Low-power Microcontroller Unit Research Articles

Optimizing Convolutional Neural Networks for Image Classification on Resource-Constrained Microcontroller Units

Running machine learning algorithms for image classification locally on small, cheap, and low-power microcontroller units (MCUs) has advantages in terms of bandwidth, inference time, energy, reliability, and privacy for different applications. Therefore, TinyML focuses on deploying neural networks on MCUs with random access memory sizes between 2 KB and 512 KB and read-only memory storage capacities between 32 KB and 2 MB. Models designed for high-end devices are usually ported to MCUs using model scaling factors provided by the model architecture’s designers. However, our analysis shows that this naive approach of substantially scaling down convolutional neural networks (CNNs) for image classification using such default scaling factors results in suboptimal performance. Consequently, in this paper we present a systematic strategy for efficiently scaling down CNN model architectures to run on MCUs. Moreover, we present our CNN Analyzer, a dashboard-based tool for determining optimal CNN model architecture scaling factors for the downscaling strategy by gaining layer-wise insights into the model architecture scaling factors that drive model size, peak memory, and inference time. Using our strategy, we were able to introduce additional new model architecture scaling factors for MobileNet v1, MobileNet v2, MobileNet v3, and ShuffleNet v2 and to optimize these model architectures. Our best model variation outperforms the MobileNet v1 version provided in the MLPerf Tiny Benchmark on the Visual Wake Words image classification task, reducing the model size by 20.5% while increasing the accuracy by 4.0%.

Tiny-Machine-Learning-Based Supply Canal Surface Condition Monitoring.

The South-to-North Water Diversion Project in China is an extensive inter-basin water transfer project, for which ensuring the safe operation and maintenance of infrastructure poses a fundamental challenge. In this context, structural health monitoring is crucial for the safe and efficient operation of hydraulic infrastructure. Currently, most health monitoring systems for hydraulic infrastructure rely on commercial software or algorithms that only run on desktop computers. This study developed for the first time a lightweight convolutional neural network (CNN) model specifically for early detection of structural damage in water supply canals and deployed it as a tiny machine learning (TinyML) application on a low-power microcontroller unit (MCU). The model uses damage images of the supply canals that we collected as input and the damage types as output. With data augmentation techniques to enhance the training dataset, the deployed model is only 7.57 KB in size and demonstrates an accuracy of 94.17 ± 1.67% and a precision of 94.47 ± 1.46%, outperforming other commonly used CNN models in terms of performance and energy efficiency. Moreover, each inference consumes only 5610.18 μJ of energy, allowing a standard 225 mAh button cell to run continuously for nearly 11 years and perform approximately 4,945,055 inferences. This research not only confirms the feasibility of deploying real-time supply canal surface condition monitoring on low-power, resource-constrained devices but also provides practical technical solutions for improving infrastructure security.

Reducing False Alarms in Wearable Seizure Detection With EEGformer: A Compact Transformer Model for MCUs.

The long-term, continuous analysis of electroencephalography (EEG) signals on wearable devices to automatically detect seizures in epileptic patients is a high-potential application field for deep neural networks, and specifically for transformers, which are highly suited for end-to-end time series processing without handcrafted feature extraction. In this work, we propose a small-scale transformer detector, the EEGformer, compatible with unobtrusive acquisition setups that use only the temporal channels. EEGformer is the result of a hardware-oriented design exploration, aiming for efficient execution on tiny low-power micro-controller units (MCUs) and low latency and false alarm rate to increase patient and caregiver acceptance.Tests conducted on the CHB-MIT dataset show a 20% reduction of the onset detection latency with respect to the state-of-the-art model for temporal acquisition, with a competitive 73% seizure detection probability and 0.15 false-positive-per-hour (FP/h). Further investigations on a novel and challenging scalp EEG dataset result in the successful detection of 88% of the annotated seizure events, with 0.45 FP/h.We evaluate the deployment of the EEGformer on three commercial low-power computing platforms: the single-core Apollo4 MCU and the GAP8 and GAP9 parallel MCUs. The most efficient implementation (on GAP9) results in as low as 13.7 ms and 0.31 mJ per inference, demonstrating the feasibility of deploying the EEGformer on wearable seizure detection systems with reduced channel count and multi-day battery duration.

