基于卷积神经网络与奇异值分解的水电机组故障诊断方法研究

Fault Diagnosis of Hydropower Units Based on Convolutional Neural Network and Singular Value Decomposition

提出了一种结合卷积神经网络,小波变换和奇异值分解理论的水电机组故障诊断方法。利用卷积神经网络提取机组轴心轨迹的图像特征;通过离散小波变换对摆度信号进行分解,获得信号的小波分解系数,对各分支系数进行重构,构造奇异值分解输入矩阵,提取矩阵奇异值作为特征向量。将两种方法提取的特征进行组合,构建包含图像特征和波形特征的混合特征向量,通过概率神经网络进行识别分类。为验证该方法的有效性,将水电机组常见故障在转子试验台上进行模拟,用上述方法进行诊断。结果表明,文中所提出的故障诊断方法能很好地识别水电机组不同运行状态,可为水电机组的故障诊断提供有效依据。

A novel method of fault diagnosis based on Convolutional Neural Network(CNN),Discrete Wavelet Transform(DWT)and Singular Value Decomposition(SVD)is proposed in this paper.CNN is used to extract features of shaft orbit images,DWT is used to transform the de-noised swing signal of rotating machinery and the wavelet decomposition coefficients of each branch of the signal are obtained by the transformation.The SVD input matrix is formed after single branch reconstructing of the different branch coefficients,and the singular value is extracted to obtain the feature vector.The features extracted from both methods are combined and then classified by Probabilistic Neural Network.The results show that this hybrid method can identify different operating states of hydropower units and provide an effective basis for fault diagnosis.