Georgios-Fotios Angelis
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Georgios-Fotios Angelis has a total of 29 co-authors. Here are the top co-authors by number of shared publications:
- Dimitrios Tzovaras
Centre for Research and Technology Hellas (CERTH)
5 shared publications Anastasios Drosou
Information Technologies Institute, Centre for Research and Technology Hellas, Thessaloniki, Greece
2 shared publications- Anastasios Drosou
Information Technologies Institute Center for Research and Technology Hellas Thessaloniki Greece
2 shared publications - Christos Timplalexis
Centre for Research and Technology Hellas, Information Technologies Institute, Thessaloniki, Greece
2 shared publications - Dimosthenis Ioannidis
Information Technologies Institute, Centre for Research and Technology Hellas, Thessaloniki, Greece
2 shared publications Alexandros Emvoliadis
Information Technologies Institute, Centre for Research and Technology Hellas, Thessaloniki, Greece
1 shared publication
Latest Publications
- 26 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (2) , 3315-3326
HMTC: A Hybrid Mamba-Transformer Model for Calibration of Low-Cost Air Quality Sensors
The assessment of Low-Cost Sensors (LCS) for air quality multisensor systems performance is a critical aspect of the Air Quality (AQ) monitoring framework. The calibration model for these devices is one of the key drivers of overall performance. Field calibration is increasingly recognized as the best approach for calibrating air quality monitoring devices. However, to achieve high-quality calibration results, low-cost sensors must be co-located with reference stations. To address this challenge, we propose a comprehensive solution that leverages self-supervised pretraining on LCS data to learn meaningful representations for pollution data calibration. In the first stage, our model undergoes unsupervised pretraining inspired by masked autoencoders. In the second stage, supervised fine-tuning is applied to enhance the accuracy of pollution data calibration. Additionally, we introduce a novel architecture that combines state-space models, specifically Mamba, with Transformer layers. Our model is rigorously evaluated on two publicly available research datasets, consistently outperforming all baseline methods. Furthermore, we present a newly designed ESP32 microcontroller based system to validate the effectiveness of our approach in real-world scenarios. The results demonstrate that our calibration technique is well-suited for microcontrollers with limited resources, significantly reducing the cost of IoT solutions for air quality and pollution monitoring.
- 26 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 26 (2) , 3315-3326
HMTC: A Hybrid Mamba-Transformer Model for Calibration of Low-Cost Air Quality Sensors
The assessment of Low-Cost Sensors (LCS) for air quality multisensor systems performance is a critical aspect of the Air Quality (AQ) monitoring framework. The calibration model for these devices is one of the key drivers of overall performance. Field calibration is increasingly recognized as the best approach for calibrating air quality monitoring devices. However, to achieve high-quality calibration results, low-cost sensors must be co-located with reference stations. To address this challenge, we propose a comprehensive solution that leverages self-supervised pretraining on LCS data to learn meaningful representations for pollution data calibration. In the first stage, our model undergoes unsupervised pretraining inspired by masked autoencoders. In the second stage, supervised fine-tuning is applied to enhance the accuracy of pollution data calibration. Additionally, we introduce a novel architecture that combines state-space models, specifically Mamba, with Transformer layers. Our model is rigorously evaluated on two publicly available research datasets, consistently outperforming all baseline methods. Furthermore, we present a newly designed ESP32 microcontroller based system to validate the effectiveness of our approach in real-world scenarios. The results demonstrate that our calibration technique is well-suited for microcontrollers with limited resources, significantly reducing the cost of IoT solutions for air quality and pollution monitoring.
- Open Access
Mem-MLP: Real-Time 3D Human Motion Generation from Sparse Inputs
Realistic and smooth full-body tracking is crucial for immersive AR/VR applications. Existing systems primarily track head and hands via Head Mounted Devices (HMDs) and controllers, making the 3D full-body reconstruction in-complete. One potential approach is to generate the full-body motions from sparse inputs collected from limited sensors using a Neural Network (NN) model. In this paper, we propose a novel method based on a multi-layer perceptron (MLP) backbone that is enhanced with residual connections and a novel NN-component called Memory-Block. In particular, Memory-Block represents missing sensor data with trainable code-vectors, which are combined with the sparse signals from previous time instances to improve the temporal consistency. Furthermore, we formulate our solution as a multi-task learning problem, allowing our MLP-backbone to learn robust representations that boost accuracy. Our experiments show that our method outperforms state-of-the-art baselines by substantially reducing prediction errors. Moreover, it achieves 72 FPS on mobile HMDs that ultimately improves the accuracy-running time tradeoff.
