TY - GEN
T1 - ML Training on a Tiny Microcontroller for a Self-adaptive Neural Network-Based DC Motor Speed Controller
AU - Funk, Frederik
AU - Bucksch, Thorsten
AU - Mueller-Gritschneder, Daniel
N1 - Publisher Copyright:
© 2020, Springer Nature Switzerland AG.
PY - 2020
Y1 - 2020
N2 - Neural-Network (NN)-based controllers have the potential to achieve better control performance than classical PID controllers. Yet NN deployment on tiny microcontrollers, which are used in DC motor control due to strict cost requirements, is challenging as NNs are computationally intensive and memory demanding. We propose a lightweight direct inverse NN-based control approach for controlling the angular speed of a permanent magnet DC motor, which runs on a tiny Arm Cortex-M0 microcontroller with only 4 kB of RAM. Moreover, the NN-based controller can self-adapt to the DC motor characteristics without the need of any external machine learning frameworks such as TensorFlow. For this, we are not deploying a pre-trained network for inference but implement a fully automated training process on the microcontroller, which also includes the dataset collection. The result is a self-adaptive control algorithm that is able to drive the motor at the desired speed after it learned the motor characteristics in an initial training phase. Furthermore, the approach is extended such that it enables the controller to constantly self-adapt to later changes in the motor characteristics caused by heating or wear-out while it is operating in standard control mode.
AB - Neural-Network (NN)-based controllers have the potential to achieve better control performance than classical PID controllers. Yet NN deployment on tiny microcontrollers, which are used in DC motor control due to strict cost requirements, is challenging as NNs are computationally intensive and memory demanding. We propose a lightweight direct inverse NN-based control approach for controlling the angular speed of a permanent magnet DC motor, which runs on a tiny Arm Cortex-M0 microcontroller with only 4 kB of RAM. Moreover, the NN-based controller can self-adapt to the DC motor characteristics without the need of any external machine learning frameworks such as TensorFlow. For this, we are not deploying a pre-trained network for inference but implement a fully automated training process on the microcontroller, which also includes the dataset collection. The result is a self-adaptive control algorithm that is able to drive the motor at the desired speed after it learned the motor characteristics in an initial training phase. Furthermore, the approach is extended such that it enables the controller to constantly self-adapt to later changes in the motor characteristics caused by heating or wear-out while it is operating in standard control mode.
KW - Extreme edge AI
KW - Microcontroller
KW - Neural network-based control algorithms
KW - On-device training
KW - Self-adaptation
KW - TinyML
UR - https://www.scopus.com/pages/publications/85101548521
U2 - 10.1007/978-3-030-66770-2_20
DO - 10.1007/978-3-030-66770-2_20
M3 - Conference contribution
AN - SCOPUS:85101548521
SN - 9783030667696
T3 - Communications in Computer and Information Science
SP - 268
EP - 279
BT - IoT Streams for Data-Driven Predictive Maintenance and IoT, Edge, and Mobile for Embedded Machine Learning - Second International Workshop, IoT Streams 2020, and First International Workshop, ITEM 2020, Co-located with ECML/PKDD 2020, Revised Selected Papers
A2 - Gama, Joao
A2 - Pashami, Sepideh
A2 - Bifet, Albert
A2 - Sayed-Mouchawe, Moamar
A2 - Fröning, Holger
A2 - Pernkopf, Franz
A2 - Schiele, Gregor
A2 - Blott, Michaela
PB - Springer Science and Business Media Deutschland GmbH
T2 - 2nd International Workshop on IoT Streams for Data-Driven Predictive Maintenance, IoT Streams 2020, and 1st International Workshop on IoT, Edge, and Mobile for Embedded Machine Learning, ITEM 2020, co-located with ECML/PKDD 2020
Y2 - 14 September 2020 through 18 September 2020
ER -