As an important and widely studied research topic, the detection of fresh milk plays a crucial role in protecting consumer health. However, due to the diversity of milk brands and environmental conditions in real-world application scenarios, the accuracy of existing detection models often drops significantly when applied to new domains. A significant amount of new data needs to be gathered in order to retrain the model for new domain, which greatly increases both time and labor costs. To tackle this problem, this paper introduces a transfer learning-based cross-domain milk freshness detection method. The method uses transfer learning to generate data for unknown categories within the target domain. Specifically, the transfer generation model trains on data from both the target and source domain categories. With the proposed model, data from other categories in the source domain can be utilized to generate corresponding target domain data. This method comprehensively considers data generation from multiple perspectives of time, frequency, and spatial, to enhance the authenticity of the generated data. It uses a transfer generation architecture, consisting of a transfer network and a decoder, to learn the mapping relationship between the source and target domain, helping to narrow the gap between domains. Additionally, a feature subspace decomposition method is introduced into the decoder, and a cross-domain consistency loss function is formulated to strengthen the model’s learning capability. The proposed method is shown to generate high-quality data for unknown target-domain categories in multiple cross-domain settings, leading to a performance improvement ranging from 17.64% to 32.44% in the detection of cross-domain milk freshness.

Under practical operating conditions, intelligent fault diagnosis of permanent magnet synchronous motors (PMSMs) is often hindered by the shortage of effective fault samples. To address this issue, this paper proposes a twin-data-driven transfer learning-based diagnostic method for PMSM inter-turn short-circuit faults. First, a finite element model of the motor is established in Ansys to generate inter-turn short-circuit twin data, thereby enriching the source-domain samples. Second, continuous wavelet transform (CWT) is employed to convert stator current signals into multi-scale time–frequency feature maps, which are then fed into a feature extraction network constructed by integrating a residual network (ResNet) into an efficient channel attention mechanism (ECA) to achieve effective fusion of local and global time–frequency features. Finally, a joint loss function combining multi-kernel maximum mean discrepancy (MK-MMD) and a domain-adversarial neural network (DANN) is introduced to align feature distributions and perform adversarial optimization, enhancing cross-domain invariance and improving fault recognition capability. Experimental results demonstrate that the proposed REDM method achieves higher diagnostic accuracy and robustness than several existing intelligent fault diagnosis approaches.

ABSTRACT The Tibetan Plateau, known as the ‘Asian Water Tower’, faces significant water security challenges due to its high sensitivity to climate change. Advancing Earth observation for sustainable water monitoring is thus essential for building climate resilience in this region. This study proposes a two-stage transfer learning strategy using the SegFormer model to overcome domain shift and data scarcity – key barriers in developing robust AI for climate-sensitive applications. After pre-training on a diverse source domain, our model was fine-tuned for the arid Zhada Tulin area. Experimental results show a substantial performance boost: the Intersection over Union (IoU) for water body segmentation surged from 25.50% (direct transfer) to 64.84%. This AI-driven accuracy is crucial for disaster risk reduction, particularly in monitoring flash flood-prone systems. Beyond quantitative improvements, our analysis demonstrates that domain-specific transfer learning enhances the model’s detection capability. Specifically, it achieves high recall rates ( > 80 % ) while maintaining robust precision ( > 73 % ) – a balance critical for risk monitoring applications. The resulting high-precision maps provide valuable support for understanding hydrological patterns and informing water management strategies in this climatically sensitive region. Our work thus demonstrates an effective technical solution for monitoring arid plateau regions and contributes to advancing AI-powered Earth observation for disaster preparedness in critical transboundary river headwaters.

