July 2021

Bonus Report: The Digital Plan

If you want AI to overcome limitations, give it more scalability

Artificial intelligence (AI) has caused a significant buzz not only within the hydrocarbon processing industry but also in everyday activities.

Brooks, M., AspenTech

Artificial intelligence (AI) has caused a significant buzz not only within the hydrocarbon processing industry but also in everyday activities. AI seems to be everywhere. It is in autonomous vehicles, it can recognize images of people and cats vs. dogs, and bank and credit card companies use it for fraud detection, among others. The application of AI is seemingly endless.

Since 2015, AI has become more pervasive and obvious as Alexa, Siri and Google Assistant help us when asked. Even AI-powered washing machines can cut laundry time, and product and services companies use AI to tell us what books we might like to read, what music we can enjoy and what movie will fit the bill for Friday night.

The essence of AI is about learning from data and using it to determine the probability of future outcomes based on the patterns learned from the past. This emulates a human characteristic. If you watch a specific area, situation or process long enough, you may develop insights. When similar things occur again, you could guess with some accuracy what is going to happen next. Other people have read the books you have and based on recognizing similar patterns, Amazon suggests you will enjoy another title just as much as others who read the same books. It sounds simple enough, but is it?

The Challenge with AI

Standard AI can recognize patterns but cannot reason why things occur the way they do without substantial help from humans. We often hear that AI sees correlations, which may be simplistic and just happen without being linked. AI cannot determine why it happens—i.e., the causation. For example, data from literature1 showed an interesting correlation: the more lemons imported to the U.S. from Mexico, the fewer U.S. highway fatalities occurred. Correlation does not necessarily mean causation—no one would seriously believe there is cause-and-effect in the referenced example; however, the problem is that AI cannot tell the difference. It turns out that humans are better at judgment and have the ability—unlike AI—to assess whether the observed correlations make sense. Consequently, despite such challenges for AI technology within industrial manufacturing, there appears to be great potential and opportunities if AI applications are tempered appropriately to overcome the limitations.

AI can create more value than is currently extracted from industrial data, especially since most data is used for viewing trend charts for scarce insight development. Intelligent data gathering and grouping with AI-driven analytics can greatly assist in helping people make more accurate judgments and better decisions. The overarching goal of AI is to uncover new and better business processes that promise significant, positive business outcomes across key areas, such as manufacturing operations profitability, safety, environmental impact and corporate social responsibility.

To be effective, AI cannot be just a technology thrown at problems. We can refer to the adage “people, process and technology” to gain solace. We believe that the importance of incorporating domain knowledge is seriously overlooked by startup companies and service providers who are trying to enter the space. Industrial AI is the realization that scalability of any solution is the most important factor.

The dimensions of industrial AI

Industrial AI incorporates several important dimensions. These include the following:

  • Many people must be able to use AI. To be pervasive, it cannot just belong to data scientists. Plant personnel must be able to build and deploy AI solutions based on what they know now. This is very similar to the iPhone, which contains deep technology, but the user does not need to know it since an elegant screen and a finger as a pointer complete the task. Consequently, with industrial AI, the competence level of workers is elevated and gravitates to making judgements on potential outcomes rather than doing the continuous and repeated heavy-lifting of deep analytics.
  • Industrial AI solutions fit current work processes well without huge departures.
  • Judgement from workers’ domain knowledge is imparted as the industrial AI solution building process begins. This is where domain knowledge from workers provides guardrails around an AI solution to ensure it will realize true causation rather than simple correlation. Embedding is simple and easy so that users can insert what they know (e.g., when an event occurs, this sensor shows a condition and then “this” will happen). Those are the clues that form the guiderails that influence the solution direction and the best outcomes.
  • With industrial AI, the step-by-step work process is built into the application building to make it simple and easy (e.g., booking a plane ticket or ordering a product on a website).
  • As much judgment as possible must be built into the application to assist the user even further, adding even more domain knowledge. Industrial AI does not use AI to just analyze the new data when a solution goes live. Instead, with or without user help, embedded AI helps define the solution strategy to select the right data sample times and data groupings for training and testing. This helps discover hidden and potential unknown relationships between events that occur and different data types—referred to as “feature selection” by the data science community.
  • Industrial AI uses a combination of AI analysis and human judgement in the beginning to determine which data is valuable and which must be discarded. Most AI detects anomalies, so getting the right baseline is extremely important for accurate detection of abnormal conditions and to eliminate false positives.
  • Industrial AI ensures that it is simple, easy and rapid to cover different types of equipment. This includes not only big machines that spin but all equipment (e.g., rotating, static, mobile, process and mechanical). Scalability also guarantees that what is learned as normal behavior and explicit failures on one machine is spread rapidly without intense engineering effort—from one asset to many that are the same (e.g., providing all boiler feedwater pumps the same safety and breakdown protection).
  • Application sustainability ensures industrial AI automatically adapts—with minimum effort—to changing process conditions, keeping the solution updated and on point without re-engaging engineering experts and data scientists.

Takeaway

Only with scalability capabilities can industrial AI deliver more accurate performance to enable better judgement and faster decision-making. The author’s company recognizes that industry needs guardrails around AI-embedded products to ensure causation is found and not just simple correlation. Those combinations mean that the products provide assertive insights immediately within the alert messages to enable rapid human judgement and remove post-alert extended analysis. Lastly, industrial AI products uncover new business processes, simultaneously evaluating thousands of different scenarios to uncover optimal processes that tie to specific business goals. HP

LITERATURE CITED

  1. “Distinguishing between correlation and causation: A key to critical thinking,” Oregon State University, February 2014, online:  https://blogs.oregonstate.edu/econ439/2014/02/03/distinguishing-correlation-causation-key-critical-thinking/

The Author

Related Articles

From the Archive

Comments

Comments

{{ error }}
{{ comment.comment.Name }} • {{ comment.timeAgo }}
{{ comment.comment.Text }}