Marine magnetotelluric and controlled source electromagnetics signal simulative generator

Ocean bottom electromagnetic receivers (OBEMs) are invaluable tools for observing seafloor electromagnetic signals, primarily used in magnetotelluric (MT) sounding and marine controlled source electromagnetics (MCSEM). However, the high operational costs of deploying OBEMs require thorough pre-deployment testing to enhance the data reliability and quality. Commercial signal generators used for MT and controlled source electromagnetic (CSEM) testing often fall short of the restricted terms of the frequency range, signal-to-noise ratio, storage depth, time synchronization, and amplitude dynamics. To address these limitations, we developed a specialized marine MT and CSEM signal simulative generator. The generator employs pseudorandom number code-based white noise to simulate MT signals and uses multi-frequency synthesis and amplitude modulation (AM) to generate CSEM signals. The system is developed using a low-power microcontroller unit and a field programmable gate array platform, integrating various circuits such as power, clock, encoding, control, data storage, chopper, and amplification circuits. The generator simultaneously outputs four-channel white noise and AM signals. The white noise feature provides a wide and adjustable frequency band and amplitude, while the AM feature provides a large dynamic amplitude range and multi-frequency synthesis at various carrier frequency points. Comprehensive testing verified the performance of the proposed simulative generator. The test results demonstrate that the MT simulated signals cover a frequency range from 1 E4s to 100 Hz, providing adjustable amplitudes between 10 μVpp and 10 mVpp. The CSEM simulated signals include standard square waves, third harmonic waves, and seventh harmonic waves, with amplitudes adjustable from 10 mVpp to 10 Vpp. The proposed simulative generator enhances the efficiency of OBEM functional testing and provides essential equipment support for maintaining the quality of marine data.

Federated learning on low-power Arduino Nano33 BLE Sense to predict the length of stay using a linear regression model

Federated learning (FL) is a collaborative learning paradigm where multiple clients are used to build the model without sharing data and preserving privacy. An FL-based linear regression model is designed to predict the length of stay for patients at hospitals using the low-power Arduino Nano 33 BLE Sense microcontroller unit (MCU). FL uses a distributed learning technique that allows model building from decentralized data sources. The Arduino Nano 33 BLE Sense is a compact and energy-efficient MCU providing an ideal platform for implementing FL in resource-constrained environments. FL algorithms aggregate model parameters from multiple Arduino clients and collectively train and build a predictive model to estimate the length of stay at the hospital by patients. Experiments were conducted to understand the performance of FL on clients with data of equal and varying sizes and heterogeneous data from multiple sources. The performance of the algorithm is evaluated based on Mean Absolute Error (MAE), Percentage Decrease in Training error (PDTE), and Percentage Difference with Optimal Testing (PDOT) value. Experimental results show that the number of local epochs and FL rounds affects the convergence of clients to the optimal value. The experimental results demonstrate the applicability of FL on low-power MCUs, preserving privacy which is a core requirement for healthcare solutions.

A 22-m Operation Range Semi-Passive UHF RFID Sensor Tag With Flexible Thermoelectric Energy Harvester

This article presents a wireless temperature sensor tag able to work in both fully passive mode and in semi-passive mode when assisted by a flexible thermoelectric generator (TEG). The sensor tag consists of an EPC C1G2/ISO 18000-6C ultrahigh-frequency (UHF) radio frequency identification (RFID) integrated circuit (IC) connected to a low-power microcontroller unit (MCU) that samples and collects the temperature from a digital temperature sensor. With a temperature gradient as low as 2.5 °C, the test results show that the TEG provides an output power of 400 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mu$ </tex-math></inline-formula> W with an output voltage of 40mV. Bymeans of an up-converter in order to boost the TEG output voltage, this harvester supplies the power required to the sensor tag for a 2-conv/s data rate in semi-passive mode. Moreover, when the tag operates in semi-passive mode, a communication range of 22.2 m is measured for a 2-W effective radiated power (ERP) reader. To the best of our knowledge, the proposed TEG-assisted sensor tag shows the longest communication range and the only one that provides stable external power at low-temperature gradients. The measured performance and the chosen architecture allow using the wireless sensor in multiple industrial or biomedical applications.