Population heterogeneity presents a significant challenge for two-sample Mendelian randomization (MR), often leading to biased estimates of causal effects. This heterogeneity arises when covariate distributions differ across populations, especially when these covariates function both as confounders and effect modifiers. To address this issue, we propose a new method, transfer learning-based Mendelian randomization (TLMR) that leverages observable effect modifiers to transfer predicted exposures from a source population to a target population. This transfer enables causal effect estimation in the target population while properly accounting for population differences. TLMR is developed under minimal modeling assumptions, allowing flexible exposure modeling, and supporting both continuous and binary outcomes. We further extend TLMR to accommodate reverse transfer in the outcome model that broadens its applicability in practical settings. Through extensive simulations, we demonstrate that TLMR yields robust and consistent estimates in heterogeneous populations, outperforming eight widely used MR methods that exhibit substantial estimation bias. Even in homogeneous populations or in the absence of effect modification, TLMR performs comparably to existing approaches. Finally, we systematically evaluate the causal relationship between body mass index and pulmonary function, demonstrating the practical utility and improved accuracy of TLMR in real-world analysis.

Traditional structure-based pre-screen compound selection relies on the assumption that chemical similarity implies similar biological activity. This paradigm narrows the exploration of chemical space and often fails to account for functional convergence, where structurally diverse compounds act through distinct targets to produce similar phenotypic effects. As a result, compounds with therapeutic potential may be overlooked. To overcome this constraint, we introduce a training-free, transfer learning-based method for large scale compound preselection that leverages deep phenotypic profiling of human cells. Notably, this enables robust pairwise comparison of phenotypic signatures across any source of the entire JUMP-CP, the largest publicly available cell painting dataset (112,480 compounds), preserving biological signals while mitigating batch effects. Validated across 65 high-throughput assays-including in vitro and in cellulo systems-our method provides efficient pre-screen enrichment of biologically active compounds, bypassing the blind spots of structure-centric approaches. Interestingly, because it is large scale, it also allows for a comprehensive analysis of structure-phenotypic activity relationships, revealing potentially thousands of compound activity cliffs, where minimal chemical changes in structure may result in profound phenotypic shifts. We show that these cliffs capture subtle, atom-level determinants of bioactivity that cannot be accessed by structure-based models. Furthermore, we demonstrate that structurally diverse compounds targeting different genes in the same biological pathway can induce either convergent or opposite phenotypes-a phenomenon validated across 30 pathways, hundreds of genes, and thousands of compounds. Finally, to support the broader community, we propose Phenoseeker, a web-based tool enabling instant retrieval of JUMP-CP compounds with similar phenotypic profiles. Together, these findings position phenotypic profiling not merely as a complementary tool, but as a transformative and scalable framework for navigating chemical space through a biological lens. By capturing rich morphological signatures that reflect functional outcomes-regardless of structural similarity-this approach enables the discovery of bioactive compounds, novel mechanisms of action, and unexpected target-pathway relationships. Applied at the scale of the JUMP-CP dataset, phenotypic profiling emerges as a powerful strategy for prioritizing compounds, illuminating activity cliffs, and accelerating the identification of therapeutically relevant candidates across diverse biological contexts.

Accurate labeling of training data is essential for reliable supervised machine learning, particularly in sensitive applications such as virus classification, autonomous driving, precision manufacturing, and medical diagnostics. However, the labeling process is labor-intensive and error-prone. Even widely used datasets such as MNIST and ImageNet contain numerous mislabeled samples. To address this challenge, we developed a transfer learning-based ensemble method that identifies mislabeled data through majority filtering and consensus filtering using fine-tuned pretrained deep neural networks, including ResNet-50, ResNet-101, VGG-16, EfficientNet, MobileNet, and Inception. Our approach was first validated on the MNIST dataset, where the ensemble detected approximately 751 label inconsistencies, which closely aligns with previously reported estimates of mislabeled samples. Additional experiments with synthetically injected mislabels demonstrated that the method could recover up to 100% of known corrupted labels using majority and consensus voting strategies. The method was then applied to a highly pure adeno-associated virus (AAV) nanopore dataset, where artificial mislabels were introduced for evaluation; the ensemble successfully identified most mislabeled samples and correctly recovered their true labels. Experiments on balanced and unbalanced AAV datasets further showed improved performance on the balanced subset, where all injected mislabels were detected. Compared to classical filtering techniques such as KNN, k-means clustering, and advanced machine learning-based mislabel detection (e.g., DivideMix), the proposed ensemble method demonstrated superior accuracy, stability, and true-label recovery, establishing it as a strong mislabel detection framework—well-suited for complex, fine-grained datasets such as nanopore signals and other biological measurement data.

