Neural Network-based Framework Research Articles

Multidimensional Financial Metrics for Corporate Financial Risk Assessment and Early Warning Mechanisms

By constructing a comprehensive multi-dimensional financial index evaluation system, this study effectively identifies, evaluates, and forewarns the financial risks of enterprises. Utilizing public financial data from S&P Global and Alpha Vantage, a BP neural network-based enterprise financial risk assessment framework is built, and the KMO method analyzes different risk factors. This comprehensive evaluation model, paired with a risk warning mechanism, assesses financial situations through financial leverage, profitability, solvency, operating efficiency, growth indicators, and cash flow indicators. The example analysis shows: 1) The RMSE value ranges from 46.71 to 64.94, indicating high accuracy and stability in predicting financial risk; 2) The R value is very close to 1, demonstrating high stability.

A deep learning-based method for the detection and segmentation of breast masses in ultrasound images

Objective. Automated detection and segmentation of breast masses in ultrasound images are critical for breast cancer diagnosis, but remain challenging due to limited image quality and complex breast tissues. This study aims to develop a deep learning-based method that enables accurate breast mass detection and segmentation in ultrasound images. Approach. A novel convolutional neural network-based framework that combines the You Only Look Once (YOLO) v5 network and the Global-Local (GOLO) strategy was developed. First, YOLOv5 was applied to locate the mass regions of interest (ROIs). Second, a Global Local-Connected Multi-Scale Selection (GOLO-CMSS) network was developed to segment the masses. The GOLO-CMSS operated on both the entire images globally and mass ROIs locally, and then integrated the two branches for a final segmentation output. Particularly, in global branch, CMSS applied Multi-Scale Selection (MSS) modules to automatically adjust the receptive fields, and Multi-Input (MLI) modules to enable fusion of shallow and deep features at different resolutions. The USTC dataset containing 28 477 breast ultrasound images was collected for training and test. The proposed method was also tested on three public datasets, UDIAT, BUSI and TUH. The segmentation performance of GOLO-CMSS was compared with other networks and three experienced radiologists. Main results. YOLOv5 outperformed other detection models with average precisions of 99.41%, 95.15%, 93.69% and 96.42% on the USTC, UDIAT, BUSI and TUH datasets, respectively. The proposed GOLO-CMSS showed superior segmentation performance over other state-of-the-art networks, with Dice similarity coefficients (DSCs) of 93.19%, 88.56%, 87.58% and 90.37% on the USTC, UDIAT, BUSI and TUH datasets, respectively. The mean DSC between GOLO-CMSS and each radiologist was significantly better than that between radiologists (p < 0.001). Significance. Our proposed method can accurately detect and segment breast masses with a decent performance comparable to radiologists, highlighting its great potential for clinical implementation in breast ultrasound examination.

A deep learning-based method for the detection and segmentation of breast masses in ultrasound images.

Automated detection and segmentation of breast masses in ultrasound images are critical for breast cancer diagnosis, but remain challenging due to limited image quality and complex breast tissues. This study aims to develop a deep learning-based method that enables accurate breast mass detection and segmentation in ultrasound images.&#xD;Approach. A novel convolutional neural network-based framework that combines the You Only Look Once (YOLO) v5 network and the Global-Local (GOLO) strategy was developed. First, YOLOv5 was applied to locate the mass regions of interest (ROIs). Second, a Global Local-Connected Multi-Scale Selection (GOLO-CMSS) network was developed to segment the masses. The GOLO-CMSS operated on both the entire images globally and mass ROIs locally, and then integrated the two branches for a final segmentation output. Particularly, in global branch, CMSS applied Multi-Scale Selection (MSS) modules to automatically adjust the receptive fields, and Multi-Input (MLI) modules to enable fusion of shallow and deep features at different resolutions. The USTC dataset containing 28,477 breast ultrasound images was collected for training and test. The proposed method was also tested on three public datasets, UDIAT, BUSI and TUH. The segmentation performance of GOLO-CMSS was compared with others networks and three experienced radiologists.&#xD;Main results. YOLOv5 outperformed other detection models with average precisions of 99.41%, 95.15%, 93.69% and 96.42% on the USTC, UDIAT, BUSI and TUH datasets, respectively. The proposed GOLO-CMSS showed superior segmentation performance over other state-of-the-art networks, with Dice similarity coefficients (DSCs) of 93.19%, 88.56%, 87.58% and 90.37% on the USTC, UDIAT, BUSI and TUH datasets, respectively. The mean DSC between GOLO-CMSS and each radiologist was significantly better than that between radiologists (p < 0.001).&#xD;Significance. Our proposed method can accurately detect and segment breast masses with a decent performance comparable to radiologists, highlighting its great potential for clinical implementation in breast ultrasound examination.

