Aiming at the difficulty of extracting vibration data under actual working conditions of rolling bearings, this paper proposes a bearing reliability evaluation method based on generative adversarial network sample enhancement and maximum entropy method under the condition of few samples. Based on generative adversarial network, data sample enhancement under few samples is carried out, and the reliability analysis model is established by using the maximum entropy principle and Poisson process. The reliability is evaluated according to the reliability variation frequency, variation speed and variation acceleration. The analysis results show that with the gradual increase of running time, the reliability variation frequency shows a nonlinear growth trend, which can be roughly divided into the initial running-in stage, the stable running-in stage and the intense running-in stage. The reliability variation speed is then used to distinguish the specific starting time of the three stages, and finally the preliminary relationship between the reliability variation acceleration and the remaining life is obtained. The experimental results of the XJTU-SY dataset show that compared with the existing reliability evaluation model, the proposed model has the advantages of less samples, no need for preprocessing and higher accuracy. The proposed model has made a beneficial supplement to the existing reliability analysis methods.
Keywords: Generative adversarial network; Lack of information; Maximum entropy principle; Reliability; Sample enhancement.
© 2024. The Author(s).