MSF-CNN: Multi-Scale Fusion Convolutional Neural Network for Tactile Acceleration Signal Classification

In recent years, with the rapid advancement of intelligent sensors and edge computing technology, material surface recognition based on tactile acceleration signals has become a hot topic in time-series signal classification. However, existing classification models encounter problems such as singlescale feature extraction, high model complexity, and excessive computational cost when handling long-sequence, complex, and multi-scale tactile signals. To address these challenges, this paper proposes a Multi-Scale Fusion Convolutional Neural Network (MSF-CNN) based on multi-scale convolutional feature extraction, attention-guided enhancement, and dynamic feature fusion mechanisms. The proposed model utilizes three parallel convolutional branches with different kernel sizes to effectively extract local, mid-scale, and global temporal features from tactile acceleration signals. Moreover, an attention gating module is introduced to adaptively enhance important channel features, and a dynamic weight fusion strategy is designed to achieve optimal combination of multi-scale features. Experimental results on the LMT108 tactile database demonstrate that the proposed MSFCNN outperforms baseline models such as single-scale CNNs, CNN+LSTM, and CNN+Bi-GRU in terms of accuracy, average recall, and $\mathbf{F 1}$-score. The experiments validate the effectiveness and advancement of the proposed model in tactile acceleration signal classification tasks, providing a novel solution that is highprecision, structurally efficient, and lightweight for tactile perception.