机电系统部件故障诊断与预测：一种基于深度学习与集成学习的方法

This proposal aims to solving existing problems in the field of fault diagnosis and prognostics for electro-mechanical system components, including high cost and uncertainty in fault feature extraction, low adaptability and generalization of single diagnosis and prognostic model, and low accuracy of long-term prognostic. In this study, deep learning and ensemble learning are combined together to improve the performance of diagnosis and prognostic model. First, unsupervised self-learning feature extraction is performed based on model-averaging deep learning methods. Then, ensemble learning methods are employed to adaptively integrate several deep-neural-network based classifiers which are trained based on the extracted fault features. Finally, an ensemble of deep recurrent neural network is utilized to realize collaborative prognostic for performance degradation. In this study, two key techniques are broken, that is, optimization of the topological structure and training parameters of the deep neural network; and ensemble mechanism construction and automatic model selection for complex diagnosis and prognostic tasks. The proposed method can effectively improve the adaptability and generalization of diagnostic and prognostic models, thus providing a higher diagnostic and prognostic accuracy for electro-mechanical system components.

针对典型机电系统部件故障诊断与预测中面临的人工故障特征提取不确定性和高成本、单一诊断预测模型适用性低与泛化能力差、以及中长期预测精度低等问题，借鉴人类认知过程的原理与集成学习思想，将认知计算领域与机器学习领域最前沿的深度学习理论与集成学习理论相结合，依次开展基于模型平均深度学习的自主故障特征提取、基于自适应选择性集成的故障诊断模型构建、以及基于集成深度循环网络的多模型协同预测方法研究。突破深度神经网络拓扑结构与训练参数的最优化问题、复杂诊断和预测任务中模型的自适应优选与集成认知机制构建等关键科学问题，实现机电系统部件鲁棒故障特征自主学习与性能衰退演化规律挖掘，增强诊断预测模型的适用性与泛化能力，提高机电系统部件故障诊断与中长期预测精度。

针对典型机电系统部件故障诊断与预测中面临的人工故障特征提取不确定性和高成本、单一诊断预测模型适用性低与泛化能力差、以及中长期预测精度低等问题，借鉴人类认知过程的原理与集成学习思想，将认知计算领域与机器学习领域最前沿的深度学习理论与集成学习理论相结合，依次开展了基于模型平均深度学习的自主故障特征提取、基于自适应选择性集成的故障诊断模型构建、以及基于集成深度循环网络的多模型协同预测方法研究。突破了深度神经网络拓扑结构与训练参数的最优化问题、复杂诊断和预测任务中模型的自适应优选与集成认知机制构建等关键科学问题。本研究所提方法以舱门驱动机构旋转作动器、柱塞式液压泵、锂电池、航空发动机等为对象开展了相关案例应用验证，试验结果证实所提方法能够有效实现机电系统部件鲁棒故障特征自主学习与性能衰退演化规律挖掘，增强诊断预测模型的适用性与泛化能力，提高机电系统部件故障诊断与中长期预测精度。