The features of nonlinearity and non-stationarity in real systems are often difficult to be extracted. This paper focuses on developing a Convolutional Neural Network (CNN) to obtain features directly from the original vibration signals of a gearbox with different pinion conditions. Experimental data is used to show the efficiency of the presented method. Support Vector Machine (SVM) is utilized to classify feature sets extracted with 1D-CNN. The obtained results show that the features extracted in this method have excellent quality for fault classification without any additional feature selection.