Realizing Value from Artificial Intelligence and Machine Learning

Manufacturers see the potential for big payoffs, but attaining them requires a disciplined approach. 

Although the concepts of artificial intelligence and machine learning have been intriguing executives for decades, manufacturing has seen tangible impact only in the last 10 or so years. As with any emerging technology, there has been considerable excitement around the potential from these advanced capabilities.

But manufacturers have found the path leading to benefits paved with complexity and obstacles. That’s not to say the payoff isn’t worth it; current use cases prove there are benefits in the form of cost reduction, revenue growth, and enhanced profitability.

But delivering those benefits requires discipline. There are right approaches and principles for introducing AI and machine learning applications successfully into your operations with an eye toward maximizing their potential value.

Defining AI and Machine Learning

AI is a broad term that can encompass any simulation of human intelligence processes by machines. These processes include learning (the acquisition of information and rules for using the information), reasoning (using rules to reach approximate or definite conclusions), and inference or prediction. Applications of artificial intelligence perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.

Machine learning, a subset of artificial intelligence, revolves around the idea that machines can adapt and learn from experience, just as people do. Applications of machine learning perform a specific task without explicit instructions, instead relying on patterns and inference. Manufacturers can apply machine learning to execute certain activities humans could do — for example, going through large data sets to find patterns — albeit much more efficiently.

One application of machine learning that has seen acceleration in recent years is around advanced process controls. An operator uses instructions from a machine learning model to adjust the settings in order to optimize inputs for a specific outcome. For example, if sustainability is a goal, then the processes are programmed not only to make sure widgets are produced on time but also for minimal energy consumption. This is beneficial in cases where there are many variables, including human variables such as new or inexperienced operators.

A Practical Approach for Introducing AI / ML

The approach outlined here largely follows the CRISP-DM (cross-industry standard process for data mining) model, considered by many to be the gold standard for developing analytics models. It includes seven key elements, with particular focus on up-front discovery and delivery of tangible business outcomes through operationalizing the insight.

1. Business understanding

AI and machine learning answer questions, but they do not solve problems. For example, we can use these capabilities to predict when a piece of equipment is going to break and prescribe how best to fix it, but they won’t keep the equipment from requiring repair. AI/ML projects can’t move forward until there is a clear and common understanding of the question(s) to be answered.

It’s important to start with a deep examination of the potential impact, partnering with cost accountants and continuous improvement managers or engineers to understand base process data and look at equipment performance.

From this analysis, you can begin to identify potential use cases that generally align with one of two goals: increasing revenue or reducing costs. Develop a value hypothesis, a statement about a single, objective question the analytical model needs to answer and the expected result when successful. Success metrics to define what’s in it for the company.

Workshops with operations managers, production managers, and operators who have shop floor experience are valuable for understanding pain points and refining potential use cases.

This is the single most important exercise you can do—and it must happen up front.

2. Ideation
Once there is a clear understanding around the business impact, then you can start to bring a potential use case to life. It is critical to involve those who will be using the application or its output.

Workshops with operations managers, production managers, and operators who have shop floor experience are valuable for understanding pain points and refining potential use cases – for example, a machine that has a higher-than-expected rate of failure. If the operator had more up-to-date information, that could be used for planning. Predictive signaling could also help procurement obtain the necessary parts to minimize downtime.

3. Data understanding
Some of the biggest challenges center around data and the pipeline for accessing it, particularly with a mix of legacy and modern equipment. This can create huge variances in data quality. You’ll need to consider the maturity of current systems to understand input and output variables and what constitutes “good” and “bad” data. From an AI/ML perspective, key indicators of data quality may include timeliness, completeness, consistency, accuracy, and relevance.

You’ll also need to understand how to access, combine, and prepare data from multiple sources in order to apply machine learning. There are new tools, such as SORBOTICS, that can connect with a variety of factory machines, even older ones, to access data.

AI/ML projects can’t move forward until there is a clear and common understanding of the question(s) to be answered.

Finally, use of AI/ML also underscores the importance of optimizing the data historian to provide visibility to data at an asset level so that you understand process performance, rather than relying on manual observations.

4. Data preparation
Manufacturing data often has not been “cleaned” and prepared for AI/ML use. Therefore, you’ll need a process for filtering data that is inaccurate, incomplete, or irrelevant and then deleting or modifying it in order to produce data sets appropriate for analysis. Otherwise, you risk a garbage-in, garbage-out situation that not only diminishes the effectiveness but increases the risk of acting based on incorrect insight. Don’t underestimate the significance of this step nor the effort required to address it.

5. Data modeling
With the right “clean” data, you can begin using machine learning models to test the value hypothesis. This will require expertise in data science and analytical technologies in order to design and build the capabilities for modeling. Note that there are plenty of established solutions that can fast-track development using out-of-the-box capabilities.

6. Evaluation
To evaluate results, you will need clearly designed success measures and criteria for evaluating both the business impact and the technical analysis capability itself. From a business perspective, this means understanding the cost of a false positive and the savings from a true positive, using test data to determine if a model is saving money or costing money.

From a technical perspective, evaluation should look at how to continue improving the model’s results. It’s important to approach this process as an iterative one, where you are continuously evaluating the model and getting better results as you train the algorithm.

7. Operationalization
Insights from AI and machine learning won’t create any value if they’re left sitting on the shelf. You’ll need a well-designed business consumption mechanism that defines how the business receives and uses the insight, as well as capabilities for measuring and reporting value. One of the easiest and most effective ways to promote business consumption is to integrate the insight into a system in the form of a record of action to be taken by a system user or operator.

Involving people with domain knowledge early on when defining problems and developing use cases reduces surprise later and prepares them to embrace the solution going forward.

It is also impossible to overemphasize the importance of change management. Involving people with domain knowledge early on when defining problems and developing use cases reduces the element of surprise later and prepares them to embrace and own the solution going forward.

It’s also important to communicate value and expectations. For example, the solution will produce information that helps operators be more effective in meeting their key performance indicators, but it won’t replace their job. Be open in acknowledging these fears as you would do when introducing any new automation solution.

Moving Ahead with Purpose

There has always been hype surrounding emerging technologies, but it can take time for expectations and reality to meet. AI and machine learning are driving tangible value for manufacturers today, but you need to get past the hype and focus on introducing it in a practical way. This requires a disciplined approach that starts with value, employs domain expertise throughout, and has the right mechanisms in place to operationalize the insight gained.  M

Footnotes:
1. “How AI Builds a Better Manufacturing Process,” Forbes, July 17, 2018
2. Press Release, Gartner, October 1, 2020
3. “New research reveals what’s needed for AI acceleration in manufacturing,” Google Cloud, Dominik Wee, June 9, 2021
4. “Machine Learning Glows Brighter,” Chemical Processing, Séan Ottewell, August 5, 2020