Multivariate Attention-Based Orbit Uncertainty Propagation and Orbit Determination Method for Earth–Jupiter Transfer

Current orbit uncertainty propagation (OUP) and orbit determination (OD) methods suffer from drawbacks related to high computational burden, limiting their applications in deep space missions. To this end, this paper proposes a multivariate attention-based method for efficient OUP and OD of Earth–Jupiter transfer. First, a neural network-based OD framework is utilized, in which the orbit propagation process in a traditional unscented transform (UT) and unscented Kalman filter (UKF) is replaced by the neural network. Then, the sample structure of training the neural network for the Earth–Jupiter transfer is discussed and designed. In addition, a method for efficiently generating a large number of samples for the Earth–Jupiter transfer is presented. Next, a multivariate attention-based neural network (MANN) is designed for orbit propagation, which shows better capacity in terms of accuracy and generalization than the deep neural network. Finally, the proposed method is successfully applied to solve the OD problem in an Earth–Jupiter transfer. Simulations show that the proposed method can obtain a similar estimation to the UKF while saving more than 90% of the computational cost.

Multi-head CNN-based activity recognition and its application on chest-mounted sensor-belt

In recent years, a great deal of research has been done on the identification, monitoring, and classification of human activities. Human activity recognition (HAR) is a term commonly used to describe the automatic identification of physical activities. For activity recognition, there are primarily vision-based and sensor-based methods available. The computer vision-based method is generally effective in lab settings, but because of clutter, fluctuating light levels, and contrast, it may not perform well in real-world scenarios. Continuous monitoring and analysis of physiological signals obtained from heterogeneous sensors attached to an individual’s body is required to realise sensor-based HAR systems. Most of the previous research in human activity recognition (HAR) is biased along with feature engineering and pre-processing which requires a good amount of domain knowledge. Application-specific modelling and time-taking methods are involved in these approaches. In this work, the multi-head convolutional neural network-based human activity recognition framework is proposed where automatic feature extraction and classification are involved in the form of an end-to-end classification approach. Experiments of this approach are performed by taking raw wearable sensor data with few pre-processing steps and without the involvement of a handcrafted feature extraction technique. 99.23% and 93.55% accuracy are obtained on the WISDM and UCI-HAR datasets which denoted the much improvement in the assessment of HAR over other similar approaches. The model is also tested on locally collected data from a chest mounted belt with fabric sensors and an accuracy of 87.14% has been achieved on that data.

A neural network-based algorithm for the reconstruction and filtering of single particle trajectory in magnetic particle tracking.

Magnetic particle tracking (MPT) is a recently developed non-invasive measurement technique that has gained popularity for studying dense particulate or granular flows. This method involves tracking the trajectory of a magnetically labeled particle, the field of which is modeled as a dipole. The nature of this method allows it to be used in opaque environments, which can be highly beneficial for the measurement of dense particle dynamics. However, since the magnetic field of the particle used is weak, the signal-to-noise ratio is usually low. The noise from the measuring devices contaminates the reconstruction of the magnetic tracer's trajectory. A filter is then needed to reduce the noise in the final trajectory results. In this work, we present a neural network-based framework for MPT trajectory reconstruction and filtering, which yields accurate results and operates at very high speed. The reconstruction derived from this framework is compared to the state-of-the-art extended Kalman filter-based reconstruction.

A Chebyshev neural network-based numerical scheme to solve distributed-order fractional differential equations

This study aims to develop a first-order Chebyshev neural network-based technique for solving ordinary and partial distributed-order fractional differential equations. The neural network is used as a trial solution to construct the loss function. The loss function is utilized to train the neural network via an extreme learning machine and obtain the solution. The novelty of this work is developing and implementing a neural network-based framework for distributed-order fractional differential equations via an extreme learning machine. The proposed method is validated on several test problems. The error metrics utilized in the study include the absolute error and the L2 error. A comparison with other previously available approaches is presented. Also, we provide the computation time of the method.

