This　paper　addresses　the　significant　challenge　of　predicting　the　Remaining　Useful　Life　(RUL)　of　mechanical　equipment,　a　critical　aspect　of　predictive　maintenance　and　reliability　engineering.　Traditional　deep　learning　methods　in　RUL　prediction　have　been　hindered　by　key　challenges,　including　the　scarcity　of　comprehensive　lifecycle　data,　the　prevalence　of　high　-frequency　noise　in　sensor　readings,　and　a　heavy　reliance　on　supervised　learning.　To　overcome　these　challenges,　we　propose　a　novel　methodology　that　synergizes　self　-supervised　and　supervised　learning.　Our　approach　uniquely　leverages　non　-full　lifecycle　data　abundant　in　industrial　settings,　thereby　bypassing　the　limitations　posed　by　data　scarcity.　The　model　undergoes　a　two　-stage　training　process,　initially　learning　from　vast　quantities　of　non　-full　lifecycle　data　in　a　self　-supervised　manner,　followed　by　finetuning　in　a　supervised　phase　with　available　full　lifecycle　data.　We　employ　Contrastive　Predictive　Coding　(CPC)　for　the　encoder　and　a　Transformer　-based　decoder,　a　combination　adept　at　extracting　low　-frequency,　significant　features　from　the　sensor　data　and　effectively　predicting　RUL.　The　paper　demonstrates　the　efficacy　of　our　approach　through　comprehensive　experiments　testing　on　both　bearing　datasets　from　experimental　setup　and　wheelset　datasets　from　urban　rail　train,　showing　superior　or　comparable　performance　against　state-of-the-art　methods.　Our　results,　supported　by　ablation　studies,　suggest　the　potential　robustness　and　innovative　aspects　of　our　model,　indicating　it　could　contribute　meaningfully　to　the　field　of　predictive　maintenance.