Virtual sensors from networks of small formulas: eight registered tests of NAI-fnet

Machines often need a quantity they cannot measure where it matters, such as the torque in a robot joint or the temperature of a rotor, and estimate it on their own microcontroller from the signals they already have. NAI-fnet builds such virtual sensors as networks of small formulas exported as dependency-free C. Its method is not disclosed; this preprint reports eight registered tests of it as a black box, each with five hypotheses written down before the single run on the test data. On four devices with real data, NAI-fnet as released in NAI v26 was compared with tuned LightGBM and an MLP 32×32. Elbow torque of a Franka Emika Panda arm: 0.062 % nMSE against 0.060 % for tuned LightGBM and 0.090 % for the dataset authors' Gaussian process, with C code 17 times smaller and 45 times faster than LightGBM's; rotor magnet temperature of a traction motor: 1.76 K² against 6.62 for tuned LightGBM and 3.12 for the MLP; power of five wind turbines: 2.26 % NRMSE against 2.14 % for tuned LightGBM and 2.95 % for the IEC power curve; drag of an airliner beyond the measured Mach numbers: 27.8 drag counts against 42.8 for tuned LightGBM and 205.8 for a look-up table. The 95 % bootstrap intervals of these error ratios all exclude 1. On four suites of 40 open engineering tables, NAI-fnet extrapolated beyond the training data about as well as tuned boosting (geometric mean error ratio 0.92, interval 0.79 to 1.06) and was less accurate inside the data. In all, 32 of the 40 hypotheses held; the eight that failed are reported with the rest. The registrations of the four device tests, their manifests, harness scripts and the outputs of the registered runs: https://github.com/mindscanlabs/NAI-track-record/tree/main/formula-network. Interactive simulators of the device examples: https://naiscience.com/fnet.html

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24163945
Primary Topic
Sensor Technology and Measurement Systems
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Virtual sensors from networks of small formulas: eight registered tests of NAI-fnet

Smirnov Mikhail
Zenodo (CERN European Organization for Nuclear Research)
Sensor Technology and Measurement Systems
preprint

Virtual sensors from networks of small formulas: eight registered tests of NAI-fnet

Smirnov Mikhail
preprint en

Abstract

Machines often need a quantity they cannot measure where it matters, such as the torque in a robot joint or the temperature of a rotor, and estimate it on their own microcontroller from the signals they already have. NAI-fnet builds such virtual sensors as networks of small formulas exported as dependency-free C. Its method is not disclosed; this preprint reports eight registered tests of it as a black box, each with five hypotheses written down before the single run on the test data. On four devices with real data, NAI-fnet as released in NAI v26 was compared with tuned LightGBM and an MLP 32×32. Elbow torque of a Franka Emika Panda arm: 0.062 % nMSE against 0.060 % for tuned LightGBM and 0.090 % for the dataset authors' Gaussian process, with C code 17 times smaller and 45 times faster than LightGBM's; rotor magnet temperature of a traction motor: 1.76 K² against 6.62 for tuned LightGBM and 3.12 for the MLP; power of five wind turbines: 2.26 % NRMSE against 2.14 % for tuned LightGBM and 2.95 % for the IEC power curve; drag of an airliner beyond the measured Mach numbers: 27.8 drag counts against 42.8 for tuned LightGBM and 205.8 for a look-up table. The 95 % bootstrap intervals of these error ratios all exclude 1. On four suites of 40 open engineering tables, NAI-fnet extrapolated beyond the training data about as well as tuned boosting (geometric mean error ratio 0.92, interval 0.79 to 1.06) and was less accurate inside the data. In all, 32 of the 40 hypotheses held; the eight that failed are reported with the rest. The registrations of the four device tests, their manifests, harness scripts and the outputs of the registered runs: https://github.com/mindscanlabs/NAI-track-record/tree/main/formula-network. Interactive simulators of the device examples: https://naiscience.com/fnet.html

Zenodo (CERN European Organization for Nuclear Research)
Sensor Technology and Measurement Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon