A general need for Wearable Robots (WRs) is to perform optimal selection of information for control and monitoring during daily assistance with an autonomous systems. A good feature selection algorithm is key to perform automated estimation of the mode of locomotion under environmental variations. Ambulatory body weight support (BWS) systems can be combined with WRs to provide safe ambulation and support in overground walking in cases of lower limb paralysis. This study aimed to develop a support vector machine (SVM) model for binary and multiclass classification that performs gait pattern recognition for different values of partial BWS during overground robot-aided walking. The principal component analysis (PCA) and kernel-based PCA (kPCA) were applied to improve the classification performance. As a result, the combination of temporal and kinematic features showed to improve the accuracy in the discrimination of gait patterns in healthy patients (88%). In SVM multiclass classification the 'one-against-one' approach showed to have a more stable performance (true positive and true negative rate are consistent) than 'one-against-all' approach and also lower computational cost both for training and SVM's decision making.