ABSTRACT Reliable indoor positioning is crucial for location‐based services, especially in complex buildings with multiple floors. Although Wi‐Fi fingerprinting is cost‐effective, it still has major limitations, including signal instability and inter‐floor errors. Clustering‐based methods are the mainstream solution to solve this problem. However, traditional distance metrics fail to capture the complexity of indoor environments and have limitations in identifying non‐linear relationships. To address these issues, we present a feature‐enhanced clustering framework built on the Wi‐Fi signal embedding neural network (WiSE‐Net) and a hierarchical‐polarised weighted K‐nearest neighbour (HP‐WKNN) algorithm. Extensive experiments on three public benchmark datasets (UJIIndoorLoc, Tampere and UTSIndoorLoc) indicate that our method achieves the lowest mean error of 6.18 m and maintains the highest floor detection rate (FDR) of 94.81%. This work provides a reliable solution for large‐scale multi‐building, multi‐floor indoor localisation, improving both positioning stability and accuracy across diverse environments.
ABSTRACT This paper proposes a Zener‐based segmented curvature‐compensated voltage reference. The reference voltage is derived from the PTAT Zener voltage through resistive division, and temperature compensation is achieved by modulating the current through the resistive networks. First‐order temperature compensation is realized by extracting a PTAT current of the trimming resistor. Segmented curvature compensation is then implemented through a digital loop in foreground mode, in which a current‐steering digital‐to‐analogue converter generates the compensation current based on voltage‐comparison‐based temperature segmentation. The circuit is designed in a 0.18‐ BCD process. Post‐simulation results validate the proposed = 3.3 V design at = 5 V, achieving a TC of 0.681–1.271 ppm/ over a temperature range of −40 to 125 .
ABSTRACT This paper presents a digital twin of a digitally controlled direct‐current (DC) servo motor. The digital twin is equipped with a position encoder and a current sensor, enabling real‐time synchronisation with the physical system through continuous updates of the shaft angular position and armature current. The proposed twin employs a physics‐based model that accurately captures the motor's electrical and mechanical dynamics. Implemented on an embedded platform, it provides a lightweight, computationally efficient solution suitable for real‐time applications. The model is fully parameterised, allowing it to emulate the behaviour of different DC motors, provided that their physical characteristics are known. This flexibility makes the proposed approach well suited to embedded industrial systems, rapid prototyping, and system‐level testing.
ABSTRACT Massive MIMO in integrated sensing and communication introduces a dimensionality curse that limits deep‐learning based joint channel estimation and localization, as convolutional architectures fail to capture long‐range inter‐antenna dependencies. This letter compares three architectural generations to address this challenge. Generation I extends a SISO CrossNet to MIMO, Generation II introduces task‐decoupled residual branches and Generation III employs angular‐delay (AD) domain tokenization with a dual‐branch transformer. Compared with Generation I, the proposed architecture reduces P90 error by 88.81% while requiring only 0.236 GFLOPS, achieving a 17.4 higher localization efficiency per GFLOP. These results demonstrate that AD sequence modelling substantially improves localization accuracy while reducing arithmetic inference cost relative to the convolutional baselines.
ABSTRACT Electroencephalography (EEG)‐based seizure detection is increasingly being considered for wearable neurological monitoring, but conventional approaches often rely on dense electrode configurations and high‐dimensional feature representations. This letter presents a machine learning‐guided co‐design framework in which feature relevance informs EEG acquisition design. From 18 bipolar derivations of the CHB‐MIT database, 270 wavelet descriptors were extracted. The available reduced inputs comprise 10 features from nine derivations and a separate four‐feature configuration comprising four bipolar derivations and eight scalp electrodes. In 75/25 sample‐level evaluations, random forest achieved 93.75%, 91.67% and 93.75% accuracy for the full, 10‐feature, and four‐feature inputs, respectively. These exploratory results demonstrate a traceable link between feature relevance and acquisition‐design decisions, supporting the proposed feature‐to‐derivation mapping as a candidate strategy for compact EEG acquisition while motivating subsequent patient‐independent and wearable validation.
