Neural networks are like puppies: you have to feed them and train them. They are currently viewed as a miracle weapon for a wide variety of tasks and issues. If you feed them enough data and train them long enough, the desired solution will come out in the end.
Our approach is somewhat different. We, EDI GmbH - Engineering Data Intelligence, combine the strengths of humans with the strengths of machines. With our experience and our AI tools, we support companies from very different industries in optimising development and production processes through digitalisation. Among other things, we also offer the automated analysis of technical drawings.
With us, human logic interacts with the machine learning of AI. Humans and machines learn hand in hand. Our focus is on using the neural networks as efficiently as possible and reaching the goal with little data and short training times. The advantages of this approach are short development times with little training data and a responsible use of resources.
The quality of the data is crucial
The example of technical drawings analysed by AI for the automated creation of quotations makes this clear: the neural networks are strong when the drawings are available as grey values and aligned straight. Slightly twisted drawings and RGB colour are a challenge for the networks. With appropriate logics based on these human insights, the machine learning of the AI can be supported or facilitated.
Pre- and post-processing of the source data play a crucial role in the effectiveness and quality of the results. Here, too, humans and machines work closely together: The human considers what the machine needs in order to find the best possible data basis. Furthermore, our employees think about how the output of the AI can be turned back into a human-understandable fact.
Many small speedboats instead of one big tanker
The analysis of the technical drawings by the AI takes place in different phases. Which type of network is used for which task is decided by our machine learning team. Instead of one big neural network trying to do everything, in this task we use several neural networks with very specific tasks. These networks are precisely tuned to each other and connected in series. As a result, this task is solved very effectively and efficiently.
Being able to react quickly to changes
If there is a change to a technical drawing, this is noticed by our microservices that communicate with each other. The changes are forwarded directly and stored in a database. Our state-of-the-art microservice architecture forms the backbone of our AI build pipeline.
Perforated like Swiss cheese or secure like a Swiss bank vault?
Security is the top priority in all our projects, including when analyzing the technical drawings. That's why we opted for the bank vault. In order to exclude errors and ensure the optimal quality of the software, which solves the customer-specific problem with artificial intelligence, we use a cascade of multi-stage tests. These tests are structured according to the so-called cheese hole theory: if an error slips through a hole in the first level, the error is caught in the subsequent levels by the further tests.
The developers' manual tests are followed by automated unit tests. Then a senior developer looks at the changes to the code in the merge request. This is followed by automated integration tests and Selenium tests, which ensure that the software continues to function as intended. At the end of the quality control process is a UX quality gate, where team members from different teams not involved in the development review the functionality.
These overlapping tests each cover different aspects and in their entirety cover the widest possible spectrum. In this way, our process allows us to offer a fast development time and a high quality of the software we custom-develop for our customers.
