Search Results (2,365)

Deep neural networks (DNNs) have shown outstanding performance in visual recognition tasks within vision sensor networks; however, they are still vulnerable to adversarial manipulations and imperceptible perturbations that can lead to erroneous predictions. To address that, this paper presents AEGIS, a semantic aware and uncertainty guided adversarial detection framework designed for robust image classification in vision sensors pipelines. At its core, a SemantiGAN module functions as a multi-class semantic discriminator, identifying and filtering visually inconsistent adversarial inputs before they propagate further in the pipeline. For inputs that pass this stage, a stochastic augmentation process generates test time variations, from which handcrafted instability metrics FlipScore, Prediction Inconsistency, Layerwise Cosine Similarity (early and mid layers), and Entropy are computed. These features are aggregated into a compact five dimensional vector and processed by an Evidential Deep Learning (EDL) classifier, which models output evidence using a Dirichlet distribution to yield both class predictions and calibrated uncertainty estimates. Evaluations on the Tiny ImageNet dataset across six categories clean, FGSM, PGD, patch-based, functional, and geometric attacks demonstrate the effectiveness of AEGIS. The proposed framework achieves an AUROC of 92.1%, an AUPRC of 90.2%, and an accuracy of 90.7%, outperforming conventional softmax-based detectors in terms of detection performance, robustness, interpretability, and uncertainty calibration. Full article

Deep neural networks (DNNs) have significantly advanced machine perception and reasoning; however, their lack of transparency in decision-making continues to pose a major challenge, particularly in high-stakes domains such as healthcare, finance, and law. This is especially concerning with the black-box nature of convolutional neural networks (CNNs), where the rationale for making a decision can be as important as the decision itself. This paper is driven by a question that is easier to ask than to answer: how can CNNs be made to explain themselves? To answer the question, we wrote a PRISMA-compliant systematic review of 154 studies published between 2017 and 2025. These studies were selected from 4421 studies retrieved through Web of Science, Scopus, IEEE Xplore, and ACM Digital Library. CNN-specific taxonomy was developed. This taxonomy organizes explainable artificial intelligence (XAI) methods on four axes: explanation timing, model dependency, output type, and target component. We found that there is a huge bias in the field regarding post hoc visual methods. Grad-CAM is the most widely cited visual explanation methodology, and within the model-agnostic framework, LIME and SHAP prevail. This research was also the first to analyze standard assessment methods. It was found that out of the 154 studies in the review, 98 used objective methods to evaluate fidelity, stability, or sensitivity. Conversely, fewer than ten of them used human-centered methods to evaluate how tasks were performed, how the users trusted the method, or how the users were prepared to interact with the system. We argue for a dual-reporting convention under which metrics should be reported together at least once, as per the family of metrics. The third contribution is an evidence-based challenge map, where we outline four issues: absence of standardized benchmarks, post hoc mechanism scalability limitations, vulnerability to adversarial perturbations, and the persistent gap between the technical descriptions and human understanding. For each challenge, we propose concrete directions: integrating causal reasoning, adopting participatory evaluation design, and building hybrid transparent architectures. We offer this review as a practical roadmap for researchers and practitioners working toward more explainable deep neural networks. Full article

Press-pack insulated-gate bipolar transistor (IGBT) devices are key components of modular multilevel converter (MMC)-based HVDC systems, yet the parallel connection of tens of chips inside one sealed housing makes their internal aging difficult to monitor—existing methods cannot determine which chip or thermal interface has degraded without invasive sensing. This paper proposes a two-layer aging condition monitoring framework based solely on non-invasive external case temperature measurements, comprising an online detection layer and a detection layer. For the diagnosis layer, a coupled thermal network state-space model of the press-pack module is established; its observability matrix is shown to be of full rank, proving that aging at any internal location is detectable from external case temperatures, and a steady-state sensitivity analysis reduces the required measurement to the module-center case-temperature pair and yields a failure-signature rule that discriminates five aging modes. A 3D electrothermal finite-element model, validated against fiber Bragg grating (FBG) measurements on an MMC sub-module with deviations below 1.2 °C at all load levels, is then used to build an aging fingerprint database of 270 scenarios that maps a measured case-temperature distribution to the aging location, the affected component, and the severity of any cooling-system degradation. For the detection layer, a deep neural network (DNN) trained only on healthy-state data provides online early warnings through its prediction residual. Experiments on a 2.2 kV/2100 A MMC sub-module platform and a solid-state DC–DC transformer platform confirm that cooling-system degradation is reliably detected and that the detection layer transfers across converter types. Even in the least sensitive aging case, the predicted case-temperature signature (0.036 °C) exceeds the 0.01 °C resolution of standard thermocouple instrumentation, which indicates that genuine device aging would be resolvable as well. The framework provides a non-invasive online health-monitoring solution that can also locate the degraded element in high-power press-pack devices. Full article

