• CdualTAL is a novel Transformer-based encoder–attention–decoder algorithm that separately processes strong and weak multi-domain features to avoid noise dilution.
• A correlation-adaptive feature selection algorithm identifies highly predictive strong features, while a dual-channel encoder handles them distinctly.
• A custom cross-attention mechanism fuses representations from both channels, and an LSTM decoder captures deep temporal dependencies for accurate wear prediction.
• CdualTAL outperforms 11 state-of-the-art methods, achieving an average R2 of 0.983 and an RMSE of 4.373 on tool wear datasets, demonstrating superior stability and accuracy.