在工業設備持續朝自動化與智慧化發展的趨勢下，設備健康監測（PHM）與剩餘使 用壽命（RUL）預測已成為提升工廠效率與安全性的重要工具。儘管傳統統計方法與機 器學習模型已被廣泛應用於此類任務，然而在處理高變異性、非線性退化與資料不平衡 問題時，常面臨準確度不足與泛化能力薄弱的挑戰。近年來，深度學習方法雖展現優異 表現，卻普遍仰賴大量完整退化資料，且缺乏明確的解釋性與物理對應性。 本研究針對此問題提出一套創新的資料生成與預測架構 —— WTTELSTMGAN（Wavelet-Temporal-Transformer Enhanced LSTM GAN）。本方法整合小波轉換的時 頻解析特性、雙向 LSTM 的時序記憶能力與多頭注意力機制的特徵聚焦優勢，能有效 模擬機械退化過程中的 RMS 演變趨勢，並補足原始資料不足所造成的學習落差。進一 步應用於結合 Weibull Distribution 與動態閾值之預測模型中，可預測失效點與剩餘壽命， 提升預測系統的實用性與準確性。 ;In the context of continuous advancements in industry and technology, Prognostics and Health Management (PHM) and Remaining Useful Life (RUL) prediction have become essential tools for enhancing operational efficiency and safety. Although traditional statistical methods and machine learning models have been widely applied to such tasks, they often struggle with high variability, nonlinear degradation, and imbalanced data, leading to limited prediction accuracy and weak generalization capabilities. In recent years, deep learning approaches have shown promising results, yet they typically rely on large amounts of complete degradation data and often lack interpretability and physical correspondence. To address these challenges, this study proposes an innovative data generation and prediction framework—WTTELSTM-GAN (Wavelet-Temporal-Transformer Enhanced LSTM GAN). The proposed model integrates wavelet transform for time-frequency feature extraction, bidirectional LSTM for temporal memory modeling, and multi-head attention for enhancing feature focus. This architecture effectively simulates the degradation trends of RMS sequences in mechanical systems and compensates for data insufficiency during model training. Furthermore, the generated data is incorporated into a prediction model that combines Weibull distribution fitting and dynamic thresholding, enabling accurate prediction of failure points and remaining useful life, thereby improving the practicality and reliability of the prognostic system.
