Neural Network Fusion Research Articles

Small-area marks realize nanoscale lithography alignment by spatial and frequency domain fusion neural network

Small-area marks realize nanoscale lithography alignment by spatial and frequency domain fusion neural network

Attention mechanism fusion neural network for typhoon path prediction

Attention mechanism fusion neural network for typhoon path prediction

A Hybrid Model for Soybean Yield Prediction Integrating Convolutional Neural Networks, Recurrent Neural Networks, and Graph Convolutional Networks

Soybean yield prediction is one of the most critical activities for increasing agricultural productivity and ensuring food security. Traditional models often underestimate yields because of limitations associated with single data sources and simplistic model architectures. These prevent complex, multifaceted factors influencing crop growth and yield from being captured. In this line, this work fuses multi-source data—satellite imagery, weather data, and soil properties—through the approach of multi-modal fusion using Convolutional Neural Networks and Recurrent Neural Networks. While satellite imagery provides information on spatial data regarding crop health, weather data provides temporal insights, and the soil properties provide important fertility information. Fusing these heterogeneous data sources embeds an overall understanding of yield-determining factors in the model, decreasing the RMSE by 15% and improving R2 by 20% over single-source models. We further push the frontier of feature engineering by using Temporal Convolutional Networks (TCNs) and Graph Convolutional Networks (GCNs) to capture time series trends, geographic and topological information, and pest/disease incidence. TCNs can capture long-range temporal dependencies well, while the GCN model has complex spatial relationships and enhanced the features for making yield predictions. This increases the prediction accuracy by 10% and boosts the F1 score for low-yield area identification by 5%. Additionally, we introduce other improved model architectures: a custom UNet with attention mechanisms, Heterogeneous Graph Neural Networks (HGNNs), and Variational Auto-encoders. The attention mechanism enables more effective spatial feature encoding by focusing on critical image regions, while the HGNN captures interaction patterns that are complex between diverse data types. Finally, VAEs can generate robust feature representation. Such state-of-the-art architectures could then achieve an MAE improvement of 12%, while R2 for yield prediction improves by 25%. In this paper, the state of the art in yield prediction has been advanced due to the employment of multi-source data fusion, sophisticated feature engineering, and advanced neural network architectures. This provides a more accurate and reliable soybean yield forecast. Thus, the fusion of Convolutional Neural Networks with Recurrent Neural Networks and Graph Networks enhances the efficiency of the detection process.

Transformer–BiLSTM Fusion Neural Network for Short-Term PV Output Prediction Based on NRBO Algorithm and VMD

In order to solve the difficulties that the uncertain characteristics of PV output, such as volatility and intermittency, will bring to the development of microgrid scheduling plans, this paper proposes a Transformer–Bidirectional Long Short-Term Memory (BiLSTM) neural network PV power generation forecasting fusion model based on the Newton–Raphson optimization algorithm (NRBO) and Variational Modal Decomposition (VMD). Firstly, the principle of the VMD technique and the Gray Wolf Optimization (GWO) algorithm’s key parameter optimization method for VMD are introduced. Then, the Transformer decoder partially fuses the BiLSTM network and retains the encoder to obtain the body of the prediction model, followed by explaining the principle of the NRBO algorithm. And finally, the VMD-NRBO-Transformer-BiLSTM prediction model and hyperparameter selection are evaluated by the NRBO algorithm. The algorithm sets up a multi-model comparison experiment, and the results show that the prediction model proposed in this paper has the best prediction accuracy and the optimal evaluation index.

Symbiotic design and numerical methods of industrial architecture and urban landscape in the context of sustainable development

The status quo of low sustainability of urban landscape symbiosis design programme and low referability of the corresponding comprehensive evaluation method, the study will take the environmental and ecological value as the symbiosis goal to design the urban landscape. The study constructs a comprehensive evaluation model of industrial heritage urban landscape ecology, researches each index affecting landscape ecology, and determines the weight of each influence index by combining qualitative and quantitative methods. The maximum influence factors of the four criteria layers of emotion, culture, material and ecology are pleasure, obvious cultural symbols, appropriate scale and microclimate regulation, and the corresponding values are 0.313, 0.404, 0.315 and 0.495, respectively. Compared with that before the introduction of the self-organised neural network fusion K-Means clustering algorithm, the ecological evaluation scores of the urban landscape of the industrial heritage are much closer to the real situation, and the overall error is much higher. The ecological evaluation model of industrial heritage buildings is closer to the real situation, and the overall error fluctuation is 0.012–0.020. The precision-recall curve of the comprehensive evaluation model of industrial heritage buildings is more advantageous than that of other comprehensive evaluation models. The results of the study provide a scientific reference basis for the design of the external landscape environment of industrial heritage buildings, and provide a strong scientific support for the ecological symbiosis between old industrial buildings and the environment under the concept of sustainability.

