Prediction and evaluation of surface roughness with hybrid kernel extreme learning machine and monitored tool wear

Abstract

In the modern manufacturing industry, surface roughness is a critical parameter to characterize surface quality. The accurate prediction of surface roughness is of great significance for data-driven intelligent manufacturing. However, it's hard to accurately predict surface roughness in the complex machining process, because of the existence of some uncontrollable factors, such as tool wear. To address the aforementioned issue, a novel hybrid kernel extreme learning machine with Gaussian and arc-cosine kernel function (RBF_Arc_HKELM) was proposed to predict surface roughness. Then an optimized whale optimization algorithm was introduced to improve the prediction accuracy. Considering that tool wear is an indirect quantity and changes dynamically with the cutting process, a novel tool wear monitoring framework with attention mechanism, weighted feature averaging, and deep learning models was proposed. Afterward, the basic cutting parameters combined with the monitored tool wear were fed into the trained RBF_Arc_HKELM model for surface roughness estimation. Finally, surface roughness was evaluated by the established RBF_Arc_HKELM model. To verify the validity and performance of the established models, milling experiments were conducted under different cutting parameter combinations and tool wear levels, and some other intelligent algorithms were also used for surface roughness prediction and tool wear monitoring. Compared with kernel extreme learning machine with RBF (RBF_KELM), support vector regression (SVR), Gaussian process regression (GPR), and light gradient boosting machine (LightGBM), in terms of mean absolute error (MAE), the prediction accuracy of RBF_Arc_HKELM is improved by 17.82 %, 15.36 %, 14.16 %, and 6.26 %, respectively. These results indicated that the proposed model has great leverage in validity and accuracy. Moreover, compared with the measured tool wear as the input, the satisfactory prediction results of surface roughness were also obtained with the monitored tool wear as the input of the RBF_Arc_HKELM model with a drop in prediction accuracy of only 8.66 %.