Search

Interview Automation of End of Line Quality Control

| Editor: Nicole Kareta

Deevio, a start-up in the field of automated quality control, presented its machine learning software to interested visitors at EUROGUSS 2020 by means of a real use case. Damian Heimel, Co-founder and COO of Deevio, explained the role and functionality of this technology in an interview.

Related Companies

Cracks, impact points, blowholes and defects such as porosity are generally difficult to define.
Cracks, impact points, blowholes and defects such as porosity are generally difficult to define.
(Source: gemeinfrei / Unsplash)

SPOTLIGHTMETAL: What exactly does Deevio do?

Damian Heimel: Deevio automates visual inspections in end of line quality control by means of machine learning technology. Hence the name Deevio, which is a combination of the terms 'Deep Learning' and 'Vision'.

How did the Deevio company develop?

Damian Heimel: Deevio was born out of the Company Builder ‘WATTx’ of the Viessmann Company. The aim of this venture builder is to promote the establishment of young industrial companies in order to introduce innovations to the market.

SPOTLIGHTMETAL: What niche does Deevio fill in the market of quality control?

Damian Heimel: Deevio automates visual inspection tasks. These kinds of visual inspections at the end of the production line are often still done manually. Quality control on large production lines is often already automated by cameras, robots and the like, but the very end of quality control or quality inspection is a highly manual task. This is because up to now this area has been difficult to cover with automation technology. That is due to the fact that defects are very specific and highly variable. They can vary greatly, making cracks, impact points, blowholes and defects such as porosity generally difficult to define. For instance, a 3 mm long crack is often still acceptable, while a crack of 4 mm length is no longer acceptable. This is precisely where our machine learning software for image recognition provides new possibilities for quality control.

SPOTLIGHTMETAL: How does Deevio's machine learning Software work?

Damian Heimel: Our software uses sample images to learn. The customer shows us sample images of cracks that are still OK - these images are added to the "OK folder". This is similar to working on a PC desktop. Pictures with more serious defects are put into the "defect folder" in reverse order. Thus, the software obtains its intelligence level from this input and from the expert knowledge of employees who have been carrying out quality control for many years, for example in a foundry. A particularly exciting aspect is that the software - as the term "learning" implies - continues to learn. This means that we can always take new images from production and feed them into the learning process. As a result, the software gets better week by week and learns in the same way as a human being. Similarly, new defects can easily be taught with new images. All the customer has to do is provide new images with this defect and we will install the new update. In short, Deevio's machine learning Software is characterized by the fact that it can handle a high variability of defects and learn continuously. The special software has the additional advantage of featuring an algorithm that creates a very high level of detail by combining several images in one frame. This means that reflective surfaces, especially on aluminum parts, are no problem at all.

SPOTLIGHTMETAL: Why is automation so important in quality control?

Damian Heimel: A company produces several thousand parts per day. In the manual process, one person has to pick up the manufactured parts several thousand times and assess its quality. I have tested it myself and my conclusion is: A person can repeat this process for several hours, but at some point, the eyes get tired, you get tired and you get hungry. In addition, people feel different each day, which makes it impossible to detect all faults. So, a person cannot perform such a repetitive task to one hundred percent. However, the end of line quality control concerns parts that are shipped directly afterwards. This means that defect parts can slip through, which are then sent to the customer. Especially if a company delivers just in time, it cannot afford this kind of error. So, the alternative is to hire more people to carry out further inspections, or to automate the process altogether. Another aspect relates to the increasing age of the workforce. People who have been working in quality control for 40 years will retire at some point. This also means that their experience will be lost. Especially in structurally less developed areas it is almost impossible to replace these employees. It is precisely these aspects that call for automating a specific process. In addition, it is also possible to introduce a company-wide standard once and for all. After all, if four different people assess a part, four different results are obtained. Automation offers the opportunity for a company to define unique standards for parts that are still OK and those that are no longer OK. We can teach these standards to the program.

Gallery

SPOTLIGHTMETAL: So, what would this process look like in a foundry?

Damian Heimel: Let's imagine a flat casting whose surface is to be inspected. In this case, we would visit the customer with the appropriate hardware - consisting of a suitable camera and lighting and all the necessary setup - and take pictures of the products. The foundry provides us with 100 parts that are OK and 100 parts that have an impact point, 100 parts that have a blowhole and 100 parts that have porosity. We take pictures of these items and archive them in the respective category. Based on these pictures, our machine learning experts in Berlin create a model that is optimized for exactly this application. Once we have done this, we return to the foundry and have the system assess 200 parts that it has not yet seen. An employee from the quality assurance department checks whether the system's assessment is correct. In this way we can judge how well the system can predict the quality. If there is a deviation between the system and the quality inspector, we further improve the model. The next step is automation. For this we work together with partners such as system integrators and automation companies, who then build a feeding system or conveyor belt and enable the rejection of defect parts. This automates the entire end of line quality control process at the customer's site.

SPOTLIGHTMETAL: How is the proper functioning of a machine learning software ensured? Is it possible that a defect part is classified as OK?

Damian Heimel: There is, for example, the possibility of setting the system a little more strictly. In this case, no bad parts will slip through, but it can happen that a good part will be sorted out. In this way, you make sure that no bad part slips through. Another measure we take is to include all other new images taken during production in the training process, update the model again and again and thus ensure that the initial 100 images have become 5,000 images after two months. This increases the accuracy further. Of course, there is also the possibility to monitor the model. Then, we would transmit the data to our company and as soon as an anomaly occurs, our experts would take a closer look and together with the foundry analyze what exactly is happening there.

SPOTLIGHTMETAL: What is the biggest challenge in developing a machine learning software?

Damian Heimel: The biggest challenge is clearly to create accurate data. It is therefore extremely important that the faults we teach the machine are actually there. So, if we tell the machine that a defect is a blowhole, we have to make sure that it is actually a blowhole and not a sand grain inclusion. This requires the highest level of accuracy in image acquisition. A second challenge that arises from time to time is the availability of parts. And of course, machine learning algorithms are a bit more robust than traditional image processing. But we still need a stable image capturing area. That means both good lighting and a good camera are important. Simply installing a webcam wouldn't work.

SPOTLIGHTMETAL: Why is Deevio focusing on the automotive, foundry and pharmaceutical industries?

Damian Heimel: We have noticed that especially the automotive and foundry industries still carry out a lot of manual inspections. In addition, foundries often produce very high volumes, which means that we have access to the required data volume and data quality. In addition, Germany as an automotive location and the important role that foundries play there represent a large market. The main reason, however, is that our new technology opens possibilities that were previously not possible with standard image processing. Regarding the pharmaceutical industry - that's where we started our first project. We received this order even before we started our company. In this industry, every company already uses image processing solutions frequently. However, there is often a high level of pseudo rejection - which means that the systems are either set too strictly or cannot detect some faults properly. This can be better solved with machine learning software. In such cases, we only take another look at the faulty images. So, by getting all the images of faulty parts, we establish a model and can then provide the company with a tool that shows exactly what kind of defects are being produced today. We see many opportunities in the pharmaceutical sector, and we are already working with several companies.

SPOTLIGHTMETAL: Thank you for the interview, Mr. Heimel.

This article is protected by copyright. You want to use it for your own purpose? Contact us at support.vogel.de (ID: 46386494)