Jul 2026· Italian National Conference on Sensors· Vol 26, pp. 4482· 0 citations· 38 references
Medicine
TL;DR
This work investigates the Convolutional Tsetlin Machine for multimodal HAR using only the raw inertial signals of the UCI-HAR dataset, rather than its pre-computed 561-feature representation, to address predictive performance, interpretability and suitability for embedded and mobile platforms.
Abstract
Human activity recognition (HAR) based on smartphone and wearable sensor data is commonly addressed using statistical learning methods and deep neural networks that often provide strong predictive performance, but at the expense of limited interpretability and substantial computational and energy requirements. Such limitations reduce their suitability for deployment in practical sensing environments where model decisions must be transparent, verifiable and executable on resource-constrained devices. In this work, we investigate the Convolutional Tsetlin Machine (CTM) for multimodal HAR using only the raw inertial signals (9 × 128) of the UCI-HAR dataset, rather than its pre-computed 561-feature representation. The Tsetlin Machine is a novel neuro-symbolic machine learning approach that offers two important advantages over many conventional machine learning methods: (i) it learns logic-based decision rules that support human inspection and provide a transparent basis for analyzing model decisions, and (ii) it operates with comparatively low computational complexity, making it well suited to efficient and low-power on-device learning. The proposed study systematically analyses the contribution of different feature modalities by decomposing the inertial signals space into semantically defined subsets according to: (i) sensor source: accelerometer and gyroscope; (ii) signal group: gyroscope angular velocity, body and total acceleration (including gravity); (iii) coordinate axis: x, y and z. A separate CTM classifier was trained for each modality and its combinations in order to determine the relative discriminative value of each modality group for activity classification. In addition to predictive performance, the study emphasizes the interpretability of the CTM model ensured by expressing each decision in the form of propositional clauses, thereby enabling visualization and direct inspection of the modality-specific patterns supporting each activity class. Owing to its symbolic structure and modest computational demands, the CTM provides a principled framework for the design of explainable, resource-efficient and deployable HAR systems. The proposed work therefore contributes toward trustworthy multimodal sensing by jointly addressing predictive performance, interpretability and suitability for embedded and mobile platforms.
Human Activity Recognition (HAR) focuses on measuring human activities from inertial sensor signals acquired through wearable devices. While machine learning and deep learning approaches achieve high classification performance, they often suffer from limited interpretability, high computational cost, and reduced generalization across users. From a measurement perspective, there is therefore increasing interest in identifying simple, physically meaningful descriptors that can be reliably derived from sensor data. In this work, we investigate jerk magnitude, defined as the time derivative of acceleration, as a standalone kinematic measurand for activity discrimination. Triaxial acceleration data were acquired using an Empatica EmbracePlus smartwatch from 20 subjects performing five Activities of Daily Living (ADLs) with different intensity levels. The jerk magnitude was computed through a signal processing pipeline and summarized over short temporal windows using basic statistical features, with particular focus on the mean jerk. To evaluate the discriminative capability of this descriptor, a fully non-parametric statistical framework was adopted, combining the Kruskal–Wallis test for global analysis and Dunn's post-hoc test with Bonferroni correction for pairwise comparisons. The results show that mean jerk magnitude exhibits statistically significant differences across all activity classes $(p<0.001$), enabling clear discrimination between static, low-intensity, and highly dynamic movements. Pairwise analysis confirms strong separability for most activity combinations, while highlighting limitations in distinguishing tasks with similar motion smoothness but different spatial orientation. These findings demonstrate that a single, computationally lightweight and physically interpretable measurand can provide robust activity discrimination without relying on complex models. The proposed approach establishes a reproducible baseline for HAR and highlights the potential of measurement-driven, physics-based descriptors for low-power wearable applications.
G. Sartini, Sara Meletani, S. Casaccia et al.· 2026 IEEE International Work...· 0 citations
This study reviews pre-2018 developments and proposes an improved HAR framework incorporating advanced feature engineering, sensor fusion, and ensemble learning techniques, which shows strong potential in healthcare, fitness, and smart environments.
