Attention last-minute decision-makers!
Conference Internet of Thing - 20 % discount
On October 11, 2022, the conference "Internet of Things - From Sensor to Cloud" will take place in Munich. The overview shows which topics and speakers await you. Book your participation now and secure a 20% discount with the discount code IOT22NLELNET.
The entire program and registration can be found on the event website of the Internet of Things - From Sensor to Cloud conference.
Opening keynote: Industrial Metaverse - AI-assisted maintenance with the help of spot robots
In their presentation, the two speakers will provide an overview of current smart product and smart factory use cases that have been implemented using Microsoft Azure IoT edge-to-cloud technology. In addition, a concrete customer example for the optimization of maintenance will be presented, in which significant savings were achieved with the help of artificial intelligence and the spot robot.
October 11, 9:15 to 10:00 a.m.
The different sessions
After the opening keynote, various sessions will take place in parallel: Session 1 (Artificial Intelligence in IoT) and Session 2 (Digital Twin and Retrofit) in the morning, Session 3 (Security and Open Source) and Session 4 (Best Practices) in the afternoon. The conference will conclude with a joint closing keynote entitled "AI, ethics and law".
Session 1: Framework for the data and AI lifecycle
In Industry 4.0, more and more data is being generated through the intelligent networking of machines and processes. This data can be used to generate knowledge - with the help of artificial intelligence (AI) - in order to improve production and services. However, there are challenges that make analysis using AI more difficult: Heterogeneity of production systems, high coordination effort for a data scientist, as well as a lack of flexibility for AI operations.
To overcome these challenges, Fraunhofer IKS is developing a framework for the data and AI lifecycle as part of the REMORA research project. The aim is to support the development of AI, to make the integration of AI (from the component level to the cloud) more flexible and to enable the automated, continuous operation of AI.
Using an example - predictive maintenance on a production line - the framework and its use will be illustrated: from data collection from machines and sensors, to AI training in the cloud, to continuous real-time analysis.
October 11, 10:0 to 10:30
Session 1: Current AI architectures for robust automatic speech recognition and transcription
ASR systems (Automated Speech Recognition) represent another interface between humans and computers alongside conventional interfaces (e.g. screens, keyboards, touchpads). ASR systems are particularly useful because they enable communication without direct physical interaction with an IT system. The scientific community and experts agree that the use of ASR systems will continue to increase. To create an ASR system, technically complex challenges need to be solved. A speech audio signal is very difficult for computers to interpret for various reasons. For example, the speech signal mixes with other noises or is more or less different from person to person, even if they speak the same language. The first practical ASR systems as market solutions use a system architecture based on probabilistic approaches (e.g. Gaussian Mixture Model) and simple language models. These system architectures have been used in the past and continue to be used today. The use of deep learning networks, often in combination with an increase in the amount of training data, has led to an important leap in the performance of these models. The use of deep learning methods in the field of automated speech recognition is constantly opening up new AI architecture possibilities due to the almost annual research breakthroughs in the field of deep learning. In addition to the technical aspects, more and more private companies are entering this field of research. The reasons for this are that they have also recognized the high value of an additional human-machine interface. For example, Google, Microsoft, Amazon and Facebook AI have now funded and published numerous research projects on this topic. For these technical and economic reasons, there are now a large number of different ASR architectures. Scientific research by DNDY. Data and Design GmbH identified and analyzed over 100 relevant scientific papers on this topic at the end of 2021. To categorize and preselect the researched models, a gradation was made based on the performance of the individual architectures. This was carried out using publicly available ASR datasets provided by the research community as test datasets.
Our presentation aims to answer the following questions:
- Are deep learning technologies finally the standard for robust automated speech recognition?
- Which deep learning methods are particularly promising in which parts of the speech recognition process?
- What influences do other deep learning research fields have on ASR research?
October 11, 10:30 to 11:00 a.m.
Session 1: Object recognition on Edge
Object recognition refers to the detection of one or more objects in sensor data. There are many fields of application here, be it the localization of other road users in autonomous driving or the recognition of products in industrial manufacturing processes. These different areas of application result in systems with more or less computing power and storage capacity, depending on their application environment.
In the case of image or video data, the aim is to localize the object using a bounding box, for example, and assign it to a class. There are already very good deep learning approaches that recognize objects with high accuracy, particularly in this area of application. They are usually based on complex convolutional neural network (CNN) architectures that recognize increasingly complex image properties over many layers. For example, colors and edges are initially recognized, which form larger shapes in later layers. However, their high complexity also requires a correspondingly large amount of training data and computing power to be available.
