Pub Date : 2026-01-07DOI: 10.1016/j.compeleceng.2026.110957
Saoueb Kerdoudi , Larbi Guezouli , Tahar Dilekh
Detecting arrhythmias via electrocardiograms (ECGs) is vital for healthcare. While deep learning has advanced classification, capturing critical patterns in complex data remains challenging. We propose Res_Bi-LSTM_MHA, a novel model integrating a multi-head self-attention (MHA) mechanism to selectively focus on relevant signal segments. This enhances the capture of subtle features often missed by conventional methods. By combining Residual Networks (ResNet) for robust feature extraction with Bidirectional Long Short-Term Memory (Bi-LSTM) for temporal dependencies, our approach significantly improves accuracy. We evaluated the model at subject and record levels using the China Physiological Signal Challenge (CPSC 2018), St. Petersburg Institute of Cardiological Technics (INCART), and Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) databases. The model achieved an F1 score of 98.01% and 99.42% accuracy on the MIT-BIH dataset. Our results demonstrate that effectively utilizing attention mechanisms offers a substantial improvement in arrhythmia classification.
{"title":"Enhanced ECG arrhythmia detection with deep learning and multi-head attention mechanism","authors":"Saoueb Kerdoudi , Larbi Guezouli , Tahar Dilekh","doi":"10.1016/j.compeleceng.2026.110957","DOIUrl":"10.1016/j.compeleceng.2026.110957","url":null,"abstract":"<div><div>Detecting arrhythmias via electrocardiograms (ECGs) is vital for healthcare. While deep learning has advanced classification, capturing critical patterns in complex data remains challenging. We propose Res_Bi-LSTM_MHA, a novel model integrating a multi-head self-attention (MHA) mechanism to selectively focus on relevant signal segments. This enhances the capture of subtle features often missed by conventional methods. By combining Residual Networks (ResNet) for robust feature extraction with Bidirectional Long Short-Term Memory (Bi-LSTM) for temporal dependencies, our approach significantly improves accuracy. We evaluated the model at subject and record levels using the China Physiological Signal Challenge (CPSC 2018), St. Petersburg Institute of Cardiological Technics (INCART), and Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) databases. The model achieved an F1 score of 98.01% and 99.42% accuracy on the MIT-BIH dataset. Our results demonstrate that effectively utilizing attention mechanisms offers a substantial improvement in arrhythmia classification.</div></div>","PeriodicalId":50630,"journal":{"name":"Computers & Electrical Engineering","volume":"131 ","pages":"Article 110957"},"PeriodicalIF":4.9,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145927375","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-06DOI: 10.1016/j.compeleceng.2025.110928
Katroth Kalyan Singh, Kirubakaran Annamalai
This article proposes a single-phase transformerless inverter for grid-tied PV installations. At the output stage, the proposed inverter can produce five levels of voltage. It features two electrolytic switching capacitors (SCs), six power switches, and two power diodes. This architecture is lighter and less expensive due to the usage of fewer power electronic components. Because the negative DC line of the suggested inverter is directly connected to the grid neutral in PV applications, leakage current is completely minimized. Another advantage of this design is that it may easily double the output voltage without the need for a transformer or inductor. Self-balancing is achieved by symmetrically charging and discharging the SCs in parallel and in series with the input voltage over time. Therefore, a complex control technique to balance the SCs is no longer necessary with the proposed inverter. The design specifications of the proposed inverter are provided. To illustrate the benefits of the proposed inverter, including the reduction of total standing voltage and cost function, a quantitative comparison analysis with similar five-level topologies is also presented. An experimental prototype of a 1 kW grid-tied system is used to validate the topology and demonstrate the capabilities of the proposed inverter with a closed-loop PR controller. Moreover, the system dynamics are tested under different loading conditions and input voltage variations.
