February 2025

embedded world 2025 // Lecture:
AI-based IoT Sensors with Autoencoders

Our software developer Kolja Bohne is giving a lecture at this year's embedded world Conference describing the development process of a machine learning model for AI-based sensors.

AI applications that use machine learning (ML) consist of two main areas: generating ML models and performing inference in deployment. Creating an ML model requires large amounts of data and significant computing power, making cloud environments a suitable choice for this task. However, for lightweight models, running inference can even be done on single-chip microcontrollers in typical edge environments. As a result, special development methodologies and maintenance concepts are needed.

In the introduction, the lecture uses an IoT sensor example to provide an overview of creating an ML model through a seven-stage development process. It addresses issues such as data acquisition, from which an ML model learns the relationships necessary for inference, and the challenges of practical use, including monitoring and model updates. Regarding the data required for training, the autoencoder architecture is introduced, and reasons are given for why this type of artificial neural network is particularly suitable for edge AI anomaly detection applications.

The lecture then explores which tasks in anomaly detection can be implemented using autoencoders in edge systems and how this approach can contribute to creating cyber-resilient IoT devices. In this context, it provides insights into a real-time embedded intrusion detection system (IDS) designed to detect potential cyber-attacks in IoT and OT data communication.

When & where?

The lecture will take place on Wednesday, March 12 at 15:30 h (SESSION 7.6 EDGE AI // Smart Sensing Based on Edge AI) in the NCC East Convention Center.

embedded world Conference Program

 

Overview News