Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / How Hardware Startups Use Acoustic Testing to Find Hidden Defects
Transcript
- Lucas: So we have a great one today. I was talking to a founder who runs a compact drone startup, and she told me about a trick that cut their field failures by forty percent. And it wasn't a new sensor or a better motor. It was a microphone. Luna: A microphone? Like, listening to the drones? Lucas: Exactly. They built a test fixture that basically listens to each unit during a burn-in period, and the acoustic signature tells them if something's off. And the thing is, this isn't just for drones. Acoustic testing is becoming a real tool for hardware startups across the board. Lucas: And look, before we go deeper — I want to mention something that's been on my mind. A handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many of these shows. Luna: Yeah, it's a small group, but it keeps us completely ad-free and independent. Which is rare in business media. Lucas: So if you get value out of these conversations, that's the way to keep them going. No pressure, just something to consider. Now, back to the microphones. Luna: So is the idea that every unit sounds slightly different, and you can train a system to know what a good one sounds like? Lucas: Right. The physics is pretty straightforward. Every mechanical and electrical component has a characteristic acoustic signature. A bearing that's slightly off, a loose screw, a solder joint that didn't reflow properly — they all make tiny noises that you can pick up with a sensitive microphone. And in a sealed enclosure, you often can't see the internals, so listening is one of the few non-destructive ways to check. Luna: I remember reading about how some car manufacturers use acoustic testing for engine diagnostics. But for a startup, isn't that expensive? Lucas: That's the best part. You don't need an anechoic chamber. You can build a test rig with off-the-shelf microphones, an audio interface, and open-source spectral analysis software. The drone startup I talked to spent under two thousand dollars on the whole setup. And they used a simple pass-fail algorithm based on frequency bands. Luna: So what exactly are they listening for? Like, a high-pitched whine or a rattle? Lucas: Both, actually. For a drone, you have the motors spinning at high RPM, and each motor has a specific harmonic pattern. If a bearing is worn or not seated right, you get a broadband noise in a certain frequency range. Also, if a propeller is slightly out of balance, you'll see a spike at the rotational frequency. The test fixture captures that during a thirty-second run and flags it. Luna: That's genius. And it catches things that a functional test might miss? Because a functional test would probably just check if it flies. Lucas: Exactly. A functional test might say, 'Yeah, it hovers fine,' but it won't catch a bearing that's going to fail after fifty hours. Acoustic testing is predictive. It catches defects early, so you can fix them before shipping, not after. Lucas: And the same principle applies to PCB assembly. Solder joints, especially in fine-pitch components, can have microscopic cracks. When the board flexes or heats up, they make a tiny clicking sound. You can trigger that with a thermal cycle or a vibration burst and listen for the signature. Luna: So it's like giving your product a stethoscope. But I imagine there's a learning curve. How do you know what a good acoustic signature looks like? Lucas: You start with a few known-good units, record their signatures, and build a baseline. Then you can use a simple machine learning model to classify new units. The drone startup used a support vector machine, and they got to ninety-eight percent accuracy on their validation set. It's not rocket science, but you need to be careful about false positives. Luna: False positives must be a pain. If you reject a good unit, you waste money on rework or scrap. Lucas: Totally. And that's why you have to tune the threshold. The startup went through a period where they were rejecting ten percent of units, but most of those were actually fine. They had to adjust the algorithm, add more training data, and set a tolerance band. Eventually they got the rejection rate down to around two percent, and the true failure rate was one point five percent. Luna: So you're catching one point five percent real failures, but you're also rejecting half a percent good units. That's a cost, but it's much cheaper than a recall. Lucas: Exactly. And you can also use acoustic testing during the design phase, not just in production. If you have a prototype that's making a weird noise, you can analyze it to find the root cause. That's how they discovered that a particular motor mount was resonating at a certain frequency. Lucas: And that led them to change the material from aluminum to a nylon composite, which dampened the vibration. They actually improved the user experience too, because the drone was quieter. Luna: So it's a design tool as much as a test tool. And I'm guessing it's especially useful for sealed devices, like smart speakers or wearables, where you can't just open it up. Lucas: Absolutely. For a smart speaker, you have microphones and speakers in a sealed enclosure, and the acoustic testing can verify that the assembly is correct. If a gasket is misaligned, it will affect the acoustic response. You can catch that before you seal it permanently. Luna: So what are the limits? Like, when would you NOT use acoustic testing? Lucas: Well, it's not great for detecting electrical failures that don't produce sound, like a short that causes a voltage drop. That's why you still need functional test fixtures. And it can be tricky in noisy environments, so you need to isolate the test area or use a soundproof box. Lucas: But the biggest challenge is interpreting the data. You need someone who understands signal processing or is willing to learn. The drone startup had a mechanical engineer who was into audio, so that helped. Luna: And there's a cost to the test time, right? Thirty seconds per unit adds up if you're making thousands. Lucas: Right, but you can parallelize. You can build a test rack with multiple stations, each with a microphone. And the software can analyze the recordings in real time. The drone startup was able to test ten units simultaneously, so the per-unit cost was minimal. Luna: That makes sense. So for a startup, what would be the first step to implement this? Lucas: Start by recording your product in its normal operating mode. Just a few units. See if there are any obvious differences. Then, if you have a unit that fails later, you can listen to its recording and see if there was a clue. That's how the drone startup started — they had a batch of drones with high failure rates, and they went back to listen to the recordings from when they were made. Luna: So it's almost like having a black box for your manufacturing process. Lucas: Exactly. And once you have that baseline, you can set thresholds and automate the pass-fail. It's a great example of using cheap sensors and software to solve a problem that would otherwise require expensive inspection equipment. Lucas: And there's another angle: predictive maintenance. If your product is a machine that runs continuously, you can embed a microphone and monitor its health over time. That's a value-add for your customers, and it also gives you data on how your product performs in the field. Luna: So it's not just about quality control, it's about building a smarter product. And it's accessible to startups because the components are cheap and the software is open source. Lucas: Right. I think we're going to see a lot more of this. The combination of cheap sensors, machine learning, and the need for reliability in connected devices is driving adoption. And it's a great example of how hardware startups can use data to compete with bigger companies. Luna: Well, I'm convinced. I'm going to look into adding a microphone to our test process. What's the key takeaway for our listeners? Lucas: If you're building hardware, don't underestimate what you can learn by listening. A hundred-dollar microphone and some software can save you from a million-dollar recall. And that's a pretty good trade.