Deep transfer learning strategy in intelligent fault diagnosis of gas turbines based on the Koopman operator

The gas turbine engine is a predominant prime mover in the transport and energy sectors, and ensuring its reliable operation holds paramount significance. While intelligent fault diagnosis (FD) approaches have seen successful advancements within the gas turbine FD landscape, many existing methods operate under the assumption of identical health states during both data collection and the FD process. Moreover, most previous studies have overlooked the diagnosis of both sensors and actuators. Another critical challenge lies in isolating simultaneous and multiple faults and providing compromising FD performance, especially in the face of continued system performance degradation. Aiming at these problems, this study develops a novel unsupervised data-driven FD strategy based on leveraging the potential of transfer learning and the Koopman operator. A deep neural network-based transfer learning framework is proposed for realizing a precise adaptive linear model called the deep transfer linear (DTL) model enabling reliable prediction of the system’s behavior in various situations and designing structured fault residuals. To this end, a deep neural network framework is used to obtain a precise Koopman model realization using the richly collected data in the source domain. Subsequently, the realized model is fine-tuned for the target domains associated with the degraded system, mitigating the adverse effects of domain shift and addressing the rich data scarcity problem in the target domain. In addition, the dedicated and generalized residual sets are designed and generated employing the geometric approach for fault isolation and a decision-making analysis is developed to diagnose simultaneous faults. The reliability of the proposed strategy is demonstrated through various experiments in the presence of noise and performance degradation, and a comparative performance analysis is conducted between the proposed strategy and another data-driven method showcasing the superiority of the proposed approach.

Deep calibration with random grids

We propose a neural network-based approach to calibrating stochastic volatility models, which combines the pioneering grid approach by Horvath et al. [Deep learning volatility: A deep neural network perspective on pricing and calibration in (rough) volatility models. Quant. Finance, 2021, 21(1), 11–27]. with the pointwise two-stage calibration of Bayer and Stemper [Deep calibration of rough stochastic volatility models. Working Paper, arXiv:1810.03399, 2018] and Liu et al. [A neural network-based framework for financial model calibration. J. Math. Ind., 2019, 9(1), 1–28]. Our methodology inherits robustness from the former while not suffering from the need for interpolation/extrapolation techniques, a clear advantage ensured by the pointwise approach. The crucial point to the entire procedure is the generation of implied volatility surfaces on random grids, which one dispenses to the network in the training phase. We support the validity of our calibration technique with several empirical and Monte Carlo experiments for the rough Bergomi and Heston models under a simple but effective parametrization of the forward variance curve. The approach paves the way for valuable applications in financial engineering—for instance, pricing under local stochastic volatility models—and extensions to the fast-growing field of path-dependent volatility models.

A physics-informed neural network framework to investigate nonlinear and heterogenous shrinkage of drying plant cells

This paper introduces a novel Physics-Informed Neural Network-based (PINN-based) multi-domain computational framework to analyse nonlinear and heterogeneous morphological variations of plant cells during drying. Here, two distinct models are involved: PINN-MT to simulate mass transfer; and PINNNS to simulate nonlinear shrinkage. The models are coupled to examine cellular morphological changes resulting from moisture loss during drying. Firstly, the coupled framework, in tandem with homogeneous conditions, operates in parallel, allowing the mutual parameters to update between models. This approach demonstrates ability to approximate homogeneous cellular shrinkage within a tissue, factoring in the influence of surrounding plant cells. Secondly, non-uniform cell wall properties and heterogeneous boundary conditions are incorporated into this computational framework through domain decomposition. Inherent capabilities of neural networks allow for seamless integration of multiple domains, with additional loss terms introduced at interfaces. The framework shows capacity to account for drastic and non-uniform morphological variations of plant cells even under extreme drying conditions, which is the key novelty and has been a challenging task for existing traditional computational methods. Hence, the proposed computational approach offers an innovative avenue for understanding nonlinear and heterogeneous morphological variations not only for plant cells, but also for soft matter in general.

PM2ECGCN: Parallelized spatial-temporal structures of multi-lead ECG with graph convolution network for multi-center cardiac disease diagnosis

Cardiovascular diseases (CVDs) are the leading global cause of death. As first-line diagnostic tool for CVDs, electrocardiograms (ECG) non-invasively assess cardiac condition across all clinical levels. Despite the strides made in automatic ECG analysis through artificial intelligence, current methods fail to utilize spatial–temporal interactions of multi-lead ECG, address multi-center data heterogeneity, and adapt to data security and computational constraints. This study proposed a graph neural network-based framework named PM2ECGCN responding to these challenges of multi-lead ECG from multi-center sources. PM2ECGCN independently extracts spatial–temporal dependencies with learning-based featurization on each isolated subset, and trains collaboratively under feature-sharing strategy that bridging isolated outcomes without exposing raw data. Benefiting from information fusion capacities of graph convolutional network with attention mechanism, PM2ECGCN enhances explicit topological representations of multi-lead ECG via parallelized cardiac graph structures, and fast spectral pooling provides linear inference complexity. PM2ECGCN promotes condition-invariant knowledge learning with domain adaptation for tackling the data heterogeneity of distrubution. Extensive experiments, conducted with various configurations under diverse scenarios, involve two public datasets, PTBXL and ICBEB, along with an external dataset from day-to-day clinical practice we collected. Empirical evaluations showcase outperformance over cutting-edge strategies for multi-center CVD diagnosis with consistent improvements, demonstrating potential for real-world clinical deployment.

