Embedded Machine Learning

Machine learning (ML) enables electronic systems to learn autonomously from existing data and to use this acquired knowledge to independently make assessments, predictions and decisions. These kinds of applications are highly compute-intensive, so they are traditionally executed on PCs and cloud servers. Thanks to new concepts and algorithms, as well as powerful dedicated processors, it is now possible to perform machine learning directly on devices used in the field (= embedded machine learning).

Embedded devices for machine learning applications can fulfill many tasks in industry. One typical example: sensor devices that detect acoustic or optical anomalies and discrepancies and, in this way, support quality assurance in production or system condition monitoring. In addition to cameras for monitoring visual parameters and microphones for recording soundwaves, these devices also use sensors for, for instance, vibration, contact, voltage, current, speed, pressure and temperature.

Why use embedded machine learning?

The Internet of Things (IoT) is the main reason for the rapidly increasing number of sensors used, and with them the amount of data collected by embedded and IoT systems. Although the transmission technology, too, continues to evolve, and the new 5G standard will provide a powerful data network, complete transmission of sensor data to the cloud is not always practical or feasible. There is a variety of reasons for this, some of which also support the use of embedded machine learning:

Power consumption relatively high power requirements of radio transmitters and receivers
Network costs overloaded networks and the cost of the resulting data traffic
insufficient coverage and data rates in some locations, e.g. in basements, tunnels, remote areas, etc.
Latency excessive round-trip transmission time for sensor data and feedback of actuator commands, especially for real-time applications that require a quick response
Privacy the frequent need to prevent external access to audio and video recordings from industry systems – on-site data processing could make data espionage more difficult
Data sovereignty system users or operators need to have full control over the data
Security und Safety risk of data and device manipulation
Form factor e.g. the need for devices to accommodate integration of suitable antennas

Options for embedded machine learning

To avoid these problems, it is often possible to process at least a portion of the sensor signals locally in the embedded device. For simple sensor data, this can be done with standard microcontrollers. Microcontrollers are well suited, for instance, for simple machine learning problems and applications with few channels and a low sampling rate, which don’t require frequent analyses. For more complex analyses, e.g. for image sensors, specialized deep learning accelerators can also be used. For other applications, such as speech recognition, an accelerator IP core on an FPGA can be useful, and ASICs are a suitable option for high-volume applications.

What we offer

Fraunhofer IIS covers both microcontroller-based machine learning and the use of embedded chips with deep learning accelerators. For a given problem, we analyze the system requirements and determine the appropriate algorithms and the best hardware options. Then we conduct a trade-off analysis for these options with our customers to find the optimum solution. Finally, we implement the chosen option on commercially available or customized hardware and integrate it into the target system.

Our approach to solving machine learning problems on embedded devices
© Fraunhofer IIS
Our approach to solving machine learning problems on embedded devices

Sensor data analysis is one core of the guiding topic “cognitive sensor technologies” at Fraunhofer IIS. As regards embedded machine learning, Fraunhofer IIS is building on its many years of experience in the fields of machine learning, embedded platforms, sensor technology, signal processing and neuromorphic hardware.

You may also be interested in

Project “SyNaKI”

SyNaKI - Synergy of ­natural and artificial ­intelligence in the swarm

The aim of the “SyNaKI” project is to map a natural swarm intelligence virtually in a network of microprocessors (digital swarm) in order to enable data analysis by artificial intelligence (AI).

Artificial intelligence

Go back to business area “Artificial intelligence“


Best paper award

Our paper “Enabling distributed inference of large neural networks on resource constrained edge devices using ad hoc networks” was awarded the “Best Paper Award” among a hundred scientific studies presented at the DCAI Conference.