The manufacturing sector has been notoriously slow to adopt new technologies, and artificial intelligence is no exception. Deep learning models have been out of reach for all but the largest manufacturers, given a shortage of internal specialized AI talent and the difficulty of harnessing complex models to optimize and automate routine tasks.
The lack of universal industrial data has been another major obstacle slowing the adoption of AI among mainstream manufacturers. Manufacturing data is often localized or specific to a particular industry domain or a company’s operations. As a result, there isn’t a lot of relevant data available for building reliable AI models, especially compared with what’s available in industries such as finance or retail, where copious transaction or stock market data is readily transferrable across the sector.
Large language models are built on readily available universal languages that many people understand and that follow specific rules and sentence structures, Kalyan Veeramachaneni, principal research scientist at MIT’s Schwarzman College of Computing, said at the recent MIT Machine Intelligence for Manufacturing and Operations Symposium. “There’s no such thing for manufacturing operations — there is no universal availability of data from turbines, cars, or other signals that we are capturing,” he said.
Closing the AI gaps
Generative AI, data-centric AI, and synthetic data make AI more accessible and suitable for solving manufacturing operations challenges. Generative AI tools, such as ChatGPT, offer a more intuitive way to model complex data sets and images that could open up AI technology to a broader set of manufacturing use cases and user types. Similarly, data-centric AI and synthetic data, which focus on engineering the data needed to build an AI system, shift the focus away from highly specialized algorithmic models to building optimal data sets to train an AI system. These approaches put AI within reach of plant workers and manufacturing engineers, who understand everyday production requirements and process challenges but aren’t necessarily versed in the language of mathematics and complex modeling.
Consider the example of a factory maintenance worker who is intimately familiar with the mechanics of the shop floor but isn’t particularly digitally savvy. The worker might struggle to consume information from a computer dashboard, let alone analyze the findings to take a particular action.
The scenario looks a lot different with generative AI. “What if that worker is able to talk to the system [through generative AI] and get the information instead of figuring out charts and metrics,” said Manoj Kothiyal, a partner at Boston Consulting Group and the tech lead for the company’s digital in AI manufacturing platform. “Now we can augment a lot of these AI and machine learning models with generative AI, and that can make adoption faster and change management easier.”
Generative AI and other advances can accelerate the use of AI for a number of manufacturing use cases, including the following:
Continuous operations, such as helping plant floor personnel quickly identify a particular machine that is operating outside of its preferred boundaries. This would allow for real-time adjustments to prevent downtime or quality issues.
A maintenance companion, which helps shop floor personnel with maintenance tasks by digitizing paper instruction manuals and using AI to provide step-by-step, real-time instructions based on the problem at hand.
Defect detection and inspection. This means augmenting or, in some cases, replacing human inspectors with AI-enabled visual inspection. This increases accuracy and shortens the time for inspections, reducing recalls and rework and resulting in significant cost savings.
Eliminating repetitive tasks and processes to increase worker productivity.
Getting started with AI in manufacturing
It’s one thing to understand the potential value AI can have in manufacturing; it’s another to actually implement it. Experts offer the following advice:
Focus on data. Compared with high-value AI initiatives in other industries, manufacturing use cases tend to be more individualized, with lower returns, and thus are more difficult to fund and execute.
An alternative to a custom-built AI solution is a data-centric vertical AI platform, which can facilitate specific use cases. For example, an automated anomaly detection tool could replace or augment human workers who are tasked with quality control.
“Instead of focusing on coding the right algorithm, data-centric AI is a systematic effort to get good data to train AI systems,” Kai Yang, vice president of product at Landing AI, said at the recent EmTech Digital conference hosted by MIT Technology Review. “This lets users who have solid domain knowledge prepare good data sets for the AI model to be trained on without having deep machine learning knowledge.”
Prepare for organizational change. Though there’s been a lot of talk about AI taking over humans’ jobs, widespread use of AI will create the need for new roles and operating models. If companies are going to rely on AI-generated insights, there will need to be a human layer that systematically governs data quality and automation results. “We are going to have to do a lot of organizational redesigning,” Kothiyal said.
Start with experimentation, keeping an eye toward ROI. It’s still early days, but things are moving quickly. Manufacturers should start applying generative AI or other technologies to targeted initiatives to learn, develop skills, and secure early wins that can be used to build organizational momentum and gain buy-in. “It’s about bringing knowledge into the organization about how to use and implement AI,” MIT Sloan professor said at the MIMO Symposium.
At the same time, AI initiatives have to be more than just a learning exercise. “Ultimately, it’s about making money,” Kothiyal said. “If you can showcase the ROI, you get the budget, and that’s when you can start doing something more experimental and higher risk.”