天然气生产过程工艺设备结构复杂、工艺过程参数及工况多样，难以进行作业过程的异常识别。针对此问题本文基于自适应高斯混合模型（Adaptive Gaussian mixture model，AGMM）和主成分分析（Principal component analysis，PCA）提出天然气脱水装置生产过程多工况时变数据异常识别方法（Anomaly identification for multi-case time-varying data, AIMT）。AIMT采用滑动窗PCA更新监测数据的主元模型，基于AGMM进行工艺过程的挖掘，通过大量历史监测数据进行AGMM中各高斯模型的参数自适应优化以实现对复杂多工况场景的有效表征，最终通过平方预测误差(Squared prediction error，SPE)和霍特林统计量（Hotelling’s T2，T2）进行工艺过程异常的识别认定。方法的有效性通过重庆气矿某扩建100万三甘醇（Triethylene Glycol，TEG）脱水系统单工况、变工况、多工况作业过程及真实故障事件的数据进行了验证。

The complexity of natural gas production processes and the variety of parameters and working conditions makes it difficult to identify abnormalities of operation processes. To address this problem, we proposed an anomaly identification method for natural gas dewatering process based on adaptive Gaussian mixture model (AGMM) and principal component analysis (PCA), namely anomaly identification for multi-case time-varying data (AIMT). AIMT uses sliding-window PCA to update the master metamodel of the monitoring data for the description of the working process based on AGMM. Adaptive optimization of the parameters of each Gaussian model in AGMM is carried out through a large amount of historical monitoring data in order to achieve effective characterization of complex multiple operating conditions. The identification of process anomalies is realized by squared prediction error (SPE) and Hotelling's T2 (T2). The effectiveness of the method proposed in this paper has been verified by data from single condition, variable condition, and multiple condition operations and real failure events in an expanded 1 million triethylene glycol (TEG) dewatering equipment at Chongqing Gas Field, PetroChina Southwest Oil and Gas Field Company.

国家自然科学基金(52275518)