ABSTRACT Multi-variety and small-batch production is widely used in modern manufacturing. Monitoring the machining process in such systems is crucial for improving efficiency and reducing costs. However, most existing monitoring methods rely on intrusive sensors which often require equipment modifications and increase costs. Additionally, many methods can only determine the machining state after all operations on a machine are completed, making real-time monitoring difficult. The small sample size characteristic also limits the effectiveness of conventional machine learning approaches.Therefore, a real-time machining state monitoring method based on transfer learning and power signal is proposed. The input power of the machine tool is used as the original data, whose acquisition is non-intrusive. The power signal is segmented using the Bayesian online change point detection algorithm and converted into recurrence plots for state monitoring model. A transfer learning–based model is then developed to identify machining steps and processes using only a small amount of training data. Power signal features are extracted using frequency domain analysis and recurrence quantification analysis. Finally, machining anomalies are detected using a Z-score variant algorithm. Case studies show the proposed method achieves a monitoring accuracy of 97.1%, enabling effective real-time state monitoring in multi-variety and small-batch production systems.

Cross-site transfer learning is a promising approach to address data scarcity at newly deployed photovoltaic (PV) stations by leveraging knowledge from data-rich source sites. However, existing domain adaptation methods align feature representations without distinguishing physically meaningful causal relationships from site-specific spurious correlations, leading to negative transfer when local environmental conditions differ substantially between stations. This paper proposes CausalTransPV, a causal invariant representation learning framework that integrates explicit temporal causal discovery with selective domain alignment for cross-site PV power forecasting. The framework comprises three synergistic modules: (i) a multi-station temporal causal discovery module that jointly learns shared and station-specific causal graphs through differentiable acyclicity-constrained optimization with a cross-station invariance regularizer; (ii) a causal-guided disentangled encoder that decomposes representations into causal-invariant and site-specific subspaces using the discovered causal graph as a structural prior; and (iii) a causal-subspace transfer and prediction module that performs maximum mean discrepancy (MMD)-based domain alignment exclusively on the causal subspace. Experiments on the Desert Knowledge Australia Solar Centre (DKASC) multi-station dataset under varying target label ratios (0–50%) demonstrate that CausalTransPV achieves relative mean absolute error (MAE) reductions of 6.9–9.9% over the strongest baseline. Ablation studies, causal graph analysis, feature space visualization, and weather-conditioned case studies further validate the contribution of each component. These results suggest that causal-guided selective transfer offers an effective paradigm for reliable PV forecasting under data-scarce cross-site scenarios.

Skin cancer is one of the most prevalent and potentially lethal diseases worldwide, with early detection being critical for patient survival. This study presents a novel framework that leverages transfer learning, pruning, SMOTE, data augmentation, and the advanced Avg-TopK pooling method to improve the accuracy and efficiency of skin cancer classification using dermoscopic images. The HAM10000 dataset was used to evaluate the performance of various transfer learning models, with Xception as the top performer. A layer-based pruning strategy was proposed to optimize the model and reduce its complexity. SMOTE and data augmentation were applied to address the class imbalance within the dataset, significantly improving the model’s generalization across all skin lesion classes. The utilization of the Avg-TopK pooling technique further enhanced model accuracy by preserving crucial image features during the downsampling process. The proposed approach achieved an overall accuracy of 91.52%, surpassing several state-of-the-art models. Following pruning, the model’s parameter count was reduced by approximately 35%, from 20.9 million to 13.5 million, improving efficiency and performance. This framework demonstrates the effectiveness of combining model pruning, oversampling, and advanced pooling methods to build robust and efficient skin cancer classification systems suitable for clinical applications.

