AI series: Putting artificial intelligence into practice

Audio and Media Technologies  – Dr. Bernhard Grill

"Application is our strength"

Whether in the cinema or at home, media entertainment often involves technology from Erlangen. In the field of digital cinema, for example, AI algorithms are helping to continuously improve the post-production software easyDCP. In home theaters, on smartphones or in VR glasses, the successors of mp3 or AAC – which put Fraunhofer IIS on the international map – provide interactive and immersive 3D sound in streaming applications and the broadcasting systems of today and tomorrow. AI methods will also play a role in this environment in the future; for example, to better recognize dialogue even amid background noise and, depending on the user’s preferences, to turn the volume up or down.

“Application is our strength,” says Bernhard Grill, Director of the Audio and Media Technologies division. And where AI is used, sophisticated data compression during recording and transmission also reduces the growing audio and image data rates without compromising quality.

Since Fraunhofer IIS is a world leader in the field of acoustic signal processing, it’s no surprise that its scientists recently announced another success: the planned voice assistant platform SPEAKER, designed for B2B applications, received an award in the AI innovation competition held by the German Federal Ministry for Economic Affairs and Energy (BMWi).

Fraunhofer IIS is taking a further step toward the future with the KISS project, which is being realized in cooperation with the Friedrich-Alexander-Universität Erlangen-Nürnberg. The project seeks to ramp up research in the field of machine learning by establishing a practice-oriented AI laboratory for system-level design of signal processing applications based on machine learning

Engineering of Adaptive Systems  – Dr. Peter Schneider

AI technologies for SMEs

All good AI things come in threes at the institute’s Dresden branch, Engineering of Adaptive Systems EAS, as it endeavors to make AI even more beneficial to users in three research areas: engineering methodology, hardware realization and intelligent data analytics. Scientists have been using AI in customer projects since the 1990s, such as the monitoring of production plants and quality assurance. Here, for example, they combine data collected by sensors with expert knowledge for the respective application. In practice, there is often a lot of data available, but it is not always the right data for solving a problem; to compensate for this, the scientists also use physical computer models to, say, teach artificial neural networks. Besides data analysis with AI algorithms, the experts also pursue other technological approaches, such as private cloud and edge solutions or the use of special, particularly powerful AI computing hardware. “With these diverse tools, we can optimally address the various requirements of our customers. This results in individual solutions for application areas where data sovereignty, data protection and high processing speed are crucial – and especially for small and medium-sized companies, too,” says Dr. Peter Schneider, Division Director of Engineering of Adaptive Systems.

He continues: “We want to build a bridge between standard applications, which many companies use today to solve problems, and the use of AI algorithms.” That’s why the institute branch is working on providing test laboratories, consulting services by experts and other modular offers. These modules grant SMEs access to the latest developments in machine learning, deep learning, explainable AI and more. To this end, the EAS branch is also active in various networks across Germany, including AI4Germany, the country’s umbrella initiative for AI. This network brings together leading regional initiatives on the topic to actively promote the use of artificial intelligence by local businesses and communities. Each partner receives support from its own state government. The institute’s Dresden branch has also been actively involved with the “Center for Explainable and Efficient AI Technologies” (CEE AI), together with other Fraunhofer partners and the TU Dresden, a German University of Excellence, since 2019. Research at this center will cover the entire AI spectrum, from hardware support and device communication to AI approaches and transfer into practice.

Development Center X-ray Technology EZRT  – Dr. Norman Uhlmann

"We know exactly what our customers need"

For Deputy Division Director Norman Uhlmann, AI-supported X-ray technology holds great promise for recycling in industry. The method makes it possible to achieve good results quickly, even in complex situations. “We’re already using AI and neural networks in evaluating 2D and 3D images,” Uhlmann says. The use cases are many and diverse: sorting facilities can employ various X-ray methods and detectors to find, say, fish bones in fillets or bones in cuts of meat. But that’s not all: the raw data is evaluated while the measurements are still being taken. And in the not too distant future, it may be possible for AI itself to propose which imaging and optical testing techniques should be used, for example in nondestructive testing. However, Uhlmann believes that AI methods shouldn’t be a “black box,” making arbitrary decisions that cannot be understood or that are obviously wrong from the human observer’s perspective.

“Some questions are too complex to be solved with conventional methods; it is precisely these that we want to answer with AI methods. The multi-energy X-ray detectors or data sets from XXL computed tomography provide complex data that cannot be solved in the usual way,” Uhlmann explains. Near-sensor AI ensures that the collected and stored data is of the highest quality and contains maximum information.

“Our systems generate data directly. We know the strengths and weaknesses of data generation, and align this with the needs of our customers", Uhlmann says. “This delivers an immediate benefit in application.”

Communication Systems  – Prof. Michael Schlicht

Where sensor technology meets physics

To make data communication as efficient as possible, the Communication Systems division began tackling the topic of artificial intelligence many years ago on two fronts. Software-based embedded machine learning and special neuromorphic hardware – the latter aimed at technical reproduction of the biological model of the human brain – are two of the main areas in which Division Director Michael Schlicht’s team is making great strides. Another area, the joint IIS AI FLEX project, demonstrates what new AI-based electronic solutions can do; make autonomous driving safer, for example. With the help of a powerful hardware platform and special software, it will be possible to precisely record the position of a vehicle and its surroundings. “And in situations where weak AI is in place, we are constantly evaluating which architectures can be used,” Schlicht says.