Development of a low-power weigh-in-motion system using cylindrical piezoelectric elements

This paper presents the development of a novel low power weigh-in-motion (WIM) system that uses cylindrical piezoelectric (PZT) elements for the dual purpose of sensing axle loads and harvesting mechanical energy for its operation. It provides details on the characterization the PZT sensing elements, the conditioning of their signals and describes the algorithms developed for determining speed, axle load and vehicle classification. These algorithms were coded in MATLAB® and converted to C in a format suitable for installing in a low power microcontroller unit (MCU). The system has the capabilities of monitoring vehicle speed, number of axles, axle spacing, axle loads and vehicle classification. It was tested in the laboratory by applying a range of loads and loading frequencies through a servo-hydraulic loading system. The results suggest sufficient accuracy and precision in measuring vehicle speeds, axle loads and determining vehicle class. Its low power requirements provide an inexpensive and sustainable method for obtaining roadway traffic data.

Design and Evaluation of Low-Power Co3O4 Gas Sensing Element as a Part of Cyber Physical Systems

Physical processes working in parallel with digital ones have transformed the way we view systems and have led to the creation of applications that boost the quality of people’s lives, increase security as well as decrease production costs of goods. Critical to this evolution is the cost decrease in the components of such systems, among which are gas sensors. In this work, a custom-made Co3O4 gas sensing element is presented, which can potentially be used as part of a cyber-physical system (CPS) for O3 monitoring. To investigate its performance, a CPS is developed using low-cost, low-power micro-controller units (MCUs) and comparisons both with the laboratory equipment and commercial off-the-shelf (COTS) ozone sensors are provided. The experiments show that the Co3O4 sensor works at room temperature with low input voltage and low power consumption when used with the proposed MCUs. Moreover, an enhanced gas sensing performance against ozone is observed under low-pressure conditions due to the detection of low ozone concentrations (85.90 ppb) and good sensor response (113.1%) towards 1100 ppb O3. However, the drawbacks that need improvement relate to the kinetics of the charge carriers, which affect the response time and recovery behavior. The effect of humidity needs to be clarified in further works.

Sub-100 $\mu$W Multispectral Riemannian Classification for EEG-Based Brain–Machine Interfaces

Motor imagery (MI) brain-machine interfaces (BMIs) enable us to control machines by merely thinking of performing a motor action. Practical use cases require a wearable solution where the classification of the brain signals is done locally near the sensor using machine learning models embedded on energy-efficient microcontroller units (MCUs), for assured privacy, user comfort, and long-term usage. In this work, we provide practical insights on the accuracy-cost trade-off for embedded BMI solutions. Our multispectral Riemannian classifier reaches 75.1% accuracy on a 4-class MI task. The accuracy is further improved by tuning different types of classifiers to each subject, achieving 76.4%. We further scale down the model by quantizing it to mixed-precision representations with a minimal accuracy loss of 1% and 1.4%, respectively, which is still up to 4.1% more accurate than the state-of-the-art embedded convolutional neural network. We implement the model on a low-power MCU within an energy budget of merely 198 μJ and taking only 16.9 ms per classification. Classifying samples continuously, overlapping the 3.5 s samples by 50% to avoid missing user inputs allows for operation at just 85 μW. Compared to related works in embedded MI-BMIs, our solution sets the new state-of-the-art in terms of accuracy-energy trade-off for near-sensor classification.

