Ultra Low Power Wake-up Radio – U-WAKE

Four paths are followed:

(1) Energy Harvesting: The increase in the energy harvesting capabilities and the low energy consumption of the wake-up radios (WURs) make feasible the solution of having a main transceiver standby mode which is independent (or in certain scenarios with a very low energy consumption) from the main power supply (battery, or main supply).

To accomplish the goal of powering the WURs with RF energy, 3 key functions were identified:
• The conversion of RF energy into a DC voltage (unavoidable).
• The management of the harvested energy.
• The supply of a regulated voltage to the WURs (constraint of the project).

(2) SNN architecture for signal recognition: Wake-up receivers wake up a main receiver only when a specific signal is detected. However, most wake-up receivers are still relying on low power microcontrollers that perform signal recognition but consume peak powers higher than 200 microW, making IoT nodes unable to reach their optimal energy efficiency. Spiking Neural Networks (SNN) begin to emerge as a low power solution for data processing. Such bio-inspired systems can compute as fast as actual devices but consume less.

We propose and evaluate a spiking neural network architecture to detect an activation sequence. We want to reproduce the potential recognition of the address of a device with this low-power architecture.

(3) Waking-up interface: Requirements
The 2 parts of the full node operate at different voltages. To achieve ultra-low power, the neuromorphic MCU uses few hundreds of mV of supply voltage while the classical IoT node is powered by a voltage of few Volts. An interface circuit must therefore be added between the output of the neuromorphic MCU and the input of the main node MCU to increase the voltage up to few volts.

(4) RF front-end and the associated neural circuit: several architectures make it possible to reduce the energy consumption of the RF-front end. The energy aspect is not the only one to be considered: the reception sensitivity must be associated with the bandwidth compatible with the targeted application.

(1) Energy Harvesting: an interesting architecture that permits to supply a regulated voltage to the WUR and also store the excess of RF energy available in a battery for periods of shortage has been proposed. The global power consumed is one order of magnitude higher than the power consumed by the WURs and this difference of magnitudes can not be reduced with currently available PMUs or voltage regulators. Although the use of a Power Managment Unit in a RF harvesting architecture is well adapted for a fluctuating RF environment, this architecture is not capable to assure a permanent source of energy for the WURs. The use of a primary battery conjointly with a rectifier and the AEM30940 PMIC was studied. This external battery relays the rectifier when there is not enough RF power. It was found a coin cell battery could last 55.5 years if the AEM30940 exits the normal mode every two minutes, however this time is overestimated with regards to the shelf time of a battery which in the best cases is around 20 years.

(2) SNN architecture for signal recognition: a compromise between accuracy and energy consumption is to be made. We observed that we cannot receive our bits too fast because of the dynamic of the neuron: for an optimal detection we face a limitation in the flow rate. This limitation is however not really an issue because we only want to wake-up the node, and not to transmit data. We discussed about an optimal threshold for our network depending on the target activation sequence. Nonetheless, this solution is even more low-power than WUR with dedicated microcontrollers for pattern recognition, which is promising. Further work will be dedicated to study a few other parameters to improve our performances, like adapting the neuron threshold to optimize the ratio between false alarm and miss detection, to study the impact of noise inside the neurons and define optimal codes for this kind of receivers.

(3) Waking-up interface: While the considered TSMC 65nm technology (used for the neuromorphic MCU) could provide at most 1.2 V, common MCU of the main node need at least 1.35V as an interrupt to be woken-up. Measurements must be performed to check theses values experimentally.

(4) RF front-end and the associated neural circuit: The sensitivity of this first version of the CMOS chip could not be evaluated experimentally because of the probes and associated cables which induce parasitic capacitive effects. These forbid to perform a frequency tuning using the matching network.
The next step consists in realizing an appropriate conditioning of the chip which allows to adapt the RF receiver to the chosen central frequency.

The project is progressing according to plan. As expected, the major difficulty lies in the interfacing of the different parts of the node that will be built. The cooperation between the three partners is key in this respect and efforts in the coming period will be focused on these aspects. The two PhD students recruited (IRISA and CITI) contribute significantly to this cohesion and their work is progressing well.

