Digitalisation and artificial intelligence enable many new applications by increasingly automating processes, services, and facilities. The key to this is the ‘Internet of Things’ (IoT) with a rapidly growing volume of data provided by a rising number of connected devices and sensors. This hardware is expected to operate more and more efficiently in terms of resources and energy while also responding quickly, especially in safety-critical applications. At the same time, it should be trustworthy from data acquisition to analysis. This is to be achieved by limiting the data traffic itself and processing data more and more near the point of origin through edge computing and thus less in the cloud.

For this purpose, IMMS is researching two issues in the HoLoDEC project: on the one hand, ultra-low-power architectures (ULP) and their efficient and automated design, and on the other hand, on edge AI systems that will enable energy-optimised power distribution between sensor and edge computing. Approaches from both focal points will be validated in two demonstrators, one for condition monitoring and one for an RFID sensor transponder IC with a ULP sensor frontend.

For design houses and chip developers, it is becoming increasingly challenging to develop optimal or energy-optimised integrated circuits and systems as these are becoming more complex and human resources are becoming scarcer. Therefore, IMMS is researching in close collaboration with Reutlingen University and Fraunhofer IIS/EAS to accelerate the development of analogue or mixed analogue/digital circuit blocks through generator-based, i.e., programmed processes.

As part of this, models should also be automatically generated to, for instance, consider non-functional properties such as energy consumption early in the system design. Both are necessary because with the currently common methods, the energy requirements of a system are only known towards the end of the design process and are therefore difficult to consider in the architectural design. At the system level, the lever for optimisations is the most significant. However, it can rarely be utilised since extensive changes in microarchitecture or topology would be required, sometimes in multiple iterations, and these, like the entire design of analogue components, would still have to be done manually.

IMMS is researching to reduce design times and uncertainties through automated designs and models. The goal is to accelerate the design process so that it can be iterated multiple times to optimise the energy consumption of a system early in the process.

With this new approach for automated design, IMMS will investigate and design particularly energy-efficient circuit concepts and validate them with a demonstrator for an RFID sensor transponder IC with a ULP sensor frontend. These results will be compared with manually designed chips. The demonstrator addresses a solution that enables the realisation of integrated sensor systems with very low energy budgets, ideally allowing for passive operation without batteries, and can be manufactured using cost-effective semiconductor technology.

To make optimal use of AI algorithms on resource-constrained devices for IoT applications, IMMS is researching energy-efficient edge AI systems with energy-optimised power distribution between as much close-to-sensor data processing as possible and minimal offloading of tasks to the wireless network in the HoLoDEC project.

IMMS takes the lead in cross-system optimisation of energy consumption in distributed sensor systems. The foci include hardware and software design, energetic modeling of embedded systems, concepts and procedures for sensor and overall system energy management, investigation and implementation of algorithms for anomaly detection on ultra-low-power components, as well as algorithms for automatic data extraction and reduction in terms of information content and energy consumption impact within the sensor. The algorithms will provide compressed sensor data, which will then be transmitted to the edge-AI platform. To combine data from different sensors, algorithms for data fusion are being explored and implemented. Additionally, IMMS is developing energy consumption models for individual sensor systems and the overall system.

To utilise AI algorithms as close to the sensors as possible without the need for extensive data transport to a cloud, IMMS is conducting research on optimisations of edge AI system architectures based on defined requirements. The background is that models and algorithms are no longer limited to running on powerful servers with ample storage space but are intended to be used on microcontrollers for which they were neither originally developed nor directly transferable, thus requiring adaptation. Current research often focuses on developing high-performance models on a server. The goal is to build upon initial investigations into optimisation for microcontrollers and energy-saving potential and to make AI algorithms accessible for resource-constrained devices. These approaches are being validated in a demonstrator for condition monitoring.