基于敏感特征的制冷系统故障扩散机理研究及早期预测

Faults of refrigeration systems may cause over 30% of extra energy consumption, the degradation of equipment reliability, safety and operating cost-effectiveness, the reduction of service life, the deterioration of indoor and outdoor environment, and the dwindling of productivity, etc. Even serious problems could be caused for a special system like cold storage where crucial substances such as medicine, gene, cell or marrow, etc. were stored for significant use. Early prediction of possible faults is important and indispensible for the prevention of abnormal operation and corresponding negative effect. Although fault prognostics is now under research home and abroad in many fields like national defense, military affairs, nuclear power, mechanical engineering, etc., it could hardly be found in the fields of heating, ventilation, air-conditioning and refrigeration...The project applicant, therefore, would like to make contribution based on her previous research on the fault detection and diagnosis (FDD) of refrigeration systems (one of the research papers has been listed among the ESI highly cited paper in recent ten years). FDD is the process of detecting and identifying a failure mode within a system or sub-system. Prognostic, in its simplest form, is to monitor and detect the initial indications of degradation in a component or system, and be able to consistently make accurate prediction. Despite the similarities, fault prognostics is quite more challenging! Based on the symptom detection and the concept of important sensors of refrigeration system faults, comprehensively using fault prognostic algorithms like support vector machine (SVM, a newly data mining methodology that shows better adaptability, classification capability and computation efficiency than the traditional artificial neural network (ANN)), genetic algorithm, principle component analysis, etc. combined with the technology of information fusion, this project would carry out study on the proliferation mechanism of the faults of refrigeration systems both in time and space, investigate the coupling of early faults and operation parameters, system performance and energy flow, probe into the non-linear relationships between early symptom and cause, and complete the prediction or forecasting of system faults (including the prediction of remaining service life), so as to provide theoretical guidance and practical basis for the intelligent maintenance and health management of refrigeration systems.

制冷系统带障运行导致能耗增加30%以上，设备可靠性、安全性及运行经济性下降，寿命缩短，室内外环境恶化，生产力降低等诸多问题。通过对故障扩散机理的深入研究，可实现故障的早期预测及控制，保证系统运行于良好状况，对其高效运行及节能具有促进作用。国外已在多个领域开展设备失效机制及故障预测研究，国内刚刚起步。本课题基于申请人对制冷系统故障检测与诊断多年的研究及其高被引论文（ESI）的热点指示，通过特殊设计的故障加速实验，结合系统的物理模型及信息融合、数据挖掘等分析方法，提取制冷系统早期微弱故障时仅对故障敏感、对工况等外在条件不敏感的特征/特征集，研究故障及其敏感特征（或信息融合后的综合特征）在时间和空间上的多级演化机理：单发故障阶段及并发故障阶段（一种故障发展诱发其他故障），对后者亦研究多故障交互作用下的故障演化机理，均建立量化模型，并基于此，实现故障短期与长期预测，评估系统可靠服务状况及剩余寿命。

本项目针对制冷系统软故障诊断及预测中的一些关键技术问题展开研究，涉及故障敏感特征、故障机理及诊断与预测性能等。基于对故障机理的研究，将制冷系统软故障（渐变故障）分为局部故障与系统故障（或称设备级故障与系统级故障），前者的诱因及症状主要集中于或体现于系统的某个局部，对全局的影响相对较弱，找到对应的敏感特征相对交易，而后者的诱因多涉及制冷剂或润滑油，随着系统的运行，症状遍及整个系统，产生的影响广泛而深入，敏感特征往往复杂而不单一，是故障诊断及预测的难点。通过不同方法、不同角度建立的制冷系统故障检测、诊断及预测模型，对上述问题进行深入研究，模型包括：基于模拟退火及深度学习（SA-DNN）的模型，基于概率神经网络（PNN）及其优化的模型，基于最小二乘支持向量机（LSSVM）的模型，基于粒子群算法（PSO）优化LSSVM的模型，基于概念漂移检测（CDD）的制冷系统故障诊断自适应模型，基于PSO优化BP的模型和基于不平衡数据处理的模型等。研究发现：通过故障敏感特征进行制冷系统的故障检测、诊断及预测是可行而有效的；与局部故障相比，系统故障（如制冷剂泄露/不足、制冷剂过量、润滑油过量等），特别是轻微的系统故障更加容易与不同工况运行的正常状态混淆，而难以检测、诊断或预测，通过选用一些相对先进的模型（如深度学习）并进行一定优化（如模拟退火优化）可以极大地改善该性能；通过运用概念漂移检测理念，可以贯通制冷系统故障检测、诊断及预测，进一步结合增量学习算法，可实现模型面对新情况、新问题、新故障的自调整、自适应学习，提升模型的推广泛化能力；利用不平衡数据技术（imbalanced data），结合一定故障诊断及预测模型可以实现故障先验知识的有效保留与推广，使模型具有在同系列、不同容量机组、不同系列机组甚或不同类型机组之间推广应用的潜力，有望突破数据驱动的智能模型的对象针对性太强以致限制其推广应用的瓶颈。