Reliable wearable near-infrared spectroscopy (NIRS) for bladder monitoring requires both a measurable volume-dependent optical signal and tolerance to structured channel loss. This study combines a 16-condition Monte Carlo eXtreme (MCX) photon-transport analysis, phantom measurements, and an acute single-animal porcine experiment. The platform used 655, 665, 810, 835, 930, and 980 nm illumination across four sectors. The MCX comparison spanned fixed and geometry-modulated overlying layers, two fat thicknesses, two bladder shapes, and two expansion anchors. Fixed-overlying-layer configurations collapsed to an intercept-only null prediction (MAE, 150 mL), whereas geometry-modulated configurations yielded MAEs of 22–51 mL. Because absorption was not independently varied under strictly fixed geometry, these findings support geometry-sensitive optical modulation rather than a definitive absorption-versus-geometry decomposition. A sector-aligned fault-aware attention model (FAA2), with gate parameters selected only from phantom development data (α = 5, β = 100, and τ = 0.30), achieved a test MAE of 6.19 ± 1.12 mL (mean ± SD across three seeds; n = 210). Under known complete LED-state loss, all 2,324 one-, two-, and three-LED configurations were compared with mean, median, and five-nearest-neighbor (5-NN) reconstruction. Five-NN was a strong baseline; 5-NN + FAA2 achieved mean configuration MAEs of 6.49, 6.53, and 6.59 mL, with corresponding 95th-percentile MAEs of 6.77, 6.97, and 7.27 mL. On an independent porcine test cycle, Lasso yielded an MAE of 14.12 mL, supporting acute single-animal feasibility. These findings establish a geometry-sensitive, fault-tolerant framework for known LED-state loss without establishing automatic fault detection or population-level clinical validity.
Traditional CAPTCHA systems are used to prevent human users and automated bots. Recent advances in machine learning and optical character recognition (OCR) have significantly reduced the utility of the conventional fixed-point design. The study introduces a curriculum-based adaptive CAPTCHA system that integrates challenge generation, multi-profile adversarial simulation, and human behaviour modelling to dynamically adjust challenge difficulty through a closed-loop feedback mechanism. During each curriculum iteration, one attacker profile is sampled from predefined weak, average, and strong adversarial profiles to provide diverse feedback for curriculum adaptation. The attacker component represents multiple adversarial capability levels and provides performance feedback for curriculum-based adaptation. Unlike existing approaches, the proposed framework integrates warp-based perturbations and entropy-conscious transformations into a curriculum learning paradigm, enabling optimisation of the security/usability balance to gradually and consistently. To achieve the quantitative performance, the Adversarial Robustness Index (ARI), a single index that simultaneously assesses the challenge’s difficulty, human success rate, and resistance to automated attacks is proposed. Further, a nonlinear formulation of difficulty is also provided to model consistency in complexity. The present study reveals that the suggested method achieves a mean ARI of 0.7775 ± 0.0278, the bot success rate drops to 0.0646 (a 78.46 percent improvement over baseline), while the human performance score stays high at 0.8515. Furthermore, calibration analysis demonstrates an Expected Calibration Error (ECE) of 0.0018, indicating strong agreement between predicted automated attack success probabilities and empirically observed attack outcomes. Extensive evaluation, including component sensitivity analysis, usability consistency analysis across increasing CAPTCHA difficulty levels, stress testing, and counterfactual assessment, has demonstrated the importance of entropy and warp components in enhancing resilience. External validation using OCR-based attack models on CAPTCHA samples generated by the proposed adaptive framework demonstrates low automated recognition accuracy (0.0645), while controlled human evaluation maintains high usability, achieving an accuracy of up to 90% with an average completion time of 3.48 seconds. These results demonstrate strong robustness against the evaluated OCR-based automated attack models while maintaining high human usability. Comprehensive evaluation against transformer-based and multimodal CAPTCHA solvers remains future work. The present study reports aggregate usability statistics rather than participant-wise demographic analysis. Comprehensive human-subject evaluation involving larger and more diverse participant groups remains future work.
