油液原子光谱信息量大且具有模糊性， 严重影响了在故障诊断中的应用效率和精度。 为选择数量少、 效率高的光谱特征， 提出了一种光谱特征选择的新方法。 基于齿轮箱实验台架， 模拟了齿轮正常磨损状态和两种典型故障， 并采集了油液样本。 将三种磨损状态视为三个Vague集， 光谱特征值视为Vague集上的Vague值。 基于Vague值之间的相似度量， 定义了平均Vague敏感度(mean vague sensitivity， MVS)， 用来描述光谱特征对不同磨损状态的敏感程度， 并据此选择出对磨损状态敏感度高的光谱特征。 此外， 针对Vague集隶属度的确定严重依赖人为经验的问题， 利用Parzen窗法分别估计出三种状态光谱数据的概率密度分布后， 结合贝叶斯公式确定出Vague集的隶属度上、 下限。 实验表明， 此方法可以有效地从大量光谱特征中选择出对故障敏感程度较高的特征。

Large quantity and ambiguity of oil atomic spectrometric information greatly affects the applicable efficiency and accuracy in fault diagnosis. A novel method for choosing less and effective spectrometric features is presented. Based on gearbox test bed，we simulated the normal wear state and two typical faults to acquire the lubricant samples. The three wear states are regarded as three vague sets，and spectrometric feature values are vague values on vague sets. Based on similarity between vague values，mean vague sensibility(MVS) is defined to describe the sensitive degree of spectrometric feature to wear state. Besides，the membership degrees of vague sets greatly depend on human experience. The probability density distribution of spectrometric data of three wear states was estimated with Parzen window. Combined with Bayesian formula, the range of vague sets membership was calculated. Experimental results verify that the proposed method is of efficient help in choosing high fault-sensitive features from so many spectrometric features.