Research On Improved Graph Convolutional Neural Network Based Algorithm And Its Application

Fault diagnosis is a key technology to ensure the safe,reliable and efficient operation of industrial systems.With industrial automation and intelligence,fault diagnosis methods are shifting from signal processing-based methods to deep learning-based methods.As a class of methods developed based on deep learning for processing graph data,graph neural networks have become a hot topic and direction in industry and academia thanks to their powerful data representation learning and analysis capabilities,which can not only help diagnosis faults,but also improve the automation and effectiveness of intelligent process control.In this thesis,the original vibration signal is modelled as a fault sample correlation map and studied based on graph neural networks,the main work of which is as follows:Traditional fault diagnosis methods are usually difficult to adaptively extract fault features of time-frequency diagrams.To solve the problem,a novel adaptive fault diagnosis method based on high-order visual graph convolutional network(HVGCN)is proposed.In the method,continuous wavelet transform(CWT)is firstly utilized to construct wavelet time-frequency diagrams of the raw vibration signals,and visual graph data of the diagrams can be obtained by lifting-dimensionality visual coding and a heuristic graph generation strategy.High-order visual graph data contains more global information while increasing the receptive field.On the basis of high-order visual graph data,we further construct a deep fault feature representation through the high-order visual graph convolutional and self-attention graph pooling,and thus HVGCN is developed.In the network,parameters are adjusted to gradually improve the network model so that establishes an accurate mapping from the raw signal features to the fault state,and adaptive fault features with hierarchical structure can be learned from high-order visual graph data.The proposed method is verified for effectiveness using two different public fault datasets.To solve the problem that GNNs cannot fully load the entire attributed graph due to limited memory resources,the binary identify-aware graph convolutional network(BID-GCN)is proposed.In this network,the nodes information is considered recursively during message passing,and then in order to obtain an embedding of a given node,the binary identify-aware graph convolutional net-work will extract the ego network centered at that node and perform multiple rounds of heterogeneous message passing,applying different parameters to the central node of the ego network to the rest of the nodes.In this process,the network parameters and input node features are binary by the network.In addition,the original matrix multiplication is modified to be binary to speed up the operation.Through theoretical analysis and experimental evaluation,BID-GCN can reduce the average approximate 36 times of both the network parameters and input data,and accelerate the inference speed by the average approximate 49 times on the citation networks.It can provide comparable performance to full precision baselines,and BID-GCN can better tackle the problem of limited memory resources.Figure 18 Table 10 Reference 80...