Only slightly kidding. I learned at a young age the sounds of unhappy machines. This skill has served me well in my engineering career. The "townies" with degrees from fancier schools than mine are often much slower on reaching for e-stop.
This project has huge economic value. Done right it is a game changer.
Same here, though I am not that enthused at the rush to get machines to do it rather than passing skills on to people. I don't see what all the formerly-employed people are going to do as their skillsets are automated away.
2) We're going to have to figure this out as a society unless we like the idea of Mad Max: Interstate 80. The incoming administration may or may not be ideal for the moment in time which we find ourselves :-/ but eventually this comes to a breaking point.
Doctors and lawyers listen to things for money, too...
I wonder though about getting enough training data - electric motors fail fairly rarely - MTBF would often be in the tens of thousands of hours (or better), I'd expect. Though possibly a hybrid approach where you record the sound of any motor that has a service call could work, I suppose. IC engines fail quite a bit faster, so that shouldn't be as hard.
That's what I'm thinking. Don't even bother to factory train the AI. Just let it learn what the machine sounds like new. After a few weeks of training data, any deviation is suspicious.
My dad was gifted with cars, and got into programming with fuel injectors back in the day. He worked as a diagnostic mechanic for a lot of years with lots of funny stories (like roping himself to a car roof with the hood taken off), but he'll sit in your car and tell you a wheel bearing is going. Jump him around the seats and he'll isolate it for you, and even today he gets it on the margin before a dealerships mechanic would be able to spot it.
IIRC a 747 produces 30TB of data per hour from its engine monitoring.
It won't be long until it's in all cars, and then in expensive appliances. A washing machine that detects an unbalanced load to save the motor?
It definitely is a game changer, and I don't think we'll fully comprehend the economic value. Say a stove that can detect a pizza box smoldering on it and cut its own power. What's the economic value of saving a house from burning down? From saving a washing machine from self destructing, or a front loader from flooding your house because a flow monitor says it put more water in than should have triggered the float valve, etc.
It's not just industrial that will benefit, it'll be everything we put moving parts in.
That would be wonderful and I hope you're right. However, most manufacturer don't benefit if you but a new washing machine every 30 years instead of every 10.
In 2000/2001 I had an office next to a couple guys that were doing failure prediction on electric motors using signal analysis of accelerometers.
The next big leap is the integration of cheap wireless force sensors in objects themselves. This will provide realtime design feedback in the field, so not just early detection of failures but root cause analysis of the design flaws themselves.
Depending on context, it also turns every object into a potential bug / movement sensor.
I think this field is pretty interesting, especially for buildings where fault finding and maintenance can be very expensive. Things like wireless strain guages in reinforced concrete, embedded temperature and humidity sensors in walls, the future does seem likely to have cheap sensors absolutely everywhere.
> The next big leap is the integration of cheap wireless force sensors in objects themselves. This will provide realtime design feedback in the field, so not just early detection of failures but root cause analysis of the design flaws themselves.
I'm aware of at least one team working towards a similar goal[1]...although cost effective (as opposed to cheap) may be more appropriate to describe it.
Yeah, I have been thinking of ways to build passive circuits into structures so they can be actively scanned by light, rf, etc. Modulate the return pulse based on the value being measured. Measuring internal forces on a parts of a crankshaft, piston connection rod, fan blade, etc would be amazing. Could actively derate a machine based on internal part fatigue. Now if we could have internal sensors using TDR we could spot cracks as they form.
I did a market research project for Honeywell back around 1997 on the use of predictive maintenance for industrial and office tower HVAC systems. I was tasked with doing focus groups and interviews with persons in charge of factories and office tower facility managers to understand the economic impact of HVAC downtime. Even at that time some factories that went out of spec on temperature and humidity could cause production lines to be shut down and each hour of downtime cost hundreds of thousands of dollars in lost productivity and costs in shutting down and restarting process manufacturing equipment.
The connection to this story is that one of the techniques they had developed was vibration and audio analysis of large electric motors and other equipment I believe to determine when a motor or part was starting to go out of spec.
I sometimes wonder about systematically mining R&D ideas that where leading edge but discarded as too expensive or impractical for one reason or another from 15 to 20 years ago.
That is the classic industrial ML advice, isn't it? Read a bunch of papers, look at the stupid obsolete method they all compare against, implement the latter.
Acoustic vibration can have surprising effects. Brendon Gregg did a demonstration of it affecting storage array performance. He literally yells at a disk array and shows measured latency spikes.
