基于深度学习与信息融合的机械系统健康评估方法研究

项目名称： 基于深度学习与信息融合的机械系统健康评估方法研究

项目编号： No.51475170

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 机械、仪表工业

项目作者： 李巍华

作者单位： 华南理工大学

项目金额： 80万元

中文摘要： 针对大型风力发电装备等服役环境恶劣的机械系统，结合项目组在故障特征提取、数据融合与半监督流形学习诊断方法等方面的研究成果，深入研究复杂机械系统运行状态健康评估的问题。提出基于深度学习的故障时频图像特征提取方法，实现对关键部件微弱故障的特征增强。提出基于深层置信网络的多源传感器数据的融合诊断方法，实现对多源异构信息的一致性融合表示。利用深度学习强化传统监督式分类、半监督分类学习算法的性能，建立逐层学习的判别性深层卷积网络结构，提出一种基于正常状态样本集的故障预测方法和多故障分类识别方法。集成隐马尔科夫模型与深层神经网络，建立基于深度学习的机械系统健康评估模型，将深度网络结构纳入隐马尔科夫状态转移的识别过程，识别模式的演化特征，实现对机械系统运行状态的健康评估。

中文关键词： 深度学习；特征提取；故障诊断；信息融合；健康评估

英文摘要： Based on the studies of feature extraction，data fusion，semi-supervised fault diagnostics and manifold learning, we investigate the problems of complex mechanical system operation state health assessment, such as wind turbine health prognosis. A deep-learning based feature extraction method will be proposed to identify the time-frequency image of incipient fault signals, and a DBN could be used to represent the multi-source isomerous sensor data. The classical supervised and semi-supervised learning algorithms can be enhanced by use of deep learning, and the discriminative deep convolutional architecture could be constructed to learn the nature of the data from insufficient samples. Based on this study, we will propose a deep CNN based fault prognosis method to diagnose the incipient failures and classify different faults. By integrating Hidden Markov Model and Deep Neural Networks, a deep-architecture model will be constructed to perform machine health assessment. The deep network will be used for the state recognition during HMM learning process, and to capture the state transition information, which could be used to predict the anomaly occurrence and its propagation. Then machine performance could be effectively evaluated by this deep-learning based prediction model.

英文关键词： deep learning;feature extraction;fault diagnosis;information fusion;health evaluation