DE

‘AI offers us a very powerful tool’

Manufacturers are gaining greater insights into industrial laser machining processes by deploying different, increasingly precise sensors. 4D Photonics GmbH from Isernhagen near Hanover is approaching multi-sensor technology with a single sensor: More than 42 individual sensors with different physical properties deliver one million samples per second. With the help of artificial intelligence (AI), it is possible to identify errors and their causes in ongoing processes. In this PHOTONICS interview, the company’s CEO Christoph Franz talks about the potential that multi-sensor technology and AI present for monitoring and understanding laser processes, about fields of application, and about an elaborate and enlightening series of experiments at DESY, an electron synchrotron research facility in Germany.

Mr. Franz, would you like to briefly introduce your 4D Photonics GmbH?

Christoph Franz: With pleasure. We are an independent supplier of systems for monitoring laser processes that are optimized for each application. With 30 employees – more than two thirds of whom are engineers – we develop and produce sensors that detect and evaluate optical radiation in a wide range of laser processes, from welding and cutting to additive manufacturing, including powder and wire deposition welding, as well as the cleaning and finishing of surfaces. In our first 25 years, we have put more than 2,000 systems into operation for more than 100 customers and now sell our solutions via a global partner network.

You presented a new type of multispectral 4D sensor technology at LASER World of PHOTONICS 2023. Can you explain what that is?

Franz: Our approach is: “One sensor for any application.” Our 4D.TWO sensor integrates more than 42 individual sensors with different physical properties in a very small installation space and can be coupled directly to the processing optics of laser systems via an optical fiber. Despite its small edge length, measuring 91 x 63 x 79 mm, it captures a wide range of wavelengths on 32 channels, each with a sampling rate of six kilohertz. Thanks to our specialized technology, it can reliably detect deviations of around 40 micrometers at feed rates of 1,000 millimeters per second. The ability to record and read multispectra in a highly dynamic way isn’t just used to optically detect errors, however. Our 4D sensor also records acoustic emissions, humidity, air pressure and temperatures in addition to any third-party sensors. We offer suitable industrial software for this purpose and can combine our sensor technology with application-specific AI.

Where is your sensor being used?

Franz: There is already a wide range of applications for lasers. Coating brake discs with extreme high-speed laser application (EHLA) is one of them. As our sensor records more than 42 channels simultaneously, it delivers around one million samples per second. It makes sense that for every individual application, we adapt the selection and mathematical combinations of the evaluation channels accordingly. We carry out tests to find the optimum configuration of channels that best capture any process deviations. Among other things, we provide targeted laser power that fluctuates in milliseconds and varies in duration and intensity. Our sensor system registers even the shortest fluctuations. They are clearly recognizable in the signals. Our 4D sensor system can also detect blocked or worn nozzles based on the reduced flow rate.

We now understand exactly how the specific error patterns in the recorded signals can be assigned to the underlying faults. However, brake disc coating is just one of many applications: In seam monitoring during the welding of bipolar plates for fuel cells, the high feed rates of up to 1,000 millimeters per second mean that it is necessary to detect even the smallest dropouts. The seams must be 100% sealed. With several hundred weld seams per panel, that’s an ambitious expectation, and our solution comes very close to meeting it. The same applies to the reliability of process monitoring when welding copper hairpins for electric motors.

How do you ensure this reliability?

Franz: One way we addressed that was through our involvement in a series of experiments at the DESY, a German electron synchrotron research facility. It was literally a matter of X-raying a wide variety of ongoing laser machining processes in very high resolution. We brought in our sensor technology and synchronized it with these X-ray images. Ultimately, we found that all relevant errors and process deviations from the signals can be read from the signals of our sensors. With our multispectral approach, we have a solution that spans many laser applications – which, compared to a single optical sensor, offers a significantly higher hit probability for detecting errors thanks to the wide spectral range examined and the 42 channels that can be switched on and off individually. At the same time, our approach paves the way for these errors to be precisely classified...

... and this is where artificial intelligence comes into play?

Franz: Exactly. We are working intensively to develop AI solutions to analyze the signals from the more than 42 channels. These signals are based on clearly defined physical measurements. We systematically analyze the spectral ranges in which relevant quality fluctuations are most visible and then combine the most suitable channels for the evaluation. This evaluation using pre-selected combination channels reduces the computational effort and facilitates the rapid assignment of errors to error classes. In addition, selecting different combination channels enables an almost unlimited range of mathematical combinations of the defined physical input variables. AI methods are ideal for these kinds of complex, multi-variant evaluation tasks. Once the errors have been classified, we have the basis for communicating not only the existence of an error to the system controller, but also its cause. By detecting the cause, it is then possible to readjust the process – or to directly eject components that are outside the fault tolerance. Although the responsibility for correct teach-in and for the limit values remains with the user, the initial values that have been worked out make it comparatively easy to get to grips with our system.

How much manual work is involved in configuring the sensors and training the AI?

Franz: We are cooperating with various research institutes to drive automation forward. The earlier mentioned experiments at DESY are part of these efforts. By monitoring the ongoing laser process in parallel with a high-speed X-ray camera and our sensors, we lay the foundation for highly automated evaluation. This is because we can precisely trace in the image data when and where exactly pores form or other errors occur in the weld pool of a weld seam – and can precisely trace how these events are reflected in the 4D sensor signals. In a four-week series of experiments at DESY, we generated more than 100 terabytes of data, which we are now systematically analyzing.

Working closely with the institutes, we are developing algorithms that lead us to the “ground truth.” In other words, they provide detailed information about every error and every deviation in the process. We correlate this set of analyzed data with the signals from our multispectral sensors. This allows us to precisely differentiate between different types of pores in the molten bath. The evaluation developed using correlations reliably guides users to errors in their welding processes. After initial training, the AI is able to independently identify errors and error classes in industrial processes. We are now planning to transfer this approach from laser welding to other laser applications. We will gradually develop presets that our customers can use to quickly and easily adapt the 4D sensor technology to their particular application.

That sounds like the future. To what extent is the multisensory approach already contributing to the quality of processes today?

Franz: There is one very specific benefit: Our sensor system synchronizes the recording with the system control data and assigns it to part, group and seam numbers. This ensures complete traceability for users and allows them to remove defective parts from the process for reworking. However, the fundamental question arises as to how many different cameras, pyrometers, and laser acoustic, or OCT sensors are useful, affordable and possible to be integrated into installation spaces for monitoring laser machining processes. Of course, a broad database paired with AI is important to systematically deepen our understanding of processes. But it’s precisely because we see limits here that we pursue the “one sensor for any application” approach. Through miniaturization and modularization, we minimize the integration effort, which is the reason for coupling the fiber optic coupling to the processing optics. In this way, power supply and data transmission are also guaranteed via the same single cable. Our customers can also integrate several 4D sensors and network them optically, mechanically and communicatively with their laser systems. We believe the future is in combining multi-sensor technology and AI, and that’s why we are working so hard to make it as easy as possible to implement.

How will the possibility of AI-supported data fusion change process monitoring?

Franz: It will lead to a much deeper understanding of the process. AI offers us a very powerful means by which to filter relevant information for process optimization even from mass-generated sensor data. This in itself is a significant customer benefit. With our approach, we are taking the idea one step further. This is because it uses the same single sensor to easily and dynamically provide the required amount of data, which can be used to analyze even very fast laser processes with the necessary precision and provide users with relevant information on the cause of errors in the process. How far we can push recognition rates and classification hit rates upward also depends on physics. But we now have an effective tool to far surpass the previous limitations of process monitoring. Smart factories of the future require lean solutions that can be flexibly integrated. AI enables us to translate sensor signals into information of huge value. That is exactly what we are aiming for.