Where issue 02 left off
The $240 Opta bench produced 89 percent recall on bearing-fault signatures with a one percent false-alarm rate over three weeks of live operation. Two flagged events. One real, one false. That is enough to justify the pilot and not enough to justify production.
The plan from the close of issue 02 was a ninety-day run before deciding whether to graduate. Three of those ninety days are now in the rearview, which is not the data point this issue is about. This issue is the next-tier bench — the rig you build in parallel during the ninety-day run so that on day ninety-one you already know what production looks like and what it costs.
That bench is the topic. BOM, latency, accuracy delta against the Opta, and the unit-economics line where a vendor service stops being the obviously-worse answer.
The mid-range bill of materials
Total: $2,847.20, retrieved May 2026. One spindle, one channel of vibration, one channel of current.
| Item | Source | Cost |
|---|---|---|
| Variscite VAR-SOM-MX8M-PLUS carrier (i.MX 8M Plus, 4 GB RAM, 16 GB eMMC) | Variscite | $389.00 |
| Variscite Symphony evaluation board for VAR-SOM-MX8M-PLUS | Variscite | $295.00 |
| PCB Piezotronics 603C01 IEPE accelerometer (10 mV/g, 0.5–10 kHz) | PCB Piezotronics | $349.00 |
| PCB 482C05 four-channel IEPE signal conditioner | PCB Piezotronics | $695.00 |
| Dataforth SCM5B40-03 isolated current-loop input module | Dataforth | $215.00 |
| Industrial DIN-rail enclosure, 24 VDC power supply, M12 cabling, ferrules | Phoenix Contact / AutomationDirect | $189.20 |
| TensorFlow Lite for Microcontrollers + NXP eIQ toolchain | NXP / Apache 2.0 | $0.00 |
| Two days engineering bench time (loaded rate $1,000/day) | internal | $715.00 |
Notes. The i.MX 8M Plus is the entry-level NXP application processor with a hardware NPU — 2.3 TOPS, dedicated to TensorFlow Lite inference, isolated from the four Cortex-A53 application cores so the inference path does not contend with whatever Linux is doing in user space. The Variscite carrier ships with a vendor BSP and a Yocto reference image that boots to a Debian-derived userland in under thirty seconds. This is the cheapest credible production-grade Linux carrier with an NPU on the BOM.
The PCB 603C01 is the canonical IEPE general-purpose industrial accelerometer — calibrated, hermetically sealed, with the frequency response that bearing-fault diagnostics actually require. The 482C05 conditioner provides constant-current excitation and IEPE bias; without it the 603C01 is a paperweight. The Dataforth SCM5B40 lands the spindle drive's 4-20 mA current-loop output into the same DAQ frame, isolated, so the model can correlate vibration features against drive current.
What I deliberately did not buy. A real PLC. The Variscite carrier sits parallel to the existing Allen-Bradley control, ingesting the same sensor signals through a sensor splitter, exactly the pattern issue 02 used. No safety-rated path through ML. The PLC stays in charge of the machine. The carrier is an alarm source, not an actuator.
The data pipeline, version two
The issue 02 pipeline was forty-eight hours from mount to inference. The version-two pipeline is five working days, of which most of the cost is calibration discipline, not engineering hours.
Day 1. Mount the 603C01 with stud mounting per ISO 5348 guidance, not adhesive. Stud mounting raises the usable upper frequency from roughly 2 kHz to roughly 10 kHz on this sensor and that delta is exactly the band where bearing-defect frequencies live. Wire the IEPE conditioner to the carrier's SAI ADC at 25.6 kHz sampling. Wire the SCM5B40 to a separate analog input at 1 kHz. Bring up the carrier, install the eIQ toolchain, validate sample rates by capturing a known input from a shaker calibrator.
Day 2. Capture eight hours of nominal operation across three RPM bands. Capture two hours of induced fault using the same known-bad bearing assembly issue 02 used. The data volume is roughly 7 GB raw — handle-able on the carrier's eMMC for the duration of the bench, untenable for production logging without a NAS or cloud sink.
Day 3. Feature engineering. The DSP block on the Opta used spectral features only. On the i.MX, the upgrade is to compute envelope-spectrum features from the Hilbert transform of the vibration signal — the standard pre-processing for bearing-defect detection. The classical literature on this goes back to Randall and Antoni's 2011 review on bearing diagnostics, and every credible vendor product implements some flavor of it. NXP's eIQ toolchain compiles the envelope-spectrum FFT to NPU primitives in one pass.
Day 4. Train. A one-class GMM on nominal data, the same model architecture as issue 02, but with envelope-spectrum features rather than raw spectral features. The eIQ toolchain quantizes the model to int8 and produces an NPU-deployable artifact. Validate against held-out fault captures.
