Multi-condition Wear Evaluation of Tool Based on T-SNE and XGBoost

On-line inspection of tool wear is a necessary function for future automated production. A good evaluation model can effectively improve machining quality and reduce economic loss. Based on the existing research, an optimized tool wear evaluation method is presented. Current signal and vibration signal of spindle in cutting process are used synthetically, which can effectively improve the deficiency of single signal analysis. Time domain, frequency domain and wavelet packet features are extracted from the collected signals, and the signal feature information is extracted as comprehensively as possible. T-SNE is used to reduce the dimension of features, and K-means is used to cluster various working conditions according to the similarity of features, which further improves the accuracy and generalization ability of model prediction. Finally, the integrated learning algorithm XGBoost is used as the estimator, and the model is evaluated by regression and classification. The results show that the problem of sample imbalance has less impact on the XGBoost algorithm. Compared with the traditional integrated learning algorithms such as random forest, XGBoost improves the prediction accuracy and reduces the prediction time by an order of magnitude. It is a more accurate and efficient tool detection algorithm and can be widely used in industry. © 2020 Journal of Mechanical Engineering.