Deep Extreme Learning Machine With its Application to Body-Conducted-Sound-Based Handwork Recognition

Technology that can digitalize handwork in detail in real time should be developed to introduce collaborative robots, improve production efficiency, and take over skilled manufacturing and maintenance work in factories. For this purpose, we focus on body-conducted sounds because they are robust to visual occlusions and surrounding noise interference, can be acquired using wrist sensors, and may help recognize hand gestures as well as hand-contact objects. In this paper, we propose a novel modeling method for handwork recognition using body-conducted sounds. This method adopts deep residual learning with dilated causal convolution extreme learning machines (DRLDCC-ELM). The DRLDCC-ELM was compared with a transformer baseline model to identify 13 types of handwork. The experimental results confirmed that DRLDCC-ELM outperformed the transformer baseline model and stably obtained models with almost the same F1 scores despite the fact that the training dataset was not large.