According to a recent survey by the digital association Bitkom, although almost three quarters of the companies surveyed consider artificial intelligence to be the most important technology of the future, only 6% of respondents are currently using it. The situation is different for start-ups, almost half of which (47%) are already using artificial intelligence.

In principle, AI-based optimization potential can be found in all departments and value chains of a company. However, in the area of human resources, for example, the use of artificial intelligence is being held back by the enormous need for anonymized training data. In the area of quality and final quality control, on the other hand, these restrictions do not exist - quite the opposite.

Here, however, the seemingly infinite error variance ensures that automated machine vision processes of the first generation have so far only been used for comparatively undemanding inspection tasks.

Rule-based machine vision solutions

In many cases, image processing software is already in use for quality control.

© Deevio

Purely rule-based processes are ideal for checking clearly defined properties programmed by the system integrator, detecting deviations and sorting out faulty parts. If a workpiece must be exactly 200 mm long, 100 mm deep and 100 mm wide in order to be further processed, these dimensions can still be checked quite easily and with an acceptable pseudo-error rate using a rule-based solution consisting of high-resolution cameras and image processing software.

Nevertheless, even the inflexible distinction between 'good' and 'bad' requires relatively complex programming, which then reaches its limits even with slightly different positions of the workpieces or other lighting effects. And even if the performance and processing speed of hardware and software components have increased rapidly over the past few years, rule-based processes have still not made the leap into end-of-line control due to their systemic weaknesses.

When it comes to more complex and variant-rich defects, highly qualified inspectors who rely on their good eyes and many years of experience are still predominantly used in final quality control. Rigorous quality criteria must be adhered to, especially in the error-sensitive pharmaceutical and automotive industries. However, this approach also has a number of disadvantages: On the one hand, even with fit and motivated inspection personnel, the respective form of the day has a negative effect on the error rate - for example, people who are already slightly dehydrated no longer look as closely - and on the other hand, there is a notoriously high demand for reliable employees in this area, which can no longer be met in rural regions. In addition, a manual visual inspection always takes a certain amount of time. Until recently, however, companies simply had no other choice if they did not want to soften their quality criteria or accept cost-intensive reject rates.

Next Step: Deep Learning Machine Vision

Industrial cameras and artificial intelligence are to automate control processes.

© Deevio

The next generation of machine vision solutions, on the other hand, impressively demonstrates its potential. Information processing is based on artificial neural networks that are taught control rules and continuously improved with training data. These systems are able to achieve a degree of accuracy of over 99% - and maintain it permanently. In combination with a high processing speed and manageable costs, this error rate of the deep learning method secures the entry ticket to the demanding pharmaceutical and automotive final quality control.

The advantages of the deep learning method can already be illustrated using a comparatively simple workpiece: In the automotive industry in particular, the focus on customers' individual equipment requirements has meant that a stamped sheet metal part can exhibit an immense range of variations and colors after just a few further processing stations. However, the smallest scratches, dents or color clouding simply cannot be reliably detected using rigid rules. If the shape also has a certain degree of complexity, even minimal fluctuations in lighting can lead to rule-based processes either detecting a pseudo defect and rejecting the workpiece - or incorrectly approving it. If safety-relevant parts are involved, in the worst case, human lives are at stake.

Step by step to the smart factory

Damian Heimel is COO of Deevio in Berlin.

© Deevio

As a pioneer in the field of AI-based automated final quality control, the Berlin-based company Deevio, for example, takes a phase-oriented approach: The first contact can be made simply by sending in a few images of the test pieces in question. The specialists then ask about other parameters such as volume, previous defect and pseudo-defect rates and variance in order to estimate the potential for optimization. The next steps are an on-site appointment and the creation of a feasibility study. Existing image databases and the software and hardware infrastructure are included in the analysis and the extent to which they can be integrated into the final solution is determined. A positive result then opens the door to the proof-of-concept phase. Here, the deep learning machine vision model is not yet switched to productive operation, but instead expands its image data pool almost incidentally and is trained further by the specialists: Artificial intelligence and trainers work together 'hand in hand'. Initially, the feed-in of images of faultless and faulty workpieces under the supervision of data scientists provides the basis. As the process progresses, more and more images from ongoing production are used in order to keep the error rate below 1%. During this optimization phase, which can take up to a few weeks, the deep learning solution ('AI box', consisting of a mini-computer and an application-optimized graphics card) provides support. When it is introduced into the production line, the deep learning solution is then able to abstract the learned defects and errors - and apply them to new products that it has not yet 'seen'.

The time savings alone are quite impressive: Where even comprehensively qualified, experienced and highly focused inspectors still need up to 15 s to check a workpiece for defects, Deevio's solution gives a 'Go!' or a 'Stop!' in less than 1 s. As soon as a deep learning machine vision implementation in the sensitive area of final quality control continuously delivers reliable and high-quality results - and also continues to optimize itself - the acceptance of the use of AI-based solutions increases throughout the company.

The experts at Deevio aim to achieve a permanent error rate of well below 1%. To achieve this, they are in regular contact with those responsible for production on site and are constantly making improvements, for example in the areas of camera positioning and lighting technology, but are also working on the further development of artificial neural networks.