In recent deep learning-based path loss prediction methods, path loss map generation is commonly formulated as an image-to-image translation task. Most state-of-the-art methods in this field incorporate convolutional layers into their deep learning architectures. However, standard convolutional layers often struggle to effectively model the inherent radial nature of radio signal propagation due to the rectangular sliding mechanism of convolutional filters. In this paper, a radial transformation of the input data is proposed in order to better align the propagation geometry in path loss maps with the convolutional sliding process. Rearranging the pixels in path loss maps according to the proposed radial transformation enables the model to learn propagation patterns more efficiently. Experimental results obtained across several state-of-the-art path loss map prediction methods demonstrate that the proposed transformation can achieve up to a 25% reduction in root mean square error. Consequently, the proposed approach can serve as a computationally efficient alternative to deterministic ray-tracing-based models in applications where both prediction accuracy and computational efficiency are of primary importance.
Medicinal plant identification plays a critical role in traditional medicine, pharmaceutical research, the nutraceutical industry, and biodiversity conservation. However, traditional identification heavily relies on expert knowledge and manual observation, which are often time-consuming and difficult to apply in the real-world. Recently, advances in computer vision, particularly deep learning methods, have contributed to the development of automated medicinal plant identification. A systematic review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. This search was conducted in Web of Science Core Collection, IEEE Xplore, ACM Digital Library, and Scopus, on 15 April 2026. Papers published by MDPI were also included with the same search strategy. Initially, 540 records were identified, of which 81 studies were ultimately included in the analysis. This review systematically analyzes the publication trends, datasets, algorithms, research tasks, evaluation metrics, data quality, operational applicability, and deployment readiness. The findings reveal that 64.2% of the studies were published within the last three years, suggesting increased research activity in recent years. The application of deep learning emerged as the mostfrequently used approach, accounting for over 86.42% of the total. The public datasets showed an upward trend recently, especially in 2025, when the utilization of public datasets first surpassed that of self-collected. But their overall coverage and representativeness remain limited. Most studies still relied on self-collected datasets acquired under controlled laboratory conditions. Although the studies generally reported high classification accuracy, only 16% of them have achieved a high level of practical deployment. This suggests anAlgorithm-to-Action gap between laboratory-level model development and practical field deployment. Overall, while computer vision technologies have demonstrated remarkable potential for medicinal plant identification, future research should increasingly focus on robustness, scalability, interpretability, and practical applicability to facilitate real-world adoption. Future directions include multimodal fusion, few-shot and continual learning, large language model integration, and edge intelligence to support the further development and practical application.
Permanent magnet synchronous motors have found wide application in high-performance systems, including electric cars and aerospace applications. Nevertheless, it is difficult to achieve accurate speed control due to the system’s nonlinearities, parameter uncertainty, and external perturbations. Traditional control approaches, such as proportional-integral (PI) and linear-quadratic regulator (LQR) controllers, are not particularly robust in such situations. Although classical sliding mode control provides enhanced robustness, it is commonly known to cause chattering and be sensitive to the choice of gain. More recent higher-order and observer-based sliding mode control methods ease some of these constraints, but they often require a priori knowledge of disturbance bounds, a low-order disturbance estimate, or complicated adaptive control. Such conditions can limit their ability to ensure convergence in finite time and discourage their application to real-time systems. To address these difficulties, this paper proposes an Arbitrary-Order Sliding Mode Controller (AOSMC), an Adaptive Arbitrary-Order Sliding Mode Differentiator (AOSMD), and a GA-based gain adaptation scheme. The AOSMC provides flexible, arbitrary-order sliding surfaces, whereas the AOSMD delivers accurate real-time estimates of high-order derivatives and lumped disturbances without prior knowledge of their bounds. The GA adaptively tunes the control and differentiator gains offline, improving performance under varying operating conditions and reducing gain overestimation. A Lyapunov-based analysis establishes finite-time convergence of the tracking and estimation errors to a small residual set whose size is determined by the measurement-noise dead zones and the a priori gain ceilings of the adaptation, and exact finite-time convergence to zero under an explicitly stated sufficient condition on those ceilings. Simulation and experimental results show that the proposed method achieves better tracking accuracy, faster convergence, and greater robustness, with less chattering, compared with traditional SMC and higher-order SMC schemes, demonstrating the effectiveness of the proposed control scheme for a practical PMSM drive.
