Accurate detection of pediatric supracondylar humerus fractures remains a diagnostic challenge in emergency settings. To address this, we developed and validated a transfer-learning model, benchmarking its performance against board-certified emergency physicians. A retrospective dataset of 1440 anonymized pediatric elbow radiographs (720 fractures and 720 normal) was preprocessed using grayscale normalization, aspect-ratio-preserving resizing, adaptive denoising, and contrast-limited adaptive histogram equalization. Data were split at the patient level into training and validation (n = 1295) and a held-out test set (n = 145). A pretrained ResNet50 backbone was fine-tuned, and model stability was assessed via five-fold stratified cross-validation. Clinical applicability was evaluated in a blinded reader study where 20 emergency physicians interpreted 200 radiographs. On the held-out test set, the model achieved 90.34% accuracy, 93.5% sensitivity, and 86.8% specificity. Five-fold cross-validation yielded a mean accuracy of 91.72%, sensitivity of 92.8%, and specificity of 88.5%. In the reader study, emergency physicians achieved a mean accuracy of 87.6%, sensitivity of 90.5%, and specificity of 84.7%. Diagnostic performance differences between the model and physicians were not statistically significant (P > 0.05). The transfer-learning model achieved clinician-comparable performance, indicating its potential utility as an initial triage tool. However, the algorithm shares human diagnostic limitations, exhibiting lower sensitivity for nondisplaced Gartland type I fractures. Therefore, based on current data, it cannot be concluded that artificial intelligence implementation will eliminate the risk of missed fractures. Prospective multicenter validation is required before clinical deployment.

Accurate age estimation and gender classification from facial images remain challenging tasks owing to substantial variations in facial appearance caused by pose, illumination, expression, occlusion, image quality, and age-related morphological changes. Although transformer-based architectures have recently demonstrated remarkable performance in facial analysis, effectively exploiting their complementary representation capabilities within a unified multi-task framework remains an open research problem. To address this limitation, this study proposes a transformer-based Multi-task Learning (MTL) framework that jointly performs age estimation and gender classification using a shared feature extraction backbone and task-specific prediction heads. Three state-of-the-art vision transformers, namely Vision Transformer (ViT), Shifted Window Transformer (Swin Transformer), and Data-efficient Image Transformer (DeiT), were independently implemented within the proposed MTL architecture to learn both global contextual information and local facial characteristics. Furthermore, two ensemble learning strategies, namely standard ensemble and Task-aware Weighted Ensemble (TAWE), were developed to exploit the complementary strengths of the individual transformer models. Experiments were conducted on the publicly available UTKFace dataset comprising 15,503 facial images, which were divided into 10,385 training, 1833 validation, and 3285 test samples, while gender balancing was applied during model training to alleviate class imbalance. Age estimation performance was evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), whereas gender classification performance was assessed using Accuracy and F1-score. Experimental results demonstrate that the proposed TAWE achieved the best point estimates among the evaluated ensemble strategies, obtaining an MAE of 4.8905, an RMSE of 6.4674, a gender classification accuracy of 94.86%, and an F1-score of 95.17% for the dataset-defined positive class. Compared with equal-weight averaging, TAWE provided a statistically significant improvement in gender classification accuracy, while the improvement in age estimation was modest and not statistically significant.

Seismic exploration has progressively expanded into urban fringe regions and areas with intensive human activities, where anthropogenic interference has increased markedly, aggravating the background noise problem and substantially affecting the accuracy of subsequent seismic data processing. Conventional denoising methods struggle to cope with such complex noise patterns and typically rely on manually designed parameter settings, which limits their ability to suppress complex background noise in real seismic data. With the rapid development of deep learning, various neural-network-based methods have been introduced for seismic data denoising. However, existing deep learning approaches are constrained by the high resolution of seismic data and limited computational resources and therefore commonly perform denoising on cropped data patches rather than on the complete seismic section. This limitation weakens the network's capability to perceive global contextual information, degrading denoising performance and potentially introducing blocking artifacts. To address these issues, we propose a fluctuation-prior-regularized denoising framework that explicitly decomposes the complete seismic data denoising task into global and local denoising subtasks, enabling globally consistent denoising under complex conditions. To overcome the limitations of CNNs and Transformers, we adopt Retentive Networks Meet Vision Transformers (RMT) as the backbone for feature extraction in this work. And a fluctuation-prior-constrained local denoising mechanism is introduced, allowing local patches to indirectly capture global information. In addition, a grouped regularization strategy for global and local tasks is proposed, enabling both optimization tasks to better capture their respective task-specific characteristics. Experimental results on noisy shot gathers constructed from field records and on field-recorded background noise demonstrate that the proposed DFRSeisnet achieves superior denoising performance while effectively alleviating blocking artifacts.

