March 23, 2020 |
International,
Land,
C4ISR
A modern mechanized military lives or dies by maintenance. But what if a computer could warn you when your weapons and vehicles were about to break, so you could fix them before they ever let you down?
By SYDNEY J. FREEDBERG JR.
WASHINGTON: Raytheon is working with the military on multiple pilot projects for AI-driven predictive maintenance.
What's that? Traditionally, military mechanics spend a huge amount of time on what's called preventive maintenance: They carry truckloads of spare parts to war, they consult historical tables of roughly how often certain parts wear out or break down, and they preemptively crack open the access hatches to check those parts on a regular basis. The idea behind predictive maintenance is to feed all that historical data into a machine learning algorithm so it can tell maintainers, vehicle by vehicle and part by part, when something is likely to fail.
It's a tremendous technical challenge that requires scanning in years of old handwritten maintenance forms, downloading digital records, and then constantly updating the database. Ideally, you want up-to-the-minute reports on things like engine temperature and suspension stress from diagnostic sensors installed in frontline vehicles.
You need to account not only for what kind of equipment you're operating, but how hard it's running for a particular mission and even where in the world it's operating, because environmental conditions like heat, moisture, dust, and sand make a huge difference to wear and tear. And you can't just push out a single software solution and call it done. You have to constantly update your data so the algorithm can continue to learn, evolve, and adapt to different situations.
But, Raytheon's Kevin Frazier and Butch Kievenaar told me, artificial intelligence and machine learning have advanced dramatically over just the last five years. Now Raytheon – a long-established defense contractor – is partnered with a flock of niche innovators to make it happen.
Currently, they told me, Raytheon is already conducting or about to launch several multi-month pilot projects, seeking to prove the technology's value to the military:
For the Army, they're working with a commercial partner on the M2 Bradley Infantry Fighting Vehicle, the mainstay armored troop transport of the heavy combat brigades, and the hulking M88 Hercules, a tracked “armored recovery vehicle” designed to tow broken-down battle tanks back for repair, if necessary under enemy fire.
For the V-22 Joint Program Office – which supports the Osprey tiltrotor for the Marines, Air Force Special Operations Command, and now the Navy – they're working on the V-22's collision-avoidance radar, a Raytheon product.
And across their customer base, they're looking at ways to do predictive maintenance on the many complex components Raytheon provides for a host of programs.
How does this work? Let's hear from Kevin and Butch in their own words (edited for clarity and brevity from a highly technical 50-minute interview):
Q: What kinds of problems can this technology help the military solve?
Kevin: Right now, maintenance is conducted either on a scheduled timeline or when something breaks. What we are trying to do is replace that one piece because you know it's about to wear out and prevent it from breaking.
Butch: One of the biggest things is you've got to understand what mission you're trying to achieve. If I'm trying to answer platform readiness questions, then I have to have certain data that's related to that topic. If I am trying to do supply chain analysis, I'm asking questions about where are critical parts and what size stockages we have to have to reduce turnaround time. So I'm answering a different question, and I'm looking at a different data set.
So the key to setting all this up is what you do on the front end with your data to give the data scientists so that we can refine the algorithm appropriately.
Q: AI/ML requires a lot of data. Is that data really available for all these different military systems?
Kevin: It is. It's in different states. Some vehicles have sensors on them. Some do self-diagnostics. Some of the older equipment, especially the support equipment, doesn't have any sensors on them — but they all have files. They all are in the maintenance system, so the data exists.
Data doesn't have to purely digital. It does have to be digitized at some point, but it doesn't necessarily have to start being digital. It could be maintenance logs that are hand-written, or the operator of a particular vehicle does a walk around and does an inspection report, writes that up — that's something that you actually can scan and input.
Now we can add so many different types of data that your whole data environment becomes much richer. It helps you get to that algorithm — and then to continue to take in that data and refine that model. You're still recording that data and getting data from both handwritten and digital sources to update your model and tune it, so that you're just that much more accurate.
Butch: What we're talking about is discrete algorithms solving for discrete problem sets. You look at the environment, and what the algorithm does is it learns.
You keep ingesting data. You can get it a bunch of different ways so your analytical tool continues to learn, continues to refine. I can do a physical download from the vehicle, or scan maintenance records, or get it all fed off of a downloader that automatically feeds to the cloud. It can be as fast as we can automate the process of that piece of equipment feeding information back.
For the Army and the Air Force especially, there is sufficient data over the last 15 that pertains to the impacts of combat. And we have it for different environments that you can then use to help train and refine the algorithms that you're using as it learns.
Kevin: You have to understand the impacts the environment has on how the vehicle is functioning and what type of a mission you're doing, because that will cause different things to wear out sooner or break sooner.
That's what the AI piece does. The small companies that we partner with, who are very good at these algorithms, already do this to some extent in the commercial world. We're trying to bring that to the military.
Butch: The really smart data scientists are in a lot of the smaller niche companies that are doing this. We combine their tools with our ability to scale and wrap around the customer's needs.
These are not huge challenges that we're talking about trying to solve. It is inside the current technological capability that exists. We have currently several pilot programs right now to demonstrate the use cases, that this capability that actually works.
https://breakingdefense.com/2020/03/raytheon-ai-fix-that-part-before-it-breaks
