Search

As an important component of trains, rolling bearing is always faced with the defection of shed oil, which inevitably threatens the train safety. Therefore, it is of great significance to conduct defection inspection on bearing shed oil. Due to the complex structure of rolling bearings, traditional signal analysis approaches cannot detect the defections of bearing shed oil with high-efficiency and low cost. In recent years, deep learning has achieved remarkable growth and been successfully applied to various computer-vision tasks. Motivated by this fact, we propose a two-stage attention aware method to recognize defections of bearing shed oil. The proposed method is based on convolutional neural networks, can automatically learn bearing defect features, and does not need manual feature design and extraction like traditional methods. The two-stage method cascades a bearing localization stage and a defection segmentation stage, to recognize the defect areas in a coarse-to-fine manner. The localization stage extracts the foremost bearing region and removes the useless part of images, so as to focus the attention of segmentation stage only on the target region. In segmentation stage, we propose a novel attention aware network APP-UNet16, to segment defect areas from extracted bearing region. APP-UNet16 stacks attention gates to enable the attention-aware features change adaptively, and thus can learn to focus on target defect areas automatically. We also utilize transfer learning in constructing the encoder of APP-UNet16, and introduce spatial pyramid pooling to connect the encoder and decoder, to improve traditional UNet. A series of comparative experiments are conducted, to compare our two-stage method with one-stage method which directly perform segmentation on original train images. The results indicate that the proposed two-stage inspection method achieves higher robustness and accuracy in recognizing defect areas with small oil spot. And the experimental results on proposed APP-UNet16 also demonstrate that a better segmentation performance is achieved, compared to traditional UNet and related state-of-art approaches. We will release the source code as well as the trained models to facilitate more research work.

Multivariate time series prediction, with a profound impact on human social life, has been attracting growing interest in machine learning research. However, the task of time series forecasting is very challenging because it is affected by many complex factors. For example, in predicting traffic and solar power generation, weather can bring great trouble. In particular, for strictly periodic time series, if the periodic information can be extracted from the historical sequence data to the maximum, the accuracy of the prediction will be greatly improved. At present, for time series prediction tasks, the sequence models based on RNN have made great progress. However, the sequence models has difficulty in capturing global information, failing to well highlight the periodic characteristics of the time series. But the this problem can be solved by CNN models. So in this paper, we propose a model called Multiple CNNs to solve the problem of periodic multivariate time series prediction. The working process of Multiple CNNs is analyzing the periodicity of time series, extracting the closeness and the long and short periodic information of the predicted target respectively, and finally integrating the characteristics of the three parts to make the prediction. Moreover, the model is highly flexible, which allows users to freely adjust the cycle span set in the model according to their own data characteristics. Tests on two large real-world datasets, show that our model has a strong advantage over other time series prediction methods.