Machining Process Learning And On-site Monitoring Methods: Toward High-precision Machining Of Low-rigidity Parts

Low-rigidity workpieces are widely applied as the key components of aero-engines,such as integrated impeller,blade,shaft,etc.The machining quality influences the aerodynamic performance of aero-engines.The control of workpiece deformation depends on process modelling-based parameter optimization and regulation during the machining process.However,Mechanism-based process models usually deviate from the actual cutting condition.The prediction errors of cutting force,system vibration will lead to the inaccurate predictions of machining stability and surface geometrical error,which makes the applications of mechanism model-based parameter optimization and regulation strongly conditioned.Take the machining of XX15 compressor stator ring as an example,the stiffness of ring blade is impacted by the matrial removal of adjacent one.Therefore,it is difficult to predict the deformationinduced machining error through mechanism modelling,which makes it hard to control the surface profile accuracy.End-to-end data-driven models can integrate measured process quantities for learning and prediction of surface stability,surface topography,as well as force-induced error without mechanism hypothesis.Purely data-driven models usually need massive data to learn and characterize the process rule of complicated cutting conditions.However,as for medium-small batch production of curved surface parts in aero-engines,the process data are partly from repeated pipelines and limitedly available.Hence,a novel idea is to develop small sample and model fused learning approaches for prediction of stability domain and surface geometrical error,so as to regulate the machining process adaptively.Concerning this,our research focuses and results are presented as follows,1.Considering the coupling effect of tool runout,system vibration and deformation on the instant uncut thickness,as well as on the tool engagement,a comprehensive dynamic equation with nonlinear multiple delays is established to predict the vibration stability domain of cutting parameters.The dynamic equation is solved by the proposed derived node method with high computational efficiency and precision.Simulations indicate that the proposed derived node approach can accelerate the error converge of stability judgement.2.A novel chatter signal enhancement processing approach,combined with threshold learning and detection,are developed to realize automatic recognition of early critical chatter in real milling process.Plenty of simulated data are generated by using the dynamic prediction method in chapter 1.The real labeled data,accompanied by the simulatedly labeled ones,are fused to driven the incremental and decremental learning of stability lobs with the help of support vector classifier.The results indicate that the proposed chatter signal enhancement approach can help to find the critical chatter as early as possible.Moreover,those simulated data with correct labels can improve the generalization performance of leant stability lobes against limited real data,while those with mistaken labels cannot.3.In order to learn the complicated relationship between surface topography and cutting parameters,cutting forces,as well as system vibrations,an end-to-end deep convolution network is designed and established.The network architecture is optimized to enhance the prediction accuracy of surface topography.Besides,adversarial learning strategy is utilized to avoid the missing of weak high-frequency characteristics on the surface topography.Experiments are conducted to prove that the proposed deep network can balance the learning of surface characteristics in different frequency scales.The predicted and measured surface topography accord well,which makes it possible to on-site visualize and detect the machined surface quality.4.Purely data-driven models usually perform unsatisfactorily under the restriction of small amount of available samples.To overcome this problem,a mechanism-fused data driven regression is established to predict the forced-induced error.Based on the mechanism priors,a group of empirically dictionary mapping functions are constructed to fit the error distribution.To improve the prediction efficiency,the sparse Bayesian learning approach is adopted to mine the optimal sparse mapping between force-induced error and cutting parameters,cutting forces and cutting locations.Experimental results indicate that the mechanism-fused data driven model performs better than purely data-driven model on the prediction accuracy and generalization capacity.The proposed method has been successfully applied in the on-site monitoring and error compensation of XX15 compressor stator ring.