Research On Terrain Classification For Robots Based On EMD And ELM

When the mobile robot needs to execute remote detection or dangerous task that humans can't do,it requires enough autonomous ability, which relies on recognition and detection ability for the terrains, so the research on terrain classification for mobile robot becomes very meaningful. Traditional method based on vision is widely used in the terrian identification and classification, which is easily influenced by illumination, surface covering,and is difficult to identify soft terrians, So this paper adopts the method based on vibration signal. Different from non-contact method, the vibration signal can reflect the bearing layer information of terrains, and is an important supplement to the visual. The research work will be carried out from three aspects: The first is to present the feature extraction methods; The second is to improve extreme learning machine; The last is to improve the fuzzy integral algorithm and uses the improved algorithm to fuse the extreme learning machine classifiers.In order to get the real vibration signals between robot wheels and terrain, this paper set up a data acquisition experiment system. A mobile robot that installed a three-direction accelerometer and a microphone in z direction on the four wheel arms respectively is used to acquisite the vibration signals of wheel-terrain interaction to classify the terrains by traversing on sand,gravel,grass,soil and asphalt terrains with five different velocities (This paper adopted the vibration signals of the left front wheel, and the rest for later research). The experiment provided data support for the latter research of classification algorithm in chapter three to five.This paper first applied Empirical Mode Decomposition (EMD) to the original signal and got a series of Intrinsic Mode Functions (IMFs). Because different IMFs contain different amount of information, the paper analyzed signal energy distribution and degree of distortion for selection, and further extracted Time Domain Amplitudes (TDA), Power Spectrum Density (PSD) and Singular Value Decomposition (SVD) features of IMFs selected.To the redundancy in hidden layer nodes of extreme learning machines (ELM) that will deteriorate classification stability and accuracy, a subtractive method of hidden layer nodes based on nodes correlation assessment has been proposed to classify five terrains at five velocities using three feature extraction methods.To the problem that ? calculation process is too time-consuming when solves fuzzy integral, the paper has presented simplified calculation method of ?; To the problem that integral function is uncertain, the paper has put forward a new integral function and given out the value of ? and ?. Then fuses two and three single improved ELMs respectively based on improved fuzzy integral.The proposed algorithms have been validated by corresponding experiments.