To address information loss from manual feature extraction and the limited reliability of single-sensor diagnosis, this paper proposes a multi-sensor deep learning method.

Sensors positioned at four different locations are first modeled as nodes in a graph structure. A parallel four-channel one-dimensional convolutional neural network (1DCNN) then extracts deep features from the raw data of each node. Meanwhile, a dynamic adjacency matrix learning module captures the spatial correlations among the nodes and yields the graph adjacency matrix. Subsequently, a graph convolutional neural network (GCN) fuses the deep features from all four nodes. Finally, the fused features are dimension-reduced and classified to obtain the diagnostic results.

Experimental results show that the proposed multi-sensor fusion model achieves a fault diagnosis accuracy of up to 98.82% on the marine screw pump dataset, significantly outperforming the compared models.

Fusing the deep features extracted by the parallel four-channel 1DCNN with the GCN improves both diagnostic accuracy and fault tolerance.