Search

The hydraulic column is a core component in the coal mine support system, however, the real-time monitoring of the hydraulic column during the service process of the hydraulic support faces challenges. To address these issues, a high-precision stress mapping method of hydraulic column is proposed. The hydraulic column loss function was constructed to guide the data-driven model training, and the cylinder stress mechanism model was established by using the elastic–plastic theory of thick-walled cylinder. The weight coefficients of the data-driven model and the mechanism model were determined by the measurement data, and the high-precision stress mapping results were obtained by combining the data-driven model data and mechanism model data. A monitoring platform was built for the unit hydraulic support column, and the unit hydraulic support column was used as the test object to carry out pressure tests of 600 kN, 800 kN, 1000 kN and 1200 kN. The test data of the four groups showed that the error between the stress mapping results and the measured value of the strain gauges was 3.086%, 1.783%, 1.182% and 1.548%, respectively. The mapping period is 0.32 s, which proves that the method can provide real-time and high-precision feedback on the stress state of the hydraulic column, realize real-time monitoring of the stress state of the hydraulic column, solve the problem that stress cannot be measured by attaching strain gauges in the complex underground environment, predict the fault and safety problems of the hydraulic column due to overload, and provide data reserve for the adaptive regulation of the hydraulic support. [ABSTRACT FROM AUTHOR]

To enhance the accuracy of motor imagery(MI)EEG signal recognition, two methods, namely power spectral density and wavelet packet decomposition combined with a common spatial pattern, were employed to explore the feature information in depth in MI EEG signals. The extracted MI EEG signal features were subjected to series feature fusion, and the F-test method was used to select features with higher information content. Here regarding the accuracy of MI EEG signal classification, we further proposed the Platt Scaling probability calibration method was used to calibrate the results obtained from six basic classifiers, namely random forest (RF), support vector machines (SVM), Logistic Regression (LR), Gaussian naïve bayes (GNB), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). From these 12 classifiers, three to four with higher accuracy were selected for model fusion. The proposed method was validated on Datasets 2a of the 4th International BCI Competition, achieving an average accuracy of MI EEG data of nine subjects reached 91.46%, which indicates that model fusion was an effective method to improve classification accuracy, and provides some reference value for the research on MI brain-machine interface. [ABSTRACT FROM AUTHOR]

Speech recognition is widely applied in fields like security, education, and healthcare. While its development drives global information infrastructure and AI strategies, current models still face challenges such as overfitting, local optima, and inefficiencies in decoding accuracy and computational cost. These issues cause instability and long response times, hindering AI's competitiveness. Therefore, addressing these technical bottlenecks is critical for advancing national scientific progress and global information infrastructure. In this paper, we propose improvements to the model structure fusion and decoding algorithms. First, based on the Conformer network and its variants, we introduce a weighted fusion method using training loss as an indicator, adjusting the weights, thresholds, and other related parameters of the fused models to balance the contributions of different model structures, thereby creating a more robust and generalized model that alleviates overfitting and local optima. Second, for the decoding phase, we design a dynamic adaptive decoding method that combines traditional decoding algorithms such as connectionist temporal classification and attention-based models. This ensemble approach enables the system to adapt to different acoustic environments, improving its robustness and overall performance. Additionally, to further optimize the decoding process, we introduce a penalty function mechanism as a regularization technique to reduce the model's dependence on a single decoding approach. The penalty function limits the weights of decoding strategies to prevent over-reliance on any single decoder, thus enhancing the model's generalization. Finally, we validate our model on the Librispeech dataset, a large-scale English speech corpus containing approximately 1000 h of audio data. Experimental results demonstrate that the proposed method achieves word error rates (WERs) of 3.92% and 4.07% on the development and test sets, respectively, significantly improving over single-model and traditional decoding methods. Notably, the method reduces WER by approximately 0.4% on complex datasets compared to several advanced mainstream models, underscoring its superior robustness and adaptability in challenging acoustic environments. The effectiveness of the proposed method in addressing overfitting and improving accuracy and efficiency during the decoding phase was validated, highlighting its significance in advancing speech recognition technology. [ABSTRACT FROM AUTHOR]

In this paper, we proposed a complete study method to achieve accurate aortic dissection diagnosis at the patient level. Based on the CT angiography (CTA) images, a classification model named DAT-DenseNet, which combined the deep attention Transformer module with the DenseNet architecture is proposed. In the first phase, two DAT-DenseNet are combined in parallel. It is used to accurately achieve two classification task at the CTA images. In the second stage, we propose a feature fusion module. It concatenates and fuses the image features output from the two classification models on a patient by patient basis. In the comparison experiments of classification model performance, DAT-DenseNet obtained 92.41 % accuracy at the image level, which was 2.20 % higher than the commonly used model. In the comparison experiments of model fusion method, our method obtained 90.83 % accuracy at the patient level. The experiments showed that DAT-DenseNet model exhibits high performance at the image level. Our feature fusion module achieves the mapping from two classification image features to patient outcomes. It achieves accurate patient classification. The experiments' results in the Discussion section elaborate the details of the experiment and confirmed that the results were reliable. [ABSTRACT FROM AUTHOR]

