Using Next generation AI for Industrial Process Optimization

The following conversation is with a process engineer learning exciting facts about AI, AutoML, data preparation, and the explainability of leading edge AI technologies. She/he works at a steel mill, but it could also be a paper manufacturer, an oil and gas company, or any other manufacturing company.

I believe that I could get more out of my operational data than I do today.

Great! The amount of operational data collected and stored has grown considerably in recent years. Still, the benefits of this data are so far only exploited to a fraction, which means losing money! AI helps to increase this benefit significantly.

OK, so I need AI.

Yes. More precisely, you need machine learning, which is a part of AI. Machine learning is a process that uses an algorithm to analyze data, learn from it, and make a statement or prediction. And even more precisely, you need automated machine learning or AutoML, which automates the process of applying machine learning to real-world problems.

Alright, so I need an AutoML solution.

Yes, but not any will do. Operational data is complex. It can include terabytes of data, thousands of unsynchronized signals and may have varying data quality. So you require AutoML that can cope with all that and handle the complexity of real-world industrial problems, like aivis. You need AutoML that produces excellent results despite those challenging conditions within few minutes, not hours or even days.

But I am an engineer, not a data scientist.

Don’t worry, an excellent AutoML technology like aivis does most of the work for you. It does not require any data science expertise from you, only some basic understanding of the process and the data coming from it.

OK, but what about data preparation?

With aivis, it is kept to a minimum. All you have to do is bring your raw data into the correct CSV data format – or not even that if you are using one of the aivis IIoT platform partners. In any case, you don’t have to filter your data according to your problem, and you don’t have to clean or synchronize your data. Just put everything in you have. This way, aivis has the best chance to answer your questions and even find hidden patterns.

But I don’t have a lot of data.

No problem, as long as you have even a little bit of data, you can start using aivis. And if the result is that what you were asking for cannot be answered from your data, this is also a valuable result.

But what about my process knowledge?

aivis doesn’t require any pre-knowledge for the analysis of the data. Instead, it follows a strict goal-oriented approach, where all you have to do is define your goal and start aivis. Everything else is done completely automated. In fact, at the beginning you have to use your knowledge to verify the results aivis presents to you. This will build up your trust when aivis presents insights that you haven’t known yet.

So I can understand the results?

Absolutely! aivis results are always fully explainable, whether it is a model or an insight report. This is because explainability and transparency are part of the basic functioning of aivis. In fact, aivis works oppositely to most conventional approaches to ensure just that.

Wait a minute! Does aivis know my processes better than I do?

No. aivis doesn’t know about the chemistry, physics, or statics of your processes. Instead, it gains process insights by looking at how different signals and sensors react and respond to each other, including time lags. This way, it can find and reveal fundamental and complex relations and dependencies, helping you to improve your knowledge. As the process and domain expert, you are still responsible for interpreting the results, deriving actions, and defining countermeasures.

What is the outcome of aivis?

aivis has two primary outcomes, models and reports. The outcome is a report if you want aivis to analyze your data and investigate a specific signal, quantity, or disruptive event. On the other hand, if you wish to monitor critical components or processes or create virtual sensors, the outcome is a model accompanied by a report explaining the model. The model is a lightweight, independent software module deployable on-premise, on-cloud, or on-edge to consume live operations data.