Rolling Bearing Fault Diagnosis Method Based On Graph Data

One of the most essential parts of rotating machinery systems,rolling bearings are used in a wide range of industrial applications.The rolling bearing will certainly sustain some damage over an extended period of usage,which will result in mechanical failure.Effectively diagnosing rolling bearing defects and ensuring the secure and reliable functioning of mechanical equipment are therefore of positive research value.With the pace of intelligentsia,the identification of mechanical faults began to shift from traditional signal processing technology to a diagnosis approach based on neural networks as intelligence increased.The fault diagnosis method based on neural network can deal with additional and higher dimensional data and has higher diagnosis accuracy.However,most neural networks can only handle regular data in Euclidean Spaces."Nothing can be done" for data in non-Euclidean Spaces.Therefore,graph neural network comes into being.Graph neural network is an efficient neural network for processing graph data.Graph data is a kind of irregular data based on non-Euclidean space.The graph data composed of nodes and edges connecting nodes has more comprehensive information expression ability.In the field of mechanical fault diagnosis,the intrinsic relationship between data has not been paid much attention in the diagnosis model.In this paper,the collected bearing vibration signals are converted into graph data,and the fault identification of rolling bearings is realized through the node classification task and graph classification task of graph neural network.The main work of this paper is as follows:(1)Fault diagnosis based on node classification task: Aiming at the problem of tiny sample size and insufficient label calibration in fault diagnosis of rotating machinery,a semi-supervised method based on graph convolution network was proposed for fault identification of rolling bearings.The foundation of this method is the node classification task of the graph convolutional network,and it uses semi-supervised learning to identify rolling bearing faults.Because the collected bearing vibration signals are one-dimensional time series,they are not topological graph data.Therefore,it is the primary problem to transform the one-dimensional time series into a suitable topology structure when graph neural networks are applied in the field of fault diagnosis.The quality of graph data is simultaneously determined by its form of production,and the classification accuracy of a graph neural network is similarly determined by the quality of the graph data.In this paper,one-dimensional time series are converted into graph data by the visibility algorithm,which better reflects the local information and the whole graph information of the data.The predicted classification results are generated using the softmax function as the input of a two-layer graph convolution network.This approach enables the identification of rolling bearing faults with a limited number of labeled samples,and avoids the problem of time consumption to label a large amount of data and the low accuracy of diagnosis under unsupervised learning.In addition,the visibility algorithm can transform one-dimensional time series without preprocessing the data.The constructed graph has a more comprehensive representation of the local information of the data,and considerably enriches the fault information and provides sufficient information for fault identification.(2)Fault diagnosis based on graph classification task: In order to verify the superiority of the graph data structure transformed by the visibility algorithm,this paper conducts an experiment on graph classification task.The graph classification task is different from the node classification task.The graph classification challenge needs the graph neural network to focus on both the graph’s structure and attribute information in addition to the nodes’ attributes.Thus,an excellent global representation can be obtained by learning the model.The data is also transformed by the visibility algorithm as the input of the graph convolution network,and the fault features are learned by supervised learning.When the graph convolution network learns and represents the global information,a one-time global pooling operation is used to extract the global information.At the same time,the influence of Max,Sum and Avg global pooling on the classification performance of the graph convolution network is analyzed.Experiments on two experimental data sets of rolling bearings are used to confirm the suggested method’s generalizability.Moreover,in order to confirm that the visibility algorithm can successfully identify rolling bearing faults,this paper carried out graph classification task experiments on two datasets with simplified graph convolution network,graph sampling and aggregation,graph attention network and graph attention network respectively.Experimental results show the effectiveness of this method.(3)Analysis of graph embedding problem : In the process of learning node embeddings to obtain graph embeddings,the node features expressed in scalar form may not be sufficient to completely express the features of nodes and the features of the entire graph,and then obtain the suboptimal global representation of the entire graph.This issue might have a negative impact on the model’s overall classification performance.Capsule graph neural network combines capsule network with graph neural network to improve upon the drawbacks of the current graph embedding.The feature of nodes is extracted by capsule,and the embedding of nodes is transformed from scalar form to vector form.Capsule graph neural network obtains the characteristics of graph hierarchy by dynamic routing mechanism and obtains multiple graph attributes from different aspects.In this paper,the visibility algorithm transforms the rolling bearing vibration signals from a one-dimensional time series to graph data,and the node states generated by graph convolution are replaced by simplified graph convolution as the vector of capsule.The graph capsule is generated by dynamic routing and attention module,and finally the fault recognition is realized by the alike capsule.Experimental results show that this method reduces the computational complexity of the model and has a certain accuracy.