High-accuracy prediction and compensation of industrial robot stiffness deformation

Industrial robots (IRs) are promising options for machining large complex structural parts due to the higher flexibility, larger operating space, and lower cost compared with multi-axis machine tools. However, the rela-tively low posture-dependent stiffness and large stiffness deformation of IRs significantly deteriorate the contour accuracy of milling in which the cutting force is large generally. It is very complex to achieve a precise stiffness model and predict stiffness deformation of IRs because of the joint clearance, drift of zero-position, and other nonlinear factors. The conventional stiffness model of IRs only takes each joint as a constant linear torsion spring into consideration and ignores other difficult-to-model factors, which leads to low-accuracy identified results and thereafter induces deformation prediction errors. The data-driven approach can be used to obtain an accurate stiffness and deformation model, but a large amount of experimental data is required and it will cost enormous time and effort. In order to circumvent the experimental data deficiency and difficult-to-model issue, a simulation-driven transfer learning method named Adaptive Domain Adversarial Neural Network with Dual -Regressions (ADANN-2R) is designed for robot deformation prediction. Amounts of coarse deformation data, which are generated by the conventional stiffness model, are regarded as source data. And few real deformation data, which are obtained by deformation experiments, are regarded as target data. The Dual -Regressions are designed after the feature extractor, and the weighting parameters are adjusted adaptively to tackle the different magnitude of the regression loss and domain discrimination loss. The ADANN-2R aligns the simulated source data and real target data to perform adversarial training, and an accurate target deformation predictor is achieved. Experimental results indicate that the proposed ADANN-2R can obtain high-accuracy prediction with few real data compared with the conventional stiffness model. Compared with the path without deformation compensation and the pre-compensated path using the conventional stiffness model, the maximum position error of the pre-compensated path using the proposed ADANN-2R is reduced by 78.12% and 32.45%, respectively.

Automated summary

Industrial robots are useful for machining large complex structural parts due to their flexibility and low cost, but stiffness deformation affects milling accuracy. Traditional stiffness models lack accuracy due to ignored factors. A simulation-driven transfer learning method is proposed for accurate deformation prediction with minimal real data compared to conventional models.