BACKGROUND AND OBJECTIVES
Recognition and treatment of neonatal seizures, primarily diagnosed using EEG, are essential to protect the developing brain, yet only 11% of seizures are treated within 1 hour of onset. Developing seizure forecasting and detection systems could help address this issue in intensive care settings with limited neurologist availability.
METHODS
We propose an ML architecture with dual functionality: (1) forecasting seizures within short prediction intervals (5, 10, and 15 min), and (2) detecting seizures in real time, using explainable entropy-, singular value decomposition-, power spectrum-, and statistical moment-based features with an AdaBoost classifier. The same feature set supports both functionalities, with forecast reliability improving when features are derived from longer EEG intervals. The model was trained and tested on a single-channel (P3-P4, n = 82) dataset and a multi-channel public dataset (n = 79).
RESULTS
The proposed ML model was able to forecast seizures at a 15-min prediction interval (86.6%/92.4% sensitivity/specificity; Matthews correlation coefficient, MCC = 0.39), with better performance at the shorter 5-min prediction interval (95.6%/88.9% sensitivity/specificity; MCC = 0.55) for the single-channel EEG seizure group. Short-range forecasting outperformed the pioneering study on the public multi-channel EEG dataset (MCC = 0.48). The proposed ML model achieved over a 30% improvement in MCC and Pearson's coefficient on the single-channel dataset compared to the state-of-the-art deep-learning architecture (ConvNeXt).
CONCLUSION
EEG segmentation and appropriate feature selection enhance ML model performance for seizure detection in small datasets. Validation of feature-based ML models is required on both single and multi-channel EEG, given that quantitative features can be sensitive to different bipolar channel pairs. Although the method has outperformed the pioneering studies in seizure forecasting, further improvement in performance and more extensive validation are needed for clinical adoption.
T. Skoric, M. Djermanovic, J. M. O'Toole et al.· Computer Methods and Program...· 0 citations
Accurate detection of forced oscillations is important for the stable operation of power systems. The method based on prior knowledge relies on manual feature extraction, which has limited ability to characterize non-stationary signals. While deep learning (DL) methods can automatically learn features, they may overlook the physical mechanisms of power systems, potentially leading to misjudgments. We propose a Hybrid Knowledge-DL network (HKD-SVM) that utilizes Support Vector Machine (SVM) as the classifier. In our method, Discrete Wavelet Transform (DWT) is used to represent the time–frequency structure of the input signals, and DL features are extracted by Convolutional Neural Network (CNN) from this time–frequency representation. These learned features are subsequently fused with prior knowledge features that carry explicit physical interpretations, thereby constructing a more discriminative feature representation space. Finally, SVM is adopted as the classifier, making the network well-suited for nonlinear, high-dimensional classification scenarios with limited training samples, which are common in power system applications. Experiments on both simulated and real-world phasor measurement unit (PMU) data demonstrate that HKD-SVM outperforms purely data-driven and purely knowledge-driven methods. The proposed method provides an effective solution for power system oscillation detection.
This paper presents a comparative study for epilepsy monitoring using EEG signals along two main axes. The first axis consists of comparing the performance of the differentiation technique, which is known to be very important for the study of non-stationarity, with wavelet transform, which is widely used for detecting different brain rhythms. The second part aims to compare the performance of different machine learning algorithms, including k-Nearest Neighbors (k-NN), Decision Trees, Random Forest, and Support Vector Machines (SVM). Used features are statistical and higher order statistics characteristics such as mean, standard deviation, median, Min-Max, Kurtosis and Skewness. We tested our approach on publicly available and widely used datasets in the literature, namely the University of Bonn dataset and the Bern-Barcelona dataset. The experimental results demonstrate the effectiveness of the differentiation method as an important tool for EEG preprocessing, leading to very high performance.
Ines Bouzouita, Zayneb Brari, S. Belghith· International Conference on...· 0 citations
Epileptic seizure detection from EEG signals remains challenging due to their non-stationary and complex nature. This study presents a comparative analysis of Discrete Wavelet Transform (DWT)-based feature extraction combined with classical machine learning classifiers (SVM, KNN, and MLP) to distinguish normal and epileptic EEG signals. Using the publicly available Bonn University dataset (Sets A and E), EEG signals were decomposed using the Daubechies-4 (db4) wavelet into five decomposition levels corresponding to standard frequency bands (Delta, Theta, Alpha, Beta, Gamma). Seven statistical features—energy, mean amplitude, standard deviation, Shannon entropy, relative wavelet energy (RWE), kurtosis, and skewness—were extracted from each sub-band. A stratified 10-fold cross-validation with a leakage-controlled record-level partitioning strategy was implemented to reduce optimistic bias. Since subject-level identifiers are unavailable in the public Bonn dataset, the validation was designed to avoid re-splitting individual EEG records across training and testing stages. Results demonstrate that kurtosis-based features consistently achieve the highest accuracy (99.8% ± 0.3) across all classifiers, significantly outperforming other features (p < 0.01). These findings underscore the potential of higher-order statistical descriptors, particularly kurtosis, for EEG-based epileptic seizure detection under a controlled benchmark setting.