Transfer learning-based method for data-driven seismic response prediction of steel moment resisting frames with different stories

Doped perovskites are widely studied in the domain of perovskite material design. However, due to the limited data available for the target materials, machine learning methods based on small datasets become particularly important. In this study, we propose a transfer learning strategy aimed at predicting doped perovskites on limited data samples. This strategy first utilizes the ABO3-type perovskite dataset to develop a deep learning source model based on its formation energies. Then, fine-tuning is performed on the doped perovskite structure dataset to obtain a model with good transferability, applicable to the doped perovskite oxide target domain. Based on the transfer learning model, we further predict the formation energies of 12,897 A2BB′O6 compounds, 10,401 AA′B2O6 compounds, and 49,723 AA′BB′O6 compounds. With the tolerance factor t ∈ [0.7–1.1], octahedral factor μ ∈ [0.45–0.7], and the modified tolerance factor τ ∈ [0, 4.18] for screening, we successfully predict 3389 A2B′BO6, 3002 AA′B2O6, and 13,563 AA′BB′O6 structures as potential stable doped perovskite candidates. Among these filtered results, 821 A2B′BO6, 69 AA′B2O6, and 6 AA′BB′O6 compounds have been reported in the OQMD database. For each doped perovskite, we select the candidate with the lowest formation energy and perform DFT validation. This resulted in three newly reported stable doped perovskite materials: CaSrHfScO6, BaSrHf2O6, and Ba2HfNdO6. The transfer learning-based perovskite material design method proposed in this study not only effectively addresses the challenges of model training on small datasets but also significantly improves the accuracy and stability of doped perovskite material predictions. Through transfer learning, the model can fully leverage the data and knowledge from the ABO3-type perovskite, effectively overcoming the problem of limited data. This strategy provides a new approach for efficient perovskite material design, enabling broader structural and performance predictions under limited experimental data conditions, and offering a powerful tool for the development of novel functional materials.

Recycling industrial solid waste phosphogypsum into phosphogypsum concrete (PGC) is a crucial pathway for achieving high-value solid waste utilization. However, the scarcity of experimental samples for PGC has led to inaccurate predictions of compressive strength by traditional models, severely hindering its application. This study proposes a dynamic weighted transfer learning-based method for predicting the strength of PGC, addressing the characterization bottleneck under small-sample conditions by transferring knowledge from the strength patterns of conventional concrete. First, feature differences between conventional concrete and PGC are eliminated through component proportion normalization and feature alignment. Then, a data augmentation technique based on Bootstrap Resampling is developed to generate enhanced samples that comply with mix proportion constraints, effectively expanding the training samples. Finally, an error feedback-driven dynamic weight calculation and weighted loss optimization framework for transfer learning is designed, prioritizing the learning of samples in the prediction blind spots of the target domain. This enables the adaptive acquisition of PGC-specific knowledge while inheriting the general knowledge of conventional concrete. Experimental results show that the transfer learning model achieves a prediction accuracy of R2 = 0.95 on the target domain test samples, a 15.9% improvement over traditional methods, while maintaining robust performance (R2 = 0.97) on an external validation samples. Feature importance analysis and Shapley Additive Explanations (SHAP) analysis reveal the nonlinear coupling effects of PGC-specific parameters on strength. This study establishes a scientific approach for accurate strength prediction of PGC under small-sample conditions.

