Visual product inspection at the end of production is considered the standard for effective final quality control: Here, the use of artificial intelligence (AI) and edge computing paired with state-of-the-art camera systems is increasingly establishing itself as the new technological standard for visual quality inspection. In recent years, the technology has made considerable progress in terms of development, miniaturization and cost reduction. Surfaces and structures as well as color variations can now be detected with a precision that far exceeds human capabilities. For example, infrared cameras make it possible to accurately detect surface temperatures down to a tenth of a degree, while microscopy cameras can identify tiny defects of less than a thousandth of a millimeter. In addition, systems such as X-rays or ultrasound can now even analyze defects within a component and analyze them at the edge.

AI and the edge are replacing the human eye

In many companies, people are still responsible for quality inspection. However, AI-supported video analytics (also known as computer vision) is one of the areas of application in which an automated inspection process simplifies and improves quality control and frees skilled workers from monotonous, repetitive work processes. After all, quality defects can cost a manufacturing company up to ten percent of its turnover, according to a study by the AFNOR Group, a French organization for standardization.

So how exactly does Computer Vision register errors? The AI analysis tool recognizes anomalies in the visual data that deviate from the standard and thus indicate damage or defects. In this way, the technology distinguishes compliant from defective components or surfaces.

To make this distinction possible, the AI algorithms have to be trained with a large amount of image data. These data points train the AI model to categorize defect-free information as well as information that indicates defects. In this way, the systems are geared towards specific quality defects, such as the detection of paint runs, micro-cracks in components or missing parts. This creates an AI model for production that manufacturing companies can use for real-time data analysis in their production facilities. The extensive training infrastructure is no longer required for the productive operation of a pre-trained model.

The technology can be used in almost any industry: whether it's paint scratches, dents or incorrectly installed components in the automotive industry, unclean solder joints on a circuit board in the electronics industry or color deviations or dented areas in the food industry - computer vision can theoretically reliably detect and report anything that deviates visually.

Why edge computing?

AI makes it possible to reliably make automated decisions based on data. But what role does edge computing play in this?

Edge computing refers to the processing of data directly at the source, i.e. in the immediate vicinity of the sensor or machine. By processing data at the edge, i.e. at the 'edge' of the network, latency times are minimized and the transfer of large amounts of data to the cloud is avoided. Direct transmission of video images to the cloud is unrealistic, primarily for cost reasons. This is why edge computing has become established, especially for real-time applications, including the industrial use of video analytics. With this concept, data processing is brought closer to the location where the data is collected, eliminating the need for transmission to a central data center in the cloud - which is dependent on parameters that are susceptible to interference. Instead, edge servers with sufficient computing power and storage capacity are used. Basic models consist, for example, of microcomputers that bundle sensors in industrial production and communicate via Industrial Ethernet or 5G mobile communications. At the other end of the spectrum are edge servers with video analysis capability and powerful graphics processors (GPUs).

Video analytics directly in the camera is itself rarely an option because the camera CPU alone is not powerful enough for advanced algorithms.

Another advantage of combining edge computing and AI in quality control is the ability to predict defects in production. Patterns and trends can be detected on the basis of continuous monitoring of products and thus indicate future deviations. In this way, computer vision warns in advance and enables companies in the manufacturing industry to take early action before a significant reduction in quality occurs.

Private networks for real-time video analytics

In theory, the detection of anomalies by sensors and cameras sounds obvious. In practice, however, transferring the real-time video streams from the cameras to the edge servers can be a considerable challenge. High-resolution cameras, which are essential for quality inspection, generate large amounts of data - a full HD camera generates up to 9 Mbit of video material per second. In addition, there are usually many cameras in use, which further increases the bandwidth requirement.

The author: Carsten Mieth is Senior Vice President of Eviden, an Atos Group company.

© Atos

In addition to cameras, edge servers and trained AI models for quality inspection, which are now also available on the market as out-of-the-box applications, computer vision therefore requires a fast - and above all very stable - data network. According to the Federal Network Agency, the expansion of high-performance public 5G networks is progressing, but more and more manufacturing companies are using private 5G networks that are independent of the public network and offer not only suitable bandwidths and latency times but also the highest level of data security. For real-time computer vision in industrial quality control, which requires maximum stability and data continuity, a private network with guaranteed quality of service is therefore currently the preferred choice.