H. Hindarto, Ade Eviyanti, A. Ahfas et al.· International journal of ele...· 0 citations
To address the issues of insufficient feature extraction and low localization accuracy in distribution network fault diagnosis, this study proposes a fault classification and localization method based on APC-SVM and PC-AZOA. The model performs a simultaneous decomposition of three-phase signals using multivariate variational modal decomposition and employs the energy entropy of each model component as the feature vector; During the classification stage, the method integrates electrical and physical constraints, introducing three-phase energy imbalance and variance into the support vector machine ’ s parameter optimization process for the first time to dynamically adjust the penalty factor and kernel parameters; finally, a traveling wave propagation time error model is constructed, and an adaptive zebra optimization algorithm constrained by physical information is proposed. By innovatively embedding prior physical knowledge into the search space constraints, the method effectively suppresses invalid searches and improves convergence efficiency. Experimental results show that the model achieves a classification accuracy of up to 98.4% with a positioning error below 1%, demonstrating both high precision and high efficiency.
Dahua Li, Xinrui Yang, Yu Song et al.· 2026 IEEE International Conf...· 0 citations
This study presents an applied comparative evaluation of automated rolling bearing fault identification using Artificial Neural Networks (ANNs) optimized through Bayesian Optimization (BO), Particle Swarm Optimization (PSO), and Genetic Algorithm (GA). Vibration signals were collected from bearings operating under five health conditions (healthy, outer ring fault, inner ring fault, ball fault, and combined faults), at three rotational speeds, and along three measurement directions. The acquired signals were preprocessed using filtering, normalization, and segmentation. Time-domain and Fast Fourier Transform (FFT)-based frequency-domain features were extracted and used to train ANN models. The ANN architectures, including hidden layers, neurons, and activation functions, were optimized using BO, PSO, and GA, resulting in six configurations. Since the problem is a multi-class classification task, performance was assessed using F1-score, Accuracy, precision, and recall. The optimized ANN models were also benchmarked against Support Vector Machine (SVM), K-Nearest Neighbors (kNN), and Random Forest (RF) classifiers using the same feature sets. Results show that FFT -based features consistently outperformed time-domain features, and ANN-PSO with FFT-based features achieved the best performance, with F1-score = 0.982, Accuracy = 0.998, precision = 0.982, and recall = 0.982. This work contributes a systematic applied comparison of ANN optimization strategies rather than a fundamentally new machine-learning architecture, highlighting optimized ANNs as competitive and computationally efficient solutions for bearing fault diagnosis.
Khoualdia Kaaïs, Khoualdia Tarek, M. Nahal· International Journal of Pro...· 0 citations
To address the challenges of conventional feature extraction methods in capturing nonlinear dependencies within fault signals and reducing feature redundancy during hydropower units fault diagnosis, this paper proposes a feature selection method integrating the Minimum Redundancy Maximum Relevance (mRMR) criterion with Copula Entropy (mRMR-CE). This method utilizes Copula Entropy (CE) to capture both linear and nonlinear dependencies in vibration signals. Combined with the mRMR criterion to suppress feature redundancy, it achieves stable selection of highly discriminative features. In noisy environment and with various classifiers, the method shows strong performance and stability. To validate the effectiveness of the proposed method, five feature selection approaches-CE, mRMR, mRMR-CE, Pearson, and PCA-were applied to the training samples during the feature selection stage. Lastly, the chosen features were input into four different types of classifiers for training and testing: Support Vector Machine (SVM), Random Forest (RF), Multi-layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). Experimental results demonstrate that the proposed method exhibits outstanding performance on both the CWRU bearing fault dataset and the Unit 3 dataset from a hydropower plant. It achieved an average accuracy of 99.87% and an F1 score of 98.25% on the CWRU dataset, and reached 100% accuracy with an F1 score of 99.31% on the Unit 3 dataset. These results significantly outperform traditional feature selection methods while demonstrating high stability and robustness.
Bo Li, Jiahao Li, Guangtao Zhang et al.· Engineering Research Express· 0 citations
This research presents an integrated condition monitoring framework for deep groove ball bearings by combining Complex Morlet Wavelet (CMW) analysis, machine learning techniques, thermographic analysis, and SKF Machine Condition Advisor tools. The proposed methodology employs Fast Fourier Transform (FFT) and Complex Morlet Wavelet-based vibration signal processing to extract discriminative time–frequency features for the early detection and diagnosis of bearing faults under both single and combined fault conditions. To evaluate fault classification performance, three machine learning algorithms, namely Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Random Forest (RF), are implemented and comparatively analyzed. Experimental investigations are conducted on a laboratory-scale bearing test rig operating under controlled conditions. The extracted wavelet-based features effectively characterize fault-induced vibration signatures, enabling accurate fault identification and classification. Comparative results indicate that Random Forest achieves the highest classification accuracy, followed by SVM and ANN. The average classification accuracies obtained using RF, SVM, and ANN are 97.48%, 95.27%, and 87.20%, respectively, demonstrating the superior robustness and generalization capability of the ensemble learning approach. Furthermore, thermographic analysis and SKF Machine Condition Advisor measurements provide complementary information for validating fault severity and machine health conditions, thereby enhancing diagnostic reliability. Although the present study is limited to constant-speed operation and controlled laboratory environments, the proposed framework demonstrates significant potential for predictive maintenance and intelligent condition monitoring applications. The integration of advanced time–frequency analysis, machine learning-based fault classification, and practical condition monitoring tools offers an effective and reliable solution for machinery health assessment in industrial environments.
