[kubark42]

Very interesting comments. Here are my responses:

Originally Posted by jfitzpat

I commend you on the effort. It is an interesting area of study and I hope you continue. However, I do have some comments/concerns about the paper.

First, and minor, reference [3]. It is very doubtful that you are achieving +/- .1% flow measurements from the electrical pulse width controlling the injectors. When commercial automotive fuel systems are tested gravimetrically, the variation is already larger.

For instance in http://ecomodder.com/forum/showthrea...tml#post152293, it seems that that is the accuracy that people on this forum are getting.

Furthermore, Luxembourg is home to Delphi’s fuel injection research center (although they might have others). I spoke with one of my colleagues who researches direct gasoline injections, and he felt that this level of repeatability sounds reasonable.

That being said, I don’t believe everything I read on the internet, and my colleague couldn’t give me another citation, so if you have some additional information on this, could you send me the reference?

Also, it is a mistake to assume that this method is readily applicable to modern fuel efficient designs. Consider, some systems intentionally alter injector opening and closing times conditionally, others will leave the injectors on some cyls closed while still cycling others.

Agreed. For instance, some cars can switch between sequential and grouped firing patterns depending on efficiency vs. power requirements. However, this is a trivial problem, as by simply adding a voltage sensor to each injector, instead of only one, the problem is solved.

A bigger problem is what to do with diesels. Most cars sold over here are diesels, and this technique does NOT work for them.

Second, I have some real concerns about your basic mathematical model. My prediction would be that it would generate a fairly inaccurate instrument of engine efficiency, and this seems to actually match your own data.

Consider figure 7, efficiency calculated from dyno readings. It does rougly match the expected V/P curve for a conventional combustion engine. And also, as expected, you have an 'efficiency island' near peak torque. However, your efficiency island is elongated from typical engine efficiency models.

Note, Fig. 7 is a 4th order curve fitting, not calculated efficiency. The curve fitting will have a tendency to reinforce this island effect. It’s somewhat worse on the 2nd order model, which is what we’re actually using now because it seems to provide acceptable results for the optimal control problem.

Mathmatically, this deviation in the curve should be exaggerated by under sampling and an EKF techique, and that is what we see in your data. Consider figure 6. You point out that peak matches, but that is the lowest possible hanging fruit, like finding lambda 1.0 on a UEGO sensor. The question is, what is the typical error as we move away from the easily identifiable set point?

Hmm… my initial thought is that the fact that a 17-dimension non-linear observer with asynchronous outputs (data measurements) converges to a correct value is not low-hanging fruit. In future work, it can be refined, augmented, etc… in order to get fruit higher up the tree, but, of course, on principle you are right, what good is this model if it’s telling us what we already know? I’ll try to answer that at the end.

Based on your included data, pretty high. Notice how much larger and elongated the efficiency island is on figure 6. The expected V/P slope is not even really visible on the chart. If we did a scatter plot of point to point deviation between the two it would appear that, aside from peak, the instrument would almost always over estimate efficiency, and in a non-trival operational envelope, extremely so.

I’m going to show my ignorance here. The people I work with do observer and optimal control research, and I am a controls engineer (See e. busvelle - Google Scholar , jp gauthier - Google Scholar , u boscain - Google Scholar), so we’re not as strong as we'd like to be in car industry experience. What is the expected V/P slope, and keeping in mind that we’re looking at total system efficiency, where should we expect to see it?

My third concern is that the paper never really establishes a link between the acheived measurement and the stated goal. Look at part of your first paragraph:

"optimal control to make great reductions in emissions and fuel consumption"

Now, we could debate rather or not your measurement of engine efficiency is accurate enough to be of value. But the more fundamental question is, is the data even really connected to either fuel consumption or emissions?

Someone once told me, “When you can say it better than anyone else, write it. When you can’t, cite it.” I think citations [6,11,14] answer your first question better than I could.

As to the second, I’m not sure I follow. Are you asking if fuel consumption data is “connected to either fuel consumption or emissions?” I think I might be missing the scope of your question.

Virtually everyone in the field knows that every conventional combustion engine has a fuel economy sweet spot - wide open throttle (so there is no induced vacuum reducing VE) at peak torque.

