A Fault Prediction Method for Water Pump Based on BP Neural Network and Improved Dempster-Shafer Theory of Evidence

Abstract

In order to improve the fault prediction accuracy of the water pump, a Hilbert-Huang Transform (HHT) is applied to the vibration signals on the horizontal, vertical directions and driving end of the pump, and the marginal spectral energy was taken as feature vectors. Then three BP neural networks are used to make local prediction according to feature vectors of three parts of the pump, and establish the fault type matrix. The output of BP neural network is normalized into the Basic Probability Assignment (BPA) of Dempster-Shafer (D-S) evidence theory propositions. Then the decision-level information fusion of D-S evidence theory is carried out for the three evidence bodies. Aiming at the limitation of D-S evidence theory in dealing with conflicting evidence, an improved fusion method based on Jousselme distance is proposed. This method takes into account the credibility of each evidence body to the recognition of proposition. The simulation test shows that the proposed method can improve the fault prediction accuracy of water pump and has the feasibility of its application in fault prediction of water pump equipment.