基于动态软聚类的航空电子部件LMKELM诊断模型

为提高小样本条件下航空电子设备模块级故障诊断精度, 基于动态软聚类的自适应特点与局部多核学习(local multiple kernel learning, LMKL)的局部特征表达能力, 提出一种新的局部多核超限学习机(local multiple kernel extreme learning machine, LMKELM)诊断模型。通过引入局部密度的概念进行自适应确定聚类数目, 并结合模糊C均值聚类对样本进行划分, 在充分体现类内多样性的同时, 约减了计算复杂度, 实现对样本的动态软聚类。通过构造选通函数解决局部权重二次非凸问题, 融合近似得到的局部权重与隶属度信息, 实现对测试样本的故障诊断。将该模型应用于某型机旋转变压器激励发生电路, 实验结果表明, 相比于4种前沿的多核学习方法, 该算法在漏警率、虚警率方面表现优异, 选用的M₁与M₂选通函数分别将诊断精度平均值提升2.78%和4.37%。

Local multiple kernel extreme learning machine fault diagnosis model with dynamic fuzzy clustering for avionics

To improve the module-level fault diagnosis rate for avionics in a small sample size,based on the adaptive feature of dynamic fuzzy clustering and the local feature of local multiple kernel learning(LMKL),a local multiple kernel extreme learning machine(LMKELM)model for avionics is proposed.The model confirms the number of clusters adaptively by introducing the concept of local density.Local density is combined with the fuzzy C-means algorithm to realize dynamic fuzzy clustering,which well reflects the diversity between clusters and reduces calculation.A gating model is constructed to solve the non-convex quadratic problem of local weight.Fault diagnosis of test samples is realized by fusing information of local weight and membership.The proposed model is applied into a certain type of rotary transformer excitation generating circuit.Compared with the four fashionable multiple kernel learning methods,the proposed model performs better in terms of false alarm and missing alarm rate.The experimental result shows that the applied M1 and M2 gating models enhance the average accuracy by 2.78%and 4.37%respectively.