The key findings of the first half are very encouraging. There is still some open issues, one being managing time and delays in analog circuits. This is probably a cornerstone to enable ultra-low power consumption in circuits with a minimum ability to recover information. This project addresses this issue and spiking neural networks could reduce energy consumption by several orders of magnitude, making IoT possible.

Publications issued from the project are under submission, not accepted yet. We can however cite a few recent related papers:

[Fumtchum2021] Fumtchum, A., Tsafack, P., Hutu, F. D., Villemaud, G., & Tanyi, E. (2021). A Survey of RF Energy Harvesting Circuits. International Journal of Innovative Technology and Exploring Engineering, 10(7), 99-106.

[Djidi2021] N.E.H. Djidi, M. Gautier, A. Courtay and O.~Berder, “MEES-WuR: Minimum Energy Coding with Early Shutdown for Wake-up Receivers”, IEEE Transactions on Green Communications and Networking, 2021.

[Danneville2019] Danneville, F., Loyez, C., Carpentier, K., Sourikopoulos, I., Mercier, E., & Cappy, A. (2019). «A Sub-35 pW Axon-Hillock artificial neuron circuit«. Solid-State Electronics, 153, 88–92.

U-Wake project aims to achieve a breakthrough in the field of IoT by developing a disruptive wake-up receiver solution based on (1) a bioinspired architecture achieved with an industrial CMOS technology and (2) Electro Magnetic energy harvesting.

Internet of Things (IoT) is becoming a reality. It will greatly impact our daily lives (city, housing, transportation, health, environment) and many economic sectors (agriculture, industry...). Unlicensed bands (868 MHz, 2.4 GHz) play an important role in this evolution with technologies like LoRa, SigFox or IEEE 802.15.4. However, energy consumption remains a major bottleneck, with many applications requiring the lifespan of objects to reach several years, even decades, without changing the batteries. Many efforts have been deployed to push the boundaries of energy autonomy, without however a full success. The radio transceiver often turns out to be the most energy consuming part of a wireless node, due to both the transmitting and also receiving phases. For instance, initiating a communication requires that the source and the destination are awake at the same time. It can be difficult to plan and usually requires some highly penalizing signalling protocols. In short range multi-hop networks, consuming MAC strategies are implemented in order to synchronize the source and the destination. Low Power Wide Area Networks solved this issue by having always turned-on base stations using single hop communications and a simple ALOHA protocol, but this only works for the uplink. Wake-up receivers (WUR) form an emerging technology, which allows continuous channel monitoring, while consuming orders of magnitude less power than traditional receivers. These receivers wake up a main transceiver using interrupts only when a specific signal is detected. Thus, fully asynchronous communication can be achieved, resulting in a huge decrease of energy waste. However, most wake-up receivers are still relying on low power microcontrollers that perform signal recognition but consume peak powers higher than 200 µW, making IoT nodes unable to reach their ultimate energy efficiency.

The proposed receiver will be woken up when detecting a dedicated off-line learned sequence and implemented in a hardware fashion using an Ultra-Low Power (ULP) SNN. The main advantage of such a design is that it requires a few mW or less for the whole wake-up receiver. Furthermore, it can work in the 868 MHz or 2.4 GHz bands and has the ability to recognize different types of signals (on-off keying, BPSK or chirp spread spectrum modulation for instance). Requiring such a low consumption opens up the possibility to be powered using RF energy harvesting or Wireless Power Transfer, and opens the way to a wide range of applications. We aim at developing a complete autonomous sensor node prototype in ISM bands (868 MHz or 2.4 GHz) with a power consumption below 10 µW and keeping a sensitivity at about -90 dBm, such performance having never been obtained altogether.

The project will be organized in three WPs: WP1, Prerequisites, led by INSA Lyon, will address the harvesting, storage and management of energy in order to power the node and make it autonomous. It will also study the choice of the appropriate waveforms, the design of the SNN and the theoretical detection performance of the proposed scheme. WP2, Neuro-inspired circuit, led by IEMN, will develop the new architectures. Two runs are planned at the end of the second year and at the end of the third year. WP3, Prototypes, led by IRISA, will address the interfacing challenges and the realization of the full node for real life IoT solutions.