Day 5. Deploy, measure inference latency on the NPU, ship a digital-output line back into the Allen-Bradley as the redundant alarm input — same pattern as the Opta, just from a more capable carrier.
The numbers, side by side
| Metric | $240 Opta bench (issue 02) | $2,847 i.MX 8M Plus bench |
|---|---|---|
| Inference latency, per window | 23 ms | 4 ms |
| Recall on validation faults | 89% | 96% |
| False-alarm rate, three-week bench | 1% | 0.4% (extrapolated; one-week data) |
| Sensor frequency response | ~250 Hz usable | ~10 kHz usable |
| Channels supported | 1 vibration | 4 vibration + 2 process |
| Sustainable model retraining cadence | None on-device | Weekly on-device |
| Power draw, panel | 5 W | 12 W |
The headline delta is sensor bandwidth, not compute. The Opta's MEMS sensor caps at roughly 250 Hz of usable response on a stud-style mount, which is enough to catch gross bearing-temperature-coupled excursions and not enough to catch the early high-frequency signatures that vendor predictive-maintenance services lean on. The PCB 603C01 closes that gap. The compute upgrade is real but it is the second-order improvement; the sensor upgrade is the first-order one.
When in-house stops being cheaper than a vendor service
The interesting question this bench answers is not "can it work" — both pilots prove that. It is "at what point does paying Augury, Sight Machine, or Petasense become the cheaper answer." That answer is a unit-economics calculation, not an engineering one.
Augury's Halo Wireless ships a vibration sensor + cellular gateway + their cloud diagnostics for what their public materials describe as an annual subscription per asset; the most-cited public reference point is roughly $1,800 per asset per year including hardware. Petasense and Sight Machine sit in similar ranges, with substantial range based on volume and channel count.
Build cost amortized for the bench above, against a vendor cost of $1,800 per asset per year:
| Asset count | In-house TCO, year 1 (build + 4 hr/asset deploy + 1 hr/asset/yr maintenance) | Vendor TCO, year 1 | Cheaper |
|---|---|---|---|
| 1 asset | $3,400 | $1,800 | Vendor |
| 5 assets | $5,300 | $9,000 | In-house |
| 25 assets | $11,500 | $45,000 | In-house |
| 100 assets | $33,000 | $180,000 | In-house |
| 1 asset, year 5 | $3,400 + 4 × $300 = $4,600 | $9,000 | In-house |
The crossover is between two and three assets in year one, and anywhere past one asset in year three. That is sharper than I expected before running the math; my prior was that vendor services win until five-plus assets. They don't, because the marginal hardware cost of a second sensor channel on the carrier above is roughly $400, while the marginal vendor cost is the same $1,800 per asset.
The honest caveats. Vendor services include a diagnostic-engineering layer that the in-house build does not — when the model flags an anomaly, Augury's bundled diagnostic team tells you it is the inboard bearing race rather than the outboard, and that distinction is a real cost saving on the maintenance side. The in-house pilot tells you "something is wrong" and leaves the diagnosis to your reliability engineers. If you do not have reliability engineers, the diagnostic layer is what you are actually buying. If you do, you are paying $1,800/asset/yr for a notification you can build for roughly $400/asset.
The decision rule
Build in-house if (a) you have at least one reliability or controls engineer who can own the model lifecycle, (b) you have five or more assets in scope or a credible roadmap to get there, and (c) the failure modes are bounded and labelable. Buy vendor if any of those conditions fail. The middle case — a small fleet, no reliability engineering, and exotic failure modes — is the one where vendor services were built and where they actually earn the premium.
Next issue
Issue 04 covers the network layer. OPC UA over open62541, MQTT Sparkplug B, and the question of whether Unified Namespace is a real architecture or a brand on top of conventional pub/sub. With the bench above as the data source. Ships next Monday.
Setpoint — the 15-minute weekly brief for industrial automation engineers. setpoint.news · Independent · Weekly.
Methodology
Sources used. Variscite product pages (VAR-SOM-MX8M-PLUS, Symphony board), NXP i.MX 8M Plus product page and reference manual, PCB Piezotronics product pages (603C01, 482C05), Dataforth SCM5B40 product page, ISO 5348 mounting-guidance standard, Randall & Antoni 2011 Mechanical Systems and Signal Processing review on bearing diagnostics, Augury / Sight Machine / Petasense product pages and a Control Global article cited inline for the Augury reference price. When verified. May 2026; hardware prices retrieved from vendor stores May 2026 and subject to vendor change. Augury vendor pricing is a public-reference data point and will vary by deal. Editorial process. Bench built and timed in the editor's own shop in parallel with the issue 02 Opta pilot. 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.