【Applied Energy最新原创论文】面向大型复杂空调系统高维传感器故障检测的热力学定律嵌入式深度学习方法

（1）本文提出了一种全新可靠的高维传感器故障检测方法。

（2）该方法利用损失函数在深度学习中嵌入热力学定律。

（3）嵌入式热力学定律能有效消除故障数据在神经网络中的负面传播影响。

（4）该方法可有效提升复杂多故障情况下的检测准确性与可靠性。

大型复杂中央空调系统中所配备的传感器日益增多，其健康运行是系统能效及室内热舒适的基本保障。传统故障检测方法难以在大量传感器构成的高维数据中高效、精确识别故障；同时新兴深度学习方法虽善于处理高维数据，但因其纯数据驱动本质，可靠性难以保障。因此，本研究提出了一种热力学定律嵌入式深度学习方法，有效提升检测可靠性，解决复杂空调系统中高维传感器故障检测难题。

In building Heating, Ventilation and Air Conditioning (HVAC) systems, sensor healthy operation is the foundation of the adopted control strategies to improve building energy efficiency and indoor thermal comfort. For large and complex HVAC systems where a large number of sensors are often installed, associated sensor fault detection is highly challenging due to the high dimensionality of the sensor data and complex multiple-fault scenarios. To address this challenging issue, this study proposes a novel method in which the thermodynamic laws (i.e., mass balance and energy conservation) are integrated with deep learning. By making use of the intelligence, flexibility, and efficiency of deep learning, the proposed method can easily handle high-dimensional sensor measurements. More importantly, the integration enables the thermodynamic laws (which govern the mass and heat transfer processes in HVAC systems) to be explicitly learned and thus can effectively reduce/eliminate unreasonable results (e.g., violations of mass balance or energy conservation) frequently observed from sole deep learning methods due to their pure data-driven nature. Reduction/elimination of such unreasonable results can improve associated high-dimensional sensor fault detection performance in terms of accuracy and reliability. In the case study, compared with a conventional sole deep learning method, the proposed method increased the fault detection rate by 27.2%, and significantly reduced the false alarm rate by 77.4% in the complex multi-fault scenarios. Associated analysis demonstrated that the integration of thermodynamic laws can substantially alleviate the adverse intercorrelation impacts induced by faulty measurements inside the deep neural network when multiple sensor faults occurred. The proposed method provides an effective and reliable means to ensure the sensor healthy operation in large and complex HVAC systems in particular as increasingly more sensors are installed nowadays.

Keywords:

HVAC   暖通空调

Fault detection   故障检测

Sensor fault   传感器故障

Physics-informed deep learning   基于物理信息的深度学习

Data-driven   数据驱动

图1 . 本文所提出热力学定律嵌入式深度学习方法总览

图2. 自编码器（Autoendocer）深度神经网络用于传感器故障检测

(a)

(b)

图3. 复杂多故障情况下，热力学定律嵌入式深度学习方法相较于单一深度学习方法实现了更高的故障检测率与更低的误报率

(a)

(b)

图4. 当多个传感器发生偏差故障时，使用单一深度学习方法生成的传感器#6的残差低于故障检测阈值，这是由传感器#22的故障数据及其在神经网络中的负面传播造成的；本文提出的热力学定律嵌入式深度学习方法克服了相关负面影响，成功检测到传感器#6与#22的故障

(a)

(b)

图5. 热力学定律嵌入式深度学习方法（b）显著抑制了多故障工况下、常见于单一深度学习方法（a）中由故障数据造成的传播性负面影响及其引起的残差增长，显著降低了误报率

(a)

(b)

(c)

图6. 进一步分析表明，在深度学习神经网络训练与测试中，随着热力学定律在损失函数中权重增加，减小热力学定律违反程度（图a）与减小自编码器的误差（图b）是存在取舍的；但更严格符合热力学定律的神经网络（既较小的违反程度）并不一定能取得更好的故障检测效果，需慎重考虑热力学定律的嵌入方式

本研究由香港城市大学、武汉科技大学、西华大学、与澳大利亚伍仑贡大学的研究人员共同完成。

通信作者简介：

孙勇军，博士，香港城市大学副教授，现从事高密度城市可再生能源智能应用、建筑节能、机器学习、数据中心浸没式冷却、零碳建筑等领域的研究。在Applied Energy、Energy Conversion and Management、Energy和Renewable & Sustainable Energy Review等期刊上发表论文90余篇，并多次受邀报告。

共同第一作者简介：

任浩山，博士，香港城市大学博士后，专注于建筑碳中和技术与深度学习交叉领域的研究，研究内容包括高密度城市可再生能源利用与碳中和规划、面向智能电网的建筑柔性用能、大型空调系统故障检测等。在Applied Energy、Sustainable Cities and Society、Renewable & Sustainable Energy Review等期刊发表论文30余篇，Scopus h-index 15。

徐成良，博士，武汉科技大学讲师，硕士生导师，从事空调系统故障诊断及区域建筑群能耗模拟领域研究，研究兴趣包括建筑空调领域中深度学习的可解释性研究以及图神经网络技术应用。近五年以第一作者或通讯作者在Applied Energy、Energy and Buildings、Applied Thermal Engineering等期刊上发表论文15篇。

关于Applied Energy

点击“阅读原文”

喜欢我们的内容？

点个“赞”或者“再看”支持下吧！