Hydrogen leakage has become the biggest bottleneck restricting the development of hydrogen energy. Accurately locating the leakage is conducive to improving the hydrogen energy safety technology system. In this paper, a deep belief network pre-training and fine-tuning (DBN-PF) leakage source positioning coordinate method based on the fully confined space hydrogen Gaussian distribution model is proposed. Firstly, a fully confined space hydrogen Gaussian distribution model is established, and then a pre-training dataset is generated. The Gaussian-Gaussian Restricted Boltzmann Machine (GGRBM) in Deep Belief Network (DBN) is unsupervised pre-trained layer by layer, and the back propagation (BP) neural network in DBN is supervised pre-trained. Secondly, the hydrogen leakage experiment in the fully confined space was carried out, in which 8 hydrogen concentration sensors (HCSs) were placed on the top of the fully confined space, and the leakage experiment was carried out in sequence at 25 leakage locations. Finally, the HCS experimental data of one leakage location is extracted to fine-tune DBN. The experimental data at the other leakage locations were used to verify the positioning results. The results show that the positioning results average error of the proposed method is 20.62 mm. The average error accounts for 2.97% of the diagonal length of the closed model XY plane. Compared with BP neural network trained directly with the same fine-tuning dataset, the positioning error is reduced by at least 82.37%. Compared with the BP neural network with the same pre-training and fine tuning, the positioning error is reduced by at least 39.31%. The positioning technology developed in this study can achieve good hydrogen leak source location accuracy in a fully confined space only with the help of HCSs and a small amount of data.

中文翻译：

---

基于全受限空间高斯分布模型的深度置信网络氢气泄漏源定位方法

氢气泄漏已成为制约氢能发展的最大瓶颈。准确定位泄漏有利于完善氢能安全技术体系。本文提出一种基于全受限空间氢高斯分布模型的深度置信网络预训练与微调（DBN-PF）泄漏源定位坐标方法。首先建立全受限空间氢高斯分布模型，然后生成预训练数据集。深度置信网络（DBN）中的高斯-高斯受限玻尔兹曼机（GGRBM）是逐层无监督预训练的，DBN中的反向传播（BP）神经网络是有监督预训练的。其次，进行全密闭空间氢气泄漏实验，在全密闭空间顶部放置8个氢气浓度传感器（HCS），在25个泄漏位置依次进行泄漏实验。最后，提取某一泄漏位置的HCS实验数据来微调DBN。利用其他泄漏位置的实验数据来验证定位结果。结果表明，该方法的定位结果平均误差为20.62 mm。平均误差占闭合模型XY平面对角线长度的2.97%。与直接使用相同微调数据集训练的BP神经网络相比，定位误差至少降低82.37%。与相同预训练和微调的BP神经网络相比，定位误差至少降低39.31%。本研究开发的定位技术仅借助HCS和少量数据就可以在完全受限的空间内实现良好的氢气泄漏源定位精度。