Brain gliomas represent one of the most aggressive cancers worldwide and remain difficult to diagnose accurately at an early stage. Although computer-aided diagnostic (CAD) approaches have progressed notably in recent years, distinguishing between high-grade glioma (HG-G), low-grade glioma (LG-G), and healthy brain tissue on magnetic resonance images is still a major challenge. To address this issue, we propose a multi-stage framework designed to push the boundaries of current classification methods. The framework begins with a preprocessing phase that integrates Adaptive Gamma Correction (AGC) for improved contrast adjustment with a Denoising Convolutional Neural Network (DnCNN) for noise removal. Feature extraction is then carried out from three representative layers across three fine-tuned transfer learning CNNs (TRCNNs), where each model is optimized by a different algorithm. These deep representations are combined with handcrafted texture measures based on the Gray-Level Co-occurrence Matrix (GLCM), producing nine unique CNN-GLCM Fused Feature (CGFF) sets. The resulting hybrid descriptors are evaluated using several strong classifiers such as Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Support Vector Machine (SVM), along with a stacked ensemble to reinforce stability and robustness. Performance significance was verified through the Friedman statistical test, with p < 0.05, confirming the reliability of the improvements. The framework achieved 99.05% accuracy, 98.99% recall, 99.52% specificity, 99.08% positive predictive value (PPV), and 99.54% negative predictive value (NPV), consistently surpassed state-of-the-art (SOTA) methods across all reported metrics.

Accurate prediction of oil formation volume factor remains a critical challenge in reservoir fluid characterization, with existing approaches limited by either a lack of interpretability in black-box machine learning models or the mathematical complexity of white-box alternatives. This study introduces a Transformer-based regression model trained on 442 data points to predict the oil formation volume factor at and below the bubble point pressure. Symbolic regression was subsequently applied to derive a simplified, interpretable mathematical correlation termed the Transformer correlation. Transfer learning was employed to assess the model’s adaptability to an independent dataset comprising 270 datapoints from different geographic region. The Transformer model achieved a coefficient of determination of 0.9902 and a root mean square error of 0.0207, demonstrating strong generalization to unseen data while reproducing the physical behavior of the target property. The Transformer correlation surpassed several well-established empirical correlations, recording a coefficient of determination of 0.9675 and offering a transparent, field deployable alternative. Transfer learning further validated and scalability of proposed, attaining a coefficient of determination of 0.9814 and a root mean square error of 0.0303. The combination of Transformer-based modeling, symbolic regression, and transfer learning presents a novel, interpretable, and adaptable framework for predicting reservoir fluid properties.

ABSTRACT In April 2024, the United Arab Emirates (UAE) experienced its heaviest rainfall in 75 years, which triggered widespread urban flooding and prolonged disruptions. This study assesses the spatial impact and post-rainfall recovery of the event using the U-Net deep learning architecture and daily PlanetScope satellite imagery captured on a pre-rainfall date and on a series of post-rainfall dates between 14 April 2024 and 27 April 2024. Multi-temporal land use and land cover (LULC) classifications were generated using a U-Net model trained via transfer learning, and the LULC categories were water, vegetation, built area, and bare ground. The multi-temporal LULC classifications were further used for change detection analyses to map flooded areas and track temporal recovery across LULC categories. The U-Net model for LULC classification achieved over 95% overall accuracy and a Kappa statistic of 0.927. The change detection and inundation recovery showed that approximately 23.8 km2 – nearly ten times the area of Downtown Dubai – was flooded. The flood affected the built area and bare ground most, while vegetation exhibited higher flood resilience. Although rainfall ended by 17 April, 95% of flooded areas remained submerged three days later, and 37% were still underwater by day 10. These findings reveal the limitations of urban drainage systems in Dubai and the value of high-temporal-resolution remote sensing and deep learning for flood monitoring. This study offers a practical and replicable framework to support urban flood risk assessment, resilience planning, and climate adaptation in the Persian Gulf region, characterized by rapid urbanization and fragile arid environments similar to Dubai.

Integrating single-cell atlases and machine learning to construct 'in silico patients' for predicting individualized drug responses.

