Abstract: Traditional rolling bearing fault warning usually adopts fixed threshold grading alarm,which exist many false alarms and missed alarms.The keys to solve this problem are how to effectively learn the indicators that can represent the health state of the equipment from vibration signals and self-learn the warning control limits.A method of bearing early fault prediction based on warning control limit self-learning was proposed.The Short-time Fourier Transform (STFT) was used to extract fault features of vibration data.A health indicator construction method based on Matrix Variate Gaussian Convolutional Deep Belief Network (MVGCDBN) was proposed,which could combine the fault features into high-level features without destroying the internal structure of two-dimensional sample space,and construct the health indicator through the full connection layer.The probability distribution of health indicators under normal operating conditions was fitted and goodness of fit was tested,and the upper quantile was used as the fault warning control limit.The effectiveness of the proposed method was verified with the international standard bearing data set.

Key words: incipient fault prediction, health indicator, distribution fitting, predictive maintenance, rolling bearing

摘要： 传统滚动轴承故障预警通常采用固定阈值分级报警,存在较多的误报警和漏报警。如何有效地从振动信号里学习能表征其健康状态的指标,自学习故障预警控制限,是解决该问题的关键所在。因此,提出一种预警控制限自学习的轴承早期故障预测方法。首先,采用短时傅里叶变换提取振动数据的故障特征;其次,提出基于矩阵变量高斯卷积深度置信网络的健康指标构建方法,在不破坏二维样本空间内部结构的同时将故障特征组合抽象成高层特征,通过全连接层构建健康指标;再次,拟合正常运行状态健康指标的概率分布的及检验拟合优度,并以上侧分位数作为故障预警控制限;最后,以国际标准轴承数据集验证了所提方法的有效性。

关键词: 早期故障预测, 健康指标, 分布拟合, 预测性维护, 滚动轴承