The aim of this presentation is to show how object recognition can be implemented with little data and hardware resources, i.e. on edge. To this end, we used Python and the NVIDIA technology stack to adapt high-performance object recognition models via transfer learning for various application scenarios and to remove parameters via pruning until the architecture can be used on an Nvidia Jetson Nano. In this way, real-time object recognition can be implemented in video streams with limited hardware. For a better understanding, we have implemented the same object detection on another technology stack from Luxonis and Intel.
October 11th 11:30 to 12:00
Session 1: Edge AI for automation: With smart hardware to smart data
Powerful neural networks are now being used in many areas and could also conquer the industrial environment in the future. Applications such as condition monitoring or predictive maintenance can be considered on the basis of "smart" collected and processed data. This will require data collection and analysis from components to entire systems with the aim of efficiently providing the right data in the right place.
Today, sensors collect data that is usually still transmitted to a powerful server where it is analyzed centrally. However, with the growing number of sensors, the volume of data is increasing and data transmission and analysis is becoming more and more time-consuming and energy-intensive.
One solution is to use smart solutions close to the sensor. Continuous measurement data can be pre-processed at an early stage and only relevant results or certain conditions need to be forwarded to the central server.
October 11 12:00 to 12:30
Session 1: Multi-criteria AutoML for TinyML with consideration of energy requirements
Philipp Woller/Fraunhofer Institute for Integrated Circuits IIS
© WEKA Fachmedien/Fraunhofer Institute for IISBackground - The wide availability of methods from the field of artificial intelligence (AI) and numerous available embedded sensor systems form the prerequisite for a large number of new, intelligent "IoT" applications in a very short time. In reality, however, the limited overlap between the skills of embedded system developers (HW) and data scientists (SW) often leads to long development cycles. The HW developer has difficulty estimating the HW requirements of data-based projects, while the SW developer is sometimes unable to ensure compliance with the boundary conditions for embedded systems, such as memory or energy requirements.
Goal - To bring these two worlds together, Fraunhofer IIS is developing various AutoML methods that take into account both the energy consumption on the target system and the performance of the AI pipeline as optimization goals.
October 11, 12:30 to 13:00
Session 2: Digital twin and retrofit
Session 2: How the Internet of Things and digital twins benefit from each other
Most systems and installations have IoT sensors whose data is stored in their own clouds. Communication between devices and installations from different manufacturers would be helpful in many cases, but is not always possible. The presentation shows how the IoT systems can be located in a common photorealistic digital twin and displayed in a visual environment. The photorealistic digital twins described here are digital representations of physical objects that are generated on the basis of digital camera photos and behave true to size thanks to mathematical methods. By using photos, these twins can be adapted to the changing reality without great expense. It is also possible to overlay the twins with 3D models/CAD drawings. Points of information are used to link the measurement and sensor data with the twin and locate it in three dimensions. Furthermore, state-of-the-art IT interfaces, such as web services, ensure problem-free connection to a wide variety of data sources. Existing evaluations, dashboards or other information can also be easily integrated into the twin. The combination of visual representation and IoT data gives the user an insight into the environment in real time, even from remote locations. Digital twins can also serve as the basis for AR applications.
October 11, 10:00 to 10:30
Session 2: Use of digital twins for quality monitoring and process optimization of metal-cutting manufacturing processes
A promising approach for a manufacturing company lies in the acquisition of existing data from existing individual processes or process chains and the subsequent intelligent evaluation to obtain new knowledge. A key role is played by digital twins for machining processes, which can be used to document, evaluate and influence the quality of the manufacturing process and the work result. These digital twins for machining processes are based on the process and planning data resulting from the manufacturing process that produces the component. To automatically generate these digital twins, all data and information sources involved in production or in the production environment are linked and merged, thereby realizing the complete digitalization of the production chain. The high-frequency acquisition of all geometric, kinematic and performance-specific process data takes place via the machine tool and is synchronized, structured and processed with the help of intelligent data processing methods based on a new information and semantic data model.
October 11, 10:30 to 11:00 a.m.
Session 2: Digital twin: from myth to guarantee of success - how we create real added value in the long term
A digital twin of a machine or an entire production line promises to solve the everyday problems of manufacturing companies - first and foremost the problem of a lack of transparency. However, reducing a digital twin solely to the creation of transparency squanders valuable potential. Looking at a digital twin in isolation on the store floor does not meet the high expectations of digitalization. With the right methodology, the many different digital twins of a production plant can be orchestrated and choreographed, thus opening up the entire value creation process for competitive innovations.
The presentation will use the results of a digitalization project to present this tried-and-tested methodology, which provides answers to the questions
* Which digital twin is the right one for me?
