Quote:
Originally Posted by arctictraveller
I spent a large part of my life diagnosing and repairing CNC machinery and controls, and a lot of what I saw in the video was beyond my current understanding, but it was still valuable to me.
|
Your sentiment is echoing my point. It may be difficult to figure out what he is doing by just watching him. I'm not saying it is not valuable, just that unless you are very technical you may not understand the formal principles involved and that will make the video harder to understand. Now maybe there is a "Trouble Shooting theory" video which he has listed. My guess is there is none else he would have referred to something like that as the video progressed.
Quote:
Originally Posted by arctictraveller
I do find you comment "He probably doesn't really even understand in a formal sense what he is actually doing." rather interesting. He has a better understanding than most professional mechanics.
|
This is the only video I watched and I only watched 10 minutes
https://www.automotivetestsolutions.com/all-ats-videos
For reference, I have an advanced degree in systems engineering and have worked as a contractor directly for big Army (LIA) on defining data analysis to optimize diagnostics and prognostics methods for Large military vehicles.
The objective was to optimize the Logistics Key Performance Parameters (KPP). Part of the dependencies is minimizing downtime which maximizes uptime or more specifically Operational Availability (a top-level KPP).
As a specific task to achieve this, I needed to find a theoretical performance framework for assessing all potential solutions whether they involved directly changing the vehicles (material solution) or doing some type of back end (non-material solution) data analysis to the data collection. Part of the model involves identifying the factors in optimal diagnostics and fault isolation performance as that directly affects downtime.
So when I saw what he is doing I immediately recognize that in this formal sense and start building a model of the process he was following. The AST mechanic has probably never seen a formal description of optimal fault isolation (under a Bayesian decision theory) and may or may not even understand it if explained.
I did not watch the full video but from what I saw he is looking for combinations of data that reflect a departure (deviation from normal) from an internalized dynamic model of the engine. This means unexpected combinations of data (as opposed to a single voltage measurement for example). Sometimes it is combinations of measurements that deviate from an expected behavior from dynamic probing input to the engine (throttle under load for example to see how AFR adjusts) set of conditions. This has to do with making a fault observable. But again that observability is predicated on what he expected to see and that is based on what I am calling his internalized model of the engine.
I don't think he has a block diagram of this model nor could he articulate well what it is although it is burned into his brain from years of experience.
Here is a similar example, in the early 1990's I developed autonomous guidance systems for parachute cargo delivery. At the time the main objective was to achieve trained parachutist levels of performance (e.g. 50 jumps of experience). Despite there being several expert jumpers (over 1000+ jumps) available to ask none of them could draw a diagram of the process they used to get to a specific target location with great accuracy.
Years later we found out that the standard approach training techniques that were taught were part of an optimal strategy (using a dynamic programming simulation). So in this case the common approach turned out to be an optimal approach but yet you could not get a good explanation of the overall target approach and much less about its optimality.
If you know anything about early artificial intelligence it was mostly dominated by what is called "expert systems". An expert system is a computer program that has had rules of thumb from experts digested into a set of rules that can be input into a computer program. The expert systems then presumably take the composite information from many experts and have a program behavior that represents the entirety of all the expert's rules of thumb. In this case, there is no model but is an example where expert knowledge has to be extracted from the experts in interviews and then codified into expert rules in the program.
Now we have AI learning systems where you skip the experts and go right to the data for training of the AI.