To address the challenges posed by non-smooth vibration signals and long time-series dependencies in the remaining life prediction of rolling bearings, this paper evaluates the predictive performance of an improved Recurrent Neural Network (RNN) algorithm, Long Short-Term Memory (LSTM) networks, and a hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) approach. Initially, ten statistical indicators from both the time and frequency domains are extracted to represent the degradation characteristics of the bearings. Next, the training set labels, corresponding to the full lifecycle of the bearings, are defined in chronological order. Finally, the degradation features are input into the LSTM and CNN-LSTM models to compare their predictive performance. Experimental results reveal that the CNN-LSTM model significantly outperforms the standard LSTM in high-noise scenarios due to the convolutional layer's ability to extract robust spatial features, despite incurring longer training times. In contrast, the LSTM model demonstrates more efficient training under low-noise conditions and exhibits reduced fluctuations in later-stage predictions. These findings illustrate the matching law between model characteristics and operational conditions: while CNN-LSTM is more suitable for industrial scenarios characterized by high noise and complex features, LSTM is preferable for real-time prediction tasks with limited computational resources.