一种迭代改进的球向量机故障诊断算法  

An Iteratively Improved Ball Support Vector Machine Algorithm for Fault Diagnosis

作　　者：杨清宇[1,2] 孙凤伟[2] 

机构地区：[1]西安交通大学机械制造系统工程国家重点实验室,西安710049 [2]西安交通大学电子与信息工程学院,西安710049

出　　处：《西安交通大学学报》2013年第10期1-6,共6页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61075001)

摘　　要：针对往复式压缩机故障数据规模大、常规诊断算法训练时间长的问题,提出了一种迭代改进的球向量机算法(IIBVM)。该算法将多个传感器采集到的压缩机不同工况下的运行数据输入球向量机进行训练,在训练过程中,增加了缓存用量,并在采样中跳过点积大于当前最远点点积的缓存点;引入缓存校正措施,每隔数次迭代即用点积和球心模计算公式对缓存中的数据进行一次校正;将距球心小于某个安全距离的点标记为无效点,并在下次采样到无效点时直接跳过;若核向量个数连续几次迭代保持不变,则提前终止迭代。训练后将新的压缩机运行数据代入决策函数,实现压缩机的故障诊断。在4个UCI标准数据集和实测的压缩机气阀故障数据集上进行的对比实验结果表明,IIBVM算法与球向量机算法相比,训练时间最多可降低50%,支持向量个数最多可减少18%。An iteratively improved ball support vector machine （IIBVM） algorithm is proposed to focus on the problem that conventional diagnosis algorithms take long training time in dealing with large scale fault data of reciprocating compressor.Operating data collected by several sensors in different working conditions are inputted into a ball support vector machine for training.More points are cached and the cache points with cached dot products larger than that of the current farthest point will be skipped in the next iteration during training procedure.Then the cache correction is performed using the dot product and the center norm formula every certain iterations.Moreover,a point will be marked as invalid and skipped in the next iteration if its distance from the center is less than a safe distance.The iteration is terminated if the number of core vectors remains the same for several iterations.The new operating data generated after training are substituted into a decision function to realize fault diagnosis.Experimental results on four UCI datasets and an actual compressor fault dataset show that the training time is reduced at most by 50％ and the number of support vectors is reduced by up to 18％.

关 键 词：往复式压缩机 故障诊断 大规模故障数据 球向量机 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]