Deep learning models for fault diagnosis demand high-end hardware, making direct implementation on edge devices challenging. Model compression is thus essential for efficient deployment. Knowledge distillation, a model compression technique, offers adaptability and has emerged as a research focus. Current studies on knowledge distillation mainly address fault diagnosis in bearings and motors, but research on gearbox-specific models is limited. Given the distinct physical characteristics and fault patterns of gearboxes and bearings, knowledge distillation approaches should be tailored to each type of device's fault diagnosis models. In this study, we propose a methodology for compressing gear fault diagnosis models using Convolutional Neural Networks in edge computing scenarios. This approach employs knowledge distillation to reduce the teacher model size, evaluating its effectiveness in terms of model size (computational parameters and storage) and performance (fault diagnosis accuracy and model complexity). Our findings indicate that the distilled model achieves nearly 80% compression while maintaining 91.48% accuracy, only a 7.62% decrease from the teacher model. This study demonstrates knowledge distillation's potential for compressing gear fault diagnosis models, offering a systematic approach for such compression efforts.