Industries / Industrial equipment and services / Industrial services / Vibration analysis
Collecting data
Collecting data is taking vibration readings from the same points on each machine at set intervals, and getting them into a form an analyst can compare over time.
4
Milestones achieved, from 1965 to 2015.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1965
Spectra by fast algorithm
A spectrum splits a vibration signal into the frequencies it contains, and many machine faults show up at a frequency of their own. James W. Cooley and John W. Tukey published an algorithm for computing spectra by machine in 1965, now called the fast Fourier transform (FFT). By 1971 an FFT analyzer built at the Johns Hopkins Applied Physics Laboratory transformed 4,096 samples in under 12 ms and was in use for vibration, seismic, radar and sonar signals. The HP Memory Project, a history of Hewlett-Packard's instruments, describes HP's 3582A of 1979 as the first real-time spectrum analyzer built around a microcomputer running the FFT, and says it measured one to two orders of magnitude faster than a swept analyzer.
Credit
James W. Cooley and John W. Tukey; A. M. Chwastyk of the Applied Physics Laboratory; Hewlett-Packard
1971
Spectra added to the rounds
E. Downham of the University of Aston and R. Woods of Shell Chemicals UK told a conference of the American Society of Mechanical Engineers (ASME) in Toronto in September 1971 that comparing a machine's overall vibration with published limits "will frequently result in quite major failures going undetected". At Shell's Carrington plant they added frequency analysis to routine readings. In 1970 the gearbox of a 1,000 hp boiler feed pump rose from 8 to 18 mm/s overall in about two and a half months, with new peaks near 1,250 Hz and 4,000 Hz that were traced to a worn drive coupling; the gearbox failed two months later. On a 500 hp gearbox, a peak at 1,750 Hz led to a shutdown and new gears, after which overall vibration fell from up to 16 mm/s to 2.8 mm/s.
LessonOne overall number compared with a limit misses many failures; the spectrum shows which part is going.
Credit
E. Downham and R. Woods
1983
Readings stored, not written
A patent filed on 14 November 1983 by Ronald G. Canada, E. Forrest Pardue, Kenneth R. Piety and Daniel G. Simpson of Technology for Energy Corp. describes how route readings were then taken. The most common method was a portable analog meter: the technician wrote down one overall number per point and took the list to an engineer. The other was a bulky tape recorder, which could take several hours per machine. Their battery-powered instrument stored readings and frequency-band data in memory and passed them to a base computer, which kept the database and produced exception reports, lists of only the readings outside their limits. The patent, US 4,520,674, was granted on 4 June 1985. Canada later owned about 22.2% of Computational Systems Inc. of Knoxville, Tennessee, which designed a complete vibration system in the 1980s and which Emerson agreed to buy in October 1997.
Credit
Ronald G. Canada, E. Forrest Pardue, Kenneth R. Piety and Daniel G. Simpson
2015
Video checked against sensors
Motion magnification is video processing that exaggerates movements too small to see, so that a camera can show where a structure moves. Justin G. Chen and colleagues at MIT filmed a vibrating cantilever beam with a high-speed camera while measuring it with a laser vibrometer and an accelerometer. The camera's displacement matched the laser's with a correlation of 99.6%, but between 0 and 2,500 Hz it showed only the first four of the eight resonances the other two sensors found, and its noise floor was 40 nm against 0.2 nm for the laser and 0.02 nm for the accelerometer. Camera systems sold for machinery, such as RDI Technologies' Iris M, treat each pixel as a motion sensor.
LessonA camera shows the shape of the motion; small, fast vibration still needs a sensor on the machine.
Credit
Justin G. Chen, Neal Wadhwa, Young-Jin Cha, Frédo Durand, William T. Freeman and Oral Buyukozturk, MIT
Watch for
Robot rounds in cement mills
ANYbotics and Vigier Cement say a four-legged ANYmal robot inspects the mills of Vigier's plant in Péry, Switzerland, every night without supervision. Interesting Engineering reported in June 2026 more than 33,000 inspections at over 450 checkpoints since the robot entered service in early 2025, a cracked crusher foundation found before it caused an estimated $630,000 of lost production, and a bearing whose temperature was heading toward 140°C. Processing magazine reported in August 2026 that the plant began using the robot in 2023, and quoted ANYbotics' Brecht Vandekerkhove giving 35,000 inspections at 350 to 400 points in 2025. The robot carries cameras, a thermal camera, a microphone that hears ultrasound and an acoustic imager, not a contact vibration sensor.