Low-light image enhancement is a longstanding challenge in computer vision, aiming to restore visibility and contrast in suboptimal lighting conditions. Recently, the illumination-based self-reinforced retinex projection (SRRP) model and its enhanced version (ESRRP) have made strides in eliminating artifact-inducing post-processing like gamma correction. However, this model often relies on the rigid and empirical exposure control parameter. This reliance restricts its adaptability, leading to under-enhancement in extremely dark scenes or over-exposure in moderately dim environments. This paper proposes an adaptive enhanced self-reinforced retinex projection (AESRRP) model to address the above limited adaptability in SRRP and ESRRP. In the proposed AESRRP model, we firstly introduce a novel adaptive mean-aware exposure mechanism that dynamically predicts a control parameter to replace the fixed empirical exposure control parameter using a non-linear sigmoid function with two saturation points, where the lower point ensures maximum brightening for dark inputs and the higher point prevents over-enhancement in brighter regions. Then, we develop a contrast-aware piecewise function incorporating illumination standard deviation and a fusion based spatially variant sub-image processing scheme with weighted aggregation to handle complex multi-exposure scenarios. Finally, we further develop a model that integrates the contrast-aware mechanism with fusion based spatially variant sub-image processing to enhance the performance of low-light image enhancement. Extensive experiments on standard benchmark LOL, LOL-v2 and SICE datasets demonstrate that the proposed AESRRP achieves a competitve performance with the state-of-the-arts methods like URetinex++ and outperforms the existing methods including modern neural networks like UTVNet, and its predecessor ESRRP in quantitative metrics.
Federated learning (FL) promises collaborative model training without centralizing sensitive operational data, making it attractive for industrial SCADA systems where data sovereignty is paramount. But when does FL actually work for rare-event detection, and when does it fail? We answer this through a matched-protocol study across three SCADA domains (pipeline leak detection, power grid fault detection, and rotating-machinery bearing fault diagnosis) with 10 random seeds each, plus two real-data validations (Tennessee Eastman chemical process and the 3W offshore oil-well dataset) under the same 10-seed, 9-method protocol. Comparisons use paired Wilcoxon signed-rank tests with Cohen’s d. Four cross-domain findings emerge. (1) Domain-aware aggregation weights provide no statistically significant advantage over vanilla FedAvg (p > 0.05 in all domains). (2) SCAFFOLD degrades performance regardless of optimizer (SGD or Adam), consistent with a violated bounded gradient-dissimilarity assumption under extreme class imbalance; FedProx and Focal Loss offer no reliable improvement; and a 10-seed FedBN experiment is bimodal on the pipeline domain and only matches FedAvg on grid and bearing. A controlled normalization ablation (BatchNorm vs. LayerNorm vs. GroupNorm) and a gradient-dynamics trace localize the cause to inter-client gradient dissimilarity rather than batch-statistic heterogeneity, and a capacity-matched recurrent (GRU) baseline yields no improvement over the pointwise model, indicating the binding constraint is data heterogeneity, not model expressiveness. (3) Natural (per-site) data partitioning degrades performance from mildly to catastrophically (F1 from 0.78 to 0.05 depending on whether fault signatures are globally coupled or localized); a per-site factor analysis shows detection quality is governed by local event availability rather than physical site descriptors, exposing a minimum per-client event-diversity threshold for viable FL. (4) Differential privacy with the Gaussian mechanism is impractical for small-model (∼2,800-parameter) rare-event FL at the client counts and ε we evaluate: a closed-form noise-to-signal calculation shows perparameter noise exceeding typical gradient magnitude by roughly an order of magnitude at ε = 10. The contributions are (i) two pre-deployment diagnostics (a gradient-dissimilarity check and the noise-to-signal calculation) that let practitioners assess FL-method viability before running experiments, and (ii) matched-protocol deployment guidelines for SCADA-based rare-event monitoring.
