In a typical visual-inspection task on the production floor, a human inspector is responsible for spotting small defects on up to thousands of products on production lines. This redundancy inevitably leads to errors: A study by Sandia National Laboratories found that human inspectors miss 20-30% of defects across multiple types of inspection tasks. This results in product recalls, soured client relationships and damaged reputations.
Traditional machine vision isn’t perfect, either. Typical rules-based algorithms have not succeeded in spotting defects on complex, objects that have acceptable variations. Rules-based deep learning systems can therefore classify a good part as defective or vice versa and can be prohibitively complex. They can also be costly, requiring engineering time to optimize models to acceptable levels, and many deep learning techniques necessitate large amounts of data before training — another time consuming, expensive task.
Elementary’s AI-powered vision platform departs from rigid, rules-based training. Instead of relying on a highly trained, high-cost operator to code instructions into a vision system, manufacturers can use Elementary’s configuration without knowing any code at all. It is also highly scalable. Manufacturers can place cameras across factory floors — or in locations around the world — and view a data roll up in one cohesive dashboard.
Because defect requirements are often ambiguous, inspectors may disagree on what constitutes a flaw. These discrepancies lead to inconsistent inspections. Elementary’s solution takes the manual guesswork out of the inspection process and allows an IoT connected, cloud-based machine learning system to do the work. Unlike traditional, rules-based machine vision algorithms, Elementary’s AI-based platform can solve many typically challenging problems, with intuitive labeling capabilities that reliably capture input and provide datasets for training models.
Small data, or anything less than 100 images, has typically held back AI visual inspection models, but this is where Elementary excels. Elementary’s solution can train models from as few as 20-30 images.
Elementary’s solution captures images via a production line-based camera. It securely sends those images from an edge computer — or a computer that lives on a production line and is linked to the camera — to the cloud. There, a data set is created, with as few as 20 images.
A human operator trains machine learning models on these cloud-based images, using simple pass/ fail selections. The operator does not need much instruction to conduct these selections, which guide the model to detect anomies including incorrect label placement, scratches, blemishes, overspray, or excess adhesive, among other defects.
Because the solution is so easy to use, manufacturers don’t need to employ expensive, specialized operators to train these algorithms. They can also conduct this work from home, which holds enormous promise as today’s workforce becomes increasingly remote.
Finally, the trained model returns to the edge computer, where it functions quickly and autonomously to alert operators via SMS or email anywhere in the world of issues. It can also connect to a programmable logic controller (PLC) to automatically reject products and remove them from the production line.
As this system collects and connects all scanned data and feeds it into a dashboard, managers can see the health of their operations at a glance. They no longer need to worry about wasting time collecting information piece meal, as they would with most modern solutions that silo data at inspection points.
Recent breakthroughs in AI have significantly heightened computer vision’s capabilities, empowering machines to achieve tasks that were previously not feasible. Elementary’s technology exemplifies this, as it helps manufacturers to either enhance their existing automated solutions or build them from the ground up, especially in highly variable, complex situations. In cases like these, it is best to employ a vision system that sees as well as thinks.