【作者】 韩华；

【导师】 谷波；

【作者基本信息】 上海交通大学 ， 制冷及低温工程， 2012， 博士

【摘要】 鉴于制冷空调系统日益复杂,系统故障难以识别,且会导致能耗增加(多达30%),室内外环境恶化,设备可靠性、安全性及运行经济性下降等诸多问题,有必要对故障检测及诊断进行相关研究,以便及时排除故障,保证系统正常运行。本研究围绕制冷系统故障检测与诊断问题,从故障指示特征智能提取,到制冷系统单发及并发故障(多故障),顺序递进,集成故障诊断、数据挖掘及模式识别领域各种智能方法,实现检测与所有故障确诊一步完成,对轻微故障亦性能良好,诊断正确率高,虚警率及诊断用时少;并基于混淆矩阵建立了故障诊断模型评价指标,探讨了能较好表征典型故障的故障指示特征。首先,对制冷系统及其典型渐变故障进行理论分析,初步了解征兆与故障(症状与原因)间的理论联系,结合ASHRAE的制冷机组故障模拟实验,探讨制冷系统故障指示特征智能提取方法,以期找到能较好表征故障的参数(集),减轻乃至消除特征间相关度,去除信息冗余,使故障更加清晰地呈现,有利于模型对故障的分离和识别,缩短诊断时间,提高诊断准确率。分别运用基于互信息(MI)的最大冗余最小相关过滤模型(MI-based mRMR filter)、基于遗传算法的封装模型(GA-LDA wrapper、GA-SVM wrapper)进行特征选择,运用主成分分析法(PCA)进行特征提取,得到不同的故障指示特征子集,并在后续章节的分析中逐步筛选出最佳子集。其次,针对制冷系统中典型的单发故障,运用故障指示特征智能提取与一种基于结构风险最小化的新型机器学习方法——支持向量机(SVM),相结合的顺序集成模型,进行故障检测与诊断,并基于混淆矩阵(Confusion matrix)建立以诊断正确率(CR)、命中率(HR)及虚警率(FAR)为核心的故障诊断模型评价体系,评价模型对于所有样本的总体性能及对包括正常及各类故障的诊断效果(分布性能),命中及虚警情况。结果表明,SVM的故障检测与诊断性能优于故障诊断决策树模型(C4.5),CR达99%以上。GA-SVM封装模型从64个原始特征中所选之8个故障指示特征较其他各种智能提取模型所选特征子集更为突出,在SVM模型及C4.5模型中均表现优良。对特征数的研究表明,不论原始特征抑或经PCA提取的综合特征,特征数越少,故障检测与诊断模型的训练及测试时间越短,但特征数与模型性能之间并非单调关系,特征太少可能造成信息缺失而降低诊断正确率,特征太多增加冗余信息而对故障诊断造成干扰,使模型鲁棒性下降。通常,所选特征数应至少等于包含正常运行及所有故障在内的独立类别数。基于PCA的特征提取只有当所选主元累计方差贡献率超过95%时,效果好于不进行任何智能特征提取的64个原始特征,但亦不及大多特征选择模型。冷凝器结垢、冷凝器水量不足、不凝性气体及蒸发器水量不足四种故障较易被检测与诊断,即使发生程度很轻微,而制冷剂泄漏及过量故障最难被命中,单纯SVM模型对该两类故障尤为难以识别,而GA-SVM模型则极大改善该性能。第三,就多故障并发时的检测与诊断,提出基于多标识(multi-label)数学解耦技术与SVM顺序集成的模型,并以制冷系统冷凝器水量减少20%、蒸发器水量减少20%的并发故障为例,研究模型性能。发现该模型仅用正常及两类单发故障数据而不用并发故障数据训练,即可对并发故障加以检测及诊断,效果良好,尤其在采用前文述及之8个最佳故障指示特征时。研究亦表明,尽管冷凝器侧水流量及蒸发器水环路阀位两个特征分别可独立表征并发故障中的两类单发故障(子故障),在并发故障的检测与诊断中却无能为力,必须借助其他参数的表征性能。另,提出采用一种较PCA有所改进的多变量统计分析法——指定元分析法(DCA),用于并发故障检测与诊断,但对故障投影方向(指定元)的定义极大地依赖于专业经验及知识。最后,以一台额定制冷量16.8kW的风冷热泵水机为受试对象,通过能量平衡并引入故障模拟管路及元件,建立了制冷系统故障诊断专用实验台,可以模拟包括制冷剂泄漏、充注过量、液体管路受阻、压缩机吸排气串通、蒸发器水量不足、冷凝器风量不足、冷凝器结垢、膨胀阀预紧力太大或太小等制冷系统典型故障,并进行了部分单发故障及双故障、三故障并发的实验模拟,分析故障发生时,制冷系统关键参数的变化,并探讨其可能的原因。对液体管路受阻、冷凝器结垢、蒸发器水量不足及其并发故障,运用前述智能集成模型,从44个原始特征中筛选出环境相对湿度或温度、冷凝器进出风温差、蒸发器进出水温差及供水温度四个特征,作为最佳故障指示特征,CR达99.58%。总之,本文所提之智能集成模型,以及讨论之故障指示特征,主要成果已在多个专业国际期刊上发表,在制冷系统故障智能检测与诊断中,具有一定应用价值与意义,值得进一步研究。更多还原

