Aero engine is essentially the heart of an airplane. However, the high temperature and high pressure working environment of the aero engine can easily lead to fatigue cracks in turbine disks, and result in serious accidents. Therefore, early disk crack diagnosis is very important to guarantee safe flight of the airplane and reduce its maintenance cost, which, however, is challenging due to the difficulty in building a complex physical model under variable operating speeds. To tackle this problem, a novel deep convolutional neural network (CNN)-based method is proposed for early disk crack diagnosis. CNN, as one of the deep learning structures, can learn deep-seated features directly and automatically from the raw data without the need of physical model or prior knowledge. It shows the potential to deal with the challenge of early disk crack diagnosis. Since the proposed diagnosis method is signal-level, the collected vibration signals can be input into the CNN architecture directly without the need of feature extractor. In this paper, the vibration signals at both the beginning and the end of the test are used for training the CNN model, then the rest signals are input into the trained model as test data to diagnose when the incipient disk crack is generated. Experimental study conducted on the fatigue test of a real turbine disk has proved the effectiveness and robustness of the proposed method for early disk crack diagnosis. Meanwhile, comparison study with some state-of-the-art methods is also performed, and further highlights the superiority of the proposed method.
