Abstract: To monitor the tool wear state, an in-process tool wear state recognition system was developed using modified variational mode decomposition (MVMD), adaptive backtracking search algorithm (ABSA) and LS-SVM. To tackle problems of mode overlap and noise sensitivity in traditional signal processing methods, the instantaneous frequency mean judgment method was used to determine the optimal number of decomposed modes and the denoised variational mode decomposition (MVMD) was introduced to decompose the signal. To improve the optimization efficiency and adaptability, an adaptive backtracking search algorithm (ABSA) was proposed, which enhanced the global and local search ability of the algorithm through parameters adaptive selection. Based on ABSA, a multiclass model of LS-SVM was used to recognize tool wear state. The experimental results show that the MVMD can effectively reduce noise and eliminate false information and prove that ABSA has stronger ability of global exploration and local optimization, which makes the LS-SVM model optimized by ABSA get higher accuracy.