Artificial intelligence (AI) is becoming the gold standard in industry. Thanks to their consistency, accuracy and efficiency, AI-based technologies are increasingly complementing human work in industry and accompanying processes from automation to quality assurance. However, artificial intelligence is only ever as intelligent as it has been trained to be, meaning that an AI system, like all intelligent systems, must first learn its tasks with the help of training data. In addition to real data, synthetic data is often used for this purpose.

Training data for AI-based systems

AI systems, such as the AI.SEE™ optical inspection system, are based on self-learning deep learning models that learn the relationships between the various data sets using state-of-the-art artificial neural networks. Just like a human brain, the system first has to learn what exactly it should and should not recognize as an error or irregularity, and the same applies to artificial intelligence: practice, practice, practice!

The more experience the artificial intelligence has, the better it can make the right decisions in the application.

As an optical inspection system in the manufacturing industry, Al.SEE™ must be flexible enough to reliably inspect cracks or weld seams on different materials, for example. To do this, it needs the right sensitivity to detect even the smallest cracks and changes and the right specificity to detect differences that are only due to the material but do not show any irregularities.

To train the algorithm, the deep learning models must therefore be fed with as much data as possible that represents all scenarios that could occur in the actual application. This means that data is required that reflects the breadth of use cases, as well as data that represents the standard case and only differs by the smallest changes in detail.

AI.SEE Solar visuelle Inspektion mit KI
Visuelle Inspektion von Solarpanelen durch KI mit AI.SEE™

Synthetic image data for training optical AI systems

Although real image data is good and reliable training data for optical artificial intelligence, it is often very expensive. Production costs, one-sided perspectives and data protection regulations of real images lead to low availability and high costs. The creation of real images is therefore disproportionate to the training resources required by an AI-based system. This is where a large area of application for synthetic images can be found.

Zudem müssen AI-Systeme in der optischen Prüfung auf alle möglichen Fehler trainiert werden, um diese im Ernstfall zu erkennen. In der Realität entstehen jedoch sehr selten fehlerhafte Bauteile, in der Luftfahrt zum Beispiel nur ca. alle 12-14 Monate. Doch besonders in diesen Bereichen ist eine zuverlässige AI zur optischen Bauteilprüfung unerlässlich und muss auf solche Fehler trainiert werden. Durch das seltene Auftreten von Produktionsfehlern, gibt es also auch einen Mangel an Fehlerbildern, die für das Training optischer Prüfsysteme essenziell sind. Hier bieten synthetische Bilder, die bestimmte Fehlerausprägungen darstellen, ein wertvolles Tool.

Neue Chancen mit synthetischen Bilddaten

Neuronale Netze können nicht nur Bilddaten verarbeiten und Anomalien erkennen, sondern auch selbst künstliche (synthetische) Bilddaten generieren. Bei elunic nutzen wir CAD Daten, um neue, synthetische Daten zu generieren. Die computergestützten 3D Modelle werden verwendet und mithilfe eines GAN Algorithmus so lange trainiert, bis sie von Originaldaten nicht mehr zu unterscheiden sind. Nach diesem Fine-tuning können die synthetischen Daten gerendert werden und als Datensatz für weitere KI-basierte Prozesse wie Training oder Tests verwendet werden.

Synthetische Daten: Mithilfe von GANs von realen Daten nicht zu unterscheiden

Zur Erstellung synthetischer Bilder werden meist General Adversarial Networks (GANs) verwendet. Hierbei lernt der Algorithmus mit einem großen Datensatz die wichtigsten Strukturen und Merkmale von Bildern. Das vom sogenannten „Generator“ erzeugte Bild wird an ein zweites System dem „Diskriminator“ gesendet, der neben den synthetischen Bildern auch mit realen Bildern (sog. Originaldaten) gefüttert wird und versucht das reale Bild vom synthetischen Bild zu unterscheiden. Ziel des GAN Systems ist, das die Bilder vom Generator nicht mehr von Originaldaten zu unterscheiden sind. Hat man dies erreicht, hat man zuverlässige synthetische Daten, die weiterverwendet werden können.

