Abstract: Aiming at the difficulty of early fault identification and accurate evaluation of degradation performance of rolling bearings, the life state identification method based on deep convolutional autoencoder (DCAE) and multi-scale residual shrinkage network (MSRSN) for the rolling bearings was proposed. Firstly, in order to obtain clear fault characteristic frequency and frequency multiplier, the original data samples were converted into the envelope spectrum input deep convolutional autoencoder to realize the automatic extraction and expression of bearing life state features, and the multidimensional scaling (MDS) algorithm was used to reduce the life state features to obtain low-dimensional features. Then the Euclidean distance (ED) between samples in the low-dimensional feature space was calculated, which was the evaluation index of bearing performance degradation. Secondly, in order to comprehensively extract the bearing performance degradation characteristics, an improved multi-scale residual shrinkage network recognition model was proposed, and a new activation scheme combining ReLU and DropBlock regularization was developed to enhance the noise resistance of the model. Finally, the proposed method and the comparison method were applied to the bearing life experiment data. The experiment results show that the proposed performance degradation evaluation index can accurately identify the starting point of bearing performance degradation and describe the trend of bearing degradation. The improved multi-scale residual shrinkage network recognition model can accurately identify the bearing life state with an average recognition rate of 91.75% in the SSNR=-4~6 dB environment. The practicability and effectiveness of the proposed method are verified.

Key words: mechatronics vehicle engineering; deep convolutional autoencoder; performance degradation index; multi-scale residual shrinkage network; life state recognition

摘要： 针对滚动轴承早期故障识别困难、退化性能难以准确评估的问题，提出了基于深度卷积自编码器(DCAE)和多尺度残差收缩网络(MSRSN)的滚动轴承寿命状态识别方法。首先，为获得清晰的故障特征频率及倍频,将原始数据样本转换为包络谱输入深度卷积自编码器中，实现轴承寿命状态特征的自动提取与表达，并基于多维尺度分析(MDS)算法约简寿命状态特征获得低维特征，然后计算低维特征空间内样本间的欧几里得距离(ED)，即为轴承性能衰退评估指标；其次，为全面提取轴承性能衰退特征，提出了改进的多尺度残差收缩网络识别模型，并开发了ReLU与DropBlock正则化相结合的新激活策略增强模型的抗噪性；最后，将所提方法及对比方法应用于轴承全寿命实验数据。实验结果表明：笔者提出的性能衰退评估指标能够精准地识别轴承性能退化起始点以及刻画轴承的退化趋势，所提出的改进的多尺度残差收缩网络识别模型在SSNR=-4~6 dB环境中平均识别正确率为91.75%，能够准确识别轴承寿命状态，验证了方法的实用性以及有效性。

关键词: 车辆与机电工程；深度卷积自编码器；性能衰退指标；多尺度残差收缩网络；寿命状态识别