| 摘要: |
| 制冷系统实际运行中，故障诊断模型可能出现诊断性能波动或下降等情况，需具备再学习能力以适应现场数据。本文设计了一种基于正确率阈值的概念漂移检测机制及支持向量机增量学习的故障诊断自适应模型，并将其应用于制冷剂过量故障的再学习。该算法通过两次优化选择、过滤数据信息，保留原有诊断知识，仅学习未知样本信息，可极大地节约模型学习时间，快速适应新环境。结果表明，新的故障种类出现时，诊断模型检测到概念漂移，进而通过增量学习进行自我更新，实现对新故障的学习与诊断。1 400个过量故障样本中诊断模型只需要学习600个，且保证最终模型对后续数据流具有较佳诊断性能，正确率高达99%。在现场制冷系统故障诊断应用中，诊断模型的再学习和自适应体现出良好的应用前景。 |
| 关键词: 故障诊断 概念漂移 支持向量机增量学习 制冷系统 自适应学习 |
| DOI： |
| 投稿时间：2018-06-08 修订日期：2018-08-02
|
| 基金项目:国家自然科学基金（51506125）资助项目。 |
|
| Adaptive Fault Diagnosis Model of Refrigeration System based on Concept Drift Detection |
| Wu Hao,Han Hua,Cui Xiaoyu,Fan Yuqiang,Xu Ling |
| (School of Energy and Power Engineering, University of Shanghai for Science and Technology) |
| Abstract: |
| A fault diagnosis model may be degraded or may fluctuate in an on-site refrigeration system. Thus, the model needs to adaptively learn the on-site data. To learn the online data stream of a refrigeration system during fault diagnosis, an adaptive diagnosis model is developed using the accurate threshold-based concept drift detection mechanism and an incremental support vector machine algorithm, which are applied to re-learn the refrigerant overcharge failure. Using two optimization processes to select and filter the data information, the algorithm retains the original diagnostic knowledge and only learns the unknown sample data information, which can greatly save training time and quickly adapt to a new environment. Simulation experiments are performed for online learning of the diagnosis of the refrigerant overcharge failure. The results show that when a new type of fault occurs, the diagnostic model detects the concept drift and then updates itself through incremental learning to learn and diagnose new faults. Three concept drifts are detected, and the diagnosis model only needs to update three times to realize learning of the refrigerant overcharge failure, that is, the diagnostic model only learns 600 of the 1400 refrigerant overcharge failure samples and ensures that the ultimate diagnosis model achieves better diagnostic performance for subsequent data streams with the correct rate reaching over 99%. In terms of on-site refrigeration system fault diagnosis application, the re-learning and self-adaptation of the diagnostic model shows good application potential. |
| Key words: fault diagnosis concept drift incremental Support Vector Machine refrigeration system adaptive learning |
