Abstract: The abnormal operating environments, human factor interference and acquisition equipment failures might cause abnormal values or missing data irrelevant to the equipment health status in monitoring data of rotating machinery, resulting in misjudgment of mechanical health status and improper formulation of maintenance strategy. Therefore, an identification method of inferior monitoring data was proposed based on adaptive bandwidth kernel density estimation. Firstly, the zero drift and local noise were “impacted” by integrating the collected data in frequency domain and the kurtosis index after integration was calculated. Then the local mean error was used to adaptively select the Gaussian kernel bandwidth, the probability density function of kurtosis index was obtained, and the boundary of 95% confidence interval was used as the identification threshold of inferior data. Finally, the extraction method was verified by the whole life data of axle durability monitoring. The results show that compared with the fixed bandwidth and the kernel density estimation method based on quadtree segmentation algorithm, the proposed method has better recognition effectiveness on poor quality monitoring data.

Key words: mechanical equipment, poor quality data identification, adaptive kernel density estimation, threshold value division

摘要： 运行环境异常、人为因素干扰及采集设备故障等问题可能导致旋转机械监测数据中出现与设备健康状态无关的异常值或缺失数据，造成机械健康状态误判及维护策略制定不当等问题，为此，提出了一种基于自适应带宽核密度估计的劣质监测数据识别方法。通过对采集数据进行频域积分从而将零点漂移与局部噪声“冲击化”，计算积分后的峭度指标；采用局部均值误差进行高斯核带宽自适应选择，获得峭度指标的概率密度函数，并将95%置信区间的边界作为劣质数据识别阈值。通过车桥耐久监测全寿命数据对提取方法进行验证，结果表明，相比于固定带宽以及基于四叉树分割算法的核密度估计方法，所提方法对劣质监测数据具有较好的识别效果。

关键词: 机械装备, 劣质数据识别, 自适应核密度估计, 阈值划分