Early and accurate detection of bearing fault is most essential for the safe and reliable working of industrial machinery units. The main problem of the traditional fault diagnosis method is manually extracting the features which requires the much experimenter’s experience; expert knowledge. Therefore, the shallow diagnostic model's classification rate does not produce good results. To address this issue, this research proposes a novel approach to detect and classify bearing fault using a convolutional neural network (CNN), which is capable of automating the extraction of features and remove the impact of expert expertise on the feature extraction process. A time-moving segmentation window is used to segment the vibration raw signal and the segmented signals are decomposed up to two levels using DWT. After that, decomposed signals are converted into gray scale images to train and test the proposed CNN model. To verify the performance of the model, CWRU bearing dataset and MFPT dataset are used. The proposed CNN model achieves the highest accuracy in terms of performance both under different load conditions as well as under noisy situations with varying SNR values. The experimental findings show that the proposed system is effective and extremely dependable in detecting bearing faults.