改进随机森林算法在电机轴承故障诊断中的应用

Applications of the Improved Random Forest Algorithm in Fault Diagnosis of Motor Bearings

电机轴承状态直接影响电机运行可靠性。随机森林算法具有较强的抗噪和适应能力,广泛应用于电机轴承故障诊断中。针对随机森林中传统决策树算法在连续特征属性值数目过大时复杂度高及易过拟合等问题,基于聚类思想构造自适应滑动步长减少其分类结点数,提出改进的C4.5决策树和分类回归树算法;针对传统随机森林算法中各决策树产生错误差异小、投票方法忽略强弱分类器差异及漏报率等问题,使用不同决策树算法进行分类,并借鉴议会制思想确定各决策树等级及权重,提出一种计及漏报率的随机森林集成投票算法。为验证所提方法的通用性及有效性,采用时域特征提取法和集合经验模态分解法分别构造特征向量,并通过凯斯西储大学轴承数据中心数据集和现场诊断试验进行验证。实验结果表明,所提算法不仅适用于多种特征提取方法,且相较于传统随机森林算法和多层感知器算法在诊断准确率和漏报率方面均更具优势,为电机轴承故障诊断提供一种新思路。

The motor bearing status directly affects the reliability of the motor operation.The random forest algorithm has strong anti-noise and adaptability and is widely used in motor bearing fault diagnosis.Given the problems of the traditional decision tree algorithm in the random forest when the number of continuous feature attribute is too large,such as high complexity and easy over-fitting,etc.,the adaptive sliding step size is constructed based on the clustering idea to reduce the number of classification nodes,and an improved C4.5 decision tree and classification regression tree algorithm is proposed.To solve the problems that the traditional decision tree in the random forest algorithm has small error difference,the voting method ignores the difference between strong and weak classifiers and have not taken the miss report rate into account,different decision tree algorithm were used for classification and drawing on the parliamentary system to determine the level and weight of each decision tree,a random forest integrated voting algorithm with miss report rate was proposed.To verify the versatility and effectiveness of the proposed method,the feature vector was constructed by time-domain feature extraction and ensemble empirical mode decomposition method respectively and verified by Case Western Reserve University Bearing Data Center data set and on-site diagnostic test.The experimental results showed that the proposed algorithm is not only suitable for multiple feature extraction methods but also has advantages in both traditional random forest algorithm and multi-layer perceptron algorithm in terms of diagnostic accuracy and miss report rate,providing a new idea for motor bearing fault diagnosis.