C. Santhosh Kumar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
C. Santhosh Kumar has a total of 59 co-authors. Here are the top co-authors by number of shared publications:
- K. I. Ramachandran
Department of Mechanical Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India
5 shared publications K. T. Sreekumar
Amrita Vishwa Vidyapeetham
5 shared publicationsK.I. Ramachandran
Intelligent Systems Research Laboratory Centre for Computational Engineering and Networking Amrita School of Engineering Coimbatore Amrita Vishwa Vidyapeetham India
4 shared publicationsA. Anand Kumar
Department of Neurology, Amrita Institute of Medical Sciences, Amrita Vishwa Vidyapeetham, Kochi, India
3 shared publications- Gopinath Ranjith
Machine Intelligence Research Laboratory, Department of Electronics and Communication Engineering, Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, Amrita University, India-641112
3 shared publications Kuruvachan K. George
Acoustic, Speech and Languages Department (ASL) Institute for Infocomm Research (I2R) Agency for Science Technology and Research (A*STAR) Singapore
3 shared publications
Latest Publications
- Open Access
- 4 August 2026
- Published by Ubiquity Press, Ltd. in Transactions of the International Society for Music Information Retrieval
- Vol. 9 (1) , 423-439
Attention‑Enhanced Convolutional Neural Network for Music Genre Classification
The complex interplay of harmonic, rhythmic, and timbral elements within musical signals poses a challenge for accurate music genre classification. Conventional approaches have limitations in effectively capturing both localized features and long‑range contextual relationships within the audio data. To overcome this, we introduce an attention‑enhanced convolutional neural network (AECNN) architecture that employs convolutional layers to efficiently extract localized features and an attention mechanism to capture long‑range temporal dependencies within the input audio segment. This integrated approach empowers AECNN to comprehensively learn both local features and segment‑level contextual relationships, resulting in enhanced accuracy and robustness in classifying diverse music genres. The performance of the proposed model is evaluated on two datasets, the widely used GTZAN dataset and the novel IMG‑Amrita dataset, encompassing 12 distinct Indian music genres. Experiments conducted with varying audio sample durations (30, 15, 10, 6, and 3 s) demonstrate the impact of sample length on classification accuracy. Notably, AECNN achieves peak accuracies of 96.02% on the GTZAN dataset and 98.94% on the IMG‑Amrita dataset with 3‑s samples, underscoring its ability to capture intricate musical nuances with limited audio data. These findings highlight the significant potential of the AECNN for robust and versatile performance across a wide range of music classification tasks.
- 4 August 2026
- Published by Ubiquity Press, Ltd. in Transactions of the International Society for Music Information Retrieval
- Vol. 9 (1) , 423-439
Open AccessAttention‑Enhanced Convolutional Neural Network for Music Genre Classification
The complex interplay of harmonic, rhythmic, and timbral elements within musical signals poses a challenge for accurate music genre classification. Conventional approaches have limitations in effectively capturing both localized features and long‑range contextual relationships within the audio data. To overcome this, we introduce an attention‑enhanced convolutional neural network (AECNN) architecture that employs convolutional layers to efficiently extract localized features and an attention mechanism to capture long‑range temporal dependencies within the input audio segment. This integrated approach empowers AECNN to comprehensively learn both local features and segment‑level contextual relationships, resulting in enhanced accuracy and robustness in classifying diverse music genres. The performance of the proposed model is evaluated on two datasets, the widely used GTZAN dataset and the novel IMG‑Amrita dataset, encompassing 12 distinct Indian music genres. Experiments conducted with varying audio sample durations (30, 15, 10, 6, and 3 s) demonstrate the impact of sample length on classification accuracy. Notably, AECNN achieves peak accuracies of 96.02% on the GTZAN dataset and 98.94% on the IMG‑Amrita dataset with 3‑s samples, underscoring its ability to capture intricate musical nuances with limited audio data. These findings highlight the significant potential of the AECNN for robust and versatile performance across a wide range of music classification tasks.
