欢迎访问过程工程学报, 今天是

过程工程学报 ›› 2024, Vol. 24 ›› Issue (2): 162-171.DOI: 10.12034/j.issn.1009-606X.223174

• 研究论文 • 上一篇    下一篇

基于多块和自注意TCN结合的冷水机组故障诊断

孙雨， 丁强*， 夏宇栋， 李聪   

  1. 杭州电子科技大学能量利用系统及自动化研究所，浙江 杭州 310018
  • 收稿日期:2023-06-21 修回日期:2023-08-18 出版日期:2024-02-28 发布日期:2024-02-29
  • 通讯作者: 丁强 dinggiang@hdu.edu.cn
  • 基金资助:
    国家重点研发计划“面向跨境互联的多能互补新型能源系统关键技术研究”

Chiller fault diagnosis based on combination of multiblock and self-attention TCN

Yu SUN  Qiang DING  Yudong XIA  Cong LI   

  1. Insitute of Energy Utilization and Automation, Hangzhou Dianzi University, Hangzhou, Zhejiang 310018, China
  • Received:2023-06-21 Revised:2023-08-18 Online:2024-02-28 Published:2024-02-29

RichHTML

PDF

摘要： 针对冷水机组故障运行数据特征参数耦合且时序特征难以提取的问题，提出一种基于多块和自注意时间卷积网络(Multiblock Self-attention Temporal Convolutional Networks, MB-SATCN)的故障诊断模型。该模型根据冷水机组传感器与系统结构的物理关系将变量划分为多个子块，在子块中使用TCN挖掘冷水机组运行数据的特征信息，并通过自注意力层加强关键特征对故障诊断结果的影响权重，然后将各子块模型输出的局部特征利用自注意力机制加权融合以构建全局特征，并使用softmax函数进行分类。与对照方法相比，MB-SATCN方法在冷水机组常见故障诊断方面表现更优，对系统故障识别能力更强，平均故障诊断精确率和平均召回率均在97%以上。

关键词: 冷水机组, 故障诊断, 时间卷积网络, 自注意力机制, 算法, 神经网络, 模型

Abstract: The energy consumed by HVAC systems accounts for 50%~60% of total building energy consumption worldwide, and various failures of chillers reduce the efficiency of HVAC systems by 15%~30%, resulting in a considerable amount of energy waste. Therefore, accurate detection of faults in chiller systems can effectively mitigate energy waste and extend the life cycle of the equipment. A fault diagnosis model based on multiblock and self-attention mechanism time convolution network (Multiblock Self-attention Temporal Convolutional Networks, MB-SATCN) architecture is proposed for the problem of difficult extraction of fault sample data feature information with high coupling and time correlation in chiller unit fault diagnosis. The model divides the overall variables into multiple sub-blocks based on the physical relationship between chiller sensors and system structure, and uses the time-convolutional network architecture to mine the feature information of chiller operation data in the sub-blocks. And by introducing the self-attention mechanism to enhance the weight of key features on the fault diagnosis results, the local features output from each sub-block are again weighted and fused using the self-attention mechanism to construct a global feature representation, and the final input global features into the fully connected layer for classification using the softmax function. The simulation results show that the introduction of MB method and SA mechanism effectively improves the feature extraction ability of highly coupled chiller unit fault samples and moreover improves the fault diagnosis performance of the model. Compared with the fault diagnosis performance of three deep learning methods dealing with time series, MB-SACNN, LSTM, and GRU, the MB-SATCN method performs the average accuracy of fault diagnosis under SL1 level of minor faults is up to 98.00%, the average recall rate is up to 97.90%, the average accuracy rate is up to 97.91%, and the F1-score is up to 98.00%, which verifies the sensitivity and stability of the method.

Key words: Chiller, Fault diagnosis, temporal convolutional network, self-attention mechanism, algorithm, neural network, model