Research On Intelligent Assembly Method Of Aero-engine Rotor Based On Deep Learning

Aero-engine is a key benchmark to evaluate the level of national comprehensive science and technology,modern manufacturing industry and even the comprehensive national strength.The core engine is the "heart" of an aero-engine,which consists of multistage rotors stacked layer by layer.Therefore,the performance and service life of aero-engine depend greatly on the assembly quality of multistage rotor.Coaxiality and perpendicularity are important parameters for evaluating the assembly quality of multistage rotors assembly.Excessive error of coaxiality and perpendicularity of multistage rotors assembly will lead to engine faults.Therefore,it is urgent to establish the prediction models for coaxiality and perpendicularity of the multistage rotors assembly to guide assembly,improve assembly quality of multistage rotors and enhance engine stability.Based on the core machine rotor assembly process,this subject has carried out research on the multistage rotors intelligent assembly method based on deep learning.The correct classification of rotor profile is the premise of a single rotor machining error evaluation and multistage rotors coaxiality and perpendicularity prediction.In order to improve the efficiency and accuracy of rotor profile classification,the rotor profile classifiers based on support vector machine and convolutional neural network are established respectively.Since the evaluation of the machining error of a single rotor is the basis for the establishment of the multistage rotors coaxiality and perpendicularity prediction models,in order to improve the current tedious and time-consuming situation of rotor machining error,the rotor depolarization and declination method based on deep belief neural network is proposed in this paper.According to the characteristics of the assembly error transmission mechanism of different profile rotors,the multistage rotors coaxiality and verticality intelligent assembly prediction models are established.The main research contents include:First,in order to solve the problem that rotor profile classification relies heavily on expert experience and low classification accuracy,the intelligent rotor profile classification method based on SVM and convolutional neural network is proposed.A SVM classifier is established based on the sampled data of the rotor axial measurement surface profile.In order to reduce the amount of calculation and ensure the classification effect of profile classification,the PCA dimensionality reduction is carried out on the sampled data,and appropriate kernel function is selected for the features of the sampled data after dimensionality reduction.The optimal hyperparameters of the SVM rotor profile classifier is obtained by combining the grid search method and 10-fold cross validation method.It has been verified that the accuracy rate of the SVM rotor profile classifier is 99%,which meets the needs of subsequent assembly process.In order to avoid excessive accumulation of subsequent rotor profile classification samples,the SVM rotor profile classifier is difficult to implement for large-scale training samples.The sampled data of the rotor axial measurement surface profile is converted into image structure data,and the rotor profile classification based on the convolution neural network is established.Using existing samples to train the network,it is verified that the accuracy of the rotor profile classifier based on the convolutional neural network is 99%,which meets the needs of subsequent assembly process.Secondly,in order to solve the problems of low measurement efficiency and poor evaluation accuracy of single rotor machining error,the rotor depolarization and declination method based on deep belief neural network is proposed.This method realizes the prediction of single rotor machining errors by establishing the concentricity prediction network,the perpendicularity prediction network of the single-dip rotor,and a relative flatness prediction network of the saddle-surface.Experiments are carried out with different optimization algorithms and weight initialization methods,and the optimization algorithm and weight initialization method that are most suitable for each machining error prediction network are selected.The experimental results show that the coefficient of determination R~2 of the concentricity value prediction network is 0.99,and the mean prediction error is 0.1 ?m,the coefficient of determination R~2 of the perpendicularity value prediction network is 0.94,and the mean prediction error is 1.6 ?m,the coefficient of determination R~2 of the flatness value prediction network is 0.94,and the mean prediction error is 1.7 ?m.The prediction accuracy of each machining error prediction network meets the requirements of on-site assembly technology,which proves the effectiveness of the rotor depolarization and declination method based on deep belief neural network.Finally,the multistage rotors coaxiality and perpendicularity intelligent assembly prediction models are established to solve the problems of unclear assembly error transmission mechanism of core machine rotor assembly and low accuracy of coaxiality and perpendicularity prediction caused by various profile.The multistage rotors perpendicularity prediction model based on the transformation of the three-dimensional coordinate system is established by analyzing the machining error transfer mechanism of single-dip multistage rotors assembly.It has been verified by experiments that the prediction error of multistage rotors perpendicularity is 1.5 ?m,which meets the requirements of the assembly technology.For the saddle-surface multistage rotors assembly with complicated deformation mechanism of the mating surface,combined with the rotor assembly process,the BP neural network prediction models for coaxiality and perpendicularity of the multistage rotors assembly are established.Particle swarm optimization algorithm is used to obtain the optimal superparameter combination of coaxiality and perpendicularity prediction network.The mean prediction error of the PSO-BP coaxiality prediction network is 1.0 ?m,and the mean prediction error of the PSO-BP perpendicularity prediction network is 0.6 ?m.The prediction accuracy meets the requirements of assembly technology.Therefore,the PSO-BP coaxiality and perpendicularity prediction network established in this paper can provide technical guidance for the optimal assembly of multistage rotors and improve the quality of multistage rotors assembly.