Commentary: Understanding the data issues that slow adoption of industrial AI

In this article, I distill portions of “Industrial AI: Applications with Sustainable Performance,” a book published in February by Jay Lee.

Lee is a board member and vice chairman of Foxconn Technology Group and Foxconn Industrial Internet. He is also an Ohio eminent scholar and the L.W. Scott Alter Chair professor and a distinguished research professor at the University of Cincinnati. He also is the founding director of the Center for Intelligent Maintenance Systems (IMS) at the University of Cincinnati.

Additionally, Lee is a co-founder of Preditronics and CyberInsight, two startups spun off from IMS. He is also a member of the World Economic Forum’s Global Future Council on Advanced Manufacturing and Production.

While reading “Industrial AI,” I was also reading “Artificial Intelligence for Industrial Applications” by Sam Charrington. Charrington is the founder and principal analyst of CloudPulse Strategies, “an industry research and analysis firm focused on enterprise adoption of machine learning and artificial intelligence and the tools, platforms and technologies that enable them.”

I augment the observations and ideas from Lee’s book with observations from Charrington’s article.

What is industrial AI?

Industrial AI is a systems discipline that integrates several technical elements to solve problems that are commonly encountered within the industrial sectors of any economy. Understanding industrial AI requires answering these fundamental questions: First, what kind of intelligence do industrial systems need? Second, what problems and challenges have not yet been solved by previous methods? Third, how can AI solve these problems?

Lee defines industrial AI as “a systematic discipline which focuses on developing, validating and deploying various machine learning algorithms systemically and rapidly for industrial applications with sustainable performance.”

Charrington defines industrial AI as “any application of AI relating to the physical operations or systems of an enterprise. Industrial AI is focused on helping an enterprise monitor, optimize or control the behavior of these operations and systems to improve their efficiency and performance.”

Lee observes that “using traditional measuring, monitoring and statistical analysis, we can only manage the visual problems that have occurred, but have no way to solve invisible problems. The real value of intelligent manufacturing should be to help us recognize, manage and avoid these invisible problems.”

He goes on to explain that “the first purpose of industrial AI is to make the hidden problems in an industrial system explicit, then to avoid those problems by managing them while they remain hidden,” and also that “the second goal of industrial AI is to accumulate, inherit and apply knowledge on a large scale.”

This leads to another question: What is industrial knowledge?

The foundations of industrial knowledge

According to Lee, “Knowledge within an industrial system is the relationship and operations of objects, environments and tasks. It is an abstract expression of comparability, relevance and purpose.” He summarizes this with the three Rs: resources — data and information from humans or machines; relationships — mathematical relationships between the visible and invisible variables that are part of every industrial process; and reference — results that form the basis of future actions within an industrial process.

“Industrial AI can achieve a significant improvement in the management of the above three Rs,” he says. “It can acquire and manage data from richer sources, model more complex relationships and provide reference and comparability in broader dimensions. In the final analysis, it is a process of optimizing real-time decision-making of management and control activities and efficient execution of employees being able to assess the state of the processes.”

How is AI different from industrial AI?

The book contains an interesting table that outlines the differences between AI and industrial AI.

In my opinion the most fundamental difference between AI and industrial AI is that AI is, in Lee’s words, “a trial-and-error judgment-driven technology,” whereas industrial AI cannot under any circumstance be successful if it’s based on trial and error. He states that AI “can be applied to a wide range of fields such as medical treatment and business, but does not have a successful usage case in engineering.”

AI functions in “divergent and opportunity-driven situations such as autonomous driving, economy sharing and facial recognition,” while industrial AI functions in “convergent and performance/efficiency-driven situations based on improving from an original basis, such as improving production efficiency, improving quality, reducing energy consumption costs, improving equipment stability and improving automobile safety.”

AI is applied in “social networks, financial sector, medical industry, among others,” while industrial AI is applied in “a broad range of industrial applications including industrial equipment and manufacturing, power grids, power generation equipment, transportation and logistics, medical systems, etc.”

Finally, the algorithmic approaches implemented in AI are “machine learning, deep learning,” while in industrial AI the algorithmic approaches are “deep learning, broad learning, fuzzy learning, augmented learning.”

While the book does not highlight reinforcement learning specifically, other sources I have read discuss reinforcement learning in terms that cause me to believe that it will become more commonly applied in industrial settings.

For example, Warren Powell, who regular readers of this column would have encountered in this #AIinSupplyChain Series on July 7, July 17 and July 28, is nearing completion of a new book, “Reinforcement Learning and Stochastic Optimization: A Unified Framework for Sequential Decisions.” He is a professor emeritus at Princeton University and co-founder of Optimal Dynamics.

