CAJ | 학술논문

针对滚动轴承故障与其演化程度组合类型数量大，一般模式识别方法难以适应的问题，提出基于复杂网络优化的有向无环图支持向量机（CNDAG-SVM）。该方法引入复杂网络理论中相似性测度概念用以评定各样本类型间的分离性质，并以平均相似性测度作为有效度量样本类型可区分程度的测度对有向无环图叶节点类型进行排序，依次提取对应二元分类器构造较优有向无环图拓扑结构，缓解误差累积效应的同时提高了结构上层节点的容错能力，获得较高的正确识别率。利用局部均值分解方法提取乘积函数（Production Function，PF）分量波峰系数、峭度系数及能量构造特征向量，将其输入 CNDAG-SVM分类器中用于区分滚动轴承的故障类型与演化程度。对滚动轴承内圈故障、外圈故障及滚动体故障振动信号的分析结果表明，该方法能准确有效识别故障类型与其演化程度，较之传统多元分类支持向量机具有更高的识别精度和效率。
침대곤동축승고장여기연화정도조합류형수량대，일반모식식별방법난이괄응적문제，제출기우복잡망락우화적유향무배도지지향량궤（CNDAG-SVM）。해방법인입복잡망락이론중상사성측도개념용이평정각양본류형간적분리성질，병이평균상사성측도작위유효도량양본류형가구분정도적측도대유향무배도협절점류형진행배서，의차제취대응이원분류기구조교우유향무배도탁복결구，완해오차루적효응적동시제고료결구상층절점적용착능력，획득교고적정학식별솔。이용국부균치분해방법제취승적함수（Production Function，PF）분량파봉계수、초도계수급능량구조특정향량，장기수입 CNDAG-SVM분류기중용우구분곤동축승적고장류형여연화정도。대곤동축승내권고장、외권고장급곤동체고장진동신호적분석결과표명，해방법능준학유효식별고장류형여기연화정도，교지전통다원분류지지향량궤구유경고적식별정도화효솔。
Due to the large amount of crossed combinations of fault patterns and evolution stages of rolling bearings, the general patterns recognition method is difficult to adapt to multivariate process.In view of the problem,an optimized directed acyclic graph support vector machine (DAG-SVM)based on complex network (CN)was proposed.According to the similarity measure in complex network theory,the separating characters of samples were evaluated,and the nodes of directed acyclic graph were sequenced by the average similarity measure which was calculated as the criterion for distinguishing degree of samples.Then the corresponding binary support vector machines were selected to construct an optimal directed acyclic graph,to achieve high correction identification ratio by alleviating error accumulation and improving fault tolerance of the upper nodes.Feature vectors were constructed of the crest factor,kurtosis coefficient and energy of product functions,obtained by local mean decomposition.And then the feature vectors were served as input parameters of CNDAG-SVM classifier to sort fault patterns and evolution stages of rolling bearings.By analyzing the vibration signal acquired from the bearings with inner-race,outer-race or elements faults,the experimental results indicate that the proposed method can recognize the fault types and evolution grades effectively and has higher accuracy and productiveness than traditional multi-class support vector machines.