29. November 2019, 11:30 Uhr | Nicole Wörner
Peter Krippner, Viscom: »One has to very selectively analyze at which points in assembly inspection meaningful and useful applications for deep learning, machine learning and artificial intelligence actually exist.«
The terms deep learning, machine learning and artificial intelligence are currently being used almost inflationarily. Based on the example of assembly inspection, we met Peter Krippner, Executive Board Member for Operations at Viscom, to discuss the practical benefits of these technologies.
Mr. Krippner, what do the terms deep learning, machine learning and artificial intelligence mean for you personally in the context of assembly inspection, and where do you see the most important differences?
The topics are not, in principle, differentiated - but they each represent different methods that can be used together in different applications. The generic term for all three is AI.
What opportunities do these technologies open up in assembly inspection?
Our customers are continually striving to operate their SMT lines with fewer personnel. On the one hand this has to do with general cost pressure, while on the other hand new technologies offer opportunities for quality improvement. With regard to assembly inspection, deep learning, machine learning and artificial intelligence are of particular importance in the verification of inspection results and programming of inspection systems. When verifying inspection results, a great deal of data, such as images from the field, is required in order to train so-called classifiers, which are then validated with additional images. Here AI can gradually take over more and more tasks: If the operator is initially supported by the AI results during verification, AI can then verify certain components automatically after successful validation. There are also applications for AI in the programming of systems, for example the automation of the component assignment and even in the evaluation of images for simple inspection tasks.
Where are these technologies currently being used?
The first pilot installations are already underway to verify inspection results. Many images and data were collected in advance to establish a basis for these field tests. AI thrives on such image data. To this end, the inspection system must provide high-quality images from all viewing directions, coupled with versatile lighting. Our XM sensor system with a combination of 2D, 2.5D and 3D provides an optimal basis for this.
Which are the challenges and where are the limits?
Collecting and structuring input data is a major challenge that almost inevitably leads to the subject of Big Data. Here you have to be sure you are on the right track in order to maintain an overview over the deluge of data. Last but not least, acceptance on behalf of the end user is important: AI is not perfect, so how should one deal with expected AI errors? There is a widespread misconception that measures taken as a result of errors will guarantee that the same error will not re-occur and without side effects. This is still very difficult in an AI environment.
What risks lie in the development towards deep or machine learning and AI?
Deep learning, machine learning and artificial intelligence are currently right at the top of the hype cycle. The risk lies in seeing the AI megatrend as the answer to all problems. One has to very selectively analyse at which points in assembly inspection meaningful and useful applications actually exist for deep learning, machine learning and artificial intelligence.
What place do these technologies have on your technology roadmap?
All three are top priorities on our roadmap. This is one of the reasons why we are not only active in these fields ourselves, but have also entered into technological cooperation with several customers and universities.
What does the future hold for these technologies - in general and in assembly inspection in particular?
They are already being applied in many areas of everyday life, for example, in language assistants and driver assistance systems. However, there is one big difference: Every AI result must be accompanied by a confidence factor, which indicates just how unambiguous the decision made by the system really is. For example, recognition reliability of 80 to 90 percent is quite sufficient for a satisfactory result with the most commonly used speech recognition systems. However, for inspecting solder joints in safety-relevant automotive electronics we need a significantly higher probability. This is the challenge we are currently facing. Furthermore, in assembly inspection, it may be possible for AI to perform tasks in addition to image evaluation of the inspection system. It is also conceivable that automatic evaluation of other system data could be undertaken - such as operating data from an axis controller or temperature data for condition monitoring or predictive maintenance.
This interview was conducted by Nicole Wörner, Markt&Technik