[Jambo]

I work in the shipping industry overseeing planned maintenance on my company's ship.  We use various tools such as oil analysis and vibration analysis.  I'm wondering if there are any thoughts on AI providing real time analysis equipment such as diesel generators, switchboards, propulsion equipment etc?  The ultimate goal would be for AI to predict failures of equipment thus saving repair costs and time making the ship more reliable.

[Korrelan]

Welcome to the forum Jambo

I would consider the following.

The first thing required is going to be a real time comprehensive sensor suite.  I would probably start out by just monitoring the key points until I had proved the system.

Vibration sensors/ transducers attached at key points through out the ship, preferably isolated from internal noise caused by normal crew activities, walking along gangways, slamming bulk head doors, etc.  They would also need to be positioned where the vibrations from the movement of water out side the hull was at a minimum.

All the major mechanical points of failure would need to be monitored, prop shaft bearings, gearboxes, valve covers, etc. 

It would also be helpful to have a digital map of throttle settings; engine vs shaft rpm’s, rudder angle, etc. If a real time analogue/ digital output/ diagnostic ports are not available for the ships controls then high resolution CCTV cameras can be used to visually read analogue gauges and controls.  The more data you can accumulate the better.

All the data will then need to be filtered and run through a Fourier transform suite to isolate the various frequencies the ship produces under varying loads/ conditions. This would help remove resonance/ harmonic signals and accumulate all the data into one stream/ format ready for analysis. 

I would design the system to produce a literal image/ picture of the data stream that could be recorded for later analysis/ comparison if required.  This would also provide the data buffer required by the system to detect frequency changes over time, regular periodic knocks, etc.  The recording buffer would also help alleviate a lack of computing power available on board the ship (I presume laptops) for real time analysis.



A single sensor data trace over time would look similar to this, the unfiltered regular knock at 1.8 Khz would be obvious to both a machine and a human, especially if its run through an intelligent Fourier filter (below).  The sensors normalized profile taken over time can be seen by the trace on the lower right.  Any signal that doesn't match the learned profile breaks through and is noted.

Once you have the data stream you can either apply a standard threshold analysis to highlight differences in frames or use a neural network to learn the regular patterns and highlight differences. 

Neural networks are usually designed to find regular patterns; you just need one that highlights exceptions… to let you know if it’s not recognising the data stream.

 

[ranch vermin]

yes, and make the ship dock itself.

[Korrelan]

yes, and make the ship dock itself.

Now that would be a fun project lol.

I made the traces using this video...



And I must admit... I do find it strangely relaxing lol.

 

[Zero]

Actually, to predict failures, the neural net would have to be trained on actual failures datasets playback, or fake datasets of failures, so it can recognize what happens before the failure as a potential danger situation.

Nice topic.

[8pla.net]

Jambo,

Yes. This is a terrific topic. "AI providing real time analysis", Jambo mentioned.  What makes A.I. a valuable hobby is knowing that Artificial Intelligence is out there doing jobs we tend to take for granted.  Perhaps the best place to start the discussion is about an ADC (Analog to Digital Convertor)?

"The ultimate goal would be for AI to predict failures of equipment thus saving repair costs and time making the ship more reliable."  The goal for the A.I. in, "Reinforced Learning" is a reward.

_____________________________________________________________________________________________________________________________________________

[Art]

Excuse me, my sensors have picked up a fault in the AE-35 unit. It is going to fail.

Oh...nevermind...just recalling moments from the past.

BTW, welcome aboard Jambo.

@ Zero - By setting appropriate thresholds wouldn't anything Outside of the normal running noises be cause for alarm? They might not know exactly what a failure sounds like or an approaching failure due to lack of actually recorded histories.
The ship needs to have the equivalent of an Airplane's or automobile's "Black Box", but those would only be useful in gaining historical data for use in learning what happened and when as well as for future events.

Interesting topic!

[Korrelan]

@Zero

Meaningful long term accurate prediction in mechanical or natural systems is impossible, you are trying to see into the future based on current information… except death of course lol.

The engineer needs warning if the system detects anything outside the normal operational sensory patterns of the ship… the temperature of bearing 12 is rising, etc.  The sensory patterns a failing bearing could produce are limitless, lack of lubricant, damaged rollers, cracked housing or race, etc, but it can say this bearing isn’t operating as normal. 

The system will have to learn through constant experience, the natural wear of mechanical parts over time would alter the base set data, so you can’t train the system on previous examples.  Parts get replaced; every slight change to an engine for example would alter the sensory maps over all the RPM ranges.  Generators and other noisy equipment will start and stop periodically, even the ambient inside/ outside temperatures will affect the sensory maps.

It would be the AI’s job to constantly monitor the whole ship at a resolution impossible for a human to achieve… and simply warn the human when and where something’s not optimal.

If Barry is releasing a flow valve by whacking it with a hammer the chief engineer should be alerted to a disturbance in (the force lol) the ships sensors, the engineer can then tell the system to ignore Barry… but better safe than sorry.

 

ED: Yeah! What Art said... 

[Zero]

Ok I understand, but Jambo's question was:

The ultimate goal would be for AI to predict failures of equipment thus saving repair costs and time making the ship more reliable.

Is that possible in your opinion?

Predicting the future, isn't that what brains do? I mean, you throw a ball, I can catch it because I can predict its trajectory.