Silvia Diallo, F. Z. Idrissi· International Journal of Mod...· 0 citations
Human Activity Recognition (HAR) has become one of the essential research areas due to the blistering development of wearable sensory devices, smartphones or the Internet of Things (IoT). HAR aims at recognizing human physical actions like walking, sitting, standing, running and lying down automatically based on the data acquired by motion sensors and physiological sensors. The ability to scale and flexibility have gradually seen the replacement of traditional rule-based systems by data-driven systems. Machine learning models and deep learning provide data science approaches that can robustly extract features, classify, and infer in real time using complicated sensor signal data streams. In the current paper, the use of data science methods in HAR is evaluated and summed up in detail. It examines data collection techniques, preprocessing techniques, feature engineering techniques and classification models. Moreover, it reviews benchmark datasets and assessment measures that are prevalent in HAR studies. The HAR methodology based on data science pipelines is offered and tested on the example of standard datasets. Findings have shown that novel machine learning and deep learning neural networks are much more effective in recognition accuracy than the classical methods of statistics. The main issues that have been identified by the study include sensor noise, user variability and computational constraints and future prospects of the study is given on context-aware systems and edge intelligence. The results lead to the realization of efficient HAR to track healthcare, intelligent environments, and human-computer interface.
Zainab Abdullahi· International Journal of App...· 0 citations
A novel lightweight cross-domain few-shot sensor-based HAR network (CFSH-Net) is proposed for cross-domain activity recognition with limited labeled samples, which demonstrates strong cross-user generalization on PAMAP2 and USC-HAD, and stable cross-dataset transfer when trained on OPPORTUNITY and evaluated on four other datasets.
Hao Zheng, Hongji Xu, Fei Gao et al.· IEEE journal of biomedical a...· 0 citations
Real-time multi-sensor human activity recognition (HAR) requires accurate models and a system architecture capable of distributing computation, exchanging compact outputs, and maintaining temporal consistency across asynchronous streams. This paper presents a distributed HAR framework in which five wearable sensors are associated with local embedded nodes that perform acquisition, windowing, preprocessing, and convolutional neural network–long short-term memory (CNN–LSTM) inference. Each node transmits a timestamped six-class softmax vector, and a central node applies approximate synchronization and learned probability-level fusion. The framework was evaluated with ten participants whose data were not used for model development. It achieved 95.868% accuracy and a 95.642% macro-F1-score. During continuous operation, the system sustained 47.949 predictions/s, with a mean post-window end-to-end latency of 33.963 ms and a mean synchronization span of 13.788 ms. Relative to complete-window transmission, the numerical payload decreased by 97.69%, and central-node energy per prediction decreased by 52.2% compared with centralized real-time processing. Under 30% independent probability-message loss, accuracy remained at 94.31%.
Jhonathan L. Rivas-Caicedo, Laura Saldaña-Aristizábal, Kevin Niño-Tejada et al.· Electronics· 0 citations
Automatic Human activity Recognition has many applications in smart environments such as in smart homes, smart cities, smart industries, smart healthcare centers, etc. While performing the activities by the participants, ambient or body-worn sensors can measure physical movements and those data can be used to develop machine learning models for recognizing those activities. In this study, we have proposed a deep convolutional neural network (DCNN) based method for recognizing human activities using body-worn sensors’ time-series data after an enormous data analysis on the data. The quality data is produced and balanced using a preprocessing chain for human activity recognition based on data analysis. The Preprocessed data is segmented using a constant-size sliding window. We developed several different DCNN models using random searches and based on validation accuracy we selected the best one for further training and testing. The outcomes of the selected model are carried out as the final predicted activities. We assessed our method on three popular and standard datasets: PAMAP2, WISDM_ar_v1.1, and UCI-HAR, and achieved 98.11%, 98.48%, and 93.25% accuracies for subject-dependent case and 90.27%, 94.51%, and 98.67% accuracies for subject-independent case. The performances of the experimental results are measured using several evaluation metrics and measures that institute the strength of the proposed model over the state-of-the-art.
S. Islam, Kamrul Hasan Talukder· International Journal of Int...· 0 citations