BRENNEN

Industries / Industrial equipment and services / Industrial services / Instrumentation and electrical

Troubleshooting

This work finds out why a loop, circuit or machine has stopped working properly, from a dead transmitter signal to a ground fault somewhere in a plant's wiring, and puts it right.

5

Milestones achieved, from 1955 to 1995.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1955

    Pulsing finds the ground fault

    Many plant power systems were run ungrounded, so a first fault from one phase to ground let the plant keep running, but the small fault current returned "by unknown paths", which made finding the fault impossible. Elof A. B. Hojding of Hägersten, Stockholm, filed in the US on 16 September 1955 (US 3,031,613, granted 24 April 1962) for a system that limits the fault current and switches it on and off rhythmically with a "ground-fault current pulser", so a technician can follow the pulsing current with a portable meter to the faulty motor or cable without disconnecting it. Pulse locating is now part of high-resistance grounding, in which the supply neutral is grounded through a resistor; a 2005 talk to an IEEE Power Engineering Society section in northern Canada says pulsing lets ground faults be "quickly located while the plant is running".

    LessonIEEE Std 242-2001, quoted in that talk, warns that ungrounded systems risk "burn downs from a second ground fault" when the first is not found, one reason they are used less than high-resistance grounded systems.

    patents.google.comewh.ieee.org

    Credit

    Elof A. B. Hojding

  2. 1978

    Troubleshooting studied as a skill

    William B. Rouse at the University of Illinois had people find the failed part in networks drawn on a screen, a stand-in for tracing a fault through a system. He reported in April 1978 that people fell further short of the best possible search as networks grew, that hurrying them pushed them to brute-force searching, and that a computer aid doing the bookkeeping cut mistaken diagnoses substantially. A 1979 study of 40 trainees in an FAA aviation maintenance certificate program found that skills learned with the aid carried over to work without it.

    ntrs.nasa.govapi.openalex.org

    Credit

    William B. Rouse, University of Illinois

  3. 1982

    Simulators train troubleshooters

    William B. Johnson and William B. Rouse compared ordinary teaching (reading, television lessons and quizzes) with two kinds of computer simulation, one of abstract problems and one of specific aircraft engines, for aviation maintenance trainees. They concluded that a mix of low- and moderate-fidelity simulations gave enough practice to be "competitive with" lecture and demonstration.

    Lessonpractice problems need not look like the real machine to teach the method of fault-finding.

    api.openalex.org

    Credit

    William B. Johnson and William B. Rouse

  4. 1990

    Live testing allowed by rule

    The US electrical safety-related work practices rule, published on 6 August 1990 (55 FR 32016), requires live parts to be de-energized before anyone works on or near them, unless the employer can show that de-energizing adds hazards or is infeasible. Its own examples of infeasible include "testing of electric circuits that can only be performed with the circuit energized" and work on circuits that are part of a continuous process in a chemical plant; other work practices must then protect the worker. Parts under 50 volts to ground need not be de-energized if that adds no exposure to burns or arc explosions.

    osha.gov

    Credit

    US Occupational Safety and Health Administration (OSHA)

  5. 1995

    Names for power problems

    IEEE 1159, approved on 14 June 1995 and published on 30 November 1995, is the recommended practice for monitoring electric power quality, the steadiness of the voltage a plant receives. IEEE describes it as giving consistent descriptions of disturbances on AC systems, defining normal conditions and departures from them, whether they start in the supply, the load or the interaction of the two, and covering monitors and how to interpret what they record. It was revised in 2009 and 2019.

    standards.ieee.orgstandards.ieee.org

    Credit

    IEEE Standards Coordinating Committee 22, Power Quality

Watch for

  1. A maintenance copilot cuts downtime

    Rockwell Automation says technicians at its Singapore factory have used an in-house GenAI-Powered Maintenance Copilot since October 2025. It runs on OpenAI's GPT-5.4-mini through Microsoft Azure, draws on notes from veteran engineers, manufacturing data and machine manuals, ranks the likely causes of an error and gives step-by-step fixes. The plant director credits this approach with 33% less machine downtime and service and parts costs down about 25%, by Rockwell's internal tracking, and Rockwell says learning to troubleshoot the plant's machines now takes three months instead of nine.

