Anomaly　detection　is　the　cornerstone　for　the　health　management　of　rolling　element　bearings.　The　unsupervised　learning　model　for　anomaly　detection　driven　only　by　normal　data　has　received　increasing　attention　in　recent　years.　In　this　article,　an　innovative　deep-learning-based　model,　namely,　memory　residual　regression　autoencoder　(MRRAE),　is　developed　to　improve　the　accuracy　of　anomaly　detection　in　bearing　condition　monitoring.　The　memory　module　and　autoregressive　estimator　are　applied　to　calculate　the　probability　density　distribution　of　the　latent　memory　residual　representation.　The　reconstruction　errors　and　surprisal　values　of　the　proposed　model　are　used　to　detect　the　abnormal　condition　of　bearing.　To　verify　the　superiority　of　the　proposed　method　in　anomaly　detection,　two　sets　of　run-to-failure　experimental　data　set　gathered　from　the　laboratories　are　studied　and　analyzed.　The　result　demonstrates　that　the　proposed　MRRAE　model　achieves　superior　performance　compared　with　several　conventional　and　deep-learning-based　anomaly　detection　methods.　Furthermore,　the　proposed　method　pays　close　attention　to　the　special　structure　of　bearing　vibration　signal　and　provides　a　new　way　for　explaining　the　decision-making　processes　of　deep　neural　networks.　©　1963-2012　IEEE.