Tool wear state identification based on CNN-LSTM-SCSA with continuous wavelet transform
Abstract
In mechanical manufacturing, tool wear directly affects machining quality and production efficiency. This study proposes a tool-wear-state identification method that combines continuous wavelet transform (CWT), coordinate-attention-weighted CWT (CA-CWT), a convolutional neural network (CNN), spatial and channel synergistic attention (SCSA), and a long short-term memory (LSTM) layer. The vibration signal from each sensor axis is divided into one-second windows and converted into 64 × 64 grayscale time–frequency images. In the implemented network, the CNN output has shape B × 128 × 8 × 8, is flattened to B × 8192, and is supplied to the LSTM as B × 1 × 8192. Because the sequence length is one, the LSTM is interpreted as a gated nonlinear feature-refinement layer rather than as a multi-step temporal model. On the fixed image-level validation split, the representative CA-CWT CNN-LSTM-SCSA run achieved 98.15% accuracy, 98.24% macro precision, 98.13% macro recall, and 98.16% macro F1-score. Across 24 archived runs using the same split and configuration, the corresponding mean ± standard deviation was 97.92 ± 1.35%, 97.98 ± 1.29%, 97.92 ± 1.35%, and 97.92 ± 1.34%, respectively. These results characterize within-tool, fixed-condition image classification and do not establish tool-independent or condition-independent generalization.