Vibration Analysis and Location

Don't know if this will work for you, but still pretty cool tech.....

https://www.nature.com/articles/srep16063

From a signal processing point of view, if you have a sound source and two sensors with a different delay due to their locations, the delay difference can be determined by a cross-correlation.

Cross-correlation can be done in the time domain or the frequency domain. Cross-correlation in the time domain is just a convolution with one of the signals time-reversed, and is rather computationally intensive.

In the frequency domain, cross-correlation is done by taking the FFT of each signal, and replacing the complex spectrum of one by its complex conjugate (change the sign of the imaginary part). Multiply cell by cell of the complex spectrum of one signal and the complex conjugate of the other, and then take the inverse FFT of this product. The result will be the cross-correlation.

Matlab can do this for you... y = crosscorr(x1, x2)

http://blog.prosig.com/2001/06/06/correlation-example/

So, your problem is locations of the sensors, and calculating the delay to each bearing location.

From a practical standpoint, I think that you would be much better off cutting your 1' x 4' piece of aluminum into 4 and monitor each with at least 4 sensors (2 pairs), mounted with the axes at right angles. The more noise sources you have to monitor, the lower your signal to noise ratio.

In sensor placement, the idea is to make sure the time difference is not close to the same for more than one bearing. I would probably write a Matlab simulation with a parameter for sensor placement and move it around until I got the delays spaced out.

I don't know how well this will work. Multiple bearings making noise will lower the SNR, and you may have to consider some sort of damping around the edges of the aluminum plate (wood frame?) to remove reflections.

Just my ideas...

GA, I agree that this is a better way to go. Maybe mount them with some sort of elastomer to isolate vibration from one another. You only need multiple sensors and electronically switch each one in turn to measure the level of vibration.

By "sweep the outputs" are you refering to a spectrum analysis? I'm just trying to understand a little more about this without having to invest in equipment I may not need.

When I did this, I used a spectrum analyzer, but I was sampling a single transducer on multiple items, pumps.

You might be able to do it by just measuring the relative amplitude at multiple points, depending on the variables.

Yes.

Thanks for the link.

This has revealed a lot of options for my project,and by other associated links,I think I have enough information to develop my project into a working model.

Thanks to all for their valuable time and input on this problem!

Bently Nevada have done this on rolling element bearings for years, REBAM I think they call it. They have application notes.

It has an accelerometer against the outer race. It is very high frequencies which reveal trouble with the bearing, 10s of kHz for contact of balls with metal fragments.

Having done vibration test on electronic equipment, I think you will find it difficult to locate an individual bearing from several fixed points. Transmission of noise on a plate is a maze of resonance patterns and reflections.

Detecting an increased general level of vibration, then probing each bearing with hand probe is probably the best you can do.

Mouser have a vibration sensors section. I have used the Murata PKGS-. They are small & cheap peizo, but you need charge amplifiers to use them.

These guys http://azimadli.com/vibman/default.htm have an online Vibration Analysis Reference manual that seems to target exactly the sort of thing you want to do. I have never contacted the company but there reference is the best single point of information I have ever found for vibration.

I suggest you access an ULTRA SOUND DETECTOR such as UE or SDT (no affiliation) since ultra sound is your first possible signature of problem of any moving device well ahead of vibration or heat.

The current devices record the sound level and you can easily trend your machinery. A little research should satisfy your questions in this area.

I've employed and evaluated numerous vibration monitoring systems, some of them automated and some not. For your application, I'd not recommend the AzimaDLI solution. I have it in service in my facility and to me it does not sound like a good fit for you.

The first question to ask is how fast are the shafts turning. In my opinion, if it's less than about 200RPM, you shouldn't use vibration. Ultrasound can detect the bearing failures reliably down to very low RPMs. UE Systems has some great technology and their staff is very knowledgeable.

If it is more than 200RPM, your best bet is to install a single axis accelerometer on each bearing. By doing this, and properly analyzing, you'll be able to tell the developing failure mode before it fails. You can typically identify inner race, outer race, ball, cage, ESD, or other issues based on the FFTs.