Copyright of World Nuclear Geoscience is the property of World Nuclear Geoscience Editorial Office and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The deep learning model has attracted widespread attention in the field of rolling bearing remaining useful life (RUL) prediction due to its advantages of less reliance on prior knowledge, high accuracy, and strong generalization. However, a large number of prediction models use very complicated artificial feature extraction and selection methods to build the original input features of the deep learning model and health indicator. These approaches do not fully exploit the capabilities of deep learning models as they continue to heavily rely on prior knowledge, The accuracy of their predictions largely hinges on the quality of the input features, and the generalization of manually crafted features remains uncertain. To address these challenges, in this paper, an end-to-end prediction model for the remaining useful life of rolling bearings is proposed, which is divided into three modules. First, a short-term Fourier transform module is incorporated into the model to automatically obtain the time–frequency information of the signal. Then, the convolutional next (ConvNext) module, which is a simple and efficient pure convolutional neural network, is utilized to extract features from the spectrogram. Finally, we capture the short-term dependence and long-term dependence by two parallel channels Transformer and self-attention convolutional long short-term memory (SA-ConvLSTM), and the self-attention mechanism is employed for the adaptive prediction of the bearing's remaining useful life. Through integration with artificial intelligence, this method proposes a high-performance solution for predicting the remaining useful life of bearings. It has minimal reliance on manual labor, stronger fitting capabilities, and can be widely used for predicting the remaining useful life of bearings. [ABSTRACT FROM AUTHOR]

Improving the ability and accuracy of intelligent state prediction of large and complex equipment is one of the important directions of current intelligent operation and maintenance technology research. Due to the influence of insufficient analysis of equipment degradation characteristics, single function of traditional prediction model, and difficulty in determining the optimal parameters of the model make the prediction effect poor. In this paper, a state prediction model fusion optimization strategy is proposed for lithium mill equipment as an example. Based on the process flow and vibration mechanism, the inherent vibration characteristics of the roller bearing system are analyzed, and the degradation characteristics of the roller bearing under resonance conditions are explored from the finite element equivalent model, so as to determine the equipment operation stage and the starting point of degradation. The state prediction task is divided into degradation phase and residual life prediction phase, and Time-Convolutional Denoising Autoencoder (TCDAE) and two-layer Sparse Auto Encoder (SAE) are designed for data feature enhancement and degradation feature fusion and dimensionality reduction. Construct BO-BiGRU state prediction model to mine the feature information hidden in the whole time series of data points and adjust the model parameters adaptively using Bayesian Optimization method. The novelty of this study is to analyze the degradation characteristics of key components, correct the theoretical degradation starting point by using the degradation trend formula, and establish a unified framework from monitoring data to condition prediction. Compared with the original model constructed by the above algorithm, the fusion model proposed in this paper has significantly improved performance. The data analysis shows that the prediction accuracy after model fusion is substantially improved, and the accuracy after TCDAE feature enhancement is improved by about 10.2%, the accuracy after two-layer SAE model fusion and dimensionality reduction improved by about 9.8%, and the state accuracy after BO-BiGRU model improved by about 11.6%. The crux to the research depends on the construction of a state prediction model, which is based on the analysis of the bearing degradation process and the effective integration of algorithms. Predictive maintenance of critical components also improves product quality. [ABSTRACT FROM AUTHOR]

Recently, remote sensing image forgery detection has received widespread attention. To improve the detection accuracy, we build a novel scheme based on Transformer and model fusion. Specifically, we model this task as a binary classification task that focuses on global information. First, we explore the performance of various excellent feature extraction networks in this task under the constructed unified classification framework. On this basis, we select three high-performance Transformer-based networks that focus on global information, namely, Swin Transformer V1, Swin Transformer V2, and Twins, as the backbone networks and fuse them. Secondly, considering the small number of samples, we use the public ImageNet-1K dataset to pre-train the network to learn more stable feature expressions. At the same time, a circular data divide strategy is proposed, which can fully utilize all the samples to improve the accuracy in the competition. Finally, to promote network optimization, on the one hand, we explore multiple loss functions and select label smooth loss, which can reduce the model's excessive dependence on training data. On the other hand, we construct a combined learning rate optimization strategy that first uses step degeneration and then cosine annealing, which reduces the risk of the network falling into local optima. Extensive experiments show that the proposed scheme has excellent performance. This scheme won seventh place in the "Forgery Detection in Multi-scenario Remote Sensing Images of Typical Objects" track of the 2024 ISPRS TC I contest on Intelligent Interpretation for Multi-modal Remote Sensing Application. [ABSTRACT FROM AUTHOR]

Background: Obstructive sleep apnea is a sleep disorder that is linked to many health complications and can even be lethal in its severe form. Overnight polysomnography is the gold standard for diagnosing apnea, which is expensive, time-consuming, and requires manual analysis by a sleep expert. Artificial intelligence (AI)-embedded wearable device as a portable and less intrusive monitoring system is a highly desired alternative to polysomnography. However, AI models often require substantial storage capacity and computational power for edge inference which makes it a challenging task to implement the models in hardware with memory and power constraints. Methods: This study demonstrates the implementation of depth-wise separable convolution (DSC) as a resource-efficient alternative to spatial convolution (SC) for real-time detection of apneic activity. Single lead electrocardiogram (ECG) and oxygen saturation (SpO2) signals were acquired from the PhysioNet databank. Using each type of convolution, three different models were developed using ECG, SpO2, and model fusion. For both types of convolutions, the fusion models outperformed the models built on individual signals across all the performance metrics. Results: Although the SC-based fusion model performed the best, the DSC-based fusion model was 9.4, 1.85, and 11.3 times more energy efficient than SC-based ECG, SpO2, and fusion models, respectively. Furthermore, the accuracy, precision, and specificity yielded by the DSC-based fusion model were comparable to those of the SC-based individual models (~95%, ~94%, and ~94%, respectively). Conclusions: DSC is commonly used in mobile vision tasks, but its potential in clinical applications for 1-D signals remains unexplored. While SC-based models outperform DSC in accuracy, the DSC-based model offers a more energy-efficient solution with acceptable performance, making it suitable for AI-embedded apnea detection systems. [ABSTRACT FROM AUTHOR]

