Seawater pump fault diagnosis based on parameter transfer and one-dimensional convolutional neural network

This paper solves the seawater pump's problem of insufficient fault samples, complex, and variable operating conditions and difficult extraction of vibration features. A fault diagnosis model based on one-dimensional convolution neural network(1DCNN) was proposed and parameter transfer was used to improve model's performance. CNN has powerful capacities of feature learning and feature expression, which was used to automatically learn fault features. Parameter transfer was added to transfer model parameters learned from source domain dataset to the identification tasks of target domain fault. Firstly, 1DCNN based model was constructed and trained on the source domain dataset to obtain model parameters. Deep learning training techniques were applied to prevent model overfitting, such as Dropout, regularization, and adaptive learning rate, etc. Then its model parameter knowledge was transferred to the target domain fault diagnosis model to construct 1DCNN parameter transfer model to identify faults by adding the step of fine-tune the parameters. Experimental results show that: compared with traditional 1DCNN, 1DCNN transfer model can significantly improve the convergence speed and classification performance. The average recognition accuracy of this method under three variable operating condition seawater pump datasets is as high as 95.93%, and it has higher recognition accuracy and stronger generalization ability. © 2021, Editorial Office of Journal of Vibration and Shock. All right reserved.