Explainable AI for rule reduction in fuzzy models for air pollution measurement adjustment

Air quality monitoring using low-cost sensors has become increasingly important, yet their measurements are often inaccurate. Traditional adjustment methods face limitations in both applicability and explainability. This paper presents an explainable artificial intelligence approach for rule reduction in adaptive neuro-fuzzy inference systems, to improve the interpretability and efficiency of fuzzy models for fine particulate matter (PM 2.5 ) measurement adjustment. We introduce two novel algorithms, the Binary Activation Method and the Weighted Activation Method, to assess and eliminate redundant rules while maintaining predictive performance, validating the approaches in multiple geographic locations. On average, rule pruning results in an increase in MAE of 0.2 on the training set and 0.1 on the test set. The simplified models retain strong correlation, with Pearson’s correlation coefficients ranging from 0.73 to 0.96 in the test set. These results support the development of reliable and interpretable artificial intelligence systems for environmental monitoring. • Fuzzy rules in ANFIS can be reduced without sacrificing performance or transparency. • Weighted activation outperforms binary activation for rule ranking. • Model complexity is reduced with minimal performance loss. • Validated across cities with different environmental conditions. • Pruning rarely used rules improves generalisation.