But this is not an available decision for routine operation of a vehicle. In fact, most people here (myself included) go out of our way to avoid this point of operation, because it is terrible from both the perspective of emissions and *operational* fuel economy.

At peak torque wide open throttle, the mixture is generally quite rich, to keep CHTs and pressure down so that destructive abnormal combustion does not occur. But this also means that the vehicle is operating way out side of the cat efficiency envelope, making emissions soar.

Interesting point about the catalytic converter. Can you give me some references, and perhaps a simplified model? This is exactly the kind of thing that we can/should include in the cost function. (More on cost functions below.)

Sorry, I missed this earlier. I would contend that that is almost wholly false. The factors tend to work additively, and many act exponentially, which is why you over estimate effiency over so much of the operational curve.

I think we’re talking about different things here. You’re interested in absolute efficiency, whereas I’m interested in the shape of the efficiency curve. I going to say we’re both right, as it pertains to our particular problems.

One of the most nagging questions for me during this project is why the efficiency is overestimated. Looking at my equations, an overestimated efficiency implies either an underestimated fuel flow (going back to the +-0.1%, I do not know if this is the case for my car or not, due to a forgetful fiancée and the fact that we refill the car every two months) or an overestimated force. Judging from the dynamometer results, we are overestimating force. Still, it’s not important if the hypothesis that the overall shape is correct and that we have a bias error holds.

I'm not saying this to discourage you. Just the opposite. You are attempting to inexpensively measure a very narrow definition of effiency. What I am suggesting is that you try an experiment to demonstrate rather or not your target metric is of any practical use in either the area of fuel economy or emissions.

I’m not taking it at all as discouraging. Quite the contrary! I appreciate the fresh voice and the experience.

We have real-world trials in the pipeline, hopefully in the coming weeks, but it will be months before we can present the data. Taking your comments into account, I think a new visit to the dynamometer test bench is in order, this time to look at emissions as a function of optimal control.

Good Luck!
-jjf

A sincere thanks for your comments, and I hope you will find the time to continue the conversation. In summation, I would like to reiterate the goal of our project, and how this fits in.

The problem comes down to this: people want to drive optimally, but do not know how. Optimal control studies are performed, but they are always constrained to working hand-in-hand with manufacturers, making publishing data difficult at best. Furthermore, the data they publish is for an ideal engine in ideal test conditions. The real-world is far different, most of all because we’re not looking at the engine as a single unit, but as a part of the whole. The transmission, differential, etc… must be included, too. And that’s even before we start talking about differences in manufacturing and wear. Our observer is a method of overcoming these obstacles on a case-by-case basis without investing in hundreds of thousands or millions of dollars of equipment.

So how does it all fit together?

One of the great weakness of optimal control is defining the cost. At this moment, our cost is only expressed in terms of gasoline used. Of course we can use far less gasoline if we run in extremely lean conditions, but then there’s too much NOx production. So it would be interesting to know what else could be included in the cost. I welcome any and all suggestions, keeping in mind that for the purposes of making a complete synthesis we need to be able to idealize.

Optimal control results are heavily dependent on the solution method. One of the most common, dynamic programming, is one of the worst when it comes to accuracy. Dynamic programming is easy, you hardly need to think about the math at all. However, it takes tremendous computing costs and is completely impractical for a small microcontroller implanted in a car. NLP (non-linear programming) comes out a little better, but still needs a fair amount of processing power.

For optimal control to work in cars, we need a way to get an answer without throwing a supercomputer at the problem. By fitting the observed efficiency map to a polynomial, we are capable of finding analytic solutions, instead of being forced to find only numeric solutions.

Our goal is not to provide a bullet-proof efficiency solution, that will have to be left to people who have better access to the sensors, and better models. We hope that by publishing this observer, researchers will be able to focus on the more interesting parts of optimal control by saying, “the solution to the problem of mapping efficiency has already been demonstrated, so now let’s get on to the fun stuff.”

So in the end, once we find the analytic solution for a given polynomial, finding the best-fit constants can become more important. For the moment, it suffices to show that our approach finds a polynomial that is sufficiently good. After that, if it is really worth it, we, or others, can spend time refining the model. After all, that’s what happened with the MPGuino: it launched a method that was lacking in certain refinement, and my project was to take it a step further. I’m certain there’s far more to be done, and I thank you for the encouragement.