Aug 7 this year will make 38 for me. I've used HP, Ono Sokki, Nicholet, Bently, et... and have software of Entek for SKF and the whole bunch plus my own software.
Recently I was in GA for a client doing PdM surveys using RTA's and on-site analysis and doing the trending thing. Finished the formal report, listed and labeled problems and provided a maintenance work scope and time line for scheduling. I had just gotten a ME-42 (machine evaluator w/artificial intelligence based on ISO 10816-1). I didn't use the ME-42 for this but took it into the field after the formal report and ran it through its paces. The 'evaluation mode' found all the faults and listed them in plain english spot on to my surprise and back to the computer and downloaded in seconds. I bought the thing and think every field engineer that has anything to do with rotating equipment needs one. If you own a calculator you can justify one of these. I'm hoping it gets expanded and built upon for PdM.
But yes; I'm kinda with you, I don't have confidence in the CSI at all or not at this point.
Sam, it's good to know that there's some software out there that does work for accurately diagnosing machine problems. In time, perhaps I'll learn to trust it. One more paradigm to overcome. Sam, you're fortunate to have the opportunity to work with various analysis equipment. I've used CSI excusively since I started 10 years ago. For the most part, I like CSI, but it took quite a while to get comfortable with it. There's still a lot about RBMware I need to learn.
I don't use nspector either for the very same reasons. I just don't trust either the programming or my ability to input all the data exactly as required to get accurate answers.
I really like the way the Rex Omega spacer couplings look when the bolts come out of one half of one side of the coupling. The stamped steel piece that is molded into the orange element is quite a sight when spinning at 1790 rpm (viewed from a distance).
Maybe your subtle humor is too subtle for some. I get it.
Thing with software is that it looks at ALL data, same way year out and in, all days of the week, humans donÂ´t do that.
And enter true data otherwise you get shit in and shit out rule activated.
So as long as you feed the silicon mind with all the info you have to your best knowledge and use it as screening tool for large amount of data to indicate what machines humans should concentrate on, it is doing itÂ´s thing, indicating what to concentrate on as "decision support tool" not a decision maker as you try to use it. Silicon can just as in implants only help to enhance and concentrate your view to what it think is essential, pun intended, at least so far.
I used also SKF, DLI and lately the Russian Dream out of S:t Petersburg, I like the latest best so far it makes sure I donÂ´t miss anything. It has happened we both missed in special cases but then my ears also did that so the "Expert" is only silicon and I am only human, together we make a hell of a team and he is more persistant than I ever be. Also I like the one I made myself but that should be obvious :-)Off the soapbox again. Olov
I have had similar experiences with nspectr, it takes a lot of time to set up and produces questionable results.
It has been my experience that inexperienced analysts that rely on it will make a lot of bad calls, and experienced analysts don't trust it.
My regular customers include: steel mills, peanut butter plant, snack bag plant, lime plant, hardwood flooring mill, rubber roofing membrane plant, trace metal recovery plant, convention center, and a wire mill.... and at one time or other I've done plants that make fish food, aluminum beverage cans, aluminum foil, MDF, and the list goes on.
I don't have a single plant that makes power, petrochemicals, plastics, or paper.... things we usually think of when we think PdM. My point is, there is a HUGE variety of equipment out there, and unless you are in a narrowly defined field (plant), there is no way an expert system is going to be efficient at diagnosing problems on such a diverse population of equipment.
Besides, as most "seasoned" analysts here will tell you, when you walk away from the machine, you're usually about 80% done with the analysis anyway.
Case in point. Up on a rooftop today, and I could feel the vibration from a big baghouse fan as soon as I hit the deck. Nearly 3 in/sec at 1x.... was shaking the world. Obviously (very obviously) imbalance... Nspectre and everyone on this board would know that right away. A quick visual revealed everything was tight, no cracked welds.... but the rubber expansion joint on the inlet of the fan was missing. Poked my little pocket strobe through the opening, tuned it, and sure enough... there is was, wrapped 1/2 way around the inside of the wheel. Snapped a few pics with my Canon A400 and had a picture of it to take down and show the maintenance super.
There are some things that software is never going to be able to do.... namely, tell you how to fix something.
(Danny, if Ludeca ever gets rid of that stupid smiley face, I'll think about buying a Rotalign. The only thing that could be worse was if 'Barney' popped up when you're in spec... or maybe a flower... a kitten... a fluffy, white cloud, perhaps?)
I agree with most of the comments too. However, there are somethings that I have learned to consider as I have used P4Pro, Checkmate, tried to use nspector, and have been involved in building 3 automated diagnostic software systems using specialized neural nets and boolean rules and some other methods. Currently working on 2 more systems as well.
Most current rules or inference engine systems do not work well. Reason being the rules are are of the canned nature and the sofware developer did not have the foresight to allow the end user to develop his own. Sure some things are common and could be relied on somewhat, but we all know that it is just not that simple. My old friend John Mitchell and I have belabored the issue. John say's the rules should be locked and not be able to be edited. I say that is the exact reason why these systems have not been successful. How many have complained just like here that they made wrong calls.
