A pruned-optimized weighted graph convolutional network for axial flow pump fault diagnosis with hydrophone signals

Due to the spatially dispersed occurrence of faults and the challenges associated with sensor installation in axial flow pump equipment, an underwater acoustic signal collection technique utilizing a hydrophone is employed. To address the task of fault diagnosis, a pruned-optimized weighted graph convolutional network (POWGCN) is proposed, leveraging the hydrophone signals. Initially, a base graph construction method using one-dimensional acoustic signals is introduced, encompassing node construction and edge establishment. Subsequently, a pruning optimization operation based on the proposed definition of the importance degree of nodes in the graph structure is implemented to enhance the quality of the input graph. And it also reduces the sensitivity of the input graph quality to the hyperparameter of k. Furthermore, an edge weighting strategy is adopted, wherein the edges are assigned varying weights based on the Mahalanobis distance between the central node and its neighboring nodes. This enables the representation of edge importance. By utilizing the pruned-optimized weighted graph as input, the fault diagnosis model, based on a graph convolutional network, can effectively extract fault-related information hidden in the hydrophone signals, despite the presence of noise. Finally, a SoftMax classifier is employed to obtain fault labels for all nodes. The experimental results of the three case studies demonstrated the effectiveness and generalization performance under noisy signals.