Mid Sweden University

Energy harvesting has been proposed to support or replace battery-based energy supplies in low-power loT systems. However, manual design and optimization is currently required to implement energy harvesters. This implies significant time and cost requirements, reducing the competitiveness of energy harvesting solutions in many applications. In this paper, we propose and study a method for the automated design of ortho-planar springs in vibration energy harvesters. The proposed method uses a Python tool that has been developed to translate parametric descriptions of the spring into 2D and 3D spring CAD models. The tool is combined with a black-box optimization algorithm in order to identify optimized spring parameters. An evaluation of the proposed method on a case study of an electromagnetic vibration energy harvester demonstrates that the approach successfully results in spring designs that perform well in the energy harvesting application. Consequently, the requirement of expert competence to adjust an energy harvester to new application constraints is significantly reduced, contributing to the availability and competitiveness of energy harvesting solutions in real world applications. 

Visual structural inspections in Structural Health Monitoring (SHM) are an important method to ensure the safety and long lifetime of infrastructures. Unmanned Aerial Vehicles (UAVs) with Deep Learning (DL) have gained in popularity to automate these inspections. Yet, the vast majority of research focuses on algorithmic innovations that neglect the availability of reliable generalized DL models, as well as the effect that the model's energy consumption would have on the UAV flight time. This paper highlights the performance of 14 popular CNN models with less than six million parameters for crack detection in concrete structures. Seven of these models were successfully deployed to a low-power, resource-constrained mi-crocontroller using Tiny Machine Learning (TinyML). Among the deployed models, MobileNetV1-x0.25 achieves the highest test accuracy (75.83%) and F1-Score (0.76), the second-lowest flash memory usage (273.5 kB), the second-lowest RAM usage (317.1kB), the fourth-fastest single-trial inference time (15.8ms), and the fourth-lowest number of Multiply-Accumulate operations (MACC) (42126514). Lastly, a hypothetical study of the DJI Mini 4 Pro UAV demonstrated that the TinyML model's energy consumption has a negligible impact on the UAV flight time (34 minutes vs. 33.98 minutes). Consequently, this feasibility study paves the way for future developments towards more efficient, autonomous unmanned structural health inspections. 

With the advent of low-cost and low-power computation, communication and sensor devices, novel instrumentation and measurement applications have been enabled, such as real-time industrial condition monitoring and fine-grained environmental monitoring. In these application scenarios, a lack of available infrastructures for communication and power supply is a common problem. In industrial applications, for example, the machine to be monitored and the monitoring system itself have significantly different technology lifespans, which requires that the monitoring system be retrofitted to machines that are already in use. In environmental monitoring, measurement systems are deployed as standalone devices in potentially remote areas. Consequently, the more autonomous the sensor system can be in terms of required infrastructure, the better it can match application and business needs.

Precision livestock farming (PLF) leverages cutting-edge technologies and data-driven solutions to enhance the efficiency of livestock production, its associated management, and its welfare. Continuous monitoring of the masticatory sound of cattle allows the estimation of dry-matter intake, classification of jaw movements (JMs), and recognition of grazing and rumination bouts. Over the past two decades, algorithms for analyzing feeding sounds have seen improvements in performance and computational requirements. Nevertheless, in some cases, these algorithms have been implemented on resource-constrained electronic devices, limiting their functionality to one specific task: either classifying JMs or recognizing feeding activities (such as grazing and rumination). In this work, we present an acoustic monitoring system that comprehensively analyzes grazing cattle's feeding behavior at multiple scales. This embedded system classifies different types of JMs, identifies feeding activities, and provides predictor variables for estimating dry-matter intake. Results are transmitted remotely to a base station using long-range communication (LoRa). Two variants of the system have been deployed on a Raspberry Pi Pico board, based on a low-power ARM Cortex-M0+ microcontroller. Both firmware versions make use of direct access memory, sleep mode, and clock-gating techniques to minimize energy consumption. In laboratory experiments, the first deployment consumes 20.1 mW and achieves an F1-score of 87.3% for the classification of JMs and 87.0% for feeding activities. The second deployment consumes 19.1 mW and reaches an F1-score of 84.1% for JMs and 83.5% for feeding activities. The modular design of the proposed embedded monitoring system facilitates integration with energy-harvesting power sources for autonomous operation in field conditions.

Electro-mechanical systems operating in periodic cycles are pivotal in the Industry 4.0, enabling automated processes that enhance efficiency and productivity. Early detection of failures and anomalies in duty cycles of these machines is crucial to ensure uninterrupted operation and prevent costly downtimes. Although the wear and damage of machines have been extensively studied, a significant proportion of these problems can be traced back to operator errors, underlining the importance of continuously monitoring the machine activity to ensure optimal performance. This work presents an automatic algorithm designed to identify improper duty cycles of industrial machines, exemplified on a mining conveyor belt. To enable the identification of duty cycles, the operational states of the machine are first categorized using machine learning (ML). The study compares six tiny ML techniques on two resource-constrained microcontrollers, reporting an f1-score of 87.6% for identifying normal and abnormal duty cycles and 96.8% for the internal states of the conveyor belt system. Deployed on both low-power microcontrollers, the algorithm processes input data in less than 106 μs, consuming less than 1.16 μJ. These findings promise to facilitate integration into more comprehensive preventive maintenance algorithms. 

