Defining Energy Indicators for Impact Identification on Aerospace Composites: A Structured Feature Selection Approach Guided by Domain Knowledge

Imagine you are a detective trying to solve a mystery, but the only clue you have is a faint, crackling sound recorded on a broken tape recorder. In the world of aerospace, this "sound" is actually a vibration traveling through a high-tech airplane wing made of carbon fiber. When a bird hits the wing or a tool drops on it, the impact sends a shockwave rippling through the material. The problem is that the most dangerous damage—the kind that can make a plane fall apart—is often hidden deep inside, invisible to the naked eye. This is called "Barely Visible Impact Damage" (BVID). To find it, engineers use "Structural Health Monitoring" (SHM), which is like giving the plane a stethoscope. These stethoscopes (sensors) listen to the vibrations and try to guess how hard the hit was. But here's the catch: the signals are messy, full of static noise, and the math to figure out the force from the sound is incredibly tricky, like trying to guess the weight of a punch just by listening to the sound of the glove hitting the air.

For a long time, scientists have tried to solve this using two main tools. One is heavy-duty physics simulations, which are accurate but take forever to run on computers. The other is "Machine Learning," where a computer learns to guess the answer by looking at thousands of examples. But the computer often gets confused because there are too many different ways to describe the sound, and it doesn't know which parts of the sound actually matter. It's like trying to bake a cake by throwing every ingredient in the kitchen into the bowl—flour, salt, soap, and sugar—and hoping the computer figure out which ones make it taste good.

This paper, written by a team of researchers, proposes a smarter way to teach the computer. Instead of letting the computer guess which ingredients to use, they act as the head chef. They use their knowledge of how physics works to carefully select only the best "ingredients" (data features) before feeding them to the computer. They found that by picking specific parts of the vibration signal—like the loudness, the pitch, and how the sound changes over time—they could build a model that is three times more accurate than the usual methods. They proved this by testing it on a real composite panel, showing that when you give the computer the right clues, it can solve the mystery of the hidden damage much better, even when the data is noisy or incomplete.

The Detective's New Toolkit

The researchers started by realizing that the raw vibration signal is a chaotic soup of information. To make sense of it, they broke the signal down into three different "flavors" or domains, just like a musician might listen to a song for its rhythm, its melody, and its harmony.

1. Time Domain: This is the "loudness" and "shape" of the signal. How high does the wave go? How long does it take to peak?
2. Frequency Domain: This is the "pitch." Does the signal sound like a low rumble or a high squeal?
3. Time-Frequency Domain: This is a mix of both, looking at how the pitch changes as the sound plays out.

The team then created a massive list of potential clues, or "features," from these three domains. They had over 20 different ways to describe the signal, such as "Peak Amplitude" (the highest point of the wave) or "Centroid Frequency" (the average pitch). But having 20 clues is overwhelming, and many of them are just repeating the same information.

The Great Filter

To find the best clues, the researchers set up a rigorous filtering process, acting like a very strict editor. They didn't just guess; they used a step-by-step workflow guided by physics.

First, they asked: "Does this clue actually change when the impact gets harder?"
They used a statistical test (called an F-test) to see which features reacted strongly to different impact energies. If a feature stayed the same whether the hit was a gentle tap or a heavy slam, it was tossed out. This left them with a smaller group of "energy-sensitive" clues.

Next, they asked: "Are these clues independent, or are they just copying each other?"
Imagine you have a clue that says "The car is red" and another that says "The car is crimson." They are saying the same thing. The researchers used a technique called correlation analysis to spot these duplicates. If two clues were too similar, they kept the better one and threw the other away.

Then, they asked: "Can we trust this clue if the signal is noisy?"
Real-world sensors are often dirty or shaky. To test this, the researchers artificially added "white noise" (static) to their data, simulating a bad recording. They checked which clues remained stable and which ones went haywire. A clue that changed wildly with a little bit of static was rejected.

Finally, they combined these scores into a single "Selection Score." This score balanced how important a clue was with how reliable it was. The result was a tiny, elite team of just seven features:

• Peak Amplitude (PA): How loud the hit was.
• Energy Peak Ratio (EPR): How the energy was distributed.
• Rise Angle (RA): How fast the signal climbed.
• Peak Centroid Ratio (PCR): A measure of the sound's shape.
• Weighted Peak Frequency (WPF): The dominant pitch.
• Peak Frequency (PF): The highest energy pitch.
• Approximation Max Energy (AME): The energy in the low-frequency waves.

The Showdown: Smart Selection vs. The Black Box

To prove their method worked, the team set up a race between four different computer models.

1. The "Physics-Inspired" Model: This used their new, carefully selected list of seven clues.
2. The "Independent Features" Model: This used a standard method that just removes duplicates but doesn't check for noise or physics.
3. The "Candidate Features" Model: This threw all the original 20+ clues at the computer, hoping it would figure out the rest.
4. The "CNN" Model: This was a "black box" Deep Learning model (a Convolutional Neural Network) that looked at the raw sound waves directly, without any human-selected clues. It tried to learn everything from scratch.

The results were clear. The Physics-Inspired Model (Model 1) was the champion. It predicted the impact energy with an error rate of just 5.23%.

• The "Independent Features" model had an error of 13.97%.
• The "CNN" black box had an error of 14.95%.
• The "Candidate Features" model (throwing everything in) had the worst error at 16.48%.

The researchers found that the "black box" CNN was actually quite good at guessing the number, but it was a mystery how it did it. It was like a magician who pulls a rabbit out of a hat but won't tell you how. The problem is, if the magician's hat breaks or the rabbit changes color, you have no idea why the trick failed.

Why the "Physics-Inspired" Way Wins

The paper argues that the real victory isn't just the lower error number; it's the transparency. Because the researchers selected specific, understandable clues (like "Peak Amplitude"), they know exactly why the computer made its guess. If the computer says, "I think the hit was 20 Joules because the Peak Amplitude was high," an engineer can check that Peak Amplitude value to see if it makes sense.

If the sensor is broken and sending garbage data, the engineer can look at the specific clues and say, "Wait, the Peak Amplitude is huge, but the Rise Angle is weird. Something is wrong with the sensor." With the black box CNN, the computer just gives a number, and if the sensor is broken, the computer might give a wrong number without anyone knowing why.

The researchers also noted that their method is more efficient. The CNN had to process huge chunks of raw data, which takes a lot of computer power. Their model only needed seven numbers. It's the difference between reading a whole library to find one fact versus having a librarian hand you the exact page you need.

The Bottom Line

This paper suggests that in the world of aerospace safety, we don't need to rely on computers to guess everything from scratch. By combining human knowledge of physics with smart data selection, we can build models that are not only more accurate but also more trustworthy. The team showed that by filtering out the noise and focusing on the clues that truly matter, they could predict impact energy with three times the accuracy of standard methods. While they admit their method relies on expert judgment to pick the initial clues, the result is a system that is robust, fast, and easy for engineers to understand and trust. It's a reminder that sometimes, the best way to teach a computer is to give it a little help from a human expert first.