【作者】 李冠男；

【导师】 陈焕新；

【作者基本信息】 华中科技大学 ， 制冷及低温工程， 2017， 博士

【摘要】 针对制冷空调系统故障诊断研究领域存在的数据利用率低、诊断模型结构单一、缺少知识解析等重点问题,本研究提出一种基于数据挖掘算法模型的故障诊断与用能模式识别的应用框架,分别以螺杆式冷水机组、离心式冷水机组、多联式机组等典型制冷空调系统作为研究对象,进行了传感器故障检测和诊断、热力故障检测、热力故障水平预测的性能优化,热力故障工况下系统能耗、运行模式识别和用能相关关联规则挖掘分析等一系列研究,为制冷空调系统的健康、高效、节能运行提供保证。首先,基于制冷空调专业知识,归纳文献中的关键故障解耦特征的统计学变化规律,为建模变量筛选提供的先验知识库。在模型选择上,深入分析数据挖掘原理,对比监督学习、无监督学习挖掘算法的数学本质,结合故障诊断、用能模式识别各个研究内容的研究目标,构建了基于数据挖掘算法模型的故障诊断与用能模式识别的应用框架,分别包括故障诊断、能耗模式识别的两个子框架。其次,故障诊断本质是对制冷空调系统的故障类别进行分类、故障水平进行预测,对应于监督学习类分类预测算法的问题范畴。依据基于监督学习算法的故障诊断子框架,针对螺杆式冷水机组温度、流量传感器故障,利用单分类的支持向量数据描述(SVDD)算法,建立故障检测和诊断模型,提出一种新的基于距离的统计监测量和贡献率变化图的故障检测和诊断方法。引入固定偏差、漂移、精度下降和彻底失效四种典型传感器故障,验证子框架下故障诊断模型的故障检测灵敏性(sensitivity)和诊断准确性(accuracy),分析影响诊断性能的重要因素。针对离心式冷水机组的七种典型热力故障,依据故障诊断子框架,提出一种新的主元分析-残差空间-支持向量数据描述(PCA-R-SVDD)故障检测方法,利用美国采暖、制冷与空调工程师学会(ASHRAE)RP-1043项目中的故障实验数据,验证该方法的有效性,并从统计监测量、数据分布和故障检测正确率(correct ratio)三方面分析该方法的故障检测灵敏性。主元分析(PCA)残差子空间有效分离故障数据和正常数据、SVDD故障检测边界紧凑,明显提高PCA-R-SVDD方法的故障检测正确率,对于其中六种常见热力故障,即使在最轻微的故障水平1条件下,检测正确率仍高于50%。对轻微故障的敏感度增加。再次,为验证故障诊断子框架在不同制冷空调系统间的通用性,针对多联式空调系统制冷剂充注量故障,提出一种新的基于支持向量回归(SVR)算法改进虚拟传感器(VRC)的制冷剂充注量故障水平预测模型。对比多联式空调系统自带传感器的测量变量与传统虚拟传感器模型预测误差之间的相关性程度,筛选大相关性变量作为改进模型新增输入,提高数据的利用率。以一台额定冷量为29.8kW的多联式空调机组为例,设计63%-130%制冷剂充注量故障实验,获取故障实验数据验证基于SVR改进预测模型的故障水平预测性能。该改进模型显著地降低了制冷剂真实充注量>90%条件下的预测误差,总体平均预测误差仅为5.51%。最后,利用多联式空调系统制冷剂充注量故障能耗数据验证无监督用能模式识别子框架,提高能耗数据的利用率。用能模式识别具有一定的开放性,适宜用无监督算法进行探索分析。依据无监督用能模式识别子框架,提出一种基于聚类数据划分与关联规则挖掘分析相结合的集成分析方法,利用内部验证指标Dunn值筛选k-means聚类方法,有效划分能耗数据为三个数据簇,识别出三种可解释的系统能耗模式;分簇在各个能耗模式下利用Aprior算法进行关联规则分析,对比挖掘出用能相关关联规则,识别出系统中的异常用能模式及相关关联规则。总之,本研究提出的基于数据挖掘的故障诊断与用能模式识别的集成应用框架,及其在螺杆式冷水机组、离心式冷水机组、多联式空调系统中的实际验证结果等主要成果已在多个国际期刊上发表,对于提升制冷空调系统数据利用率,提高模型故障检测正确率、灵敏性和诊断准确率,以及向系统用能模式分析模型延伸拓展,具有一定应用价值与意义,值得进一步研究。更多还原