* How much digital twin is just right for me?
* Where do I start?
provides. Furthermore, a connection will be made between "digital twins" and the trend topic of "hyperautomation" and an outlook will be given on the new fields of action for IT and production managers.
October 11 11:30 to 12:00
Session 2: AI retrofit for control systems: Condition monitoring for drive elements with TinyML
Countless control solutions in machines and systems take insufficient account of the current status of the mechanical components of a drive train, such as the wear of motors, bearings, fans, pumps, etc. The entire behavior of a control system can be fundamentally changed if not only the classic reference variable is available on the input side, but also, for example, a categorical variable for the condition of the drivetrain components in the respective control section. High-quality condition monitoring input data enables intelligent system behavior. The control variable at the control output can be dynamically adapted to the current state of the mechanics. Faults, such as the imbalance of a rotating component, acoustic anomalies due to bearing damage at an early stage, thermal anomalies of a slip ring assembly, a completely failed drive element and other wear-related disturbances in the control section are detected in good time and can be taken into account by the PLC software.
The article shows how various sensors can be used to record the raw real-time data required to determine the condition in order to expand the reference variable of a control system with high-quality condition monitoring data. For example, capacitive MEMS-based inertial sensors for measuring acceleration and angular velocity or infrared sensor arrays for measuring temperature in surfaces can be used as sensor elements. In both cases, multidimensional data is generated that is particularly suitable for analysis using supervised machine learning (ML). This involves using regression or classification to calculate a categorical state variable that can be forwarded to a PLC input.
Since ML is to be used for real-time sensor data analysis in the immediate vicinity of the sensors and controller (i.e. not somewhere in a cloud or on a central, high-performance edge system), special software modules are required to create an inference function that can run on resource-limited microcomputers. The presentation will therefore answer the following three questions:
- What is the procedure for using a supervised machine learning algorithm to generate a condition monitoring input variable for controllers from multi-dimensional sensor data
- What needs to be considered when creating machine learning models and what are the typical errors in practice?
- What does a typical example look like and why should TensorFlow, TensorFlow Lite or TinyML be used as software modules for machine learning?
October 11 12:00 to 12:30
Session 2: Retrofitting existing machines with IoT functionality
One of the biggest hurdles when integrating existing machines into the Internet of Things is that they are often not equipped with an adequate interface. Replacing these machines is neither sustainable nor economically viable, which is where retrofitting systems come into play.
The prototyping phase of a new retrofitting system usually involves connecting the key system components such as sensors, microcontrollers and the wireless module. This can be done either by using a single board with all components, or by connecting individual evaluation boards via jumper cables. The first, monolithic approach is expensive and inflexible, while the second, modular approach is error-prone and complicated. To overcome both disadvantages, the Adafruit FeatherWing form factor and its standardized pin assignment are used. With this approach to componentization, there are a number of stackable boards that can be combined with each other, providing high flexibility and minimizing sources of error. This approach of component-based system development is already widely used in software development and is applied in this case through the use of software libraries.
The presentation will demonstrate modularity for a retrofit use case by connecting an RS232-capable machine to a cloud via MQTT over Wi-Fi, where the microcontroller and the interfaces are interchangeable modules.
October 11, 12:30 to 13:00
Session 3: Security and Open Source
Session 3: EU Cyber Resilience & Cyber Security Act - An Overview
The Cybersecurity Act strengthens the EU Agency for cybersecurity (ENISA) and establishes a cybersecurity certification framework for products and services.
The core business of critical sectors such as transport, energy, health and finance is increasingly dependent on digital technologies.
Connected devices, including machines, sensors and networks that make up the Internet of Things (IoT), as well as their security, will play a key role in shaping Europe's digital future.
October 11 14:00 to 14:30
Session 3: Edge computing also for security in the IIoT
Edge computing is not just computing power on or near systems and machines. Edge computing can connect, calculate and operate. But it can also decouple, secure and hide. Edge computing can not only make machines and systems smarter.
It can make them more secure. And not just the machines and systems. The applications that are operated on an edge computing device can also be protected. Not only can, but must be protected. Why? Because we rely on the applications being tamper-proof. Because the results of these applications are crucial for our business. secunet shows in the presentation "Secure Edge Computing" why IT security is also necessary here. And what secure edge computing can look like.
October 11, 14:30 to 15:00
Session 3: Protection and security for machine learning
Machine learning is the training of an artificial intelligence (AI), usually with a lot of data. The trained model can then be used to make predictions for further data.
An example would be an examination for tuberculosis using X-ray equipment. With the help of machine learning, this can be automated and also made accessible to general practitioners. The trained model is the intellectual property of the manufacturer of the device.