In wireless networks with severe blockage, path loss critically limits communication coverage. Reconfigurable Intelligent Surfaces (RIS) offer a promising remedy. However, the fine-grained control of massive reflecting elements incurs prohibitive computational overhead, which hinders real-time deployment. To address these challenges, this paper proposes a low-complexity scheme integrating RIS block selection with adaptive beamforming. The large-scale RIS is partitioned into multiple sub-arrays to enable block-wise phase control. By activating only those blocks with dominant channel gains, the system maximizes reflection gain while minimizing control overhead. To avoid the exponential complexity of exhaustive search, we develop a deep neural network (DNN)-based prediction architecture. By learning the mapping from channel states to optimal configurations, the DNN enables instantaneous selection of near-optimal RIS block combinations. Simulation results show that the proposed data-driven scheme achieves near-optimal bit error rate (BER) performance compared to exhaustive search. Notably, it avoids the exponential complexity growth typically associated with an increasing number of reflecting elements. The proposed mechanism extends reliable coverage range and improves link stability, offering an efficient solution for future wireless networks. Full article

This study presents a novel approach for model-based calibration transfer and drift compensation for metal oxide semiconductor (MOS) gas sensors. The sensors are lab-calibrated and different calibration models for each of the eight volatiles contained in the calibration are trained to allow for an interpretable quantification of individual volatiles in complex mixtures. Calibration transfer and drift compensation are used to compensate for domain shifts that particularly affect the model accuracy. Here, several domain shifts are considered, e.g., sensor-to-sensor variation among different production batches (calibration transfer) or time-related changes in sensor response like poisoning and aging (drift compensation). Such domain shifts can lead to a substantial performance degradation and are critical for reliable field deployment. Since interpretable and robust machine learning algorithms based on feature extraction, feature selection, and regression (FESR) are not inherently capable of model-based calibration transfer and drift compensation, recalibration typically requires time-consuming and labor-intensive laboratory calibration procedures. To address this challenge, a novel approach represents the interpretable FESR machine learning models as a deep neural network (IDNNRep), enabling the application of transfer learning techniques from the field of deep neural networks (DNNs). This allows the reuse of knowledge gained in an initial calibration domain and facilitates model transfer using only a small amount of new calibration data, thereby reducing calibration effort and time. The proposed method is evaluated across multiple gases, including acetone and toluene, for four domain-shift scenarios and compared with FESR models retrained exclusively on data from the new domain and orthogonal signal correction (OSC). The results demonstrate that the proposed approach reduces the root mean square error (RMSE) compared to the initial model, achieving values of 18.0–28.0 ppb (normalized RMSE: 6.3–9.3%) for both gases with only 0.1 of the calibration data, resulting in a reduction of up to 93% compared to the initial calibration model and 89% compared to the OSC. Furthermore, due to the interpretable nature of the underlying FESR structure, the calibration transfer enables additional sensor- and gas-specific insights. Full article

Near-fault pulse-like ground motions produce coupled intensity, duration, and period-matching effects, making nonlinear seismic assessment of multi-ribbed composite slab structures (MCSS) computationally expensive and difficult to generalize. To address this problem, a 4000-case OpenSees nonlinear time-history analysis (NLTHA) database is generated from Wenchuan ground motions through Latin hypercube sampling, and a mechanically informed feature-enhanced deep neural network (MIFE-DNN, previously denoted as PE-DNN in the first submission) is trained using equivalent stiffness, equivalent yield strength, mass proxy, demand-capacity ratios, period-matching ratio, normalized duration, and energy-capacity proxy; a validation-weighted stacked surrogate is further constructed from multi-seed MIFE-DNN and residual learners. On the independent test set, the mean $R^2$ increases from 0.9645 for the ordinary deep neural network (DNN) and 0.9686 for the single MIFE-DNN to 0.9782 for the stacked mechanically informed surrogate, while the maximum inter-story drift-ratio $R^2$ reaches 0.9541. Additional checks include 16 active-learning OpenSees enrichment cases, 12 analyses under two external near-fault records, 3 out-of-domain parameter cases, 100 cross-story tests, SHAP-based interpretation, and multi-EDP fragility post-processing. These checks show that the surrogate is reliable for interpolation and screening within the calibrated equivalent-model domain, but direct OpenSees recalculation is required for boundary, out-of-domain, and cross-configuration use. Parameter-importance, SHAP, and fragility analyses identify peak ground acceleration (PGA), pulse index, period matching, rib height, rib spacing, and damping ratio as dominant factors, indicating that mechanically informed feature-enhanced surrogate modeling provides an interpretable and efficient tool for MCSS response prediction and conditional fragility assessment within the sampled structural and ground-motion domain. Full article

