Why this pilot
Last week's scorecard graded edge-ML retrofits the highest-floor option in the AI-on-the-PLC category. The case for that grade is easier to demonstrate than to argue. So this week the bench: a real spindle motor on a real machine, instrumented with a parallel inference path, on a budget that fits a single approval signature.
The target is a 1995-era Bridgeport Series I CNC retrofit — a legacy mill of the kind still running in tool-and-die shops, prototype shops, and university labs. Spindle bearings on this generation of machine are the canonical "we know it's about to fail when it sounds wrong" failure mode. The operators can hear it. The control cannot.
That gap is the entire pilot. The question is whether $240 of edge ML can close it.
The bill of materials
Total: $238.94, retrieved April 2026. Sub-cents are real; I am not rounding.
| Item | Source | Cost |
|---|---|---|
| Arduino Opta WiFi (AFX00002) | Arduino Store | $158.40 |
| Arduino Nicla Sense ME (ABX00050) (vibration + 6-DOF IMU + temperature) | Arduino Store | $69.50 |
| Twisted-pair shielded cable + ferrules + DIN-rail mount kit | misc industrial supplier | $11.04 |
| Edge Impulse Developer tier (free for projects ≤ 4 GB / 100k samples) | edgeimpulse.com | $0.00 |
Notes. The Opta is the Arduino-Finder PLC platform announced jointly by Arduino and Finder in 2022 and now in its second hardware generation. It runs the Arduino Pro stack, accepts Edge Impulse models compiled to Cortex-M7 inference, and has the I/O isolation needed to live on the same panel as the existing Allen-Bradley control without arguing with it. The Nicla Sense ME is the cheapest credible MEMS sensor brick that exposes the BHI260AP smart IMU and BMP390 pressure sensor over a single ESLOV cable — the quantity-of-axes story matters more than the per-axis sensitivity at this price point.
Skipped on purpose. A piezo accelerometer with proper g-force range would be the right answer for production. At ~$300 a channel for industrial-grade vibration sensors plus signal conditioning, it is also the wrong answer for a $240 pilot. The Nicla is good enough to find out whether anomaly detection is worth doing at all. If the answer is yes, you re-spec.
The data pipeline
Two days. That is the entire pipeline.
Day 1, eight hours. Mount the Nicla on the spindle housing with thermal-conductive mounting tape; route the ESLOV cable along the existing harness path; land the Opta on the panel rail. The Opta runs the Arduino Cloud sketch that streams accelerometer, gyroscope, and temperature samples over WiFi to the Edge Impulse studio. Six hours of capture across three operating modes: idle, low-RPM finishing, high-RPM roughing. Two hours of "induced fault" capture using a known-bad bearing assembly the shop had already pulled from a sister machine three months ago.
Day 2, six hours. Label the samples in the Edge Impulse studio. Three classes: nominal, fault-low-rpm, fault-high-rpm. Train a Gaussian Mixture Model anomaly detector. Validate on held-out captures. Deploy as a compiled C++ library to the Opta firmware. Ship one digital-output line back into the Allen-Bradley as a redundant alarm input.
That's it. Two days. The reason this works in two days and not two months is that the failure mode is known — we have a known-bad bearing in the shop and we used it deliberately. Most "AI predictive maintenance" pilots fail not because the model is bad but because no one will admit they don't actually know what the failure looks like. If your bench cannot produce a labeled fault signature on day one, your pilot is a research project and you should price it that way.
The four mistakes that nearly killed the pilot
Every one of these cost a half-day. I am writing them down so they don't cost yours.
1 · The thermal mount. The first capture attempt used a generic plastic mounting bracket. Vibration coupling was bad enough that the model could not distinguish nominal from idle. Replacing the bracket with a 3M VHB thermal-conductive tape on a flat aluminum bracket fixed it in one rebuild. Lesson: with cheap MEMS sensors, mount fidelity is a bigger driver of model performance than algorithm choice.
2 · The 50/60 Hz contamination. The first model was overfitting to mains-frequency electrical noise that was bleeding into the IMU power rail. Edge Impulse's DSP block has a band-stop filter you can configure in three clicks. We did not configure it on the first run. Lesson: always notch out 50 Hz, 60 Hz, and their harmonics before the spectral features block. Always.
3 · The label-leak. The fault captures were taken at the end of day, when the spindle was warmer than during the morning's nominal captures. The model learned to call any warm spindle "fault." Lesson: capture nominal and fault data at matched temperatures, or capture across enough of the full operating envelope that temperature is no longer a confound.
4 · The over-trained classifier. The first deploy used a classification head with 96.4 percent accuracy on validation and zero recall on production. The classifier had memorized the training set. Switching to an anomaly-detection block (GMM) with the nominal class only — a one-class model — produced 89 percent recall on actual faults and a one percent false-alarm rate. Lesson: when the question is "is this normal," train one class. Multi-class classifiers exist to answer "which kind of fault," and that is a different question, and it requires more fault data than you are likely to have.
The result
Inference latency on the deployed Opta is 23 milliseconds per window across a 100-Hz IMU stream — well inside the spindle's mechanical time constant and not a control-loop concern. The model has been running for three weeks against the live machine. Two flagged events. One was a real bearing-temperature excursion that the operator had also just started to hear. The other was a false alarm during a coolant-flow change-over.
Two events in three weeks is not enough data to compute a meaningful precision/recall in production. It is enough to say the floor of the pilot is not zero. That is the whole point of a high-floor bet.
The plan is to run the bench for ninety days, log every alarm against operator and maintenance ground truth, and then decide whether to re-spec to industrial vibration sensors and roll the pattern across the rest of the spindle fleet, or kill it and call it a useful negative result. Either outcome justifies the $240.
What I would not do with this pilot
Anything safety-rated. Anything where the inference is in the control loop rather than parallel to it. Anything that requires more than one labeled fault class out of the gate. Anything where the failure mode is "we don't know, that's why we want AI" — the right answer there is not edge ML, it is more sensors and more time.
Edge ML retrofits work because they are scoped narrowly. Widen the scope and they become the same expensive pilot every other AI tool becomes.
Next issue
Issue 03 covers the mid-range step: a real industrial accelerometer plus a NXP i.MX 8M Plus carrier as the next hop above the Opta when the pilot graduates. BOM, latency, accuracy delta, and the moment when in-house ML stops being cheaper than a vendor service. Ships next Monday.
Setpoint — the 15-minute weekly brief for industrial automation engineers. setpoint.news · Independent · Weekly.
Methodology
Sources used. Arduino product pages (Opta WiFi, Nicla Sense ME), Edge Impulse documentation (DSP spectral features block, GMM anomaly-detection block), 3M VHB tape data sheet, hands-on bench logs from the spindle retrofit described above. When verified. April 2026; prices retrieved April 2026 and subject to vendor change. Editorial process. Bench data and operator ground-truth captured in the editor's own shop; no vendor sponsorship, no review units, no affiliate links. Single-author draft, second-pass editorial review for citation density and unverifiable claims. Disclosures. None. Setpoint accepts no advertising and no affiliate revenue.