Deep learning-based fault diagnosis of rolling bearings is a hot research topic, and a rapid and accurate diagnosis is important. In this paper, aiming at the vibration image samples of rolling bearing affected by strong noise, the convolutional neural network- (CNN-) and transfer learning- (TL-) based fault diagnosis method is proposed. Firstly, four kinds of vibration image generation method with different characteristics are put forward, and the corresponding pure vibration image samples are obtained according to the original data. Secondly, using CNN as the adaptive feature extraction and recognition model, the influences of main sensitive parameters of CNN on the network recognition effect are studied, such as learning rate, optimizer, and L1 regularization, and the best model is determined. In order to obtain the pretraining parameters, the training and fault classification test for different image samples are carried out, respectively. Thirdly, the Gaussian white noise with different levels is added to the original signals, and four kinds of noised vibration image samples are obtained. The previous pretrained model parameters are shared for the TL. Each kind of sample research compares the impact of thirteen data sharing schemes on the TL accuracy and efficiency, and finally, the test accuracy and time index are introduced to evaluate the model. The results show that, among the four kinds of image generation method, the classification performance of data obtained by empirical mode decomposition-pseudo-Wigner–Ville distribution (EP) is the best; when the signal to noise ratio (SNR) is 10 dB, the model test accuracy obtained by TL is 96.67% and the training time is 170.46 s.

Ferrography analysis(FA) is an important approach to detect the wear state of mechanical equipment. Ferrographic image recognition based on deep learning needs a large number of image samples. However, the ferrographic images of mechanical equipment are difficult to obtain enough high-quality samples in a short time due to the complexity and low efficiency of the ferrogram making. Therefore, the recognition method for small sample ferrographic images based on the convolutional neural network(CNN) and transfer learning(TL) is proposed. Based on the similarity of samples, the virtual ferrographic image set is designed as the source data of the pretraining model, the tested CNN model is constructed by using the TL. Based on the AlexNet frame, this paper studies the influence of the CNN internal factors including network structure, convolution parameters, activation function, optimization mode, learning rate and the external factors on the classification effect of test samples, and the L2 regularizer is added to solve the overfitting. According to the classification result of test samples, an optimal parameter combination is obtained to establish an intelligent recognition model of ferrographic images based on CNN and TL with the recognition accuracy of 93.75%. Moreover, the t-SNE is used to realize the wear particle recognition process visualization, which proves the effectiveness of the proposed algorithm. This work provides an effective way for the ferrographic image recognition of wear particles under small samples.

No abstract

No abstract

Ferrography analysis is one of main means to identify wear state of mechanical equipment, and its key is the intelligent recognition of wear debris ferrographic images. Ferrographic image acquisition is a complex and time-consuming work, so the direct deep learning cannot been carried out for the small tested samples. A virtual ferrographic image dataset is prepared firstly and then two-level transfer learning scheme is proposed to improve the identification rate of the tested samples based on the deep learning model trained by the virtual samples. A combined network of YOLOv3 and DarkNet53 is constructed, and the application effect of model is improved by two-level transfer learning of virtual dataset to open dataset and then open dataset to tested dataset, and the model errors before and after twice transfer learning are analyzed. The average identification accuracy of the model in the validation dataset is 86.1%, which is 44.5% higher than that without two-level transfer learning, and the average recall reaches 95.8%. The experimental results prove the proposed method have a high identification rate for the tested ferrographic images of an actual gearbox.