【Abstract】 Heating, ventilation, air-conditioning and refrigeration (HVAC&R) systems are becoming increasingly complex with various faults happening during operation. If not being fixed in time, system operation parameters will deviate even far from their original design value and accordingly, a series of negative effects will arise --- uncomfortable people indoor, low productivity efficiency, more system energy consumption, shorter equipment lifecycle and even worse atmospheric environment, etc. Fault detection and diagnosis (FDD) may help in timely finding and fixing faults, so as to improve system security, reliability and stability, prevent or avoid faults from happening and spreading. In order to improve accuracy & sensitivity and save computational time for the intelligent detection and diagnosis of typical individual faults and multi-simultaneous faults (MSF) for refrigeration systems, this study put forward a variety of sequentially integrated models, established FDD model evaluation guidelines based on confusion matrix, investigated all kind of possible fault indicative features.Firstly, refrigeration system and its typical soft faults were first theoretically analyzed with the relationship between symptom and faults, results and cause primarily understood. Intelligent methods for feature selection and extraction were studied for the purpose of finding better fault indicative features set by reducing or removing correlation between features, cutting redundancy, making the faults‘appear’clearly and easy to be identified, shortening FDD time span and improving FDD accuracy simultaneously. Filter models based on mutual information (MI) and minimum- redundancy-maximum-relevance (mRMR), wrappers based on genetic algorithm (GA), linear discriminant analysis (LDA) and support vector machine (SVM), feature extraction model based on principal component analysis (PCA) were widely investigated and carefully applied to the historical normal and faulty data for a 90 tons centrifugal chiller. Fault indicative features sets were obtained and further studied in the later chapters to single out the best one.Secondly, for the seven typical individual faults in refrigeration systems, intelligent feature selection and extraction methods were sequentially integrated with SVM, a newly concerned machine learning method based on minimizing structural risk, to perform detection and diagnosis. FDD model evaluation system with correct rate (CR), hit rate (HR) and false alarm rate (FAR) as its core was established based on confusion matrix commonly used in pattern recognition field. CR is for the evaluation of the model’s overall FDD performance for all samples; HR and FAR are guidelines for the evaluation of model’s individual performance for each class, normal or each fault. What about the ratio of samples that are hit or correctly reported and what about those falsely alarmed. The results showed that SVM model was better than the famous decision tree (C4.5) in the FDD of refrigeration system, with test CR over 99%. The eight-feature subset selected by the GA-SVM wrapper from the original 64 features behaved much better than other subsets selected by other schemes, even in the FDD by C4.5 model. Investigation on the number of features for fault indication demonstrated that that no matter the selection of the original features or the extraction of the comprehensive features by PCA, the fewer the features, the less training and testing time consumed by the FDD model, but the performance would not be that consistent with the increasing or decreasing of the feature numbers. Excessively fewer features might cause lack of information and undermine the performance accordingly, while too many features would add excessive redundancy, cause interference for FDD and harm model’s robustness. In fact, the features should better be at least equal to or more than the number of individual class including normal and all types of faults concerned. Only when the cumulative variance contribution rate was a little bit more than 95%, did the integrated model with feature extraction by PCA perform better than the SVM model without integration (64 features), but still, it could not surpass those models that integrated with feature selection schemes. Four faults such as condenser fouling, reduced condenser water flowrate, non-condensables and reduced evaporator water flowrate were easy to be detected, isolated and identified, even with the most slight level (level 1), whereas refrigerant leakage/undercharge or overcharge were the most difficult to be hit, especially when the SVM model without integration was employed, but GA-SVM model performed much better for these two faults.Thirdly, for the detection and diagnosis of MSF, put forward a sequentially integrated model that combined multi-label (ML) decoupling technique with SVM. Model performance was investigated for the MSF of the reduced condenser water and evaporator water flowrate simultaneously, both about 20% less than the rated. It was found that the integrated model had an excellent behavior in the FDD of MSF even when it was trained just by the normal and the individual faults instead of the MSF, especially when the eight fault indicative features previously stated were employed. Although the condenser water flowrate (FWC) and the valve position in evaporator water loop (VE) could independently indicate the individual faults (sub-faults) well, they were incapable of indicating MSF and must get assistance from other features to obtain a better performance. Moreover, designated component analysis (DCA), a multivariate statistical analysis method better than PCA in a sense, was adopted for the FDD of MSF in refrigeration systems. The method was effective as long as the prior knowledge and experience for the investigated systems were enough.At last, a dedicated fault detection and diagnosis test stand has been established by designing energy balance system and introducing fault simulation lines and components for and into an air-source heat pump of 16.8kW rated cooling capacity. Typical faults that could be simulated include refrigerant leakage/undercharge, overcharge, liquid line restriction, compressor valve leakage, reduced evaporator water flowrate, reduced condenser air flowrate, condenser fouling, thermal expansion valve over or less pre-tightened, etc. Experiments have been done for some types of the individual faults and two or three faults happening simultaneously. Variation of the critical parameters while faults happening was analyzed and possible cause or reasons were discussed. To the individual faults of liquid line restriction, condenser fouling, reduced evaporator flowrate and the combinations of two or three of them, the sequentially integrated models previously studied was applied, and four features selected from the original 44 features, including the environment relative humidity or temperature, the temperature difference between the inlet and outlet air of condenser, the temperature difference between the inlet and outlet water of evaporator, and the supply water temperature, were regarded as the best fault indicative features for the faults investigated and the test CR was as high as 99.58%.In general, the sequentially integrated models put forward in this study and the concept of fault indicative features for the intelligent FDD of refrigeration systems are effective, having a promising perspective and worthy of further investigation.(This research was supported by the Chinese National Natural Science Foundation under No. 50876059.)更多还原