Fault diagnosis has great influence on mechanical equipment and even the whole industrial system. Applying inverse problems to fault diagnosis is a new challenge. This paper proposes a method for fault diagnosis of rolling bearings based on TubeNet. Firstly, the vibration signal feature of rolling bearing is extracted and used for the training of a TubeNet. The trained TubeNet achieves high accuracy in test. Lastly, fault feature is calculated from fault categories by trained parameters. The error of the fault feature in reverse calculation within an acceptable range verifies the reliability of the TubeNet further. This application makes the fault diagnosis results of neural network more convincing.