Cassava cultivation is important to the nation because of the products and raw materials that are supplied to industries. However, Its cultivation is affected by diseases such as cassava brown leaf and cassava mosaic, which have been classified using a Support Vector Machine (SVM). SqueezeNet pretrained Convolutional Neural Network (SPCNN) is adjudged a very effective image detection and classification algorithm with limited usage areas. This research, therefore was used to detect and classify cassava leaf diseases using SPCNN and SPCNN with multiclass SVM. Performance evaluation of both techniques showed that SPCNN effectively detected and classified cassava leaf diseases with optimal accuracies of 92.22% and 98.89% for SPCNN and SPCNN-SVM, respectively. Keywords- SqueezeNet, SVM, Cassava diseases, Feature extraction, Transfer learning Arinola I. O., Oke O. A. &amp; Falohun A. S. (2026): Detection and Classification of Cassava Leaf Diseases using Squeezenet Pretrained Convolutional Neural Network and Support Vector Machine. Journal of Advances in Mathematical &amp; Computational Science. Vol. 14, No. 1. Pp 1-12 Available online at www.isteams.net/mathematics-computationaljournal. dx.doi.org/10.22624/AIMS/MATHS/V14N1P1

Autism Spectrum Disorder (ASD) and Attention Deficit and Hyperactivity Disorder (ADHD) are two psychiatric disorders frequently encountered in children. ADHD is further categorized into three subtypes. The diagnostic processes for these conditions are complex and often prone to misclassification. We proposed a lightweight deep neural network, ADBrainNet, to differentiate ASD, ADHD combined, ADHD hyperactive/impulsive, ADHD inattentive and neurotypical individuals. Our methodology was benchmarked against prevalent ImageNet transfer learning methods, including AlexNet, MobileNet, ResNet18, and Xception, for training on resting-state fMRI images sourced from ABIDE and ADHD-200 datasets. ADBrainNet achieved superior performance on the independent external testing set through five-fold cross-validation, with a mean (± standard deviation) accuracy, precision, recall, and F1 score of 61.87% (± 5.59%), 65.72% (± 6.98%), 61.87% (± 5.59%), and 62.50% (± 5.78%), respectively. Furthermore, the explainable artificial intelligence algorithm LIME was employed to explore the most significant features during ADBrainNet's decision process. Our model provides an interpretable computational framework for neuroimaging-based classification between ASD and ADHD subtypes. This approach may inform future research and, upon further validation and comparison with clinician performance, could potentially aid in patient assessment, stratification, and management of psychiatric disorders.

Transferable deep learning models for the estimation of daily potential evapotranspiration across altitudinal forest gradients in the Mediterranean

The emergence of the Internet of Things (IoT) has introduced major security challenges. Deep learning models have shown strong potential for intrusion detection. However, they often require large datasets and high computational resources. In contrast, IoT environments are resource-constrained and lack sufficient labeled data. This paper proposes a lightweight intrusion detection approach based on Long Short-Term Memory (LSTM) networks and homogeneous transfer deep learning. The model is first trained on a subset of the BoT-IoT dataset as a source domain. It is then fine-tuned on a disjoint subset containing a rare attack type. This setup represents adaptation to unseen attack behaviors within the same environment. By freezing earlier layers and fine-tuning only the final layers, the method reduces training overhead while preserving performance. This is important to meet the IoT requirement for frequent, lightweight model updates on resource-constrained devices. The proposed model achieved 99.9% accuracy, a macro F1-score of 0.96, and a 47.8% reduction in training time compared to training from scratch. Extensive experiments confirm that it maintains balanced detection across both common and rare classes.

