Abstract: Infrared thermal imaging technology, when combined with image processing techniques, enables long-distance, non-destructive, high-precision, and intelligent fault diagnosis and condition monitoring. Its application to the fault diagnosis and condition identification of axle box bearings—components with a high train failure rate—is therefore of significant practical importance. In this study, a train infrared thermal image dataset is utilized to train, validate, and test the single-stage deep learning object detection model YOLOv8, which incorporates a Convolutional Neural Network (CNN), for axle box detection. Subsequently, infrared thermal images corresponding to different bearing fault conditions of four axle boxes are simulated using COMSOL finite element analysis to generate an additional dataset. Three classification approaches—Bag of Visual Words (BoVW) with HOG Characteristic Gradient extraction, Support Vector Machine (SVM), and YOLOv8-based classification—are then employed for image classification and fault recognition. The results show that the YOLOv8-based classification model, which integrates a convolutional neural network for deep learning-based object detection, achieves the highest classification accuracy, reaching 100%. In comparison, the BoVW model attains an accuracy of up to 99.39%. In contrast, the combination of HOG features with SVM demonstrates relatively poor performance, with a classification accuracy of only 37.00%.