一种基于BPNN和SVM-PDE的旋转机械变工况预警方法

针对传统固定报警限未考虑时变工况的影响,易造成设备在高工况下虚警、低工况下漏警的问题,提出了一种基于BPNN(BP神经网络)和SVM-PDE(支持向量机概率密度估计)的旋转机械变工况故障预警方法。利用BPNN识别设备运行工况,结合信号处理方法从各工况振动数据中提取出多维特征并利用PCA(主成分分析)约简特征维度。将传统支持向量机(SVM)核函数改造为概率密度函数,将运行工况和低维特征输入SVM求解不同工况下正常样本的概率密度。以各个工况下正常样本概率密度值的边界值作为振动阈值进行故障预警。利用双转子试验台振动数据进行验证,结果表明,相较于固定阈值预警方法,基于BPNN和SVM-PDE的旋转机械变工况预警方法能有效降低漏警率和虚警率,验证了该方法的有效性。

An Early Warning Method of Rotating Machinery Based on BPNN and SVM-PDE Under Variable Working Conditions

Traditional fixed alarm limit is easy to cause false alarm at high working conditions and missed alarm at low working conditions due to the influence of time-varying working conditions.An early warning method of rotating machinery based on BP neural network(BPNN) and probability density estimation of support vector machine(SVM-PDE) under variable working conditions is proposed.BPNN is used to identify the working conditions of the equipment.Combined with signal processing method,multi-dimensional features are extracted from vibration data of various working conditions,and principal component analysis(PCA) is used to reduce the feature dimension.The traditional support vector machine (SVM) kernel function is transformed into a probability density function,and the working conditions and low-dimensional features are input to the SVM to acquire the probability density of normal samples under different working conditions.The boundary of the probability density value of the normal sample under each working condition is taken as vibration threshold for fault early warning.The results of dual-rotor test rig show that compared with the fixed threshold early warning method,the early warning method of rotating machinery based on BPNN and SVM-FDE under variable working conditions can effectively reduce the false alarm rate and miss alarm rate,which verifies the effectiveness of the method.