Trends in Intelligent Communication Systems: Review of Standards, Major Research Projects, and Identification of Research Gaps

The increasing complexity of communication systems, following the advent of heterogeneous technologies, services and use cases with diverse technical requirements, provide a strong case for the use of artificial intelligence (AI) and data-driven machine learning (ML) techniques in studying, designing and operating emerging communication networks. At the same time, the access and ability to process large volumes of network data can unleash the full potential of a network orchestrated by AI/ML to optimise the usage of available resources while keeping both CapEx and OpEx low. Driven by these new opportunities, the ongoing standardisation activities indicate strong interest to reap the benefits of incorporating AI and ML techniques in communication networks. For instance, 3GPP has introduced the network data analytics function (NWDAF) at the 5G core network for the control and management of network slices, and for providing predictive analytics, or statistics, about past events to other network functions, leveraging AI/ML and big data analytics. Likewise, at the radio access network (RAN), the O-RAN Alliance has already defined an architecture to infuse intelligence into the RAN, where closed-loop control models are classified based on their operational timescale, i.e., real-time, near real-time, and non-real-time RAN intelligent control (RIC). Different from the existing related surveys, in this review article, we group the major research studies in the design of model-aided ML-based transceivers following the breakdown suggested by the O-RAN Alliance. At the core and the edge networks, we review the ongoing standardisation activities in intelligent networking and the existing works cognisant of the architecture recommended by 3GPP and ETSI. We also review the existing trends in ML algorithms running on low-power micro-controller units, known as TinyML. We conclude with a summary of recent and currently funded projects on intelligent communications and networking. This review reveals that the telecommunication industry and standardisation bodies have been mostly focused on non-real-time RIC, data analytics at the core and the edge, AI-based network slicing, and vendor inter-operability issues, whereas most recent academic research has focused on real-time RIC. In addition, intelligent radio resource management and aspects of intelligent control of the propagation channel using reflecting intelligent surfaces have captured the attention of ongoing research projects.

Development of a feedback interface for in-situ soundscape evaluation

Studies involving subjective evaluation require feedback from human participants to assess the performance of a system or an environment. A participant is typically presented with a set of metrics to be observed and they present their assessment accordingly. Investigator-led in-situ soundscape evaluation in ISO 12913-2 collects perceptual responses along with other acoustical and locale information. This is a labor intensive and time-consuming processes. To alleviate and complement investigator-led evaluations, a portable and compact feedback system with an e-ink display and capacitive buttons was designed, and is undergoing field tests to address the aforementioned requirements. The system employs a low-cost, low-power microcontroller unit (MCU) with necessary hardware interfaces to enable capacitive sensing. Capacitive buttons provide an intuitive interface and avoid the inherent wear and tear of mechanical buttons. This digitized feedback interface affords the flexibility to synchronize (wired or wirelessly) with a playback system to evaluate an augmented soundscape, and is suitable for both supervised and unsupervised subjective assessments.

Cost-Effective Test Screening Method on 40-nm Embedded SRAMs for Low-Power MCUs

Embedded static random access memories (SRAMs) with cost-effective test screening circuitry are demonstrated for low-power microcontroller units (MCUs). The probing test step at the low temperature (LT) of -40 °C is obviated by imitating pseudo-LT (PLT) conditions in the package test, where a sample is measured at room temperature (RT). Monte Carlo simulation is carried out considering local V <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</sub> variations as well as contact soft open failure (high resistance), confirming good minimum operating voltage ( V <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> ) correlation between LT and PLT conditions. Test chips with two types of 4-Mbit single-port SRAM macros and 1-Mbit dual-port SRAM macro are designed and fabricated using low-power 40-nm CMOS technology. Measurement results demonstrate that the proposed test method reproduces LT conditions and screens out LT failures with less overscreening. The proposed test method eliminates 1/3 or more of the test costs.

SSVM: An Ultra-Low-Power Strain Sensing and Visualization Module for Long-Term Structural Health Monitoring.