CTConline is a great company for sensors and cabling for vibration monitoring. They also have switch boxes that work great for manual data collection. IMI Sensors is another option.

For automatic data collection, I've seen systems by AzimaDLI, Augury, Flare Labs (Shinkawa), IMI Sensors, Pruftechnik, CSI (Emerson). Ensure that the chosen system has enough lines of resolution. If you go with something that can only do 800 lines for example, it may not be enough to adequately capture the data. I think these days 6400 lines is pretty typical on the low end. IIRC, my AzimaDLI system can do up to 32,000, which is not uncommon for a good vibration data collector.

One last option would be to install bearing fault detectors. These are typically DIN mounted devices that attach to an accelerometer mounted on a bearing. They can interface with a PLC to tell you when a specific bearing fails, though you typically do not get the FFTs or other data analysis possible with a standard monitoring system.

Hope that info proves helpful.

Thanks everyone for your thought provoking replies.

However,after due consideration,what I am seeking may not be currently possible.I cannot put a sensor on each bearing because they are moving,not simply rotating,and getting power to the sensor would be problematic.

I have considered a laser at a fixed location to analyze each bearing as it passes by,similar to extracting audio from a window pane,except at a higher frequency.

The scan on/off could be triggered by a photocell or other proximity detector to trigger a reading at each bearing.This would minimize extraneous noise,although I think that a very robust software would be needed to analyze the wave forms from the bearings.

I have used ultrasonic hand held detectors,and it takes a while to learn to decipher the signals from bearings.

A brand new bearing is very noisy at first,while the grease is being pushed out of the ball path,but it gets quieter after a while,then as wear sets in,the sound changes drastically.Even a good bearing sounds like a freight train,but there are certain characteristics that are indicators for eminent failure;after a while,it is much easier to tell good from bad but there is still a learning curve.

Perhaps AI software could learn these differences and sift the data from the background.

I will continue my search in the meanwhile,or perhaps my needs will inspire a company to develop a method.

Thanks again to all contributors.

While this may be extreme, it is possible to conduct an inspection through a stream of water or other incompressible fluid while the sensor and area of interest are connected by the stream. It may be too messy for your application but it is certainly used for ultrasonic inspection without contact (except for the stream of water). It may be limited, however, to vertical water flow but it is considered to be a type of waveguide or conductor. This technique was used to look for problems with a carbon fiber wing section of a certain fighter plane.

It is incorrect to assume that a regular rolling element bearing wears out. If you install a good bearing properly, align it very well, and maintain it with the proper amount/type of lubrication, the bearing will last indefinitely if operated within its design parameters. It will only deteriorate after a defect is introduced.

do you know the root cause of your failures? Perhaps understanding that will give you a more proactive solution instead of just waiting until the bearings are already failing.

I am familiar with the expected life of bearings,and the main reason for reduced bearing life is overloading and excessive grease.

The manufacturer's MTBF is based on ideal conditions,not the variables that occur in the real world,such as vibration in multiple axes,seal/shield failure,excessive temperature,hostile environment, and improper lubricant.

Doubling the load on a bearing will decrease bearing life by 90%,while doubling the rpm will decrease life by only 10%,which illustrates the need for proper tensioning of belts.

Too much grease causes churning and overheating of the grease,which carbonizes the grease,and results in premature failure.

The problem when dealing with literally 1000's of bearings is proper installation and maintenance.It is hard to get consistency among hundreds of maintenance employees across multiple shifts.

I agree that a scheduled replacement is often desirable over a one-by-one replacement,but the scale of this method,and the required down time is not available.

A sub assembly replacement is not practical for this application,but it is a well recognized time saver in other applications.

I do agree with your statements about installation and maintenance,but these conditions are hard to achieve in the real world.

And I certainly appreciate your input on this.

Maybe you could adapt something like this....

"A High-Speed Vision-Based Sensor for Dynamic Vibration Analysis Using Fast Motion Extraction Algorithms"

http://www.mdpi.com/1424-8220/16/4/572/pdf

Thanks for the link,it may be useful.

I have glanced briefly at the basic principles involved,and they seem valid.

Perhaps this will be an answer to the problem.

Thanks!