Personalized federated learning (PFL) is a framework that targets individual models for optimization, providing better privacy and flexibility for clients. However, in challenging intelligent sensing applications, the heterogeneous client's data distributions make the aggregation of local models in the server unstable or even hard to converge. To deal with the performance degradation caused by the preceding problem, existing PFL methods focus more on how to fine-tune the global model but ignore the impact of the global model fusion algorithm on the results. In this article, we propose a new explainable neural-aware decoupling fusion based PFL framework, p-FedADF, to address the preceding challenges. It contains two carefully designed modules. The local decoupling module, deployed on the client, utilizes the architecture disentangle technique to decouple the feature extractors in the client's local model into sub-network according to data categories. It obtains the inference process of feature extraction for different categories of data by training. The global aggregation module, deployed on the server, aligns the sub-network positions for multiple clients and implements a fine-grained generic feature extractor aggregation. In addition, we provide a mask encoding scheme to reduce the communication overhead of transmitting the sub-network sets between the server and clients. Our p-FedADF obtains 1.6%, 0.2%, 2.3%, and 4.5% improvement on a real-world dataset and three benchmark datasets, compared to state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Abstract The hydraulic column is a core component in the coal mine support system, however, the real-time monitoring of the hydraulic column during the service process of the hydraulic support faces challenges. To address these issues, a high-precision stress mapping method of hydraulic column is proposed. The hydraulic column loss function was constructed to guide the data-driven model training, and the cylinder stress mechanism model was established by using the elastic–plastic theory of thick-walled cylinder. The weight coefficients of the data-driven model and the mechanism model were determined by the measurement data, and the high-precision stress mapping results were obtained by combining the data-driven model data and mechanism model data. A monitoring platform was built for the unit hydraulic support column, and the unit hydraulic support column was used as the test object to carry out pressure tests of 600 kN, 800 kN, 1000 kN and 1200 kN. The test data of the four groups showed that the error between the stress mapping results and the measured value of the strain gauges was 3.086%, 1.783%, 1.182% and 1.548%, respectively. The mapping period is 0.32 s, which proves that the method can provide real-time and high-precision feedback on the stress state of the hydraulic column, realize real-time monitoring of the stress state of the hydraulic column, solve the problem that stress cannot be measured by attaching strain gauges in the complex underground environment, predict the fault and safety problems of the hydraulic column due to overload, and provide data reserve for the adaptive regulation of the hydraulic support.

Remote patient monitoring has emerged as a prominent non-invasive method, using digital technologies and computer vision (CV) to replace traditional invasive monitoring. While neonatal and pediatric departments embrace this approach, Pediatric Intensive Care Units (PICUs) face the challenge of occlusions hindering accurate image analysis and interpretation. Goal: In this study, we propose a hybrid approach to effectively segment common occlusions encountered in remote monitoring applications within PICUs. Our approach centers on creating a deep-learning pipeline for limited training data scenarios. Methods: First, a combination of the well-established Google DeepLabV3+ segmentation model with the transformer-based Segment Anything Model (SAM) is devised for occlusion segmentation mask proposal and refinement. We then train and validate this pipeline using a small dataset acquired from real-world PICU settings with a Microsoft Kinect camera, achieving an Intersection-over-Union (IoU) metric of 85%. Results: Both quantitative and qualitative analyses underscore the effectiveness of our proposed method. The proposed framework yields an overall classification performance with 92.5% accuracy, 93.8% recall, 90.3% precision, and 92.0% F1-score. Consequently, the proposed method consistently improves the predictions across all metrics, with an average of 2.75% gain in performance compared to the baseline CNN-based framework. Conclusions: Our proposed hybrid approach significantly enhances the segmentation of occlusions in remote patient monitoring within PICU settings. This advancement contributes to improving the quality of care for pediatric patients, addressing a critical need in clinical practice by ensuring more accurate and reliable remote monitoring.

The deep learning model has attracted widespread attention in the field of rolling bearing remaining useful life (RUL) prediction due to its advantages of less reliance on prior knowledge, high accuracy, and strong generalization. However, a large number of prediction models use very complicated artificial feature extraction and selection methods to build the original input features of the deep learning model and health indicator. These approaches do not fully exploit the capabilities of deep learning models as they continue to heavily rely on prior knowledge, The accuracy of their predictions largely hinges on the quality of the input features, and the generalization of manually crafted features remains uncertain. To address these challenges, in this paper, an end-to-end prediction model for the remaining useful life of rolling bearings is proposed, which is divided into three modules. First, a short-term Fourier transform module is incorporated into the model to automatically obtain the time–frequency information of the signal. Then, the convolutional next (ConvNext) module, which is a simple and efficient pure convolutional neural network, is utilized to extract features from the spectrogram. Finally, we capture the short-term dependence and long-term dependence by two parallel channels Transformer and self-attention convolutional long short-term memory (SA-ConvLSTM), and the self-attention mechanism is employed for the adaptive prediction of the bearing’s remaining useful life. Through integration with artificial intelligence, this method proposes a high-performance solution for predicting the remaining useful life of bearings. It has minimal reliance on manual labor, stronger fitting capabilities, and can be widely used for predicting the remaining useful life of bearings.

