To address the issues of low data quality and poor adaptability in deep learning methods for infrared image analysis in gearbox fault diagnosis, this paper introduces an enhanced deep prototype network model (MSPNet). This model employs a multi-scale strategy to improve fault diagnosis accuracy and algorithm generalization, especially with small sample sizes. First, infrared image data of six fault types under five operating conditions are collected using a rotating test bed. Gaussian noise is added to simulate real operating conditions. Next, the fault data are processed using a multiscale module to extract multiscale fault features and reduce feature value fluctuations. Finally, the proposed model is used to process the image data and is experimentally compared with five other algorithms. The experimental results demonstrate that the proposed method outperforms the other algorithms under various operating conditions.