Open AccessMem-MLP: Real-Time 3D Human Motion Generation from Sparse Inputs
Realistic and smooth full-body tracking is crucial for immersive AR/VR applications. Existing systems primarily track head and hands via Head Mounted Devices (HMDs) and controllers, making the 3D full-body reconstruction in-complete. One potential approach is to generate the full-body motions from sparse inputs collected from limited sensors using a Neural Network (NN) model. In this paper, we propose a novel method based on a multi-layer perceptron (MLP) backbone that is enhanced with residual connections and a novel NN-component called Memory-Block. In particular, Memory-Block represents missing sensor data with trainable code-vectors, which are combined with the sparse signals from previous time instances to improve the temporal consistency. Furthermore, we formulate our solution as a multi-task learning problem, allowing our MLP-backbone to learn robust representations that boost accuracy. Our experiments show that our method outperforms state-of-the-art baselines by substantially reducing prediction errors. Moreover, it achieves 72 FPS on mobile HMDs that ultimately improves the accuracy-running time tradeoff.
- Open Access
- 11 January 2024
- Published by Springer Nature in Universal Access in the Information Society
- Vol. 24 (3) , 2025-2041
Engagement and accessibility tools for pro-environmental action on air quality: the SOCIO-BEE paradigm
The involvement of citizens and all stakeholders is crucial in tackling environmental and social matters. This, addressing equity and diversity issues, although challenging, is a necessary condition for achieving positive outcomes and ensuring that no one is left behind. To help ease this challenge, this work presents a systematic approach to ensure inclusive participation and leverage non-technical and technical elements to maximise stakeholder engagement in scientific activities to successfully address sustainability concerns. For that, it builds on the interim results of the H2020 SOCIO-BEE project, a Citizen science (CS) proposal to reduce air pollution through inclusive community engagement and social innovation. As part of an interdisciplinary CS project, an abductive systematic combining methodology was employed, which allowed for dialogue and collaboration between theory and practice throughout the whole process, during which separate groups of experts and potential end-users were involved. The article presents (i) the stakeholder engagement strategy codified in the SOCIO-BEE toolkit as a robust, actionable and inclusive foundation of engagement to CS activities; and (ii) the digital platform UX that allows setting up campaigns for measurements and assignment to citizens, incorporating the requirements for flexibility, accessibility, limited digital literacy, inclusion and legal and ethical considerations. Their combination and mutual interaction aim to leverage the pros of CS and technology whilst reducing their cons to ensure the four pillars of applicability, scalability, actionability, and inclusion. This is supported by the presented hybrid model which combines physical and virtual spaces and individual and collective action.
- 11 January 2024
- Published by Springer Nature in Universal Access in the Information Society
- Vol. 24 (3) , 2025-2041
Open AccessEngagement and accessibility tools for pro-environmental action on air quality: the SOCIO-BEE paradigm
The involvement of citizens and all stakeholders is crucial in tackling environmental and social matters. This, addressing equity and diversity issues, although challenging, is a necessary condition for achieving positive outcomes and ensuring that no one is left behind. To help ease this challenge, this work presents a systematic approach to ensure inclusive participation and leverage non-technical and technical elements to maximise stakeholder engagement in scientific activities to successfully address sustainability concerns. For that, it builds on the interim results of the H2020 SOCIO-BEE project, a Citizen science (CS) proposal to reduce air pollution through inclusive community engagement and social innovation. As part of an interdisciplinary CS project, an abductive systematic combining methodology was employed, which allowed for dialogue and collaboration between theory and practice throughout the whole process, during which separate groups of experts and potential end-users were involved. The article presents (i) the stakeholder engagement strategy codified in the SOCIO-BEE toolkit as a robust, actionable and inclusive foundation of engagement to CS activities; and (ii) the digital platform UX that allows setting up campaigns for measurements and assignment to citizens, incorporating the requirements for flexibility, accessibility, limited digital literacy, inclusion and legal and ethical considerations. Their combination and mutual interaction aim to leverage the pros of CS and technology whilst reducing their cons to ensure the four pillars of applicability, scalability, actionability, and inclusion. This is supported by the presented hybrid model which combines physical and virtual spaces and individual and collective action.