Accurate monitoring of anesthesia status is very important in surgery, as it can guide anesthesiologists, reduce drug usage, and reduce postoperative adverse effects. However, due to the complex interactions between anesthetic drugs and the central nervous system, there is no perfect monitoring method. In recent years, the development of artificial intelligence technology has offered the possibility of using machine learning algorithms to achieve more accurate monitoring of anesthesia depth. In this paper, four levels of anesthesia states were classified and multifaceted feature values were extracted from Electroencephalogram (EEG) signals, a convolutional neural network-based KRDGB-CNN model was constructed, which was based on K-nearest neighbor (KNN), Random Forest (RF), Decision Tree (DT), Gaussian Naive Baye (GNB), and Back propagation Neural Network (BP), and fused by Convolutional Neural Network (CNN) algorithm for decision layers. By evaluating the model performance on the collected data, the results show that the model outperforms existing algorithms in terms of classification accuracy and specificity, and can effectively improve the robustness and accuracy of the algorithm. [ABSTRACT FROM AUTHOR]

Accurate categorization and timely control of leaf diseases are crucial for citrus growth. We proposed the Multi-Models Fusion Network (MMFN) for citrus leaf diseases detection based on model fusion and transfer learning. Compared to traditional methods, the algorithm (integrating transfer learning Alexnet, VGG, and Resnet) we proposed can address the issues of limited categories, slow processing speed, and low recognition accuracy. By constructing efficient deep learning models and training and optimizing them with a large dataset of citrus leaf images, we ensured the broad applicability and accuracy of citrus leaf disease detection, achieving high-precision classification. Herein, various deep learning algorithms, including original Alexnet, VGG, Resnet, and transfer learning versions Resnet34 (Pre_Resnet34) and Resnet50 (Pre_Resnet50) were also discussed and compared. The results demonstrated that the MMFN model achieved an average accuracy of 99.72% in distinguishing between diseased and healthy leaves. Additionally, the model attained an average accuracy of 98.68% in the classification of multiple diseases (citrus huanglongbing (HLB), greasy spot disease and citrus canker), insect pests (citrus leaf miner), and deficiency disease (zinc deficiency). These findings conclusively illustrate that deep learning model fusion networks combining transfer learning and integration algorithms can automatically extract image features, enhance the automation and accuracy of disease recognition, demonstrate the significant potential and application value in citrus leaf disease classification, and potentially drive the development of smart agriculture. [ABSTRACT FROM AUTHOR]

Federated learning is a new learning paradigm that decouples data collection and model training via multi-party computation and model aggregation. As a flexible learning setting, federated learning has the potential to integrate with other learning frameworks. We conduct a focused survey of federated learning in conjunction with other learning algorithms. Specifically, we explore various learning algorithms to improve the vanilla federated averaging algorithm and review model fusion methods such as adaptive aggregation, regularization, clustered methods, and Bayesian methods. Following the emerging trends, we also discuss federated learning in the intersection with other learning paradigms, termed federated X learning, where X includes multitask learning, meta-learning, transfer learning, unsupervised learning, and reinforcement learning. In addition to reviewing state-of-the-art studies, this paper also identifies key challenges and applications in this field, while also highlighting promising future directions. [ABSTRACT FROM AUTHOR]

A fusion algorithm named permutation entropy weighted and bias adjustment rule fusion (PEW-BAR) was proposed to enhance the accuracy of speech emotion recognition by exploiting the emotional information in the spectral characteristics of speech signals. The algorithm was based on the integration of wavelet scattering transform and Mel-frequency cepstral coefficients (MFCC). Firstly, wavelet scattering features and MFCC-related features from speech signals were extracted. Then, the wavelet scattering features were expanded in the scale dimension and applied support vector machines to obtain posterior probabilities for emotion recognition. And permutation entropy was calculated and a weighted fusion based on this entropy was subsequently applied. Finally, a bias adjustment rule was utilized to refine the integration results obtained from the MFCC-related features. Experimental results on various datasets, including EMODB, RAVDESS, and eNTERFACE05, demonstrate notable improvements. The proposed algorithm outperforms traditional wavelet scattering coefficient-based methods, achieving accuracy improvements of 2.82%, 2.85%, and 5.92%, respectively. Additionally, it shows enhancements of 3.40%, 2.87%, and 5.80% in terms of unweighted average recall (UAR), and a 6.89% improvement on the IEMOCAP dataset.

The Face Anti-Spoofing (FAS) methods plays a very important role in ensuring the security of face recognition systems. The existing FAS methods perform well in short-distance scenarios, e.g., phone unlocking, face payment, etc. However, it is still challenging to improve the generalization of FAS in long-distance scenarios (e.g., surveillance) due to the varying image quality. In order to address the lack of low-quality images in real scenarios, we build a Low-Quality Face Anti-Spoofing Dataset (LQFA-D) by using Hikvision's surveillance cameras. In order to deploy the model on an edge device with limited computation, we propose a lightweight FAS network based on MobileFaceNet, in which the Coordinate Attention (CA) attention model is introduced to capture the important spatial information. Then, we propose a multi-scale FAS framework for low-quality images to explore multi-scale features, which includes three multi-scale models. The experimental results of the LQFA-D show that the Average Classification Error Rate (ACER) and detection time of the proposed method are 1.39% and 45 ms per image for the low-quality images, respectively. It demonstrates the effectiveness of the proposed method in this paper. [ABSTRACT FROM AUTHOR]