- Open Access
- 3 August 2026
- Published by Springer Nature in Discover Applied Sciences
- Vol. 8 (9) , 959
Physics informed spectrogram normalization and domain adaptation for intermittent fault diagnosis
- 3 August 2026
- Published by Springer Nature in Discover Applied Sciences
- Vol. 8 (9) , 959
Open AccessPhysics informed spectrogram normalization and domain adaptation for intermittent fault diagnosis
- Open Access
- 26 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 80928-80945
Minimizing Person-Specific Variations Using Cepstral Features and Supervised Clustering for Improving Stress Identification Accuracy
The escalating prevalence of stress disorders necessitates innovative and automated diagnostic tools. In electrocardiogram (ECG)-based stress identification systems (SIS), person-specific physiological variations do not contribute to stress-related information and often degrade classification performance. This work focuses on improving SIS performance by minimizing non-informative, person-specific variations in ECG signals through cepstral processing and supervised clustering. Person-specific characteristics in ECG signals are manifested in the frequency domain as multiplicative factors. By extracting cepstral features, these multiplicative components are transformed into additive factors using logarithmic scaling and subsequently decorrelated using the discrete cosine transform (DCT). Cepstral mean subtraction (CMS) and cepstral smoothing are then applied to further minimize person-specific variations and enhance the performance of SIS. To reduce residual intra-class dispersion, supervised clustering of deep features was performed. ECG signals from 104 participants (53 meditators and 51 individuals with clinically diagnosed stress disorders) were analyzed. The baseline system using raw spectral features achieved a mean accuracy of 71.62%, sensitivity of 52.48%, specificity of 88.74%, and AUC of 0.804. The final system incorporating cepstral smoothing and supervised clustering achieved a mean accuracy of 91.18%, sensitivity of 87.10%, specificity of 91.25%, and AUC of 0.922, representing an improvement of 19.56%, 34.62%, 2.51%, and 0.118, respectively. The proposed approach was further validated on the publicly available Wearable Stress and Affect Detection (WESAD) dataset, and similar performance improvements were observed under induced stress conditions. These results emphasize the importance of mitigating person-specific variations through cepstral processing, leading to a robust and generalizable ECG-based stress identification framework.
- 26 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 80928-80945
Open AccessMinimizing Person-Specific Variations Using Cepstral Features and Supervised Clustering for Improving Stress Identification Accuracy
The escalating prevalence of stress disorders necessitates innovative and automated diagnostic tools. In electrocardiogram (ECG)-based stress identification systems (SIS), person-specific physiological variations do not contribute to stress-related information and often degrade classification performance. This work focuses on improving SIS performance by minimizing non-informative, person-specific variations in ECG signals through cepstral processing and supervised clustering. Person-specific characteristics in ECG signals are manifested in the frequency domain as multiplicative factors. By extracting cepstral features, these multiplicative components are transformed into additive factors using logarithmic scaling and subsequently decorrelated using the discrete cosine transform (DCT). Cepstral mean subtraction (CMS) and cepstral smoothing are then applied to further minimize person-specific variations and enhance the performance of SIS. To reduce residual intra-class dispersion, supervised clustering of deep features was performed. ECG signals from 104 participants (53 meditators and 51 individuals with clinically diagnosed stress disorders) were analyzed. The baseline system using raw spectral features achieved a mean accuracy of 71.62%, sensitivity of 52.48%, specificity of 88.74%, and AUC of 0.804. The final system incorporating cepstral smoothing and supervised clustering achieved a mean accuracy of 91.18%, sensitivity of 87.10%, specificity of 91.25%, and AUC of 0.922, representing an improvement of 19.56%, 34.62%, 2.51%, and 0.118, respectively. The proposed approach was further validated on the publicly available Wearable Stress and Affect Detection (WESAD) dataset, and similar performance improvements were observed under induced stress conditions. These results emphasize the importance of mitigating person-specific variations through cepstral processing, leading to a robust and generalizable ECG-based stress identification framework.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 94874-94898
Improving Epileptic Seizure Prediction Using ECG From Wearable Sensors Through Spectral Whitening and Cepstral Processing
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 94874-94898
Open AccessImproving Epileptic Seizure Prediction Using ECG From Wearable Sensors Through Spectral Whitening and Cepstral Processing
- 4 March 2025
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 189, 109929
Seizure information enrichment in ECG through spectral whitening for improving epileptic seizure prediction
- 4 March 2025
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 189, 109929
Seizure information enrichment in ECG through spectral whitening for improving epileptic seizure prediction
- Open Access
- 5 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 107944-107958
Deep Discriminative Feature Learning and Feature Space Transformation for Scalable Machine Fault Diagnosis
Machine fault diagnosis for expensive high-capacity machines may be challenging as it is infeasible to induce faults and collect data. In this work, we propose a scalable fault modeling approach to address this challenge. The proposed method utilizes a low-capacity machine (source) to generate synthetic fault data representative of the high-capacity machine (target), facilitating fault modeling for the target machine. This is achieved by learning a feature space transformation from the source machine to the target machine in the healthy condition, employing constrained maximum likelihood linear regression (CMLLR). Feature spaces of the source and target machines in healthy condition are modeled using two Gaussian mixture models (GMMs) to estimate the CMLLR transformation parameters. This transformation is then applied on the source system’s fault data to synthesize the data representative of the high capacity machine’s fault data. For this feature space transformation to work effectively, the feature spaces should be linearly transformable. We employ a deep neural network with discriminative feature extraction capabilities to derive features. To validate the scalability of fault diagnosis across varying machine capacities, we conducted experiments on three distinct datasets. The scalable fault diagnosis system for synchronous generators achieved accuracy improvements of 4.65% (R-phase), 4.75% (Y-phase), and 5.11% (B-phase) compared to the baseline system. The system for geared motors showed a 23.76% accuracy improvement and the cross-machine scalable system achieved a 32.5% accuracy improvement over the baseline system. The results establish the effectiveness of the proposed approach for machine fault diagnosis across systems with varying capacities.
