Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / How Hardware Startups Use Accelerometers for Predictive Maintenance
Transcript
- Lucas: If these conversations are useful for what you're building or running, you're in good company. Today we're talking about accelerometers — specifically how hardware startups are using cheap MEMS accelerometers to predict when a motor or a pump is going to fail. Luna: And when I hear 'predictive maintenance,' I usually think of expensive wired vibration systems that cost tens of thousands per machine. So I'm curious what 'cheap' actually means here. Lucas: It means you can now buy a three-axis MEMS accelerometer for less than two dollars in volume. The kind that's inside your phone to detect orientation. Startups are putting those on industrial motors, running FFT — fast Fourier transform — right on a small microcontroller, and detecting bearing wear weeks before a catastrophic failure. Luna: So the sensor itself is cheap, but what about the rest of the system? The enclosure, the battery, the wireless module, the firmware? Lucas: That's where the startup's engineering challenge really sits. Let's take a real example. There's a startup called VibeSense — not a household name, but they supply a midwestern pump manufacturer with wireless vibration sensors. Their bill of materials for a single sensor node runs about eighteen dollars. That includes the MEMS accelerometer, a Nordic Bluetooth SoC, a small coin cell battery, and a 3D-printed enclosure. Luna: Eighteen dollars per node, and they just stick that on a pump? How long does the battery last? Lucas: About eighteen months with a 500-milliamp-hour coin cell, if they sample at one kilohertz for one second every fifteen minutes. That's the design trade-off: you can't stream continuous high-fidelity data and expect battery life. So they do edge processing. The sensor captures a short vibration burst, runs an FFT on the Cortex-M0 core, and only transmits a few summary metrics — like the peak energy in the bearing fault frequency band. Luna: And that's enough to catch a failing bearing? You're not getting the full spectrum? Lucas: It is, if you know the fault signature. A healthy bearing produces vibration at specific frequencies — ball pass frequency, cage frequency, that sort of thing. When the bearing starts to spall, energy shifts into higher harmonics. VibeSense's firmware tracks the ratio of energy in those bands relative to the baseline. If that ratio crosses a threshold, it sends an alert. Their reported result: a 38 percent reduction in unplanned downtime for that pump manufacturer. Luna: Thirty-eight percent — that's a real number. But how confident are we that an eighteen-dollar sensor can compete with a five-thousand-dollar wired system? Lucas: It doesn't compete on accuracy. A high-end piezoelectric accelerometer with a twenty-kilohertz bandwidth and a twenty-four-bit ADC will always give you richer data. But for the vast majority of industrial motors — pumps, fans, conveyors — the fault frequencies are below two kilohertz. A MEMS accelerometer with a three-kilohertz bandwidth can capture those just fine. The trade-off is in noise floor. MEMS sensors have higher noise, so you need to calibrate the threshold carefully to avoid false positives. Luna: Right, false positives kill trust. If the system sends an alert every week and nothing's wrong, the maintenance team stops paying attention. Lucas: Exactly. And that's the number one mistake I see hardware startups make: they set the threshold too aggressively because they want to catch everything. But the right approach is to baseline each machine individually. You install the sensor, let it run for a week, learn the normal vibration profile — which varies by mounting location, by load, by temperature — and then set the alarm bands at, say, two standard deviations above that baseline. Luna: That sounds like a lot of upfront work. How many sensors can a startup realistically deploy in a factory without a dedicated team? Lucas: VibeSense found they could deploy about fifty nodes per technician per week. Their trick was to design a peel and stick mounting with a pre-applied adhesive pad — no screws, no drilling. The accelerometer is already calibrated to the enclosure. So the technician just peels, sticks, walks away, and the node self-registers via Bluetooth to a gateway. That's where the startup mind-set matters: it's not just the sensor, it's the entire installation workflow. Luna: So the hardware is almost commodity, but the differentiation is in the firmware and the deployment system. That's a pattern we've seen in earlier episodes on modular design and design for assembly. Lucas: It is. And there's another layer: what do you do with the data once it hits the cloud? VibeSense sends each node's daily health score to a dashboard. But they also built a simple machine-learning model that correlates vibration changes with temperature and runtime. So if a pump runs hotter than usual and vibration increases, the system flags it as 'accelerated wear' rather than just an anomaly. Luna: And that reduces false positives further, because it's looking at two signals together. Lucas: Right. One pump manufacturer using VibeSense saw their false positive rate drop from 18 percent to 4 percent after they added the temperature correlation. And that was a firmware update — no hardware change. Luna: I want to push back on something, though. A lot of industrial facilities are wary of wireless — they worry about interference, about range, about security. How do startups overcome that? Lucas: It's a real concern. In a factory with welding equipment and variable-frequency drives, the 2.4 gigahertz band can be noisy. So some startups are moving to sub-gigahertz radios like LoRa or 868 megahertz. Lower bandwidth, but much better penetration through concrete and steel. You can cover an entire factory floor with a single gateway. Security-wise, they use AES-128 encryption at the link layer, and many now support certificate-based authentication for each node. Luna: So the technology exists. What's holding startups back from broader adoption? Cost? Trust? Lucas: I think it's the integration — or rather, the lack of it. A startup can sell you a hundred sensor nodes, but if you don't have a system to act on the alerts, it's just data. The successful startups are the ones that either partner with a CMMS — computerized maintenance management system — or offer a simple API to push alerts into existing workflows. VibeSense built a direct integration with Fiix, a popular CMMS, so when a pump health score drops below fifty, a work order is automatically created. Luna: That's the kind of thing that makes a purchasing decision. The maintenance manager doesn't want another dashboard — they want a system that fits into what they already do. Lucas: Exactly. And that's where I think hardware startups have a huge opportunity. The sensor is cheap, the compute is cheap, but the value is in the closed loop: detect, decide, act. If you can shorten that loop, you save real money. The pump manufacturer VibeSense works with estimates they saved about two hundred thousand dollars in the first year from avoided downtime alone. Luna: Against a cost of, what, a few thousand dollars for sensors and gateways? That's a compelling return. Lucas: And honestly, if today was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. Listener support is what keeps these episodes ad-free and focused on the real engineering decisions. Luna: Yeah, it really does make a difference. Every contribution goes straight into making the show better. Lucas: So back to accelerometers. One frontier I'm watching is the shift to running neural networks directly on the sensor node. There are already MEMS accelerometers with an embedded TinyML accelerator — like the Bosch BHI360 — that can classify gestures or vibration patterns without sending raw data anywhere. Luna: That could eliminate the need for a separate microcontroller. You'd have a single chip that senses, processes, and decides. Lucas: Right. And that brings the BOM down even further. I think within two years we'll see wireless vibration sensors at a ten-dollar total cost. At that price, it becomes economical to put them on every motor in a facility, not just the critical ones. Luna: Which means the next challenge isn't the hardware — it's the software to make sense of all that data. Startups that can scale their analytics without scaling their support team will win. Lucas: Exactly. And that's where we're headed. Next episode, we'll talk about how some hardware startups are using tiny ultrasonic sensors to detect gas leaks. Should be interesting. Luna: Looking forward to it. Until then, keep building.