Abstract: In addressing the challenge of extracting multi- periodic features for lifetime prediction based on time- series inputs, we propose a residual life prediction method for rolling bearings based on time- domain signal decomposition. We designed a dimension transformation module to decouple one- dimensional vibration signals based on prominent frequencies and converted them into a two- dimensional space. Utilizing a multi- scale residual attention mechanism, we extracted local and global deep semantic information from the transformed two- dimensional features. This multi- scale feature extraction captured dependencies both within and between periods, while the residual attention mechanism effectively filtered out irrelevant frequency bands and features. By employing bidirectional gated recurrent units, we amalgamated the semantic information from both forward and backward directions, enhancing the capture of deep degradation features through self- attention mechanisms. Experimental results on the PHM2012 dataset demonstrate superior predictive performance of our proposed method compared with existing state- of- the- art techniques.