A dual-branch spatial-frequency fusion network for tool-wear prediction under limited training data

Abstract Reliable tool wear prediction is critical for achieving high precision and stable productivity in modern manufacturing systems. However, in real industrial environments, available data are often limited to a single sensor stream and only a small set of labeled training samples, which significantly constrains the accuracy of data-driven methods. To address these challenges, this paper presents a dual-branch spatial-frequency fusion network (DSFFN) designed for the small-sample regime. First, the Stockwell transform converts raw sensor signals into time–frequency representations that capture both transient and stationary components. A spatially separable module at the frontend decouples temporal and spectral processing and enlarges the receptive field with large kernels, reducing overfitting in small-sample training. A channel-group embedding module structures cross-channel interactions into group-wise subspaces and employs adaptive soft-thresholding to denoise while retaining wear-related features. The spatial-frequency aware fusion module integrates a low-frequency branch for degradation trend extraction with global-dynamic modeling branch for long-range dependencies. Then a multi-feature fusion attention mechanism adaptively emphasizes their complementary information. Experiments on real-world industrial data demonstrate that DSFFN achieves state-of-the-art accuracy and robustness, with the biggest gains at small training sizes.