The natural "life" of a mechanical part (wear, replacement), and the changes it implies, would perhaps be part of what the system expect? Maybe the lack of previous failure examples in the datasets could be solved by connecting every system to a central world-wide "failure database", which could then be used as training ground?

[Korrelan]

Is that possible in your opinion?

Yes up to a point.  Mechanical systems will usually show symptoms well before a total failure occurs, a good engineer will use his/ her own senses to detect obvious symptoms… the AI would just be using signal analysis  to achieve similar results only ship wide… and ever vigilant.

Predicting the future, isn't that what brains do? I mean, you throw a ball, I can catch it because I can predict its trajectory.

That’s why I was careful to say ‘Meaningful long term accurate prediction’.

Your brain creates a set of possible futures/ predictions and only over a very short time span and very narrow domain.  We are always 80 ms behind reality; we do make predictions based on the posture of the person throwing the ball, but only one person and one ball.  Through experience we learn the posture of the person usually leads to the ball arriving at roughly X,Y location… can you tell by the way the person throws the ball what he is going to do next?

Like a computer has a base frequency measured in Ghz, the smaller the organic brain the faster its base frequency.  Dogs have a higher base frequency than humans, they get more ‘thought frames’ per second, they have faster reflexes etc, and yet they fall for the ‘fake ball throw’ every time… bless em.

The natural "life" of a mechanical part (wear, replacement), and the changes it implies, would perhaps be part of what the system expect? Maybe the lack of previous failure examples in the datasets could be solved by connecting every system to a central world-wide "failure database", which could then be used as training ground?

It might be possible to build a database of symptoms and solutions based on the engineer’s interpretations and fixes.  So basically any type of periodic disturbance can be detected as a symptom of baring failure, etc. though I’m not sure how helpful this would be. 

I suppose if all sensory maps of symptoms are logged accurately to a particular failure then a database could be created to diagnose possible failures from the given symptoms.

The logs of the symptoms and resulting repairs could certainly be used to train new engineers.

 

[Zero]

  korrelan

I think we all would love a postcard Jambo, if one day you end up using Ai in your work! There are certainly Ai things to be done in the maintenance area.

[ranch vermin]

heres the coolest thing since sliced bread in my opinion.

https://worldmodels.github.io/


With one of these, you can "sorta" get long distance predictions, it has a problem that it can only predict ahead what its already seen before.

[Art]

I tend to view most of this as having an A.I. be Reactive rather than Predictive. Meaning that predicting is a summation based on collections of historical data whereas reactive is an operation to correct an event (unexpected/undesired/-expected/normal) that has just happened.

If an industry had a number of wheeled containers in service that suddenly began having issues, they would send individuals to investigate the incidents, assess the damage if present and collect as much information as possible.

The individuals would then examine those containers and note which numbered series of these containers had these events/failures.
They would likewise inspect the other operational series in their company to compare.

One of the pieces of collected information might be the age or length of time in service of that wheeled container prior to the failure.
Another item of concern would be the part or area of the failure. Some forensic examination by a structural/mechanical specialist might be necessary.

Once all data has been collected, it can be entered into a computer to see resultant data plots of the affected containers.

If there is indeed, historical evidence that a certain series of containers failed after a given period of time, then a Predictive analysis of them could be made. This might necessitate all wheeled containers in a certain series being periodically inspected, removed from service or have appropriate repairs done to them to prevent the failure(s) from happening.

The Reactive part is doing something After one of the containers fails. This usually happens at the very beginning of a line of impending failures. It is the first domino that topples in a long line of dominos. Once we learn why being reactive isn't enough or isn't the best possible path, we can then work out being more predictive. This is usually true of many things.

Having said that, I still do not believe it would be practical or possible to place sensors on every location of every container (or item) and be able to predict when something is going to break or fail.  Historical analysis is the "Crystal Ball" of Predictive behavior.

[Jambo]

First of all, thanks guys for all the comments and positive feedback.  A lot of the above went over my head but I think I grasp the general idea.

A bit more information on the ship's I operate.  We have online oil and vibration condition monitoring of our propulsion equipment and engine / alternator sets, we also have monitoring systems, Kongsberg K- Chief in most cases which constantly monitors the running parameters of all the critical equipment.  I think we have all the information coming in already.  It is more the continuous monitoring of the K- Chief system with inputs from the online condition  monitoring as well that I think AI could assist with.

Any thoughts on web sites where I could read more about AI, I am a complete novice but really interested in the possibilities.

Thanks again,

Scott

[8pla.net]

Jambo wondered, "Any thoughts on web sites where I could read more".  Wow AiDreams! 100% FREE "thoughts", as seen in, $100 an hour engineering!  You could read more about korrelan's thoughts, "Fourier transform suite", subsequent to sampling analog signals to convert to digital signals that  "isolate the various frequencies the ship produces".

In preparation for what's next, inspiration is an important ingredient in Artificial Intelligence. Watch season 2, episode 13 of Star Trek Voyager: Prototype.  With the completion of your ADC and DSP,  to quote Star Trek actor Hugh Hodgin who played the Pralor Automated Personnel Unit (a sentient robot).  Your, "Prototype Unit 0001 is ready to accept programming" Jambo.   So, it is important to keep the fun in Artificial Intelligence, since even the easiest A.I. is never easy.

Reference: http://memory-alpha.wikia.com/wiki/Hugh_Hodgin