Towards Explainable Detection of Alzheimer's Disease: A Fusion of Deep Convolutional Neural Network and Enhanced Weighted Fuzzy C-Mean.

Alzheimer's disease (AD) is a progressive neurodegenerative disorder characterized by cognitive decline, posing a significant challenge for individuals and society. Early detection and treatment are essential for effective disease management. The objective of this research is to develop a novel and interpretable deep learning model for rapid and accurate Alzheimer's disease detection, incorporating Explainable Artificial Intelligence (XAI) techniques. The model aims to ensure generalizability through cross-validation and data augmentation, while enhancing interpretability and transparency by using Explainable Artificial Intelligence methods such as Grad-CAM, SHAP, and LIME, alongside an Enhanced Fuzzy C-Means (FCM) algorithm to clarify feature categorization and improve understanding of the model's decision-making process. The proposed model employs a multi-stage approach. Initially, MRI scans are transformed into feature vectors suitable for input into a Deep Convolutional Neural Network (CNN). Subsequently, an Enhanced Fuzzy C-Mean (FCM) algorithm, incorporating spatial information, refines these features to improve clustering precision. The model integrates Explainable Artificial Intelligence techniques, including Grad-CAM, SHAP, and LIME, to elucidate critical features and regions influencing classification outcomes. The performance metrics such as Accuracy, Recall and Specificity are used for assessing the performance of the model. The XAI-DEF Alzheimer's disease detection model consistently demonstrated exceptional performance across both the ADNI and OASIS datasets. On ADNI, the model achieved an accuracy of 99.39%, recall of 99.47%, and specificity of 99.3%. Similarly, on OASIS, the model attained an accuracy of 99.36%, recall of 99.53%, and specificity of 99.15%. These results underscore the model's effectiveness in accurately classifying Alzheimer's disease cases while minimizing false positives and negatives. Through the development of this model, we contribute to the advancement of dependable diagnostic tools tailored for the detection and management of Alzheimer's disease. By prioritizing interpretability alongside accuracy, our approach provides valuable insights into the decisionmaking process of the model, ultimately improving patient outcomes and facilitating further research in neurodegenerative disorders.

Aerodynamic Angle Estimation in Rotating Projectiles: A NBCA Fusion Approach

Abstract The angle of attack (AOA) is a crucial parameter for describing the flight state of a projectile and is one of the key elements in external ballistic testing. In light of the inherent difficulty and substantial cost associated with the direct detection of the AOA during the flight of guided projectiles, this study introduces an estimation method for the AOA that is predicated on a neuro-behavioural cognitive architecture (NBCA) neural network fusion model. This approach leverages data from geomagnetic sensors to calculate the geomagnetic azimuth angle related to the AOA. Subsequently, by fitting the geomagnetic azimuth angle and angle of attack-related quantities into the fusion model and training the neural network, the angle of attack is estimated. Research results indicate that the root mean square error of the angle of attack estimation is 0.0149°. This method achieves optimal estimation of the projectile’s angle of attack without the need for extensive detection equipment, providing a novel perspective for the practical engineering application of projectile angle of attack detection.

Uncertainty quantification analysis using multi-fidelity deep neural network

Abstract In order to solve the highly-dimensional and highly-nonlinear problems in uncertainty quantification of CFD, the paper develops a multi-fidelity neural network model and proposes a residual connection-based multi-fidelity deep neural network fusion model. Based on the small sample size and low computational cost of multi-fidelity models, the residual structure is further utilized to accelerate the convergence rate of the model and reduce the possibility of inducing systemic risks. Finally, the feasibility and accuracy of the algorithm are verified by six typical numerical examples, and the uncertainty quantification analysis of the NACA0012 airfoil’s yaw moment coefficients is completed.