The cyber-physical data of equipment whole life cycle are interrelated and play an important role in decision-making for business activities. However, multi-stage cyber-physical data are numerous, heterogeneous, and massive; non-real-time information data and real-time physical data are challenging to correspond to each other because of the inconsistency in space-time and granularity. As a result, equipment multi-stage heterogeneous cyber-physical data are difficult to effectively integrate and apply. Therefore, multi-stage cyber-physical data fusion of equipment is a key problem to be solved in manufacturing. To the best of our knowledge, a practicable framework and method to effectively fuse and apply equipment multi-stage data are still missing. To overcome this gap, this study proposes a multidimensional and multilevel cyber-physical fusion method for equipment. First, a four-dimensional cyber-physical fusion framework based on time, data, model, and structure is put forward. Then, the four kinds of fusion, including data fusion, model fusion, knowledge fusion, and digital twin (DT) application fusion, are discussed detailedly. Besides, the five types of algorithms set to support the fusion process are given. Finally, equipment life cycle activity is regarded as a DT application process using dynamic association of data and models. Through cyber-physical fusion, the data of equipment multi- stages could be applied for different activities to solve problems by means of DT. As a primary verification of the feasibility of the proposed approach, a case study of metro vehicle performance evaluation has been carried out, and the results have well confirmed that the multi-stage cyber-physical fusion method is feasible. [ABSTRACT FROM AUTHOR]

Forest fire risk prediction is a crucial link in maintaining forest ecological security. Machine learning, due to its powerful non-linear modeling capabilities, has been widely applied in forest fire risk prediction research. However, existing studies often focus on the direct information provided by multiple environmental factor features when constructing the feature space, while overlooking the deeper information conveyed by feature cross-correlations. Additionally, fire risk prediction predominantly relies on single-model forecasting, exhibiting slightly insufficient generalization and stability in models. Model fusion algorithms (MFA) can combine the advantages of multiple models to compensate for this limitation. In this study, a machine learning framework, FC-StackGNB, combining feature crosses (FC) and model fusion, is proposed. This framework employs the FC method to analyze the temporal trends of various environmental factors influencing fire occurrence, constructing multiple seasonal cross features (SCFs) capable of effectively capturing the non-linear relationship between environmental factors and time. Moreover, the framework develops a Gaussian Naive Bayes (GNB) optimized stacking MFA to fully leverage the strengths of different ML algorithms. Results demonstrate that the introduction of SCFs effectively enhances the prediction performance of six machine learning models, with the mean values of five evaluation metrics (Accuracy, Precision, Recall, F1-score, and ROC_AUC) increasing by 1.58% to 6.30%. The fusion model constructed based on the StackGNB algorithm can effectively handle the multicollinearity issue of features, exhibiting significantly better prediction performance than single models, particularly in improving the Recall metric (increasing by around 3% and 5% compared to the top two ranked single models respectively), which signifies the model’s ability to predict positive samples (i.e., high-risk fire areas). The proposed modeling framework effectively enhances the robustness and prediction performance of the models, offering new modeling insights for subsequent research. This study holds significant importance for enhancing the level of forest fire risk warning.

Abstract The application of near-infrared spectroscopy (NIRS) for rapid quantitative analysis of soil total nitrogen (STN) is of great significance to recycling nitrogen in the ecosystem and crops growth. However, collecting thousands of soil samples and chemical analysis are impracticable, more importantly a deviation from NIRS advantages of rapid, inexpensive and nondestructive. To more efficiently improve the estimation performance and reduce uncertainty of the model when working with small sample sizes (less than 100), solutions from soil particle size decomposition and model fusion were investigated. Elaborately, 123 Latosols samples were collected and decomposed them according to particle sizes to extend limited data at multiple scales. Based on all soil groups decomposed, a hyperspectral data recapture and model decision fusion method were implemented. The results demonstrated that the proposed method increased the scale of spectral data, extracted more STN-related spectral information, improved estimation accuracy, and reduced uncertainty. The fused model based on data from all decomposed groups yielded the best estimated results (root mean square error $$(RMSE) = 0.075g.kg^{-1}$$ ( R M S E ) = 0.075 g . k g - 1 , $$R^2 = 0.784$$ R 2 = 0.784 , and a ratio of performance to inter-quartile distance $$(RPIQ) = 3.787$$ ( R P I Q ) = 3.787 ) on the validation set. Through a tenfold cross-validation, the weighted fusion model with six groups of particle sizes data showed an improvement of 0.307 in $$R^2_cv$$ R c 2 v and an improved RPIQ of 1.015 compared to models constructed using conventional machine learning (ML) techniques and limited pristine data ( $$R^2_cv = 0.442, RMSE = 0.119$$ R c 2 v = 0.442 , R M S E = 0.119 ). Therefore, when utilizing NIRS to build rapid and accurate STN predictive models, the proposed method demonstrates great potential in improving the reliability of soil spectral models under small sample sizes. Graphical Abstract

Speech recognition is widely applied in fields like security, education, and healthcare. While its development drives global information infrastructure and AI strategies, current models still face challenges such as overfitting, local optima, and inefficiencies in decoding accuracy and computational cost. These issues cause instability and long response times, hindering AI’s competitiveness. Therefore, addressing these technical bottlenecks is critical for advancing national scientific progress and global information infrastructure. In this paper, we propose improvements to the model structure fusion and decoding algorithms. First, based on the Conformer network and its variants, we introduce a weighted fusion method using training loss as an indicator, adjusting the weights, thresholds, and other related parameters of the fused models to balance the contributions of different model structures, thereby creating a more robust and generalized model that alleviates overfitting and local optima. Second, for the decoding phase, we design a dynamic adaptive decoding method that combines traditional decoding algorithms such as connectionist temporal classification and attention-based models. This ensemble approach enables the system to adapt to different acoustic environments, improving its robustness and overall performance. Additionally, to further optimize the decoding process, we introduce a penalty function mechanism as a regularization technique to reduce the model’s dependence on a single decoding approach. The penalty function limits the weights of decoding strategies to prevent over-reliance on any single decoder, thus enhancing the model’s generalization. Finally, we validate our model on the Librispeech dataset, a large-scale English speech corpus containing approximately 1000 h of audio data. Experimental results demonstrate that the proposed method achieves word error rates (WERs) of 3.92% and 4.07% on the development and test sets, respectively, significantly improving over single-model and traditional decoding methods. Notably, the method reduces WER by approximately 0.4% on complex datasets compared to several advanced mainstream models, underscoring its superior robustness and adaptability in challenging acoustic environments. The effectiveness of the proposed method in addressing overfitting and improving accuracy and efficiency during the decoding phase was validated, highlighting its significance in advancing speech recognition technology.

