Manufacturers know variability is a big problem. But sometimes they don’t know how big. Manufacturing variability isn’t just an isolated anomaly. It’s a tangible and often widening gap between the intended objectives of a process and the actual result. It reflects inconsistencies in manufacturing operations that undercut reliability, increase costs, and degrade quality.
Quite simply, variability shouldn’t happen. Whether fabricating moving parts or producing chemical reactions, for example, the process should result in the same outcome, day after day. You want consistency. Knowing where the inconsistencies are coming from, however, is the challenge. The creeping causes that lead to variability are rarely obvious to the human eye. I’ve seen countless instances where small variations are overlooked and wind up becoming larger, costly failures. Fortunately, tools like AI can perceive them where people can’t.
Consider this recent case of a company that uses melting furnaces. The standard operating practice on their factory floor was to check and adjust the air and gas ratio of their burners only intermittently. As a monitoring strategy designed to uncover issues, it revealed next to nothing.
But when the company deployed sensors linked to AI and machine learning-enabled technology to monitor furnace internals and forecast temperature and glass quality, they discovered plenty of room for improvement. Via these digital tools, they learned that if they more frequently adjusted the air and fuel ratios in burners, electric boosters, and other factors, they could achieve markedly better quality and yield.
This approach, not incidentally, is saving the company more than $1 million per year for every degree of temperature that AI helps to consistently control. Since the company has furnaces in facilities around the globe, that’s a hefty benefit.
AI-Enabled Multivariant Analyses Gets to the Root of Plant Stoppages
Say you have a plant manager who knows there’s an issue and has a hunch about the cause — a common, yet frustrating situation. What the manager needs — more than anything — is more analytical muscle to really find out what’s going on. AI can incorporate many information sources — unlike the typical siloed approach of focusing on just one — and track the history of developments and cascades of events that include “unknown unknowns” — those elements of a machine’s work that plant managers would never consider under normal conditions. It can discern whether the problem has been developing over time and to what extent, providing a fuller picture not just of the problem, but the solution.
This happened recently at a manufacturing plant where we worked. Their compressor kept breaking down. The plant manager suspected a mechanical problem, but how to solve it was a mystery. Then the manager started to take advantage of the underused data coming from their instrumentation and other equipment and began to track every detail of their machine and facility’s operations.
What this data-driven investigation discovered was that radial vibration was increasing in the compressor and that the vibration was due to a spike in the shaft speed. The spike itself was caused by elevated temperatures in the intercooler — which originated in the record-breaking high temperatures then happening outside the plant. The plant used seawater to cool the compressor. On extraordinarily hot days when the water was too warm, the compressor malfunctioned. But it could be modified to work in higher temperatures.
Given that the higher vibration and shaft speeds were certainly straining the compressor’s bearings, shaft, and gearbox, that could lead to a serious catastrophe in time. In the ensuing crisis, plant managers would have rightly discerned the mechanical issue, but not known why — allowing the same cycle to just continue and scale as a problem throughout the company’s operations worldwide where they used the same equipment.
By running a multivariate analysis that combined mechanical data and process data in a multivariate model, however, the company spotted creeping degradation in machine performance. The root cause and recommendations provided by the system enabled the control room operator to intervene before failure occurred.
An Explosion of Manufacturing Data
It’s not that the plant managers in the above examples needed more data. They needed a means to gather and analyze the data, explore a hypothesis from every angle and discover the answers hidden within. At present, there’s an incredible amount of industrial data available to manufacturers. In fact, the data generated by machines in manufacturing plants and other facilities is growing 50 times faster than business data. But the industry isn’t exploiting around 98 percent of it — which means as far as shedding light on variability, they’re operating in the dark.
What plant managers need are computational power, automation, and analytical strength to monitor signals such as flow, temperature, pressure motor current, vibrations, and more. They need a way to glean a history and detect concealed patterns, dig into the data in real or near-real time, and use the conclusions for better decisions. Being able to train on the recurring issues spotted by AI means they will be able to spot the telltale signatures as they occur — not after. The full spectrum of information revealed by AI removes the limitations of the linear models that plants often had to use for investigating any issues. That’s where people come in, of course. People need to know how to conduct these analyses using AI and data, and they need to know what to do with the new information at hand.
There’s another need as well, particularly for plants in more volatile markets. Once an issue is identified, AI can also help pinpoint the least cost-intensive times for making the requisite adjustments and fixes. The approach enables the plant to solve problems without creating another one: the cost of downtime during market phases when prices (or costs) are at their peak.
By becoming multi-disciplinarians and learning how to train on and work with AI, machine-enabled learning, and automation, however, we can quickly develop the means to fully investigate manufacturing variability — reducing breakdowns, inefficiencies, concealed glitches, and other hurdles to peak efficiency and productivity. This approach also builds bridges across traditional silos in the organization — such as process and maintenance —bringing both together to share and optimize the same data-driven foresight.
Truth be told, it’s not as complicated as it sounds. The Internet of Things has permeated countless arenas. It’s already driving plant operations. It’s not an anomaly. The industrial equipment sensor market will grow by more than $21 billion between 2021 and 2025, — a massive jump in the Internet of Things (IoT) data ecosystem that’s already driving factory automation. Manufacturing is well aware of the importance of not being left behind and is waking up to the value of deep learning and advanced forecasting techniques for process optimization.
So, it’s time for engineers to become “citizen data scientists” — fluent in not only applications like Python and tools like Excel macros that explore data and its correlations, but also in AI. Being a plant manager or an engineer in the future will require data awareness and a comfort level with AI that’s the same as with a spreadsheet.
Two Snapped Bolts, $2 million
It’s certainly simpler to deploy technologies that are inherently more sophisticated than human judgment — and arrive at solutions that are clear, cost-effective, and fast — than try to adapt to flawed performance with workarounds or a just-keep-going mindset. I’ve seen mystifying variables turn into painfully obvious factors when run through AI.
In one case, once AI was let loose to swim through and analyze a trove of historical data, it predicted that failure would happen 19 days before it actually did, losing a yield of $1 million per day for two days. To those concerned about adding complexity to the already complicated state of manufacturing today, I’d counter that AI is not adding complexity, it’s overcoming it.
Variability isn’t new in manufacturing. But it is getting more expensive. When a plant doesn’t capitalize on an innovation that exceeds humans’ capacity to tame it, the plant’s competitors probably will.
Dominic Gallello is CEO of SymphonyAI Industrial.