Aiming to address challenges in predicting the remaining useful life ( RUL) of harmonic drives-such as difficulties in selecting degradation nodes, poor physical interpretability of degradation indicators, and large prediction deviations, a novel approach was proposed. The method combined a one- dimensional stacked convolutional autoencoder ( SCAE) integrated with deep convolutional embedded clustering (DCEC) for degradation point extraction, along with an improved dung beetle optimization (DBO) algorithm to enhance the performance of a CNN-LSTM-based RUL prediction model. The vibration signals were processed by using the SCAE DCEC framework to identify degradation nodes, addressing issues related to the difficulty of node selection and the low compatibility between degradation indicators and the predictive network. Secondly, a modified dung beetle optimization (MDBO) algorithm was developed, incorporating SPM chaotic mapping, adaptive probability thresholds, and differential mutation perturbations, with its performance rigorously evaluated. Thirdly, the MDBO algorithm was applied to optimize the hyperparameters of the CNN-LSTM model, forming the MDBO-CNN-LSTM-RUL prediction model. An accelerated life test and validation experiment were conducted by using a harmonic drive test bench. The experimental results demonstrated that the MDBO CNN LSTM model significantly outperformed CNN, LSTM, CNN-LSTM, DBO-CNN-LSTM, fully convolutional networks, and Bayesian-optimized LSTM models in terms of goodness of fit. The proposed model achieved a prediction accuracy of 91.33% and exhibited superior recognition capability for capturing the degradation trends during the late stages of the lifecycle of harmonic drive.