- 5 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 107944-107958
Open AccessDeep Discriminative Feature Learning and Feature Space Transformation for Scalable Machine Fault Diagnosis
Machine fault diagnosis for expensive high-capacity machines may be challenging as it is infeasible to induce faults and collect data. In this work, we propose a scalable fault modeling approach to address this challenge. The proposed method utilizes a low-capacity machine (source) to generate synthetic fault data representative of the high-capacity machine (target), facilitating fault modeling for the target machine. This is achieved by learning a feature space transformation from the source machine to the target machine in the healthy condition, employing constrained maximum likelihood linear regression (CMLLR). Feature spaces of the source and target machines in healthy condition are modeled using two Gaussian mixture models (GMMs) to estimate the CMLLR transformation parameters. This transformation is then applied on the source system’s fault data to synthesize the data representative of the high capacity machine’s fault data. For this feature space transformation to work effectively, the feature spaces should be linearly transformable. We employ a deep neural network with discriminative feature extraction capabilities to derive features. To validate the scalability of fault diagnosis across varying machine capacities, we conducted experiments on three distinct datasets. The scalable fault diagnosis system for synchronous generators achieved accuracy improvements of 4.65% (R-phase), 4.75% (Y-phase), and 5.11% (B-phase) compared to the baseline system. The system for geared motors showed a 23.76% accuracy improvement and the cross-machine scalable system achieved a 32.5% accuracy improvement over the baseline system. The results establish the effectiveness of the proposed approach for machine fault diagnosis across systems with varying capacities.
- 28 June 2024
- Published by Elsevier in Epilepsy Research
- Vol. 204, 107403
AI in ECG: Validating an ambulatory semiology labeller and predictor
- 28 June 2024
- Published by Elsevier in Epilepsy Research
- Vol. 204, 107403
AI in ECG: Validating an ambulatory semiology labeller and predictor
- 15 March 2023
- Published by Springer Nature
- p. 415-428
Improving Sleep Apnea Screening with Variational Mode Decomposition and Deep Learning Techniques
- 15 March 2023
- Published by Springer Nature
- p. 415-428
Improving Sleep Apnea Screening with Variational Mode Decomposition and Deep Learning Techniques
- 27 August 2021
- Published by Springer Nature
- p. 435-445
Bottleneck Features for Enhancing the Synchronous Generator Fault Diagnosis System Performance
- 27 August 2021
- Published by Springer Nature
- p. 435-445
Bottleneck Features for Enhancing the Synchronous Generator Fault Diagnosis System Performance
- 1 June 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 52-57
Stator Inter-turn Fault Diagnosis and Fault Location in Synchronous Generator using Dual Extended Kalman Filter and Linear Regression Analysis
Contemporary society is very much dependent on electric power, and so good fault detection schemes are necessary for the smooth functioning of the grid. Synchronous generators are very essential in power generation from industries to military applications. Stator inter-turn or turn to turn short is the most likely short to occur at the early stage which leads to further complications resulting in damage to the machine. In this paper, we propose a method to detect and find the number of turns shorted in the stator winding using Dual Extended Kalman filter and Linear regression.
- 1 June 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 52-57
Stator Inter-turn Fault Diagnosis and Fault Location in Synchronous Generator using Dual Extended Kalman Filter and Linear Regression Analysis
Contemporary society is very much dependent on electric power, and so good fault detection schemes are necessary for the smooth functioning of the grid. Synchronous generators are very essential in power generation from industries to military applications. Stator inter-turn or turn to turn short is the most likely short to occur at the early stage which leads to further complications resulting in damage to the machine. In this paper, we propose a method to detect and find the number of turns shorted in the stator winding using Dual Extended Kalman filter and Linear regression.