Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / How Hardware Startups Use Frequency Response Analysis for Reliability
Transcript
- Lucas: If these conversations are useful for what you're building or running every day, we want to make sure they stay easy to access and free of interruptions. Luna: We deliberately don't run ads on these episodes, and that's a choice. If you'd like to support that, you can do it at buy me a coffee dot com slash fexingo. Lucas: Totally. It's a small thing that helps us keep the focus on the engineering and the decisions that actually matter for hardware founders. Lucas: All right, so today I want to talk about something that often gets overlooked in early-stage hardware: frequency response analysis. Or FRA for short. Luna: I've seen that term pop up in reliability testing docs. But what does it actually tell you that a simple vibration test doesn't? Lucas: Great question. A standard vibration test shakes the product at a set level across a broad range of frequencies and sees if it breaks. But FRA is different — it's about identifying the specific natural frequencies of your structure. Lucas: Think of a tuning fork. If you hit it, it rings at one pure frequency. Every physical object has those resonant frequencies, and if your product's operating environment hits one of those frequencies, the vibration amplitude can amplify by ten or even a hundred times. Luna: So you're saying a drone flying at a certain RPM could excite a resonance in its arm, and suddenly the arm is flexing way more than the motor's raw vibration would suggest? Lucas: Exactly. I was talking to a team last year building a delivery drone. They had a prototype that kept cracking the motor mounts after about fifty flight hours. They'd done standard vibe testing — passed with margins. But when they ran an FRA, they found a strong resonance at 110 hertz. Lucas: Turns out the propellers were producing a 110 hertz excitation at cruising speed. The arm was amplifying that vibration by about eight times. The crack wasn't from static overload — it was fatigue from resonance. Luna: And that's something a random vibration test might not catch because it's averaging energy across a wide band. The peak at that one frequency gets smoothed out. Lucas: Right. A random vibration profile might have a flat power spectral density from 20 to 2000 hertz. It's like playing white noise. But the drone's actual environment is more like a single note. FRA reveals that note. Luna: So how do you actually run an FRA? Is it expensive equipment or can a startup do it in a small lab? Lucas: You need three things: an excitation source — like a small shaker or even an instrumented hammer — an accelerometer to measure the response, and software to compute the transfer function. The transfer function shows you the gain at each frequency. Lucas: An entry-level setup with a modal hammer, a single accelerometer, and open-source software like ModalVIEW can run you under ten thousand dollars. That's a fraction of what a full environmental test chamber costs. Luna: And the insight is huge. I remember a robotics startup I advised — they were building a robotic arm for assembly lines. The arm would develop a high-pitched whine after a few hundred cycles. They checked bearings, lubrication, everything. Finally did an FRA and found that the wrist joint had a natural frequency at exactly the motor's commutation frequency. Lucas: That's a classic. The motor driver's pulse-width modulation creates a tiny torque ripple at a specific frequency. If that matches a structural resonance, you get noise and accelerated wear. In your case, did they fix it by changing the stiffness or the motor tuning? Luna: They stiffened the wrist housing with a ribbed design. Moved the resonance up to 2.3 kilohertz, well above the operating range. The whine disappeared, and the gearbox life doubled. Lucas: That's a perfect example of using FRA as a design tool rather than just a pass-fail test. You identify the problem frequency, then you shift it. You can add mass, increase stiffness, or add damping material. Luna: How early in development should a startup do this? I imagine doing it on a prototype is better than waiting for the first production run. Lucas: As soon as you have a physical representative prototype. Ideally during the first engineering validation test, before you commit to tooling. Changing a plastic part's wall thickness by a millimeter is cheap in CAD. It's expensive in a steel mold. Lucas: I've seen startups order injection molding tooling, then discover a resonance issue in the first production samples. Now you're either scrapping the tool or adding a costly secondary operation like adding a tuned mass damper. Luna: So FRA is really about avoiding that kind of late-stage pain. It's a few thousand dollars and a week of testing that could save you months of rework. Lucas: Exactly. And the data is useful beyond just fixing resonances. The frequency response function gives you a fingerprint of your product's structural dynamics. If you run the same FRA after shipping a few thousand units and the peaks have shifted, that tells you the material or assembly process has changed. Luna: So you can use it for incoming quality control too. Like, we tested ten units from the first batch, and the resonance peaks were all at 210 hertz plus or minus two percent. If batch two shows a peak at 190, something's off. Lucas: That's a great application. I know a medical device company that uses FRA as a go/no-go check on every production lot of their infusion pump enclosures. The plastic resin's modulus can vary by batch depending on moisture, filler content. The FRA catches that variance in minutes. Luna: And it's nondestructive, right? Unlike a static load test that might crack the part. Lucas: Right. You're just tapping it with a hammer and measuring the ring. The part is perfectly usable after. That's a big advantage for low-volume production where you can't afford destructive tests on every unit. Luna: Are there any pitfalls? Things a hardware startup might get wrong when running their first FRA? Lucas: Several. The biggest is boundary conditions. If you test the product sitting on a foam pad, the resonance frequencies will be different than when it's bolted into its final enclosure. You need to replicate the actual mounting as closely as possible. Lucas: Another one is robbing the accelerometer mass. If you use a heavy accelerometer on a lightweight plastic part, the added mass shifts the resonance down. You're measuring the system plus the sensor, not the part alone. Luna: So use a tiny MEMS accelerometer or a laser vibrometer if you can afford it. Lucas: Yeah. And the other pitfall is not testing enough directions. A resonance in the Z-axis might not show up if you only tap in X. You need to do a triaxial measurement or test each axis separately. Luna: So FRA isn't a silver bullet, but it's a high-leverage tool that fits the hardware startup timeline and budget. Lucas: It really is. And I think as more startups move into high-performance applications — drones, robotics, medical devices — FRA will become as standard as thermal cycling. It's just good engineering. Luna: I'm wondering, are there any open-source or low-cost FRA software tools you'd recommend? Lucas: ModalVIEW is one. There's also PyFRA, a Python library that works with common data acquisition hardware. For under five hundred dollars in parts, you can build a setup that gives you 80 percent of the insight. Luna: That's the kind of approach that makes sense for a bootstrapped startup. You don't need a fifty-thousand-dollar polytec laser vibrometer to start learning. Lucas: Exactly. And the learning itself — understanding your product's resonant fingerprint — can prevent failures that are nearly impossible to diagnose after the fact. A cracked housing in the field could be resonance, or it could be a material flaw. FRA tells you which one. Luna: It turns guesswork into data. And for a startup, that's a huge competitive edge. Lucas: Absolutely. So if you're building hardware, put FRA on your test plan. Start simple, iterate, and save yourself the headache of field failures. Luna: Well said. And thanks everyone for listening. We'll be back next week with another deep dive.