User regret psychology-driven electric vehicle charging navigation strategy based on deep reinforcement learning and transfer learning

Abstract Transfer learning, leveraging the gaps through learned knowledge from source domain to target domain recognition, has achieved impressive performance in bearing fault diagnosis with two domains which have only weak distribution discrepancy. From different devices, collected vibration signal always suffered from big distribution discrepancy, which has limited the generalization of these existing transfer learning-based methods significantly. To overcome this challenge, the partial domain transfer learning methods are studied latest. However, most of these available techniques only focus on reducing the discrepancy of two domains using statistical distance metrics, which cannot consider the Riemannian manifold hidden in distribution space that impacted the accuracy of discrepancy measurement during testing. This paper proposes a Log-CORAL–based deep residual shrinkage network for partial domain adaptation bearing fault diagnosis scenario, where the domains data from different machinery are selected. Specifically, the log-correlation alignment (Log-CORAL), as a domain discrepancy metric, is explored to weaken the influence by Riemannian manifold, which disturbed the discrepancy measurement dependability. In addition, adversarial domain discriminator is embedded into deep residual shrinkage network to reduce the discrepancy between the two domains by maximizing the loss of domain discriminator. Comparison experiments with the SOTA methods on three well-known bearing datasets are conducted to verify effectiveness of the proposed method.

IntroductionAlzheimer’s disease (AD) is one of the most common neurodegenerative disabilities that often leads to memory loss, confusion, difficulty in language and trouble with motor coordination. Although several machine learning (ML) and deep learning (DL) algorithms have been utilized to identify Alzheimer’s disease (AD) from MRI scans, precise classification of AD categories remains challenging as neighbouring categories share common features.MethodsThis study proposes transfer learning-based methods for extracting features from MRI scans for multi-class classification of different AD categories. Four transfer learning-based feature extractors, namely, ResNet152V2, VGG16, InceptionV3, and MobileNet have been employed on two publicly available datasets (i.e., ADNI and OASIS) and a Merged dataset combining ADNI and OASIS, each having four categories: Moderate Demented (MoD), Mild Demented (MD), Very Mild Demented (VMD), and Non Demented (ND).ResultsResults suggest the Modified ResNet152V2 as the optimal feature extractor among the four transfer learning methods. Next, by utilizing the modified ResNet152V2 as a feature extractor, a Convolutional Neural Network based model, namely, the ‘IncepRes’, is proposed by fusing the Inception and ResNet architectures for multiclass classification of AD categories. The results indicate that our proposed model achieved a standard accuracy of 96.96%, 98.35% and 97.13% for ADNI, OASIS, and Merged datasets, respectively, outperforming other competing DL structures.DiscussionWe hope that our proposed framework may automate the precise classifications of various AD categories, and thereby can offer the prompt management and treatment of cognitive and functional impairments associated with AD.

High-frequency steady-state visual evoked potential-based brain-computer interface (SSVEP-BCI) systems offer improved user comfort but suffer from reduced performance compared to their low-frequency counterparts, limiting their practical application. To address this issue, we propose a transfer learning-based method that leverages low-frequency SSVEP data to enhance high-frequency SSVEP performance. A filtering mechanism is designed to extract informative components from low-frequency signals, and the least squares algorithm is employed to generate high-quality synthetic high-frequency data. Experiments conducted on two public datasets using TDCA, eTRCA, and advanced TRCA-based algorithms demonstrate significant performance improvements. Our approach requires only two calibration trials, achieving 9.03% and 14.49% accuracy increases for eTRCA and TDCA in Dataset 1, and 13.91% and 14.53% improvements in Dataset 2, all within 1.5s. Moreover, our approach effectively addresses the issue of single calibration data for high-frequency SSVEP-BCI systems. These results support the feasibility of fast calibration and improved performance in real-world high-frequency BCI applications.

