There’s lots of talk about machine learning, but can it help you improve the performance of your machines? It’s becoming easier and cheaper to gather and store data related to the operation of your machines installed at customer sites, but this creates a problem which we’ll label as Data Rich, Information Poor (Figure 1). Machine learning is being promoted as the solution, and the cover story in the May 2016 issue of Control Design magazine titled “Is machine learning smart enough to help industry?” investigates these claims by examining the experiences of machine builders and their suppliers.
As related in the cover story, “From the industry or manufacturing side of business, machine learning can be applied to just about any control system smart enough to actually alter how it controls a machine in response to changing conditions.”
Can machine learning address issues?
“Machine learning will help machine builders, integrators and end users by allowing the machines to solve the problems that typically can only be done by humans and, in some cases, can’t even be done by humans,” said Matt Wicks in the cover story. Wicks is the vice president, product development, manufacturing systems for Intelligrated (www.intelligrated.com), a provider of automated, intelligent conveyance and robotic handling systems in Mason, Ohio.
“When discussing machine learning, the top item is the ability to process large amounts of data analytics to identify patterns and trends not readily visible using traditional statistical tools. This information can be leveraged for preventive maintenance and/or machine and system optimization improvements. When considering machine learning for robotic and computer vision tasks, such as object recognition, pose estimation and complex motion involving perception, it provides improved robotic operation and performance (Figure 2)”, related Wicks.
Machine-learning techniques can be applied to the aspects of perception, allowing the equipment to respond to changing and novel scenarios, continued Wicks. “Perception is a component and, relatively speaking, low-hanging fruit for AI,” he said. “Coupling vision, motion and machine learning can provide more impressive results, allowing not only the perception tasks to perform at a higher level, but allowing the perception to be combined with the machine motion, yielding performance levels that may be more optimal and capable of functioning in a much wider variety of scenarios.”
While we are still a long way from leveraging machine learning to help design material-handling equipment, in the short term, it’s feasible to see AI/machine learning techniques applied to optimize certain operations, explained Wicks. “Robotic technology is moving away from the traditional ‘program it and let it do the repetitive operation over and over again’ to a ‘move, see, think, act’ model,” he said. “This roughly translates into machine vision, processing algorithms and physical motion. Machine learning can help individually with each of these steps, but, when taken as a whole, it can yield some very impressive results.”
Not only is machine learning being developed at the machine level, interest is increasing across industry, claimed the cover story. “One thing you might want to take a look at is the Google Trends comparing the search volumes for machine learning vs. artificial intelligence vs. neural networks vs. late-comer deep learning,” noted Michael Risse, vice president and CMO at Seeq (www.seeq.com), in the cover story.
“There might be other terms to consider—prescriptive analytics, for example—and then there are the process-industry-specific analytics tools such as advanced process control (APC), statistical process control (SPC), multivariate analysis and even application performance management (APM),” added Risse. “It’s a long list. What is machine learning exactly? Is it one thing, or is it all of these things? And why the resurgence
of interest?”
The IoT, cloud and big-data technology are likely pushing the interest in machine learning and how it will be used, according to the cover story. “Without machine-learning techniques, the future of IoT solutions would certainly be limited,” noted Nikunj Mehta, founder and CEO at Falkonry (www.falkonry.com) in the cover story. “IoT solutions are complex in the way biological systems are complex. Systems are composed of large numbers of interacting things, each possessing their own complex behaviors; structure and behavior are not fixed, but evolve over time. IoT systems are great at producing data, but without automated learning techniques, the data volumes drown effective use. Builders and users of IoT solutions need systems that learn and that are adaptive.”
Machine learning defined
As related in the cover story, “How machine learning will influence and improve machines and manufacturing is a tough question, but the definition of machine learning is not science fiction, and it depends on who you ask.