    Achieved whenan outside party publishes repair times before and after, by fault type, for the copilot alone.

    news.microsoft.com

    Claimed by

    Rockwell Automation, in a Microsoft feature of 24 September 2026

Open

  1. 01

    A loop fault bench

    A bench of 24 V DC, 4–20 mA loops, the current signal most field instruments send, with faults switched in by relays: open wire, short across the transmitter, either wire to ground, reversed polarity, a corroded high-resistance terminal, low supply voltage, a blown input fuse and a wrong range setting. (Shield and earthing faults that show up only as noise are the Installation item above.) For each fault, record what a technician would measure at each test point, publish the readings with answer keys, and score students, technicians and AI models on naming the fault from the readings.

    The brief
    Credit · unclaimed
  2. 02

    Rouse's tests for AI agents

    Rebuild as software the network fault-diagnosis tasks that William Rouse gave people from 1978, run AI agents on the same network sizes under the same paced and unpaced conditions, and compare their number of tests and wrong diagnoses with the human results he published.

    The brief
    Credit · unclaimed

Open 01 · the brief

A loop fault bench

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A bench of 24 V DC, 4–20 mA loops, the current signal most field instruments send, with faults switched in by relays: open wire, short across the transmitter, either wire to ground, reversed polarity, a corroded high-resistance terminal, low supply voltage, a blown input fuse and a wrong range setting. (Shield and earthing faults that show up only as noise are the Installation item above.) For each fault, record what a technician would measure at each test point, publish the readings with answer keys, and score students, technicians and AI models on naming the fault from the readings.

Why it is still open

The method is written down in makers' guides, such as Fluke's on using a loop calibrator's mA simulate mode to tell a faulty transmitter from faults in the wiring, power supply or input card, but searches found no public set of recorded loop faults with answers. The nearest AI tests use other equipment: CoffeeBench (Lucas McCabe, Sarah Lukens and H. Howie Huang, October 2026) has AI agents diagnose a simulated appliance from its sensors.

fluke.comarxiv.org

Built with what exists

A current-limited 24 V DC supply; loop-powered transmitters or a loop calibrator; a PLC or data-acquisition analog input with its 100 to 250 ohm input resistor; a relay board run by a Raspberry Pi to insert faults; a multimeter with a data port; AI models through their APIs.

  1. First step

    Build one loop on a bench with relay-switched open and short faults and log the readings at four test points.

  2. Done looks like

    The eight fault types and a healthy loop, each recorded five times on each of three transmitters, published as an open dataset with a scoring script, and baseline scores for at least ten people and three AI models.

Open 02 · the brief

Rouse's tests for AI agents

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Rebuild as software the network fault-diagnosis tasks that William Rouse gave people from 1978, run AI agents on the same network sizes under the same paced and unpaced conditions, and compare their number of tests and wrong diagnoses with the human results he published.

Why it is still open

Rouse's papers give human baselines, including the 40 FAA maintenance trainees of 1979. CoffeeBench (October 2026) measures how efficiently AI agents diagnose a simulated appliance, but its abstract reports no comparison with people. A search of Semantic Scholar for AI fault-diagnosis studies found work on optical networks, chemical processes, wheeled robots and aircraft maintenance, and none that reuses Rouse's tasks. NIKA (December 2025), a public benchmark of hundreds of computer-network incidents, scores AI agents on real network faults, not on Rouse's abstract networks, and its abstract reports no comparison with people.

arxiv.orgntrs.nasa.govarxiv.org

Built with what exists

The 1978 to 1982 papers in IEEE Transactions on Systems, Man, and Cybernetics and Human Factors; Python to generate the networks; an agent harness that lets a model test one connection at a time; general and open-weight AI models.

  1. First step

    Get the 1978 and 1979 papers through a library and write the network generator and scoring.

  2. Done looks like

    At least three AI models run on 100 problems at each network size used in the papers, with tests per diagnosis and error rates set beside the published human figures, and code and logs public.