A neural network-based model framework for cell-fate decisions and development

Gene regulatory networks (GRNs) fulfill the essential function of maintaining the stability of cellular differentiation states by sustaining lineage-specific gene expression, while driving the progression of development. However, accounting for the relative stability of intermediate differentiation stages and their divergent trajectories remains a major challenge for models of developmental biology. Here, we develop an empirical data-based associative GRN model (AGRN) in which regulatory networks store multilineage stage-specific gene expression profiles as associative memory patterns. These networks are capable of responding to multiple instructive signals and, depending on signal timing and identity, can dynamically drive the differentiation of multipotent cells toward different cell state attractors. The AGRN dynamics can thus generate diverse lineage-committed cell populations in a robust yet flexible manner, providing an attractor-based explanation for signal-driven cell fate decisions during differentiation and offering a readily generalizable modelling tool that can be applied to a wide variety of cell specification systems.

A novel neural network-based framework to estimate oil and gas pipelines life with missing input parameters

Dry gas pipelines can encounter various operational, technical, and environmental issues, such as corrosion, leaks, spills, restrictions, and cyber threats. To address these difficulties, proactive maintenance and management and a new technological strategy are needed to increase safety, reliability, and efficiency. A novel neural network model for forecasting the life of a dry gas pipeline system and detecting the metal loss dimension class that is exposed to a harsh environment is presented in this study to handle the missing data. The proposed strategy blends the strength of deep learning techniques with industry-specific expertise. The main advantage of this study is to predict the pipeline life with a significant advantage of predicting the dimension classification of metal loss simultaneously employing a Bayesian regularization-based neural network framework when there are missing inputs in the datasets. The proposed intelligent model, trained on four pipeline datasets of a dry gas pipeline system, can predict the health condition of pipelines with high accuracy, even if there are missing parameters in the dataset. The proposed model using neural network technology generated satisfactory results in terms of numerical performance, with MSE and R2 values closer to 0 and 1, respectively. A few cases with missing input data are carried out, and the missing data is forecasted for each case. Then, a model is developed to predict the life condition of pipelines with the predicted missing input variables. The findings reveal that the model has the potential for real-world applications in the oil and gas sector for estimating the health condition of pipelines, even if there are missing input parameters. Additionally, multi-model comparative analysis and sensitivity analysis are incorporated, offering an extensive comprehension of multi-model prediction abilities and beneficial insights into the impact of various input variables on model outputs, thereby improving the interpretability and reliability of our results. The proposed framework could help business plans by lowering the chance of severe accidents and environmental harm with better safety and reliability.

An Automated Heart Shunt Recognition Pipeline Using Deep Neural Networks.

Automated recognition of heart shunts using saline contrast transthoracic echocardiography (SC-TTE) has the potential to transform clinical practice, enabling non-experts to assess heart shunt lesions. This study aims to develop a fully automated and scalable analysis pipeline for distinguishing heart shunts, utilizing a deep neural network-based framework. The pipeline consists of three steps: (1) chamber segmentation, (2) ultrasound microbubble localization, and (3) disease classification model establishment. The study's normal control group included 91 patients with intracardiac shunts, 61 patients with extracardiac shunts, and 84 asymptomatic individuals. Participants' SC-TTE images were segmented using the U-Net model to obtain cardiac chambers. The segmentation results were combined with ultrasound microbubble localization to generate multivariate time series data on microbubble counts in each chamber. A classification model was then trained using this data to distinguish between intracardiac and extracardiac shunts. The proposed framework accurately segmented heart chambers (dice coefficient = 0.92 ± 0.1) and localized microbubbles. The disease classification model achieved high accuracy, sensitivity, specificity, F1 score, kappa value, and AUC value for both intracardiac and extracardiac shunts. For intracardiac shunts, accuracy was 0.875 ± 0.008, sensitivity was 0.891 ± 0.002, specificity was 0.865 ± 0.012, F1 score was 0.836 ± 0.011, kappa value was 0.735 ± 0.017, and AUC value was 0.942 ± 0.014. For extracardiac shunts, accuracy was 0.902 ± 0.007, sensitivity was 0.763 ± 0.014, specificity was 0.966 ± 0.008, F1 score was 0.830 ± 0.012, kappa value was 0.762 ± 0.017, and AUC value was 0.916 ± 0.006. The proposed framework utilizing deep neural networks offers a fast, convenient, and accurate method for identifying intracardiac and extracardiac shunts. It aids in shunt recognition and generates valuable quantitative indices, assisting clinicians in diagnosing these conditions.