With the rapid growth of electric vehicles (EVs), high charging demand has increased uneven station utilization and pressure on distribution networks. Therefore, this paper proposes a collaborative method for capacity planning and charging scheduling of multiple charging stations (CSs) considering time-of-use (TOU) pricing and energy storage systems (ESSs). A total system cost model is established, including transformer and ESS costs. Secondly, a deep neural network-guided improved sparrow search algorithm (DNN-ISSA) is proposed to optimize the number of chargers and parking spaces by predicting the initial capacity center. Furthermore, a charging scheduling algorithm is proposed to optimize user charging time by introducing a TOU price response function to modify charging probabilities. A case study of 36 CSs in Jinan shows that the proposed method reduces average charging time by 15.7, 15.4, and 15.2 min for 1000, 5000, and 10,000 demand points, while lowering the total system cost from 73.92 to 70.36 million yuan. The convergence value of DNN-ISSA reduces by 15.05%, 21.67%, and 11.61% compared with the improved sparrow search algorithm (ISSA), particle swarm optimization algorithm (PSO), and sparrow-particle swarm optimization algorithm (SSA-PSO), respectively. The proposed method enhances energy utilization, mitigates peak loads, and supports low-carbon EV charging operation. Full article

Short-term rental housing plays an important role in the housing market by increasing property utilization and generating income opportunities for property owners. This study investigates the key attributes associated with Airbnb occupancy rates using listing data from five major U.S. cities. Data mining and machine learning techniques, including Random Forest, XGBoost, Deep Neural Networks (DNN), and hierarchical cluster analysis, were applied to identify factors associated with occupancy rate variation. Random Forest achieved the best performance (R² = 23.92%). Feature importance analysis identified price, location (city), listing capacity, amenities, and host response rate as the variables most strongly associated with occupancy rates. Cluster analysis supported these findings by identifying a dominant group of moderately priced listings with smaller accommodation capacity, more amenities, higher host responsiveness, and an average occupancy rate of 57%. These findings provide data-driven insights that may help property owners optimize listing performance and improve occupancy rates. Full article

Long-term wildfire susceptibility mapping represents a significant component of disaster prevention and the protection of human communities, public health, and local ecosystems. In this study, a wildfire inventory was developed through multi-sensor fusion of satellite data (MODIS and VIIRS), comprising 153,305 fire events across China for the period 2001–2024. In addition to historical incidents, 14 predictive variables were processed, representing geomorphological, climatological, hydrological, vegetative, and anthropogenic conditions. This study evaluates long-term spatial wildfire susceptibility based on long-term mean environmental and climatic conditions. Methodologically, the research applies six models from machine learning (ML), deep learning (DL), and transformer-based approaches: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Deep Neural Network (DNN), Fourier Multi-Layer Perceptron (F-MLP), Kolmogorov–Arnold Network (KAN), and Feature Tokenizer (FT) Transformer. The results were integrated into an ensemble susceptibility map with a spatial resolution of 500 m using Geographic Information Systems (GIS), indicating that 7.4% of China’s territory is classified as having a very high wildfire susceptibility. In addition to the national-scale assessment, a local differentiation was conducted across 34 province-level divisions, revealing that Fujian Province (86.8%) and the Guangxi Zhuang Autonomous Region (82.9%) had the largest shares of areas classified as high and very high wildfire susceptibility. Performance evaluation under spatial block-based validation demonstrated that the Random Forest model achieved the highest predictive power, with an area under the curve (AUC) of 87.8%, followed by XGBoost (87.3%) and Fourier MLP (86.6%). Based on the combined SHAP (Shapley additive explanations) analysis of all applied models, soil moisture, elevation, and terrain slope were identified as the most influential factors affecting wildfire occurrence in China. Overall, the findings contribute to more effective wildfire prevention and risk management strategies at both the local and national levels. Full article

