NQPSO-SLFN: A Q-Learning enhanced PSO framework with neighborhood rough sets for failure mode recognition of reinforced concrete columns

To address the challenges of data heterogeneity and incompleteness in reinforced concrete (RC) column failure mode prediction, this study proposes a novel Q-learning-enhanced particle swarm optimization (PSO) algorithm integrated with neighborhood rough set theory, termed NQPSO. A single-hidden layer feedforward neural network (SLFN) is employed as the classifier, resulting in a unified framework: NQPSO-SLFN. Firstly, a neighborhood construction scheme is designed to effectively handle incomplete datasets with mixed-type attributes. Secondly, Q-learning is incorporated to dynamically adjust the PSO parameters, thereby improving the algorithm’s convergence behavior. Thirdly, the particle encoding scheme of PSO is redesigned, and a new fitness function is formulated, which jointly considers relevance, redundancy, and interaction among features, as well as classification accuracy. The proposed NQPSO algorithm was validated on several UCI benchmark datasets, where it demonstrated superior feature selection capability. Subsequently, comparative experiments were conducted on a real-world RC column failure mode dataset. The NQPSO-SLFN framework significantly outperforms machine learning methods such as SVM, RF, CatBoost, and GNN, achieving an accuracy of 0.9394, an F1-score of 0.9195, and a kappa of 0.8929. Furthermore, additional experiments were performed to assess the model’s performance using individual features, all features, and the selected feature subset. Results indicate that individual features yield poor performance, while using all features offers a moderate improvement. The selected feature subset delivers the best classification performance, thereby confirming the effectiveness of the proposed feature selection strategy in enhancing accuracy while reducing dimensionality. Overall, the findings underscore the practical applicability and robustness of the proposed NQPSO-SLFN framework in structural engineering tasks involving complex and incomplete data. • Developed a neighborhood rough set framework for mixed-type incomplete datasets. • Introduced Q-learning enhanced PSO for feature selection and classifier tuning. • Designed a fitness function based on relevance, redundancy, interaction, and accuracy. • Validated NQPSO on UCI datasets and RC columns, achieving high recognition accuracy.