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

With the increasing requirements for air quality in cleanrooms, gas filtration media, as a core purification component, have a direct impact on system operating efficiency and product quality when their performance degrades. Under complex pollution sources and dynamic operating conditions, traditional replacement strategies based on empirical rules are often inadequate, highlighting the urgent need for the accurate prediction of the remaining useful life (RUL) of filtration media. In this study, a parallel predictive optimization model integrating Random Forest (RF), Bidirectional Long Short-Term Memory (BiLSTM), self-attention mechanism, and Particle Swarm Optimization (PSO), referred to as the PSO-RF-BiLSTM-Attention model, is proposed. This model can extract temporal features and key variables from the operational data of filtration media, enabling automated parameter optimization and dynamic performance prediction. An experimental platform for gas filter degradation was independently designed to simulate the long-term corrosion process of activated carbon filtration media under different SO2 concentrations, and multidimensional monitoring data were collected for model training and validation. Experimental results indicate that, compared with multiple baseline models, the proposed model reduces the mean absolute error (MAE) by ∼58.7% and increases the coefficient of determination (R2) by about 5.6%. Existing studies largely focus on single models or single prediction targets. In this study, a unified framework is proposed to achieve joint prediction of filter media performance and RUL, demonstrating its effectiveness and feasibility in complex degradation scenarios, thereby providing effective support for intelligent maintenance strategies of critical cleanroom components.