Department of Automation, Faculty of Mechanical Engineering, Lublin University of Technology, Nadbystrzycka 36, 20-618 Lublin, Poland.
Institute of Production Techniques and Systems, Leuphana University of Lüneburg, Universitätsallee 1, 21335 Lüneburg, Germany.
Sensors (Basel). 2023 Jun 25;23(13):5875. doi: 10.3390/s23135875.
This article classifies the dynamic response of rolling bearings in terms of radial internal clearance values. The value of the radial internal clearance in rolling-element bearings cannot be described in a deterministic manner, which shows the challenge of its detection through the analysis of the bearing's dynamics. In this article, we show the original approach to its intelligent detection through the analysis of short-time intervals and the calculation of chosen indicators, which can be assigned to the specific clearance class. The tests were carried out on a set of 10 brand new bearings of the same type (double row self-aligning ball bearing NTN 2309SK) with different radial internal clearances corresponding to individual classes of the ISO-1132 standard. The classification was carried out based on the time series of vibrations recorded by the accelerometer and then digitally processed. Window statistical indicators widely used in the diagnosis of rolling bearings, which served as features for the machine learning models, were calculated. The accuracy of the classification turned out to be unsatisfactory; therefore, it was decided to use a more advanced method of time series processing, which allows for the extraction of subsequent dominant frequencies into experimental modes (Variational Mode Decomposition (VMD)). Applying the same statistical indicators to the modes allowed for an increase in classification accuracy to over 90%.
本文根据径向内部游隙值对滚动轴承的动态响应进行了分类。滚动轴承的径向内部游隙值不能以确定性方式描述，这表明通过轴承动力学分析来检测它具有挑战性。在本文中，我们展示了通过短时间间隔分析和选择指标计算的智能检测原始方法，这些指标可以分配到特定的游隙等级。测试是在一组 10 个相同类型（NTN 2309SK 双列自调心球轴承）的全新轴承上进行的，这些轴承具有对应于 ISO-1132 标准各个等级的不同径向内部游隙。分类是基于加速度计记录的振动时间序列进行的，然后进行数字处理。计算了广泛用于滚动轴承诊断的窗口统计指标，这些指标作为机器学习模型的特征。分类的准确性不尽如人意；因此，决定使用更先进的时间序列处理方法，该方法允许将后续主导频率提取到实验模式中（变分模态分解 (VMD)）。将相同的统计指标应用于模式可以将分类准确性提高到 90%以上。