Achieved whenthe plant or an outside party publishes, for a stated period, which confirmed faults the robot flagged first, which the human rounds or fixed sensors caught, and which were missed.
Claimed by
ANYbotics and Vigier Cement, June and August 2026
Open
01
Measure how much readings vary
A published measurement-system study for route vibration readings. Factories call this a gauge repeatability and reproducibility (R&R) study: it splits the scatter in repeated measurements into the part due to the instrument, the part due to who takes the reading and, here, the part due to how the sensor is fixed. Five or more technicians each measure ten marked points on a small, guarded, running test rig three times, with a magnet on bare paint, a magnet on a flat pad and a screwed-in stud.
The briefCredit · unclaimed02
Contact readings by robot
A mobile robot that follows a set route around a small, guarded test rig and places an accelerometer on marked points as a technician would. Its readings are compared with sensors screwed permanently to the same points and with a technician's readings taken the same day.
The briefCredit · unclaimed
Open 01 · the brief
Measure how much readings vary
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A published measurement-system study for route vibration readings. Factories call this a gauge repeatability and reproducibility (R&R) study: it splits the scatter in repeated measurements into the part due to the instrument, the part due to who takes the reading and, here, the part due to how the sensor is fixed. Five or more technicians each measure ten marked points on a small, guarded, running test rig three times, with a magnet on bare paint, a magnet on a flat pad and a screwed-in stud.
Why it is still open
Practitioners report large effects, but no measured study was found. Bill Hillman, writing for Ludeca in 2010, says he has seen readings "vary by as much as 50%" when the sensor was moved a quarter of an inch, and an article credited to DLI Engineering on Maintenance World says that with a magnet mount "one can never be sure that data is being collected in the same location every time". A search in November 2026 for a formal R&R study of route readings found practitioner articles and vendor pages, none with measured variance figures.
Built with what exists
A bench rig of motor, shaft and bearings, such as a SpectraQuest Machinery Fault Simulator or a home-built one; an industrial accelerometer with magnet and stud mounts, such as PCB Piezotronics' 603C01; a data acquisition board for such sensors, such as Digilent's MCC 172 on a Raspberry Pi; Python with SciPy and statsmodels for the analysis of variance.
First step
Build or borrow the rig, mark one point, and take 30 repeat readings yourself with each mount to see how large the full study must be.
Done looks like
At least 450 readings (5 technicians, 10 points, 3 mounts, 3 repeats), with the share of total variation due to each source given for overall velocity and for a high-frequency acceleration band, and the raw data and code published.
Open 02 · the brief
Contact readings by robot
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A mobile robot that follows a set route around a small, guarded test rig and places an accelerometer on marked points as a technician would. Its readings are compared with sensors screwed permanently to the same points and with a technician's readings taken the same day.
Why it is still open
Robots on plant rounds listen and look rather than touch. Boston Dynamics added acoustic vibration inspection to its Spot robot in September 2024, using an acoustic imager so that customers can "capture more data without installing more inline sensors or requiring more contact sensors". ANYbotics lists a thermal camera, a microphone covering sonic and ultrasonic ranges, a zoom camera and an acoustic imager for its ANYmal robot. Searches in November 2026 found vendor articles saying arm-mounted contact sensors are possible, but no published test of robot-placed readings against reference sensors.
Built with what exists
A legged or wheeled base with an arm, such as Boston Dynamics' Spot with Spot Arm or a wheeled base carrying a Universal Robots UR5e; ROS 2 and MoveIt 2 to plan the arm's moves; AprilTag markers at each point; a magnet-mount accelerometer such as PCB Piezotronics' 603C01 and a data acquisition board such as Digilent's MCC 172. A learned arm model with open weights, such as Runway's Praxis-1 (see the Note "Open robot model weights"), could be tried for the placing motion.
First step
Put AprilTags on an existing rig and measure how repeatably the arm reaches each point with the rig switched off.
Done looks like
Twenty filmed runs of a ten-point route, with the share of robot readings within 10% of the stud-mounted reference in overall velocity, the placement error in millimetres, and the same figures for a technician on the same days.