广度自适应的深度迁移度量学习方法及其在复杂机械设备故障诊断与预测中的应用研究

In the face of the mechanical big data considering the effect of strong nonlinearility and low SNR, aiming at the difficulty of that are brought to the fault high-accuracy diagnosis and prediction of mechanical equipment by the imbalance and difference of probability distribution of data samples, less of fault samples and underutilized data samples of different equipment of the same type, on the basis of deep learning, transfer learning and metric learning, a deep transfer metric learning method based on Yu norm is proposed. Based on the method, the deep transfer metric learning diagnosis model with wideness self-adaption is built and applied to the fault diagnosis of mechanical equipment, which is combined with multi-variables information and correlation analysis method and bayes belief method. And using the optimal muliti-variables information selected by the correlation analysis, a series of deep transfer metric prediction models based on Yu norm are built respectively. And through the combination of these models and the SOM the deep transfer metric learning prediction model with widenss self-adaption is also built and applied to the fault prediction of mechanical equipment. At last,with rolling mill as the object of the research a set of system for fault diagnosis and fault prediction is built and used to validate the effectiveness of these supposed methods. The research project gives a new way for the fault diagnosis and fault prediction of complex mechanical equipment and is of vital importance for the improvement of the safety and reliability and the avoidance of terrible accident.

本项目面向受强非线性和低信噪比影响的机械大数据，针对数据样本不均衡和概率分布不同，且故障样本少、同型号不同设备的状态数据没被利用，导致设备故障高精度诊断与预测难的问题，以深度学习、迁移学习和度量学习为基础，提出基于Yu范数的深度迁移度量学习方法，在此基础上，结合多元信息谱，利用相关性分析方法和贝叶斯置信法，建立广度自适应的深度迁移度量诊断模型，实现机械设备的故障诊断；结合基于相关性分析优选的多元信息谱，依此分别建立基于Yu范数的深度迁移度量预测模型，利用自组织映射模型SOM，建立广度自适应的深度迁移度量预测模型，实现设备的故障预测。最后，以轧机机组为研究对象，建立了一套故障诊断和预测系统，用于验证本研究所提出方法的有效性。本项目研究将为大数据下复杂机械设备故障诊断和预测提供了一条新的途径，对提高机械设备运行安全性和可靠性、避免恶性事故发生具有重大意义。

本项目面向受强非线性和低信噪比影响的机械大数据，针对数据样本不均衡和概率分布不同，且故障样本少、同型号不同设备的状态数据没被利用，导致设备故障高精度诊断与预测难的问题，以深度学习、迁移学习和度量学习为基础，建立了深度置信神经网络、深度度量网络和深度Yu范数度量网络模型，并结合基于EEMD分解的多元信号，构建了基于贝叶斯置信法的广度自适应深度Yu范数度量学习模型和基于投票表决法的集成深度神经网络模型；建立了深度迁移学习和深度迁移度量学习以及集成迁移学习网络模型，有效实现了轴承和齿轮等设备的高精度故障诊断分析；建立了基于深度学习/深度度量学习和自组织映射网络模型的健康因子，并结合多个轴承性能退化数据，提出了基于深度迁移的LSTM预测模型以及集成深度迁移预测模型，实现了轴承性能退化趋势的高精度预测分析。本项目研究将为大数据下复杂机械设备故障诊断和预测提供了一条新的途径，对提高机械设备运行安全性和可靠性、避免恶性事故发生具有重大意义。