【作者】 张强；

【导师】 田喆；

【作者基本信息】 天津大学 ， 土木工程， 2023， 博士

【摘要】 随着暖通空调系统(Heating,ventilation,and air-conditioning,HVAC)的运维管理模式向精细化和智能化转变,发展能够实时感知系统健康状态的智能故障诊断方法是保障系统节能高效运行的重要举措。数据驱动的故障诊断方法具备高准确性、高建模效率的优点,契合智能故障诊断对于诊断模型的需求,因此已成为智能故障诊断领域的研究重心。数据驱动故障诊断模型的构建依赖丰富的标注数据(带有正常状态和各类故障状态标签的监测数据)。然而,HVAC系统虽有大量的监测数据,但标注数据是有限甚至缺失的,在此情形下,所得故障诊断模型的泛化能力差,甚至无法支撑故障诊断模型的构建。仿真模型作为实际系统的数字孪生体,可描述实际系统个性化的故障特性知识,且具备搭建容易和故障实验成本低的优点,使其成为一种优质的标注数据来源。因此,如果能在实际系统的故障诊断过程中合理融入仿真数据源,则有可能在标注数据短缺场景下获得具有良好性能的故障诊断模型。在这种背景下,本研究面向实际系统仅有限正常数据可获取的诊断场景,提出了一种以仿真模型作为标注数据源、融合仿真和实际数据建模的HVAC系统跨数据源故障诊断方法。主要研究工作如下:(1)为了获得正常状态和故障状态下的运行数据,本研究开发了面向标注数据增强需求的仿真模型。首先,搭建能够模拟实际系统水力、热力和电力特性的全性能仿真模型,以充分反映故障的影响。其次,使用可获取的实际正常数据对全性能仿真模型实施校准,提升仿真模型的精度。所用校准方法为面向具有多校准变量且变量间耦合特性的全性能仿真模型提出的分层模型校准方法。最后,以校准模型为基础,引入故障模拟技术,使模型具备故障模拟能力。基于上述方法为一处HVAC系统案例搭建仿真模型,结果表明所得模型既可准确描述实际系统正常状态下的性能,又可真实复现实际系统故障状态下的征兆。因此可利用该模型生成正常和各类故障运行数据,为故障诊断建模提供标注训练数据。(2)为有效利用仿真标注训练数据中的故障特性知识,本研究提出了基于特征变换迁移学习的故障诊断模型融合构建方法。首先,使用实际系统可获取的正常数据辅助仿真数据实施特征变换,以减小两类数据之间的潜在差异。其次,利用特征变换后的仿真数据训练数据驱动故障诊断模型。所提方法基于先验知识(描述仿真数据和实际数据的共性结构和数值偏差)分别规定了仿真正常和故障数据的特征变换方式,从而保障仿真数据在无实际故障数据约束条件下也可实施有效的特征变换。案例结果表明,所提融合建模方法能明显减小仿真数据和实际数据之间的潜在差异,获得准确的故障诊断模型(平均准确性为81.7%),并较传统的机器学习建模方法和预训练-微调迁移学习方法有明显的性能优势(平均准确性提升幅度为9.6%)。所提方法可有效利用仿真标注数据为实际系统搭建故障诊断模型,获得的诊断模型可对实际系统的健康状态进行预测。(3)为了判断所得故障诊断模型在实际应用时的诊断准确性以及掌握故障诊断模型对于仿真模型精度的需求,本研究提取了仿真模型精度和故障诊断模型准确性之间的定量关联。具体包括故障诊断场景的构建、仿真模型精度-诊断准确性数据集的构造以及仿真模型精度和诊断准确性之间定量关联的挖掘三个步骤。方法实施过程中会引入聚类算法和关联规则挖掘算法提升定量关联的提取效率。案例结果表明,所提方法能以低成本获得高可靠度的定量关联。具体地,聚类算法可在保证场景丰富性的同时缩减诊断场景数量,减小了计算负担;关联规则挖掘算法可从大量的仿真模型精度和诊断准确性结果中自动挖掘二者关联,提升了知识归纳效率;所得定量关联可适用于近80%的真实诊断任务,表现出高可靠性。定量关联知识既可判断故障诊断模型在实际应用时可能达到的准确性水平,又可指明为达到工程需求的诊断准确性水平时仿真模型应达到的精度水平。综上,全性能仿真模型提供了丰富的标注训练数据,基于特征变换迁移学习的故障诊断模型融合构建方法则完成了由仿真至实际的跨数据源建模任务,结合二者,可获得一套行之有效的面向标注数据短缺问题的故障诊断方法。而提取出的仿真模型精度和诊断准确性之间的定量关联规律则为该方法的实施提供了精准指导。更多还原