The training configuration and the trained model require special protection, regardless of whether the model is used in the cloud or runs on hardware directly at the GP's practice.
CodeMeter offers a generic solution for both cases by encrypting the data model and thus protecting it against unauthorized use, disclosure and analysis.
The medical example also shows the importance of the original training data, which contains personal data. Further threats need to be considered here. The theft of data on one side and the falsification of data on the other.
Here too, CodeMeter can help by encrypting and signing the data.
In the first part of the presentation, we will briefly show you the specific threats and dangers that occur in machine learning. In the second part, we will give a brief overview of CodeMeter Protection Suite and how it can be used to protect data models.
Machine learning is an option for new products and protecting the generated data models against unauthorized use, disclosure and analysis is essential.
October 11 15:00 to 15:30
Session 3: Sustainable consumption of open source for the IoT
For many companies, the concept of "open source" is still relatively new. It is only just starting to gain a foothold, especially in the manufacturing industry.
The part that most companies quickly understand is that open source is available free of charge. However, that is only one facet.
My aim in this talk is to highlight the other facets as well:
Because open source is so infinitely more. So join me on a journey into a world in which competitors can work together profitably, in which small companies can achieve what is otherwise only possible for large corporations. I would also like to show that despite open source, this does not mean things like: Commercial support, SLAs and consulting need not be ruled out.
However, I also want to show that I need to change something. The way in which open source is currently used will not be sustainable for much longer. Situations like Log4Shell and the Faker.js/Color.js incident, as well as other cases of open source activists dropping out, are symptoms of a systemic problem. A problem for which we urgently need to find solutions.
October 11th 16:00 to 16:30
Session 3: Using Elastic - Stack for transferring, storing and analyzing machine metrics and state data
The transmission of machine metrics and status data for mobile machinery faces the major challenge of mobile network coverage, which is not always optimal. For this reason, data storage and the automated reloading of historical data are particularly important in order to provide seamless after-sales service, optimize products based on the data and offer new digital products and services. The specialized use of the ELK stack and the associated components such as Filebeat, Elastic Stack and Kibana can make a valuable contribution here. In addition, the Azure IoT Hub or Digital Twin can be used to remotely parameterize the frequency and quantity of data points sent.
October 11 16:30 to 17:00
Session 4: Best Practice
Session 4: All roads lead to the cloud: concepts for data transmission in IoT scenarios
A large number of architectural decisions have to be made when implementing IoT projects. An important part of this is communication between devices and the cloud. Various synchronous and asynchronous concepts for data transmission have become established in the IoT world. We use practical examples to show you which concept is particularly suitable for which use case. Learn with us how integrated IoT hub communication, device twins and other exciting concepts can be used correctly. We look forward to sharing experiences and best practices from our projects with you.
October 11 14:00 to 14:30
Session 4: Device-2-Cloud - From embedded Linux to the Azure cloud with Qt or .NET
Embedded devices are becoming increasingly networked: integration into local Wi-Fi networks, wireless connections to mobile end devices or the use of scalable cloud interfaces are often the focus when implementing networked embedded devices. But which mix of frameworks offers an optimal environment to reach the goal efficiently?
A case study will be presented in the lecture: An embedded Linux-based Linux device is to be connected to the Microsoft Azure IoT Hub cloud solution. The necessary steps with an implementation from the device to the cloud are demonstrated with 2 different technologies: In the first case, the application, which was implemented with C++ Qt, is made cloud-enabled in combination with the Microsoft Azure IoT C-SDK. The second case is an application that was implemented with .NET for embedded Linux. Finally, the technical advantages and disadvantages will be compared.
October 11 14:30 to 15:00
Session 4: Practical report: Behavior Driven Development (BDD) for an agile embedded IoT development project with international partners
There is a saying in software development that sometimes the greatest difficulty lies in finding out what to develop in the first place. This may sound surprising to developers working in hardware development and make them smile. However, for innovative IoT projects with a high proportion of software, this difficulty is often the determining factor in budget overruns, delivery delays or even failed projects. BDD or Behavior Driven Development is an agile approach that starts with the requirements process and ensures seamless specification and very efficient communication between stakeholders, project management, developers and validation.
This article is a field report of a successfully completed agile innovation project for an embedded IoT application in the machine tool sector. The implementation was carried out by an internationally distributed team. The key to success was the use of methods from modern software development with BDD (Behavior Driven Development). These were adapted and applied for requirements management, project management and the accompanying validation for the hardware-related development of embedded systems. The practice-oriented contribution describes not only the concrete project organization and approach, but also the open source tool infrastructure used for continuous integration and automated testing.