Structural health monitoring (SHM) is crucial for quantitative behavioral analysis of structural members such as fatigue, buckling, and crack propagation identification. However, formerly developed approaches cannot be implemented effectively for long-term infrastructure monitoring, owing to power inefficiency and data management challenges. This study presents the development of a high-fidelity and ultra-low-power strain sensing and visualization module (SSVM), along with an effective data management technique. Deployment of 24-bit resolution analog to a digital converter and precise half-bridge circuit for strain sensing are two significant factors for efficient strain measurement and power management circuit incorporating a low-power microcontroller unit (MCU), and electronic-paper display (EPD) enabled long-term operation. A prototype for SSVM was developed that performs strain sensing and encodes the strain response in a QR code for visualization on the EPD. For efficient power management, SSVM only activated when the trigger-signal was generated and stayed in power-saving mode consuming 18 mA and 337.9 μA, respectively. The trigger-signal was designed to be generated either periodically by a timer or intentionally by a push-button. A smartphone application and cloud database were developed for efficient data acquisition and management. A lab-scale experiment was carried out to validate the proposed system with a reference strain sensing system. A cantilever beam was deflected by increasing load at its free end, and the resultant strain response of SSVM was compared with the reference. The proposed system was successfully validated to use for long-term static strain measurement.

Gate Oxide and Implantation Process Co-Optimization for Low-Power MCU Applications

The fast development in microcontroller unit (MCU) technology has urged continuous decreasing in power consumption by different assignment of operating status among devices. In this work, we focused on the ultrahigh Vth (UHVT) transistor and used gate oxide thickness and Vth implantation co-optimization to minimize the gate leakage current towards low-power MCU applications. Based on the 55 nm node, it has been found both theoretically and experimentally that the leakage level has been significantly reduced at different temperature in n-FET, p-FET, data flip-flop (DFF) and inverter (INV). Upon MCU testing under active, sleep and deep sleep modes, obvious decrease in the power consumption is also achieved, providing a promising optimization approach towards a better balance between the speed and power in modern MCU technology.

Design and implementation of low power bio signal sensors for wireless body sensing network applications

Continuous human surveillance alone will be useful for elderly people to realize the highest standard of living in human society. However, it is difficult to get such a setting as long as it can operate long hours without maintenances of its battery limitation. Biosensor signals, which are referred to as a device that consumes a lot of power in the system of a Microcontroller Unit (MCU), must be made continuously to signal samples of the biosensor. To decrease the power utilization of the Microcontroller Unit, and a low power application specific integrated circuit for biosensors feature extraction is designed for the replacement of a role of the microcontroller unit to extract the characteristic values of bio signal. In this investigation, an Application-Specific Integrated Circuit (ASIC) of Wireless Body Sensor Networks (WBSNs) is intended to fill in as a convention for medical use. The proposed wireless body sensor network comprise an newly designed sensor interfaces, integrated passive Radio Frequency (RF) receiver, contact ports, low-power Microcontroller Unit (MCU), wireless transistor control, a Power Management Unit (PMU), and an with low power harvesting capacity contains. The proposed method has been designed using 90 nm and 45 nm CMOS technology. Simulation results exhibit up to the change in noise immunity and lessening in power utilization contrasted with the traditional wireless body sensor networks design at a similar delay.

Enhancing the performance of electromagnetic damper cum energy harvester using microcontroller: Concept and experiment validation

Abstract An electromagnetic damper cum energy harvester (EMDEH), which consists of an electromagnetic damper (EMD) connected to an energy harvesting circuit (EHC), aims to suppress structural vibrations by harvesting the vibration energy of structures. Maintaining a steady equivalent resistance of an EHC is often desirable for achieving the target performance of vibration control and energy harvesting functions; however, it remains an unresolved challenge in the EHC design. This study proposes an improved version of EMDEH with a specially designed EHC, a buck–boost converter with a low-power microcontroller unit (MCU). Compared with the existing EMDEH using an EHC with a fixed duty cycle, the improved version employs the MCU to adjust the duty cycle based on the feed-forward signal so as to continuously maintain a nearly constant resistance characteristic of the EHC. The effectiveness of the proposed EHC was validated through numerical simulations and experimental tests. A prototype of the improved EMDEH was installed in a 135 m-long bridge stay cable and tested under harmonic excitations of different levels. An output power of 200 mW and a nearly constant resistance of the proposed EHC were observed in this full-scale experiment. The results of this study show how the use of a MCU in a feed-forward controlled EHC benefits the performance of a dual-function EMDEH with respect to its simultaneous vibration control and energy harvesting performance.

