Everything we have published, filterable by topic, year and kind.
Results: 8
Average current, not peak MIPS, decides how long a sensing node lives. This note walks through the arithmetic for four common duty cycles, shows where datasheet sleep figures mislead, and gives a measured comparison of eight candidate parts under an identical one-per-cent duty workload with the same radio and sensor attached.
Second-sourcing every line is unaffordable; second-sourcing nothing is how a programme stops. We classify BOM lines by substitution cost rather than by unit price, and show on three real designs that qualifying alternates for the top twelve per cent of lines covered eighty per cent of the historical shortage exposure.
Datasheet bias-stability figures are quoted at room temperature on fresh parts, which is not how industrial equipment ages. We cycled forty six-axis IMUs from four manufacturers between minus 40 and plus 85 degrees Celsius for 500 cycles and re-characterised bias, scale factor and cross-axis sensitivity after each 100. Gyroscope bias degraded three to eleven times faster than the accelerometer channel, and the ranking between manufacturers changed after cycle 200, so a selection made on fresh-part data can be wrong within a year of field service.
Layout-induced gate-loop inductance remains the dominant cause of false turn-on in fast GaN half-bridges. We measure sixteen board revisions across a 3 nH to 21 nH loop range and show that overshoot scales almost linearly with loop inductance until the sink impedance of the driver dominates. A practical layout rule is derived, together with the measurement fixture needed to verify it without disturbing the loop under test.
A 2400-node metering pilot in a high-rise district produced a packet-error rate four times the link-budget prediction. Spectrum capture over six weeks shows the shortfall is not path loss but self-interference from synchronised reporting windows. Staggering the schedule by a node-derived offset recovered most of the margin without changing hardware, and the resulting scheduling rule is given in full.
A sealed enclosure can sit 25 K above ambient long before anything looks wrong, and electrolytic life halves for every 10 K. This note gives the internal-rise measurement method, a worked derating example for a 40 W supply, and the ripple-current budget that follows from it.
Vendor low-light claims are not comparable because the illuminant, exposure and denoising stage all differ. We define a fixed test scene, illuminant spectrum and raw-capture protocol, then publish results for eleven sensors so the numbers can be reproduced rather than trusted. The reference target is dimensioned in the appendix so any lab can build one.
Returns from three years of desert deployment were disassembled and traced to root cause. Sixty-one per cent of failures resolved to a single connector series whose retention degraded under daily thermal cycling, not to the semiconductors that the original FMEA had ranked highest. The revised FMEA weighting is included.
