燃气轮机启动过程故障诊断方法研究

Research on Fault Diagnosis Method of Gas Turbine Start- up Process

在诊断燃气轮机启动故障时,常存在平均故障诊断错误率较高的问题。为此,提出基于深度信念网络(DBN)的燃气轮机启动过程故障诊断方法。在选取与典型启动故障相关的样本数据后,从机理的角度入手获取故障特征。根据DBN结构构建启动过程故障诊断模型,结合列文伯格-马夸尔特(L-M)算法设计自适应学习率改进算法,实现故障诊断精度的提升。按照参数表征性和冗余性选取试验样本,确定排气温度、电机电流、燃机转速和电机转速这4种参数类型作为故障诊断方法测试的数据基础,经由模型得出诊断结果。试验结果表明:所设计方法的平均故障诊断错误率为3.3%,与交叉全局人工蜂群和支持向量机(CGABC-SVM)方法、深度学习与信息融合诊断方法相比,所设计方法的平均故障诊断错误率分别降低了28.7%、37.0%,有效提升了燃气轮机启动过程故障诊断结果的准确性。该研究为燃气轮机的安全启动提供了保障。

When diagnosing gas turbine start-up faults, there is often a problem of high average fault diagnosis error rate. For this reason, a deep belief network-based fault diagnosis method for gas turbine start-up processes is proposed. After selecting sample data related to typical start-up faults, fault characteristics are obtained from the perspective of the mechanism. A start-up process fault diagnosis model is constructed according to the deep belief network structure, and an adaptive learning rate improvement algorithm is designed in combination with the Levenberg-Marquardt(L-M) algorithm to realize the improvement of fault diagnosis accuracy. Test samples are selected according to parameter representativeness and redundancy, and four types of parameters, namely exhaust gas temperature, motor current, combustion engine speed and motor speed, are identified as the data bases for testing the fault diagnosis method, and the diagnostic results are derived from the model. The results of the practical tests show that the average fault diagnosis error rate of the designed method is 3.3%, which is 28.7% and 37.0% lower than that of the crossover global artificial bee colony and support vector machine(CGABC-SVM) method and the deep learning and information fusion diagnosis method respectively, effectively improving the accuracy of fault diagnosis results during gas turbine start-up. This research provides assurance for the safe start-up of gas turbines.