In this paper, an improved method based on HOG-SVM (histogram of oriented gradient characteristic and support vector machine) is proposed for fault diagnosis. First, by converting mechanical vibration signals to 3-D (three dimensional) images, this proposed method can extract the T-HOG (improved HOG) feature of 3-D images precisely. With the optimal method, all characteristic information of mechanical vibration signal, including fault characteristic signal and health characteristic signal, are converted into characteristic 3-D image. Then, fault information can be accurately recognized though R-SVM's (optimal SVM) classification. Furthermore, the new method which is tested on two kinds of field tests, including rail and gear box fault diagnosis, has achieved high detection accuracy of 97.3% and 96.7% respectively. Finally, compared with other ML and signal feature extraction methods, the proposed method shows superiority in fault diagnosis, which is significant for industry safety and reliability.