大数据构架下模式辨识和集成学习的制冷空调故障诊断方法研究

The data-driven method has shown increasingly obvious advantages on online modeling analysis and fault diagnosis applications for refrigeration and air conditioning systems. Due to the big data characteristic of the actual measured data and its low quality as well as the continuous changes of complicated operational patterns in actual systems, the existing data-driven fault diagnosis models have some shortcomings such as low precision, weak sensitivity and poor adaptability..According to the theories and methods on research domains like thermodynamics, mathematical statistics and data mining, this project attemps to develop a statistical-thermodynamic coupled data quality evaluation and optimization method, build an unsupervised cluster-based system operational patterns patitioning and recognition model, establish the ensemble learning-based fault diagnosis strategy and validate it under the framework of big data. The research adopts an overall data-based analysis framework which depends on field opeartioanl data principally and experimental and simulation data accessorily. Firstly, the overall data quality is improved by eliminating the abnormal data. The abornormal data are determined using a combined criterion taking both the data statistical distribution similarity and thermodynamic energy balance into account. Secondly, the system operational patterns are adaptively partitioned and recognized through the clustering model which considers bothe the relativity of the probability density function and the accuracy of the distance threshold. The critical fault indicators are then extracted using the minimum redundancy maximum relevance feature selection approach. Finally, the extracted indicators are inputted into the ensemble learning-based classification model in each identified system operation pattern to eliminate the missed and delayed diagnosis problems, which as a result providing an accurate, sensitive and adaptive online fault diagnosis method for refrigeration and air conditioning systems.

数据驱动方法在制冷空调在线建模、故障诊断应用中优势日益明显。由于实际测量数据的大数据特性且质量偏低，系统运行工况模式复杂，数据驱动故障诊断方法存在模型精度低、灵敏度弱、模式适应性差等不足。.本项目结合热力学、统计、数据挖掘等领域的理论和方法，利用制冷空调运行数据，重点研究大数据构架下统计-热力学耦合机制的数据质量评价与优化、无监督聚类的运行模式适应性划分与辨识以及集成学习分类故障诊断策略的构建与验证等内容。计划以工程实测数据为主、实验仿真数据相结合，基于统计分布相似性和热力学能量平衡耦合判据剔除异常数据以提高数据质量，利用聚类算法结合概率密度相对性和距离阈值准确性实现运行模式的适应性划分，通过最大相关最小冗余提取关键故障特征，在各模式下建立集成分类的故障诊断模型，分析数据质量、运行模式差异性的定量影响规律，消除误诊、漏诊问题，为制冷空调系统提供精准、灵敏、适应性强的在线故障诊断模型。

针对制冷空调系统故障诊断研究领域存在的质量提升不足，复合故障特征分析及集成故障诊断研究等重点问题，本研究提出基于运行模式识别和集成学习分类的应用框架，以多联机、冷水机组为主的制冷空调系统作为研究对象，进行了充注量故障、脏污故障、阀类故障及故障耦合的多故障试验和特征灵敏度分析、不平衡数据质量增强研究，及基于深度学习和算法集成的多故障诊断和故障诊断迁移学习等系列研究，为制冷空调系统的健康、高效、节能运行提供保证。.多故障诊断诊断任务，关键问题为对制冷空调系统的不同故障特征进行分析和识别，从而实现分类型、故障水平的故障检测和诊断。研究以基于特征分析、特征选择和样本生成的数据质量优化方法为框架，利用多联式空调系统制冷剂充注量故障和脏污故障试验数据，研究故障对参数的影响趋势，提高多故障耦合情况下参数灵敏度的分析水平。依据基于生成对抗学习和过采样的数据强化，针对多联机系统、冷水机组系统的充注量故障、脏污故障等多种典型制冷空调系统，利用带有置信度评价的数据生成和诊断策略，建立故障诊断模型，提出一种新的数据增强和系统故障的故障检测和诊断方法。.研究以深度学习方法和多种算法集成为主要思想，对多故障类型的耦合故障情况利用卷积神经网络、长短期记忆网络、深度网络等方法进行深度特征提取,并利用注意力机制、联合分布适应算法和多种算法集成的原理，对故障诊断模型的知识迁移算法进行探索分析，研究故障模型的泛化机制，建立关键故障诊断知识的迁移策略。