Intelligent tool wear monitoring approach in milling of titanium alloys

• This paper addresses the wear behavior of ball-end milling cutters and introduces a BWO-SVMD signal processing method. An MSCNN-LSTM-Attention model was developed for tool wear monitoring. In titanium alloy milling with micro-textured coated cutters, vibration signals and wear data were collected, wear stages identified, and signals processed. The model reliably evaluated tool wear, demonstrating its applicability. • The main conclusions are as follows: A BWO-SVMD signal processing method is proposed, which, through a bimodal feature selection criterion (correlation coefficient–variance contribution), effectively balances noise separation and component retention. An MSCNN-LSTM-Attention model is developed for tool wear monitoring, enabling robust feature extraction. Experimental results show an average accuracy of 96.25 %, a computation time of 320.64 s, and an average Kappa coefficient of 0.9686. • The innovation of this study lies in that the use of SVMD avoids the problem that the number K of modal components is difficult to determine. This wear monitoring framework innovatively combines MSCNN, LSTM and attention mechanism to collaboratively extract wear features. • This research provides a reference for the identification of tool wear in actual engineering machine tools. Tool wear exerts a critical influence on machining stability and workpiece quality, making its accurate, intelligent monitoring indispensable for preventing tool failure and ensuring product consistency. Although direct assessment via wear imagery is possible, it requires interrupting the machining process and thus is impractical for real‐time production. A more viable solution is to leverage in‐process signals—such as vibration—to enable continuous monitoring. Here, we present a Signal processing method that Beluga whale optimization‐Successive variational mode decomposition (BWO‐SVMD) for noise suppression, followed by the S‐transform to produce high‐resolution time–frequency representations. Based on these denoised spectrograms, we develop an intelligent monitoring model that integrates a multi‐scale convolutional neural network (MSCNN), long short‐term memory (LSTM) units, and a channel–spatial attention mechanism. Experimental results demonstrate that our model achieves 96.25 % classification accuracy, a Kappa coefficient of 0.9686, and a total computation time of 320.64 s. Compared with CNN‐LSTM‐Attention, MSCNN‐Attention, and MSCNN‐LSTM baselines, it improves average accuracy by 1.89 %, 8.02 %, and 6.67 % and Kappa by 0.0732, 0.1374, and 0.2009, respectively. Although training time increases by 10.2 %–14.2 %, the substantial gains in predictive performance justify the additional computational cost.