In the in-flight transfer alignment of a Precision Airdrop System (PADS) equipped with inertial navigation systems (INS), achieving rapid high-precision alignment under unknown installation angles and limited carrier maneuvering remains challenging. This study proposes a unified nonsingular rapid transfer alignment scheme for arbitrary misalignments. An inertial frame is defined, and the attitude matrix of the slave INS (S-INS) is decomposed into a chained product of the master INS (M-INS) attitude matrix, installation matrix and body attitude variation matrices. A matching scheme is constructed using inertial-frame integrals of specific force and angular velocity. By taking the installation attitude quaternion as the state, a model comprising a linear state equation and second-order nonlinear measurement equation is established. Time updates follow standard Kalman filter procedures, whereas measurement updates adopt the the second-order Extended Kalman Filter (SEKF) framework with only two extra correction terms. Furthermore, a simplified linear Kalman filter (LKF) scheme is developed by discarding error terms for short-duration transfer alignment. Accordingly, the time update process is eliminated, and the algorithm essentially becomes a weighted recursive least-squares method. Simulation and flight-test results show that the proposed SEKF converges within 10 seconds, yielding horizontal alignment RMSE below 0.035° and heading alignment RMSE below 0.045°. It outperforms conventional methods in both convergence speed and estimation accuracy, which fully satisfies practical alignment requirements for PADS.
The lack of specific standards for evaluating the significant variation in children’s speech and the limited availability of annotated speech samples make it hard to automatically identify misarticulation in children speaking low-resource Indian languages. While techniques for assessing pronunciation using Automatic Speech Recognition (ASR) have shown positive results for high-resource languages, there has been little focus on applying them to identify phoneme-level misarticulation in Marathi children’s speech. Marathi is considered one of the major Indo-Aryan languages spoken by millions of people in India. But in speech and language technology research, it remains underrepresented as compared to high-resource languages such as English. This research proposes a new hybrid framework that combines machine learning methods with Kaldi-based ASR to automatically detect misarticulated speech in children who speak Marathi. The study used speech samples from children aged five to fifteen, including both typically developing speakers and those with articulation challenges, to create a specialized speech corpus. The Kaldi toolbox was employed to develop a Marathi ASR system with acoustic and linguistic models tailored for phoneme-level analysis. The proposed method combines forced alignment, phoneme location information, Goodness of Pronunciation (GOP) ratings, MFCC statistical characteristics, and phonetic confidence metrics based on posterior probability to capture articulation error more precisely. The classification of phonemes as correctly or incorrectly articulated using these qualities is trained and assessed using various machine learning methods, including K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and ensemble models. Empirical results confirmed that combining ASR-derived phonetic features with data-driven machine learning techniques is effective for resource-constrained speech therapy applications. For the early detection of speech sound issues in Marathi children, the suggested architecture provides an unbiased, scalable, and adaptable method. This work enhances speech assessment techniques for under-resourced Indian languages and opens the door for future AI-supported speech therapy and clinical evaluation tools.
