基于近邻传播聚类的航空电子部件 LMK诊断模型

针对小样本条件下，航空电子部件功能模块故障诊断精度不高的问题，将局部多核学习（LMKL）算法的多分辨率解释与局部特征自适应表示能力和超限学习机（ELM）运算高效的特点相结合，提出一种新的局部聚类MK-ELM（LCMKELM）诊断模型。通过引入近邻传播（AP）聚类，在挖掘训练样本局部特征信息的同时，有效约减了局部算法的计算复杂性，避免了过学习问题的出现；通过分别分析输入空间与特征空间的聚类特征，构造了相应的2种选通函数M₁、M₂，以优化选通函数的模型参数取代优化局部权重，有效解决了核超限学习机（KELM）的对偶优化形式关于局部权重二次非凸的问题。将本文模型应用于某型机旋转变压器激励发生电路功能模块故障诊断，结果表明：相比于4种常用的多核诊断算法，模型在实现低漏警、低虚警的同时，采用M₁选通函数的诊断算法将诊断精度平均提升了3.80%，采用M₂选通函数的诊断算法将诊断精度平均提升了5.98%。同时，模型在实现与流行的LMKL算法相近的训练时间的同时，测试时间更短。

Localized multi-kernel diagnosis model for avionics based on affinity propagation clustering

In consideration of the low diagnosis accuracy for avionics functional module fault, a new offline localized clustering multi-kernel extreme learning machine (LCMKELM) diagnosis model is proposed in this paper by combining the capabilities of multi-resolution interpretation and local feature self-adaptive representation from localized multi-kernel learning (LMKL) with the characteristic of high-performance operation from extreme learning machine (ELM). In order to avoid overfitting issue, affinity propagation (AP) clustering is used to make full use of the underlying localities in the training data and effectively reduce the computational complexity. Considering that the updating of localized kernel weights in dual optimization form of kernel ELM (KELM)is a difficult quadratic nonconvex problem, gating function M₁ and M₂ are respectively constructed to approximate localized weights by analyzing the clustering characteristics in input space and feature space. The proposed method is applied to actual fault diagnosis task of rotary transformer excitation generating circuit, and the experimental results show that the proposed method has the lower false alarm rate and missing alarm rate in comparison with four state-of-the-art multi-kernel learning algorithms, and meanwhile the diagnosis accuracy is averagely increased by 3.80% when M₁ gating model is used, and increased by 5.98% when M₂ gating model is used. Moreover, compared with canonical LMKL algorithms, the proposed method obtains similar training time cost, but it has less testing time cost.