基于PCA相似系数与SVM的涡轮泵故障检测算法

Turbopump fault detection algorithm based on PCA similarity coefficients and SVM

提出了一种基于主元分析法(principal component analysis,PCA)相似系数与支持向量机(support vector machine,SVM)的故障检测算法用于液体火箭发动机涡轮泵试车后故障检测。该算法将历史信号按合理的步长分段,对信号段进行小波去噪预处理;再将每个步长信号平分为多段,采用主元分析法对每段信号进行降维处理;以第1段信号为基准,计算其他各段信号与第1段信号的相似系数,将所得相似系数组成向量作为故障特征;提取历史信号的所有故障特征向量构建SVM的训练样本集,并由此得到SVM决策函数用于待检信号的故障检测。用某型号涡轮泵振动加速度信号对算法进行验证,结果表明对时长21.50 s的待检信号,算法检出故障发生时刻为20.0076 s,非常接近故障真实出现时刻(约20.007 s),未出现虚警和漏警。算法具备了良好的准确性。

A fault detection algorithm based on principal component analysis （PCA） similarity coefficients and sup- port vector machine（SVM） was proposed for liquid rocket engine （LRE） turbopump fault detection after test run. Firstly the algorithm divides the historical signal into some segments by appropriate step, and does pretreatment of wavelet de- noising for each step. Secondly the algorithm divides every step of signal into some average segments, reduces the dimen- sion of every segment by PCA, chooses the first segment as basis, computes the similarity coefficient between the first segment and every other segment, uses all the similarity coefficients in the step to construct a vector as fault feature. Fi- nally the algorithm selects all the fault feature vectors of the historical signal to construct the SVM training sample set, obtains SVM classifier for fault detection of the test signal. A part of the vibration acceleration signal of a certain type of turbopump was chosen as the test object to validate the algorithm. The test results showed that for the test signals within 21.50s＇ duration, the algorithm detected the faults at 20.0076s without false alarm and missing alarm. The result was very close to the time that the faults really occurred（about 20.007s）. The algorithm has good accuracy performance.