seattle951

All,

I just took a job as a consulting practice manager with a Seattle-based analytics firm. We develop machine learning solutions on Microsoft Azure ML. I have a MS in Predictive Analytics but need some hands-on experience with Azure ML. My MS program was based on SAS, SPSS, R, Weka and Python.

Azure ML is a cloud-based service from Microsoft that allows developers to build and publish predictive analytics models that can be marketed trough an on-line store.

In order to improve my skills, I am considering building and publishing a model that would be used to diagnose 951 problems. Porsche 951 enthusiasts could use this web-based application to identify potential causes of system failures. The model would be dynamic and would re-train as more data becomes available. This would be a free application. Users may be asked to provide data about their cars, their project objectives and product preferences. This data could be analyzed (Conjoint Analysis) and provided to the aftermarket vendors.

The challenge in building such an application would be generating enough data to train machine learning algorithms. Rennlist has huge amounts of unstructured data that could be used with some difficulty. The project would be much easier with structured data.

If I were to launch this project, would Rennlist users be willing to document their troubleshooting experiences in a web-based template? The template would gather symptoms (independent variables), configurations (independent variables) and outcomes (dependent variables). Machine learning algorithms such as random forests, neural networks and support vector machines would then attempt to model the system.

At this point, I am in brainstorming mode. As mentioned, data collection will be the challenge and would require cooperation of the 951 community over a period of a year.

Note: No Rennllist data will be used without permission.

What do you think?

Tom M'Guinn

I'm in. Would be great for new-comers especially...

Dare

Sounds interesting. I'd contribute.

SummitP

I'd contribute.

Van

I'd try and help you out.

George D

Neat idea. The only issue, in my opinion, is consistent data from exact same sampling showing relevant data. Most of us have modified the original 944 Turbo. Some taking the original engineering towards potential limits, and others still using the first MAF/Chips from Autothority. Google that

I truly like your idea, and would support you with relevant data, but these cars have too many different setups towards a software design that's relevant today worthy of your time/effort.

Kindest Regards,

G

Willard Bridgham 3

Many use Clark's Garage forum for troubleshooting and this should help you as well.

seattle951

Modifications would be built into the model as independent variables. The challenge would be getting enough data to support the number of variables involved.

We could limit the scope of the project to reduce the number of variables. This would not be an "all up" troubleshooting tool but instead could focus on a specific problem such as the engine cutting out, misfiring or rough idle. If the project is successful, we could expand into other areas.

My team has many car enthusiasts (mostly BMW) and would likely join in on this project in their spare time.

To make this project viable, we would need a way to capture large amounts of data. This would require significant support from the 951 community.

seattle951

Clarks Garage could provide more unstructured data. This data cold be leveraged, but I am not sure I am willing to invest the time required.

bumflick

I like the idea, and I'm willing to contribute. I concede the modifications to many of the cars will create some level of variance, but before we all kill the idea, may I propose a trail with one specific issue? I nominate 'no start'

Van

Can you give us some examples? I assume you're thinking like:
-No start
-is engine turning over?
-No
-voltage at battery?
-Yes
-voltage at ignition?
-Yes
-voltage at starter when key is turned?
-Starter bad
-Yes
-is the engine getting fuel
-No
-is there fuel pressure at the fuel rail?
-No
-is there voltage at the fuel pump?
-yes
-fuel pump bad
-Yes
-check injectors or DME
Etc
Etc

seattle951

You are not too far off. The user would select a problem from a pull-down menu using a web interface. The system would then prompt the user with questions that would be answered with dropbox's or buttons. The variable list would be dynamic and self generate as the user provides inputs. Each input would map to a coefficient in a logistical regression model. The system might use regression or some other algorithm based on performance.

After the process is complete, the tool would suggest a troubleshooting methodology starting with the most likely cause of the failure.

With small data sets, the human brain can do the same thing. However, when hundreds or thousands of data sets are involved, computer algorithms are needed.

Van

Does there need to be some feedback loop for the system to "learn"? Won't the user have to come back and say if the recommended fix was the actual fix required?

seattle951

In its most rudimentary form, machine learning systems use a portion of a data set for training (build the model) and for testing (evaluate the model). If the user provides the outcome, then the data is added to the database and reused the next time a model is built (training). If the outcome is not recorded, the data is not reused. In order for the system to improve itself, outcomes will need to be recorded.

ehall

Cool idea.