大数据驱动基于深度强化学习的智能装备预测性维护研究

Predictive maintenance is one of the key and common techniques for equipment machining industry to be resolved. The purpose of this proposal is to do research on new methods of big data-driven predictive maintenance for intelligent equipment by implementing full integration of big industrial data and artificial intelligence with intelligent equipment. Its main research contents are as follows: First of all, methods of massive monitoring data cleaning and reconstruction, and early weak fault features enhancement will be proposed, which eliminate random missing, noise interference or low signal-to-noise ratio in monitoring data. Then, a deep convolution neural network model will be built to excavate weak fault features by layer-by-layer and identify early fault symptom. And, an adaptive useful life prediction method based on deep reinforcement learning will be proposed to reveal its change over time and realize automatical adjustment over different environment. Further, a generative adversarial nets will be constructed and diagnosis of fault location and cause will be realized. On this basis, a multi-objection optimization and decision model for predictive maintenance will be established to explore the optimal maintenance mode and time interval by considering various uncertainties. Finally, a software and hardware platform for predictive maintenance of intelligent equipment will be developed so as to provide the application support for the above methods. Predictably, it has important theoretical significance and engineering application value for creating the operation safety guarantee system of intelligent equipment, promoting the ability of intelligent equipment to run stable and reliable for a long time, reducing the cost of equipment operation and maintenance and so on.

预测性维护是装备制造业亟待突破的关键共性技术。本申请项目将工业大数据、人工智能与智能装备进行深度融合，针对性地研究大数据驱动基于深度强化学习的智能装备预测性维护性新技术，主要包括：探索海量异构数据清洗与重构以及微弱故障特征增强方法，消除监测数据中存在的随机缺失、噪声干扰或早期故障信号弱；建立深度卷积神经网络模型，逐层挖掘微弱故障特征与感知早期故障征兆；提出基于深度强化学习的运行寿命自适应预测方法，实现不同环境下运行寿命变化的自主调节；设计一种具有真假互搏的生成对抗网络，破解故障诊断中遇到的故障样本匮乏难题；建立不确定条件下基于多目标优化的预测性维护决策模型，揭示装备最佳的预测性维护方式与时机；研制智能装备预测性维护软硬件平台，为上述技术的广泛应用提供支撑。研究结果对于从纵深方向发展装备运行安全的科学保障体系，提升装备可靠运行能力和降低装备运营维护成本等都具有重要的理论意义和工程应用价值。

预测性维护是装备制造业亟待突破的关键共性技术。本项目将工业大数据、人工智能与智能装备进行深度融合，针对性地开展了大数据驱动基于深度强化学习的智能装备预测性维护新技术研究，主要的研究内容与取得的重要结果包括：研究了多源异构数据清洗与重构及增强等数据预处理方法，考虑智能装备失效机理不明、早期故障信号微弱，且易受噪声干扰或掩盖，提出了基于残差收缩关系网络的智能装备早期故障感知方法；发明了基于Stacking集成学习的智能装备无监督故障感知方法，实现了对未知标签故障样本深度感知；设计了孪生数据与监测数据驱动的智能装备早期故障感知方法，解决了早期故障样本数据匮乏的问题。研究了不同类型监测数据的劣化特征提取与消减方法，发明了基于嵌套式长短期记忆神经网络和高斯滤波的剩余寿命预测方法，解决了现有方法因梯度消失而导致预测精度低的问题；针对智能装备运行寿命同时受到内部劣化和外部环境变化影响的问题，提出了变工况下基于多维度循环神经网络的智能装备剩余使用寿命预测方法；同时，设计了基于集成双向循环神经网络的寿命预测方法，实现了智能装备寿命的精确预测。研究了基于深度强化学习的智能装备自适应故障诊断方法，大幅提升了诊断模型在变工况下的诊断性能；提出了基于域自适应对抗网络的故障诊断方法，实现对极度匮乏故障样本的在役装备故障的精准诊断；针对监测数据存在变分布的问题，设计了基于残差深度子域自适应网络的跨多域故障诊断方法；考虑智能装备运行工况多变、劣化过程随机、运行寿命波动等不确定因素的影响，提出了基于健康监测的多状态智能装备预测性维护策略。研制了基于云-边-端协同的智能装备预测性维护软硬件平台，为上述技术的广泛应用提供支撑。研究结果对于从纵深方向发展装备运行安全的科学保障体系，提升装备可靠运行能力和降低装备运营维护成本等都具有重要的理论意义和工程应用价值。