The rapid advancement of edge computing and Tiny Machine Learning (TinyML) has created new opportunities for deploying intelligence in resource-constrained environments. With the growing demand for intelligent Internet of Things (IoT) devices that can efficiently process complex data in real-time, there is an urgent need for innovative optimisation techniques that overcome the limitations of IoT devices and enable accurate and efficient computations. This study investigates a novel approach to optimising Convolutional Neural Network (CNN) models for Hand Gesture Recognition (HGR) based on Electrical Impedance Tomography (EIT), which requires complex signal processing, energy efficiency, and real-time processing, by simultaneously reducing input complexity and using advanced model compression techniques. By systematically reducing and halving the input complexity of a 1D CNN from 40 to 20 Boundary Voltages (BVs) and applying an innovative compression method, we achieved remarkable model size reductions of 91.75% and 97.49% for 40 and 20 BVs EIT inputs, respectively. Additionally, the Floating-Point operations (FLOPs) are significantly reduced, by more than 99% in both cases. These reductions have been achieved with a minimal loss of accuracy, maintaining the performance of 97.22% and 94.44% for 40 and 20 BVs inputs, respectively. The most significant result is the 20 BVs compressed model. In fact, at only 8.73 kB and a remarkable 94.44% accuracy, our model demonstrates the potential of intelligent design strategies in creating ultra-lightweight, high-performance CNN-based solutions for resource-constrained devices with near-full performance capabilities specifically for the case of HGR based on EIT inputs.

State of Charge (SoC) estimation is critical for battery management systems, yet existing methods face practical limitations. Electrochemical impedance spectroscopy requires 60-1500 s measurements and specialized hardware, while datadriven approaches need excessively long windows ( 60 − 3000 s ) or train separate models per temperature. This paper presents a hybrid CNN-Transformer architecture achieving 0.86% RMSE using three seconds windows of voltage, current, and temperature signals sampled at 2 5 0 points. Systematic evaluation across window lengths ( 1 − 1 0 s ) and sampling resolutions ( 1 0 0 − 5 0 0 samples) reveals that three seconds windows optimally capture charge transfer dynamics without redundant low-frequency content. The architecture combines convolutional layers for local feature extraction with transformer encoders for long-range dependencies, with ablation studies confirming synergistic contributions. The unified model generalizes across 1 0 ∘ C , 2 5 ∘ C , and 4 0 ∘ C without temperature-specific training, achieving 35.8% improvement over Random Forest and 6 4. 8 % over LSTM methods while reducing measurement time by 95 % compared to recent transformer approaches. With 1.2 s inference time on GPU hardware, the method enables real-time deployment in battery management systems using standard sensors.