Machine learning is any number of algorithms that use an optimization objective function to help a computer interpolate or extrapolate trends from a learning data set to apply to unknown data, explained Anthony Skjellum, PhD, professor of computer science and software engineering at Samuel Ginn College of Engineering, Auburn University (www.eng.auburn.edu) in Auburn, Alabama. Skjellum’s view is more computational. “Correlations can be determined. Identifying causation—that the correlations mean something— is still the human’s job.”
“Humans make and break models constantly,” said Skjellum in the cover story. “That’s a key aspect of human intelligence. Machine learning tries to show correlations. Humans then abstract models, do further experiments and determine if the model is a useful abstraction. It is a closed-loop process. There can’t, at present, be a purely AI data scientist.”
But as the cover story reveals, others prefer a wider software view of machine learning. “To us machine learning is the ability for software systems to use observations of the world around it,” said Mike Haley, senior director, emerging products and technology at Autodesk (www.autodesk.com) in the cover story. “The physical, virtual and textual worlds are used to understand and predict behaviors and semantics that the program was never explicitly programmed to understand. In that way, these machine learning systems are truly dynamic.”
The cover story said perhaps the better way to look at machine learning is to consider the computer and market angles separately. “There is a computer-science and a market answer to what machine learning is,” commented Seeq’s Risse. “The computer-science answer is machine learning uses automated and iterative algorithms to learn patterns in data, so you don’t program the endpoint solution at the outset. Instead the algorithm adjusts itself—by learning from one data point to the next—to solve a particular problem as part of the process, using either a supervised, training-set or unsupervised starting point.”
The market answer is that machine learning is on the cusp of joining big data and the IoT as a marketing necessity for modern software offerings, such that the technical definition or correctness of any particular solution is lost in the hype, continued Risse. “And that is just within machine-learning offerings,” he explained. “There are many other computer-aided insight tools vying for attention at the same time: deep learning, machine intelligence, artificial intelligence. The answer is getting more marketing-focused over time, given the competition within and across the ecosystem.”
Machine learning is here to stay, said the cover story, and David White, senior research analyst at ARC Advisory Group (www.arcweb.com), agreed. “Machine learning is going to be an essential technology moving forward,” he predicted. “In practice, data from the Industrial Internet of Things is going to be big data for industry. We have reached the point where data visualization and the human eyeball aren’t going to be enough. We are generating a much greater volume of data at a much greater speed than ever before. This presents challenges and opportunities.”
The challenge is to make sense of all the data in time to make the right management or operational decisions, explained White. “Machine learning can help here,” he said. “For example, machine learning algorithms executing close to assets and processes—at the edge—will be able to work with this complex data and make intelligent decisions in real time. The opportunities then are to improve productivity by reducing scrap or cut maintenance costs by moving to predictive maintenance, rather than preventive maintenance. This is a clear benefit to end users. However, machine builders can also take advantage of machine-learning technologies to gain competitive advantage.”
Applying machine learning
“The key thing about machine learning is that the performance of the algorithms improves over time,” added White. “There is a long-standing rule of thumb among machine-learning experts that a weaker algorithm with more data will ultimately outperform a stronger algorithm with less data. In this way, I think machine learning can help to make control systems much more agile and responsive in meeting changing needs.”
But adoption of machine learning in heavy industrial settings is still very limited, largely due to its only recent emergence as a viable control method, notes the cover story. “You’re seeing it begin to appear in some robotic systems—mainly related to computer vision and path planning—and object-avoidance drone technology, but it’s only just beginning,” said Autodesk’s Haley. “Perhaps the biggest uptake in machine learning is in the IoT sector, where learning predictors of failure modes in devices is of significant value. In all of these cases, the inherent ability of new machine-learning systems to find and predict patterns in many often-messy data signals is driving the state of the art.”
Human intuition is limited; machine learning can point out correlations that are subtle and that humans can’t see; it makes people smarter when used wisely, explained Auburn’s Skjellum.
