One of my favorite quotes comes from the IEEE Signal Processing magazine in 2010. They attempted to answer the question: What does ultra-low power consumption mean? And they came to the conclusion that it is where the “power source lasts longer than the useful life of the product.”[1] It’s a great answer because it’s scalable. It applies equally to signal processing circuitry inside an embedded IoT device that can never be accessed or recharged and to signal processing inside a car where the petrol for the engine dominates the operating lifetime, not the signal processing power. It also describes exactly what a lot of science fiction has always envisioned: no changing or recharging of batteries, which people forget to do or never have enough batteries for. Rather, we have devices that simply always work.
My research focuses on healthcare applications and creating “wearable algorithms”—that is, signal processing implementations that fit within the very small power budgets available in wearable devices. Historically, this focused on data reduction to save power. It’s well known that wireless data transmission is very power intensive. By using some power to reduce the amount of data that has to be sent, it’s possible to save lots of power in the wireless transmission stage and so to increase the overall battery lifetime.
This argument has been known for a long time. There are papers dating back to at least the 1990s based on it. It’s also readily achievable. Inevitably, it depends on the precise situation, but we showed in 2014 that the power consumption of a wireless sensor node could be brought down to the level of a node without a wireless transmitter (one that uses local flash memory) using easily available, easy-to-use, off-the-shelf-devices.[2]
Today, there are many additional benefits that are being enabled by the emerging use of ultra-low power signal processing embedded in the wearable itself, and these new applications are driving the research challenges: increased device functionality; minimized system latency; reliable, robust operation over unreliable wireless links; reduction in the amount of data to be analyzed offline; better quality recordings (e.g., with motion artifact removal to prevent signal saturations); new closed-loop recording—stimulation devices; and real-time data redaction for privacy, ensuring personal data never leaves the wearable.
It’s these last two that are the focus for my research now. They’re really important for enabling new “bioelectronic” medical devices which apply electrical stimulation as an alternative to classical pharmacological treatments. These “bioelectronics” will be fully data-driven, analyzing physiological measurements in real-time and using this to decide when to optimally trigger an intervention. Doing such as analysis on a wearable sensor node though requires ultra-low power signal processing that has all of the feature extraction and signal classification operating within a power budget of a few 100 µW or less.
To achieve this, most works do not use any specific software platform. Instead they achieve very low-power consumption by using only dedicated and highly customized hardware circuits. While there are many different approaches to realizing low-power fully custom electronics, for the hardware, the design trends are reasonably established: very low supply voltages, typically in the 0.5 to 1 V range; highly simplified circuit architectures, where a small reduction in processing accuracy leads to substantial power savings; and the use of extensive analogue processing in the very lowest power consumption circuits.[3]
— ADVERTISMENT—
—Advertise Here—
Less well established are the signal processing functions for ultra-low power. Focusing on feature extractions, our 2015 review highlighted that the majority (more than half) of wearable algorithms created to date are based upon frequency information, with wavelet transforms being particularly popular.[4] This indicates a potential over-reliance on time–frequency decompositions as the best algorithmic starting points. It seems unlikely that time–frequency decompositions would provide the best, or even suitable, feature extraction across all signal types and all potential applications. There is a clear opportunity for creating wearable algorithms that are based on other feature extraction methods, such as the fractal dimension or Empirical Mode Decomposition.
Investigating this requires studying the three-way trade-off between algorithm performance (e.g., correct detections), algorithm cost (e.g., false detections), and power consumption. We know how to design signal processing algorithms, and we know how to design ultra-low power circuitry. However, combining the two opens many new degrees of freedom in the design space, and there are many opportunities and work to do in mapping feature extractions and classifiers into sub-1-V power supply dedicated hardware.
[1] G. Frantz, et al, “Ultra-low power signal processing,” IEEE Signal Processing Magazine, vol. 27, no. 2, 2010.
[2] S. A. Imtiaz, A. J. Casson, and E. Rodriguez-Villegas, “Compression in Wearable Sensor Nodes,” IEEE Transactions on Biomedical Engineering, vol. 61, no. 4, 2014.
[3] A. J. Casson, et al, “Wearable Algorithms,” in E. Sazonov and M. R. Neuman (eds.), Wearable Sensors, Elsevier, 2014.
[4] A. J. Casson, “Opportunities and Challenges for Ultra Low Power Signal Processing in Wearable Healthcare,” 23rd European Signal Processing Conference, Nice, 2015.
Alex Casson is a lecturer in the Sensing, Imaging, and Signal Processing Department at the University of Manchester. His research focuses on creating next-generation human body sensors, developing both the required hardware and software. Dr. Casson earned an undergraduate degree at the University of Oxford and a PhD from Imperial College London.
