摘要
This letter proposed a robust computational framework for object recognition based on a force sensor array. To eliminate weight-based interference, a normalization strategy was implemented, forcing models to prioritize spatial geometric features. A comprehensive benchmark was established by evaluating 33 machine learning methods alongside a customized lightweight convolutional neural network (CNN) optimized for tensor-formatted data. To address the challenge of limited sample variations, a ten-fold data augmentation pipeline with spatial translation, rotation, interpolative scaling, and Gaussian noise injection was developed to expand the training set. Experimental results across four scenarios demonstrated that although typical models showed significant performance deteriorations, the proposed CNN achieved superior generalization with consistent 100% accuracy. This research demonstrated that data augmentation and deep representation learning can effectively compensate the hardware limitations of low-resolution responses, providing a definitive benchmark for future intelligent sensing.