摘要
针对采集的船舶发电柴油机有标签状态数据集为小样本而造成的分类精度较低的问题,本文提出了一种新型的自监督学习框架用于机电设备的故障诊断,挖掘无标签数据集中的特征信息,以提高模型的分类能力。首先,通过KNN算法,将采集到的无标签数据集划分为正类样本和负类样本,并通过添加噪声的方法对原始数据进行数据增强,以此构造自监督任务。然后,设计基于卷积神经网络的编码器,根据正类、负类的伪标签,来提取无标签数据中的监督信息。最后,基于小样本的标签数据,通过编码器得到新的特征表征,对分类模型进行参数微调,提高模型精度。船舶柴油发电机故障实验证明,该自监督学习框架下的分类模型的准确率、精确率和召回率均高于直接用小样本标签数据训练的分类模型。
To solve the problem that the collected marine diesel engine condition data is a small labeled dataset,this paper proposes a novel self-supervised learning framework for the fault diagnosis of marine diesel genset equipment,which can mine feature information in the unlabeled datasets to improve the classification ability of model.Firstly,the collected unlabeled data is classified into positive samples and negative samples by the KNN algorithm,and the data enhancement is carried out to the raw dataset by adding noise,so as to construct the self-supervised task.Then,the encoder based on the convolutional neural network is designed to extract the feature information from the unlabeled datasets based on the positive and negative pseudo tags.Finally,based on the labeled data of small sample,a new feature representation is obtained through the encoder,and the parameters of the classification model are fine tuned to improve the performance.Experiments prove that the accuracy,precision and recall of the classification model under the self-supervised learning framework are higher than those trained directly with small sample label data.
作者
胡继敏
罗梅杰
Hu Jimin;Luo Meijie(No.1 Military Representative Office of the Navy in Shanghai District,Shanghai,201913,China;Naval research Institute,Shanghai,200030,China)
出处
《船电技术》
2022年第9期19-24,共6页
Marine Electric & Electronic Engineering
关键词
船舶机电设备
小样本数据集
自监督学习
故障诊断
marine electromechanical equipment
small datasets
self-supervised learning
fault diagnosis
