针对矿井通风阻变故障诊断观测特征维度高、存在较多与阻变故障诊断无关和冗余特征的问题，以最小的观测点覆盖率和交叉验证误差为目标，建立基于多目标优化的观测特征选择模型，用于矿井通风阻变故障诊断的观测点选择。以k-近邻算法作为故障诊断模型的求解方法，利用非支配排序遗传算法Ⅱ对观测特征选择模型进行求解，解决仅在必要位置布设观测点，降低阻变型故障诊断的成本；剔除易导致观测特征之间的冗余或不相关特征，提高阻变型故障诊断学习器的性能；一定程度上缓解过拟合或欠拟合问题的发生，提高模型的泛化能力等关键问题。分别对7组特征组合方案进行观测特征选择比对试验，并以无样本模型进行验证试验，结果表明：相对观测点覆盖率降至0.2～0.5后，故障诊断准确率收敛于稳定的状态；风量单一特征类型在所有比对试验中均获得最低准确率；在确保故障诊断性能的情况下，观测特征经优化选择后很大程度上降低了阻变故障诊断模型的复杂度；优化选择的观测特征在无样本模型中具有与监督模型同样的有效性。因此，基于多目标优化的观测特征选择模型可有效地剔除冗余或不相关特征，从而提高阻变故障诊断模型的性能，为实时网络解算提供新的理论方法，为智能通风的实现提供技术支撑。

The observation feature dimension of resistance variant fault diagnosis in mine ventilation is high, and there are many irrelevant and redundant features. A multi-objective optimization model of mine ventilation resistance variant fault observation feature selection was established by taking the minimum coverage of observation points and cross-validation error as the goal. The k-nearest neighbor algorithm was used as the solution method for the fault diagnosis model. A multi-objective optimization feature selection method of resistance variant fault observation based on non-dominated sorting genetic algorithm Ⅱ was proposed. The critical problems are solved by deploying observation points only at necessary locations to reduce the cost of resistance variant fault diagnosis, eliminating redundant or irrelevant features between observation features to improve the performance of resistance-variant fault diagnosis learners, and alleviating the occurrence of over-fitting or under-fitting to enhance the model's generalization ability. Seven feature schemes were used to carry out the observational feature combination selection and comparison experiments and a sample-free model for verification experiments. The results show that after the relative observation point coverage falls to 0.2-0.5, the rate of fault diagnosis accuracy is converged to a stable state. A single feature type obtains the lowest accuracy rate in all comparison tests. Under the condition of ensuring fault diagnosis performance, the optimized selection of observation features dramatically reduces the complexity of the resistive fault diagnosis model. In the sample-free model, the optimized selected observation features have the same effectiveness as the supervised model. The proposed method can effectively eliminate redundant or irrelevant features, improving the fault diagnosis model′s performance. The proposed method can provide a new theoretical approach for real-time network calculation. The optimization of the locations of a small number of observation points provides technical support for the realization of intelligent ventilation.

1 故障诊断模型及观测特征选择模型

   1.1 阻变故障问题及其样本描述

   1.2 k-近邻阻变故障诊断模型

   1.3 观测特征选择模型

2 观测特征组合试验

   2.1 试验实例

   2.2 试验设计

3 试验结果分析

   3.1 观测特征优化选择结果

   3.2 监督学习模型验证结果

   3.3 无样本模型验证结果

4 结论