As the core component of mechanical transmission system, planetary gearbox is widely used in various mechanical equipment. Due to the working environment of gearbox is usually more complex, long-time operation, coupled with the influence of temperature, lubrication and other factors, the gearbox gear is prone to pitting, broken teeth, wear failure, gearbox bearing part is prone to wear, bending and other failures. If the abnormality can not be found in time, the long-term operation will cause damage or the abnormality will accumulate gradually, resulting in irreparable consequences. Therefore, it is necessary to monitor the gearbox status in real time, accurately judge the fault type and location, and maintain it in time, so as to minimize the expected loss.
In gearbox fault diagnosis, it generally includes three parts: signal collection and processing, feature extraction and fault diagnosis. The vibration signal contains rich characteristics of the operating conditions of gearbox components, and is often used as the input information of various fault diagnosis models. The common signal processing methods include short-time Fourier transform, wavelet transform and empirical mode decomposition (EMD). For example, Li Heng used short-time Fourier transform to transform one-dimensional vibration signal into two-dimensional signal, and then input it to classifier to realize bearing fault diagnosis; Liu et al used EMD to transform non-stationary vibration signal of gearbox to obtain stationary sub signal for gearbox fault diagnosis, even if the working environment changes, it still has a good recognition effect.
Deep learning is a hot topic in recent years, and it is an important branch in the field of machine learning. Among them, deep neural network (DNN) can effectively extract deep-seated feature information of data through deep network structure and various nonlinear transformations, and has strong data expression ability. Convolutional neural network (CNN), stacked denoised autoencoder (sdae) and deep belief network (DBN), as representatives of deep neural network, have played an important role in feature extraction and recognition. They overcome the shortcomings of manual selection of data features and make the diagnosis process of mechanical equipment more intelligent. Chen Baojia et al. Used the powerful feature extraction ability and complex mapping representation ability of DBN to identify the fault signal. CNN is a typical deep neural network. Its powerful feature extraction ability makes it widely used in the field of image recognition. Some scholars have applied CNN to fault diagnosis. Chen et al. Used wavelet transform to transform vibration signal into two-dimensional signal containing time-frequency information, and then input it into CNN for fault diagnosis. Jing et al. Developed CNN to automatically extract fault features of signals. However, the above literature using two-dimensional CNN to process the fault signal does not maximize the advantages of CNN in processing the original vibration signal. Recently, one-dimensional convolutional neural network (1-dcnn) has been successfully applied to fault diagnosis. Zhou Qicai and others proposed 1-dcnn model, which used one-dimensional convolution pooling layer to process one-dimensional time-domain signal, and solved the problem of bearing and gearbox condition monitoring. Wu et al. Used 1-dcnn model to study the fault diagnosis of fixed shaft gearbox and planetary gearbox, and proved that it has strong feature extraction ability and classification ability. However, in the above methods, the method of using single channel signal input to CNN for fault diagnosis does not give full play to the ability of CNN to extract the vibration signal features, and cannot fully mine the equipment fault feature information.