基于优化核函数带宽SVDD的机械振动预警模型  

Mechanical vibration warning model based on optimized kernel bandwidth SVDD

作　　者：刘晓金 陈文武 王庆锋[2] LIU Xiaojin;CHEN Wenwu;WANG Qingfeng(Equipment Safety Research Division,Sinopec Safety Engineering Research Institute Co.Ltd.,Qingdao 266000,China;Beijing Key Laboratory of Health Monitoring and Self-Recovery of High-End Machinery Equipment,Beijing University of Chemical Technology,Beijing 100029,China)

机构地区：[1]中石化安全工程研究院有限公司设备安全研究室,山东青岛266000 [2]北京化工大学高端机械设备健康监控及自愈化北京市重点实验室,北京100029

出　　处：《机电工程》2023年第11期1641-1654,1672,共15页Journal of Mechanical & Electrical Engineering

基　　金：中国石油化工股份公司科技部项目(323031);中石化安全工程研究院项目(Y315)。

摘　　要：基于高斯核函数的支持向量数据描述(SVDD),因其具有良好的异常检测性能,常被用于机械振动故障预警领域,但其性能的好坏受限于核函数带宽的取值是否适宜。为此,针对常规高斯核函数支持向量数据描述(SVDD)存在需要负类样本训练模型、计算量大、不收敛、不适用于小数值数据等问题,提出了一种不需要专家经验知识和负类样本训练SVDD超球体的优化核函数带宽计算方法,构建了基于优化SVDD核函数带宽的机械振动故障预警模型。首先,根据空间矩阵复杂度的信息熵,量化表征核函数带宽的取值对SVDD超球体的影响;然后,采用粒子群优化(PSO)算法寻找空间矩阵复杂度最大时对应的核函数带宽σ取值,实现了目标函数的快速收敛目的;综合考虑惩罚参数对SVDD超球体描述边界的影响,引入惩罚参数对寻优结果进行了修正,完成了对历史正常运行状态数据驱动的机械振动故障预警模型的构建任务;最后,应用辛辛那提大学智能维护中心轴承试验数据集等6项公开实验室数据和4项工程案例数据,对上述方法的实用性和可靠性进行了验证,并将其结果与采用常规方法所得结果进行了对比验证。研究结果表明:与常规方法相比,采用优化核函数带宽计算方法训练出的机械振动故障预警模型的合格率为100%,超球体描述边界拟合良好,并且不存在不收敛的问题。Support vector data description(SVDD)based on Gaussian kernel function was often used in the field of mechanical vibration fault warning field,because of its excellent anomaly detection performance,however its performance was limited by the appropriate value of kernel bandwidth.Therefore,aiming at the problems of conventional Gaussian kernel support vector data description(SVDD),such as the requirement of negative class data training model,complicated calculation,non-convergence,and inapplicability to small value data,an optimal kernel bandwidth calculation method was proposed which could get rid of dependence on expert experience knowledge and negative class data to train SVDD hypersphere.A mechanical vibration fault early warning model based on optimizing SVDD kernel function bandwidth was constructed.Firstly,the influence of kernel bandwidth value on SVDD hypersphere was characterized by information entropy of spatial matrix complexity.Then,the particle swarm optimization algorithm(PSO)was used to find the value of kernel function bandwidth parameterσwhen the spatial matrix complexity was maximum,and the convergence of the objective function was realized fleetly.Considering the influence of penalty parameter on the description boundary of SVDD hypersphere,the penalty parameter was introduced to correct the optimization results,and the mechanical vibration fault warning model driven by historical normal operating state data was constructed.Finally,the practicability and reliability of the proposed method was verified by six public laboratory data and four engineering case data,and the proposed method was compared with the conventional SVDD kernel function bandwidth calculation method.The research results show that comparing with conventional methods,the incipient fault warning model trained by the optimized kernel function bandwidth calculation method has a 100%qualification rate,the hypersphere description boundary is well fitted,and there is no problem of non-convergence.

关 键 词：机械设备故障预警 高斯核函数 支持向量数据描述 核函数带宽 惩罚参数 超球体 空间矩阵复杂度 粒子群优化算法 

分 类 号：TH133[机械工程—机械制造及自动化] TH113.1[自动化与计算机技术—控制理论与控制工程] TP18[自动化与计算机技术—控制科学与工程]