Rolling bearing fault diagnosis remains constrained by three persistent challenges arising from limited labeled fault data, distribution shifts across machines and operating conditions, and privacy restrictions on centralized data collection. Transfer learning (TL) and self-supervised learning (SSL) address these challenges in complementary ways. TL facilitates knowledge transfer across related domains and tasks, whereas SSL learns transferable representations from abundant unlabeled vibration and acoustic signals. This survey provides a unified review of TL and SSL methods for bearing fault diagnosis, organized according to learning settings, transfer strategies, deployment objectives, benchmark datasets, and evaluation protocols. Recent developments are further categorized into three integration paradigms, namely self-supervised transfer (SST) learning, semi-supervised domain adaptation (SSDA), and federated TL–SSL, and their potential to support privacy-preserving, resource-efficient, and interpretable fault diagnosis in Industry 5.0 environments is examined. The synthesis establishes several quantified trends across the reviewed literature: SSL pretraining enables diagnostic accuracy comparable to fully supervised baselines with only 5–15% of the labels, federated TL–SSL foundation models operate competitively with as little as 1% labeled data, and source-free domain adaptation recovers 82–91% accuracy under domain shift without access to source data. At the same time, the analysis indicates that existing studies frequently investigate TL and SSL independently, rely predominantly on controlled benchmark datasets, and provide limited validation of cross-machine and cross-condition generalization. Key research challenges are subsequently identified, including reliable transfer under nonstationary conditions, label noise and class imbalance, computational efficiency, uncertainty quantification, and model explainability. Finally, a research agenda is presented for developing scalable, trustworthy, and deployment-ready rolling bearing fault diagnosis systems.
Accurate photovoltaic (PV) power forecasting is crucial for reliable grid operation, energy management, and the effective integration of PV systems into modern power grids. In this study, a short-term PV power forecasting framework is designed based on a residual multi-scale network enhanced with a modified coordinate attention module (MCAM). The active PV power is decomposed into intrinsic mode functions (IMFs) using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and the obtained components are combined with radiation, temperature, and wind speed variables. The feature matrix is transformed into a multi-channel tensor through a sliding-window-based representation, enabling the network to learn feature–time dependencies in a structured form. The proposed architecture includes three residual multi-scale stages with cascaded convolutional branches to extract local and contextual patterns at different kernels. The MCAM is integrated after each multi-scale block to emphasize the spatial–temporal features while suppressing less relevant features. Experiments are conducted on real-world PV data from the Yulara solar power system for 1-h, 2-h, and 3-h forecasting horizons. The proposed model achieves R² values of 0.9950, 0.9902, and 0.9836, with RMSE values of 0.4400, 0.6185, and 0.7994 for 1-h, 2-h, and 3-h horizons, respectively. The effectiveness of the proposed method is validated through comprehensive comparisons with benchmark models and evaluations across seasonal and weather conditions.
Handwriting has attracted increasing attention as a noninvasive behavioral indicator of neurological and learning disorders. However, dysgraphia screening remains challenging because current assessments rely heavily on expert judgment and lack objective, scalable, and reproducible evaluation criteria. This study proposes an end-to-end deep learning framework for dysgraphia screening using handwriting data collected from digital tablets. The framework converts online pen-trajectory signals into offline handwriting images by integrating dynamic cues such as pen pressure and velocity, thereby producing visually interpretable representations that preserve handwriting dynamics while remaining accessible to educators and clinicians. Within this framework, a lightweight self-enhanced residual-attention convolutional neural network (SERA-CNN) is employed for end-to-end classification. To examine the contribution of each architectural component, the proposed model was compared with various CNN variants. Experiments conducted on a publicly available dysgraphia benchmark dataset demonstrated that SERA-CNN achieved a classification accuracy of 88.9%, exceeding all previously reported results on the same benchmark while maintaining a compact model size of approximately 0.29 M parameters. Visualization techniques were employed to provide complementary insights into the learned feature representations and prediction mechanisms. These visualizations support model understanding and may assist practitioners in educational and clinical screening settings. By bridging online handwriting signals and visually interpretable handwriting representations through an end-to-end deep learning pipeline, the proposed framework provides an objective, scalable, and practically deployable approach for dysgraphia screening.
