For the assembly of multi-stage rotors, this paper proposes the coaxiality predicting model of multi-stage rotors based on neural network. The model takes the complicated operation of centering and tilting during the measurement of single-stage rotor machining error and the indeterminacy of saddle surface error transmission mechanism during the process of multi-stage rotors assembling into consideration. First of all, the paper proposes the depolarization and declination model of single-stage saddle surface rotor based on deep confidence neural network. And then, the single-stage rotor machining error is taken as the input amount into the BP neural network to establish the coaxiality predicting model of multi-stage saddle surface rotors. Finally, experimental measurements of the level-four core engine rotor are performed to verify the accuracy of multi-stage rotors coaxiality prediction model. The result shows that the coaxiality of multi-stage rotors can be effectively predicted by the neural network. The average error of coaxiality prediction is 1.0μm, the standard deviation is 0.7μm, compared to the traditional method, the mean error and standard deviation decreases by 81.8% and 73.1%, respectively, which can reflect the advantages of the coaxiality prediction model of BP neural network.

When implementing the traditional assembly method, the rotor is affected by machining errors. The morphology of the rotor is complex, and the machining error of the rotors at all levels are transmitted step by step through the stop mating surface, which affects the performance and service life of the aero-engine. The evaluation of machining error of single-stage rotor is the basis of assembly quality of multi-stage rotor. In order to improve the current situation of complicated and time-consuming rotor machining error evaluation, this paper proposes to establish a deep belief neural network (DBNN) to replace the traditional procedure of depolarization. The network takes the relative evaluation error of the rotor profile data without depolarization as the input and takes the machining error of the rotors obtained after depolarization as the output. First, the evaluation mechanism of the rotor’s machining error is analyzed, and the corresponding machining error influence source is selected as the input source of the deep belief neural network. Second, as DBNN is trained, and the appropriate weight initialization method and the optimization algorithm of the prediction network are selected to ensure the optimization of the whole network for feature mapping extraction of the training set. Finally, the assembly of multi-stage rotors is simulated and analyzed. It is shown in the experiments that after the iteration, the prediction network, with good training effects, has converged, and its prediction results tend to be consistent with the real values. The mean prediction error of the concentricity is 0.09 µm while the mean difference of angle of concentricity error value is 0.77°, and the mean difference of perpendicularity error value is 0.21 µm while the mean difference of angle of perpendicularity error value is 1.4°, the corresponding R2 determination coefficients were 0.99, 0.98, 0.91, and 0.94, respectively. It meets the requirements of field assembly and fully embodies the effectiveness of the procedure of depolarization based on deep confidence neural network.