基于高斯过程混合模型的瓦斯安全状态分类研究

针对目前瓦斯浓度预测与瓦斯安全状态分类方法中主观性较强、超参数难以选取、解释性差、无法有效地利用样本之间时序信息等问题，本文提出了基于高斯过程混合模型的瓦斯浓度预测与安全状态分类方法。高斯过程是机器学习领域中解决非线性回归问题的典型方法，能够有效地利用数据之间的相关性，常用于时间序列的建模与预测。然而，单个高斯过程存在着一定的局限性，难于对非平稳、多模态的数据进行有效地建模和回归分析。在高斯过程的基础上引入其混合模型，则可增强模型的表达能力，能够对有复杂结构的数据进行建模。我们将瓦斯安全状态根据风险由高至低分成红橙黄蓝四个等级，在每个风险等级上瓦斯浓度数据采用单个高斯过程进行建模。由于一般瓦斯浓度数据包含着各个风险等级的数据，高斯过程混合模型则可用于对整体数据进行建模和回归分析。根据对数据的参数学习结果，高斯过程混合模型便可自适应地得到每个时刻对应的风险等级，并在预测瓦斯浓度时对各个高斯过程分量的预测进行加权，得到更为鲁棒的预测结果。实验结果表明，基于高斯过程混合模型的方法可有效地预测瓦斯浓度、评估安全状态。 

Gas Safety State Classification Based on the Mixture of Gaussian Processes

In order to solve the problems of subjectivity, difficulty in selecting super parameters, poor interpretability and ineffective use of temporal information among samples on the existing methods of gas concentration prediction and gas safety state classification, this paper proposes a novel gas concentration prediction and safety state classification method based on the mixture of Gaussian processes. In fact, Gaussian process model is a classic method to solve nonlinear regression problems in machine learning. It can effectively get the correlations between temporal data and thus is often used in time series modeling and prediction. However, a single Gaussian process has certain limitations, and cannot model the data generated from a non-stationary source. The Mixture of Gaussian processes (MGP) can enhance the model capacity, and fit the data with complex structure. We try to divide the gas safety status into four levels according to the risk from high to low, namely, red, orange, yellow and blue. Since the gas concentration data in each risk level are generated by their specific time sequence characteristics, they can be modeled by a single Gaussian process. Because the general gas concentration data come from four risk levels, the MGP can be used to model the whole data. According to the MGP with the learned parameters, the risk level in each time can be adaptively computed. As for the gas concentration prediction, we can weight the prediction results of four Gaussian processes together to get a more robust prediction. The experimental results demonstrate that the MGP based method can effectively predict the gas concentration and evaluate the safety state.