基于迁移学习的往复压缩机故障诊断机制及预测预警模型研究

The intelligent equipment diagnosis centered on artificial intelligence mostly relies on data driven models, with the hypothesis of identical/similar probability distributions between training and testing data. However, the complex operating conditions of oil and gas equipment leads to the variability and different probability distributions of monitoring data, resulting in difficulty to reuse monitoring data and poor universality of diagnosis model. This project introduces transfer learning theory to break the hypothesis of identical/similar probability distributions, and takes reciprocating compressor (key equipment in oil and gas industry) as research object. This project starts with response properties and components failure of reciprocating compressor under complex operating conditions, studies the effective methods to extract and mine fault representative features, puts forward a new diagnostic mechanism based on transfer learning theory, and establishes fault early warning model based on uncertainty analysis of fault prediction. This project focuses on the dynamic properties of monitoring data under complex operating conditions, transferability measure of transferred diagnosis, symbiosis transfer mechanism between training domain and testing domain. The fundamental framework of transferred fault diagnosis and prognosis under complex operating conditions is then constructed and validated via fault simulation, experimental study and theoretical analysis. It is anticipated to provide a new approach and new method for monitoring data reuse and accumulated-learning based fault diagnosis model. It is of significance to make maintenance strategy and ensure the equipment safety in oil and gas industry.

以人工智能为核心的设备智能诊断多依据数据驱动模型，并以训练数据与测试数据满足概率同分布假设为前提；而油气设备运行的复杂工况使得监测数据呈现出多变性和异分布，从而导致监测数据重用难、诊断模型普适性差。本课题引入迁移学习理论破除数据同分布的假设，以故障频发的油气关键设备-往复压缩机为研究对象，从分析复杂工况下的故障响应特性和部件失效过程入手，研究多变性特征的有效提取与挖掘方法，提出基于迁移学习的故障诊断机制，构建基于趋势发展不确定性分析的故障预测预警模型。重点研究复杂工况下监测数据的多变性规律、设备故障迁移诊断的准则、训练域与测试域间的共生迁移诊断机制等关键问题。最终通过故障仿真、模拟实验和理论分析，构建复杂工况下设备迁移诊断与预测基础性研究框架。项目预期为复杂工况下监测数据重用及诊断预测模型可积累式学习提供新思路、新方法，对制定科学的维修决策、保障设备安全可靠运行具有重要的理论和现实意义。

以人工智能为核心的设备智能诊断多依据数据驱动模型，并以训练数据与测试数据满足概率同分布假设为前提；而油气设备运行的复杂工况使得监测数据呈现出多变性和异分布，从而导致监测数据重用难、诊断模型普适性差。本项目基于迁移学习，研究往复压缩机故障诊断机制及预测预警模型，主要成果包括：基于虚拟样机技术进行动力学仿真实验，分析了复杂工况下曲轴间隙故障和气阀故障的特性及监测数据的多变性规律；提出基于互信息的非监督式特征选择方法和等概率关联规则的特征挖掘方法，对复杂工况下故障特征进行了有效挖掘和提取；提出基于辅助模型的领域自适应诊断模型及定量评判迁移效果的指标；利用迁移学习实现跨场景迁移诊断，并利用累积特征和主成分分析方法构建设备性能退化指标；提出基于卷积神经网络和长短期记忆网络的深度模型来进行故障预测预警，为部件的性能退化及预测预警研究提供了新的方法。共发表期刊论文21篇（其中SCI论文10篇），会议论文13篇；出版学术专著1部；授权发明专利2件、授权实用新型1件、受理发明专利4件；登记软件著作权2件；获国家技术发明二等奖1项、教育部科技进步一等奖1项；培养毕业博士2人、毕业硕士8人、在读博士2人、在读硕士4人。