Intelligent operating monitoring of pipelines helps to detect anomalies in time to ensure pipeline safe, reducing potential risk. However, the operating conditions of the multi-product pipeline change frequently, and the recognition and monitoring by on-site personnel are easy to cause misjudgment, so the operating conditions of the pipeline cannot be accurately recognized. Noticeably, operating condition recognition is an important part of pipeline safety and risk management. Although ample operating data are stored in SCADA system, these data are lack of corresponding condition labels, making it hard to be mined. In this work, a semi-supervised learning for operating condition recognition is proposed to overcome aforementioned issues. Firstly, the operating parameters of each station are preprocessed and collected to construct into data matrices to overcome transient disturbance considering the pipeline space characteristics and time series of the operating data. Then stacked autoencoder (SAE) is used to pre-train the network parameters of multi-layer neural network (MLNN) based on a large amount of unlabeled operating data. After that, MLNN is fine-tuned based on a small amount of labeled data annotated by referring to the operation log. To verify the effectiveness of the semi-supervised learning, a real multi-product pipeline is taken as an example for operating condition recognition. The accuracy, precision, recall and F1 score is 95 %, 95 %, 80 % and 80 %, respectively. Results show that the condition recognition accuracy of the proposed model is better than other machine learning models. Finally, the sensitivity analysis is conducted to illustrate the importance of SAE in this classification model.

管道的智能运行监控有助于及时发现异常，以确保管道安全，降低潜在风险。然而，多产品管道的运行条件变化频繁，现场人员的识别和监控容易造成误判，因此不能准确地识别管道的运行条件。值得注意的是，运行状态识别是管道安全和风险管理的重要组成部分。尽管SCADA系统中存储了足够的运行数据，但是这些数据缺少相应的条件标签，因此很难进行挖掘。在这项工作中，提出了一种用于操作状态识别的半监督学习方法，以克服上述问题。首先，考虑到管道空间特征和运行数据的时间序列，对每个站的运行参数进行预处理和收集，以构建成数据矩阵，以克服瞬态干扰。然后，基于大量未标记的操作数据，使用堆叠式自动编码器（SAE）对多层神经网络（MLNN）的网络参数进行预训练。之后，基于少量的标记数据（通过参考操作日志进行注释）对MLNN进行微调。为了验证半监督学习的有效性，以实际的多产品流水线为例来进行操作条件识别。准确性，准确性，召回率和F1分数分别为95％，95％，80％和80％。结果表明，该模型的条件识别精度优于其他机器学习模型。最后，进行敏感性分析以说明SAE在此分类模型中的重要性。