Issues and Challenges with Machine Intelligence: An example scenario

This article is intended to describe and bring some context to issues and misconceptions of machine intelligence that crop up all across industries. This article is not covering AI specifically, but many of the same issues apply across the entire 'machine intelligence' realm.

Many years ago, I was working as a Controls Engineer for a midwestern utility on an R&D project working to lower NOx emissions of coal fired power plants.

For one part of the project, we had a Neural-Network system that implemented a genetic algorithm which caused slight random ‘mutations’ to the combustion controls logic and then analyzed the plant-wide results and allowed the most beneficial ‘mutations’ to survive each day (very similar to the natural selection process, or 'survival of the fittest').

All major combustion and efficiency parameters were monitored and analyzed by an elaborate plant performance monitoring system. 

We had carefully established weighted point values for each key variable, to include things such as overall thermal efficiency, NOx, SOx, and other emissions, and other parameters that were important for financial, maintenance, or safety reasons. We also carefully set the constraints and rates of change for how much we would allow the system to mutate.

One of our first objectives was to study the NOx factors, so we weighted that parameter very heavily. Once the system was up and running, we made several observations:

At first, most of the changes were things we somewhat predicted. That progressed for some time. But then, things got really interesting...

The system began manipulating the new X-Y axis style burner nozzles we had installed in strange ways. At first, the nozzle adjustments were in ways we would have expected – but then, the system began positioning the nozzles in directions that made no sense to us from an engineering or combustion perspective. Yet our NOx numbers were dropping… And they continued dropping (to surprisingly low levels). 

After some analysis, we realized that the combustion mass & energy balance didn’t add up with the numbers we were seeing. About that same time, we discovered that the stack emissions system (which is downstream of where our NOx measurements were being taken) was seeing elevated values for NOx (along with some other undesirable changes). And suddenly, it started to become clear…

CAUSE:

The NOx numbers we were using came from an analyzer array at the discharge of the boiler section (just prior to the superheater if I recall correctly). It was an array of 3 evenly distributed analyzers/probes
with several tap points distributed vertically on each probe. Our flawed assumption was that the combustion flow would be turbulent enough to be pretty well mixed and consistent, or at least mixed enough that the averaging effect of the NOx analyzers/probes would produce a reasonably accurate overall reading.

But, the system had figured out that by making certain odd adjustments to the burner nozzles, it could lower the NOx readings (by diverting most of the NOx production AROUND the sensors).

We subsequently performed a cross-sectional NOx analysis of the duct and confirmed that the system had learned to ‘cheat’ our NOx analyzer measurements.   

THE POINT:

This was a talented team with diverse backgrounds and good leadership - yet we were so focused in on the cool technology and data, that we failed to do some of the basics such as ensuring data integrity and failed to pick up on the changes as they were happening, among other things. It was a great lesson to me and several other engineers on the project and helped us learn not to focus excessively on the spreadsheets and number crunching until we had verified the accuracy and integrity of the incoming data and made sense of the how's and why's of the results. 

In my opinion, too many (uninformed) managers just blindly trust some of these systems as black boxes, without understanding how they acquire the source data; how they process the information or make decisions or conclusions; or what the driving mechanisms are.

Machine intelligence has so much potential to change the game across nearly all industries in everything from plant automation to maintenance planning to training and development - but from my observations, it is very often misapplied or not fully utilized, or has had inadequate expertise in the implementation and is operating in 'black-box' mode. 

Based on recent advances and trends, it is safe to say that any engineer or technician in the I&E or I&C fields, would be wise to begin learning and understanding machine intelligence, smart systems, and AI implementations in automation at every opportunity.