Due to the complex internal structure and multiple interference factors for aero-engine， it is difficult to extract features of the blade cracks and diagnose crack fault accurately. Hence， the paper proposes a new improved deep belief networks （DBNs） for feature learning， which is used to fault diagnosis for the blade cracks based on three-dimension blade tip clearance （3-DBTC）. Firstly， based on the law of reconstruction error transmitting in hidden layers， the paper creates a new variant of DBNs to adaptively adjust the depth with global back-reconstruction （GBR） mechanism， aiming at avoiding feature degradation with an increase in depth of DBNs and then causing insufficient of the representation ability in features； Then， the variant of DBNs is adopted to adaptively learn the blade crack deep-level features from the multi-dimensional signal 3-DBTC involving characteristic information of the blade cracks； Finally， the softmax regression model is used to build the complex mapping between the deep-level features and the blade cracks， accomplishing the accurate diagnosis for the blade cracks. The experimental results demonstrate that the proposed method can fully mine characteristic information of the blade cracks， and effectively improve the diagnosis accuracy， yielding average accuracy of 98.43%， and its standard deviation is only 0.092%， showing its stability and generalization ability are also pretty good and achieving effective fault diagnosis for the blade cracks.