For the field of structural health monitoring (SHM), acoustic emission (AE) technology is important as a damage identification technique that does not cause secondary damage to concrete. Nowadays, applications of non-destructive concrete damage identification are mostly limited to commercial software or identification algorithms running on desktop computers. It has so far not been deployed in low-power embedded devices. In this study, a lightweight convolutional neural network (CNN) model for online non-destructive damage type recognition of concrete materials is presented and deployed on a resource-constrained microcontroller unit as a tiny machine learning (TinyML) application. The CNN model uses raw acoustic emission signals as input and damage recognition types as output. 15,000 acoustic emission signals are used as data sets divided into training, validation, and test sets in the ratio of 8:1:1. The experimental results show that an accuracy of 99.6% is achieved on the nRF52840 microcontroller (ARM Cortex M4) with only 166.822 ms and 0.555mJ for a single inference using only 20K parameters and 30.5KB model size. This work demonstrates the effectiveness and feasibility of the proposed model, which achieves a trade-off between high classification accuracy and deployability on resource-constrained MCUs. Consequently, it provides strong support for online continuous non-destructive structural health monitoring. 

Precision livestock farming allows farmers to optimize herd management while significantly reducing labor needs. Individualized monitoring of cattle feeding behavior offers valuable data to assess animal performance and provides valuable insights into animal welfare. Current acoustic foraging activity recognizers achieve high recognition rates operating on computers. However, their implementations on portable embedded systems (for use on farms) need further investigation. This work presents two embedded deployments of a state-of-the-art foraging activity recognizer on a low-power ARM Cortex-M0+ microcontroller. The parameters of the algorithm were optimized to reduce power consumption. The embedded algorithm processes masticatory sounds in real-time and uses machine-learning techniques to identify grazing, rumination and other activities. The overall classification performance of the two embedded deployments achieves an 84% and 89% balanced accuracy with a mean power consumption of 1.8 mW and 12.7 mW, respectively. These results will allow this deployment to be integrated into a self-powered acoustic sensor with wireless communication to operate autonomously on cattle. 

The feasibility of energy harvesting as a viable alternative for powering low-energy electronics has been demonstrated through advancements in transduction mechanisms. Energy harvesters incorporating counterweights have gained attention in rotational energy harvesting to develop single-anchored devices with flexible placement and easy installation. In this work, a three-phase variable reluctance energy harvester (VREH) with low torque ripple is combined with a counterweight to facilitate a single-anchored design, specifically targeting low rotational speed applications. The energy harvester is integrated with a low-power sensor system to enable energy-neutral operation. We present the design, implementation, and evaluation of an on-rotor RPM sensor system powered by the single-anchored three-phase VREH. Experimental evaluations on a laboratory test bench demonstrate the system performance under varying conditions, with the ability to supply the sensor system at low speeds achieving, for example, a 3.5 Hz sample rate at a low speed of 3 rpm. Evaluations of the system illustrate that pendulum effects induced by the interaction of the cogging torque and the gravitational torque improve the output power of the harvester under low-speed conditions. This promises for the proposed design to be suitable to power wireless sensors for industrial condition monitoring, providing a flexible solution for energy-neutral sensor systems with reduced installation complexity.

Acoustic emission (AE) is a widely used non-destructive test method in structural health monitoring applications to identify the damage type in the material. Usually, the analysis of the AE signal is done by using traditional parameter-based methods. Recently, machine learning methods showed promising results for the analysis of AE signals. However, these machine learning models are complex, slow, and consume significant amounts of energy. To address these limitations and to explore the trade-off between model complexity and the classification accuracy, this paper presents a lightweight artificial neural network model to classify damage types in concrete material using raw acoustic emission signals. The model consists of one hidden layer with four neurons and is trained on a public acoustic emission signal dataset. The created model is deployed to several microcontrollers and the performance of the model is evaluated and compared with a state-of-the-art machine learning model. The model achieves 98.4% accuracy on the test data with only 4019 parameters. In terms of evaluation metrics, the proposed tiny machine learning model outperforms previously proposed models 10 to 1000 times. The proposed model thus enables machine learning in real-time structural health monitoring applications. 

Postural sway is a critical measure for evaluating postural control, and its analysis plays a vital role in preventing falls among the elderly. Typically, physiotherapists assess an individual's postural control using tests such as the Berg Balance Scale, Tinetti Test, and time up-and-go test. Sensor-based analysis is available based on devices such as force plates or inertial measurement units. Recently, machine learning methods have demonstrated promising results in the sensor-based analysis of postural control. However, these models are often complex, slow, and energy-intensive. To address these limitations, this study explores the design space of lightweight machine learning models deployable to microcontrollers to assess postural stability. We developed an artificial neural network (ANN) model and compare its performance to that of random forests, gaussian naive bayes, and extra tree classifiers. The models are trained using a sway dataset with varying input sizes and signal-to-noise ratios. The dataset comprises two feature vectors extracted from raw accelerometer data. The developed models are deployed to an ARM Cortex M4-based microcontroller, and their performance is evaluated and compared. We show that the ANN model has 99.03% accuracy, higher noise immunity, and the model performs better with a window size of one second with 590.96 us inference time.