{"title":"Single-phase switched-capacitor based common ground five-level inverter for grid-tied PV systems with double gain","authors":"Katroth Kalyan Singh, Kirubakaran Annamalai","doi":"10.1016/j.compeleceng.2025.110928","DOIUrl":"10.1016/j.compeleceng.2025.110928","url":null,"abstract":"<div><div>This article proposes a single-phase transformerless inverter for grid-tied PV installations. At the output stage, the proposed inverter can produce five levels of voltage. It features two electrolytic switching capacitors (SCs), six power switches, and two power diodes. This architecture is lighter and less expensive due to the usage of fewer power electronic components. Because the negative DC line of the suggested inverter is directly connected to the grid neutral in PV applications, leakage current is completely minimized. Another advantage of this design is that it may easily double the output voltage without the need for a transformer or inductor. Self-balancing is achieved by symmetrically charging and discharging the SCs in parallel and in series with the input voltage over time. Therefore, a complex control technique to balance the SCs is no longer necessary with the proposed inverter. The design specifications of the proposed inverter are provided. To illustrate the benefits of the proposed inverter, including the reduction of total standing voltage and cost function, a quantitative comparison analysis with similar five-level topologies is also presented. An experimental prototype of a 1 kW grid-tied system is used to validate the topology and demonstrate the capabilities of the proposed inverter with a closed-loop PR controller. Moreover, the system dynamics are tested under different loading conditions and input voltage variations.</div></div>","PeriodicalId":50630,"journal":{"name":"Computers & Electrical Engineering","volume":"131 ","pages":"Article 110928"},"PeriodicalIF":4.9,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145927369","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Self-supervised learning (SSL) models are increasingly used in speech processing tasks, where they provide powerful pretrained representations of speech. Most existing methods utilize these models by either fine-tuning them on domain-specific data or using their output representations as input features in conventional ASR systems. However, the relationship between SSL layer representations and the severity level of dysarthric speech remains poorly understood, despite the potential for different layers to capture features that vary in relevance across severity levels. Furthermore, the high dimensionality of these representations, often reaching up to 1024 dimensions, imposes a heavy computational load, highlighting the need for optimized feature representations in downstream ASR and keyword spotting (KWS) tasks. This study proposes a severity-independent approach for dysarthric speech processing using SSL features, investigating three state-of-the-art pretrained models: Wav2Vec2, HuBERT, and Data2Vec. We propose: (1) selecting SSL layers based on severity level to extract the most useful features; (2) a Kaldi-based ASR system, that uses an autoencoder to reduce the size of SSL features; and (3) validating the proposed SSL feature optimization in a KWS task. We evaluate the proposed method using a DNN–HMM model in Kaldi on two standard dysarthric speech datasets: TORGO and UAspeech. Our approach shows that selecting severity-specific SSL layers, combined with autoencoder (AE)-based feature optimization, leads to significant improvements over both zero-shot and fine-tuned SSL baselines. On TORGO, our method achieved a WER of 23.12%, outperforming zero-shot (60.35%) and fine-tuned SSL model (40.48%). On UAspeech, it reached 50.33% WER, surpassing both the fine-tuned (51.04%) and MFCC-based systems (58.67%). Layer-wise analysis revealed consistent trends: lower layers were more effective for very high-severity speech, while mid-to-upper layers performed better for low/medium-severity cases. Further, in the KWS task, later SSL layers showed the best performance, with our proposed system outperforming the MFCC baseline. These findings highlight the generalization of our proposed method, which combines layer-specific selection and autoencoder-based optimization of SSL features, for dysarthric speech processing tasks.
{"title":"Role of SSL models: Finetuning and feature optimization for dysarthric speech recognition and keyword spotting","authors":"Paban Sapkota, Hemant Kumar Kathania, Subham Kutum","doi":"10.1016/j.compeleceng.2025.110921","DOIUrl":"10.1016/j.compeleceng.2025.110921","url":null,"abstract":"<div><div>Self-supervised learning (SSL) models are increasingly used in speech processing tasks, where they provide powerful pretrained representations of speech. Most existing methods utilize these models by either fine-tuning them on domain-specific data or using their output representations as input features in conventional ASR systems. However, the relationship between SSL layer representations and the severity level of dysarthric speech remains poorly understood, despite the potential for different layers to capture features that vary in relevance across severity levels. Furthermore, the high dimensionality of these representations, often reaching up to 1024 dimensions, imposes a heavy computational load, highlighting the need for optimized feature representations in downstream ASR and keyword spotting (KWS) tasks. This study proposes a severity-independent approach for dysarthric speech processing using SSL features, investigating three state-of-the-art pretrained models: Wav2Vec2, HuBERT, and Data2Vec. We propose: (1) selecting SSL layers based on severity level to extract the most useful features; (2) a Kaldi-based ASR system, that uses an autoencoder to reduce the size of SSL features; and (3) validating the proposed SSL feature optimization in a KWS task. We evaluate the proposed method using a DNN–HMM model in Kaldi on two standard dysarthric speech datasets: TORGO and UAspeech. Our