Deep learning technique has dominated computer vision benchmarks in many industries for recent years. Numerous data-driven techniques have demonstrated significant promise in intelligent fault identification. However, the deep learning methods need a large annotated dataset to generate an explicit result, but the machinery bearing contains a small number of data samples. It is still challenging to collect a vast amount of bearing datasets due to cost and time constraints. To solve this problem, we have proposed a novel Fault Diagnosing Adversarial Network (FDAN) and a Deep Convolution Neural Network (CNN) that classify the bearing fault accurately. Firstly, the semi-supervised GAN generates more sample-bearing data to balance the training set with the Case Western Reserve University (CWRU) dataset. Secondly, the CNN is trained on the balanced dataset to classify the bearing faults into four categories such as inner race fault (IF), outer race fault (OF), roller fault (RF), and normal. The proposed FDAN and Deep-CNN techniques have better effectiveness and robustness than existing fault diagnosis methods.