Exploration of the Application of Multimodal Model in Psychological Analysis

Multimodal sentiment analysis is one of the important research areas in the field of artificial intelligence today. Multimodal sentiment analysis is to extract features from various human modalities such as facial expressions, body movements, and voice information, perform modal fusion, and finally classify and predict emotions. This technology can be used in multiple scenarios such as stock prediction, product analysis, movie box office prediction, etc., especially psychological state analysis, and has important research significance. This paper introduces two important datasets in multimodal sentiment analysis, namely CMU-MOSEI and IEMOCAP. It also introduces the feature-level fusion, model-level fusion, decision-level fusion and other fusion methods in multimodal fusion methods, and also introduces the semantic feature fusion neural network and sentiment word perception fusion network in multimodal sentiment analysis related models. Finally, the application of multimodal sentiment analysis models in depression and other related mental illnesses and the challenges of multimodal sentiment analysis models in the future are introduced. This paper hopes that the above research will be helpful for multimodal sentiment analysis.

Resource-efficient artificial intelligence for battery capacity estimation using convolutional FlashAttention fusion networks

Accurate battery capacity estimation is crucial for optimizing lifespan and monitoring health conditions. Deep learning has made notable strides in addressing long-standing issues in the artificial intelligence community. However, large AI models often face challenges such as high computational resource consumption, extended training times, and elevated deployment costs. To address these issues, we developed an efficient end-to-end hybrid fusion neural network model. This model combines FlashAttention-2 with local feature extraction through convolutional neural networks (CNNs), significantly reducing memory usage and computational demands while maintaining precise and efficient health estimation. For practical implementation, the model uses only basic parameters, such as voltage and charge, and employs partial charging data (from 80 % SOC to the upper limit voltage) as features, without requiring complex feature engineering. We evaluated the model using three datasets: 77 lithium iron phosphate (LFP) cells, 16 nickel cobalt aluminum (NCA) cells, and 50 nickel cobalt manganese (NCM) oxide cells. For LFP battery health estimation, the model achieved a root mean square error of 0.109 %, a coefficient of determination of 0.99, and a mean absolute percentage error of 0.096 %. Moreover, the proposed convolutional and flash-attention fusion networks deliver an average inference time of 57 milliseconds for health diagnosis across the full battery life cycle (approximately 1898 cycles per cell). The resource-efficient AI (REAI) model operates at an average of 1.36 billion floating point operations per second (FLOPs), with GPU power consumption of 17W and memory usage of 403 MB. This significantly outperforms the Transformer model with vanilla attention. Furthermore, the multi-fusion model proved to be a powerful tool for evaluating capacity in NCA and NCM cells using transfer learning. The results emphasize its ability to reduce computational complexity, energy consumption, and memory usage, while maintaining high accuracy and robust generalization capabilities.

Identification of pregnancy in women based on fingertip pulse using a multi-feature fusion neural network model

This study proposes a rapid method for determining pregnancy status based on fingertip pulse signals. A finger pulse sensor collects data, which is processed into unified multimodal signals. The Bamboo-Net model, combining ResNet, LSTM, and 1D-CNN, extracts key features from time, frequency, and time-frequency domains. Tested on 346 training and 138 testing samples, the model achieves 91% accuracy with 6 s input, outperforming mainstream methods. Recognition rates for mid and late pregnancy are higher than for early pregnancy, highlighting its potential for practical applications.

Music Sequence Creation Technique in Piano Music Production Based on Embedded Neural Network

With the rapid development of computer technology and artificial intelligence, especially the rise of embedded neural network technology. The traditional arrangement methods are facing increasing challenges in the field of music production, and their complexity and time-consuming nature are no longer sufficient to meet the urgent needs of modern audiences for efficient and high-quality computer piano music works. This paper proposes an innovative solution — a piano music production matching method based on embedded neural network fusion, and particularly focused on intelligent creation technology for music sequences. In the system, embedded neural networks are core applied in the process of creating music sequences. By utilizing their powerful feature learning and generalization abilities, neural networks can capture complex patterns and dynamic changes in music sequences, providing infinite possibilities for melody generation and innovation. At the same time, combined with association rule algorithms, the system can discover potential association patterns between notes, further optimizing the fluency and harmony of the melody; The clustering algorithm helps extract common features from massive music data, providing a rich material library for melody innovation.

