A new cause-mechanism independence estimation based cross-domain learning method for machining deformation prediction

Abstract

Monitoring data-based machining deformation prediction is fundamental for accurate deformation control and product quality guarantee. For problems where involved unobservable variables like residual stress that can lead to data distribution bias, causal cross-domain learning methods have prominent advantages over other pure data-driven methods by shifting cause distributions and mechanisms. However, existing causal methods are based on the hypothesis that cause and mechanism are independent, which ignores the corresponding changes of mechanism across domains and can limit accuracies. This paper proposes a new causal cross-domain learning method based on cause-mechanism independence estimation, where the hypothesis is broken by taking the dependence of cause and mechanism into consideration. A cause-mechanism independence estimator is established by introducing the structural integral of mechanism derivative multiplies cause distribution, and the estimation value can measure the cross-domain changes of mechanism. As a result, the proposed method based predicting model can make efficient distribution shifts according to the estimation. The machining of aero-engine casings is taken as a case study, and experimental results show that the proposed method could predict the deformation well with limited target domain data. Besides, the proposed method can be readily extended to other cross-domain regression problems involved with unobservable variables.