This paper presents a study on the development of a deep neural network (DNN)-based surrogate model for rapid induced voltage prediction in current transformer (CT)-based magnetic energy harvesting systems. CT-based magnetic energy harvesters are promising self-powered energy sources for low-power electronic devices, but their output performance is strongly affected by the nonlinear magnetic behavior of the core material. Therefore, the accurate prediction of the induced voltage is important for device design. However, evaluating the voltage response under various magnetic material characteristics and operating conditions through repeated electromagnetic simulations requires considerable computational effort. In this study, nonlinear B–H curves were parameterized using an arctangent-based model, and electromagnetic simulations were performed by varying the magnetic material parameters, primary current, and load resistance. The resulting dataset was used to train and validate the DNN surrogate model. The trained model showed high prediction accuracy, with an R² value greater than 0.99 and low prediction errors. It also reproduced the RMS induced voltage trends for different magnetic material characteristics and operating conditions and was further used for maximum power point analysis within the investigated parameter range. These results indicate that the proposed surrogate model can reduce the need for repeated electromagnetic simulations and support the efficient design exploration of CT-based magnetic energy harvesting systems. Full article

Background: The high costs, long development timelines, and low clinical success rates in oncology highlight an urgent need for reliable computational strategies for drug repositioning. Current machine learning approaches often integrate heterogeneous pharmacogenomic datasets, which may lose biological specificity and limit model interpretability. Methods: In this study, we propose Separate XAI, an explainable artificial intelligence framework that retains dataset-specific biological features by adopting separate preprocessing and training pipelines for the Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) datasets. Different deep learning architectures such as Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs) were used to predict the drug response in the cancer cell lines. We also used SHapley Additive exPlanations (SHAP) to improve interpretability and identify biologically relevant features. Results: The developed framework showed good predictions with 94.49% accuracy in the CCLE dataset and a mean squared error of 0.0725 in the GDSC dataset. Explainability analysis identified important biomarkers and signaling pathways such as TP53 and KRAS, providing mechanistic insights into drug sensitivity and therapeutic response. Conclusions: The distinct XAI presented here offers an interpretable, biologically grounded framework for cancer drug repositioning by integrating dataset-specific modeling and explainable artificial intelligence. However, integration-based approaches often suffer from confounding effects of experimental and biological heterogeneity, but the proposed framework explicitly preserves dataset-specific characteristics, which potentially could lead to more robust predictions and higher interpretability for precision oncology and translational cancer research. Full article

Lightweight, privacy-aware and adversarial robust intrusion detection is required for the proliferation of Internet of Things (IoT) devices. In the Industrial Internet of Things (IIoT), centralized detectors can be compromised by adversarial perturbations via gradient-based attacks, making them susceptible to raw traffic. We suggest Federated Learning-Adaptive Gated Recurrent Unit (FL-AdGRU), a Federated approach that combines a lightweight Gated Recurrent Unit (GRU) classifier with alternating adversarial fine-tuning on each client using FGSM and PGD, without any communication overhead. A two-stage resampling scheme (UCAS-SMOTE) reduces the class-imbalance ratio from 4081:1 to ≈4:1, followed by 61 features being reduced to 40 by a mutual-information selector (MI-SelectK). Under this scenario, FL-AdGRU achieves 99.9% accuracy and 0.999 weighted F1 (+6.5 p.p. over the federated DNN baseline), with no loss of accuracy when facing clean attacks, and boosts Fast Gradient Sign Method FGSM/Projected Gradient Descent (PGD) robustness by +19.3/+19.0 p.p. at the same level of $ϵ$ = 0.1, thus effectively balancing the accuracy–robustness trade-off. It is robust (97.8%/84.2% on UNSW-NB15) and generalizes well to UNSW-NB15, while decaying slowly in skeptical scenarios (≈99.9% weighted F1 for moderate skew, 93.9%/86.7% for severe). Assuring data-locality privacy through exchange of only model weights; defenses against inference attack are left for future work. FL-AdGRU, with a total communication of 43.8 MB (≈50× less than centralized training), is deployable on bandwidth-constrained IIoT networks. Full article

