On the Object Recognition Via Free-Fall Impact Vibration Signals

Object recognition through physical interaction provides an alternative sensing approach for robotic systems and intelligent devices when visual information is unavailable or insufficient. This paper investigates object recognition using free-fall impact vibration signals. We collect vibration data using an inertial sensor module measuring acceleration and angular velocity during object impacts on a spring-mounted wooden platform, and evaluate our approach on five datasets with increasing difficulty: 10 objects with distinct weights (Set 1), 15 objects with similar weights but different shapes (Set 2), 20 highly similar screwdriver bits (Set 3), a unified 45-object dataset, and an extended 60-object dataset. We propose a Temporal Convolutional Network (TCN) architecture and examine various approaches to process acceleration and angular velocity signals along the vertical axis. Session-based 10-fold cross-validation, where each fold corresponds to a complete data collection session, is used to ensure evaluation of real-world generalization capability. Experimental results demonstrate that TCN significantly outperforms baseline architectures (1D-CNN, CNN-LSTM, SVM) across all datasets and experiments. On the challenging unified dataset, single-sensor TCN achieves 94.27% accuracy with angular velocity data, while dual-sensor fusion achieves 94.51%. TCN maintains robust performance from Set 1 to Set 3, demonstrating practical applicability even with highly similar objects. Remarkably, TCN achieves this performance with the parameter efficiency (0.6M parameters for single-branch, 1.2M for dual-branch), outperforming 1D-CNN dual-branch (8.3M parameters, 92.80% accuracy) with much fewer parameters. Analysis also reveals that angular velocity consistently outperforms acceleration by 1.71% to 15.69% depending on the model and validates the effectiveness of TCN for impact vibration classification with moderate training data. The method demonstrates robustness across signal lengths (1-2s optimal) and maintains low cross-session variance (1-2%), indicating stable performance across environmental conditions.