Machine health monitoring is used to forecast remaining useful life in predictive maintenance (PdM) to optimize production efficiency. Given the substantial cost and critical operational role of ion pumps (IPs) in synchrotron beamlines, in this study, we developed a deep learning framework for PdM to obtain failure predictions, allowing for maintenance scheduling to be optimized. We extracted historical data over a one-year period before an IP experienced failure. The developed framework uses sensor data from the previous 5 min to predict the target machine’s health status (normal/abnormal) in the next minute. A prediction of abnormality indicates that the machine may fail within a maximum planning window of five months, which provides engineers sufficient time to schedule maintenance. The following four models were trained for use in the proposed framework and then evaluated: long short-term memory (LSTM), gated recurrent unit, independent recurrent neural network, and one-dimensional convolutional neural network. The results indicated that among these models, the LSTM model achieved the lowest false negative rate of 5%. This trained LSTM model performed well when applied to seven additional IPs on an experimental beamline. Furthermore, a proposed iterative mechanism for interpreting results considerably reduced the false positive and false negative rates of the LSTM model from 19 to 0.69% and from 5 to 0.01%, respectively. Thus, the proposed real-time anomaly detection framework is suitable for practical industrial application in beamline maintenance. Future enhancements will focus on integrating data from additional sensors to improve predictive accuracy.