Network intrusion detection systems based on deep learning have shown strong performance in modeling complex traffic patterns. However, recent studies indicate that these models can be vulnerable to carefully crafted adversarial perturbations, in which small modifications to traffic features lead to incorrect classification. This work proposes a sensitivity-driven adversarial generation framework (AGF) that identifies and perturbs the most influential traffic features that affect the classifier’s decision boundary. The framework integrates feature-level sensitivity estimation, stochastic sensitivity-guided feature partitioning, and threshold-guided optimization to generate statistically consistent adversarial samples with constrained perturbation magnitude. Experimental evaluation on CIC-IDS2017, CIC-IDS2018, and CIC-DDoS2019 achieves attack success rates of 94.82%, 93.56%, and 92.71%, respectively, while maintaining low perturbation distortion, KL divergence values below 0.013, and protocol violation rates close to 1–2%. A comparative analysis against FGSM, PGD, C&W, BotDefender, VMFCVD, and Boundary attacks demonstrates improved evasion capability with smaller, more structured perturbations.
Natural Language Inference (NLI), processes pairs of sentences to extract their semantic relations. NLI has been a hot research topic, integrated as a main component in other NLP applications, from morphological spelling correction tasks to higher-level tasks such as machine translation and information extraction. Despite significant advancements in textual inference across various languages all around the world, Arabic language still suffers from limited resources in this domain, specifically, when it comes to the scarcity of robust well-constructed benchmarks necessary for effective model tuning and generalization. To address this gap, this paper introduces E-CONAN benchmarks that are composed of sentences pairs from various sources: (1) automatically-translated pairs, (2) human-validated machine-translated pairs, (3) hand-crafted pairs from teaching Arabic as foreign language books, and (4) headlines pairs from different news channels containing rumors. E-CONAN contains two benchmark datasets, E-CONAN-2, a 2-way dataset (RTE) and E-CONAN-3, a 3-way dataset (NLI). Additionally, we have used E-CONAN benchmarks to evaluate 9 state-of-the-art multilingual pretrained models using zero-shot classification. Models were evaluated across the ArNLI, XNLI, and E-CONAN datasets. Results show that E-CONAN is a potentially valuable resource for evaluating model generalization and even for fine-tuning pre-trained models. Its diverse composition, derived from a combination of sources, offers a broader and more robust assessment compared to XNLI and ArNLI. Furthermore, results show that mDeBERTa model, pre-trained on 100 languages and fine-tuned on a combination of four machine-translated datasets, demonstrated superior performance on all datasets. It achieved accuracies of 71% and 86% on the E-CONAN-3 and XNLI datasets, respectively, outperforming all other evaluated models. In addition, we have evaluated 5 LLMs on E-CONAN-3 dataset. Best results were achieved by Gemma with an accuracy of 68%. Analyzing these results showed that most errors were between neutral and contradiction classes, thus we calculated results after reformulating E-CONAN-3 using 2-way labeling. Best results are achieved by Gemma and Qwen with an accuracy of 95%, 94% respectively. Moreover, we incorporated MARBERT as a representative Arabic-specific baseline and conducted performance evaluation comparison to demonstrate how Arabic-specific models scale against cross-lingual and LLM-based approaches on the E-CONAN benchmarks. Furthermore, we conducted detailed qualitative and quantitative error analysis to analyze frequent error patterns. Results show that while pretrained models are often misled by lexical overlap, LLMs are often misled by topical familiarity. E-CONAN benchmarks will be publicly available, we hope that it will enrich research community in Arabic textual entailment and natural language inference.
