Abstract: Low-speed centrifugal pumps are prone to stress concentration issues during operation, which, along with dynamic stress variations, may lead to fatigue damage. Traditional finite element methods are limited in predicting fatigue life accurately. This paper describes a new fatigue life prediction method based on Long Short-Term Memory (LSTM) networks. It predicts the pump's operational life under different conditions, using LSTM networks to train and predict dynamic stress time series signals under rated and low-flow conditions, and integrating fatigue life models. Results show that this LSTM model captures stress variations accurately and can achieve fitting performance under rated conditions superior to that under low-flow conditions. The prediction set under rated conditions has a Mean Absolute Percentage Error (MAPE) of 21.3% lower than that under low-flow conditions. The operational life achieves 3×108 cycles under predicted stress of 2.28×107 Pa, meeting the life requirements of high-intensity operation. As a reliable, useful tool for pump management and maintenance, our method improves significantly the accuracy of centrifugal pump operational life prediction.

Key words: low specific speed centrifugal pump, long short-term memory network, dynamic stress, fatigue life