Automatic modulation recognition (AMR) is a key enabling technique for intelligent spectrum sensing, non-cooperative wireless signal analysis, and communication monitoring. However, its reliability degrades significantly under low signal-to-noise ratio (SNR) conditions, where noise weakens local I/Q waveform patterns, blurs time-frequency structures, and limits the robustness of single-domain representations. Although multimodal AMR methods have improved recognition by combining I/Q and time-frequency representations, many of them still use fixed fusion rules, which may not reflect the changing reliability of different modalities across SNR levels. In addition, most existing backbones model signal sequences in a Euclidean manner and lack an explicit mechanism to capture adaptive relational structures among signal tokens under strong noise. To address these limitations, this paper proposes MPCG-Net, a modulation-prior-enhanced cross-modal graph learning framework for low-SNR AMR. MPCG-Net first performs modulation-prior-enhanced cross-modal encoding, where learnable multi-resolution spectro-temporal representation learning enhances the input representation and training-stage SNR-aware cross-modal consistency alignment improves the complementary interaction between I/Q and spectro-temporal tokens. A modulation context memory module then refines the fused tokens before graph construction. In the graph-based relational learning stage, a temporal similarity graph with self-loops, local temporal-chain edges, and top-k feature-similarity edges is built to learn more discriminative graph-level signal representations. Experiments on RadioML2016.10A, RadioML2016.10B, and RML22 validate the effectiveness of the proposed design. Compared with the strongest baseline on each dataset, MPCG-Net improves the average accuracy in the -20 to 0 dB low-SNR range by 6.90, 6.45, and 1.66 percentage points on RadioML2016.10A, RadioML2016.10B, and RML22, respectively, while maintaining a lightweight model scale.

To address the challenges of low localization accuracy for insulator defects in complex inspection scenarios and excessive model redundancy that hinders edge deployment, this paper proposes FS-YOLO, a lightweight defect detection algorithm. First, a C3k2 ConvFormer Gated Linear Unit (C3k2-CFGLU) module replaces the original backbone, using depthwise separable convolution and gated linear units to enhance defect feature representation while reducing computation. Second, a Multi-Branch Multi-Scale Feature Pyramid Network (MBMSFPN) neck improves defect target perception via multi branch auxiliary connections that strengthen high-level and low-level feature interaction. Third, a Shared Convolution Detection Head (SCDH) reduces parameters through multi-scale shared convolutions and group normalization, and incorporates a localization quality estimation mechanism to offset lightweight induced accuracy loss. Finally, layer adaptive sparsity for magnitude-based pruning (LAMP) removes redundant channels, compressing model size and improving efficiency. Experiments show that FS-YOLO achieves 93.0% precision, 88.3% recall, 92.9% mAP@0.5, and 62.9% mAP@0.5:0.95, with only 0.88M parameters and 4.1 GFLOPs. Compared with the baseline, parameters and computation drop by 65.9% and 34.9%, while mAP@0.5 increases by 5.7%. Against other mainstream YOLO variants, FS-YOLO offers superior accuracy efficiency trade offs, offering a promising reference for the intelligent development of power line inspection.