Cochlear Implant to "Hear" Good Bearings

HiTekRedNek said, "

I have considered a laser at a fixed location to analyze each bearing as it passes by, similar to extracting audio from a window pane,except at a higher frequency.

The scan on/off could be triggered by a photocell or other proximity detector to trigger a reading at each bearing.This would minimize extraneous noise,although I think that a very robust software would be needed to analyze the wave forms from the bearings.

"

Consider gluing to your plate a sort of cochlear implant. Glue traffic sign retro reflective glass beads to the loose ends of short fibers (plastic, glass, metal, whatever) mounted on thin mounting strips which you later glue to your plate. As the plate passes the laser or the laser scans the strip, a video camera or an optical scanner sensor strip mounted beside the laser receives the reflected beams. This setup achieves an acoustic isolation of your source and optical sensor from your vibration sensor(remote sensing). It also achieves a scan of each bearing(indirectly via the plate and fiber strip) which you mentioned. The laser could be left on or strobed at desired(or scanned) frequencies. The implant would be sensitive to both amplitude and frequency. The bearing vibrations will be producing both amplitude and frequency changes in the fiber motions relative to bearing health.

There are two major modes one might choose, sensing good bearings(or all bearings on one plate good) or bad ones. I like sensing good ones which may get you more false bads but will likely not produce so many false good signals. In either case some knowledge of bearing behavior(which you seem to have) and some empirical experimentation will likely be necessary. Getting a bead(including glue) mass, a fiber stiffness and length, a laser strobe frequency, etc. which works well is the goal but the simple temporary(hot glue) attachment of vibrating fiber arrays and sweeping the laser strobe frequency makes an approach to this experimentation straight forward. Individual fibers in an implant strip may not necessarily be uniform in their characteristics and indeed may be designed to present a gradient of characteristics for a shmoo result. After some experimental work, I suspect that you will be able to home in on an implant which will rapidly and easily spot the majority of good bearings and you can, perhaps, sort the chaff for further intensive analysis if you wish.

Basically, the robust software you mention is massively parallel-ized and mostly embodied mechanically in the fiber array(implant) design. A bearing on the edge of failing will drift out of the normal good amplitude and frequency range and the sensor near the laser will fail to see it anymore as a good bearing. One can statistically run the experiments with known good and (make them rare) bad bearings to converge on manufacturing specs for the cochlear implants. Your software may need to characterize individual sensing fibers on your implant fiber arrays but that can be done incrementally based on other known implants applied nearby geographically or temporally since it is trivial to substitute implants which require no sensor nor power wire tethers.

The fixed optical sensor(not on the plate) will inherently get a very jumbled signal. Pick those signal characteristics which are easy to perceive such as overall amplitude, amplitude at offset relative to results with a stationary plate and non-rotating bearings. Then statistically compare those results with moving results for all-good bearings, then find distinguishing differences when a single good bearing is replaced with a beginning-to-fail bearing. Do not corrupt the statistical analysis with human bias if you have that much self control since the results can easily be counter-intuitive. In your early data collection and analysis, pay statistical attention to fibers which may not be the closest to the bearing since the bad bearings may set up standing waves on the plate. Also, pay attention to the phase position of the plate relative to the laser. Attempt to identify not just that a bearing is failing but which plate position it happens to occupy. Be aware that one bad bearing may poison the data from the entire plate and may present different results in different plate locations. List me as a contributor on any related patents please.

thewildotter

Visualization

Note that with some implant design techniques one might under certain circumstances enable a human visualization mechanism that makes bearing status pop right out to a human observer standing near the scanning laser and looking at the implant due to stroboscope, heterodyne, visual persistence, or other possibly interacting phenomena. I would use a human safe laser so that this could be easily discovered, celebrated, and possibly commercialized. Tufte(Envisioning Information and other seminal monographs author) would love it !

thewildotter

A man was standing with one foot in a bucket of boiling water,and the other foot in a bucket of ice water.

Someone asked him how he felt,and he said

"On the average,I feel okay."

Averages do not tell the whole story.

Averages ?

