Abstract: Tool wear is one of the critical factors affecting the quality and efficiency of computer numerical control (CNC) machine tools. To address the current issues of signal singularity and insufficient prediction accuracy in milling cutter wear predictions, a novel method for tool wear prediction is proposed, which involves the fusion of multi-sensor information based on a stacked long short-term memory (LSTM) network with attention mechanism. Initially multiple sensor signals are preprocessed, subsequently multi-domain features are extracted. These features are fused at the feature level using Kernel Principal Component Analysis (KPCA), providing the input for the subsequent network. A stacked LSTM network model with attention mechanism is used to enable adaptive learning of crucial information, achieving a predictive accuracy of 99.9% on the PHM2010 dataset. Comparative experiments are conducted with other algorithms and incorporating artificial noise to verify the high accuracy and robustness of the proposed model.