Rolling　bearings　are　critical　in　mechanical　systems,　and　their　failure　can　cause　significant　economic　and　safety　risks.　Timely　fault　diagnosis　is　essential　to　prevent　unplanned　downtime.　This　paper　introduces　a　new　method　for　rolling　bearing　fault　diagnosis　using　a　cross　-attention　network　and　multi-scale　feature　fusion.　The　original　vibration　signal　is　decomposed　using　Variational　Mode　Decomposition　(VMD)　for　time-domain　features,　and　frequency-domain　features　are　extracted　via　Fast　Fourier　Transform　(FFT).　A　Bidirectional　Temporal　Convolutional　Network　(BiTCN)　extracts　time-domain　features,　while　a　Bidirectional　Gated　Recurrent　Unit　(BiGRU)　captures　frequency　-domain　patterns.　These　features　are　integrated　through　a　cross　-attention　mechanism　and　classified　for　fault　identification.　Evaluated　on　the　Case　Western　Reserve　University　(CWRU)　bearing　dataset,　the　method　achieves　98.67%　accuracy,　outperforming　traditional　techniques.