In modern society, various kinds of products are produced by machines where the gear mechanism relies on the teeth on the rim to mesh with other gears to transmit torque. Due to its high transmission efficiency, accurate transmission ratio, large power range, and the ability to change the speed or direction of movement, gears are widely used in machinery. In recent years, with the advancement of technology and the concept of unmanned factories, the use of machines to replace workers can not only reduce labor costs but also reduce human errors and dangers caused by long hours of work. Thus, it increases the production rate and stability.
The research on the fault diagnosis of the rotor system can be traced back to Randall [
1] where some common types of gear faults and frequency spectrum characteristics were introduced. Generally, the accelerometers or laser vibrometers were the most used for signal detection; and the spectrum distribution of the signal can be interpreted by the experienced technicians to determine the type of fault. However, human judgments sometimes make uncertain errors. In recent years, with the progress of science and technology, the computing speed of computers has increased dramatically, so some machine learning methods that require high-speed computing, such as convolutional neural networks (CNN), support vector machines (SVM) [
2], etc., once again attract everyone’s attention. Comparing CNN and SVM, the classification accuracy of CNN would be higher than that of SVM, especially for the complex systems like the compound fault diagnosis of rotating machinery. Therefore, the goal of this paper is to develop a CNN model to achieve the fault diagnosis of a gearbox; and with the combination of the local area network, the fault diagnosis can be predicted from the remote site.
The artificial neural networks (ANN) was first inspired by Hebbian theory [
3] proposed by Donald Olding Hebb, and after Rumelhart et al. [
4] proposed the back-propagation method to automatically correct the weights of the multilayer perceptron (MLP), the ANN method became concrete and began to be widely used. Although the multilayer perceptron relies on its huge internal variables and nonlinear functions to handle highly complex nonlinear problems, it cannot effectively learn the relationship between spatial data. And, this issue was effectively improved when LeCun et al. [
5] proposed CNN approach. Furthermore, Krizhevsky et al. [
6] used CNN method for image recognition to win the championship in the ImageNet competition held in 2012. After that, CNN related research began to flourish in various fields, including object recognition [
7,
8], portrait recognition [
9,
10], action recognition [
11,
12], etc. In addition to the use of the above-mentioned imaging field, neural networks are also used for fault diagnosis in the mechanical field. Chen et al. [
13] used graph of vibration signal as input of CNN network for fault diagnosis of gearboxes. Jia et al. [
14] used the frequency-domain signal, converted from fast Fourier transform (FFT), as input to the neural network for fault diagnosis of bearings and planetary gearboxes. The weighting matrix of the network was corrected by a specific way. In addition, Janssens et al. [
15] and Jing et al. [
16] also used frequency-domain signal as input to CNN for machine diagnosis. They all showed the spectrum signal was suitable for the extraction of fault characteristics. Zhang et al. [
17] used WDCNN (wide first-layer kernels deep CNN) method for fault diagnosis. In the first convolutional layer, they used a larger kernel for noise attenuation and used AdaBN (adaptive batch normalize) skill to improve its robustness. Zhao et al. [
18] converted the data into 2D images through wavelet-transform and used DRN (deep residual network) to diagnose faults. Furthermore, a dynamic adjustment layer was used, which can be automatically corrected during training process to find out the more obvious frequency characteristics. In contrast to accelerometer-based diagnosis, the motor encoder signal was used in the work of Jiao et al. [
19]. With some signal processing for obtaining angular speed and acceleration signals, the stitching data are fed to the one dimension CNN for fault diagnosis. Wu et al. [
20] used the raw vibration data as input of one-dimensional CNN for fault diagnosis of rotating machinery and much better performance than traditional methods was shown. Liang et al. [
21] converted the vibration data into a two-dimensional picture through wavelet-transform to perform compound fault diagnosis. The disadvantage of this method is that the conversion of 2D images would increase the computational cost during CNN training, and it will also increase the computational time during actual diagnosis.