How Machine Learning Models Can Benefit Aerospace Manufacturing
There’s a common misconception around machine learning that it’s a ‘magic’ technology that can be applied anywhere to improve everything. That being said, as a data-heavy industry, there are many ways that aerospace can reap the benefits of machine learning: improved speed and accuracy in design, manufacturing and services activities, to name a few. At Acubed, Project ADAM is investigating how to use machine learning as part of a flexible, overarching digital approach in these areas.
Machine learning is a type of artificial intelligence (AI). It enables systems to learn from data, identify patterns and then make decisions more efficiently than humans. On a basic, everyday level, it’s how streaming services know what you’ll want to watch next or how your email system filters out spam.
Let machine learning reduce unnecessary strain
So how does this apply to aerospace? Because of its long timeframes and emphasis on safety, the industry has a lot of very experienced, highly qualified workers. In comparison to, say, the car industry, many of the processes in aerospace manufacturing are manual, rather than automated. We believe machine learning can change this.
To be clear, this doesn’t mean replacing humans with machines. Our goal is to apply machine learning models that let computers take on some of the repetitive, time-consuming tasks in order to free up time for people to contribute more meaningfully in other areas.
Check out the three examples below to see how we’re using machine learning to improve design, manufacturing and services.
Machine learning and design
A design for a new part may look great in theory, but what about when it comes to actual manufacturing? How can machine learning models improve design processes?
We have worked with researchers at a leading research university to create a faster method of assessing manufacturability factors such as tooling, lead time and cost, leveraging machine learning.
First, we teach a program to identify mechanical parts based on topological information—data that describes geometrical properties and spatial relations. The program then applies that knowledge to automatically work out potential manufacturing processes for that part, including factors such as time, cost and materials. Traditionally, this task is split across numerous teams in different locations using unconnected tools. It’s a necessary but time-consuming process. We’re discovering that applying machine learning creates a more homogenous, streamlined process, enabling design and manufacturing teams to work closer together and optimize part design more quickly.
Machine learning in manufacturing and production
Aerospace manufacturing is highly distributed, meaning full, real-time visibility of the entire system is limited without digital tools that connect the dots. We are currently working with our A320 Final Assembly Line in Mobile, Alabama, to introduce a connected toolset that can detect production anomalies and address them before they become a blocker. Strictly speaking, this isn’t machine learning but rather data analytics.
There are typically three stages in creating machine learning models:
- Data collection
- Data engineering to build meaningful features
- Machine learning algorithms to deliver predictive insights
An aircraft production line is an extremely dynamic environment with a lot of moving parts. Capturing structured, accurate data isn’t easy, so we’re having to walk before we can run. But this is still potentially very valuable. By tracking production anomalies in real time and reporting them in a clear, digestible way, we’re enabling final assembly line managers to quickly see where the problem lies. Armed with that knowledge, they can reduce non-conformities and improve quality and efficiency. In the long term, we can use this data to apply machine learning in our lean production processes and more accurately predict the duration of a specific task, for example.
Machine learning in services
Designing an aircraft interior is a complex task with a huge number of variables and data sources that can take weeks. But what if you could design it at the touch of a button? Thanks to the power of machine learning, we are developing a solution that can automatically assess the feasibility of the engineering and even produce drawings, part lists, and pricing to speed up the process.
At the heart of this is a subset of machine learning called deep learning. While machine learning handles structured data like numbers and categories, deep learning can handle unstructured data like text, images or videos. Instead of being trained by a human, deep learning determines by itself the relevant features for the model to be trained on. To come back to our everyday analogy, its deep learning that powers the voice-activated virtual assistant in your home or enables autonomous cars—or aircraft—to navigate themselves.
By applying deep learning, we’re creating a program that can automatically recognize which part changes are needed to go from one specific customer configuration to another, indicating almost instantly what’s feasible and what isn’t.
What’s next?
Autonomous flight may grab many of the headlines when it comes to machine learning technology and aerospace, but as you can see, it also has the potential to bring benefits to design teams, manufacturing halls and assembly lines. In my role as a data scientist, I’m often operating in this fascinating space between software engineers, industrial engineers and the managers interested in the time and cost savings our solutions can bring. The understanding and appreciation of AI, machine learning and deep learning is increasing all the time but we still have quite a way to go, and I’m thoroughly enjoying finding ways to capture the data we need to drive continuous improvement in the future.