A TinyML Deep Learning Approach for Indoor Tracking of Assets

Empowering stroke recovery with upper limb rehabilitation monitoring using TinyML based heterogeneous classifiers

Stroke is one of the leading causes of disability worldwide, with approximately 70% of survivors experiencing motor impairments in the upper limbs, significantly affecting their quality of life. Home-based rehabilitation offers a cost-effective approach to improving motor function, but it faces challenges, including inaccurate movement reporting, lack of real-time feedback, and the high cost of rehabilitation equipment. Therefore, there is a need for affordable, lightweight home-based rehabilitation monitoring systems. This paper presents an intelligent wearable sensor system that utilizes TinyML AI technology to classify eight upper limb rehabilitation movements with minimal sensors. The system is designed for patients with upper limb impairments who retain antigravity voluntary movement, enabling them to monitor rehabilitation progress at home. The study recruited 10 healthy volunteers to perform rehabilitation movements, creating a standardized dataset for model training. Data normalization, preprocessing, model training, and deployment were carried out using the Edge Impulse platform. A hybrid classifier, combining multilayer perceptron and k-means clustering, achieved 96.1% training accuracy, 95.09% testing accuracy, and 88.01% deployment accuracy. The proposed TinyML-based system shows promising potential for home-based rehabilitation of stroke patients.

Optimising TinyML with quantization and distillation of transformer and mamba models for indoor localisation on edge devices

This paper proposes small and efficient machine learning models (TinyML) for resource-constrained edge devices, specifically for on-device indoor localisation. Typical approaches for indoor localisation rely on centralised remote processing of data transmitted from lower powered devices such as wearables. However, there are several benefits for moving this to the edge device itself, including increased battery life, enhanced privacy, reduced latency and lowered operational costs, all of which are key for common applications such as health monitoring. The work focuses on model compression techniques, including quantization and knowledge distillation, to significantly reduce the model size while maintaining high predictive performance. We base our work on a large state-of-the-art transformer-based model and seek to deploy it within low-power MCUs. We also propose a state-space-based architecture using Mamba as a more compact alternative to the transformer. Our results show that the quantized transformer model performs well within a 64 KB RAM constraint, achieving an effective balance between model size and localisation precision. Additionally, the compact Mamba model has strong performance under even tighter constraints, such as a 32 KB of RAM, without the need for model compression, making it a viable option for more resource-limited environments. We demonstrate that, through our framework, it is feasible to deploy advanced indoor localisation models onto low-power MCUs with restricted memory limitations. The application of these TinyML models in healthcare has the potential to revolutionize patient monitoring by providing accurate, real-time location data while minimising power consumption, increasing data privacy, improving latency and reducing infrastructure costs.

TinyML-Based In-Pipe Feature Detection for Miniature Robots

Miniature robots in small-diameter pipelines require efficient and reliable environmental perception for autonomous navigation. In this paper, a tiny machine learning (TinyML)-based resource-efficient pipe feature recognition method is proposed for miniature robots to identify key pipeline features such as elbows, joints, and turns. The method leverages a custom five-layer convolutional neural network (CNN) optimized for deployment on a robot with limited computational and memory resources. Trained on a custom dataset of 4629 images collected under diverse conditions, the model achieved an accuracy of 97.1%. With a peak RAM usage of 195.1 kB, flash usage of 427.9 kB, and an inference time of 1693 ms, the method demonstrates high computational efficiency while ensuring stable performance under challenging conditions through a sliding window smoothing strategy. These results highlight the feasibility of deploying advanced machine learning models on resource-constrained devices, providing a cost-effective solution for autonomous in-pipe exploration and inspection.

A Selection of Starting Points for Iterative Position Estimation Algorithms Using Feedforward Neural Networks

This article proposes the use of a feedforward neural network (FNN) to select the starting point for the first iteration in well-known iterative location estimation algorithms, with the research objective of finding the minimum size of a neural network that allows iterative position estimation algorithms to converge in an example positioning network. The selected algorithms for iterative position estimation, the structure of the neural network and how the FNN is used in 2D and 3D position estimation process are presented. The most important results of the work are the parameters of various FNN network structures that resulted in a 100% probability of convergence of iterative position estimation algorithms in the exemplary TDoA positioning network, as well as the average and maximum number of iterations, which can give a general idea about the effectiveness of using neural networks to support the position estimation process. In all simulated scenarios, simple networks with a single hidden layer containing a dozen non-linear neurons turned out to be sufficient to solve the convergence problem.

Human-Robot Interaction Using Learning from Demonstrations and a Wearable Glove with Multiple Sensors

Human-robot interaction is of the utmost importance as it enables seamless collaboration and communication between humans and robots, leading to enhanced productivity and efficiency. It involves gathering data from humans, transmitting the data to a robot for execution, and providing feedback to the human. To perform complex tasks, such as robotic grasping and manipulation, which require both human intelligence and robotic capabilities, effective interaction modes are required. To address this issue, we use a wearable glove to collect relevant data from a human demonstrator for improved human-robot interaction. Accelerometer, pressure, and flexi sensors were embedded in the wearable glove to measure motion and force information for handling objects of different sizes, materials, and conditions. A machine learning algorithm is proposed to recognize grasp orientation and position, based on the multi-sensor fusion method.

Device-Free Tracking through Self-Attention Mechanism and Unscented Kalman Filter with Commodity Wi-Fi

Recent advancements in target tracking using Wi-Fi signals and channel state information (CSI) have significantly improved the accuracy and efficiency of tracking mobile targets. However, there remains a gap in developing a comprehensive approach that combines CSI, an unscented Kalman filter (UKF), and a sole self-attention mechanism to accurately estimate the position, velocity, and acceleration of targets in real-time. Furthermore, optimizing the computational efficiency of such approaches is necessary for their applicability in resource-constrained environments. To bridge this gap, this research study proposes a novel approach that addresses these challenges. The approach leverages CSI data collected from commodity Wi-Fi devices and incorporates a combination of the UKF and a sole self-attention mechanism. By fusing these elements, the proposed model provides instantaneous and precise estimates of the target's position while considering factors such as acceleration and network information. The effectiveness of the proposed approach is demonstrated through extensive experiments conducted in a controlled test bed environment. The results exhibit a remarkable tracking accuracy level of 97%, affirming the model's ability to successfully track mobile targets. The achieved accuracy showcases the potential of the proposed approach for applications in human-computer interactions, surveillance, and security.