Vorteile synthetischer Daten

Mit dem wachsenden Bedarf an Daten, gehören synthetische Daten zu einem der wichtigsten Themen im KI-Bereich. Nicht nur kann der Datenbedarf nicht mehr in der nötigen Geschwindigkeit real generiert werden, auch Kosten und Themen wie Datenschutz spielen bei dieser Entscheidung eine große Rolle. Ein synthetischer Datensatz kann AI-Systemen die nötigen Bilder und Informationen liefern, ohne auf Datenschutz oder Einwilligung von Urhebern achten zu müssen. Dieser neue Weg Daten zu generieren und KI-Systeme zu trainieren bringt einen entscheidenden Vorsprung für Unternehmen in der produzierenden Industrie.

Mit langjähriger Erfahrung ist elunic Spezialist beim Generieren and Implementieren synthetischer Daten und unterstützt Sie gerne auf Ihrem Weg zu den optimalen synthetischen Daten für Ihre AI.

Showcases.

Your path to automatic visual inspection with elunic

Consulting

Counselling

Advice on business areas, procedures and technical challenges.

Development

Architecture, programming and maintenance of customised software solutions.

Integration

Integration of the most important platforms and on-premises solutions.

Get in touch.

Do you have a project enquiry? Please do not hesitate to contact us - we will be happy to answer you promptly.

    You are:*Mechanical engineerMachine operatorOther
     

    *By using this form, you consent to the storage and processing of your data by this website.

     

    Munich Map

    elunic AG

    Erika-Mann-Str. 23

    80636 Munich

    Germany

    089 / 4161737 - 30

    info@elunic.com

    EN

    shopfloor.io Logo

    Register now for a free demo

      *By using this form, you consent to the storage and processing of your data by this website.

      shopfloor.io Logo

      Register now for a free demo

        *By using this form, you consent to the storage and processing of your data by this website.

        shopfloor.io Logo

        Register now for a free demo

          *By using this form, you consent to the storage and processing of your data by this website.

          Download now for free

          shopfloor.io Logo

          Register now for a free demo

            *By using this form, you consent to the storage and processing of your data by this website.

            With the shopfloor.io asset ping, you can connect existing machines in the field, record and visualise machine data, distribute digital services and increase your after-sales.

            Download now for free

            shopfloor.io Logo

            Register now for a free demo

              *By using this form, you consent to the storage and processing of your data by this website.

              shopfloor.io Logo

              Register now for a free demo

                *By using this form, you consent to the storage and processing of your data by this website.

                shopfloor.io Logo

                Register now for a free demo

                  *By using this form, you consent to the storage and processing of your data by this website.

                  shopfloor.io Logo

                  Register now for a free demo

                    *By using this form, you consent to the storage and processing of your data by this website.

                    shopfloor.io Logo

                    Register now for a free demo

                      *By using this form, you consent to the storage and processing of your data by this website.

                      shopfloor.io Logo

                      Register now for a free demo

                        *By using this form, you consent to the storage and processing of your data by this website.

                        shopfloor.io Logo

                        Register now for a free demo

                          *By using this form, you consent to the storage and processing of your data by this website.

                          shopfloor.io Logo

                          Register now for a free demo

                            *By using this form, you consent to the storage and processing of your data by this website.

                            shopfloor.io Logo

                            Register now for a free demo

                              *By using this form, you consent to the storage and processing of your data by this website.

                              shopfloor.io Logo

                              Register now for a free demo

                                *By using this form, you consent to the storage and processing of your data by this website.

                                shopfloor.io Logo

                                Register now for a free demo

                                  *By using this form, you consent to the storage and processing of your data by this website.

                                  Register now for a free demo

                                    *By using this form, you consent to the storage and processing of your data by this website.

                                    After the first screening, we will send you an invitation to our Talentcube tool. We have prepared 3-5 short questions that you can answer in a personal video (we have also prepared a short introduction video). If you don't feel comfortable with this, please let us know - we'll find an alternative!

                                    We want to get an authentic first impression of you and at the same time have the opportunity to review your video with our team and make the right decisions quickly.