Background: Obstructive sleep apnea is a sleep disorder that is linked to many health complications and can even be lethal in its severe form. Overnight polysomnography is the gold standard for diagnosing apnea, which is expensive, time-consuming, and requires manual analysis by a sleep expert. Artificial intelligence (AI)-embedded wearable device as a portable and less intrusive monitoring system is a highly desired alternative to polysomnography. However, AI models often require substantial storage capacity and computational power for edge inference which makes it a challenging task to implement the models in hardware with memory and power constraints. Methods: This study demonstrates the implementation of depth-wise separable convolution (DSC) as a resource-efficient alternative to spatial convolution (SC) for real-time detection of apneic activity. Single lead electrocardiogram (ECG) and oxygen saturation (SpO2) signals were acquired from the PhysioNet databank. Using each type of convolution, three different models were developed using ECG, SpO2, and model fusion. For both types of convolutions, the fusion models outperformed the models built on individual signals across all the performance metrics. Results: Although the SC-based fusion model performed the best, the DSC-based fusion model was 9.4, 1.85, and 11.3 times more energy efficient than SC-based ECG, SpO2, and fusion models, respectively. Furthermore, the accuracy, precision, and specificity yielded by the DSC-based fusion model were comparable to those of the SC-based individual models (~95%, ~94%, and ~94%, respectively). Conclusions: DSC is commonly used in mobile vision tasks, but its potential in clinical applications for 1-D signals remains unexplored. While SC-based models outperform DSC in accuracy, the DSC-based model offers a more energy-efficient solution with acceptable performance, making it suitable for AI-embedded apnea detection systems.

Recently, remote sensing image forgery detection has received widespread attention. To improve the detection accuracy, we build a novel scheme based on Transformer and model fusion. Specifically, we model this task as a binary classification task that focuses on global information. First, we explore the performance of various excellent feature extraction networks in this task under the constructed unified classification framework. On this basis, we select three high-performance Transformer-based networks that focus on global information, namely, Swin Transformer V1, Swin Transformer V2, and Twins, as the backbone networks and fuse them. Secondly, considering the small number of samples, we use the public ImageNet-1K dataset to pre-train the network to learn more stable feature expressions. At the same time, a circular data divide strategy is proposed, which can fully utilize all the samples to improve the accuracy in the competition. Finally, to promote network optimization, on the one hand, we explore multiple loss functions and select label smooth loss, which can reduce the model’s excessive dependence on training data. On the other hand, we construct a combined learning rate optimization strategy that first uses step degeneration and then cosine annealing, which reduces the risk of the network falling into local optima. Extensive experiments show that the proposed scheme has excellent performance. This scheme won seventh place in the “Forgery Detection in Multi-scenario Remote Sensing Images of Typical Objects” track of the 2024 ISPRS TC I contest on Intelligent Interpretation for Multi-modal Remote Sensing Application.

In order to improve the control performance and reliability of the pulse-width modulation (PWM) rectifiers in electric vehicle (EV) charging systems, the evaluation of DC bus capacitor health status is critical. In order to accurately monitor the health status of DC bus capacitors, a data-driven model fusion method is developed. In the method, multi-layer perceptron, random forest, and XGBoost are adopted as the base learners that produce separate row predictions. The second-level learner, support vector machine (SVM) accepts the outputs of the previous learners and integrates them into the final health status prediction. Meanwhile, the feature vector is constructed by only collecting the grid voltage, the grid current and the AC component of DC bus voltage. With the feature vector as input, the proposed method is able to accurately predict the health status of DC bus capacitor. Finally, we built a three-phase PWM rectifier as an experimental platform for validation. The experimental results verify that the proposed method fully utilizes the advantages of data-driven and model fusion, and achieves a high accuracy in the health state evaluation of DC bus capacitor. [ABSTRACT FROM AUTHOR]

We investigated the application of ensemble learning approaches in geotechnical stability analysis and proposed a compound explainable artificial intelligence (XAI) fitted to ensemble learning. 742 sets of data from real-world geotechnical engineering records are collected and six critical features that contribute to the stability analysis are selected. First, we visualized the data structure and examined the relationships between various features from both a statistical and an engineering standpoint. Seven state-of-the-art ensemble models and several classical machine learning models were compared and evaluated on slope stability prediction using real-world data. Further, we studied model fusion using the stacking strategy and the performance of model fusion that contributes to slope stability prediction. The results manifested that the ensemble learning model outperformed the classical single predictive models, with the CatBoost model yielding the most favourable results. To dive deeper into the credibility and explainability of CatBoost composed of multiple learners, the compound XAI fitted to CatBoost was formulated using feature importance, sensitivity analysis, and Shapley additive explanation (SHAP), which further strengthened the credibility of ensemble learning in geotechnical stability analysis. [ABSTRACT FROM AUTHOR]

