Anomaly　detection　for　the　control　rod　drive　mechanism　(CRDM)　is　key　to　enhancing　the　security　of　nuclear　power　plant　equipment.　In　CRDM　real-time　condition-based　maintenance,　most　existing　methods　cannot　deal　with　long　sequences　and　periodic　abnormal　events　and　have　poor　feature　extraction　from　these　data.　In　this　paper,　a　learning-based　anomaly　detection　method　employing　a　long　short-term　memory-based　autoencoder　(LSTM-AE)　network　and　an　extreme　gradient　boosting　(XGBoost)　algorithm　is　proposed　for　the　CRDM.　The　nonlinear　and　sequential　features　of　the　CRDM　coil　currents　can　be　automatically　and　efficiently　extracted　by　the　LSTM　neural　units　and　AE　network.　The　normal　behavior　LSTM-AE　model　was　established　to　reconstruct　the　errors　when　feeding　abnormal　coil　current　signals.　The　XGBoost　algorithm　was　leveraged　to　monitor　the　residuals　and　identify　outliers　for　the　coil　currents.　The　results　demonstrate　that　the　proposed　anomaly　detection　method　can　effectively　detect　different　timing　sequence　anomalies　and　provide　a　more　accurate　forecasting　performance　for　CRDM　coil　current　signals.