Accurate interpretation of chest radiographs is central to the early diagnosis and management of pulmonary disorders. This study introduces an explainable deep learning framework that integrates biomedical signal fidelity analysis with transfer learning to enhance diagnostic reliability and transparency. Using the publicly available COVID-19 Radiography Dataset (21,165 chest X-ray images across four classes: COVID-19, Viral Pneumonia, Lung Opacity, and Normal), three architectures, namely baseline Convolutional Neural Network (CNN), ResNet-50, and EfficientNetB3, were trained and evaluated under varied class-balancing and hyperparameter configurations. Signal preservation was quantitatively verified using the Structural Similarity Index Measure (SSIM = 0.93 ± 0.02), ensuring that preprocessing retained key diagnostic features. Among all models, ResNet-50 achieved the highest classification accuracy (93.7%) and macro-AUC = 0.97 (class-balanced), whereas EfficientNetB3 demonstrated superior generalization with reduced parameter overhead. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed anatomically coherent activations aligned with pathological lung regions, substantiating clinical interpretability. The integration of signal fidelity metrics with explainable deep learning presents a reproducible and computationally efficient framework for medical image analysis. These findings highlight the potential of signal-aware transfer learning to support reliable, transparent, and resource-efficient diagnostic decision-making in radiology and other imaging-based medical domains.

Transfer learning (TL) is used to extrapolate the physics information encoded in a generative adversarial network trained on synthetic neutrino-carbon inclusive scattering data to related processes such as neutrino-argon and antineutrino-carbon interactions. We investigate how much of the underlying lepton-nucleus dynamics is shared across different targets and processes. We also assess the effectiveness of TL when training data is obtained from a different neutrino-nucleus interaction model. Our results show that TL not only reproduces key features of lepton kinematics, including the quasielastic and Δ -resonance peaks, but also significantly outperforms generative models trained from scratch. Using data sets of 10,000 and 100,000 events, we find that TL maintains high accuracy even with limited statistics. Our findings demonstrate that TL provides a well-motivated and efficient framework for modeling (anti)neutrino-nucleus interactions and for constructing next-generation neutrino-scattering event generators, particularly valuable when experimental data are sparse.

INTRODUCTION. Convolutional neural networks (CNNs) show strong promise for automating dental diagnosis from radiographic images. Robust head-to-head comparisons across tasks and datasets are needed to guide model selection for clinical use. MATERIALS AND METHODS. We compared three lightweight CNNs – EfficientNet-B0, ResNet-18, and MobileNetV3 – for two classification tasks: enamel caries on intraoral images and periapical lesions on panoramic radiographs. Data comprised the Caries-Spectra dataset (2,000 intraoral images; advanced, earlystage, and no caries) and a panoramic radiograph set with 13,071 images labeled with periapical lesion scores (PAI 3–5). For the periapical task, data augmentation was applied to the panoramic training split only, increasing its size to 17,004 training instances; validation and test splits (as well as all Caries-Spectra splits) remained at their original sizes. Models were trained via transfer learning with early stopping and evaluated using accuracy, precision, recall, F1-score, and confusion matrices. RESULTS. EfficientNet-B0 achieved the best overall performance on both tasks, reaching 99.74% accuracy for caries detection and 69.65% accuracy for periapical lesion detection, outperforming ResNet18 and MobileNetV3 across the reported metrics. CONCLUSIONS. Lightweight CNNs – particularly EfficientNet-B0 – are effective for dental image classification and are suitable candidates for integration into clinical diagnostic workflows. Model architecture choice and data quality materially influence performance

The Greater Caribbean manatee faces significant conservation challenges due to a lack of demographic data in low-visibility habitats. To address this, we present a refined automated manatee counting method pipeline integrating deep learning-based call detection with unsupervised individual counting. We resolved significant computational bottlenecks by implementing an offline feature extraction strategy, bypassing a 13-h processing lag for 43,031 audio samples. To mitigate overfitting in imbalanced bioacoustic datasets, non-parametric bootstrap resampling was employed to generate 100,000 balanced spectrograms. Benchmarking revealed that transfer learning via a VGG-16 backbone achieved a mean 10-fold cross-validation accuracy of 98.92% (±0.08%) and an F1-score of 98.08% for genuine vocalizations. Following detection, individual counting utilized k-means clustering on prioritized music information retrieval descriptors—spectral bandwidth, centroid, and roll-off—to resolve distinct acoustic signatures. This framework identified three individuals with a silhouette coefficient of 79.20%, demonstrating superior cohesion over previous benchmarks. These results confirm the automatic manatee count method as a robust, scalable framework for generating the scientific evidence required for regional conservation policies.

