Channel-attention 1D-CNNs for Real-time Collision Detection in Human-robot Interaction via Audio Spectral Features

초록

The deployment of collaborative robots has led to more frequent physical interactions with human collaborators, occasionally resulting in collisions. For collaborative robots in cooking applications, various vibrations are generated from human collisions and routine tasks, such as shaking a fryer basket. The challenge is that existing emergency stop mechanisms struggle to differentiate between these events, leading to frequent false activations. To solve this problem, this paper introduces an audio-based machine learning model for the real-time classification of vibrations into two distinct categories: benign vibrations from cooking operations and hazardous ones originating from human collisions. The system employs a 1D convolutional neural network (1D-CNN) with a channel-attention mechanism to classify vibration signals extracted from raw audio using the short-time Fourier transform (STFT). Events are categorized into three classes: no vibration, task-induced vibration, and human collision. To address the class imbalance, a hop-length adjustment strategy is applied, increasing the sampling resolution for minority classes. This technique outperforms other data-augmented methods, achieving 89.6% recall for human collisions and 95.8% overall accuracy. A 30 to 50 Hz bandstop filter is applied to suppress environmental noise, further improving robustness. Real-time evaluation on unlabeled continuous audio achieves 100% recall for human collisions and 89.1% accuracy, demonstrating strong performance under realistic conditions. The results highlight the method’s scalability, safety, and potential for practical deployment in human–robot collaboration settings using only audio data.

키워드