Objectives Marine water pump bearings operate in complex environments, and the fault features in the acquired signals are easily submerged by noise, resulting in low fault diagnosis accuracy. This paper proposes a Butterworth mean filtering Markov transition field (BM-MTF) technique combined with the ResNet-18 network to solve the problem.

Methods First, the BM filter is employed to improve the fault impulse waveform of the signal, suppressing noise interference and amplifying fault characteristics. Then, a two-dimensional image is generated through MTF to effectively visualize and enhance the signal characteristics. The MTF images, after BM filtering, are input into the ResNet-18 network for fault diagnosis. Finally, the method is verified using the public bearing fault dataset from Western Reserve University, the laboratory bearing fault dataset, and marine water pump bearing fault dataset, with comparisons to other methods.

Results The proposed method demonstrates a 100% accuracy on three bearing fault datasets. The comparative experiments show that the proposed method can effectively extract fault features and achieves higher recognition accuracy.

Conclusions This paper presents a novel method for fault diagnosis of marine water pump bearings.