Rolling element bearings are among the most crucial parts of rotating machinery and the most failureprone components, so estimating their remaining useful life before failure is essential for planning maintenance rather than reacting to breakdowns.The proposed work develops a system to predict the Remaining Useful Life (RUL) of high-speed bearings that combines the conventional machine learning methods with a custom deep learning architecture.For the Experiments, the PRONOSTIA dataset from the FEMTO-ST Institute was used, which contains vibration signals recorded at 25.6 kHz under varying speeds and loads.Time-domain and frequency-domain features are calculated from the Raw signals that RE splits into overlapping windows.The features are standardized and arranged into fixed-length sequences to preserve temporal patterns.Baseline models SVM, Random Forest, XGBoost, LSTM, and BiLSTM are trained for comparison.The proposed Hybrid-BiLSTM model builds on these by combining a multi-scale convolutional layer, dilated Temporal Convolutional Network blocks, bidirectional LSTM layers, and multi-head attention, followed by pooling and a regression layer.Trained with a Huber loss and early stopping.The MAE, RMSE, and R², are effectively evaluated for estimating model performance and for RUL estimation.