Soluble solids content (SSC) is one of the main quality indicators of apples, and it is important to improve the precision of online SSC detection of whole apple fruit. Therefore, the spectral pre-processing method of spectral-to-spectral ratio (S/S), as well as multiple characteristic wavelength member model fusion (MCMF) and characteristic wavelength and non-characteristic wavelength member model fusion (CNCMF) methods, were proposed for improving the detection performance of apple whole fruit SSC by diffuse reflection (DR), diffuse transmission (DT) and full transmission (FT) spectra. The modeling analysis showed that the S/S- partial least squares regression models for all three mode spectra had high prediction performance. After competitive adaptive reweighted sampling characteristic wavelength screening, the prediction performance of all three model spectra was improved. The particle swarm optimization–extreme learning machine models of MCMF and CNCMF had the most significant enhancement effect and could make all three mode spectra have high prediction performance. DR, DT, and FT spectra all had some prediction ability for apple whole fruit SSC, with FT spectra having the strongest prediction ability, followed by DT spectra. This study is of great significance and value for improving the accuracy of the online detection model of apple whole fruit SSC. [ABSTRACT FROM AUTHOR]

Despite its wide applications in criminal investigations and clinical communications with patients suffering from autism, automatic micro-expression recognition remains a challenging problem because of the lack of training data and imbalanced classes problems. In this study, we proposed a meta-learning-based multimodel fusion network (Meta-MMFNet) to solve the existing problems. The proposed method is based on the metric-based meta-learning pipeline, which is specifically designed for few-shot learning and is suitable for model-level fusion. The frame difference and optical flow features were fused, deep features were extracted from the fused feature, and finally in themeta-learning-based framework, weighted summodel fusion method was applied for micro-expression classification. Meta-MMFNet achieved better results than state-of-the-art methods on four datasets. [ABSTRACT FROM AUTHOR]

The complexity of raw materials and multi-step purification processes presents considerable technical challenges in establishing universally applicable process parameters for the production of high-purity metals. Machine learning has emerged as an indispensable tool in the field of materials science, facilitating the accurate prediction of target variables and accelerating process optimization, thereby yielding substantial reductions in both experimental costs and time. This study explores the utilization of high-precision machine learning models to predict the residual impurity content in high-purity indium after vertical zone refining. A dataset comprising 82 experimental datasets was employed to determine the optimal hyperparameters for XGBoost and LightGBM models through Bayesian optimization. The XGBoost and LightGBM models demonstrated mean absolute errors (MAEs) of 0.022 and 0.023, respectively, as determined via leave-one-out cross-validation (LOOCV). Their comparable predictive performance to the previously established Ridge regression model (MAE = 0.024) prompted the exploration of fusion techniques, including mean, weighted, and stacking fusion, to further enhance accuracy. Remarkably, the weighted fusion model exhibited the most optimal predictive capabilities, supported by comprehensive evaluation metrics, including an MAE of 0.020, root mean squared error (RMSE) of 0.026, and a coefficient of determination (R2 score) of 0.830. Furthermore, the SHapley Additive exPlanations (SHAP) analysis revealed a significant correlation between lower initial arsenic (As) content and reduced total post-refining impurity levels in both the XGBoost and LightGBM models. This study underscores the precision of ensemble learning in predicting residual impurity content in vertically zone-refined indium products.

With the rapid development of information communication and mobile device technologies, smart devices have become increasingly popular, providing convenience to households and enhancing the level of intelligence in daily life. This trend is also driving innovation and progress in various fields, including healthcare, transportation, and industry. However, as technology continues to proliferate, network security concerns have become increasingly prominent, making the protection of digital life and data security an urgent priority. Intrusion detection has always played an important role in the field of network security. Traditional intrusion detection systems predominantly rely on anomaly detection technology to identify potential intrusions by detecting abnormal patterns in network traffic. With technological advancements, machine learning-based methods have emerged as the cornerstone of modern intrusion detection, enabling more precise identification of abnormal behaviors and potential intrusions by learning the patterns of normal network traffic. In response to these challenges, this paper introduces an innovative intrusion detection model that amalgamates the Attention-CNN-BiLSTM (ACBL) and Temporal Convolutional Network (TCN) architectures. The ACBL and TCN models excel in processing spatial and temporal features within network traffic data, respectively. This integration harnesses diverse neural network structures to elevate overall model performance and accuracy. Furthermore, a unique approach inspired by dung beetles’ natural behavior, incorporating Tent mapping-enhanced Dung Beetle Optimization Algorithm (TDBO), is leveraged for both optimizing feature selection parameters and searching for optimal model hyperparameters. The feature selection parameters obtained from TDBO are then combined with the importance ranking from the Random Forest algorithm, ensuring optimal features can be better selected to enhance model performance. This paper introduces a novel intrusion detection model, the TDBO-ACBLT model, and validates its performance using the UNSW-NW15 dataset. TDBO excels in feature selection compared to common algorithms and achieves superior parameter optimization accuracy over Harris’s Hawk Optimization (HHO), Particle Swarm Optimization (PSO), and Dung Beetle Optimization (DBO). The proposed model achieves higher accuracy than prevalent machine learning models.

In the era of digital communication, social media platforms have experienced exponential growth, becoming primary channels for information exchange. However, this surge has also amplified the rapid spread of hate speech, prompting extensive research efforts for effective mitigation. These efforts have prominently featured advanced natural language processing techniques, particularly emphasizing deep learning methods that have shown promising outcomes. This article presents a novel approach to address this pressing issue, combining a comprehensive dataset of 18 sources. It includes 0.45 million comments sourced from various digital platforms spanning different time frames. There were two models utilized to address the diversity in the data and leverage distinct strengths found within deep learning frameworks: CNN and BiLSTM with an attention mechanism. These models were tailored to handle specific subsets of the data, allowing for a more targeted approach. The unique outputs from both models were then fused into a unified model. This methodology outperformed recent models, showcasing enhanced generalization capabilities even when tested on the largest and most diverse dataset. Our model achieved an impressive accuracy of 89%, while maintaining a high precision of 0.88 and recall of 0.91.

