Feature: AI
Sparse modelling: perfect for embedded systems
By Zeljko Loncaric, Marketing Engineer, Congatec T
he nature of the factory floor is that conditions are constantly changing, which is proving challenging for inspection systems,
prone to false positives. They also produce errors when re-calibrating to new environments and conditions, requiring the assistance of human inspectors. But, these resources are rare, and the manual inspection of flagged defects wastes time and money, which makes artificial intelligence (AI) an attractive proposition. AI will certainly help, but is not a
panacea since it requires massive training data. Such data is currently rarely available, since quality production is not programmed to produce massive quantities of faulty parts to teach AI systems. To make matters worse, the computing power required to train AI cannot be hosted at each inspection system. One AI approach that can solve this
problem is Sparse Modelling, which recently evolved from the academic space and entered the business world. This new kind of AI needs only a small
data set, resulting in both training and inference algorithm execution tasks fitting into ultra-low-power embedded computing platforms. Test kits are already available, ready to run standalone or as add-on to existing vision systems to analyse pictures identified as ‘not good’.
Improved performance? Artificial intelligence has great potential to improve the performance and accuracy of modern visual inspection systems. But conventional AI approaches have some drawbacks: • Vision-based deep learning must process every detail of a picture to be able to provide reliable results. This is power- and compute-intensive, with masses of data moving between processor and memory. In one hour, a 60fps camera with UHD resolution and 8-bit colour depth can produce up to 5.18 terabytes of uncompressed data that needs to be analysed.
• Another problem is that AI needs a massive number of AI-classified pictures to make reliable predictions. This also consumes many clock
54 November/December 2020
www.electronicsworld.co.uk
cycles and power. Recent studies have indicated that a single AI model based on deep-learning technology pollutes the environment up to as much as five cars during their entire lifecycle. Local embedded systems cannot provide such computing performance, only data centers can. This could be a reason not to work with AI, but even if we take this into account, and want to leverage such a system because wastage from production failures is even more expensive and harmful to the environment, there is another critical pitfall with conventional deep- learning-based AI:
• In reality, the required training data is not available within a few days or weeks. It can take a year or longer to collect images of thousands of defective parts from a production line. If the production process is modified during that year, and the definition of what is good, or the nature of possible defects changed (even slightly), will make any previously-gathered data virtually useless.
• Conventional AI always needs server- grade training via deep learning
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68
