基于增强层次对称点图像分析和深度残差网络的水电机组故障诊断

图像转化在水电机组故障诊断领域具有一定的潜力，传统将一维数据转化为图像的方法存在图像特征单一性、一张图像难以表示多种信号且图像识别精度偏低等问题。为此，提出一种基于增强层次对称点图像分析（Enhanced Hierarchical SDP，EHSDP）和深度残差网络（Deep Residual Network，Resnet50）的水电机组故障诊断方法。首先，利用移动差分和移动平均过程代替传统的层次分解，提出EHSDP的图像转化方法，在克服信号特征表现单一性问题的同时图像转化效率提高27.42%；其次，将分解过的振动信号图像化得到水电机组的图像数据库，划分EHSDP图像为训练集和验证集，利用训练集训练Resnet50模型得到最优模型参数；然后，将验证集图像输入训练好的Resnet50模型中，借助TSNE对提取到的特征降维可视化，各状态特征信号无混叠；最后，输出图像特征分类实现水电机组故障诊断，并用某水电站SK-3#真实机组数据进行验证。仿真实验和实例验证结果均表明，所提方法在所有对比模型中优势明显，验证了本文所提方法的有效性和实用性。

Fault diagnosis of hydropower units based on enhanced hierarchical Symmetrized Dot Pattern and Resnet50

Image transformation has certain potential in the field of fault diagnosis of hydropower units.The traditional methods of transforming one-dimensional data into images has the problems, such as image feature singularity, the difficulty of representing multiple signals with one image, and the low accuracy of image recognition.Therefore, this paper proposes a fault diagnosis method based on enhanced hierarchical symmetrized dot pattern(EHSDP) and deep residual network(Resnet50).Firstly, instead of the traditional hierarchical decomposition, the image transformation method of EHSDP is proposed by utilizing the moving difference and moving average process, which can overcome the problem of single signal feature while improving the image transformation efficiency by 27.42%.Secondly, the decomposed vibration signals are imaged to obtain the image database of the hydropower units, and the EHSDP images are divided into a training set and a validation set.The Resnet50 model is trained with the training set to obtain the optimal model parameters.Then, the validation set images are input into the trained Resnet50, and the extracted features are visualized with TSNE.There is no aliasing in the characteristic signals of each state.Finally, the output image feature classification realized the fault diagnosis of hydropower units and the real machine data of SK-3# units of a hydropower plant was used for validation.The results of both simulation experiments and example verification show that the proposed method has obvious advantages in all the comparison models, which verifies the effectiveness and practicality of the proposed method in this paper.