Accurate categorization and timely control of leaf diseases are crucial for citrus growth. We proposed the Multi-Models Fusion Network (MMFN) for citrus leaf diseases detection based on model fusion and transfer learning. Compared to traditional methods, the algorithm (integrating transfer learning Alexnet, VGG, and Resnet) we proposed can address the issues of limited categories, slow processing speed, and low recognition accuracy. By constructing efficient deep learning models and training and optimizing them with a large dataset of citrus leaf images, we ensured the broad applicability and accuracy of citrus leaf disease detection, achieving high-precision classification. Herein, various deep learning algorithms, including original Alexnet, VGG, Resnet, and transfer learning versions Resnet34 (Pre_Resnet34) and Resnet50 (Pre_Resnet50) were also discussed and compared. The results demonstrated that the MMFN model achieved an average accuracy of 99.72% in distinguishing between diseased and healthy leaves. Additionally, the model attained an average accuracy of 98.68% in the classification of multiple diseases (citrus huanglongbing (HLB), greasy spot disease and citrus canker), insect pests (citrus leaf miner), and deficiency disease (zinc deficiency). These findings conclusively illustrate that deep learning model fusion networks combining transfer learning and integration algorithms can automatically extract image features, enhance the automation and accuracy of disease recognition, demonstrate the significant potential and application value in citrus leaf disease classification, and potentially drive the development of smart agriculture.

Aiming at the problem of the variety of plant leaf diseases and how to extract effective features, an attention network model fused with channel information is proposed to identify a variety of plant leaf diseases. Firstly, a residual structure based basic network is built for feature extraction, and in order to extract effective information, the feature is re-calibrated by integrating multiple channel information through the attention network. Then, the constraint information is added into the cross entropy function to accelerate the convergence of the model. Finally, the model is tested on the data sets of 16 diseases of four different plants. The results show that the recognition accuracy of the basic network model is 83.13%, while the accuracy increased by 4.64% after fusing the channel information network. Compared with other models, the fusion model improves the recognition accuracy by 9.72% and the model complexity is less than twice that of the optimal model in the comparison experiment. [ABSTRACT FROM AUTHOR]

Federated learning facilitates collaborative data analysis among multiple participants while preserving user privacy. However, conventional federated learning approaches, typically employing weighted average techniques for model fusion, confront two significant challenges: 1. The inclusion of malicious models in the fusion process can drastically undermine the accuracy of the aggregated global model. 2. Due to the heterogeneity problem of devices and data, the number of client samples does not determine the weight value of the model. To solve those challenges, we propose a trustworthy model fusion method based on reinforcement learning (FedDRL), which includes two stages. In the first stage, we propose a reliable client selection mechanism to exclude malicious models from the fusion process. In the second stage, we propose an adaptive model fusion method that dynamically assigns weights based on model quality to aggregate the best global models. Finally, we validate our approach against five distinct model fusion scenarios, demonstrating that our algorithm significantly enhanced reliability without compromising accuracy. [ABSTRACT FROM AUTHOR]

Navigation map generation based on remote sensing images is crucial in fields such as autonomous driving and geographic surveying. Style transfer is an effective method for obtaining a navigation map of the current environment. However, there is lack of robustness of the map-style transfer model, resulting in unsatisfactory quality of the generated navigation maps. To address these challenges, we average the parameters of generators sampled from different iterations with a dense sampling strategy in the Generative Adversarial Network (CycleGAN). The results demonstrate that the training efficiency of our method on the MNIST and generation quality on the Google Map dataset are significantly superior to traditional style transfer methods. Moreover, it performs well in multi-environment hybrid mapping. Our method improves the generalization ability of the model and converts existing navigation maps to other styles of maps precisely. It can better adapt to different types of urban layout and road planning, bringing innovative solutions for traffic management and navigation systems. [ABSTRACT FROM AUTHOR]

Massive amounts of data drive the performance of deep learning models, but in practice, data resources are often highly dispersed and bound by data privacy and security concerns, making it difficult for multiple data sources to share their local data directly. Data resources are difficult to aggregate effectively, resulting in a lack of support for model training. How to collaborate between data sources in order to aggregate the value of data resources is therefore an important research question. However, existing distributed-collaborative-learning architectures still face serious challenges in collaborating between nodes that lack mutual trust, with security and trust issues seriously affecting the confidence and willingness of data sources to participate in collaboration. Blockchain technology provides trusted distributed storage and computing, and combining it with collaboration between data sources to build trusted distributed-collaborative-learning architectures is an extremely valuable research direction for application. We propose a trusted distributed-collaborative-learning mechanism based on blockchain smart contracts. Firstly, the mechanism uses blockchain smart contracts to define and encapsulate collaborative behaviours, relationships and norms between distributed collaborative nodes. Secondly, we propose a model-fusion method based on feature fusion, which replaces the direct sharing of local data resources with distributed-model collaborative training and organises distributed data resources for distributed collaboration to improve model performance. Finally, in order to verify the trustworthiness and usability of the proposed mechanism, on the one hand, we implement formal modelling and verification of the smart contract by using Coloured Petri Net and prove that the mechanism satisfies the expected trustworthiness properties by verifying the formal model of the smart contract associated with the mechanism. On the other hand, the model-fusion method based on feature fusion is evaluated in different datasets and collaboration scenarios, while a typical collaborative-learning case is implemented for a comprehensive analysis and validation of the mechanism. The experimental results show that the proposed mechanism can provide a trusted and fair collaboration infrastructure for distributed-collaboration nodes that lack mutual trust and organise decentralised data resources for collaborative model training to develop effective global models. [ABSTRACT FROM AUTHOR]