A data-driven method based on deep belief networks for backlash error prediction in machining centers

Abstract

Backlash error occurs in a machining center may lead to a series of changes in the geometry of the components and subsequently deteriorate the overall performance of the equipment. Due to the uncertainty of mechanical wear between kinematic pairs, it is challenging to predict backlash error through physical models directly. An alternative method is to leverage data-driven models to map the degradation. This paper proposes a data-driven method for backlash error predication through Deep Belief Network (DBN). The proposed method focuses on the assessment of both current and future geometric errors for backlash error prediction and subsequent maintenance in machining centers. During the process of prognosis, a DBN via stacking Restricted Boltzmann Machines is constructed for backlash error prediction. Energy-based models enable DBN to mine information hidden behind highly coupled inputs, which makes DBN a feasible method for fault diagnosis and prognosis when the target condition is beyond the historical data. In the experiment, to confirm the effectiveness of deep learning for backlash error prediction, similar popular regression methods, including Support Vector Machine Regression and Back Propagation Neural Network, are employed to present a comprehensive comparison in both diagnosis and prognosis. The experimental results show that the performances of all these regression methods are acceptable in the diagnostic stage. In the prognostic stage, DBN demonstrates its superiority and significantly outperforms the other models for backlash error prediction in machining centers.