Remaining Useful Life (RUL) prediction is critical for ensuring equipment safety and optimizing maintenance schedules, directly impacting system reliability and maintenance efficiency. However, in real-world industrial scenarios, factors such as operating condition fluctuations and load variations lead to inconsistent data distributions, making it challenging for existing models to achieve satisfactory adaptability and accuracy. To address this issue, this paper proposes a deep learning framework based on a multi-branch serial-parallel fusion of CNN-BiLSTM-Transformer architectures. Through innovative model architecture design and optimized training strategies, the framework aims to enhance prediction accuracy and robustness under cross-condition and cross-distribution scenarios. Specifically, a training strategy combining snapshot ensemble and cosine annealing learning rate scheduling is introduced to improve model generalization through multi-model ensemble prediction. Experiments conducted on the C-MAPSS aero-engine dataset and the A123 lithium battery dataset validate the effectiveness of the proposed method. The results show that on the FD004 dataset, the proposed model reduces RMSE by 7.61% and Score by 26.39% compared to the best baseline methods; on the main test set of the A123 battery dataset, RMSE is reduced by 16.18%. Overall, the proposed model outperforms traditional machine learning models under both single and complex operating conditions, and its prediction results are comparable to those of mainstream models, demonstrating the feasibility and effectiveness of the proposed method in improving RUL prediction accuracy under cross-condition and cross-distribution scenarios.