I would literally pay to subscribe to a comments feed on posts like this. It's so obvious and so powerful, I'm sitting here kicking myself for not thinking of it first.
Nearly every time I've put in a successful patent, that last bit ("god damn it! Why didn't I think of that?") comes up. The other times are derived from clinical trials in rare populations (whereupon other trialists say "god damn it! Why didn't I think of measuring that?").
The idea being, look for simple ideas like this, jump into the market second, and split the market for a while. Then when it gets too crowded, jump to the next-best-idea in the queue. Could be very profitable in this big ML transition.
I'm not sure how that follows. Folks have been using statistical machine learning techniques to analyze machine sounds for at least a decade in my experience, and probably since the 80s according to other comments here. Therefore, if you subscribed to a newsletter of articles like this one and tried to launch competitor companies, you wouldn't be second to market, but yet another in a crowded field.
Possibly. I was thinking along the lines of "look at idea, see if easily, reliably, and cheaply implemented, assess market & competition, decide accordingly".
Hell, even that could probably be automated. Hmm. ;-)
I might be wrong, but Uptake seems to focus on cleaning and analyzing data from existing sensors on heavy machinery, whereas the firm in the article installs their own sensors (perhaps in addition to that).
34 comments
[ 3.7 ms ] story [ 66.5 ms ] threadOnly slightly kidding. I learned at a young age the sounds of unhappy machines. This skill has served me well in my engineering career. The "townies" with degrees from fancier schools than mine are often much slower on reaching for e-stop.
This project has huge economic value. Done right it is a game changer.
2) We're going to have to figure this out as a society unless we like the idea of Mad Max: Interstate 80. The incoming administration may or may not be ideal for the moment in time which we find ourselves :-/ but eventually this comes to a breaking point.
Doctors and lawyers listen to things for money, too...
IIRC a 747 produces 30TB of data per hour from its engine monitoring.
It won't be long until it's in all cars, and then in expensive appliances. A washing machine that detects an unbalanced load to save the motor?
It definitely is a game changer, and I don't think we'll fully comprehend the economic value. Say a stove that can detect a pizza box smoldering on it and cut its own power. What's the economic value of saving a house from burning down? From saving a washing machine from self destructing, or a front loader from flooding your house because a flow monitor says it put more water in than should have triggered the float valve, etc.
It's not just industrial that will benefit, it'll be everything we put moving parts in.
The next big leap is the integration of cheap wireless force sensors in objects themselves. This will provide realtime design feedback in the field, so not just early detection of failures but root cause analysis of the design flaws themselves.
I think this field is pretty interesting, especially for buildings where fault finding and maintenance can be very expensive. Things like wireless strain guages in reinforced concrete, embedded temperature and humidity sensors in walls, the future does seem likely to have cheap sensors absolutely everywhere.
I'm aware of at least one team working towards a similar goal[1]...although cost effective (as opposed to cheap) may be more appropriate to describe it.
[1] http://www.secnav.navy.mil/innovation/Documents/2016/05/SEAM...
Yeah, I have been thinking of ways to build passive circuits into structures so they can be actively scanned by light, rf, etc. Modulate the return pulse based on the value being measured. Measuring internal forces on a parts of a crankshaft, piston connection rod, fan blade, etc would be amazing. Could actively derate a machine based on internal part fatigue. Now if we could have internal sensors using TDR we could spot cracks as they form.
The connection to this story is that one of the techniques they had developed was vibration and audio analysis of large electric motors and other equipment I believe to determine when a motor or part was starting to go out of spec.
I sometimes wonder about systematically mining R&D ideas that where leading edge but discarded as too expensive or impractical for one reason or another from 15 to 20 years ago.
That is the classic industrial ML advice, isn't it? Read a bunch of papers, look at the stupid obsolete method they all compare against, implement the latter.
https://www.youtube.com/watch?v=tDacjrSCeq4
Nearly every time I've put in a successful patent, that last bit ("god damn it! Why didn't I think of that?") comes up. The other times are derived from clinical trials in rare populations (whereupon other trialists say "god damn it! Why didn't I think of measuring that?").
The idea being, look for simple ideas like this, jump into the market second, and split the market for a while. Then when it gets too crowded, jump to the next-best-idea in the queue. Could be very profitable in this big ML transition.
For every "why didn't I think of that?", there's probably a bunch of companies already doing that.
Hell, even that could probably be automated. Hmm. ;-)