Ruslan Grumble - Fotolia
Manufacturing hasn't earned as many artificial intelligence headlines as other vertical markets, but experts believe AI in manufacturing is making healthy inroads. "Cognitive capabilities include natural language processing, object detection, face detection, sentiment analysis, logo detection and translation, all of which can be applied to an industry, regardless of their vertical, from media and entertainment to legal e-discovery and compliance to manufacturing," noted Tyler Schulze, vice president of strategy and development of cognitive engines at Veritone Inc., a provider of AI insights and cognitive technology based in Costa Mesa, Calif.
One of the world's largest adhesive manufacturers recently used AI to tackle a $300 million waste and quality problem. The company worked with Flutura Decision Sciences and Analytics, a Palo Alto, Calif., IoT intelligence company, and used its AI features to identify abnormalities in sensor stream data. Newly discovered sensor/operational abnormalities were correlated to quality outcomes.
Derick Jose, co-founder and chief scientist at Flutura, explained that the AI algorithms used were real-time anomaly detectors on high-frequency sensor electromechanical-hydraulic data streams. "The algorithm learns the typical behavior by processing states for over a year of sensor data," Jose said. Once the system observes a point that is not in the "normal healthy zone," it sends a signal back to close the loop in real time -- for instance, on an operator's paint gun.
In another instance of AI in manufacturing, a U.K. car maker used the Cognitive Predictive Maintenance Platform from DataRPM, based in Redwood City, Calif., to spot potential problems with production equipment and help reduce or eliminate expensive downtime. The manufacturer wanted to analyze its factory to identify which external factors could impact its production, as well as to predict how it could optimize the efficiency of a machine and reduce maintenance costs. DataRPM combined the data from the company's production log with external factors to provide actionable insights for day-to-day decision-making and increasing machine efficiency.
Where AI in manufacturing fits
Generally speaking, the impact of AI in manufacturing is twofold, said Mark Hung, research vice president and lead analyst of IoT at Gartner.
"On the back end, machine learning is used and applied against data coming in from the manufacturing process to then help enhance the operation of the manufacturing line, reducing downtime and generally improving performance of equipment," he said. Likewise, on the operational side, AI is starting to make an impact on activities such as visual inspection to help reduce defects in products.
Mark Hungresearch vice president and lead analyst of IoT, Gartner
Hung said that the back end has seen the most development because those analytic models have been around the longest and are now simply continuing to evolve. "Adoption on the operations side is off to a good start, but [it] will probably see some speed bumps," Hung added. "For many years in manufacturing, the focus was on operational tech. Now, with IoT, it is bridging that existing operational tech infrastructure and integrating it with the rest of IT."
"So far, the success of AI in manufacturing has depended on domain expertise, the quality of the data and code, and best practices," noted Said Tabet, Ph.D., chief architect for IoT solutions at Dell EMC and a leader in the Industrial Internet Consortium (IIC). He explained that many companies already have big data sets, so they are trying to improve business processes not by new workflows, but new capabilities in existing systems, shifting to the interpretation of data in a way that will have the greatest impact on automation while reducing human involvement.
"Some of that is already successful, for instance, in using deep learning to detect anomalies and figure out the state of a specific process, because abnormal signals of failure can be detected much faster with deep learning," Tabet said.
While predictive maintenance has already become relatively common in both field service and manufacturing, the emerging field for AI in manufacturing is predictive quality, said Tanja Rueckert, Ph.D., executive vice president of the IoT and customer innovation units at SAP and a vice chair of the IIC Steering Committee. For example, she noted, organizations are starting to adopt a range of visualization and machine intelligence capabilities to spot quality issues and predict potential problems in time to prevent them. She said SAP recently showcased another AI capability at Hannover Messe -- namely, the concept of autonomous agents for dynamic routing in production activities. "We have seen a lot of interest in that, especially from the automotive industry," she said.
Of course, manufacturing is a particularly good candidate for sophisticated analytics because it is already well-instrumented with sensors and controls, said Mike Gualtieri, vice president and principal analyst at Forrester Research. However, he noted, "the pieces and parts don't necessarily talk to each other." Indeed, according to DataRPM, its U.K. implementation had to contend with 30 different air handling unit file structures in temperature data and inconsistency in time granularity of data sources with multiple data formats, along with similar inconsistencies in electrical data.
But the means to overcome that issue are plentiful. Then, injecting AI into those processes, specifically machine learning, can enable, at a minimum, better predictive maintenance, Gualtieri said.
Some of that has been done in the past, he said, through scheduled maintenance and rules of thumb, but further advances were stymied by false positives or negatives based on an inability to adequately cull through and interpret raw sensor data. Because machine learning correlates disparate data, it opens doors to processes becoming more accurate and keeping machines in service more of the time, he said.
The current and future states of AI in manufacturing
Despite the fact that the challenges of combining AI and manufacturing are not insurmountable, Gualtieri said there is little that is "commercial off the shelf" in terms of applying AI in an industrial or IoT environment. "At the moment, it is all separate pieces," he said. "This is a huge issue from a vendor standpoint because it may slow adoption. But if you are a manufacturer with a specific use case, you can still implement something; it just might take more effort."
Gualtieri said startups and established companies alike are moving toward providing middleware or cloud-based platforms that can help tie the parts together.
There is one other important thing to consider, Gualtieri said: Consider the need for data science talent. Although the AI building blocks are getting more user-friendly and powerful, data scientists are indispensable to help refine the setup of AI in manufacturing and garner the most value from the output.
For now, he classifies AI in manufacturing as "nascent." Most everyone seems to recognize the potential, but relatively few companies have gone ahead and applied it, he said.
A similar observation, but with more optimistic conclusions backed by survey data, came from Kimberly Knickle, research vice president at IDC Manufacturing Insights. Although the results haven't been finalized and put into a report yet, she said the survey IDC conducted revealed a level of AI adoption that surprised her. For instance, across the U.S., about 10% of packaged goods manufacturers are already employing AI. "What was even more surprising is that many more were either evaluating AI or starting to implement it," she added.
Other verticals covered in the IDC survey -- notably, high tech -- also seem to be making a move toward adopting AI in manufacturing, she added. So, for now, it may be a nascent movement, but, perhaps, not for much longer.