Bidirectional transfer learning-based evaluation method for controlling femtosecond laser-induced porous structures of polymers

Aerodynamic noise is an important evaluation indicator for high-pressure ratio centrifugal compressors. However, using traditional numerical methods to predict aerodynamic noise requires significant computational resources and time, making it challenging to quickly assess the aerodynamic noise of compressors. This study proposes a transfer learning-based method for predicting the aerodynamic noise of centrifugal compressors. A rich set of aerodynamic noise data from the baseline compressor and a small amount of data from the target compressor (TC) were first obtained through experimental measurements. The transferability between datasets was evaluated using the maximum mean discrepancy method. Then, the pre-trained model was trained using data from the baseline compressor, and its generalization performance was validated. Finally, the pre-trained model was fine-tuned using noise data from the TC, and the model's performance was validated through mean squared error analysis. The results show that the proposed method can effectively and rapidly predict the aerodynamic noise of serial centrifugal compressors, with the overall sound pressure level error of the predicted frequency spectrum being less than 3 dB. Compared with traditional methods, this approach achieves high prediction accuracy with a small amount of training data.

Cutting force is a critical indicator reflecting the interaction between the cutting tool and the workpiece in machining processes. Conventional measurement methods using dynamometers are accurate but costly and challenging for real-time applications. This study proposes a novel transfer learning-based method for estimating milling cutting forces using only spindle vibration signals without direct force sensors. The proposed approach consists of two stages: First, an autoencoder is trained with measured cutting force data to construct a latent feature space. Second, a target encoder aligns spindle vibration signals to this latent space, allowing the decoder to reconstruct estimated cutting forces. To reflect machining parameters into the learning model, the input dataset was constructed by integrating material type, cutting speed, and cutting direction as additional inputs into each model’s inputs. Experiments were conducted on Ti-6Al-4V and STS316L workpieces under various machining conditions. Under normal conditions, the proposed method achieved an average Pearson correlation coefficient (PCC) of 0.9213 (Fx) and 0.9072 (Fy). Under abnormal transient conditions, robust performance was maintained, with PCC values of 0.8573 (Fx) and 0.9202 (Fy). The results demonstrate that the proposed method can effectively monitor cutting forces and reflect changes across a variety of machining environments using only vibration signals.

Single-cell RNA sequencing (scRNA-seq) analysis relies heavily on effective clustering to facilitate numerous downstream applications. Although several machine learning methods have been developed to enhance single-cell clustering, most are fully unsupervised and overlook the rich repository of annotated datasets available from previous single-cell experiments. Since cells are inherently high-dimensional entities, unsupervised clustering can often result in clusters that lack biological relevance. Leveraging annotated scRNA-seq datasets as a reference can significantly enhance clustering performance, enabling the identification of biologically meaningful clusters in target datasets. In this article, we propose Single Cell MUlti-Source CLustering (scMUSCL), a novel transfer learning method designed to identify cell clusters in a target dataset by leveraging knowledge from multiple annotated reference datasets. scMUSCL employs a deep neural network to extract domain- and batch-invariant cell representations, effectively addressing discrepancies across various source datasets and between source and target datasets within the new representation space. Unlike existing methods, scMUSCL does not require prior knowledge of the number of clusters in the target dataset and eliminates the need for batch correction between source and target datasets. We conduct extensive experiments using 20 real-life datasets, demonstrating that scMUSCL consistently outperforms existing unsupervised and transfer learning-based methods. Furthermore, our experiments show that scMUSCL benefits from multiple source datasets as learning references and accurately estimates the number of clusters. The Python implementation of scMUSCL is available at https://github.com/arashkhoeini/scMUSCL.

Flatfoot (pes planus) is a condition defined as the flattening of the curved structure as a result of the collapse of the foot or the weakening of the structures, such as ligaments and muscles that hold the bones and tissues in the foot in a certain order and a curve due to various reasons. If left untreated, this condition can lead to calf, knee, hip, and lower back pain and even postural disorders due to foot deterioration. In this study, a transfer learning-based method is presented using the Dilation filter for flatfoot detection from X-ray images. The X-ray image dataset contains 402 flatfoot images and 440 control images. For image preprocessing, dilation filtering is used, and the images are enhanced with the dilation method. After image preprocessing, the performance of transfer learning approaches, DarkNet19, GoogLeNet, DenseNet-201, ResNet-101, and MobileNetV2 architectures, were compared. The holdout method was used for performance measurements. The experimental results show that the DenseNet-201 model performs the best with an overall accuracy of 0.9802 and a Cohen's Kappa value of 0.96. The results show that the combination of dilation filtering and transfer learning methods provides an effective approach for automatic flatfoot detection. Compared to similar studies in the literature, the accuracy of the proposed model is significantly higher.