【Abstract】 As for the fault detection and diagnosis(FDD)researches on refrigeration and air conditioning systems,there exist some major problems,i.e.,the under-utilization of operating data,the lack of energy consumption patterns recognition models and the shortage of inadequate interpretation for data-driven fault diagnosis results,etc.Therefore,this study attempts to overcome these limitations by proposing an integrated framework for fault detection and diagnosis and energy consumption patterns recognition for refrigeration and air conditioning systems based on data mining(DM)techniques.We selected some refrigeration and air conditioning systems like,screw chillers,centrifugal chillers and variable refrigerant flow(VRF)systems as main research objects to validate the data mining-based framework.After collecting the sensor fault data,thermdynamic fault data and energy consumption data,we conducted studies such as,sensor fault detection and diagnosis,thermdynamic fault detection and fault severity levels prediction,power consumption patterns identification and association rules mining,which as a result can ensure the system operating reliably and cost effectively.Firstly,based on expert knowledge and fault experimental results in the literature,critical decoupling features for multiple faults and their statistics characteristics were summarized and prepared for developing the integrated framework for fault detection and diagnosis and energy consumption patterns recognition.The entire framework for refrigeration and air-conditioning systems were established using two different DM-based models,supervise and unsupervised learning models.According to the different targets,two sub frameworks,FDD and ESA were developed to improve the FDD performance,identify power consumption patterns and fault energy saving energy consumption analysis,respectively.Secondly,fault diagnosis is essentially to classify the fault types and predict the fault severity levels.The supervized learning data mining algorithms are most suitable for solving the classification and prediction problems.In accordance with the data mining-based sub-framework for fault detection and diagnosis,we proposed a sensor fault detection and diagnosis model for screw chiller systems using the supervised one-class classification algorithm,support vector data description(SVDD).This model uses a distance-based D-statistic plot to detect sensor faults and a new distance variation-based DV-contribution plot diagnose the sensor faults.Four typical sensor faults including fixed bias,drifting,precision degradation and complete failure with different fault severity levels were introduced into the temperature and water flow rate sensors,respectively.The fault data were used to assess the fault detection sensitivity and fault diagnosis accurucay of the SVDD-based model.Also,the effects on FDD results were considered and evaluated in terms of fault types,severity levels and the statistical property of training data.Thirdly,in order to detect seven typical thermodynamic faults in the centrifugal chiller system,this study proposed a novel principle component analysis-residual-support vector data description(PCA-R-SVDD)method based on the the data mining-based sub-framework for fault detection and diagnosis.Centrifugal chiller experimental data from the ASHRAE Research Project 1043(RP-1043)was used to evaluate the proposed model.The sensitivity for fault detection was analyzed on three aspects,monitoring statistic,data distribution and fault detection correct ratio.Instead of principle component subspace(PCs),the proposed model constructed the SVDD model in the residual subspace(Rs).The SVDD based distance based monitoring statistic was used for fault detection.The proposed method showed significant improvements of fault detection correct ratio comparing with the traditional methods due to the better fault data distribution and tighter monitoring statistic.For six common faults of the seven,at least 50%of the fault data can be correctly detected even at the least severe fault level.Forthly,this study further performed a to verify the commonality of the DM-based FDD framework between different refrigeration and air conditioning systems,this study proposed a Pearson-SVR-based method to improve the prediction performance.Traditional virtual refrigerant charge(VRC)sensor models perform well at undercharge situations but produce large prediction errors at overcharge situations.When the refrigerant charge level(RCL)is over 90%,the Person correlation coefficient data-based method was introduced to select the additional input variables to modify the VRC models.Results reveal that the overall prediction errors for SVR based modified VRC model(SVR-VRC)is 5.53%in the range of 63.64%-130%RCL.The SVR-VRC model improves the VRC models and extends the application in the VRF system when only the system self-provided sensor measurements are used.Finally,the tasks for ESA are open-ended.The unsupervized learning DM algorithm is just suitable for exploratory data analysis.Based on unsupervised ESA framework,an integrated clustering and association rules mining analysis(ICAA)method was put forward for VRF refrigerant charge fault energy consumption analysis.The agglomerative hierarchical clustering method was selected using the Dunn index.It was employed to divide the original data set and identify the operation mode of energy consumption pattern for VFR system.The Aprior algorithm was seletectd for association rules analysis.The abnormal operating data and interesting energy consumption patterns were obtained by comparing the useful rules hidden in various data clusters,thus improve the data utilization ratio.Overall,the proposed data mining-based framework for fault diagnosis and energy consumption pattern recognition application and its validation results have been published in some international journals.The framework has been proven to be quite effective on improving the data utilization ratio,the fault detection ratio and the diagnostic accuracy.Moreover,it has extened the application of data-driven models from pure fault diagnosis to energy consumption pattern recoginition.It is a promising perspective and worthy of further investigation.更多还原