Convolutional Neural Network (CNN) has been widely used in bearing fault diagnosis and many satisfying results have been achieved. However, nearly all the published researches only focus on the classification accuracy, without considering another equally or even more important characteristic, classification uncertainty. The target of this research is to bridge this gap and to provide a comprehensive understanding of the uncertainty for CNN. Three kinds of uncertainties from training data, training CNN and bearing operating load were defined as data uncertainty, algorithm uncertainty and working condition uncertainty respectively. After definition, three uncertainties were separated by variable-controlling method at first, and then quantitated by Gaussian process. Moreover, the influence on classification uncertainty from operating load was analyzed, which brought insight into how to determine an appropriate operating load to reduce uncertainty. A simple CNN was built as the test network and features were extracted from wavelet packet energy entropy and wavelet singular entropy. The effectiveness of the proposed method was verified on the Case Western Reserve University bearing dataset.