AI Comes to Factory Quality Control
Among the challenges of “re-shoring” manufacturing from China are the vagaries of consumer demand. For example, demand during the pandemic for personal protection gear and toilet paper has far outstripped supplies. Previously, North American manufactures would spend heavily to ramp up production only to find demand disappearing after several months.
The long-term solution for manufacturers, of course, is stable demand. In the meantime, new tools are emerging to help producers cope with demand swings while turning on a dime to introduce new products that will keep manufacturing lines humming.
AI vendors are betting automation technology can play a role as more manufacturing and assembly operations are re-shored in response to supply chain vulnerabilities exposed by the pandemic. Deep learning specialist Neurala, for example, unveiled automation software this week designed to help manufacturers improve product quality inspection on production lines.
The Boston-based company insists its approach also would help manufacturers scale production in response to shifting consumer demand.
Thus far, AI tools have proven too costly and complex to scale on the factory floor. Neurala and other vendors are betting the pandemic will not only boost regional manufacturing hubs but also boost requirements for factory floor automation. Automating quality inspections via computer vision is seen as an early AI application that can scale along with fluctuating demand.
“Manufacturers are facing irregular patterns in consumer demands, and heightened pressures on machine utilization, production efficiencies and quality control—and they need to address all of this with fewer people on the factory floor,” said Max Versace, Neurala’s co-founder and CEO.
The company’s vision inspection automation software targets manufacturers with little previous experience with production line automation. Hence, Neurala emphasizes ease-of-use and the ability to quickly train its vision platform that is based on its “lifelong” deep neural network technology.
The approach, battle-tested on NASA’s Curiosity Mars Rover, is billed as allowing neural nets to become the brains of a device like an inspection camera rather than learning through on-device inference. That framework is touted as reducing the amount of training data and the time needed to train a system.
For industrial applications, Neurala said its visual inspection software can train and run multiple AI models while operating with existing hardware on the factory floor, including widely used GigE industrial inspection cameras and factory touchscreens.
Along with boosting product inspection rates and reducing human intervention, the vision software is promoted as helping manufacturers inspect smaller batches as demand fluctuates.
Early industrial automation deployments relied on Internet access or cloud deployments to connect devices like inspection cameras. As agile production and assembly lines return to North America, Neurala is also promoting the option of keeping production and quality data in-house as domestic competition grows.
Given slow adoption rates for factory AI, vendors like Neurala must also convince manufacturers they will get a quick return on investments in automation. Neurala’s pitch includes an “anomaly recognition” feature that addresses the paucity of defective product images. The relative lack of data on product defects among “high-yield” manufacturers creates a use case for factory-floor AI, the company argues.
The anomaly detection framework spots any product that deviates from an “acceptable” image used to train inspection systems. Hence, the argument goes, producers could roll out AI-based vision systems without having to specify multiple features used to detect defects via traditional machine vision systems.