The　complexity　of　the　internal　structure　of　rolling　bearings　and　the　harshness　of　their　operating　environment　result　in　strong　non-stationarity　and　nonlinearity　of　the　vibration　signals.　It　remains　a　challenging　and　attractive　task　to　accomplish　more　accurate　classification　through　signal　processing　techniques　and　pattern　recognition　methods.　To　realize　this　aim,　a　novel　one-dimensional　improved　self-attention-enhanced　convolutional　neural　network　(1D-ISACNN)　with　empirical　wavelet　transform　(EWT)　is　proposed　for　rolling　bearing　fault　classification.　Firstly,　the　EWT　algorithm　is　employed　to　decompose　the　raw　signal　into　three　frequency　components,　allowing　for　further　extraction　of　multi-frequency　components　to　enhance　signal　characteristics.　Subsequently,　a　creative1D-ISACNN　leverages　the　merits　of　a　newly　developed　attention　mechanism　and　an　optimized　meta-activation　concatenation　function　in　feature　learning　to　better　capture　and　map　crucial　information　within　the　signal.　Furthermore,　label　smoothing　regularization　is　designed　as　the　loss　function　of　the　1D-ISACNN,　which　takes　into　account　not　only　the　loss　of　correctly　labeled　positions　in　the　training　samples　but　also　the　loss　of　other　mislabeled　positions.　Finally,　the　adaptive　moment　projection　estimation　is　designed　to　ensure　a　more　robust　gradient　update　strategy　for　updating　the　parameters　of　the　proposed　model.　The　developed　model　tested　on　three　different　sets　of　bearing　data,　has　achieved　a　classification　accuracy　of　100%.　In　ablative　experiments　and　other　comparative　experiments,　the　proposed　method　demonstrates　higher　recognition　accuracy　and　more　robust　generalization　capabilities　compared　to　other　excellent　approaches.