“If you start with a big bunch of data, and machine learning gives you a correlation, the scientist or engineer hypothesizes a causation, does experiments, maybe more machine learning, and concludes whether the hypothesis, concept and relationships are real or coincidental,” said Skjellum.
Autodesk’s Haley saw benefits mainly by accelerating the design and development process through learning from everything that has already been done. “In this way designers and engineers can rely on a smart system to guide them through a design, making sure they are incorporating the best approaches, avoiding duplication and tracking most closely to the desired solution,” he explained. “One simple example here is the amount of duplication of designs or components that can occur over many years in a large firm. This can be almost entirely eliminated through machine learning and real-time guidance provided to a designer.”
As sensors become cheaper, along with networking hardware and storage, machine learning offers more opportunities. “Well-understood patterns discerned via machine learning from product-usage data can point to design and operational improvements,” pointed out Falkonry’s Mehta. “Likewise, machine learning can use data collected from production processes to identify conditions and guide process improvements.”
Most manufacturers struggle with creating and maintaining accurate process data, such as production time standards, yields, run times and setup times, said Jim Cerra, cofounder and CEO of PlanetTogether (www.planettogether.com). “This data is instrumental in creating optimized production plans and schedules and driving higher productivity and on-time delivery,” he noted. “Predicting machine downtime, absenteeism and other probabilistic data is even more difficult, but it would be helpful in assessing the risk of customer-service issues due to delayed delivery and revenue-attainment shortfalls for the company.”
If the computer can learn from watching these manual changes, systems can then begin to suggest or even automate the work, freeing more of the planner’s time to make the tough decisions that only a human can make, said Cerra.
“As computers get faster and algorithms continue to be refined for machine learning, greater inferences become possible in near real time or real time,” related Auburn’s Skjellum. “Combining machine learning with predictive simulation and feedforward and feedback control can help to address complex control-system objectives, while also enhancing the potential for detecting cyber threats or other kinds of disturbances.”
Lots of system designs done with machine learning produce results that are not intuitive to the best human designers, explained Skjellum. “There are many process variables and figuring out the most important ones—a kind of identification problem—is not new; it is an important part of systems and control,” he said. However, looking at the vastness of historical data or the vast number of sensors in a plant for correlations and failure prediction is a new and emerging machine learning example.
“Autodesk is experimenting with robotic systems that watch and learn from skilled craftspeople, and then work alongside these people to add precision, repeatability and/or safety to their work,” said David Thomasson, principal research engineer at Autodesk. “As the new breed of collaborative industrial robot arms, such as the Universal Robots models designed to work in close proximity to people, become more common, there is a need for more intuitive interaction with these machines (Figure 3). This is being enabled by machine-learning systems that automatically determine the best course of action based on their understanding of human preferences and abilities, along with a more complete awareness of the design and the situation in which it is to be realized.”
The right tool for the job
According to the cover story, “Seeq provides an application dedicated to time-series data investigation. It allows Google-like searches, collaboration in real time and interaction with analog data series like never before. Seeq is intuitive; it’s visual; and it connects to just about any process historian to find answers in the data.”
“A CoPilot software feature uses available server computational capacity during idle times to find a better schedule for companies automatically,” noted Cerra at PlanetTogether.
As the cover story relates, “If plans are found that result in higher KPIs, such as better on-time delivery, increased cash flow or reduced costs, then the new scenarios are automatically presented to the planners who can compare the scenario to the current plan and choose the best option. This requires massive amounts of computer computations, especially in complex factories with tens of thousands of orders in the schedule and quadrillions of possible solutions to choose from.”
“There is a wealth of information available on the techniques associated with machine learning,” noted Intelligrated’s Wicks. “Google has recently open-sourced its machine-learning software called TensorFlow. Google believes machine learning is a key ingredient to the innovative products and technologies of the future. The research in this area is growing fast but lacks standard tools. By sharing, Google believes it can create an open standard for exchanging research ideas and putting machine learning in products.”
