摘要
•This study proposes a production loss-guided cost matrix with self-attention model.•The model boosts fault detection by combining economic impact into classification.•Real-world validation is performed using data from a glycerin purification process.•The model achieves 95.31% accuracy, ensuring reliable fault detection and diagnosis.•The model prevents 99.08% of losses in fault classification traded by 3% accuracy.
Product loss is an ongoing critical challenge in industrial processes, particularly in chemical systems where undetected faults can have major economic consequences. Traditional fault detection models, designed primarily for mechanical systems, often prioritize accuracy and system performance but fail to account for the economic impact of faults in chemical systems. To address this gap, this study proposed a production loss-guided cost matrix self-attention, long short-term memory (PLSA-LSTM) model, which integrates a production loss-guided cost matrix to align fault classification with operational priorities. The cost matrix assigns higher penalties to faults with major production losses, guiding the model to focus on economically critical faults. The self-attention mechanism emphasizes critical input features and Bayesian optimization fine-tunes hyperparameters to balance accuracy and production loss minimization. The PLSA-LSTM model was applied to a glycerin purification process and achieved a fault detection accuracy of 95.31% while reducing production loss by 99.07% per fault occurrence, notably outperforming traditional methods. Compared to the traditional self-attention model, the PLSA-LSTM model reduced unprevented production loss from 94.97 kg/h to 0.87 kg/h while maintaining competitive classification performance. The results demonstrated the ability of the model to handle complex fault scenarios, prioritize faults with high economic impact, and minimize production losses, making it highly applicable to fault-prone industrial environments.