- 13 July 2023
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 83 (6) , 18309-18332
AI-enabled Underground Water Pipe non -destructive Inspection
- 13 July 2023
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 83 (6) , 18309-18332
AI-enabled Underground Water Pipe non -destructive Inspection
- 18 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 69 (3) , 308-320
Energformer: A New Transformer Model for Energy Disaggregation
In recent years, a lot of progress has been reported in the field of energy disaggregation, also referred to as Non-Intrusive Load Monitoring (NILM). Despite the fact that there are many studies focusing on the residential sector, there is considerably less research interest for the industrial sector. In this paper, we present a deep neural network based on Transformers, targeted towards capturing complex patterns in long sequences of data. The proposed transformer architecture employs 1D spatial convolutions in self-attention, and modifications inside the attention computations manage to reduce computational complexity without any loss in predictive accuracy. In order to evaluate the performance of the proposed deep learning architecture, a set of experiments has been conducted using a publicly available dataset. The experimental results indicate that the proposed model achieves better disaggregation accuracy compared to other state-of-the-art NILM models.
- 18 January 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 69 (3) , 308-320
Energformer: A New Transformer Model for Energy Disaggregation
In recent years, a lot of progress has been reported in the field of energy disaggregation, also referred to as Non-Intrusive Load Monitoring (NILM). Despite the fact that there are many studies focusing on the residential sector, there is considerably less research interest for the industrial sector. In this paper, we present a deep neural network based on Transformers, targeted towards capturing complex patterns in long sequences of data. The proposed transformer architecture employs 1D spatial convolutions in self-attention, and modifications inside the attention computations manage to reduce computational complexity without any loss in predictive accuracy. In order to evaluate the performance of the proposed deep learning architecture, a set of experiments has been conducted using a publicly available dataset. The experimental results indicate that the proposed model achieves better disaggregation accuracy compared to other state-of-the-art NILM models.
- 12 March 2022
- Published by Taylor & Francis in Energy Sources, Part A: Recovery, Utilization, and Environmental Effects
- Vol. 44 (1) , 493-514
Low frequency residential non-intrusive load monitoring based on a hybrid feature extraction tree-learning approach
Non-intrusive load monitoring (NILM), or else energy disaggregation, aims at decomposing the aggregated energy consumption data, estimating the contribution of each appliance to the total building energy consumption. The current paper introduces a supervised algorithm for the accurate estimation of both power consumption and operational state of the appliances. A hybrid feature set consisting of electrical, temporal, and statistical features is utilized, focusing on steady state information, due to the low sampling rate (1 minute). The contribution of each feature category is experimentally validated. A predictive model is built for each appliance, utilizing Light Gradient Boosting Machine (LGBM) regression algorithm, which is a tree-based technique with accelerated training time and low computational cost. The results obtained, exploiting both private and public datasets, indicate that the proposed algorithm is able to disaggregate most of the appliances with high accuracy. The average f-score across all the disaggregated appliances is calculated at 0.77 and 0.86 for the private and the public dataset, respectively. The proposed models have also been benchmarked against other top state-of-the-art approaches, achieving high performance with high rank.
- 12 March 2022
- Published by Taylor & Francis in Energy Sources, Part A: Recovery, Utilization, and Environmental Effects
- Vol. 44 (1) , 493-514
Low frequency residential non-intrusive load monitoring based on a hybrid feature extraction tree-learning approach
Non-intrusive load monitoring (NILM), or else energy disaggregation, aims at decomposing the aggregated energy consumption data, estimating the contribution of each appliance to the total building energy consumption. The current paper introduces a supervised algorithm for the accurate estimation of both power consumption and operational state of the appliances. A hybrid feature set consisting of electrical, temporal, and statistical features is utilized, focusing on steady state information, due to the low sampling rate (1 minute). The contribution of each feature category is experimentally validated. A predictive model is built for each appliance, utilizing Light Gradient Boosting Machine (LGBM) regression algorithm, which is a tree-based technique with accelerated training time and low computational cost. The results obtained, exploiting both private and public datasets, indicate that the proposed algorithm is able to disaggregate most of the appliances with high accuracy. The average f-score across all the disaggregated appliances is calculated at 0.77 and 0.86 for the private and the public dataset, respectively. The proposed models have also been benchmarked against other top state-of-the-art approaches, achieving high performance with high rank.