FDA与ML-KELM结合的机电系统故障识别

为准确识别坦克自动装弹机中的机电系统故障,提出了一种结合函数型数据分析(Functional data analysis,FDA)和多层核极限学习机(Multi-layer kernel extreme learning machine,ML-KELM)的故障识别方法。首先,以函数的视角对机电系统运行过程中具有平滑特性的时序数据进行特征信息挖掘,利用函数型主成分分析和主微分分析从不同空间将时序数据的变化特性表征为特征参数；其次,对提取的多传感器时序数据的特征进行Relief-F特征筛选,得到与分类强相关的特征；最后,采用ML-KELM对强相关特征进行深度特征学习,获取更抽象的特征表达,进而实现准确的故障识别。结果表明: 采用与某坦克自动装弹机中的链式输送机原理一致的实验装置进行故障识别实验,函数型主成分分析和主微分分析能够从不同的特征空间中提取时序数据中的有效故障特征,并且两种方法提取的特征具有互补性; 基于多传感器时序数据特征中的强相关特征,使用3层隐含层的ML-KELM能够实现较为准确的故障识别,所提方法具有可行性和有效性,为坦克自动装弹机中的机电系统故障识别的研究提供了一种参考。

Fault identification of electromechanical system combined with FDA and ML-KELM

To accurately identify the faults of electromechanical systems in a tank autoloader, a fault identification method combining functional data analysis (FDA) and multi-layer kernel extreme learning machine (ML-KELM) is proposed. Firstly, the feature information of time series data with smooth characteristics during the electromechanical system operation is mined from a functional perspective, and the change features of time series data are characterized as feature parameters from different spaces by functional principal component analysis and principal differential analysis. Secondly, the extracted features of the multi-sensor time series data are screened by Relief-F to obtain features strongly correlated with the classification. Finally, ML-KELM is used to perform deep feature learning on the strongly correlated features to achieve a more abstract feature representation, thereby realizing accurate fault identification. The fault identification experiment is carried out using an experimental setup consistent with the principle of the chain conveyor in a tank autoloader. Experimental results show that functional principal component analysis and principal differential analysis can extract effective fault features of time series data from different feature spaces, and the features extracted by the two methods are complementary. The ML-KELM with three hidden layers can realize more accurate fault identification based on the strongly correlated features in the multi-sensor time series data features. The proposed method proves to be feasible and effective, providing a reference for the research on fault identification of the electromechanical systems in the tank autoloader.