College of Safety Science & Engineering, Liaoning Technical University
Key Laboratory of Ministry of Education for Mine Thermo-motive Disaster and Prevention, Liaoning Technical University

及时准确判断故障分支的位置对保障矿井通风系统的可靠性和安全性意义重大。针对实际工况下，矿井通风系统故障样本数据存在不平衡性，导致传统的机器学习模型诊断能力与泛化能力差的问题，提出了一种面向通风系统不平衡数据集的WGAN-div-RF故障诊断模型。以简单通风网络为例构造了不平衡比分别为2∶1、5∶1、10∶1、20∶1的故障数据集，深入分析了不平衡样本集对通风系统故障诊断的影响。搭建了基于Wasserstein距离生成对抗网络(WGAN-div)对不平衡数据集进行数据增强处理，在构建网络时创新性地加入了残差块，提高了生成数据的质量，实现原始样本的有效扩充。结合集成学习中的随机森林(RF)模型实现通风系统故障分支诊断。以东山煤矿通风系统为实验对象，分别进行了不同数据增强模型、不同分类模型以及不同数据生成率下的故障诊断对比实验，以多种评价指标及t-SNE可视化对模型有效性进行评估。结果表明：加入残差块的WGAN-div模型生成数据与真实数据具有很好的相似性，相较于GAN模型、WGAN模型和WGAN-gp模型，WGAN-div模型更具优越性；应用WGAN-div模型进行数据增强后，机器学习分类模型的性能提升明显；当扩充数据集达到平衡时，与其他集成模型及常用的矿井通风系统故障诊断SVM模型相比，RF模型在Re、Pr、Gmean和F1指标上均占优势。

It is of great significance to determine the location of fault branch timely and accurately to ensure the reliability and safety of mine ventilation system. To solve the problem that the traditional machine learning model has the poor diagnostic ability and generalization ability due to the imbalance of sample data in mine ventilation system under actual working conditions, a WGAN-div-RF fault diagnosis model is proposed. Taking a simple ventilation network as an example, the fault data sets with the imbalance ratios of 2∶1, 5∶1, 10∶1, 20∶1 are constructed, and the impact of imbalanced samples on the ventilation system fault diagnosis is analyzed indepth. The Wasserstein divergence for GANs (WGAN-div) is built, and the residual blocks are added innovatively to improve the quality of the generated data and expand the original sample set. Combined with the RF model, the fault diagnosis of ventilation system is realized. Taking the ventilation system of the Dongshan Coal Mine as the experimental object, the comparative experiments are carried out respectively with different data enhancement models, different classification models, and different data generation rates. The effectiveness of the model is evaluated with various indexes and t-SNE visualization. The results show that the data generated by the WGAN-div model with residual blocks has a good similarity to the real data. Compared with GAN, WGAN, and WGAN-gp, the WGAN-div model is superior. After applying the WGAN-div model for data augmentation, the performance of the machine learning classification model is significantly improved. When the expanded data set is balanced, compared with other integrated models and the commonly used SVM model for mine ventilation system fault diagnosis, the RF model is superior in Re, Pr, Gmean and F1 indexes.

国家自然科学基金资助项目(52374202)

赵丹，沈志远，宋子豪. 面向不平衡数据集的矿井通风系统智能故障诊断[J]. 煤炭学报，2023，48(11)：4112−4123.

ZHAO Dan，SHEN Zhiyuan，SONG Zihao. Intelligent fault diagnosis of mine ventilation system for imbalanced data sets[J]. Journal of China Coal Society，2023，48(11)：4112−4123.