Extreme learning machine oriented surface roughness prediction at continuous cutting positions based on monitored acceleration

Abstract

Surface roughness of mechanical parts plays an essential role in the practical application. Due to the complexity of the machining process, establishing a comprehensive theoretical model that can accurately simulate the real machined surface is challenging. Therefore, data-driven prediction methods are both practically useful and scientifically sound. This paper proposes an approach of using extreme learning machine (ELM) to predict surface roughness based on the monitored vibration signals, which can achieve the prediction of the surface roughness at any position of the machined surface, rather than using a single value to evaluate the processing quality under specific situations. To ensure the accuracy of the vibration intensity representation, the amplitude of monitored signals is compensated, accounting for varying positions and timings. The linear relationship between the spindle speed and acceleration is investigated to mitigate the impact of machining parameters on the monitored acceleration. Combined with grey relational analysis and mutual information coefficient, 18 features substantially linked to the surface roughness are selected from all extracted features and then used as the inputs for an ELM-based predictive decision-making system. The results indicate that the average and minimum errors of the developed method are 5.09% and 1.03%, respectively, demonstrating its accuracy and feasibility. In contrast to other machine learning models of SVR and ANN, the proposed method in this paper exhibits optimal predictive precision and efficiency, suggesting its potential application in the fast response of quality control.