【作者】 王超；

【导师】 吴军；

【作者基本信息】 华中科技大学 ， 船舶与海洋工程， 2023， 硕士

【摘要】 船舶是海洋运输的基础和载体,是国家装备制造业中不可缺少的组成部分。然而,受载荷多变以及海洋高盐雾环境等因素影响,船舶核心机械装备性能退化相对较快且故障频发,若不及时检测和修复,极易引起灾难性事故。因此,为了保障船舶运行的安全性与可靠性,研究智能化、高精度的机械装备故障诊断技术刻不容缓。随着数据监测技术的发展和人工智能技术的普及,基于机器学习的智能故障诊断方法已经成为本领域的研究热点之一。但目前智能诊断模型大多是建立在理想状态监测数据驱动上,即数据分布一致且标注信息完备,并未考虑到服役机械装备运行工况多变、数据标注信息缺失、故障样本稀少等因素,导致智能模型普遍存在鲁棒性差、自适应性低等缺陷。然而,弱监督学习可以采用各种不完备或受损的标注数据来建立高性能诊断模型,提高故障诊断精度。因此,本文基于弱监督学习理论,开展典型机械装备智能故障诊断关键技术研究,主要工作如下:(1)针对机械装备监测数据标注信息缺失的问题,提出一种基于对抗自监督学习(Adversarial Self-Supervised Learning,ASSL)的故障诊断方法。该方法通过构造一种新的对比式学习范式,实现模型对无标注样本间的隐层相关性的自主挖掘与自动聚类,避免了模型坍塌,同时增强了编码器的自学习能力与泛化能力。与传统监督式学习相比,ASSL训练过程中对样本标注信息的依赖性显著下降,只需在下游任务中通过少量样本对网络参数进行微调,即可实现高精度故障诊断。通过搭建的高速轴向柱塞泵实验台对提出的方法进行验证。实验结果表明:该方法在标注样本稀少场景下的诊断性能优于有监督式方法。(2)针对机械装备变工况导致监测数据变分布以及监测信号隐含噪声的问题,提出一种基于自注意力子域自适应对抗网络(Self-Attention and Subdomain Adaptive Adversarial Network,SASAAN)的故障诊断方法。首先,构造带有自注意力模块的特征提取器,通过其权重决策机制消除监测信号中的噪声干扰,提升模型鲁棒性;接着,将域对抗技术与子域自适应技术相结合,减小源域与目标域中样本的全局分布与相关子类分布差异,增强模型对高维域不变特征的提取能力;最后,通过Ottawa轴承案例验证所提出方法的有效性。结果显示:在强噪声和变工况条件下,该方法在目标工况上的平均诊断精度超过96%,优于其他迁移学习方法。(3)针对机械装备故障样本稀少的问题,设计一种自监督宽度残差孪生网络(Self-Supervised Wide Residual Simaese Network,SS-WRSN)。该方法在ASSL模型的基础上,结合元学习与特征融合技术,提升ASSL的自学习能力,减小其对样本信息的依赖,同时利用图卷积神经网络实现对样本特征信息的整合与蒸馏,使模型充分利用有效的故障信息,增强模型的泛化性能。通过搭建的船舶推进轴系实验台和Paderborn轴承数据集验证所提出方法的有效性。结果显示:该方法在故障样本稀少场景下的平均诊断精度超过97%,相比匹配网络和原型网络两种方法,诊断性能提升3%,具有良好的工程应用价值。更多还原

【Abstract】 Ships are the basis and carrier of ocean transportation and an indispensable part of the national industry of equipment manufacturing.Due to operation under variable loads and marine environment with high salt spray,the performance of core mechanical equipments on ships degrade relatively quickly and are accompanied by frequent failures,which easily cause catastrophic accidents without timely detection and repair.Therefore,in order to ensure the safety and reliability of ship operation,it is urgent to develop intelligent and highprecision fault diagnosis technology for mechanical equipment.With the development of data monitoring technology and the popularization of artificial intelligence techniques,the intelligent fault diagnosis methods based on machine learning have become one of the research hotspots in this field.However,most of the current intelligent diagnosis models are driven by ideal condition monitoring data,which means that the distribution of data is required to be consistent and fully annotated.These methods do not take into account the mechanical equipment,such as the lack of data labels,the scarcity of fault samples and variable operating conditions,which will lead to general defects in intelligent models including poor robustness and low adaptability.Weakly supervised learning can employ a variety of incomplete or damaged labeled data to build high-performance diagnostic models and improve fault diagnosis accuracy.Therefore,this dissertation conducts research on key technologies of intelligent fault diagnosis of typical mechanical equipment based on weakly supervised learning theory.The main works are presented as follows:As for the problem of lack of annotation information in mechanical equipment monitoring data,a fault diagnosis method based on Adversarial Self-Supervised Learning(ASSL)is proposed.By constructing a new contrastive learning paradigm,the model is able to independently mine and automatically cluster the hidden layer correlation between unlabeled samples,which avoids the collapse of model and enhances the self-learning ability and generalization ability of the encoder.Compared with traditional supervised learning,ASSL’s dependence on sample labeling information is significantly reduced during training,which only needs fine-tuning on a small number of samples to achieve highprecision fault diagnosis in downstream tasks.The proposed method is verified by building a high-speed axial piston pump test bench.The experimental results show that the this method performs better than existing supervised learning methods in the case of dignosis with rare labeled samples.As for the problem of variable distribution of monitoring data and hidden noise in monitoring signals caused by changing working conditions of mechanical equipment,a fault diagnosis method based on Self-Attention and Subdomain Adaptive Adversarial Network(SASAAN)is proposed.Firstly,a feature extractor with a self-attention module is constructed,which eliminates noise interference in the monitoring signal through a weight decision-making mechanism and improves the robustness of the model.Secondly,the domain adversarial techniques are combined with subdomain adaptation techniques to reduce the difference between the global distribution of samples in the source and target domains and the distribution of related subclasses,which enhances the model’s ability to extract high-dimensional domain-invariant features.Finally,the effectiveness of the proposed method is verified through the Ottawa bearing dataset.The results show that in the case of strong noise and variable working conditions,the average diagnostic accuracy of this method on the target working condition exceeds 96%,which is better than other transfer learning methods.As for the problem of scarce fault samples of mechanical equipment,a Self-Supervised Wide Residual Simaese Network(SS-WRSN)is designed.This method combines metalearning and feature fusion technology on the basis of the ASSL model,so as to improve the self-learning ability of ASSL and reduce its dependence on sample information.Further,the graph convolutional neural network is used to integrate and distill the feature information of samples,which enables the model to make full use of effective fault information to enhance its generalization performance.The proposed method is verified by the ship propulsion shafting test bench and the Paderborn bearing dataset.The results show that the average accuracy of this method exceeds 97% in the case of diagnosis with few fault samples.Compared with the matching network and prototype network,the diagnostic performance is improved by 3%,which has good engineering application value.更多还原