approach shows that selecting severity-specific SSL layers, combined with autoencoder (AE)-based feature optimization, leads to significant improvements over both zero-shot and fine-tuned SSL baselines. On TORGO, our method achieved a WER of 23.12%, outperforming zero-shot (60.35%) and fine-tuned SSL model (40.48%). On UAspeech, it reached 50.33% WER, surpassing both the fine-tuned (51.04%) and MFCC-based systems (58.67%). Layer-wise analysis revealed consistent trends: lower layers were more effective for very high-severity speech, while mid-to-upper layers performed better for low/medium-severity cases. Further, in the KWS task, later SSL layers showed the best performance, with our proposed system outperforming the MFCC baseline. These findings highlight the generalization of our proposed method, which combines layer-specific selection and autoencoder-based optimization of SSL features, for dysarthric speech processing tasks.</div></div>","PeriodicalId":50630,"journal":{"name":"Computers & Electrical Engineering","volume":"131 ","pages":"Article 110921"},"PeriodicalIF":4.9,"publicationDate":"2026-01-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145927370","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-02DOI: 10.1016/j.compeleceng.2025.110926
Muhammad Farhan Khan , Sile Hu , Yuan Gao , Yu Guo , Yuan Wang , Maryam Saeed , Yucan Zhao , Jiaqiang Yang
Accurate and adaptive multi-horizon electricity load forecasting is essential for secure operation of modern power systems and for the integration of variable renewable generation. This paper proposes DWRNet, a Dynamic Weighted Residual Network that combines statistical decomposition, deep residual learning, and reinforcement learning (RL)-based adaptive fusion. A Fruit Fly Optimization-tuned Holt-Winters model first extracts the dominant seasonal-trend component, while a Long Short-Term Memory (LSTM) network learns the nonlinear residual structure. A continuous-action policy-gradient controller then produces horizon-dependent convex weights that balance the statistical and neural forecasts, enabling the ensemble to adapt to changing load regimes while remaining lightweight enough for EMS/SCADA deployment. DWRNet is evaluated on four years of hourly load data from two structurally different power systems (Inner Mongolia, China and Germany) over 24 h, 168 h, and 720 h horizons, and compared against strong baselines including SVR, LSTM, GRU, CNN, CNN-LSTM, and recent Transformer-based models (Informer, FEDformer) under a common rolling-origin protocol. Across both regions and all horizons, DWRNet consistently achieves the best or near-best MAE, RMSE, sMAPE and R² values, with particularly notable gains on weekly and monthly forecasts. Robustness is assessed through cross-validation with varying training fractions, bootstrap-based confidence intervals, ablation studies, and residual diagnostics, which collectively indicate that the improvements are stable and not attributable to overfitting. A complexity analysis and runtime benchmarks further show that the RL-based blending stage adds only modest offline training cost and negligible inference overhead. DWRNet offers a practical and scalable solution for real-time energy forecasting, with strong potential for use in energy management systems, dispatch operations, and smart grid planning.
{"title":"A hybrid reinforcement learning framework for adaptive multi-horizon electricity load forecasting: The DWRNet approach","authors":"Muhammad Farhan Khan , Sile Hu , Yuan Gao , Yu Guo , Yuan Wang , Maryam Saeed , Yucan Zhao , Jiaqiang Yang","doi":"10.1016/j.compeleceng.2025.110926","DOIUrl":"10.1016/j.compeleceng.2025.110926","url":null,"abstract":"<div><div>Accurate and adaptive multi-horizon electricity load forecasting is essential for secure operation of modern power systems and for the integration of variable renewable generation. This paper proposes DWRNet, a Dynamic Weighted Residual Network that combines statistical decomposition, deep residual learning, and reinforcement learning (RL)-based adaptive fusion. A Fruit Fly Optimization-tuned Holt-Winters model first extracts the dominant seasonal-trend component, while a Long Short-Term Memory (LSTM) network learns the nonlinear residual structure. A continuous-action policy-gradient controller then produces horizon-dependent convex weights that balance the statistical and neural forecasts, enabling the ensemble to adapt to changing load regimes while remaining lightweight enough for EMS/SCADA deployment. DWRNet is evaluated on four years of hourly load data from two structurally different power systems (Inner Mongolia, China and Germany) over 24 h, 168 h, and 720 h horizons, and compared against strong baselines including SVR, LSTM, GRU, CNN, CNN-LSTM, and recent Transformer-based models (Informer, FEDformer) under a common rolling-origin protocol. Across both regions and all horizons, DWRNet consistently achieves the best or near-best MAE, RMSE, sMAPE and R² values, with particularly notable gains on weekly and monthly forecasts. Robustness is assessed through cross-validation with varying training fractions, bootstrap-based confidence intervals, ablation studies, and residual diagnostics, which collectively indicate that the improvements are stable and not attributable to overfitting. A complexity analysis and runtime benchmarks further show that the RL-based blending stage adds only modest offline training cost and negligible inference overhead. DWRNet offers a practical and scalable solution for real-time energy forecasting, with strong potential for use in energy management systems, dispatch operations, and smart grid planning.</div></div>","PeriodicalId":50630,"journal":{"name":"Computers & Electrical Engineering","volume":"131 ","pages":"Article 110926"},"PeriodicalIF":4.9,"publicationDate":"2026-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145885937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
IF 4.9 3区 计算机科学Q1 COMPUTER SCIENCE, HARDWARE & ARCHITECTURE
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号