How a startup plans to revolutionize quality assurance

At the mechanical engineering summit, the expert audience awarded an acoustic AI solution that automatically detects defects in metallic components by tapping. This allows scrap to be reduced without complex methods like CT scans - starting from goods receipt inspection.

Originally, the founding duo came together in 2022 to actively solve production problems using AI. They started with a solution that measures fill levels in containers. Based on a specific customer problem with stainless steel drums and hazardous materials, where visual methods were not an option, the idea of using sound was born. The container was tapped, and the sound data was collected. The tapping is done by small linear motors, with a small "hammer" automatically extending.

"The collected data had significantly more potential than we expected. We can not only measure fill levels but also know if a drum is securely closed, if it has dents, or if it is intact," says founder and CTO Fabian Oberndorfer, who is also a passionate musician.

"For thousands of years, people have tapped on clay pots, for example, to hear if there are cracks in them. This experiential knowledge can now be mapped by an algorithm based on audio data," says Oberndorfer. The idea of using audio - to analyze, for example, whether a machine is operating within optimal parameters - has been around for a while. Here, algorithms learn from the experiential knowledge of machine operators, who already know from the sounds if the machine has a problem. 

Many companies have problems detecting defects in metal components

But the startup goes a step further: “By actively knocking, we can suddenly listen to objects that otherwise would not make any noise or provide any information,” explains the Resoniks CTO.

Through the partner Dassault Systèmes, the young company pitched its idea early on at the “Factory of the Year” event. “Our solution can measure many parameters by knocking. Thanks to the event, we received direct feedback from practice on where this could be helpful,” recalls Oberndorfer. The answer was surprisingly clear: In quality assurance, many companies apparently have great difficulty identifying cracks, pores, or voids. “This led to one of our first proof-of-concept projects for acoustic quality assurance with a large OEM,” reports Oberndorfer.

In the meantime, the startup has created a complete system in which the user - or a robot - only needs to insert the component to be tested. Conveyor belt constructions are also feasible. The sound is recorded, and the algorithms examine whether there is a defect in the component. Since the corresponding sensors are primarily built for laboratory environments, it was decided to develop their own hardware, including highly specialized microphones that can capture everything up to the ultrasonic range and resolve the entire spectrum from 20 hertz to 80 kilohertz with high precision.

Rittal is the factory of the year 2025

The winners of the production competition factory of the year have been announced: Rittal is the overall winner for 2025. What distinguishes the plant and who the other winners are.

Thanks to the mechanical engineering summit: Startup receives significantly more inquiries

On the recommendation of the VDMA, the startup participated in the mechanical engineering summit startup award this year - and emerged as the winner from the pitches. This has taken the young company, based in Holland and Finland, where 20 people now work, a step further. Much of what was hoped for from participating has already been fulfilled: "Since then, we have recorded significantly more inquiries and, among other things, were able to show an investor in the current financing round that there is a great demand for our solution," Oberndorfer is pleased to say. 

In projects with German industrial customers, a professionalism in collaboration has been repeatedly experienced that is not a given, says the trained mechanical engineer, who had already worked in product development for ten years before founding the company: One reason why it was decided to choose Germany as the most important core start market.

Significantly shorten lengthy quality assurance

Particularly great benefits can be generated through automated quality assurance, for example in the metal parts sector. This includes stamping systems, forging systems, CNC milling systems. The technology is also interesting for welding systems from the startup's perspective. Because especially with high-quality components, quality assurance is still very slow and laborious today.

“Ultimately, the part must be destroyed to test it; alternatively, a CT scan or X-ray is done, or a person must manually check and subjectively assess whether the part is okay,” explains Oberndorfer. The Resoniks system can also be directly integrated into their product by machine and plant manufacturers. This allows for checking during the manufacturing process whether the component is defect-free, thus avoiding waste. 

At the same time, the machine parameters can be adjusted live to produce more effectively. Oberndorfer sees another area of application in pre-sorting during incoming goods inspection, where in some projects, 70 percent of a batch of components are defective - which is often only discovered at the end of a manufacturing process.

The process is particularly worthwhile when highly complex, safety-relevant components are manufactured in sufficiently high quantities per year - starting from several thousand. This is the case in industries such as automotive, energy sector, aviation, and marine. Although the time savings vary depending on the case study, the automated method is 10 to 100 times faster. “There are components for which we can deliver a result within 20 seconds, where the only alternative currently is a twenty-minute CT scan,” says Fabian Oberndorfer.

Data as the pivot point

To evaluate in a statistically significant analysis whether the solution is suitable for the present problem, labeled data is initially required - on average about 100 good parts and around 40 parts with as many different defects as possible. The system primarily relies on training with good parts. "We essentially create an acoustic fingerprint of a good part and can then detect deviations. This means we can also identify previously unknown error patterns," says the founder. The data topic is often complex, as the assessment of "good and bad" in quality assurance is a rather rough, sometimes subjective gray area.

"In our practical projects, we have seen that it is always a bigger step from a successful pilot project to operation in the line," the founder also reports. Because at this point, topics such as integration and certifications are relevant. Therefore, the startup now offers a plug-and-play system for purchase or rental, which does not need to be deeply integrated into a line but can be placed next to it. "In collaboration with our customers, one pilot usually leads to the next. We often start with OEMs, who then 'pass us on' to their next supplier levels and ensure acceptance," reports Oberndorfer.

Usability and security as central prerequisites

In view of the shortage of skilled workers especially in the area of quality assurance, the startup placed a major emphasis from the beginning on the easy use of its solution. "In my opinion, software as a link is often greatly underestimated. Most of our employees are actually concerned with making the software as easy to use as possible and ensuring that the processes around data acquisition and labeling function as automatically as possible," states the founder.

In terms of safety and cyber security, a lot of thought has also been given. The cloud is only used for training the algorithms. After that, the trained model runs autonomously and is secured externally at the edge on an industrial PC. Communication with other machines is possible via an OPC-UA connection.

AI revolution in acoustic data?

If computing power continues to increase in the future and prices fall, it would also be possible to generate an acoustic fingerprint from synthetically obtained data. This would not only shorten the relatively complex labeling process, but the method would also be worthwhile with a small number of components.

Industry struggles with competence gaps in GenAI

Generative AI promises revolution - but in the industry, lack of know-how slows down implementation. New figures show: The reality is complex.

Anja Ringel

There is still as much potential for the technology as there was and is for image and text-based AI, believes the CTO. In the future, it could be quite normal for a phone to analyze by the sound of a tap whether a carbon frame has micro-cracks after a bicycle fall, whether the helmet is still intact, or what might be wrong with the heating. "Theoretically, you know if something is structurally intact as long as there is a reference to whether the object still sounds as it should. I find that very exciting," summarizes Fabian Oberndorfer.

By the end of the year, the startup's current funding round is expected to be completed. The next step is to focus on getting more pilots into the line to increasingly solve more complex problems. There are also plans to work more closely with mechanical engineers who want to integrate the system into their product. In the long term, the startup is also looking for partners in the machinery and plant engineering sector who can build parts of the hardware, such as large measuring devices.

Edited by Anja Ringel.