Authors - Thuan Nguyen Dinh, Truong Nguyen Xuan Abstract - In this paper, the authors suggest and compare models for industrial oven Time-to-Failure (TTF) prediction within a critical 60-minute timeframe using sensor data. They compare traditional methods such as LSTM, GRU + Attention, and XGBoost with a new hybrid approach: CNN-Autoencoder + XGBoost (CNN-AE+XGBoost). Experimental outcomes, in terms of RMSE, MAE, and R², confirm the performance superiority of the proposed system. For the complete dataset, R² for the hybrid model was 0.89, well ahead of LSTM (0.27), GRU + Attention (0.47), and XGBoost (0.83). Importantly, for the focused TTF ≤ 60 minutes frame, it also had a low Mean Absolute Error (MAE) of 6.57. These results present the CNN-AE+XGBoost model as an effective predictive maintenance tool for curbing production downtime within the food processing sector.
