Grinding is used to improve surface roughness and dimensioning precision in the metal industry. A large amount of heat is released during grinding. Most of this heat is transferred to the workpiece. In this case, a grinding burn may occur on the workpiece. Grinding burn is a significant issue in the production of bearings. If a burn occurs on the workpiece during grinding, the surface quality deteriorates and the internal structure and mechanical qualities of the material are adversely affected. Therefore, detecting grinding burn is critical for bearing manufacturers. In this study, during the grinding of the bearing parts, the machine conditions were changed in accordance with the real grinding scenario, and burnt and non-burned bearing data were obtained with the acoustic emission sensor. Then, time-frequency representations were obtained from these data with the continuous wavelet transform. These images have been classified in the GoogLeNet Network by transfer learning. Combinations of faulty/normal data processed under different machine settings and combinations of faulty/normal data processed with the same machine parameters were classified with the proposed method and compared. Faulty bearings processed with the same machine characteristics were detected with 100% accuracy using the proposed method. This percentage gives a reliable result for bearing producers. This study contributes to the literature in three ways: (a) It is based on data collected under real-world grinding situations. (12 different machine settings were employed.) (b) Various machine conditions were compared. (c) Faulty bearings were detected with high accuracy.