Research On Terrain Classification Methods For Wheeled Robots Based On Vibration Signals

In order to explore in the planets’(such as the moon and Mars) surface and work in thedangerous environment (such as desert, marsh, the scene of the fire, nuclear radiation area, etc)of the earth’s surface, autonomous mobile robots should be able to independently identifyenvironment, complete the mission without a dangerous situation. Terrain identification orterrain classification is an important part of environmental identification. Correspondingcontrol strategy is necessary for robot to travel on different terrain safely and effectively,when the terrain changes, autonomous mobile robot must be able to adapt to the terrain whereit is traversing. Terrain classification can solve the issue of trafficability of autonomousmobile robot in complex terrain. It is very important to improve robot autonomous mobileperformance.Based on in-depth analysis and synthesis of similar studies home and abroay, the theoryand techniques are researched from the two aspects, i.e. terrain classification featureextraction and classification method.In this dissertation，experiments for data acquisition are designed. The experimentalplatform is a four-wheeled mobile robot on which arm accelerometers in x, y, zdirections and a microphone in z direction are installed in left front wheel. When the robotis traversing respectively on sand, gravel, grass, soil and asphalt terrain with six differentvelocities, the acceleration and sound pressure signals of wheel-terrain interaction arecollected as the original data.By time domain amplitude analysis of original data, several parameters of amplitudedomain are selected as the terrain features for each sensor data. To the conventionalk-nearest neighbors (kNN) algorithm, it is necessary to deal with the choice of k, and nowthere is no best decision strategy for the situation when number of votes is the same in theprocess of multi-classification based on voting decisions, though a practical strategy selectedis random method, which is reduces classification accuracy. To the two problems, animproved kNN method was proposed, i.e. the choice method of k was proposed and kNNcycle optimization method was also investigated to deal with the problem that more than twokinds (including two) of terrains have the same number of votes.To the conventional probabilistic neural network (PNN) method, there is a problem aboutthe estimation of smoothing factor σ which is important to improve the classification accuracy. Previous scholars considered that the same σ was chosen for all samples orsamples of the same dimensions, but it could not make sure that the σ was the best orsub-optimal for all test samples, even there was no result. For the problem, an improved PNNmethod was proposed to deal with the choice of smoothing factor σ by iterativeoptimization method.The traditional one-against-one support vector machine (SVM) method, now there is nobest decision strategy for the situation when number of votes is the same in the process ofmulti-classification based on voting decisions, an improved one-against-one SVM methodwas rendered to deal with the problem that more than two kinds (including two) of terrainshave the same number of votes based on two-classification program of LIBSVM. Yetimproved kNN，improved PNN and improved one-against-one SVM methods were comparedin terms of classification accuracy and data processing time.In the field of fault diagnosis, a method based on singular value decomposition of trackmatrix of attractor reconstructed by time series is always used to reduce the noise in originalsignal. Based on singular value decomposition (SVD), a feature extraction method wasproposed using the fore several singular values of track matrix of attractor reconstructed byvibration signals time series as eigenvalues, and better classification effect was achieved.Feature extraction methods based on the fast Fourier transform (FFT) and the power spectraldensity (PSD) were studied, and both feature selections methods were described. Yet the threefeature extraction methods were compared in terms of classification accuracy and dataprocessing time.Based on measured data, the proposed methods have been validated by correspondingclassification experiments.