<p>Machine learning is advantageous for the online monitoring of tool wear conditions. However, current algorithms encounter limitations in proper extraction and application of features in high-dimensional data, which degrade accuracy and efficiency in the identification of tool wear states. In this study, a Kernel Principal Component Analysis-Sparse Principal Component Analysis (KPCA-SPCA) feature is proposed, which overcomes the limitations of conventional statistical models in managing high-dimensional nonlinear data. A Deep-Kernel Gaussian Process Regression (DKGPR) method is proposed for online tool wear monitoring, which integrates long short-term memory into radial basis function, extracts critical time-dependent features, and reduces sensitivity to short-term abnormal fluctuations. The performance of the DKGPR is compared with different machine-learning-based algorithms, and results show that the proposed algorithm has higher accuracy. The average Root Mean Squared Error (RMSE) of the DKGPR model is 1.686, which is 41.8% less than that of the conventional Gaussian process regression; the KPCA-SPCA reduces RMSE by over 55% and compresses the average confidence interval width by more than 70%. The KPCA-SPCA effectively improves the identification of multi-scale features and robustness to non-stationary signals, and the DKGPR is capable of learning via small-batch data. The combination of modified data-processing and machine-learning algorithms provides a highly efficient solution for tool wear monitoring.</p>