Fero Labs has developed software to help certain types of process manufacturing plants improve quality output economically when given a random mix of feedstock. I wrote about the company last August—A Better Way to Control Process Quality.
They sent a new press release, and I must admit that I understood almost nothing in it:
Fero Labs, the only Profitable Sustainability Platform for industrial optimization, announced the release of their ground-breaking feature ‘ExplainIt for Live Predictions’ which expands a factory’s production knowledge in real-time. This advanced feature for cross-functional teams increases trust in AI predictions by disclosing real-time text explanations about abnormal factors influencing their live production.
There were way too many marketing-type phrases in there. Worst of all was the concept of “trust in AI predictions.” So, I asked the very patient publicist. She suggested that I talk with Berk Birand, Fero Labs Co-founder and CEO. And, I did. He was most helpful.
We caught up from my last article about their ability to use the huge data sets manufacturers have accumulated over the past decade using advanced statistical methods and “white box machine learning (ML)” to help engineers optimize their plants. Make them more profitable and reduce waste (sustainability). Therefore the “Profitable Sustainability” company.
Birand took me through an example that I could understand, since I had a customer in the 90s who did this sort of process.
Imagine a plant with piles of scrap steel in a yard. They have an electric arc furnace that melts all that disparate steel that will be poured out eventually to make their final product. Given that the feedstock has high variability as to the composition of the steel, the typical plant overdesigns the process to allow for variations. This, of course, is wasteful on the surface. But if the final chemical analysis shows that the output will not make the desired tensile strength or other spec, then the waste is even higher.
What if you accumulated the data (feedstock, process, finished steel) over time built a modern AI model? Its predictions could be used to drive profits, reduce waste, save time. But, would anyone trust yet another advanced process control system? We all know that models eventually goes out of whack sometimes and sometimes gets the wrong answer.
Here comes the “trust” part of the trust in AI model. They built an explainable model from the beginning. It can predict characteristics, say tensile strength of the mix because of chromium or carbon levels and so forth. Since we know that every model is wrong sometimes, they built in confidence levels in the prediction engine. Their AI looks at the material composition and suggests adding chemicals to the mix, but it gives an explanation and a confidence level. The engineer looks at the confidence report (I am confident in this prediction or I’m not confident in this prediction) and can decide whether to go with the AI or to go with gut feel based on years of experience.
He convinced me. Fero Labs has developed an AI engine that gives the engineer a level of trust in the prediction.
More explanation from the press release:
Expanding on Fero Labs’ white-box ML, which provides full transparency of Fero’s powerful machine learning models, the new ExplainIt feature provides a contextual explanation of anomalous factors involved in each live production optimization.
This type of analysis is typically addressed through linear Root Cause Analysis (RCA) tools. Unlike traditional methods, Fero Labs’ solution is non-linear, much like process operations, and delivers results in seconds rather than the hours or days typically needed. Traditional methods generally require the engineer to preselect a small sample of factors to investigate, which can introduce potentially misleading biases. Fero Labs’ software has the power to evaluate all relevant factors which improves insight and prediction accuracy.