A mutual cross-attention fusion network for surface roughness prediction in robotic machining process using internal and external signals

Abstract

Compared with machine tools, industrial robots exhibit low, position-dependent stiffness. This dynamic compliance leads to inconsistent surface roughness under identical machining parameters when the robot configuration changes, thereby significantly complicating roughness prediction. Therefore, to address the challenge of predicting surface roughness in robotic machining processes and provide reference for its effective surface roughness monitoring, this paper proposes a Mutual Cross-attention Fusion Network (MCFN) for surface roughness prediction in robotic machining process using internal and external signals. Firstly, the machined surface roughness data set is obtained through the robotic machining experiments with different workpiece placements and machining parameters. The internal torque signals and external vibration signals of the robot are acquired to better reflect the state information during the machining process. Secondly, Uniform Manifold Approximation and Projection(UMAP) is used to reduce the dimension of time domain, frequency domain and time-frequency domain features extracted by signal channel to reduce the interference of redundant features. The features after dimension reduction are used to form a double-branch structure, and the dynamic interaction between different channels features is realized by Parallel Multi-channel Feature Enhancement Module(PMFEM). Then, the mutual fusion module based on the Dual Multi-head Cross-attention Mechanism(Dual-MCM) is used to realize the collaborative interaction of cross-modal information, to complete the bidirectional deep collaborative representation between the robot internal and external signals features in the fusion process. And the features are segmented and aggregated to predict the robot machined surface roughness. Finally, based on the performance evaluation index, the effectiveness of the MCFN is verified through hyperparameter adjustment, ablation experiment, comparison experiment of different dimension reduction techniques and data-driven methods. The verification results show that MCFN can realize the prediction of robot machined surface roughness at different postures and machining parameters, which provides an effective method for the accurate prediction and monitoring of robot machined surface roughness.