In the realm of intelligent manufacturing, accurately predicting the remaining useful life (RUL) of rolling bearings is essential for maintaining the high reliability and optimized performance of rotating machinery. To address the challenges associated with efficiently representing degradation states and capturing temporal dependencies in RUL prediction, this paper proposes a deep learning-based approach. The proposed method integrates a one-dimensional deep convolutional autoencoder (1D-DCAE) for high-quality feature extraction and a multilevel bidirectional long short-term memory (Bi-LSTM) network with a temporal pattern attention (TPA) mechanism to capture temporal dependencies. The 1D-DCAE extracts high-quality health indicators (HIs) from vibration signals, which serve as high-quality representations of the degradation state. These HIs, along with self-labelled data, are fed as inputs into the Bi-LSTM + TPA model, enhancing the quality of the data used in the prediction network. Experimental results on the PHM2012 bearing dataset demonstrate that the proposed method effectively extracts signal features and outperforms traditional labelling methods, achieving higher prediction accuracy and robustness. Furthermore, the model exhibits strong generalizability and transferability across diverse operating conditions, underscoring its potential for real-world applications.