Due　to　the　limited　bearing　degradation　data　under　actual　working　conditions,　it　is　impossible　to　obtain　enough　degradation　data　to　train　the　neural　network,　it　is　difficult　to　obtain　good　prediction　results　in　the　deep　learning　network,　so　a　new　fusion　method　was　proposed.　Firstly,　the　features　of　the　original　vibration　signal　was　extracted,　dozens　of　dimensional　features　were　obtained　through　the　ensemble　empirical　mode　decomposition　(EEMD)　and　the　singular　value　decomposition　(SVD),　and　the　effective　features　such　as　kurtosis　and　mean　value　commonly　used　in　remaining　useful　life　prediction　were　added,　then　the　decision　tree　to　filter　out　15-dimensional　features　was　used　the　data　was　obtained　by　double　exponential　model　fitting　and　the　degraded　signal　was　reduced　to　a　linear　trend　through　t-SNE.　The　linear　degradation　trend　has　better　generalization　in　prediction　than　the　exponential　trend,　and　the　prediction　accuracy　is　superior　to　support　veotor　regression(SVR)　and　deep　belief　network　(DBN)　model.　©　2024　Chinese　Mechanical　Engineering　Society.　All　rights　reserved.