Well, no doubt averages are lossy compared to the original sets of numbers you choose to average.. This setup however is measuring the energy transfer into multiple periodic oscillators resonant at a distribution of frequencies. That goes far beyond averaging. It has to do with spectral distribution of energy at some set of fundamental frequencies and at harmonics of those frequencies. The retro reflector deflections will each exhibit an amplitude related to the energy transferred from the plate at that location into the fiber. That amplitude will result in movement of the reflectors which can be observed to characterize the spectral distribution. The fiber will impart a two dimensional movement to its attached reflector which can be observed from a perspective near the laser source. For example, if a certain normal bearing vibration is related to the rotational velocity of the bearing due say to the center of mass of a bearing not being exactly on the axis of the bearing then a fiber resonant at that frequency will display a large displacement. Another fiber may be resonant close to a frequency related to a bearing ball not being exactly round. Specific fibers will pick up various manufacturing tolerance extremes and oscillate wildly if a bearing happens to have a vibration very close to the fiber resonance.

A failing bearing might easily be producing a high amplitude vibration which might cause all of the fibers to vibrate nearly to their limits. This would disrupt the normal distribution of amplitudes one observes for a plate full of only good bearings making the pattern across the fibers completely different than observed without the failing bearing installed.

I called this fiber device a cochlear implant because the cochlea is a resonator in the ear which excites hairs in certain places corresponding to the frequencies the listener is hearing. Your plate is like the resonator. The strip of fibers is an array of fibers sensitive to specific frequencies. In the ear nerves attached to the hairs relay what frequencies at what amplitudes are present to the brain. A laser beam substitutes for the nerves for this implant and a visible light pattern displays the spectral distribution in patterns of reflected laser light which would look generally like a row of ellipses of various major and minor axis sizes, major axis orientations, and speeds of axis rotation. You would lose a lot of information cramming all that data into an average just like the man reporting how he felt.

I do not recall talking before about averages but only about statistics which is far broader including such things as distributions and transforms. The fibers arguably do what amounts to a real time transform from the time domain to the frequency domain but that is even a rather simplistic view of what all is going on. Pushing this transfomation processing into materials provides a rapid response to an otherwise compute intensive process that you might have to perform if you, for example, measured the instantaneous distance to the bearing with a laser and had to extract complex periodic behavior from a single(or more) rapid stream of high precision displacement numbers.

Statistics is also the word I used to describe converting the analyzed sensor data stream into decisions about the health of bearings. This might involve population statistics, correlations and similar activities which might need to be done regardless of whether you analyzed the motion or the energy spectrum of the assembly. Characterizing the parameter ranges of a good or bad bearing might arguably be easier from vibration spectral energy distributions than from pure motion analysis but in a sense it is still a correlation run against a known population. You get into questions about how bad a bearing must get to be rejected.

thewildotter

I am in no way impugning the veracity of what you have to say - but, a quote "If you can't explain something in a few words, try fewer".Robert Breault

Hey, you're making it harder and harder...

I've often dreamed of having a device that could localize sound and display the position on a video picture.

Here is something I found that may be of some help. I have no idea what it costs...

https://www.nissan-global.com/EN/LICENSE/PDF/technology10.pdf

https://www.nissan-global.com/EN/LICENSE/TECHNOLOGY/TECHNOLOGY10/

You will have serious crosstalk and reflected signal path issues. I'm used to using fairly standard three axis quartz piezoelectric accelerometers, such as http://www.extech.com/display/?id=14797

A side issue is that bearings tend to fail on a statistical basis, unless you have one that is subjected to particularly damaging conditions. If the bearings are failing statistically, then you may be well ahead swapping out all the bearings at once as soon as you start getting failure acoustics. You might even want to swap out the plate with all the bearings mounted and then rebuild the bearing array at leisure while the working installation is running.

If the bearings are failing statistically, then you may be well ahead swapping out all the bearings at once as soon as you start getting failure acoustics.

Depending on the width of his bell curve, that could be wasteful. If it has to get to the point of being audible before it is treated then it is past the point of need for replacement and probably shows up as a defect in some process or product. In that case, the defect itself could be the indicator instead of worrying about detecting bearings. There may be other approaches to direct reading of the acoustics that are less expensive. A camera/vision system comes to mind. Each duplicate part of a multipart machine is an individual and thereby subject to error in positioning that is generally significant enough to detect with a camera. Removing that process variation using a camera can yield impressive results.

