基于改进支持向量机的水电机组多类轴心轨迹智能识别

在水电机组故障诊断系统中，轴心轨迹是判断机组状态的一个重要特征。而水电机组实际运行中轴心轨迹故障样本数量较少，依据其进行故障智能诊断无法准确完成，需结合相应频谱特性才可做出诊断。论文针对此问题，采用改进的支持向量机（support vector machine，SVM）多故障分类算法，建立了多故障轴心轨迹分类器，并应用于水电机组的故障诊断。结果表明，改进的SVM在样本数较少时取得较好的分类效果，样本数为16和50时，分类准确率达到了96.3%和91.2%，；并且在分类数增多时，分类准确率得到提高，而样本数增多时，分类准确率骤减。该故障分类器可实现多故障的识别和诊断，并且具有算法简单和对多故障轴心轨迹图形分类能力强的优点。该研究可为水电机组少样本轴心轨迹故障的智能诊断提供参考。

Intelligence identification for multi-class shaft centerline orbit of hydropower unit based on improved SVM model

Abstract: In the fault diagnosis system of hydropower units, the shaft centerline orbit is an important feature for the recognition of the unit operating condition, and different types of shaft centerline orbits reflect different operation state and fault information of shaft centerline orbit. In the actual operation of hydropower unit, there are few fault samples for shaft centerline orbits. Hence, the intelligent fault diagnosis cannot be performed accurately, and this problem must be solved with the combination of the corresponding spectral characteristics. Aimed at this problem, based on the improved support vector machine, a multi-fault classification algorithm was presented, the Hu invariant moment data of shaft centerline orbit graph were selected as training sample of the classification system, the error threshold level was inducted to effectively control category interference phenomenon, and a multi-fault shaft centerline orbits classifier was set up. Furthermore, it was applied to carry out the fault diagnosis of hydropower units. Results of the fault diagnosis application showed that just a few measured samples of shaft centerline orbits and a certain number of stimulated samples were needed in order to establish a fault classifier with superior performance, when the number of samples was 16 and 50, the classification accuracy was up to 96.3% and 91.2%, and the four different shapes of shaft centerline orbit graphs such as double ring-shaped, eight-shaped, ellipse-shaped and banana-shaped can be clearly distinguished. Meanwhile, the classification accuracy increased with an increase in the number of classification and the classification accuracy decreased rapidly with an increase in the number of sample, that is to say, the number of classification and the number of sample had an important influence on the classification accuracy. In addition, the optimum classification surface of invariant line moment can be obtained by adjustment of kernel function coefficient, the ability of multi-category classification can be obviously improved by introduction of distinct matrix, and it has been successfully verified in four different classifications. This fault classifier can realize the identification and diagnosis of multi-faults. And it has both the advantages of simple algorithm and strong capacity in pattern classification for multi-fault shaft centerline orbits. So the result provides a reference for the intelligent fault diagnosis of shaft centerline orbits of hydropower units with few fault samples.