Research on filter life and performance prediction integrating attention mechanism and optimization algorithms

Area of Science:

• Environmental Engineering
• Chemical Engineering
• Data Science

Background:

• Cleanroom air quality is critical for system efficiency and product quality.
• Degradation of gas filtration media necessitates advanced predictive maintenance.
• Traditional replacement strategies are insufficient for complex pollution scenarios.

Purpose of the Study:

• To develop an accurate Remaining Useful Life (RUL) prediction model for gas filtration media.
• To integrate multiple machine learning techniques for enhanced predictive capabilities.
• To provide a unified framework for joint prediction of filter performance and RUL.

Main Methods:

• Developed a PSO-RF-BiLSTM-Attention model integrating Random Forest, BiLSTM, self-attention, and PSO.
• Designed an experimental platform for simulating activated carbon filter degradation under SO2.
• Collected multidimensional operational data for model training and validation.

Main Results:

• The PSO-RF-BiLSTM-Attention model significantly outperformed baseline models.
• Achieved a ~58.7% reduction in Mean Absolute Error (MAE).
• Increased the coefficient of determination (R2) by approximately 5.6%.

Conclusions:

• The proposed model effectively predicts gas filter media performance and RUL in complex degradation scenarios.
• Demonstrated the feasibility of a unified framework for joint prediction.
• Offers effective support for intelligent maintenance strategies in cleanroom environments.