October 11th 15:00 to 15:30
Session 4: RUL prediction of a geared motor in the cloud
This presentation shows how the Remaining Useful Life (RUL) of a servo gearbox can be predicted by real-time streaming of the servo motor data using an Arduino-based data acquisition system to the cloud (ThingSpeak) to a MATLAB-based RUL model.
Motor current signature analysis (MCSA) of the current signal driving the servo motor used is used to extract features in the frequency domain (spectral domain) over several frequency ranges of interest that indicate motor and gearbox faults. A combination of features is used to create a health indicator (HI) for subsequent RUL prediction.
MCSA is a useful method for diagnosing faults that cause torque or speed variations in servo gearboxes, which in turn lead to correlated motor current changes. MCSA has proven to be ideal for analyzing motor faults as only the motor current signal is required for the analysis and therefore no additional and expensive measurement hardware is needed. Detecting transmission faults with conventional vibration sensors is challenging, especially in cases where the transmission train is not easily accessible for instrumentation with accelerometers or other vibration sensors.
This presentation will illustrate how to build a real-time data streaming, feature extraction and RUL estimation system using simple, off-the-shelf components suitable for both industrial application prototyping and educational lab exercises.
October 11, 16:00 to 16:30
Session 4: Building a robust IoT system for monitoring wastewater pumps with pattern recognition
Challenge: Toilets on ICE trains break down too often
Solution: Networked system of battery-operated sensor nodes with integrated pattern recognition
pattern recognition and reporting of faults
The system I designed is successfully in use and fully automatically monitors the use of pumps and the disposal of rail vehicles at Deutsche Bahn. I report on my practical experience in project management and the numerous technical and organizational difficulties that we overcame together.
Architecture:
- Fuzzy logic instead of deep learning
- Edge computing instead of cloud computing
- Radio standard LTE-M instead of WiFi, Bluetooth or LoRa
Requirements:
- Stability of hardware and software for months of non-stop use
- Minimal energy consumption
- Maintenance-free operation
- Automatic switch-on and switch-off during use
- Functions for remote maintenance
- Reliable classification of pump behavior
- Signal processing locally in the module
- No transmission of all raw data, only reliable results
October 11 16:30 to 17:00
Closing keynote: AI, ethics and law
Particularly in connection with the use of artificial intelligence, there are efforts to discuss or even standardize ethical aspects. This presentation provides background information on ethical concepts as well as existing and upcoming legal bases (keyword: EU AI Act), debates the conflicts and problems behind certain approaches and analyzes possible effects on society.
October 11, 17:00 to 17:30
You might also be interested in
Embedded community votes for favorites
In addition to the coveted 'embedded award', the 'community choice award' will also be presented for the first time at embedded world. The special feature: The award winner is determined by the embedded community.
read more...Preview on embedded world 2024
Inference can be as Compute Intensive as the Training
With CPUs, GPUs and FPGAs, AMD has all architectures that are used for AI. embedded world Keynote Speaker Dr. Salil Raje shares his view on AI.
read more...International meeting places for the embedded community
From April 9 to 11, the embedded industry will meet in Nuremberg - but not only there, because embedded world will also take place in China and the USA in 2024. How did it go in China? What can we expect in Nuremberg? And a look ahead to the 1st...
read more...In focus: Artificial intelligence
The e-paper of the January 2024 issue
ChatGPT has given artificial intelligence a huge boost over the past year - especially in terms of public perception. In this issue, we want to take a closer look at AI and categorize it in more detail for potential industrial users.
read more...AI boosts the performance of software engineers
New "superpower" for software developers: An AI-supported programming assistant is set to revolutionize the development landscape.
read more...ERP trend forecast for 2024
Unit4 forecasts the most important ERP trends for the year 2024. Claus Jepsen, Chief Product and Technology Officer, names nine points that he considers relevant.
read more...These are the AI trends for 2024
Artificial intelligence is changing the entire IT world forever and is having a huge impact on how data is understood, analyzed and used. AI is also taking data management to a new level. Five AI trends have been identified for the year 2024.
read more...Open Industrie 4.0 Alliance on App Store
The OI4.0 Alliance and the app stores
Shaping the future of Industry 4.0 together is the goal of the Open Industry 4.0 Alliance. Flecs has been chairing the 'Application Management' working group for a year now. Samuel Greising, Managing Director at Flecs and head of the working group,...
read more...Software in online direct sales
Flecs entered the market two years ago with a new business model: it wants to automate the installation and updating of industrial apps on control systems. Patric Scholz, Co-Founder, explains the innovations for the SPS.
read more...