基于时频图与改进图卷积神经网络的异步电机故障诊断方法

An induction motor fault diagnosis method based on the time-frequency image method and an improved graph convolutional network

由于传统故障诊断技术依赖于人工提取特征,造成方法的泛化能力及应用受限。针对该问题,提出一种基于时频图与改进图卷积神经网络的异步电机故障诊断方法。首先,通过小波分析方法将电机振动信号转换为时频图,构建不同工况的图像样本;再基于超像素分割法处理图像生成超像素块,将其作为节点,并根据其纹理、颜色、距离特征生成图结构数据;然后将图结构数据输入改进网络,算法可以自适应地提取故障特征、得到诊断结果,其中,网络通过结构学习方法进行改进。该方法通过对节点相似度计算打分,以重构图连接结构,从而克服传统图卷积神经网络在池化操作后存在的图结构完整性缺失问题,实现卷积层和池化层的层层堆叠及图级分类。试验结果表明,所提方法可实现对转子断条故障、轴承故障、单相短路故障的有效诊断,与传统方法相比,具有较高的故障识别准确率。

Traditional fault diagnosis methods rely on the handcraft feature extraction,so their performance highly depends on the expert knowledge and their application is limited.To solve this problem,a fault diagnosis method based on the time-frequency image method and an improved graph convolutional network(GCN)was proposed in this paper.Firstly,the vibration signals were transformed into time-frequency images by the wavelet transform method.Then,based on the simple linear iterative clustering(SLIC),the images were segmented adaptively to generate superpixels,which were regarded as nodes in the graph.Next,based on the color and texture features in the superpixels,the connection relationships and the node features were formed.Then,the graphs were input into the network to diagnose the fault type.The algorithm can extract the features automatically and make the diagnostic classification.To overcome the limitations in traditional GCN,the network is improved by the structure learning method.This method reconstructs the connect relationship by calculating the similarity between nodes,thereby keeping the structure integrity after pooling.By stacking the convolutional and pooling operations,the diagnosis can be realized in the graph level.The results show that the proposed method can effectively classify different motor statuses under varying working condition,and the fault detection accuracy is the highest compared with the traditional deep learning method mentioned in the paper.