If one took the time to develop his own rules for a fault and it missed who does he blame?
It takes discipline and many hours to sit down and calculate the specific frequency locations for all that may be present. Then what are valid alarm levels? Some frequencies may overlap and require other data input to differentiate the faults. What about data resolution, if too course then you cannot discriminate discrete frequencies. What is the manufacturer really doing with the data processing? You know they all claim proprietary data rights. Do the applications utilize phase data and other possible operation related data?
Looks bleak, not really. It's in the methodology.
How well do you think one would work if you sat down and determined the data resolution required to capture discrete data related to all faults while considering all other faults. Set your data acquisition resolution and data types to the appropriate resolution and type. Then condition indicators are developed (narrow bands, demod spectrum with band extraction applied, cepstrum, crest factors, phase, motor current, flows, pressures, etc.) These values are then available to the user to develop his own rules and his own desired response to these faults. This data is also processed in parallel via special neural nets (same one the Navy acutally uses for automated sonar diagnostics). The neural net overcomes the pitfalls of the rules and vice versa. Voting logic could be used in a second tier neural net to compare both sub-sytem outputs prior to notification of a fault which is e-mailed to responsible individuals who view the actual data and make the final decision. The system may process data in near real time or from static databases. The rules can be modified as time goes on to refine them. The neural nets are trained as new data is collected that may modify the diagnostic response. If online systems are used then you could detect transients if data collection rates were adequate.
What you wind up with is a high speed intelligent data filter that is an enourmous tool for the analyst. They can process vast amounts of data unbelievably fast. The Neural nets can detect hidden patterns, overlaping faults are shown with a degree of probability (bayesian), will detect unknown faults and identify them as such.
Some of the major companies out there have developed similar systems that process up to 1500 data points every couple of seconds. (can't tell you who non-disclosure and confidentialty agreements) Their quality requirements were no missed faults and only one false alarm per machine week. This come out to one false alarm in 7.56 million opportunities. They met the criteria and deployed the system in 2000.
The system with the neural nets was developed in 1998 and could pull in 30-40 MB of data per second and would do all the processing above to the tune of 50 points per second. The data was analyzed manually and found that the outputs were accurate in the high 90's percentile.
This is not your Daddy's Oldmobile for darn sure, nor is it a tool for a novice.
The problem is the big boys back then had their heads in the sand. The old we didn't invent it so it can't work. Look at what they are starting to do now. They still don't understand how to do it. Also the data interface issue in gaining access to the data via a third party application. So far we have been able to gain access to Prism4, Vibrometers system, Entek Odyssey and several others.
I'll be glad to share what I can and I also have the only functional prototype of the Neural net based system. Now you know pretty much what I have been doing for the last 13 years of my life.
This technology will work but not in its present form! My $5.00 worth.
I have always been thinking of how you could effectively train those neurals. In humans we here say that regardless of background you have a 3 year apprentice time to do before you have trained your personal neurals enough, and they fortunately never finish training. So my problems is to see how you squeeze those basic 3 years training in to the silicon in a shorter time. Olov
The other method is to use a rule based system with pre-established, well designed, and extensively tested rules.
Because neural nets are learning systems which require time and data to train they are like an infant in that they are born with a certain amount of intelligence but know basically nothing until educated and trained. (By the way archived data works well in training NN's. I also developed an artificial data trainer.) I guess you could say it has a learning capacity. Like you sai it may take some time to train, but maybe not as much as you might think.
Rules based systems are knowledge systems which operate based on some rule or set of rules. The rules require one to program in known fault parameters. They are good at linear faults but are not well suited for detecting overlapping faults, hidden patterns or unrecognized faults.
This is why the both of them were utilized.
Sensor validation, data persistance and rolling averages may incorporated as well.
In the case of my company's expert system, what is required to be programmed in is the physical properties of the machine components, i. e. the physical parameters of a bearing's elements, the number of teeth in gearbox gears, the number of slots in an AC motor, etc. The expert system also weighs the effects of combination of faults on the safe operating life of the machine components.
I appreciate your quality comments, you are affortunately to have worked with Jhon S. Mitchell, one of my vibrations book was written for him, edited in 1981. For this date hi was an experienced analyzer,
John is one of a kind.
running speed in.
Rusty Cas posted:(Danny, if Ludeca ever gets rid of that stupid smiley face, I'll think about buying a Rotalign. The only thing that could be worse was if 'Barney' popped up when you're in spec... or maybe a flower... a kitten... a fluffy, white cloud, perhaps?)LOL!Smiley face gone. Ludeca has "Easy Laser" Now
What is your point of opening a 14-year old posting? Duncan Carter has passed away years ago RIP. Sam has quietly retired.
It's dementia Walt
I was just doing some research on automated analysis and ran across the thread. Sent the reply just having some fun and THEN saw the dates. I was hoping nobody would notice
It's dementia Walt
I was hoping nobody would notice
Perhaps next time click the “Take Action” button in the lower right corner followed by the “Delete” option
I did develop the ME42, still making them.