Bearing fault diagnosis method based on convolutional neural network and DS evidence theory:

With the rise of intelligent manufacturing, the demand for intelligent fault diagnosis is becoming more and more urgent. The traditional fault diagnosis method relies only on single sensor data for fault diagnosis, which will have the problems of being vulnerable to the interference of noise and low accuracy of fault diagnosis. Therefore, this paper proposes wide large CNN-SA-IDS (WLCNN-SA-IDS), a multi-sensor data fusion bearing fault diagnosis model based on convolutional neural network and improved Dempster-Shafer (IDS) evidence theory. By combining large convolutional kernel and self-attentive mechanism of convolutional neural network to automatically extract features and perform diagnosis. The final result is used as the basic probability assigment (BPA) of the IDS to achieve "end-to-end" intelligent fault diagnosis. Multiple sensor data fusion through improved DS evidence theory, improved Pignistic probabilistic distance and iterative replacement of the least credible evidence BPA. Finally, the experimental validation on CWRU bearing dataset shows that this paper WLCNN-SA-IDS can efficiently fuse three sensor data and obtain higher fault diagnosis accuracy than existing machine learning methods, while avoiding the influence of conflicting evidence and improving the reliability of this method.

. Time Series Prediction with Events via Causal Representation Learning.:

The value of time series prediction is getting more and more attention. While the prediction of time series data under event disturbance has been difficult, the different distribution of data before and after the event and the different distribution of datasets will lead to the poor prediction accuracy, robustness, and generalization ability of the prediction model. In this paper, based on causal representation learning, we propose the causal representation prediction model(CRP), which is divided into CRP_Encoder and CRP_Decoder. CRP_Encoder is composed of the causal factor extractor and the causal representation decoupler to extract causal representations and distinguish them according to whether they are affected by events, which are designed by four properties proposed: causal representations are independent of non-causal factor, can be distinguished, causally sufficient for the prediction, dimensions of causal representations are independent of each other. CRP_Decoder is composed of the CNN network and causal representation coupler to learn the causal mechanism and predict data after the event, which are designed based on the equivalence of conditional structure and causal mechanism we proved. The experimental of the three datasets results show that the CRP model has high prediction accuracy, good robustness, and strong generalization ability.