Machine learning makes for more intelligent inspections
November 29, 2017
Andrew Dykes discusses the future of autonomy and asset integrity with Brad Tomer of Avitas Systems
ROVs, UAVs and crawlers are already the backbone of the modern asset inspection fleet, but they are only as useful as their human operators. In many cases, data gathered digitally must still be processed and reported on manually, accounting for considerable time and resources even after the inspection itself is complete. However, one GE venture company intends to alter the industry’s expectations radically.
“We’re a company that is trying to digitise the inspection market from end to end,” Avitas Systems’ VP of operations Brad Tomer told InnovOil. “We do this on three levels: we’re automating the collection of data through robotics and permanent sensors; we fuse various data sources together and use data analytics together with a deep learning platform for automated defect recognition, and finally, we have automated reporting and workflow so maintenance can be initiated right away.”
The inspection market itself is worth some US$40 billion a year, Tomer said, of which oil and gas alone makes up about US$10 billion. Avitas Systems was conceived as a start-up within GE, with the aim of capitalising on that sizable opportunity. Using the “GE Store” concept – a corporate philosophy which encourages the transfer of technology and knowledge across the company’s many units – Avitas Systems combines some of the most cutting-edge technologies available from the company’s various units, including GE Global Research Center.
Bringing together data analytics, robotics and artificial intelligence expertise from aerospace, transportation, renewables and oil and gas, the goal is to fully automate the inspection process, with the world’s “first end-to-end solution.” Doing so, Tomer said, could reduce the costs of inspection by up to 25%.
An inspector crawls
The inspection equipment and routines used by Avitas Systems are reassuringly familiar. Manned inspections teams are dispatched to survey assets in much the same way, using the company’s technicians and agreed mission plans. The group works with providers of robotic technology to employ state-of-the-art underwater vehicles (remote and autonomous), as well as aerial drones (UAVs), and magnetic/tracked crawlers. Sensor packages include inspection staples like LiDAR, infrared/HD cameras and laser measurement devices, among others.
Tomer instead explained that the company’s defining capabilities were its automation and artificial intelligence systems: “The differentiation of our autonomy system is that it not only controls the flight and mission of the drone but it controls the sensors themselves. It’s a full autonomy package that will allow us to take repeatable measurements from the same distance, location and angle every time.”
This capability is what helps Avitas Systems gain an edge over traditional manual inspection routines. The accuracy of these machine-led measurements means the quality of data gathered is unsurpassed, and can be used to track the life of an asset in more detail than ever before. Tomer added: “Once we have a mission plan developed we can repeat it over and over again. That allows us to see changes over time, and our deep learning system can see these patterns, what is changing and where you should place your attention.”
Autonomous robotic inspection offers a number of intrinsic benefits, which InnovOil has covered widely in the past. The ability to inspect assets while they remain in production means no lost revenue, and potential safety risks to personnel are reduced with less need for rope access. But the ability to source better data quality is also important – one UAV can perform multiple runs with different sensor packages, allowing operators a full suite of datasets from infrared imaging to point-cloud modelling, in a substantially reduced amount of time. This advantage only grows in the data-processing stage.
The GE Predix platform, on which the system is built, can analyse inspection data and identify defects automatically, and can also integrate other disparate external data sources, such as weather, to recommend future inspection and maintenance schedules. Data are then made available via a secure online platform, ensuring that customers have access to – and ownership of – their asset data.
The implications of AI and machine learning in this environment are, even for the technically minded, impressive. By training the AI to recognise specific components – achieved by showing it thousands of images of, say, bolts – the resulting neural network can identify the location and integrity of those components in data gathered from new inspections, and with a remarkable degree of detail.
“If you want to find all the bolts on a flare stack, for example, you can do that,” Tomer explained. “We can then run our corrosion [monitoring] system to identify how severe corrosion is, to six different levels of severity. You can do that today, and then next time you inspect the asset you can see the change in severity and act accordingly – and the system does all that automatically.”
The more images and data the system gathers, the more accurate it becomes. He noted that the same flare stack bolt dataset was then used to train the system to identify bolts and corrosion on a set of transmission towers. And unlike manual inspections, that learning process is compressed into a tiny window of time – Tomer said processing for the transmission towers took just ten minutes once the data were collected.
“That’s the key. Some people are doing robotic collection of data and then a lot of manual interrogation of that data,” he continued. “We’re trying to disrupt the inspection process from end to end, to truly digitise the process for inspection as a service.”
Awaken the Kraken
As these capabilities grow, Avitas Systems has incorporated more technologies into the platform. In September it announced a partnership with chipmaker NVIDIA to enhance its AI capability, and even more recently signed an agreement with subsea technology vendor Kraken Robotics. This will see the latter’s SeaVision sensor adapted as a payload for autonomous underwater vehicles (AUVs).
SeaVision is a highly accurate 3D laser scanning system, capable of producing very high-resolution scans in full colour. Borrowing techniques from the company’s Synthetic Aperture Sonar system, SeaVision uses a twin-scanner configuration to produce a 1.2 million point scan in four seconds, either from a stationary position, or in motion. Processing the data enables the creation of a detailed 3D model of an asset for existing autonomous systems from Avitas Systems.
“3D modelling with SeaVision is faster and higher resolution than most systems on the market today,” added Tomer. “We’re going to mount the laser system on underwater vehicles for 3D modelling of subsea assets. The combination of their laser system and our autonomous system allows us to inspect the same points on assets, identify defects and detect changes over time,” he said.
The result is an even more detailed picture of subsea assets over time, and faster notification in the event of a failure.
Avitas Systems has already had “an overwhelming response” to the technology from operators, Tomer noted, though he admitted to a familiar frustration that the contracting structure of many international oil companies (IOCs) means it can be uphill battle getting new services like this approved. Nevertheless, with the level of detail and time saving on offer, it seems that it will not take long before the autonomous inspection model becomes an industry standard.
With autonomous systems now capable of handling much of the legwork of asset inspections, Avitas Systems will focus on improving efficiency even more in future.
One avenue will be to devise payloads capable of gathering multiple datasets in a single run – simultaneous capture of infrared and RGB images, for example. Tomer said that the company was already working on that capability, and that the team was just a few months away from being able to undertake a mission with multiple sensors in one payload.
In addition to larger and more complex payloads, some improvements will be enabled by regulatory changes. Few environments currently allow UAV operations beyond line of sight (LOS), for example. That makes inspecting linear assets such as pipelines and transmission lines slower, Tomer said, largely because of the difficulty in conducting long-distance scans while maintaining visual contact. Yet as these technologies becomes more accepted, autonomous systems such as Avitas Systems will likely be some of the first into the breach.
“We’re taking a systems-by-systems approach. In the next five years, we will be working on things that will make us even more efficient than today. As our deep learning system learns more, and as our robotic capabilities extend to untethered and beyond line of sight, we will become more efficient and accurate,” he added.
What is clear, however, is that these technologies are already reshaping how the industry approaches challenges associated with asset integrity. Rather than cyclical programmes of arbitrary inspection and maintenance, Avitas Systems offers a more intelligent and predictive solution, informed by learning gathered from thousands of other assets. “Today’s inspections are time-based, and they are based on rules of thumb. Some are too late, and some are too soon,” he explained. “But five years from now, with all the data we can bring in, I think we’ll be able to predict when you need to do those inspections – so you can do more frequent inspections on high-risk areas and less frequent on the low-risk areas, and tailor your strategy to the asset itself.”