Research · Results · Experiment

Reading machines from the meter.

What a machine is doing, how often, and whether something is wrong with it, read from the averaged power reports an ordinary meter sends every few seconds. No waveform, no neural network, no GPU: the readers are fitted in one closed-form pass on a laptop CPU.

26 September 2026 · Apple M5 Max CPU, no GPU · eight data sources, every experiment preregistered · Report · Paper (PDF)

The question

How much of a machine is in its meter?

A clip-on current sensor or a revenue meter is the cheapest sensor a factory can add. It reports averaged power every one to ten seconds and keeps no waveform. Machine-monitoring products turn that signal into running, idle and off, cycle counts, utilization and downtime. We asked what else it carries, and whether the people who run a machine can teach a reader what they care about by labelling a few periods after the fact, in their own words, without anyone training a model.

We treat each operating period as a phrase of work: a load cycle, a compressor on-period, a milling pass. States and phrases are found with resolution rules read off the instrument's own noise, so no threshold is tuned. Conditions are read by inscription, our name for closed-form fitting: class means and one shrinkage covariance are estimated once from labelled phrases (shrinkage linear discriminant analysis) and new phrases are read by whitened distance.

Headline

Findings, with their controls attached.

0.9955running vs idle from compressor current by a label-free threshold reader against the machine's own valve signal, 1.5 million reports, 0.031 false state changes per hour

4 / 4air-leak failures recognised afterwards from motor current with each failure held out, at most 0.9 % false positives the week before; the open state-of-the-art classifiers flagged at most half the failure phrases at their default thresholds. Four events: a 95 % lower bound of 0.40, and not yet an early warning

6 msto fit and apply a condition reader on the CPU, where the open state-of-the-art classifiers took 0.7 to 400 s on the same fold, with accuracy on valve lag equivalent within 2 to 2.5 points

0.746accumulator pre-charge from 1 Hz motor power with time held out, by inscribing our read on a borrowed quantile feature map; the best any method reached before was 0.513

The operations metrics are in the meter. A label-free threshold baseline (and k-means) recovers running and idle, cycle counts within about 1.4 % a day, daily utilization within 0.18 points (median) and 99.6 % of stops over five minutes. Stops under a minute are not recoverable from 10 s reports by any reader we tried.
Condition is there when it changes load or cycle shape. Valve lag, air leaks, impeller damage, cavitation and tool wear move power; bearing and alignment faults raise it by about 1 %, inside a healthy machine's session-to-session spread, and were not reliably detectable.
Closed-form reads match the state of the art where our features fit, and beat its heads on borrowed features. On valve lag the inscribed reads match MiniRocket, MultiRocket-Hydra, QUANT and HIVE-COTE 2 (no difference detected, p = 0.75; equivalent within 2–2.5 points). On pump leakage and the accumulator those classifiers' feature maps are better than ours; holding a map in the Frame and reading it by inscription beat the maps' own heads (preregistered, p = 0.019) and read accumulator pre-charge above a shuffled-label null, which no read of our own features did. The winning feature maps are borrowed, not ours.
How you evaluate decides what you conclude. Random resampling of cyclic machine data inflated accuracy by up to 59 points over time-blocked evaluation; most published figures on these datasets use random splits.

The result

Read the report, or the paper.

Report

The full paper, with 17 figures

Methods, eight data sources, the operations metrics, condition across machines, the comparison with open state-of-the-art classifiers, the feature-map lift, negative results and limits.

Paper (PDF)

The same paper as a PDF

For reading offline and citing.

What was measured

On what, and how.

Two simulators built with the methods (they establish mechanisms, not field performance), the PLAID appliance corpus, the UCI hydraulic test rig, a metro train air compressor, a CNC milling machine, household appliances from a repair centre, and a centrifugal pump test bed. Every experiment was preregistered and sealed by SHA-256 before it ran; analyses added after a result was seen are labelled post hoc; failed gates and a stopped comparator run are reported. Held-out units are the strictest the data allow: blocks of time, the future, unseen homes, tools, failures, dates and operating speeds.

Evidence label: experiment. Results hold under the stated protocols on these machines. They are adaptively developed and retrospective: methods were revised on the same data, condition reads use complete phrases, and a frozen pipeline tested once on new machines is the next step. The operations metrics come from one compressor; published figures in the product category concern other machines.