Analysis on significance of various statistical texture features in vision-based surface roughness prediction in end milling process

The aim of this study is to identify the significant statistical texture features involved in surface roughness characterization. The vision-based approaches for assessing the surface roughness involve features extracted from the sampled surfaces. Conventional models use all the extracted features in prediction, thereby resulting in computational complexity and larger execution time. Hence, the development of an effective method which reduces the feature space to optimum is attempted in this work by conducting CNC end milling experiments on aluminum 3025 alloy material. Taguchi’s orthogonal array of L 27 is used along with the cutting parameters spindle speed, feed/tooth and depth of cut in designing the test samples. The image acquisition is done by CMOS sensor and the statistical features extracted from the machined surface image after preprocessing are optical roughness value, mean, standard deviation, skewness, kurtosis and entropy. Significant features influencing the surface roughness are identified by applying Analysis of Variance on the extracted feature data set with p-value less than 0.01 with 99% confidence level. Multilayered feed forward back propagation Neural Networks are trained using the cutting parameters and the extracted features to predict the surface roughness. The performance of the ANN architecture 5-3-1 with 3 hidden layers gave better results. For checking the suitability of the proposed methodology, in terms of prediction accuracy, confirmation tests are performed. The proposed approach gave a better prediction efficiency of 90.65% compared with 85.82% from conventional approach. The presented approach can be attempted in situations where the number of features involved in conventional approaches are more, to optimize the feature size based on significance and to improve the prediction efficiency.