It sounds like bearings are only a small part of a larger machine. I wish we knew more about it.

I agree that knowing more about the equipment would be nice.

What I am referring to about statistical failure is the L10 life for the bearing. That is the point that 10% of the bearings are expected to fail in normal use. At that point, if the bearings are equally loaded, then replacing the failed bearing is the beginning of a cycle of more frequent replacements as the rest of the bearings hit their end of life. This is usually an accelerating cycle of more frequent breakdowns.

At the point that the first one has failed, then you are best going into preventative maintenance mode and change out all the bearings. Then you have another time period where you can expect and plan on reduced failures for the reasonable future. It's the same rationale for changing out all your brake pads in one go, rather than replacing them one at a time when they finally hit metal to metal.

Good point. In some workplace environments management does not want to suffer downtime until the machine hits the floor in pieces. Then they want it rebuilt in the time it takes to do preventative maintenance. Wishful thinking at its best. I sometimes think that certain managers would drive hundreds of miles on a flat tire before stopping to fix it. But I still have one point to make.

In the statistical world, the wear of a bearing on a single shaft may exhibit a statistical performance with respect to the replacement, but the wear of several bearings on different shafts makes the assumption that the alignment and tolerances are exactly the same in all occurances. Since we know that discription is impossible, then the best we can say is that they exhibit a behavior that resembles statistical performance and are therefore similar to the breakpads. They are individuals that each have their own issues and possible interaction which we can not know.

In other words, it is entirely to have one or more particularily bad performers that do not represent the group as a whole. It really depends on the machine which we can't see or fully understand. It could be that some early failures lie outside the bell curve of normal statistics (a.k.a. outliers). We would just have to have more information to trust the L10 point as being meaningful in this case. Using the general rule of thumb that 80% of your problems come from 20% of the aggrigate, I'd go with the L20 point rather than the L10 point. In most cases, the best practice is to remove a poor performer as soon as it is detected. For that, a quick change procedure is useful.

The Value of Dramatic Visualization

And dramatic visualization is more compelling to many managers than logical statistical arguments.

thewildotter

This is a very expensive and ineffective maintenance practice. L10 life being the number of hours run by which point 10% of the bearings have failed. Are you suggesting that he change out all the bearings after the first failure or after 10% have failed. The other 90% may well continue to run happily for many years instead of the new bearings which again will have 10% failures.

Condition based as per the OP's question is the way to go, though I do not have any insight into the OP's specific question.

It sounds like you have an ongoing problem with repeated random failures in the bearings affixed to the plate. Have you maintenance logged the failure history of the particular bearing locations to find one particular bad acting location?

Acoustics in complex geometry is a tough nut to crack. Usually a failing bearing develops a noise signature, but also develops a heat signature. Have you looked into an infrared camera to look for development of a hot spot (could be as little as .3 degrees elevated temperature depending on the camera).

I was involved in an acoustic study to locate a buzz in a nuclear reactor core. Multiple sensors were used in any number of places. We went so far as to insert a tool through gauge and sensor conduits to tap in various locations to calibrate the sensors and software, but after three years came up dry. I would suggest that if you must monitor the bearings continuously that you use Lyn's contact microphone suggestion on each bearing. If you can monitor on a periodic basis, there are bearing vibration meters that have a probe shaft that you use to touch each bearing housing to get individual readings. I suppose you could use a triangulation system, but the study would need to start with waiting for a known failure and catalog the signatures for future reference.

After due consideration of all suggestions,I guess the best way is still using an ultrasonic scanner by maintenance personnel.

I guess when the IOT is fully implemented,all bearings will have a sensor embedded and be able to generate an alarm when failure is eminent.

MEM's are becoming smaller and cheaper,so that is the logical next step for mission-critical applications.

A MEMS could be embedded in the shield of the bearing,using RFID for communication.

What better place to monitor vibration and noise?

I'll just have to wait and see.

Thanks all,for the valuable time and suggestions on this matter.