I did not ask Powell if there’s something inherent about industrial AI that would make reinforcement learning unsuitable. However, my reading of his December 2019 paper, “From Reinforcement Learning to Optimal Control: A unified framework for sequential decisions,” leads me to conclude that there’s no reason to think reinforcement learning and industrial AI are incompatible.

For example, in the introduction to the paper, Powell says, “There is a vast range of problems that consist of the sequence: decisions, information, decisions, information. … Application areas span engineering, business, economics, finance, health, transportation and energy. It encompasses active learning problems that arise in the experimental sciences, medical decision-making, e-commerce and sports. It also includes iterative algorithms for stochastic search, as well as two-agent games and multi-agent systems. In fact, we might claim that virtually any human enterprise will include instances of sequential decision problems.”

He goes on to say, “We present a unified framework for all sequential decision problems.” It seems to me that industrial AI is made up of a large number of sequential decision problems.

The five major challenges of data and AI in industrial applications

AI depends on data that is obtained from traditionally digital sources, whereas industrial AI depends on data obtained from physical sources and processes that have been enhanced with some digital capabilities through sensors and other connected devices.

Lee describes five problems, challenges and limitations practitioners of industrial AI encounter. First, “training data is heavily dependent on manual work, otherwise it is difficult to obtain a large and comprehensive training data set, and the quality of labeling is heavily dependent on human experience and ability.” Second, “the transparency of the model needs to be improved since AI algorithms cannot explain how conclusions are reached step-by-step.” Third, “models are not very general, and it is hard to replicate from one application to the next. This means lots of money and energy is needed to train new models for new problems.” Fourth, “the risk of deviation in data and algorithms, much like the differences between societies and culture, requires extensive steps to solve.” And, fifth, “It is difficult to reach agreement on data privacy and attribution.”

He suggests that enterprises in the process of implementing an industrial AI strategy should be cognizant of the three Bs with respect to data: The first is bad quality — industrial organizations must invest in the creation of the necessary infrastructure to obtain data of the required quality. The second is broken — to be useful for industrial AI, data must be comprehensive, meaning that all the data required to solve the problems at hand must be considered upfront so that it can be collected. The third is background — industrial AI requires that attention be paid to hidden correlations as well as surface statistical features. This requires that auxiliary data about the industrial process be collected as well. Auxiliary data is information not directly related to the industrial process but that might affect the outcomes that are observed.

Industrial processes are typically mission-critical in nature. So the bar for reliability of industrial AI systems is very high since the tolerance for failure is very low. Failure in this case leads to loss of life or property, or loss of both life and property. Safety and security are closely related to reliability.

According to Charrington, organizations that seek to implement industrial AI should expect to encounter at least seven problems. First, data acquisition and storage is more difficult because the data is more noisy and much more voluminous, creating problems related to the development of training datasets. Second, the amount of training data required to adequately train algorithmic systems for industrial AI is costly. Third, it is very difficult to test industrial AI systems on production lines of any kind. Fourth, high regulatory and compliance requirements with respect to the validation of changes to industrial processes makes it difficult to easily implement industrial AI. Fifth, the cost of failure is too high. Sixth, industrial systems are extremely complex, making it difficult to cost-effectively develop an industrial AI approach to solving problems. Seventh, the talent required to solve industrial AI problems is hard to find and expensive to acquire.

I went back to Powell’s work to see how he would approach some of these problems, since he and I have briefly discussed related topics on data, decision-making and uncertainty in supply chain logistics in the past.

In a post on LinkedIn titled Batch vs. Online Learning, he says, “The most common approach to estimating a statistical model (especially neural networks) is to take a large data set and then run specialized algorithms to fit the best model. When you get new data, you do this all over again with a new batch of data. With more complex models (such as neural networks) small changes to the data can actually result in significant changes to the estimated model (an annoying property of nonlinear models such as neural networks). An alternative approach is to use adaptive learning where old models are updated with new data, but without reoptimizing from scratch. Also, in operational, or online, settings, data is arriving continuously over time, and the underlying process creating the data (such as demand responding to online advertising) is often changing, which means we should discount estimates based on older data (which is not possible with batch learning).”

This is Topic 13 in his Decision Analytics Series. It was posted Saturday.           

Conclusion

There is still a lot of work to be done before industrial AI can be reliably and scalably implemented across industrial supply chains and applications for monitoring, optimization and control in legacy industries like agriculture, logistics and transportation, energy, manufacturing, construction, distribution and warehousing. However, the conditions appear to be ripe for organizations to experience more success in industrial AI than has been possible in the past.

Amazon acquired Kiva Systems in 2012, and that has formed the foundation on which Amazon Robotics was built. It is one of the sources of Amazon’s significant advantage over its competitors. John Deere acquired Blue River Technology in 2017 for its precision crop spraying technology. In 2018, BP invested in Beyond Limits in order to apply AI and machine learning to BP’s business operations with respect to exploration for oil and gas — BP is both an investor and a customer.

