The data-driven intelligent diagnosis of rolling bearing status suffers from low recognition rate due to the poor quality of learning samples in the process of identification model construction. To address this problem, a method is proposed to improve the recognition rate of the rolling bearing intelligent diagnosis model by selecting learning samples using the deep belief neural network. First, aiming at the time-varying modulation characteristics of the rolling bearing vibration signal, a finite number of intrinsic mode function components were decomposed using variational mode decomposition, which represents different components of the original signal. Secondly, according to the fault energy fluctuations and correlation of each component, the proportion of fault information in each intrinsic mode function components is quantitatively evaluated. On the basis of this, the vibration signal is screened and reconstructed to obtain learning samples. Finally, the obtained samples are used as the input of the deep belief network, and the fault identification model of the rolling bearing is constructed accordingly. Experimental results show that the proposed method is capable of screening out the intrinsic mode function components of the rolling bearing vibration signal, which contains the main components of the fault and realizes the construction of learning sample sets. Moreover, it improves the fault recognition rate of the rolling bearing state identification model based on vibration data.