基于物理引导神经网络的蒸发器制冷剂泄漏故障诊断模型
Fault diagnosis model of refrigerant leakage in evaporators based on physics-guided neural network
摘要：
制冷剂泄漏是制冷和空调系统中的常见问题,目前主流的基于数据驱动的诊断方法普遍存在缺乏物理信息的指导、对数据依赖性强、外推性能差等问题。本文提出了一种基于物理引导神经网络(PGNN)的蒸发器制冷剂泄漏故障诊断模型,根据换热过程能量守恒方程选择相关的特征,并在神经网络中加入物理损失函数,以更好地捕捉制冷剂泄漏引起的物理变化。实验结果表明,PGNN模型在训练集和测试集上的平均正确率分别达到99.73%和99.70%,最佳结果可达100.00%,相比之下,传统的多层感知机(MLP)模型的正确率仅为81.74%,且存在高达30.89%的误报率。将本文方法应用于轻度提升机(LightGBM)和支持向量机(SVM)模型,其正确率分别从98.31%和81.01%提升至99.58%和94.09%,表明本文方法具有良好的外推性能。
Abstract：
Refrigerant leakage is a common problem in refrigeration and air conditioning systems, and the current mainstream data-driven diagnostic methods generally have defects such as insufficient physics-informed guidance, high dependence on data, and poor extrapolation performance. In this paper, an evaporator refrigerant leakage fault diagnosis model based on physics-guided neural network (PGNN) is proposed, in which the relevant features are selected according to the energy conservation equation of the heat exchange process, and the physical loss function is incorporated into the neural network to better capture the physical changes caused by refrigerant leakage. The experimental results show that the PGNN model achieves average accuracy rates of 99.73% and 99.70% on the training and testing datasets, with the best results reaching up to 100.00%. In contrast, the traditional multi-layer perceptron (MLP) model only achieves an accuracy rate of 81.74% and has a high false alarm rate of 30.89%. When the proposed method is applied to the LightGBM and SVM models, the accuracy rates are increased from 98.31%and 81.01% to 99.58% and 94.09%, respectively, indicating that the proposed method has good extrapolation performance.
Keywords:physical loss function; feature selection; physics-guided neural network; refrigerant leakage; fault diagnosis; machine learning