The Yili River Valley in Xinjiang is characterized by complex geological structures and frequent geological hazards, which seriously threaten local lives, property, and infrastructure. Improving the accuracy and interpretability of geological hazard assessment is therefore of great significance. To address this, nine factors, including elevation, distance from fault, and slope, were selected to construct a WOE-BPNN-SHAP coupled model. The weights of evidence (WOE) method was first used for factor correlation testing and to optimize the input of the BP neural network. The evaluation accuracies of WOE, WOE-DNN, and WOE-BP models were then compared, and the SHAP model was introduced to analyze the coupling relationships among factors. Results show that the WOE-BP model achieves the best predictive performance, with an AUC of 83.65%. Areas of extremely high-risk account for 8.63% of the study area, while higher-risk areas account for 15.39%. Elevation (1688–2847 m), distance from fault (<3000 m), precipitation (192.6–290.8 mm), and slope (>16°) are identified as the main driving factors. This coupled method provides a new technical approach for regional geological hazard assessment and offers a theoretical basis for disaster prevention, mitigation, and resilience building in the Yili River Valley. Full article

Bridge health monitoring and machine learning are increasingly intertwined for civil engineers and artificial intelligence experts. Bridges’ poor health can result in severe outcomes if not addressed in time. Therefore, continuous monitoring is required to detect any anomaly or damage. Sensors, such as accelerometers, inclinometers, thermistors, etc., can help actively monitor these bridges. The signals from these sensors help record physiological activities. Such activities are helpful for anomaly detection, damage localization, and bridge health predictions with the help of machine learning algorithms. The proposed method extracts features from the dynamic response of a bridge to ambient excitation. It focuses on processing the signal received from different accelerometers installed on a steel railway bridge to determine the location of the damage and the level of the damage predictions. Initially, features are extracted from time-series data; then, they are fed to a deep neural network after some pre-processing. Normal and augmented data are used with different parameter tuning for results. Original data is also subdivided, and the effect of data slicing on the predictions is investigated. The results show that one-fourth of the slicing of the original data gives the best results for training and testing accuracy with a deep neural network. The results show that the reduced matrix representation, particularly the 40 × 40 feature slicing, improved the classification performance for the predefined bridge scenario classes under the considered experimental settings. For bridge scenario classification, the best reported accuracy was 93.54%, while for damage intensity classification the best reported accuracy was 98.21%. In the DNN-based optimizer comparison, the Adam optimizer achieved higher and more stable performance than Stochastic Gradient Descent (SGD), with test accuracies of 92.3% and 93.7% compared with 75.2% and 86.4%, respectively. It is also observed that the Adam optimizer outperformed Stochastic Gradient Descent (SGD) in terms of both damage localization and damage intensity estimation. Full article

Background/Objectives: Ankle dorsiflexion plays a fundamental role in gait stability, impact absorption, and the stance-to-swing transition, and its impairment is a major limitation in transtibial prostheses. This study proposes and evaluates a lightweight two-stage pipeline for generating ankle-dorsiflexion references using a single shank-mounted inertial measurement unit (IMU). Methods: In the first stage, a deep neural network (DNN) estimates the shank pitch waveform from raw three-axis accelerations and angular velocities. In the second stage, the estimated shank pitch is transformed into an ankle-dorsiflexion waveform using a temporal mapping model. The approach was evaluated on a multisubject subset of the NONAN GaitPrint database comprising 35 healthy young adults, 598 walking trials, and approximately 122,468 gait cycles, using a strict subject-held-out protocol. Results: A feature-based Random Forest baseline showed limited performance, whereas the waveform-based DNN achieved high accuracy for shank pitch estimation, with test R² values up to 0.97. A conventional polynomial mapping between shank pitch and dorsiflexion yielded weak performance, whereas a temporal mapping model substantially improved the estimation of ankle dorsiflexion, with test R² values up to 0.85. The resulting ankle reference was integrated into a Gazebo/Robot Operating System 2 (ROS 2) simulation of a transtibial prosthesis, where the generated trajectories were executed in a software integration test under open-loop position control, confirming stable and consistent trajectory execution. Conclusions: These results indicate that combining accurate shank pitch estimation with temporal mapping enables feasible ankle-dorsiflexion reference generation from a single sensor in able-bodied gait, offering a preliminary, simulation-based pathway for single-sensor artificial intelligence (AI) pipelines in prosthetic development. The framework supports waveform-level feasibility, not clinical readiness or functional prosthetic control. Full article