Artificial intelligence is much more flexible and energy-saving when it is coupled with near-sensor (and thus variable) signal processing, and when algorithms, for example, act right at the sensor – especially since AI can then also respond directly to environmental influences. This is important for energy suppliers who, for example, want to keep their quality of service with regard to condition monitoring and security of supply at the desired high level and, just like their customers, want to keep costs under control at the same time.

“At the interface of sensor technology and physical data transmission, we provide the necessary and correct information so that we can transfer it efficiently to a server or to the cloud,” Schlicht explains. In addition to data transmission, however, Schlicht’s team also optimizes the transmission routes themselves so that applications – for example in public safety, IoT or automotive – save energy in their overall system

Positioning and Networks – Dr. Günter Rohmer

A step closer to people

For the wireless positioning and locating of people and objects, for example in industrial applications in which humans collaborate with machines, the Positioning and Networks division relies on AI methods to implement positioning and networking technologies as communication solutions. “In this way, AI or machine learning can help make production environments more efficient, especially where wireless sensor communication and networks are used to make interaction with production staff more secure,” says Günter Rohmer, Director of the Localization and Networking division. From assistance systems for automated transport systems to intelligent tool tracking, “a broad range of applications is bringing us a step closer to the people who interact with the machine, so networking and communication must be precise, extremely fast and secure,” Rohmer says. “The systems must react to the complex and changing environmental conditions in production and, for each location, determine the best correction of the physical measurement. This can be achieved in real time and with optimum results only by using the right self-learning AI methods.” Industry partners who require IoT solutions for production, smart cities or smart buildings benefit from the Positioning and Networks division’s technical know-how. Add in the use of the new 5G mobile communications standard, and it is possible to achieve high-precision positioning accurate to within a few centimeters, with latency times of a few milliseconds and data rates of up to 20 gigabits per second.

AI processes are also suitable for classic SMEs, allowing them to maintain the high quality standards of their products. This is because AI paves the way from industrial mass production using robots to customized high-quality production with small batch sizes. “Building on high-precision localization in real time, we want to use AI in such a way that lengthy training cycles are no longer necessary,” Rohmer says.

Still, the new key technology will not replace expert knowledge and human creativity. As Rohmer puts it, “AI helps us, and therefore our customers as well, to implement extraordinary ideas faster and more effectively.”

Smart Sensing and Electronics – Josef Sauerer

Smart AI for a wide range of applications

Constant monitoring of chronically ill patients, image-based emotion recognition in autistic children or the digitalization of scents: “The range of AI applications is very broad and illustrates the potential of deep learning,” says Josef Sauerer, Division Director of Smart Sensing and Electronics. Embedded AI reveals its decisive advantage wherever a detour via the cloud is not possible; for example, because the use case requires real-time capability. “Our unique selling point is the combination of applied knowledge with sensor technology and microelectronic implementations. Because mobile operation calls for energy-efficient, low-cost, small and lightweight devices, it requires highly integrated microelectronic systems.”

One core area for Smart Sensing and Electronics is determining exactly how AI decisions are made so that it can provide users with certainty and, above all, secure and reliable results – including in robust use and under a wide range of conditions. This includes the question of how to facilitate machine learning even with comparatively little data – and which learning methods are generally best suited for which application.

Smart Sensing and Electronics works for international clients on the development and implementation of needs-based, pioneering solutions in the fields of medical technology, image sensors, integrated sensor systems and IC design. Some examples include solutions for reliably determining the condition of production machines, for safety-relevant applications in autonomous driving, or for intelligent 3D magnetic field sensor technology. The latter can be used for, say, better resource utilization in washing machines or for the continuous mobile monitoring of chronically ill people.

Supply Chain Services – Prof. Alexander Pflaum

AI from the management perspective

“Today, data is an essential source of higher sales and efficiency,” says Prof. Alexander Pflaum, “provided that it is correctly selected, analyzed and used. And data analytics and AI are important underlying technologies for this.” The required data is not always available in the right format or in sufficient quality, but this is necessary if the digital transformation and using data to generate added value are to succeed.

This is where the Fraunhofer SCS Working Group contributes its expertise: when, for example, data is available only in an unstructured form or has to be regenerated, the group develops data spaces based on knowledge graphs or uses IoT technologies to close the gap. If the necessary data is available, the group’s experts combine mathematical optimization methods with data-driven processes to add value. This added value is realized in a further step by developing new services and corresponding business models.

Pflaum says: “The key lies in the combination of domain knowledge, technology, mathematics, statistics and business processes.” As a result, the Working Group approaches AI from the technological and especially the management perspective; a clear USP. The group applied this method to BSH Hausgeräte GmbH, making it possible to predict the need for spare parts, even for products that won’t need any until many years down the line. This significantly reduces storage and recycling costs, among other things. This comprehensive approach is also reflected in the collaboration with BHS Corrugated Maschinen- und Anlagenbau GmbH, the world market leader for corrugated cardboard systems. If the digital transformation is to succeed, it is important that companies build up their own in-house AI competencies as well. For this reason, the Working Group has developed a new cooperation format: Joint Lab Data Analytics. Scientists and BHS employees have been working there together on AI applications since 2019, thus advancing the company’s own data analytics expertise.

Article by Ilona Hörath.