Dynamic evidence fusion neural networks with uncertainty theory and its application in brain network analysis

Deep learning has demonstrated significant potential and advantages, achieving notable success in the medical field, particularly in the application of brain network analysis. However, most models ignore the uncertainty caused by inconsistent view quality and fail to effectively leverage the potential correlations and temporal sequences present in multi-view data, preventing neural networks from fully showcasing their strengths. To this end, this paper proposes dynamic evidence fusion neural networks (DEF-NNs) with uncertainty theory, and applies it to brain network analysis. Our model is established within a multi-view learning framework that considers the functional connections under each window as a view. We employ a dynamic evidence learning module to capture the evidence for each time window of the dynamic brain network, utilizing three distinct convolutional filters to extract feature maps. Then, a dynamic evidence fusion mechanism is designed and a dynamic trust function is constructed according to the temporal nature of dFC data. The evidence generated by multiple windows is fused at the decision level of classification, dealing with the uncertainty caused by inconsistent view quality and improving the classification performance. We verified the effectiveness of DEF-NNs through comparison with advanced algorithms on three schizophrenia datasets, and the results show that DEF-NNs significantly improved the classification performance of brain disease diagnosis tasks.

Characterizing Subsurface Structures From Hard and Soft Data With Multiple‐Condition Fusion Neural Network

AbstractAccurately inferring realistic subsurface structures poses a considerable challenge due to the impact of morphology on flow and transport behaviors. Traditional subsurface characterization relies on two primary types of data: hard data, derived from direct subsurface measurements, and soft data, encompassing remotely sensed geophysical information and its interpretation. Existing deep‐learning‐based methodologies predominantly focus on the transition from multiple observations to subsurface structures. However, implicit non‐linear correlations among diverse data sources often remain underutilized, leading to potential bias and errors. In this study, we introduce a multiple‐condition fusion network (MCF‐Net) to characterize subsurface structures based on both hard and soft data. To harness the full potential of multiple‐source subsurface observations, two distinct neural networks extract implicit features from hard and soft data. The integration of these features is achieved through multiple‐condition fusion blocks, designed to capture representative characteristics. These blocks are also adept at reconstructing heterogeneous structures and facilitating hydrological parameterization. MCF‐Net exhibits accuracy in estimating subsurface structures across various types of subsurface observations. Experimental results underscore the utility and superiority of MCF‐Net in applications of hydrogeological modeling.

Joint image dehazing and denoising for single haze image enhancement

AbstractOutdoor haze images are typically degraded by noise due to the external environment and imaging equipment. The existing haze image enhancement methods ignore the interrelation between haze and noise, which cannot suppress the noise and remove the haze simultaneously. To address these intractable problems, a dual‐branch architecture that combines dehazing and denoising is proposed here to restore clear images. First, dark channel prior and unsupervised networks in the image dehazing branch to remove the image blur are adopted. Then, the image denoising branch removes the image noise in parallel by constructing a mean/extreme sampler and a self‐supervised network. Finally, a convolutional neural network fusion strategy is presented to fuse output images from the aforementioned two branches to generate the final qualified results. Extensive experiments reveal that the proposed haze image enhancement method outperforms other state‐of‐the‐art methods in terms of peak signal‐to‐noise ratio (PSNR) and structural similarity index measure (SSIM).

A new fusion neural network model and credit card fraud identification.

Credit card fraud identification is an important issue in risk prevention and control for banks and financial institutions. In order to establish an efficient credit card fraud identification model, this article studied the relevant factors that affect fraud identification. A credit card fraud identification model based on neural networks was constructed, and in-depth discussions and research were conducted. First, the layers of neural networks were deepened to improve the prediction accuracy of the model; second, this paper increase the hidden layer width of the neural network to improve the prediction accuracy of the model. This article proposes a new fusion neural network model by combining deep neural networks and wide neural networks, and applies the model to credit card fraud identification. The characteristic of this model is that the accuracy of prediction and F1 score are relatively high. Finally, use the random gradient descent method to train the model. On the test set, the proposed method has an accuracy of 96.44% and an F1 value of 96.17%, demonstrating good fraud recognition performance. After comparison, the method proposed in this paper is superior to machine learning models, ensemble learning models, and deep learning models.

Power substation load forecasting using interpretable transformer-based temporal fusion neural networks

Demand forecasting is one of the biggest challenges facing the power system, given that anticipating future consumption is fundamental to guarantee an adequate supply, without wasting resources or overloading the power grid. Despite numerous techniques developed for this purpose, the unstable nature of historical data makes electricity demand time series highly complex, making forecasting a critical and challenging problem. This work develops and explores the predictive potential of the temporal fusion transformer (TFT), an innovative approach capable of performing temporal fusions and forecasts over different time horizons. This architecture features interpretability capabilities, allowing the understanding of which variables and patterns are most relevant to predictions. It can help identify problems, provide insights and improve system reliability. The efficiency of the proposed model is assessed using historical data on the overall load of the New Zealand Electrical Company, which is made up of two thermoelectric power stations and seven substations. The results highlight the TFT as a promising approach since it shows significant indicators in the forecasts, including the mean absolute percentage error of less than 1.5% along 48 h in advance. The obtained performance metrics underscore the accuracy and robustness of the TFT architecture.

