类别不平衡的通信基站空调故障诊断

Air Conditioning Fault Diagnosis for Communication Base Station with Class Imbalance

为提高通信基站空调故障诊断的准确率,特别是小类样本故障的识别率,本文针对类别不平衡的数据集提出一种单隐层前馈多标签分类算法SLF-CIB.首先在特征空间的低秩假设上构建单隐层前馈神经网络.其次,在最小化误差损失阶段引入非对称阶式损失函数替代平方误差损失函数,通过截断参数和边界参数动态改善类别不平衡问题.根据SLF-CIB模型在训练过程的每一轮迭代的凸特性,应用交替方向乘子方法进行优化.测试过程中多标签输出层可提供故障源偏序向量为软故障的早期排查提供多维度的参考.在UCI标准数据集和通信基站空调数据集上的实验表明,SLF-CIB算法有效地提高了故障诊断精度特别是少数类的识别率.

In order to increase the accuracy of air conditioning fault diagnosis in communication base stations,especially the identification rate of minority class sample faults,this paper proposes a single hidden layer feed-ward multi-label classification algorithm( SLFCIB) for class imbalance data set. Firstly,a single hidden layer feed-ward neural network was constructed based on the lowrank hypothesis of feature space. Secondly,an asymmetric stage wise loss function was introduced to replaced square error loss function in the phase of minimizing error loss,and the class imbalance problem can be solved by setting the ramp parameter and margin parameter dynamically. According to the convex characteristic of SLF-CIB in each iteration of training process,alternating direction multipliers method can be used for optimization. Finally,the multi-label output layer can provide a multi-dimensional support for the early detection of soft faults. Experiments on UCI standard data sets and real communication base station air conditioning data set show that SLFCIB algorithm can effectively improve the accuracy of fault diagnosis,especially the recognition rate of minority category.