Machine learning can certainly accelerate the design process all the way through to physically realizing a product, said Haley at Autodesk. “It will mean less time designers or engineers spend on tasks not directly related to the creative aspects of their jobs,” he said. “One of the most important aspects of machine learning is to have a sufficient quantity of data for it to learn from. Making sure the data in your company is gathered somewhere, reachable and understandable will go a long way to successful adoption of future machine-learning technology. Most of the human-computer interaction models that relate to machine learning strive to avoid any particular understanding or skills on the part of the user.”
The functionality surfaces as suggestions, recommendations, alerts or even something completely transparent to the user, explained Haley. “The design and engineering software world is still working on this, so just keeping up to date on the latest software available would be a good start,” he added.
There are a lot of packages that will help, and the fundamental algorithms are well documented, said Auburn’s Skjellum. But the cover story cautions, “However, it is as important or more important for engineers to understand the limitations of machine learning and how to interpret results, compared to learning to use canned packages.”
“Fortunately, it has never been easier to get educated on machine learning,” said David White at ARC Advisory Group. “First, there are some great courses available online, offered by organizations like Coursera and edX, which are mostly free. Second, open-source software and the cloud make it much easier to experiment with machine learning at little or no cost. There are a number of cloud-based platform-as-a-service (PaaS) solutions where you can just sign up and get started with machine learning for free. Most of the solutions have tutorials, and they are quite modern and visual, as well. If you’re cloud-phobic, as an alternative, there are open-source solutions you can also download and play with for free.”
Machines as mentors
Several of your work colleagues may be computers in the future and will possibly be great mentors, the cover story speculated. “I think that cognitive AI can certainly be a source of mentoring, and IBM Watson could certainly be giving advice sooner than later in specific niches,” said Skjellum from Auburn University. “Also, machine learning that helps you to discover patterns that are constructive and destructive to personal and professional success, such as procrastination and eating poorly, are on the near horizon. This is not a long-term thing; it is coming shortly—months not years.”
These same things can apply in industry, continued Skjellum. “Mining the behavior of a company over time to show success and failure patterns in hiring, promotion and business processes seems likely,” he speculated. “Now, to mentoring, that is simply a decision support approach based on what works versus what fails. That can be not only for a given company but also learned over comparable industries and cross-correlated with successful firms in adjacent spaces.”
So mentoring of sorts from machines is a logical thing, predicted Skjellum. “In control systems, which depend heavily on math modeling and understanding nonlinearities of systems and also the complexity of human interactions, there is clearly room for operators to be mentored by computers in complex decision making scenarios,” he said. “Collision avoidance systems for aircraft, and coming now for cars, are a kind of decision support. Humans deal only with a few variables at a time; while control operators gain intuition over time, they don’t necessarily make the right choices under stress, or even the right choices when confronting situations that are uncommon. Think Chernobyl.”
Brave new world
As more and more computing power emerges and computational elements are also optimized for the machine learning algorithms—such as special-precision massively parallel general-purpose graphical processing units are doing nowadays—faster and faster machine learning will become a reality, said Auburn’s Skjellum. “Distributed learning on handheld devices will complement that for quick and dirty decision making under uncertainty, right at the user’s fingertips,” he explained. “Technologies such as Siri and Cortana will become digital assistants that help more and more, while engineers will build the machine learning directly into the online processes of industrial systems as integral, rather than off-line or near-line.”
Machine learning will gradually become more competent at sensing—taking lots of signals, including video and sensor data and understanding patterns in them—to the point that they’re far better at it than humans could ever be, predicted Autodesk’s Haley.
The cover story concluded: “This will allow software systems to more seamlessly understand and interact with the world around them. That said, this is only a tiny step toward realizing a machine-learning future where the systems themselves can perform truly intelligent reasoning. That’s still a long way off, but, along the way, there will be a lot of interesting and useful solutions such as generative design systems, adaptive industrial systems and cognitive computing.”
By Dan Hebert, PE
Originally Published: March 1, 2017
