在大型化工厂中，输送液体的管道的泄漏是一个重要的问题。管道的破损不仅会影响工厂的正常运行，同时也增加了维护成本。此外，还会使操作人员的生命安全受到威胁。因此，管道泄漏的检测与定位是维护和状态监测中的关键任务。近年来，大型工厂利用红外（IR）相机进行泄漏检测。红外相机可捕捉温度比周围环境温度高（或低）的液体泄漏。本文针对化工厂中的管道泄漏，提出了一种基于红外视频数据和机器视觉技术的检测与定位方法。由于所提出的方法是以视觉技术为基础，无需考虑泄漏液体的物理性质，因此其适用于任何类型的液体（水、油等）泄漏检测。在本方法中，首先对后续帧进行减影和分块处理，然后对每一分块进行主成分分析，提取特征；接着将分块内所有减影帧都转换为特征向量（作为块分类的依据），根据特征向量，采用k-最近邻算法将块分为正常（无泄漏）和异常（泄漏）两类；最后在各异常块上确定泄漏的位置。本文使用了两种不同格式的数据集（由红外相机拍摄的实验室工厂演示装置的视频图像组成）对上述方法进行评估。结果表明，本文提出的利用红外视频进行管道泄漏检测与定位的方法前景可观，具有较高的检测精度以及合理的检测时间。本文最后讨论了该方法在工厂进行实际推广的可能性及局限性。

Liquid leakage from pipelines is a critical issue in large-scale process plants. Damage in pipelines affects the normal operation of the plant and increases maintenance costs. Furthermore, it causes unsafe and hazardous situations for operators. Therefore, the detection and localization of leakages is a crucial task for maintenance and condition monitoring. Recently, the use of infrared (IR) cameras was found to be a promising approach for leakage detection in large-scale plants. IR cameras can capture leaking liquid if it has a higher (or lower) temperature than its surroundings. In this paper, a method based on IR video data and machine vision techniques is proposed to detect and localize liquid leakages in a chemical process plant. Since the proposed method is a vision-based method and does not consider the physical properties of the leaking liquid, it is applicable for any type of liquid leakage (i.e., water, oil, etc.). In this method, subsequent frames are subtracted and divided into blocks. Then, principle component analysis is performed in each block to extract features from the blocks. All subtracted frames within the blocks are individually transferred to feature vectors, which are used as a basis for classifying the blocks. The k-nearest neighbor algorithm is used to classify the blocks as normal (without leakage) or anomalous (with leakage). Finally, the positions of the leakages are determined in each anomalous block. In order to evaluate the approach, two datasets with two different formats, consisting of video footage of a laboratory demonstrator plant captured by an IR camera, are considered. The results show that the proposed method is a promising approach to detect and localize leakages from pipelines using IR videos. The proposed method has high accuracy and a reasonable detection time for leakage detection. The possibility of extending the proposed method to a real industrial plant and the limitations of this method are discussed at the end.