The Industrial Internet of Things (IIoT) is being used a lot in important areas like advanced manufacturing, smart energy systems, and intelligent cities. Watching for intrusion detection is very important for the safety of the IIoT. Nevertheless, multivariate sensor sequences frequently demonstrate pronounced physical coupling, non-stationarity, and periodicity concurrently, rendering it prone for detection models grounded in statistical correlation to erroneously classify normal collaborative variations as anomalies. Moreover, prevailing methods predominantly concentrate on single-domain representations within either the time or frequency domain, posing challenges in addressing both burst and periodic attacks concurrently. To address these challenges, this article introduces the Time-Frequency Dual-Branch Fusion and PCMCI-Based Association-Guided Network (TFPAG-Net) for IIoT Intrusion Detection. This model initially constructs a lightweight temporal convolutional network backbone employing depthwise separable convolutions. Subsequently, parallel branches in the time and frequency domains are established to respectively model local abrupt changes, long-range dependencies, and periodic spectral structures, with time-frequency feature fusion facilitated through a sample-dependent gating mechanism. Building on this, multi-scale temporal pyramids are employed to amalgamate fine, intermediate, and coarse-scale information. Furthermore, as an auxiliary refinement, a lagged conditional-dependence prior estimated from the training data via PCMCI is projected into a bounded attention bias to provide supplementary guidance for channel feature reweighting. Evaluations on Edge-IIoTset, X-IIoTID, and SWaT yield mean Macro-F1 scores of 0.9886, 0.9466, and 0.9607, respectively, over five predefined random seeds. Under the unified training protocol, TFPAG-Net ranks second on Edge-IIoTset and achieves the highest mean Macro-F1 on X-IIoTID and SWaT. Ablation experiments show dataset-dependent effects of the proposed components. On Edge-IIoTset and X-IIoTID, the final attention-stage improvement reflects the joint effect of SE-based modulation and the PCMCI-derived association prior. Accordingly, PCMCI is treated as an auxiliary association refinement rather than a principal contribution of TFPAG-Net. Additionally, TFPAG-Net maintains a moderate computational footprint, with approximately 0.29 M parameters, providing a favorable balance between model complexity and intrusion detection performance.

Ultrasonic guided wave (UGW) inspection of small-diameter pipes is complicated by the narrow usable frequency band and by strong susceptibility to measurement noise, both of which degrade the reliability of damage identification. This study proposes a multi-channel attention Mamba-2 (MCA-Mamba-2) network that couples a convolutional front end with a bidirectional Mamba-2 state-space backbone and combines multi-head self-attention with channel attention for feature fusion. The network takes as input the four waveforms recorded simultaneously by a circumferential receiver array so that the input tensor has four physical channels that are subsequently expanded into learned feature channels. A dataset of ten damage states, spanning non-through-wall abrasions, through-wall holes and one multi-defect case, was acquired on a 20 mm outer-diameter steel pipe specimen. Averaged over three independent runs, MCA-Mamba-2 reaches 98.01 ± 0.85% accuracy on noise-free signals and 81.09 ± 1.74%, 83.65 ± 1.34%, 82.49 ± 0.20% and 84.15 ± 1.04% on 0 dB test sets contaminated with white, pink, Gaussian and Laplacian noise respectively, exceeding the Mamba-2 baseline in all sixteen tested noise-type and SNR combinations, by up to 17.87 percentage points. An ablation study attributes the improvement principally to the convolutional mixing block. A cross-noise evaluation reveals that accuracy drops by 59.80 percentage points, on average, at 0 dB when the training and testing noise distributions differ, indicating that the matched-noise accuracies are not indicative of robustness to unknown noise conditions; however, a single model trained on a mixture of all noise types and SNR levels recovers an average accuracy of 91.21% across the sixteen test conditions. Because each damage state was recorded in a single acquisition session, the dataset was partitioned at the level of individual records rather than by session, and the reported accuracies should be interpreted as an upper bound on the performance attainable on independently acquired data. The results were obtained on a single pipe geometry with fixed defect positions and synthetically added noise; extension to other geometries and to field measurements remains to be demonstrated.

