To address challenges in multi-time-step health prediction for lifting machinery, such as short data spans, high-frequency measurements, multi-dimensional feature complexity, and limited labeled data, this paper proposes a hybrid method combining convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) networks with an encoder-decoder architecture (ED-BLSTM). The method begins by chronologically organizing monitoring data, followed by segmenting and reconstructing the dataset while maintaining consistent input-output time step sizes. The processed data is first fed into a CNN to extract the main features, generating a multi-dimensional feature matrix. This matrix then trains a BiLSTM network within an encoder-decoder framework to build a predictive model for multistep forecasting of machinery health status. Comparative experimental results show that the method reduces validation loss by 0.097% to 0.474% and prediction loss by 1.230% to 1.411%, outperforming current mainstream approaches. These results demonstrate its potential to advance predictive maintenance in industrial equipment.