Stroke stands as a primary reason for prolonged motor dysfunction that produces long-term dependence and reduced use of upper limbs throughout everyday activities. The established rehabilitation techniques need ongoing therapist guidance together with clinical facility access which restricts their availability to numerous patients. This research introduces a voice-operated exoskeleton system which enables upper arm rehabilitation for stroke survivors at home. The core of the system features a lightweight convolutional neural network (CNN) model that operates under memory and processing constraints to perform real-time voice command detection of “Up” and “Down” for initiating therapeutic flexion and extension movements. The model framework was built to deliver optimal performance and minimize computational resources. The quantized CNN model achieved an accuracy of 96.87%. The system functions with complete offline and hands-free operation on resource-constrained low-power hardware. The study demonstrates that embedded CNN-based voice recognition systems can create scalable customizable rehabilitation technologies which enhance accessibility and effectiveness for patients needing limited clinical support.
The application of scan-based test sets targeting advanced fault models is necessary for the detection of defects after fabrication and during the lifetime of a chip. Advanced fault models require large test sets, and test compaction is important for reducing the test data volume and test application time. For delay fault models, a comprehensive scan-based test set may consist of both launch-on-capture (LOC) and launch-on-shift (LOS) tests. Earlier test generation and test compaction procedures exist that optimize the two types of tests in a mixed test set together. These procedures are referred to as unpartitioned test compaction procedures. This article suggests that additional test compaction can be achieved by a test compaction procedure that partitions the set of faults into subsets F0 that consists of faults that are detected only by LOC tests, F1 that consists of faults that are detected only by LOS tests, and F2 that consists of faults that are detected by both types of tests. The new test compaction procedure, referred to as a partitioned test compaction procedure, optimizes the subset of LOC tests based on F0, and the subset of LOS tests based on F1. It then ensures that the faults in F2 are detected by the mixed test set. The benefit to test compaction results from reducing the duplication in the detection of F2. The article presents experimental results for benchmark circuits to demonstrate the effectiveness of the new partitioned test compaction procedure.
Android devices face a continuously evolving malware ecosystem that demands on-device intrusion detection capable of continual adaptation without privacy compromise. Combining Federated Learning (FL) and Continual Learning (CL) addresses this but raises two practical costs: catastrophic forgetting as malware evolves, and the bandwidth of uploading model updates every round. We present FL-CL-IDS, an on-device framework that couples a Page–Hinkley drift detector, a per-client adaptive Elastic-Weight-Consolidation (EWC) coefficient with online Fisher consolidation, and a drift-triggered skip protocol, retaining neither raw data nor model checkpoints. On four real Android benchmarks, FL-CL-IDS significantly outperforms fixed-λ EWC on the longer task sequences (CCCS-CIC-AndMal-2020, +2.1 pp ACC, paired p=0.026; and DREBIN, +2.4 pp, p=0.013) and is statistically tied on the shorter CICMalDroid 2020 and TUANDROMD, with no significant accuracy loss on any benchmark, while all continual methods improve on unregularized FL by up to 9 pp through reduced forgetting. Crucially, FL-CL-IDS attains this accuracy at 0.6× the client upload cost (40% fewer uploads), scales to 100 clients, and stays robust under heavy client dropout. An ablation identifies online Fisher consolidation as the principal driver of the accuracy gain, and an analytical on-device cost model shows the skip protocol yields up to ∼40% lower communication and energy at negligible compute overhead.
Indirect Time-of-Flight (iToF) cameras provide dense depth measurements for a wide range of sensing and imaging applications; however, their depth outputs are often degraded by noise, scattering effects, depth discontinuities, and unreliable measurements in low-reflectivity regions. Improving the quality and robustness of iToF depth data therefore remains an important challenge for their practical deployment. This paper presents a model-based depth enhancement framework that improves iToF sensor output quality by fusing the depth map (D) with the active near-infrared (NIR) image, both acquired by the same sensor, thus inherently aligned in space and time. Depth enhancement is formulated as an energy minimization problem that combines a confidence-weighted depth fidelity term, spatial regularization, and a joint infrared-depth (NIR-D) histogram-based conditional entropy term to exploit structural consistency between depth and NIR signals. The proposed formulation effectively suppresses different noise artifacts while preserving depth discontinuities and structural details. The method is learning-free and does not require external calibration or training, enabling a straightforward deployment across different iToF sensing configurations. Experimental results on both synthetic data, with noise characteristics modeled after real iToF sensors, and real-world iToF measurements demonstrate consistent improvements in depth accuracy and visual quality compared to classical depth enhancement and fusion baselines.
