Real-time leak detection using an infrared camera and Faster R-CNN technique

Highlights

• A novel approach for real-time leak detection method using Infrared camera.

• Transfer learning approach is employed considering the limited training dataset.

• Sensitivity analysis to extract features from pre-trained model.

• The Faster R-CNN model exhibits the better performance compared to SSD models.

Abstract

Real-time hydrocarbon leak detection is an essential part of process safety and loss prevention program. Optical gas imaging (OGI) is one of the attractive methods to monitor hydrocarbon leak in the processing system. The manual analysis of a video frame to detect a potential leak is cumbersome and error-prone. The purpose of this study is to develop the automated hydrocarbon leak detection using appropriate technology and numerical technique. This is achieved by integrating Faster Region-Convolutional Neural Network (Faster R-CNN) technique with the OGI technology. The application of the procedure is demonstrated using the videos of the hydrocarbon leaks from an Ethane cracker plant. The videos are used to train the Faster R-CNN and subsequently used for the testing. The performance of the proposed integrated approach is compared with the Single Shot MultiBox Detector (SSD) models. The results confirm the proposed optimal model is superior compared to the SSD models.

Introduction

Hydrocarbon fire and explosion accidents are outcomes of leak incidents. Detecting the hydrocarbon leaks at the early stage is critical to prevent the causation of fire and explosion accident. Leak detection generally has two tasks, identifying the leak (often termed as classification task) and localization of the leak (often called localization task). The classification task also checks the hydrocarbon leak type, while the localization task also establishes the hydrocarbon leak sources. Both tasks are critical for a quick accident prevention decision.

Numerous studies have been conducted to realize the leak detection tasks of varied objectives. Some of the previous works combined the traditional sensors with the mobile robots (Ishida et al., 2004; Russell, 2004a,b; Lochmatter et al., 2008; Ferri et al., 2009). However, such integrated approaches exhibit poor performance in the complicated terrain of the chemical plant (Cho et al., 2018). In addition, some works integrated the traditional sensors with varied analytical approaches(Nofsinger et al., 2004; Huseynov et al., 2009; Li et al., 2011; Neumann et al., 2013; Chraim et al., 2015). More recently, the high focus has been given to application of the deep learning approaches for the chemical process area (Lee et al., 2018; Wu and Zhang, 2018; Ning and You, 2019; Shin et al., 2019). Kim et al. (2019) integrated the Computational fluid dynamics (CFD)-based sensor data with the deep-learning-based analytical approach, which plays a competitive role in the chemical leak detection.

In addition to the above traditional sensor-based approaches, Optical gas imaging (OGI) becomes the promising alternative to visually detect hydrocarbon leaks of the chemical plant (Ravikumar et al., 2016, 2018; Golston et al., 2018). The OGI's advantage lies in that it can realize the real-time hydrocarbon leaks detection from several meters away without the need to shut down the chemical process. However, some drawbacks still exist (Ravikumar et al., 2016). For example, manually investigating the camera is much labor-intensity since the camera cannot provide the real-time feedback of the detection results without operators' judgment (Wang et al., 2019). As such, Wang et al. (2019) proposed the machine vision-based approach to detect the natural gas methane emission to reduce the labor cost as well as the operator-related uncertainty. This recently developed CNN-based approach only realizes the classification task of the detection for the specific equipment, however, could not visually localize the hydrocarbon leaks, which could not provide the quick guide to trace the leaking sources among several chemical equipment.

Recently, several advanced object detection approaches have been developed to handle both classification and localization tasks for real-time object detection. These modern approaches can be generally classified into two types namely one-stage object detection methods such as You Only Look Once (YOLO) (Redmon et al., 2016), Single Shot MultiBox Detector (SSD) (Liu et al., 2016), and two-stage object detection methods such as Faster R-CNN (Ren et al., 2015). Among the modern approaches, Faster R-CNN approach exhibits competitive performance. In addition, several types of research have been recently performed to investigate the Faster R-CNN approach's capacity for varied real-time object detection applications (Sun et al., 2018; Pham et al., 2018; Zhang et al., 2018; Chen et al., 2018a,b; Xu et al., 2018; Chang et al., 2018; Hong et al., 2019). Generally, the researches demonstrated the approach's accuracy is relatively larger for all sizes of objects. However, its detection speed is slower compared to the one-stage object detection methods. Furthermore, for the generic datasets, Huang et al. (2017a,b) concluded that the detection speed could be effectively increased without harming the detection accuracy by changing the CNN architectures of the Faster R-CNN approach. This motivates us to determine the trade-off Faster R-CNN model between the detection speed and accuracy for real-time automated hydrocarbon leaks detection by altering the CNN architectures.

This study aims to handle both the classification and localization tasks of the real-time hydrocarbon leaks detection with no need of the operator's intervention. The Faster R-CNN approach and the OGI are accordingly integrated. Additionally, a procedure is proposed to develop the optimal trade-off Faster R-CNN model between the detection accuracy and speed for the first purpose. The OGI camera takes videos of the hydrocarbon leaks from the breather values in an Ethane cracker plant. Pre-trained model approach as one of the commonly used transfer learning approaches is employed considering the limited number of videos for the Faster R-CNN model development. Sensitivity analysis of varied pre-trained models and CNN architectures (both are denoted as pre-trained model configurations in the context below) on the integrated approach's performance is conducted. Optimality frontier presenting the optimal model is constructed. Comparison among the trade-off model and SSD models is performed. This study can contribute to selecting an appropriate model when employing the Faster R-CNN approach for the real-time automated hydrocarbon leaks detection.