M. Maurya, Chandrabhanu Malla, I. Panigrahi et al.· F1000Research· 0 citations
Deep-OCR (DeepSeek-OCR) advances document recognition by treating the visual modality as an optical compression medium, enabling long-context OCR at low token cost. However, its increased complexity may introduce new security vulnerabilities. In this paper, we present, to the best of our knowledge, the first pure black-box adversarial attack against a generative OCR vision-language model, where only the decoded string can be queried and no gradients, logits, or model internals are available. We recast the attack as a zeroth-order optimization problem driven by a bounded scalar loss defined directly on the string output via sequence similarity, and estimate the gradient with a random-direction finite-difference scheme whose query cost is independent of the image dimension. An Adam update with ell_infinity projection yields imperceptible perturbations for both untargeted and targeted objectives. Pilot experiments on Deep-OCR validate the string-only attack and evaluation pipeline and expose severe qualitative decoder failures, including repetition, truncation, and prompt leakage. They also show that controlled targeted rewriting remains substantially harder than untargeted degradation; we avoid claiming targeted success until the pre-registered evaluation is complete.
Wenbo Sun, Hongzong Li, Yanyun Wang et al.· 0 citations
Large language models (LLMs) may still produce unsafe responses under optimized adversarial prompts, even after safety alignment. Existing gradient-based jailbreak evaluation methods usually use raw gradient signals for discrete prompt search, but raw gradients may be unstable in both magnitude and direction. This paper proposes a gradient-calibrated adversarial prompt optimization framework for LLM safety evaluation. The framework adopts REINFORCE-style randomized discrete prompt optimization as the parent objective and inserts a unified calibration interface before gradient-derived token-search signals are used. We instantiate this interface with two complementary primitives: LSGM for layer-wise magnitude calibration and LiLA for representation-space directional calibration. Experiments on HarmBench show that joint calibration improves attack success rate on multiple open-source safety-aligned models. Ablation results further suggest that the two primitives are more stable when used jointly than when used alone. The framework provides a compact white-box diagnostic tool for analyzing gradient quality in discrete adversarial prompt optimization, while transfer and defense results reveal clear limitations in generalizability and practical robustness.
Multimodal Large Language Models (MLLMs) are increasingly deployed for nuanced content safety and moderation tasks, yet they remain vulnerable to adversarial attacks and out-of-distribution edge cases. Traditional active learning and manual annotation fail to scale against the complexity and volume of novel multimodal threats. In this paper, we propose an automated, agentic red-teaming framework that systematically synthesizes difficult examples using an iterative strategy that proposes novel hypotheses as well as mutating on past attempts. Leveraging a multi-agent architecture that consists of a high-reasoning Architect agent, an advanced image generator, and a multi-level verification committee of LLM raters, our system autonomously uncovers boundary-pushing violations and ambiguous policy edge cases without any human intervention. By employing these carefully synthesized adversarial examples as in-context demonstrations via test-time Retrieval, we substantially improve the target model's robustness, reducing the False Negative Rate (FNR) from 41.2% to 24.5% in a public image safety benchmark without relying on any human labeling.
Genglin Liu, Muye Zhang, Krishnamurthy Viswanathan et al.· 0 citations
While prompt engineering enhances the capabilities of Large Language Models (LLMs), it also exposes critical safety concerns. Due to the inherent brittleness of their static safety boundaries, LLMs are vulnerable to jailbreak prompts , i.e. adversarial inputs designed to by-pass safeguards and induce the generation of harmful content. Existing detection mechanisms rely on static model components or fixed decision thresholds, limiting their ability to generalize to evolving attack patterns and continual model updates. To bridge this gap, we propose RLShield , a dynamic jailbreak detection framework that employs reinforcement learning for adaptive threshold selection. RLShield incorporates three key innovations: ( i ) a dynamic retrieval and LLM-based rewriting module to simulate diverse adversarial contexts; ( ii ) a cross-layer representation analysis to pinpoint safety-critical parameters; and ( iii ) a Soft Actor-Critic (SAC) based agent that learns to predict optimal, sample-specific detection thresholds. Experimental results demonstrate that RLShield consistently outperforms state-of-the-art base-lines in detection performance while maintaining high computational efficiency. Notably, it improves F1 by up to 7.3%, while achieving an average of 3 × gain in inference efficiency across multiple LLM backbones. Our codes are available at this website.
Zhao Tong, Pengfei Yang, Yimeng Gu et al.· Annual Meeting of the Associ...· 0 citations