Abstract: Aimming at the problems of complicated mechanism of fault and lack of faulty samples in fault detection of electronic equipment, a fault detection method based on Safe Level Synthetic Minority Oversampling TEchnique（SL-SMOTE） and Cost Sensitive Relevance Vector Machine（CS-RVM） is proposed. The proposed method considers the fault detection of electronic equipment as an imbalanced binary-class classification problem. Firstly, SL-SMOTE is used in data layer to expand the fault samples. Then RVM detector is trained by these optimized samples. At last, cost sensitive learning is introduced to the detection in order to get the results at the minimun cost of loss. The experimental results of the UCI dataset and application case show that the proposed method effectively improves the detection accuracy.

Key words: fault detection, imbalanced data, oversampling, cost sensitive, relevance vector machine

摘要： 针对电子设备故障检测问题中故障机理复杂、故障样本贫瘠的问题，提出一种SL-SMOTE(Safe Level Synthetic Minority Oversampling TEchnique）和代价敏感相关向量机（Cost Sensitive Relevance Vector Machine，CS-RVM）结合的电子设备故障检测方法。所提方法将电子设备的故障检测视为一个非平衡的二分类问题，首先在数据层采用SL-SMOTE对故障样本进行拓展，然后根据优化后的样本训练得到RVM检测器，最后将代价敏感学习引入到检测结果的判别中，得到损失代价最小的检测结果。UCI数据集以及应用案例的实验结果表明所提方法有效提高了检测正确率。

关键词: 故障检测, 非平衡数据, 过采样, 代价敏感, 相关向量机