Machine learning model for windage alteration fault diagnosis of mine ventilation system under unbalanced samples

Purpose. The purpose of this paper is to improve the accuracy of resistance variation fault diagnosis in mine ventilation systems under unbalanced datasets. Methods. Based on WGAN-div, the unbalanced dataset is enhanced to achieve effective expansion of the original samples. A Bagging integrated learning and ResNet deep learning model is integrated to facilitate fault diagnosis of the ventilation system. Findings. Taking the simple T-shaped ventilation network as an example, fault datasets with unbalance ratios of 1:2, 1:8, 1:10, and 1:20 are constructed. The influence of unbalanced samples on windage alteration fault diagnosis (WAFs) of the ventilation system is deeply analyzed. Taking the ventilation system of Dongshan coal mine as the experimental object, fault diagnosis comparison experiments are conducted using different data augmentation models and classification models. Multiple evaluation indicators, along with t-SNE visualization, are used to assess the validity of the models. The results show that the data generated by the WGAN-div model has a good similarity to the real data. Compared to the GAN, WGAN, and WGAN-GP, the WGAN-DIV is superior. The performance of the ResNet deep learning model has improved significantly. Originality. This paper conducts research on fault diagnosis of ventilation systems using unbalanced datasets from both the data level and the network system level, effectively addressing the issue of sample imbalance in the actual working conditions of mine ventilation systems. Practical implications. The proposed method can provide technical support for the application of intelligent ventilation, enhancing both the reliability of monitoring and the overall safety performance of mine ventilation systems.