Bringing Smarts to Methane Emissions Detection An Update on the DOE Smart Methane Emissions Project Maria Araujo Manager R&D Southwest Research Institute
Background Fugitive emissions from compressors is one of the largest sources of methane emissions in the midstream sector 50% of fugitive emissions are from major compressor equipment A small number of sites have a disproportionate amount of emissions, particularly within the midstream segment 10% of emitting sites contributed to 50% of the overall methane emissions
Limitations on Existing Methane Detection Technologies Many existing methane detection technologies have several advantages: Ability to detect very small leaks (high sensitivity) Ability to locate leaks Low costs But there are also many gaps: High False Alarm Rates Limited Autonomous, Real-Time Capabilities
Smart Methane LEak Detection System (SLED/M) Funded by the U.S. Department of Energy, National Energy Technology Laboratory (NETL) Early detection of fugitive methane emissions Designed for reliable fenceline monitoring: Remote sensing Autonomous Onboard classification and reporting Deployable Light enough for mobile and fixed-wing applications
Project Objectives Develop an autonomous, reliable, real-time methane leak detection technology (SLED/M) SLED/M applies machine learning techniques to passive optical sensing modalities to mitigate emissions through early detection Phase I focused on the development of the prototype methane detection system with integrated optical sensors and the embedded processing unit. Phase II will focus on integration and field-testing of the prototype system along with documentation and demonstration within a representative controlled environment SLED/M is a passive technology that can monitor various regions of a pipeline facility and easily integrate with existing and new suites of natural gas mitigation technologies
SLED/M Details The Smart Methane Leak Detection (SLED/M) system provides methane emissions mitigation by providing autonomous early detection of sources of methane emissions in failing midstream infrastructure SLED/M consists of a mid-wave infrared (MWIR) camera and an embedded computer mounted inside of an enclosure SLED/M use case: installed at a compressor station and other similar-type facilities for monitoring midstream equipment for leak events and subsequently reporting these events SLED/M builds upon a similar system developed by Southwest Research Institute (SwRI ) to detect hazardous liquids using optical sensors
How is SLED/M Different? Remote Sensing + Artificial Intelligence State of the art medium wave infrared imager (COTS) Passive, intrinsically safe Expands on the limitations of existing methane sensors by using machine learning-based detection algorithms to autonomously and reliably (low false alarm rates) detect methane Powerful deep learning algorithms (AI) for reliable plume identification No human in the loop Reliable false positive rejection 7
SLED/M Key Features Feature Low False Alarm Rates Autonomous Detection Near Real-Time Detection Non-Intrusive, Passive Technology Details Less than 0.5% (number of events incorrectly classified as leaks). No need for a human to be in the loop the system acquires, process and makes autonomous decisions on whether or not a hazardous substance was observed, using machine learning algorithms. The time between acquiring data and obtaining an output from the system is only a few minutes. No need to retrofit existing equipment and facilities. The proposed technology is passive in nature, thus eliminating safety and operational restrictions.
Where Did SLED/M Come From? Smart Leak Detection (SLED) System A combination of optical sensor modalities to find liquid hydrocarbon leaks COTS components Visible (0.4μm 0.9μm) Long-wave Infrared (7.5μm 13μm) Developed under SwRI internal research Machine Learning + Optical Sensing
SLED: Are There Hazardous Liquids In These Images? 1 10
SLED: Hazardous Liquids Detected and Identified! Mineral Oil Diesel Gasoline Crude
Leak Detection in Real Time! Mineral Oil Diesel Gasoline Crude
SLED/M Project Team Funded by the U.S. Department of Energy National Energy Technology Laboratory (NETL) Southwest Research Institute Principal Investigators Cost Share Partners Falcon Inspection Contributing the lease of one of the MWIR cameras (FLIR A6604) IRCameras Contributing the lease of one of the MWIR cameras (Niatros Optical Gas Imaging Camera) 13
SLED/M Development 14
Sensor Setup and Testing SwRI performed a variety methane release tests under realistic conditions to establish a database containing methane leaks of various concentrations, distances, and scenarios
Data Collection Broad data collection in order to diversify the training sets for both methane imagery and false positives Total 1.4T of Raw Data 50mm FLIR Data - 9 days of data collection - 1,308,484 image files - 91 sets of methane and corresponding backgrounds 25mm FLIR Data - 83,958 image files - 23 sets
Algorithm Results Overview 17
Algorithm Overview Video allows for the use of temporal information Algorithm loop: Detect change in scene over time Classify change as methane / no methane Track classifications over time in order to increase confidence Pre- Processing Scene Model model update Difference Extraction Classification Interpreter 18
Detection: Uniform Background Raw Camera Video SLED/M Processed Video Methane Released Here 19
Detection: Dark Background Raw Camera Video SLED/M Processed Video 20
Detection: Noisy Background Raw Camera Video SLED/M Processed Video 21
Detection & Network Input SLED/M Processed Video Pre-processed Camera Video 22
Performance Metrics Overall pixel accuracy: 0.872 Overall frame accuracy: 0.970 Calculated for 4,722 frames (512 x 640 pixels each) across 4 different validation videos of varying background scenery Real-Time Performance The effective rate of the software is currently one classification (possible detection) per ~1.4 seconds Running on an embedded NVIDIA Tegra TX1 board 23
Embedded System Overview 24
Embedded System Overview This image shows the embedded system block diagram, with pictures of the actual components being used for ongoing testing. 25
SLED/M User Interface The SLED-M user interface, available via a web browser, displays of the raw camera feed and the result of the methane detection algorithm (methane in red overlay). The effective rate of the software is currently one classification (possible detection) per ~1.4 seconds. This video shows data being displayed through the embedded software via HTTP Right: Raw Camera Data Left: SLED/M Processed Data 26
SLED/M Mechanical Overview This animation shows a 3D orbit view of the assembly. The enclosure is NEMA certified, designed for outdoor use, and measures 26.65 x 18.77 inches. 27
Plans for Phase 2 28
Plans for Phase 2 Phase II Focus: Field experimentation and demonstration Field experiments will be performed within representative midstream pipeline infrastructure environments. It is anticipated that much of this testing will occur at SwRI s Metering Research Facility (MRF). Algorithm will be further tested and refined SLED/M will be integrated with a user-monitoring interface to quantify the efficacy of the system under a variety of conditions Functionality of the prototype system will be demonstrated 29
Questions? Maria Araujo Manager R&D Southwest Research Institute +1-210-522-3730 Maria.Araujo@swri.org 30