Industries / Industrial equipment and services / IoT and industrial automation
Predicting failures
Predicting failures means watching machines for early signs of wear, such as a change in vibration, so that repairs can be planned before a breakdown.
5
Milestones achieved, from 1967 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1967
Watching the shaft spin
Bently Nevada, based in Minden, Nevada since 1961, was, in the words of the American Society of Mechanical Engineers (ASME), "the first to manufacture a commercially successful eddy current proximity probe," a sensor that measures the vibration of a turbine or compressor by watching its spinning shaft without touching it. ASME calls the launch of its standardized 5000 series monitors in 1967 "a watershed moment" that led to automatic shutdown protection for such machines, and named the sensors an engineering landmark in June 2024.
Credit
Donald E. Bently and the Bently Nevada engineering team
1978
Most failures are not age
In a report for the US Department of Defense, F. Stanley Nowlan and Howard F. Heap of United Airlines found that "some 89 percent of the items analyzed had no wearout zone," so an age limit could not improve them, and that only 6% showed pronounced wear-out. They added that for many complex items scheduled overhaul raises the failure rate, by adding failures of newly reworked parts. A past chairman of the Society for Maintenance and Reliability Professionals told Chemical Processing that "all reliability-centered maintenance came out of that study."
LessonReplacing parts on a fixed schedule does not prevent most failures; watching their condition can catch more of them.
Credit
F. Stanley Nowlan and Howard F. Heap, United Airlines
2001
Monitoring systems share a plan
In 2001 a team of Boeing, Caterpillar, Rockwell Automation, Newport News Shipbuilding, Penn State's Applied Research Laboratory and the standards group MIMOSA, partly funded by the US Navy, developed OSA-CBM, an open specification for condition-monitoring software. It implements ISO 13374, the standard that divides a condition-monitoring system into six blocks of work, from collecting data to advising what to do, so that parts from different makers fit together.
Credit
MIMOSA and the OSA-CBM team; ISO
2008
Shared data to practise on
Prediction methods became comparable once researchers could test them on the same public records of machines run until they failed. NASA's prognostics data repository holds a bearing dataset from the University of Cincinnati's Center for Intelligent Maintenance Systems and Rexnord Technical Services (J. Lee, H. Qiu, G. Yu and J. Lin, 2007) and simulated jet-engine run-to-failure data made at NASA Ames with the C-MAPSS engine simulation (A. Saxena and K. Goebel, 2008). Crossref counts more than 1,300 citations of each team's paper.
Credit
the Center for Intelligent Maintenance Systems, University of Cincinnati; NASA Ames Research Center
2024
A cloud vendor steps back
Amazon Web Services launched Amazon Monitron in December 2020, wireless vibration and temperature sensors with a gateway and machine-learning software to detect abnormal behaviour in machinery. On 31 October 2024 AWS closed it to new customers, said device sales would continue until July 2025 with no new features planned, and pointed customers to partner companies.
LessonA plant that builds its maintenance on a vendor's monitoring service depends on that vendor's decision to keep it.
Credit
Amazon Web Services
Watch for
Agents that do maintenance
IBM Research's AssetOpsBench tests AI agents on more than 140 human-written requests about sensor data, failure modes, forecasts and work orders. In IBM's runs GPT-4.1 completed the most tasks, 65%, and no model exceeded 70%; the authors say their measures track whether agents follow the workflow, not how accurate their diagnoses are.
Achieved whena plant operator or an outside group reports how often such an agent's diagnoses and work orders were right on its own equipment.
Claimed by
IBM Research, June 2025, revised August 2026
Maintenance models on the chip
MIPS says its Actus platform for robot joints lets "sensor fusion and predictive maintenance models run on the same chip that drives the motor" (see the Note "Maintenance inside the joint"). It names no customer and gives no test results or silicon dates.
Achieved whenthe chips ship and an outside party measures how well models running on them detect wear in working joints.
Claimed by
MIPS, November 2026
Open
01
Wear robot joints out deliberately
A test bench that runs small robot joints fitted with strain-wave gears (the compact, low-backlash gearboxes used in most collaborative robot arms) until they fail, recording motor current, vibration and temperature throughout, published as an open dataset with a model that warns before each failure.
The briefCredit · unclaimed02
Count the false alarms
A public tally of how often a predictive-maintenance system's alerts were right in routine use: every alert over a year on a set of real machines, each marked by the technician who checked it as a real fault, a sensor problem or nothing, together with the failures that came without an alert.
The briefCredit · unclaimed
Open 01 · the brief
Wear robot joints out deliberately
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test bench that runs small robot joints fitted with strain-wave gears (the compact, low-backlash gearboxes used in most collaborative robot arms) until they fail, recording motor current, vibration and temperature throughout, published as an open dataset with a model that warns before each failure.
Why it is still open
A 2024 failure analysis of the strain-wave gears in a Universal Robots UR5 arm, from Politecnico di Torino, says field data are "often restricted to harmonic drive and robot manufacturers." The open run-to-failure gear data found come from other machines, such as a three-stage planetary gearbox that the University of Ferrara ran for 1,006 hours to failure and published in December 2025.
Built with what exists
small strain-wave gear units sold for research robots; motor drivers that report current (chips that drive a motor and run models on the same silicon are described in the Note "Maintenance inside the joint"); low-cost accelerometers; PyTorch to train the warning model.
First step
Build one stand in which a strain-wave joint swings a fixed load through a repeated arc, log motor current at 10 kHz, and run it for a week.
Done looks like
At least ten joints run to failure, with the data published under an open licence, and a model trained on some joints that warns at least 24 hours before failure on at least eight of the others, with its false warnings counted.
Open 02 · the brief
Count the false alarms
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public tally of how often a predictive-maintenance system's alerts were right in routine use: every alert over a year on a set of real machines, each marked by the technician who checked it as a real fault, a sensor problem or nothing, together with the failures that came without an alert.
Why it is still open
Published results report detection methods and relative gains, not the share of alerts that proved real. A 2021 study at a coal-fired power plant and a steelworks reported a 90.25% average cut in false alarms against a baseline method, and warned that "frequent alarms can lead to discouragement and disregard for the system by the user." Searches found no field study giving that share for wireless monitoring.
Built with what exists
the alert export of any monitoring system and the plant's work-order history; a small language model tuned to sort technicians' free-text notes into the three outcomes, checked by a person (see the Note "One GPU to fine-tune"); a shared spreadsheet.
First step
Ask one plant already using wireless monitoring for last year's alert export and work orders, and match the two by machine and date.
Done looks like
Twelve months of alerts from at least 100 machines at two or more sites, each with an outcome, published with the share of alerts that were real and the count of failures that came without one.