Traditional data-driven methods need to be based on manually designed features to achieve health state recognition and thus predict the remaining useful life (RUL) of a bearing. Under complex industrial conditions, prediction methods based on expert a prior knowledge have obvious limitations. A data-driven Dual-stage hybrid forecasting method is proposed in the paper. We feature extract noise reduction data to construct health indicators (HI), utilize a weighted feature density clustering algorithm based on HI, label and convert a small amount of data into a nested scatter plot (NSP), which is used to train the Nested Scatter Plot Classification Network (NSCNet), and the output features are analyzed using FWDBSCAN to complete the division of the health stages as well as the first prediction time (FPT) determination. Then, the degraded image data are fed into a hybrid Spatio-Temporal Convolutional LSTM (ST-ConvLSTM) for RUL prediction. Experimental results validated from publicly available datasets show that the method in this paper is able to classify health stages and improve prediction accuracy by accurately identifying the FPT with a two-stage hybrid model.