Automotive Neural Networks Pt. 2 - Early Usage to Modern Day
Part 2 of the Series. See Part 1 for a primer.
See Part 3 Here
The history of neural networks goes all the way back to the 1958 invention of Perceptron, by Frank Rosenblatt. We're going to skip a head a couple decades and get into early usage of various neural networks/machine learning models that automotive manufacturers experimented with, and may even utilize currently, to varying degrees of success and failure.
Neural networks really weren't even thought of in automotive usage until the early to mid 1980's as vehicle technology was really hitting the beginning of it's rapid technological advancements. Interest and experimentation with them began around the mid to late 1980's as a means to more economically and efficiently develop and manage on-board Failure Detection & Isolation (FDI). FDI was essentially a strategy that was designed to meet the up-coming OBD and later, OBD-II standards. Traditional computer modeling seemed to be more labor intensive, complex, and failure prone than originally expected. Great excitement lay around the supposed ability to take a stack of data, feed it into one of these neural networks, and have all the right answers pop out with little to no human intervention. Needless to say, things proceeded a bit differently than manufacturers expected.
One of the earliest mentions of the predictive and modeling potential of NNs in automotive usage was published in 1991 (Modeling, Fault Detection & Diagnosis of an Automotive Engine Using Artificial Neural Networks by Reza Afrashteh). The focus of that particular paper was an early version of the 3800 GM V6. Modeling of the various non-linear systems within the engine (think air flow, fuel trims, EGR valve operation, etc.) was examined to see if it beat other potential methods of engine parameter mapping and fault detection. It is understandably costly and time intensive to develop fuel/air/spark maps and the like for every possible combination of engine & environmental variables, then factor in potential failure points, their symptoms, and of course the most unpredictable of all parameters: the driver.
Early methods of fault detection included:
- Temporal Redundancy: limit and trend checking on inputs. and outputs based on some prior information about the system. This was/is the commonly used method due to ease of implementation and cost. (I interpreted this as the basic equivalency of $ Mode 6 data)
- Hardware Redundancy: Cross checking of at least 3 separate sensors to determine if a fault is indeed a fault (think of when you do a sensor drift test with a cold vehicle, you want to see if the IAT and ECT are within a degree or so of each other. Now picture having to have at least 3 sensors monitor any given parameter). Prohibitively expensive and complex as one would imagine.
- Analytical Redundancy: Some form of mathematical knowledge of the process along with inputs and outputs from that process are used to develop a model and then the difference between the output of the system and that of the model is determined. Deviations from the model flagged trouble codes
Analytical redundancy involves first flagging deviations from the model and then interpret and isolate the faults based on previous experiences of the fault, and through the use of a logical decision-making unit (sounds like a lot of our jobs doesn't it?). When performing engine diagnostics, these neural network frameworks can easily be moved from one engine type to another with little to no modification and calibration compared to starting from scratch with every engine variant.
Ford was experimenting around this time with using neural networks for misfire monitor modeling; Kenneth A. Marko of the Ford Research Group authored a paper titled Neural Network Application to Diagnostics and Control of Vehicle Control Systems. Paul Baltusis, a member here, mentioned in a post that in 2005 Ford was still experimenting with this technology on the 6.8L V10; the results were less than optimal and abandoned.
Fast forward a few years to … Daimler Chrysler and the University of Michigan co-authored at least 2 SAE papers directly related to the usage of neural networks:
- SAE … : Using Artificial Neural Networks For Representing the Air Flow Rate through a 2.4 Liter VVT Engine.
- SAE …: Cam-Phasing Optimization Using Artificial Neural Networks as Surrogate Models.
GM R&D department along with California Institute Of Technology co-authored a paper around the same time:
- Neural Networks For Engine Fault Diagnostics
The John Deere Foundation supplied partial funding to the Artificial Intelligence Research Group at Iowa State University for their paper:
- Intelligent Diagnosis Systems
Once again, the ability to rapidly and accurately model non-linear systems (of which an internal combustion engine has many), was at the forefront of their research. There are countless other instances of OEM's investigating, experimenting with, or actually releasing some variation of systems that incorporated neural networks into their design or operation. Jim Cokonis, mentioned in his post that Toyota specifically references neural network control in a 2016 Tacoma HVAC system.
This is just a brief introduction to some of the usages of neural networks leading up to the modern day, this doesn't even touch on manufacturers such as Tesla and their plans for using neural networks and deep learning. That topic was pretty well discussed here, in case you missed it.
Once again, I hope this starts some discussions, or at least leads to some research on the topic. This technology has been around for about 60 years now, and is only getting better, it surely is not disappearing anytime soon. I believe that by understanding the basics of the system, and manufacturers intentions for the systems, we are better equipped to attempt to utilize, diagnose, and repair components/systems that incorporate them. I can think of multiple instances where I would not be the least bit surprised to find out that part of a modules logic was determined by a neural network.
I would love to get some OEM representatives in here so we can really pick their brains as to ANN usage and start to leverage them for our own uses on the repair side. We will be exploring what I see some of those uses being, as well as explain how groups like this are actually on course to become invaluable to engineers designing and modeling automotive control systems in the very near future. Valuable to the point that we can finally leverage it to get some truly equal footing with the OEM's.
sae.org/publications/t… To be honest.. A lot of this is over my head, and I don't feel much like reading into this much until it starts to get a foothold (which will be soon enough). We live in very interesting times. When I am done in this business there will be nothing left of what attracted me to it in the first place.
Bill, First off thank you for that link, full paper isn't published yet but I pre-ordered. Sounds like they will be exploring some concepts I have been kicking around. I fully admit I have tenuous grasp of the concepts here but it is still interesting nonetheless. The foothold is already here, I think what is fascinating about these systems is that, at least for some purposes, this network
You go Chris! I sure don't understand this stuff either, but am intrigued never the less. Because as networks in general become more and more complex... and people "take their car to the shop to get fixed" ..... well its not their problem that its complicated, but the car needs to get fixed somehow. And how can it get fixed without some serious undertaking? So I get excited as I see complexity