Remaining Useful Life Prediction Method Based on Improved MLMA-Net and GMM

Predicting the remaining useful life (RUL) of key components is an important basis for formulating equipment maintenance plans, and the construction of reliable health indicators (HIs) is the key to achieving RUL prediction. By addressing the scarcity of samples, the unreasonable unified failure threshold, and the deficiencies of the Multi-level Long-Term Memory Enhancement Network (MLMA-Net) in long-term memory and boundary modeling, this study proposes a gearbox remaining life prediction method based on the Gaussian Mixture Model (GMM) and the improved Multi-level Long-Term Memory Enhancement Network (IMLMA-Net). The probability distribution of vibration signals is modeled through GMM, and the HI is constructed in combination with the distributed distance measurement method DCRM to overcome the problem of unreasonable setting of a unified threshold. To alleviate the difficulties of feature extraction and state modeling in small samples, a multi-level memory enhancement structure and the Gumbel activation function are introduced to construct IMLMA-Net, enhancing the perception of degraded information and the modeling ability of long-term dependence. The combination of GMM and IMLMA-Net further enhances the prediction performance and generalization ability of the system. Multiple sets of comparative experiments were conducted through two common gearbox datasets, indicating that the constructed HI can more accurately characterize the degradation process, and IMLMA-Net has better RUL prediction accuracy under different working conditions.