Accurate Indoor Sound Level Measurement on a Low-Power and Low-Cost Wireless Sensor Node.

Wireless sensor networks can provide a cheap and flexible infrastructure to support the measurement of noise pollution. However, the processing of the gathered data is challenging to implement on resource-constrained nodes, because each node has its own limited power supply, low-performance and low-power micro-controller unit and other limited processing resources, as well as limited amount of memory. We propose a sensor node for monitoring of indoor ambient noise. The sensor node is based on a hardware platform with limited computational resources and utilizes several simplifications to approximate more complex and costly signal processing stage. Furthermore, to reduce the communication between the sensor node and a sink node, as well as the power consumed by the IEEE 802.15.4 (ZigBee) transceiver, we perform digital A-weighting filtering and non-calibrated calculation of the sound pressure level on the node. According to experimental results, the proposed sound level meter can accurately measure the noise levels of up to 100 dB, with the mean difference of less than 2 dB compared to Class 1 sound level meter. The proposed device can continuously monitor indoor noise for several days. Despite the limitations of the used hardware platform, the presented node is a promising low-cost and low-power solution for indoor ambient noise monitoring.

Tracking Elite Swimmers in Real Time with Wearable Low-power Wireless Sensor Networks

Monitoring elite athletes in real-time remains challenging. A lot of research has been conducted towards the application of inertial measurement systems for the quantification of performance in sports applications. The challenge however, remains in creating systems that meet high comfort standards in order to be able to apply them on a daily basis. The problem is mainly related to the requested volume of existing systems, which can be linked to the high power requirements and hence large batteries. This paper presents research towards the performance assessment of elite swimmers using wearable low-power sensor networks. This is the first work to focus on the reduction of power consumption by optimizing the power cost of signal processing and sensor use on a small and comfortable system. A built-in algorithm exploits the sensor data in order to predict which information is valuable to assess the motion of the athlete in real-time. Moreover, the algorithm is designed to run on a low-power microcontroller unit and uses lightweight computational techniques that further reduce the power requirements. By intelligently putting the sensor system into a low-power mode whenever possible, a reduction in sensor power consumption of 96% is estimated over minimally 80% of the swimming time, depending on the swimmer's lap times. Stroke types and turns were detected with high accuracy for each tested swimmer, despite different stroke techniques.

On-demand Customizable Wireless Sensor Network

In this paper, we propose a wireless sensor network (WSN) whose behavior can be dynamically customized by injecting programs or roles specified by the user. To enable easy specification of the roles, a role-generation mechanism is also proposed. To realize the WSN, we introduce a reconfigurable wireless sensor node that has an ultra-low-power field-programmable gate array (FPGA) as well as a low-power microcontroller unit (MCU). By injecting several different roles into the sensor nodes, we confirmed that the behavior of the WSN can be changed on demand.

Towards a wireless optical stimulation system for long term in-vivo experiments.

This paper presents our recent progresses towards the development of a wirelessly powered head mountable optical stimulator for enabling long-term optogenetic experiments with small freely moving transgenic models. The proposed system includes a wireless power transmission chamber with uniform power distribution in 3D and a wireless head mountable optical stimulator prototype with power recovery. The wireless power link, which includes the inductive chamber and power recovery circuits, is robust against subject movements in all directions, and against angular misalignment. Such link provides uniform power distribution without the need for a closed-loop control system, and can localize the transmitted power towards the receiver, without using additional detection and control circuitry compared to other systems. Additionally, the chamber is equipped with a camera for capturing the animal motion and behavior after applying optical stimulation patterns. A low-power microcontroller unit is embedded with the stimulator prototype to generate arbitrary light stimulation patterns. Measurement results show that the inductive chamber can continuously deliver 70 mW to the stimulator prototype with a power efficiency of 59%.