【Abstract】 With the transformation of the operation and maintenance mode of heating,ventilation,and air conditioning(HVAC)system to refinement and intelligence,the development of intelligent fault diagnosis methods that can continuously perceive the health status of the HVAC system is an important measure to ensure the energy-saving and efficient operation of the system.The data-driven fault diagnosis method has the advantages in accuracy and flexibility,which meets the requirements of intelligent fault diagnosis for fault diagnosis models.Therefore,it has gained increasing popularity in the field of intelligent fault diagnosis.The development of data-driven fault diagnosis models relies on rich labeled data(the data with health condition labels,such as the normal condition or various fault conditions).However,although the HVAC system has a large amount of measured data,the labeled data is limited or even missing,which leads to poor generalization ability of the fault diagnosis model and even cannot support the development of the fault diagnosis model.The simulation models,as digital twins of actual systems,can describe personalized fault characteristics knowledge of actual systems,and have the advantages of easy development and low cost of fault experiments,making it a high-quality labeled data source.Therefore,if simulation data sources can be reasonably integrated into the fault diagnosis process of actual systems,it is possible to obtain a fault diagnosis model with good performance in scenarios of labeled data shortage.In this context,this study establishes a cross-domain fault diagnosis method for HVAC systems with the simulation model as the labeled data source and fusion simulation and actual data modeling for the diagnosis scenarios where only limited normal data can be obtained from the actual system.The main research work is as follows:(1)In order to provide operational data under normal and fault conditions,this study establishes a simulation model for the augmentation of labeled data.Firstly,develop a full-performance simulation model that can simulate the hydraulic,thermal,and electrical characteristics of the actual system to fully reflect the impact of faults.Secondly,calibrate the full-performance simulation model using available actual normal data to improve the accuracy of the simulation model.The calibration method used is a hierarchical model calibration method proposed for a full-performance simulation model with multiple calibration variables and coupling characteristics between variables.Finally,based on the calibration model,the fault simulation technique is introduced to enable the model to have fault simulation capabilities.Based on the above method,a simulation model was built for an HVAC system case,and the results showed that the obtained model can accurately describe the normal performance of the actual system and reproduce the fault symptoms of the actual system.Therefore,the obtained simulation model can be used to generate normal and various types of fault operation data,providing labeled training data for fault diagnosis.(2)In order to effectively use the fault characteristic knowledge in the simulation training data,this study proposes a fusion modeling method for fault diagnosis based on feature transformation transfer learning.Firstly,use normal data available from the actual system to assist simulation data in implementing feature transformation to reduce potential differences between the two types of data.Secondly,the simulation data after feature transformation is used to train the data-driven fault diagnosis model.The proposed method specifies the feature transformation methods for normal and fault data in simulation based on prior knowledge(describing the common structure and numerical deviation between simulation data and actual data),thereby ensuring that simulation data can be effectively transformed without actual fault data constraints.The case results showed that the proposed fusion modeling method can significantly reduce the potential difference between the simulation data and the actual data,obtain an accurate fault diagnosis model(with an average accuracy of 81.7%),and has obvious performance advantages over the traditional machine learning modeling method and the pre-training and fine-tuning transfer learning method(with an average accuracy improvement of 9.6%).The proposed method can effectively utilize simulation-labeled data to develop a fault diagnosis model for the actual system,and the obtained diagnostic model can determine the health status of the actual system.(3)In order to determine the diagnostic performance of the obtained fault diagnosis model in practical application and to understand the requirements of the fault diagnosis model for simulation model accuracy,this study extracts the quantitative associations between simulation model accuracy and fault diagnosis model accuracy.Specifically,it includes three steps:constructing fault diagnosis scenarios,constructing the datasets including simulation model accuracy and diagnostic accuracy,and mining quantitative associations between simulation model accuracy and diagnostic accuracy.During the implementation of the method,clustering algorithms and association rule mining algorithms are introduced to improve the efficiency of association extraction.The case results indicated that the proposed method can achieve high-reliability quantitative associations at a low cost.Specifically,clustering algorithms can reduce the number of diagnostic scenarios while ensuring scene richness,thereby reducing computational burden;The association rule mining algorithm can automatically mine the association between a large number of simulation model accuracy and diagnostic accuracy results,improving the efficiency of knowledge induction;The obtained quantitative associations can be applied to nearly 80%of real diagnostic tasks,demonstrating high reliability.Quantitative correlation knowledge can not only determine the level of accuracy that a fault diagnosis model may achieve in practical application,but also indicate the level of accuracy that a simulation model should achieve in order to meet the diagnostic accuracy level required by engineering requirements.In conclusion,the full-performance simulation model provides rich labeled data,and the fusion modeling method for fault diagnosis based on feature transformation transfer learning completes the task of cross-domain modeling from simulation to reality.Combining the two,an effective cross-domain fault diagnosis method for labeled data shortage can be obtained.The extracted quantitative associations between the simulation model accuracy and diagnostic accuracy can provide precise guidance for the implementation of this method.更多还原