Enhanced surface roughness discrimination with optimized features from bio-inspired tactile sensor

Abstract The ability of surface roughness discrimination is of great significance in the development of biomimetic robots and prosthetic limb as well as the research on humanoid tactile sensing. In our previous research, a bio-inspired tactile sensor was developed to discriminate different surface roughness. But its discrimination accuracy decreased greatly as the roughness became smaller. Furthermore, the discrimination accuracy declined when signals from an extra PVDF film parallel to the sliding direction were included. In this paper, enhanced surface roughness discrimination using the same bio-inspired tactile sensor was achieved, including: 1) discrete wavelet transform was applied at first to decompose the original sensing data into different scale and frequency components; 2) the most discriminative features were selected from as many as 80 features defined both in time and frequency domains based on components via the algorithm of sequential forward selection; 3) instead of kNN or SVM, extreme learning machine (ELM) was applied to discriminate surface roughness. It was found that the extra PVDF film parallel to the sliding direction could provide the feature that affects the discrimination accuracy the most significantly if its signals were decomposed and the feature was selected from the high order component. Furthermore, the neural network based ELM was shown to have better discrimination accuracy than kNN or SVM using the same dataset and features. With ELM, the discrimination accuracy was improved from 82.6% to 97.88%. The result indicates the importance of signal decomposition and feature selection for surface roughness discrimination based on the bio-inspired tactile sensor. The proposed framework could also be applicable to other researches in sensor development and signal processing.