Introduction: Monkeypox (MPox) is a continuing global public health concern for zoonotic disease. Accurate and prompt diagnoses of MPox are key to programmatic disease control. Established key diagnostic modalities are resource intensive, requiring technology and facilities, clinical examination and laboratory testing are thus time-consuming. Conversely, machine learning and deep learning techniques provide fast and automatic diagnostic solutions. Objectives:To diagnose monkeypox from clinical imagery this work proposes a transfer learning-based method utilizing the EfficientNetB3 and ResNet50V2 models. These models go very well with image classification and prediction. Although this approach is intended to be useful for therapeutic usage in the detection of monkeypox at low resources and inadquate healthcare facilities. This research demonstrates how transfer learning can be utilized to implement pre-trained models for Monkeypox detection with high accuracy thereby reducing the necessity of fully labeled datasets. Methods: Using transfer learning, this study explores two pre-trained efficient model architectures – EfficientNetB3 and ResNet50V2, to classify skin lesion images exhibited in MPox. Pre-trained on large scale datasets, these models are fine-tuned using the Monkeypox Skin Lesion Dataset (MSLD) v2. 0 to improve prediction accuracy. Simple rotating, scaling, scaling, brightness adjustment, and other methods are used for data augmentation to enrich the dataset and promote generalization. Convolutional neural network is a deep learning architecture which, especially transfer learning based network, significantly improves detection and robustness of MPox with high accuracy and less qualification of the data set labels. Results: The results of this study demonstrate that deep learning—more especially, transfer learning—can be an effective instrument for managing and detecting outbreaks of monkeypox early on, perhaps leading to better patient outcomes and less strain on healthcare systems. It is essential to combine cutting-edge transfer learning techniques with well-established deep learning algorithms to increase prediction accuracy and clarify the complexities of the ongoing worldwide monkeypox outbreak. Our work presents novel approaches to address this significant health issue while highlighting the ongoing significance of ground-breaking methods. Our proposed models with accuracy of 0.89 and 0.99 have outperformed the existing models in MPox detection. Conclusions: The suggested transfer learning architecture outperforms the most recent models in terms of MPox detection capabilities. Better early MPox identification, more effective treatment planning, and ultimately better patient care could be the outcomes of these developments.

Intelligent fault diagnosis (IFD) for mechanical equipment based on small and imbalanced datasets has been widely studied in recent years, with transfer learning emerging as one of the most promising approaches. Existing transfer learning-based IFD methods typically use data from different operating conditions of the same equipment as the source and target domains for the transfer learning process. However, in practice, it is often challenging to find identical equipment to obtain source domain data when diagnosing faults in the target equipment. These strict assumptions pose significant limitations on the application of IFD techniques in real-world industrial settings. Furthermore, the temporal characteristics of time-series monitoring data are often inadequately considered in existing methods. In this paper, we propose a cross-machine IFD method based on a residual full convolutional neural network (ResFCN) transfer learning model, which leverages the time-series features of monitoring data. By incorporating sliding window (SW)-based data segmentation, network pretraining, and model fine-tuning, the proposed method effectively exploits fault-associated general features in the source domain and learns domain-specific patterns that better align with the target domain, ultimately achieving accurate fault diagnosis for the target equipment. We design and implement three sets of experiments using two widely used public datasets. The results demonstrate that the proposed method outperforms existing approaches in terms of fault diagnosis accuracy and robustness.