Multi-fidelity neural network-based aerodynamic optimization framework for propeller design in electric aircraft

The energy efficiency of the electric propulsion system in electric aircraft is significantly influenced by the propeller's efficiency, making the design of a high-efficiency propeller critical. Currently, the Blade Element Momentum Theory (BEMT) is widely used for propeller design. The BEMT method exhibits high efficiency but is limited by its low accuracy due to its inability to capture three-dimensional effects of rotating flow. On the other hand, the CFD simulation can enhance the accuracy of simulating complex rotating flows, but it is computationally expensive. To alleviate the contradiction between accuracy and design efficiency in propeller optimization design, this paper describes a Multi-Fidelity Neural Network (MFNN)-based optimization framework for the optimization design of electric aircraft propellers. This method is based on high-fidelity CFD numerical simulations for the optimization design of propellers. In order to enhance design efficiency and accuracy, it fuses low-fidelity BEMT knowledge, aiming to achieve higher model accuracy with fewer CFD simulations. Based on the MFNN, an adaptive sample-infilling approach is employed to update the model, and a teaching–learning-based optimization is used to build the surrogate-based optimization framework. The proposed method improves the cruise efficiency of the propeller from 82.3% (designed by BEMT) to 87.1% and shows advantages in optimization effectiveness and efficiency compared with single-fidelity optimization.

Artificial neural network-based home energy management system for smart homes

&lt;p&gt;Energy efficiency is widely recognized as one of the most significant and economical ways to lower greenhouse gas (GHG) emissions. The aims and goals are that smart meters can evaluate and communicate in-depth real-time electricity usage, enable remote real-time monitoring and management of power consumption, and provide consumers with real-time pricing and analyzed usage information. The house energy management controller decides which loads will be powered based on the real home energy demands and the predefined load priorities. Artificial intelligence (AI) is being used increasingly in control applications due to its great effectiveness and efficiency. As a result, in this work, the author designed, simulated, and optimized an artificial neural network-based model simulation framework that simulates a home with a variety of home appliances and optimizes the total energy consumption of the home realistically through intelligent control of home appliances. The MATLAB application was used to model and examine the performance of four common household appliances: the water heater (WH), washing machine (WM), air conditioner (AC), and refrigerator (RG). The result shows a considerable reduction and savings in energy consumption without a decrease in consumer comfort.&lt;strong&gt;&lt;/strong&gt;&lt;/p&gt;

MTBullyGNN: A Graph Neural Network-Based Multitask Framework for Cyberbullying Detection

MTBullyGNN: A Graph Neural Network-Based Multitask Framework for Cyberbullying Detection

Risk sharing with deep neural networks

We consider the problem of optimally sharing a financial position among agents with potentially different reference risk measures. The problem is equivalent to computing the infimal convolution of the risk metrics and finding the so-called optimal allocations. We propose a neural network-based framework to solve the problem and we prove the convergence of the approximated inf-convolution, as well as the approximated optimal allocations, to the corresponding theoretical values. We support our findings with several numerical experiments.

Improving rainfall-runoff modeling in the Mekong river basin using bias-corrected satellite precipitation products by convolutional neural networks

Accurate rainfall-runoff (RR) modeling is crucial for effective Mekong River Basin (MRB) water resource management. Satellite precipitation products (SPPs) can offer valuable data for such modeling; however, these products often exhibit biases that may adversely affect hydrological simulations. This study aimed to improve RR modeling using bias-corrected SPPs and the Soil and Water Assessment Tool (SWAT) model for MRB. A convolutional neural network-based deep learning framework was employed to correct biases in four SPPs (TRMM, PERSIANN-CDR, CHIRPS, and CMORPH), resulting in four respective bias-corrected SPPs (ADJ_TRMM, ADJ_CDR, ADJ_CHIR, and ADJ_CMOR). The bias-corrected products were compared against a gauge-based dataset in terms of rainfall analysis, and their performance within the SWAT model was assessed over calibration (2004–2013) and validation (2014–2015). Bias-corrected products demonstrated superior performance in rainfall analysis, with ADJ_TRMM outperforming other products. The SWAT model calibration results showed satisfactory performance across all stations, with a Nash-Sutcliffe Efficiency (NSE) ranging from [0.76–0.87]. