Eliminating errors with fused sensor data

Manufacturing errors are expensive and annoying – and all the more so the later they occur in the process. This is because the product value increases with each production step – whether it’s an automobile, a battery cell, a wafer full of microchips or elaborately ground and coated high-end lenses. And a fault that goes undetected before delivery opens up the risk of accidents, recalls and reputational damage.

That’s why many industries rely on close-meshed quality controls and, increasingly often, on inline process monitoring. In many cases, however, the measures only serve to identify faulty parts, remove them from the process and – if possible – rework or repair them. This could change if it were possible to take the signals and data from the sensors of the inline monitoring system, and derive relevant information about deviating process parameters and the causes of fluctuating quality from them in real time. The dream of first-time-right production or zero-defect manufacturing is becoming a tangible prospect. There are three main reasons for this: The exponential increase in computing power and data transfer rates is paving the way for multi-sensor approaches; these in turn create the necessary database for using artificial intelligence (AI) to filter relevant information from the heterogeneous sensor data, allowing for the continuous readjustment of quality-critical process parameters.

Multi-sensor technology detects faults earlier and more reliably

All three factors are important. For example, when laser welding processes or construction processes in additive manufacturing are monitored using multiple sensors, it doesn’t take long before many terabytes of data are being generated. The key to more deeply understanding processes lies in combining high-resolution and in some cases high-speed cameras, multi- and hyperspectral sensors, pyrometers, photodiodes, acoustic sensors, interferometric approaches like optical coherence tomography (OCT) as well as the tracking of laser power and scanner. In order to be able to use this key, all the data collected has to be synchronized using time stamps – and then it must be searched for defined error classes using pre-trained algorithms. By using different sensors, deviations from the target are not only detected more quickly, but can also be verified immediately by comparing the synchronized sensor data.

In the past, inline process monitoring has focused on detecting production defects after they occurred based on defined symptoms. Now, a different interest is coming to the fore: Suppliers and users of laser material processing processes want to find out the causes of defects in detail and recognize the principles behind them so that they can use this knowledge to take countermeasures in ongoing processes. In addition to using different sensors, this requires a powerful infrastructure for real-time synchronization, transmission and evaluation of the measurement and camera data. These are collected in different formats, and most of the time also at different frequencies, sometimes up to the kilohertz (kHz) range. However, truly effective process monitoring requires more than just synchronizing data. In order to be able to interpret sensor signals, relevant signals need to be reliably distinguished from process-related noise, and it’s necessary to conduct systematic sensitivity analyses of the individual sensory approaches.

Multi-sensor technology and AI have a major impact on photonics

This can be seen, for example, in a dissertation on the multi-sensor process monitoring of laser powder bed fusion (LPBF) by researcher Emil Duong from Aachen. In it, he shows how the interaction of different optical and acoustic sensors can be used to effectively monitor additive manufacturing processes. At the same time, the work shows that multisensory process monitoring has long since arrived on the market. All leading suppliers of AM systems use this approach, although their solutions differ in detail. Some rely on pyrometers, others on photodiodes for NIR and visible spectral ranges, on solid borne sound sensors, or on sCMOS cameras. The goal is always the same: more control over the laser-based layer construction processes. The sensor data flows together in emission maps and can be analyzed in detail using the software supplied.

In his work published in mid-2023, Duong not only provides an up-to-date overview of the market penetration of multisensory approaches, but also shows just how much this topic is receiving attention from researchers all over the world. Sensor fusion was already commonplace in laboratories before the breakthrough of generative AI and large language models. Researchers use NIR sensors to monitor the stability of the melting bath and investigate the formation of tiny pores; they use cameras to check the build platform and powder bed; and they track the heat development and cooling behavior of the AM components in the laser-based layer construction process using thermographic cameras. Optical microphones and solid borne sound sensors are also used. Their signals provide information about process stability and clear indications of deviations. Machine learning methods are used for the purpose of evaluation.

Utilizing the opportunities of AI-supported process monitoring

This insight into research was clearly reflected at the AKL International Laser Technology Congress 2024 in Aachen. The extent to which digitalization and AI are already influencing value creation and business models in photonics was brought to light in many presentations and discussions. It also became clear just how widespread the use of AI in process monitoring already is – and how urgent this development is. The bottom line: Access to data and the ability to derive added value from it with AI is already synonymous with competitive advantages. It’s a matter of who will control the photonics markets of the future. If suppliers of photonic hardware do not act quickly, there is a risk that control will pass to software companies, which will integrate lasers as a commodity in overarching digital manufacturing platforms and consistently shift the focus toward digital services. At the same time, AKL’24 also sent out an extremely positive message: The majority of the numerous LASER exhibitors have long since recognized the signs of the times. From TRUMPF, to COHERENT, to Precitec, to Blackbird Robotersysteme, to Scansonic, to 4D Photonics, to XARION Laser Acoustics – they have all been harnessing the possibilities and potential of AI for a long time – and are creating real added value for their customers.