Fault diagnosis of seawater pump based on transfer kernel locality preserving projection

Abstract

Due to the strict data conditions of machine learning for fault diagnosis, the application of fault diagnostic method based on the transfer kernel locality preserving projection is proposed. It solves the seawater pump's problem of low diagnostic accuracy due to insufficient fault samples and complex and variable operating conditions. This method uses the vibration signal of seawater pumps as the object, and the historical data came from different working conditions of the pump to prepare for the proposed model. By preserving the prior distribution structure of seawater pumps fault training data, the fault data is mapped into high-dimensional space. Then, transfer learning minimizes the distribution discrepancy between different datasets by the maximum mean discrepancy (MMD) in the Hilbert space. By this means, the samples with same class in different datasets could cluster together. Resulting with a classifier SVM trained to diagnose the fault class of the seawater pump by the different datasets combined. Through experiments, the results show that the proposed algorithm is effective, having better diagnostic accuracy than several learning algorithms.