Intelligent fault diagnosis of rolling-element bearings is essential for predictive maintenance of rotating machinery. Although deep convolutional neural networks (CNNs) achieve high accuracy on public benchmarks, their large parameter counts and computational costs hinder deployment on resource-constrained edge devices. This paper proposes a frequency-aware lightweight CNN (FAL-CNN) that reconciles competitive accuracy with extreme efficiency. A fixed, non-trainable pipeline (fast Fourier transform, log compression, linear resampling, and min-max normalization) converts each 1024-sample vibration segment into a compact 128-dimensional spectrum, reducing the input eightfold. A frequency-aware module applies three parallel one-dimensional convolutions with kernels of 3, 7, and 15, matched to single-peak, harmonic-pair, and harmonic-family fault signatures, followed by a depthwise-separable backbone that sharply reduces parameters and floating-point operations (FLOPs). All Case Western Reserve University (CWRU) experiments use a leakage-controlled protocol with time-blocked splits, a sample-isolation gap, and five random seeds. Across four operating loads, FAL-CNN reaches a 99.78% mean accuracy with 65,122 parameters and 3.88 M multiply-accumulate operations (MACs) (7.76 M FLOPs), inferring in 0.17 ms on a single CPU thread (0.27 ms end-to-end) within a 288.8 KB Open Neural Network Exchange (ONNX) footprint. Because full-data accuracy saturates for all compact models, we evaluate where architecture matters: under operating-load shift, FAL-CNN outperforms a parameter-matched plain CNN by 13.27 percentage points (88.42% vs. 75.15%); frequency-masking and attribution analyses confirm that the multi-scale filters concentrate on fault frequency bands (a 21.91-percentage-point targeted-masking gap, roughly twice that of structural controls); under 10:1 training class imbalance, it maintains a macro-F1 above 99.2%; and on the Paderborn real damage benchmark with bearing-level nested validation, it attains (59.86 ± 18.86)% accuracy on unseen bearings, the best mean among the compared pipelines, although the difference from the parameter-matched baseline is not statistically significant. These results position FAL-CNN as a compact, edge-oriented classifier whose measured feasibility is currently limited to the development CPU; physical edge-device profiling remains required. Its demonstrated value lies in reliability under the evaluated distribution shifts rather than in saturated benchmark accuracy.

(1) Background: To address the challenges in ground-penetrating radar (GPR) B-scan profiles-where targets typically exhibit hyperbolic signatures, small scales, low contrast, and severe background clutter-this paper proposes an efficient and accurate target detection algorithm based on YOLOv8n, termed YOLOv8n-PCPG. (2) Methods: First, a Pinwheel-shaped Convolution (PSConv) is integrated into the backbone to capture directional hyperbolic boundaries. Second, a lightweight CSP-PMSFA module replaces the standard C2f structures in the neck, reducing computational redundancy while preserving feature integrity. Finally, a structurally reparameterized GC-downsampling module is introduced to enhance cross-channel feature interaction without adding inference latency. (3) Extensive experiments on a self-constructed GPR B-scan dataset show that the proposed model achieves a Precision of 97.2%, a Recall of 97.3%, an mAP@0.5 of 98.8%, and an mAP@0.75 of 84.3%. Compared with the baseline YOLOv8n, the parameter count and GFLOPs are reduced by 13.0% and 7.4%, respectively, while achieving a real-time detection speed of 155.78 FPS, outperforming existing mainstream detectors. (4) Conclusions: These results demonstrate that YOLOv8n-PCPG achieves a strong trade-off between detection accuracy and computational efficiency, providing effective technical support for subsurface hyperbolic signature identification.

Unmanned aerial vehicle (UAV) acquired imagery provides a flexible non-contact visual sensing modality for monitoring inland waterway locks, yet reliable perception remains challenging due to significant scale variation among targets, as well as partial visibility, low contrast, water-surface texture variation, and complex backgrounds. To address these issues, this study proposes BISRF-Net, a baseline-preserving scale-guided residual feature routing network built upon the CBv2 Faster R-CNN framework. The method retains the original backbone, feature pyramid network, and region proposal network, while introducing scale-guided residual routing at the region-of-interest (ROI) refinement stage. Through identity-preserved residual addition, the baseline ROI representation is preserved and enhanced with complementary scale-sensitive information without altering the original feature pathway. Experiments conducted on the UAV image subset of the TROUT lock-monitoring dataset demonstrate that BISRF-Net maintains comparable overall detection performance while improving medium-scale target representation. Repeated trials with different random seeds further assess the robustness of the proposed method and indicate that its main benefit lies in medium-scale target refinement. Ablation and computational analyses further show that the proposed design enables targeted ROI-level refinement with additional computational cost. These findings highlight the potential of scale-guided ROI feature refinement for robust UAV-based visual sensing in complex inland waterway environments.