Abstract Cervical cancer remains a major global health challenge, particularly in low-resource settings where access to timely and accurate diagnostic services is limited. Although Pap smear screening has significantly reduced cervical cancer mortality, manual cytological assessment is labor-intensive, prone to inter-observer variability, and dependent on expert availability. To address these limitations, this study presents a hybrid classification framework that integrates deep feature extraction with interpretable rule-based learning for cervical cancer diagnosis from Pap smear images. High-level visual features are extracted using a pre-trained InceptionV3 convolutional neural network, while classification is performed using Convolutional Tsetlin Machines (CTMs), which employ logical clauses learned through Tsetlin automata to enable transparent decision-making. The proposed framework was evaluated on a publicly available liquid-based cytology dataset comprising four diagnostic classes: Negative for Intraepithelial Malignancy, Low Squamous Intraepithelial Lesion, High Squamous Intraepithelial Lesion, and Squamous Cell Carcinoma. Experimental results obtained under stratified cross-validation demonstrate strong and consistent performance, achieving an accuracy of 99.96%, precision of 98.99%, recall of 98.96%, and an F1-score of 98.98%. Beyond high classification accuracy, the intrinsic interpretability of CTMs allows inspection of learned logical rules, supporting clinical transparency and trust. These findings indicate that combining deep transfer learning with interpretable machine learning offers a promising direction for reliable and explainable cervical cancer screening systems, particularly in resource-constrained healthcare environments.

A transfer learning modeling approach for three-axis instantaneous milling forces based on theoretical model priors

Human Activity Recognition (HAR) is becoming a hot topic for research at the intersection of artificial intelligence, computer vision, and sensor analysis. It analyses and classifies different human behaviors from diverse data inputs in an automated way. This paper provides a thorough study of Human Activity Recognition (HAR), following its development from early machine learning models which were handcrafted feature-based systems to modern deep learning models which process multimodal inputs. A number of methods such as wearable sensor-based techniques, vision-based approaches, radar and non-contact sensing, transfer learning techniques and domain adaptation are surveyed. An overview of real-world applications which include healthcare, sports, surveillance, and human-computer interaction is provided. These real-world applications reveal HAR’s visible impact on society. Lastly, we discuss important future challenges such as robustness and generalizability, explainability and interpretability, multi-activity and complex behavior recognition, privacy concerns, and real-world &amp; open-world recognition. This overview focuses on the current advancements and recognizes open research directions mandatory for reliable, interpretable, and ethically responsible HAR systems

Introduction This article proposes a fault diagnosis algorithm for mechanical rolling bearings based on transfer learning. Methods The proposed algorithm enhances the traditional conventional convolutional neural network (CNN) algorithm by introducing a domain category judgment module and an inter-domain conditional probability distribution difference module, thereby achieving transfer learning between source domain samples and target domain samples. Simulation experiments were performed. On a PT100 bearing fault simulation test platform, vibration signals of bearings were collected in cases of normal operation, inner race faults, outer race faults, and ball faults at motor speeds of 1,000, 1,500, and 2,000 r/min. The diagnostic performance of support vector machine (SVM), back-propagation neural network (BPNN), and the proposed algorithm was evaluated in operating condition transfer tasks. Moreover, ablation experiments were conducted. Results It was found that the proposed algorithm could effectively and accurately identify bearing faults in the face of changes in operating conditions. Discussion Both the domain category judgment module and the inter-domain conditional probability distribution difference could effectively achieve transfer learning of the diagnostic model.

Advances in transfer learning for smart wastewater treatment plants: Learning frameworks and emerging pathways.