Graph Neural Networks (GNNs) are widely used in recommendation systems, but most methods focus solely on positive feedback, neglecting the avoidance preferences inherent in negative interactions. Although recent sign-aware methods have attempted to integrate bidirectional feedback, they suffer from two major limitations: 1) directly incorporating negative feedback into graph topological propagation causes the homophily assumption to fail, leading to feature contamination; 2) when dealing with long-tail distributions, the underlying aggregation mechanisms of traditional GNNs lack flexibility, making them prone to passively absorbing polarity noise during deep propagation, which further exacerbates feature contamination. To enhance model robustness in complex polarity scenarios, this paper proposes a Decoupled Adaptive Normalized Graph Convolutional Network (DANG). The framework is designed from three perspectives: structural propagation, layer-wise aggregation, and optimization objective. First, positive interactions are retained within the graph topology for collaborative propagation, while negative interactions are modeled independently using a Multilayer Perceptron (MLP). This structurally mitigates feature contamination caused by the mixing of positive and negative semantics. Second, parameterized normalization and tunable layer-wise pooling mechanisms are introduced to dynamically adjust neighbor aggregation strength and cross-layer information allocation, thereby reducing the dilution of personalized user representations caused by deep smoothing signals. Finally, a sign-aware Bayesian Personalized Ranking (sBPR) objective is constructed to more accurately capture users' preference and avoidance intentions. Experiments on three real-world datasets—Amazon-Music, Epinions, and KuaiRec—demonstrate that DANG outperforms 12 representative baseline models, including LightGCN and GFormer, in both Recall@20 and NDCG@20. Compared to the strongest baseline, DANG achieves an average improvement of 3.43% in Recall@20 and 2.42% in NDCG@20 across the three datasets. Triplicate experiments and ablation studies further indicate that the performance improvement of DANG is stable, and that decoupled modeling, adaptive normalization, and tunable layer-wise pooling all contribute positively to alleviating feature contamination and over-smoothing issues.
The performance of a battery is characterized by its parameters. One of the parameter is the state of health (SoH). In this work, SoH of lithium-ion batteries is being estimated using supervised learning algorithms. Linear Regression (LR), Decision Tree (DT), Support Vector Regressor (SVR), Random Forest (RF), and XGBoost (XGB) are some of the approaches are being taken into consideration. The performance of the algorithms has been evaluated on NASA (B0005 & B0006), and a synthetic datasets. Model performance has been checked for the metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), coefficient of determination (R²), Mean Absolute Percentage Error (MAPE), Median Absolute Error, and explained variance. It has been seen that out of all the models, the performance of Linear Regression proved to be consistent showcasing MAE of 0.0030 and 0.0033, RMSE below 0.0042, and R² exceeding 0.999 while performing on NASA datasets. Similarly for the synthetic dataset, Linear Regression retained the best R² of 0.802. From results it is obvious that while competing with complex models; simple linear models can also be reliable and computationally efficient for SoH estimation under diverse conditions. To prevent target leakage, the models use six cycle inputs only i.e. mean voltage, mean current, mean temperature, cycle duration, ambient temperature and cycle type. It is seen that Linear Regression maintains robust performance under leave-one-battery-out (LOBO) cross cell test achieving R² values around 0.991-0.993. It also achieves the R² of 0.79-0.95 under chronological data splits while other models exhibit degradation in performance including negative R² values in some cases. These findings demonstrate that a simple linear model can provide a reliable and efficient baseline for lithium-ion battery SoH estimation. However, these should be interpreted as benchmark performance on the evaluated datasets rather than universal generalization across different battery chemistries or operating conditions.