Fusion Technology-Based CNN-LSTM-ASAN for RUL Estimation of Lithium-Ion Batteries

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) not only prevents battery system failure but also promotes the sustainable development of the energy storage industry and solves the pressing problems of industrial and energy crises. Because of the capacity regeneration phenomenon and random interference during the operation of lithium-ion batteries, the prediction precision and generalization performance of a single model can be poor. This article proposes a novel RUL prediction based on data pre-processing methods and the CNN-LSTM-ASAN framework. The model is based on a fusion technique for optimizing the tandem fusion of the Convolutional Neural Network (CNN) and the Long Short-Term Memory Network (LSTM). Firstly, the improved adaptive noise fully integrates empirical mode decomposition (ICEEMDAN) and the Pearson correlation coefficient (PCC), which are used to estimate the global deterioration tendency component and the local capacity restoration component, to reconstruct the dataset and eliminate the noise. Then, the Adaptive Sparse Attention Network (ASAN) is added in the model construction stage to improve the training efficiency of the model. The reconstructed degraded data are features extracted by the CNN-LSTM-ASAN model. Finally, the proposed method is validated against models such as DCLA, using the NASA public datasets, the CALCE public datasets, and the self-use datasets. And the results show that the root mean square error (RMSE) of the model is below 1.5%.

Multi-Scale and Multi-Network Deep Feature Fusion for Discriminative Scene Classification of High-Resolution Remote Sensing Images

The advancement in satellite image sensors has enabled the acquisition of high-resolution remote sensing (HRRS) images. However, interpreting these images accurately and obtaining the computational power needed to do so is challenging due to the complexity involved. This manuscript proposed a multi-stream convolutional neural network (CNN) fusion framework that involves multi-scale and multi-CNN integration for HRRS image recognition. The pre-trained CNNs were used to learn and extract semantic features from multi-scale HRRS images. Feature extraction using pre-trained CNNs is more efficient than training a CNN from scratch or fine-tuning a CNN. Discriminative canonical correlation analysis (DCCA) was used to fuse deep features extracted across CNNs and image scales. DCCA reduced the dimension of the features extracted from CNNs while providing a discriminative representation by maximizing the within-class correlation and minimizing the between-class correlation. The proposed model has been evaluated on NWPU-RESISC45 and UC Merced datasets. The accuracy associated with DCCA was 10% and 6% higher than discriminant correlation analysis (DCA) in the NWPU-RESISC45 and UC Merced datasets. The advantage of DCCA was better demonstrated in the NWPU-RESISC45 dataset due to the incorporation of richer within-class variability in this dataset. While both DCA and DCCA minimize between-class correlation, only DCCA maximizes the within-class correlation and, therefore, attains better accuracy. The proposed framework achieved higher accuracy than all state-of-the-art frameworks involving unsupervised learning and pre-trained CNNs and 2–3% higher than the majority of fine-tuned CNNs. The proposed framework offers computational time advantages, requiring only 13 s for training in NWPU-RESISC45, compared to a day for fine-tuning the existing CNNs. Thus, the proposed framework achieves a favourable balance between efficiency and accuracy in HRRS image recognition.

Artificial intelligence based leak detection in blended hydrogen and natural gas pipelines

In the transition towards a hydrogen-based energy system, the strategic use of pipelines is crucial for efficient hydrogen distribution. Leveraging existing natural gas pipelines to carry a mix of hydrogen and natural gas offers a cost-effective alternative to building new infrastructure. This study explores the development of a leak detection system compatible with existing pipelines, specifically tailored for the blended hydrogen and natural gas mix. Given the scarcity of leak data for blended hydrogen-natural gas pipelines, the study introduces a Real-Time Transient Model (RTTM) for blended gases, simulating leak dynamics to generate necessary data. Additionally, a leak detection system (LDS) is developed using a fusion of Convolutional Neural Network (CNN) and Explainable Artificial Intelligence (XAI) through Adaptive Neuro-Fuzzy Inference Systems (ANFIS). This LDS framework overcomes the “black box” issue common in AI-driven systems, enabling reliable detection. The integration of Explainable and traditional AI techniques holds promising implications for blended hydrogen pipelines by enhancing the safety and efficiency of hydrogen transportation, thereby mitigating economic and environmental impacts, and addressing public concerns.