Neural network controller for internal combustion engine

[Jaco]

I'm currently doing a project on neural network controller for an internal combustion engine to reduce emissions level. Can anyone assist me in training such models (neural networks)

 

[Mech_Engineer]

What's the advantage of using a neural network in this situation? It seems to me a standard PID loop is perfectly sufficient...

 

[Phrak]

What's the advantage of using a neural network in this situation? It seems to me a standard PID loop is perfectly sufficient...

Perhaps a neural network can be trained to optimize PID gains.

 

[russ_watters]

Most halfway decent PID controllers are self-tuing. Your car already does this.

 

[Cyrus]

Most halfway decent PID controllers are self-tuing. Your car already does this.

Cars are self-tuning, how do they do this?

 

[russ_watters]

Cars are self-tuning, how do they do this?

For starters, by keeping track of operating parameters and matching the next one to the previously logged parameters. How much further it goes beyond that, I'm not sure, but a home thermostat does similar things. It keeps track of overshoot and adjusts the timings to compensate.

Digital thermostats with PID controller
More expensive models have a built-in PID controller, so that the thermostat knows ahead how the system will react to its commands. For instance, setting it up that temperature in the morning at 7am should be 21 degrees, makes sure that at that time the temperature will be 21 degrees (a conventional thermostat would just start working at that time). The PID controller decides at what time the system should be activated in order to reach the desired temperature at the desired time. It knows this by remembering the past behavior of the room, and the current temperature of the room.

It also makes sure that the temperature is very stable (for instance, by reducing overshoots at the end of the heating cycle) so that the comfort level is increased.

http://en.wikipedia.org/wiki/Programmable_thermostat

I'm not sure exactly how much a car's computer does, but if you've ever run the battery completely dead on your car, you may notice that it has trouble idling for a while until it figures out how to properly control the fuel/air mixture.

 

[Cyrus]

Mmmm...I don't believe this explanation. A PID controller has a set of gains which are determined based on a linear model of the system.

All you have described is how the PID controller works, not how the controller figures out which gains to use. Perhaps they are doing gain scheduling where the map a series of gains based on the closest conditions of the system at the time, but none of these explanations talk about how it (if at all) 'figures out' what gains to use.

 

[D H]

I hope you guys do know that you aren't helping the OP one iota. And you haven't rolled out the welcome mat, either! Bad form!

That said, this is not a good opening post:

I'm currently doing a project on neural network controller for an internal combustion engine to reduce emissions level. Can anyone assist me in training such models (neural networks)

So, first off, rolling out the official welcome mat: Welcome to PhysicsForums, Jaco!

Now I'm going to pick on you a bit, Jaco. Don't take it too personally. Hopefully this will help other newbies learn to phrase things a bit better (and you too, Jaco; this can be a very helpful place).

We do have some written rules here; you violated an unwritten one. A very big one, too. The rule is don't ask us to write a book. That is precisely what you have done by asking us "Can anyone assist me in training such models (neural networks)". The question is too open ended, too vague. The answer would require us to write a book on neural nets. There are in fact many books whose main topic is this very question.

So, Jaco, if you are still around, I'll help you in terms of questions. First question, very important:

What research have you done on this topic? We are a lot more willing to help students who have shown some work of their own.

Now some more questions that you should investigate during your research:

• Why would someone want to use neural nets when conventional control theory does a bang-up job for the most part?
• Where does conventional control theory start to unravel?
• Are neural nets already in use in the auto industry? By whom?