Slingshot Aerospace helps customers like the U.S. Air Force, NASA, Boeing, BAE Systems and Northrop Grumman gain situational awareness of physical assets using its situational intelligence platform that “brings together heavy data streams from different sources — such as satellites, airplanes, drones, and ground-based sensors — to create decisive context for organizations.”

In October 2019, Charles Yeomans, CEO of AtomBeam Technologies, was a showcase presenter at #TNYSCM17: Secure Supply Chains. He described how “AtomBeam’s advanced software technology uses machine learning to reduce the size of individual internet of things (IoT) data files by 70% or more, while adding security” and described commercial pilots and proofs-of-concept that cover nearly every industry mentioned in this article.

Redis Labs has created RedisAI, an “AI serving engine for real-time applications” and RedisGears, a platform for “infinitely programmable data processing in Redis.” Ofer Bengal is founder and CEO of Redis Labs. When I met him in New York in November 2012, our conversation focused on how processing and storage needs for streaming data would continue to be an acute pain point for enterprises of all kinds, and how Garantia Data, which was the company’s name at the time, was building a product and a platform to solve those types of problems.

Citrine Informatics accelerates new product development by embedding AI in smart data infrastructure for materials and chemicals manufacturers in the Fortune 500.

Desktop Metal is going public and will become the first publicly traded additive manufacturing company, in a deal that is expected to raise about $575 million, giving the company a market cap of more than $2.5 billion. The transaction is expected to close in Q4 2020.

There is clearly more work being done in the area of industrial AI than would appear at first blush.

The partnership between FedEx and Microsoft, which we discussed in this column on June 10, Commentary: FedEx/Microsoft partnership points to era of supply chain platforms, is ultimately about industrial AI. Maersk has inked a similar partnership with Microsoft.

In 2017, J.B. Hunt announced that it was spending $500 million on disruptive supply chain technologies — one could assume at least a portion of that investment will be dedicated to exploring industrial AI in the context of J.B. Hunt’s business operations.

The Center for Intelligent Maintenance Systems (IMS) at the University of Cincinnati counts Applied Materials, Hitachi, Mitsubishi Electric, P&G, ALSTOM, Bosch, Caterpillar, Cisco, Chrysler, Ford, GM, Goodyear, Harley-Davidson, Intel, Johnson Controls, Modlex, Nissan, National Instruments, Pratt & Whitney, Rockwell Automation, Samsung Austin Semiconductor, Siemens, Toshiba, the U.S. Postal Service, and United Technologies, among many others, as current or former members.

Given the pace of advancements in the enabling technologies for industrial AI, I expect to see more incumbents and emerging startups begin exploring practical implementations of the methods and techniques developed, discussed and championed by Jay Lee, Sam Charrington and Warren Powell. It is an exciting time to be studying developments in industrial supply chains, industrial intelligence and applied artificial intelligence.

In June, I was having a conversation with Elisa Chiu, founder and CEO of Anchor Taiwan, “a platform for world-class entrepreneurs and professionals to experience and succeed in Asia through Taiwan.”

She made the observation that when people in Asia talk about supply chain, they are typically talking about manufacturing, and when people from the U.S. talk about supply chain, they often really mean supply chain logistics.

Perhaps this means that industrial AI in Asia will focus primarily on applications of AI, machine learning, and decision analytics to manufacturing, while in the U.S., industrial AI will focus mainly on applications to supply chain logistics.

We’ll have to give things a bit of time to unfold. 

If you are a team working on innovations that you believe have the potential to significantly refashion global supply chains, we’d love to tell your story in FreightWaves. I am easy to reach on LinkedIn and Twitter. Alternatively, you can reach out to any member of the editorial team at FreightWaves at media@freightwaves.com.

Dig deeper into the #AIinSupplyChain Series with FreightWaves.

●     Commentary: Optimal Dynamics – the decision layer of logistics?

●     Commentary: Combine optimization, machine learning and simulation to move freight

●     Commentary: SmartHop brings AI to owner-operators and brokers

●     Commentary: Optimizing a truck fleet using artificial intelligence

●     Commentary: FleetOps tries to solve data fragmentation issues in trucking

●     Commentary: Bulgaria’s Transmetrics uses augmented intelligence to help customers

●     Commentary: Applying AI to decision-making in shipping and commodities markets

●     Commentary: The enabling technologies for the factories of the future

●     Commentary: The enabling technologies for the networks of the future

Author’s Disclosure: I am not an investor in any early stage startups mentioned in this article, either personally or through REFASHIOND Ventures. I have no other financial relationship with any entities mentioned in this article.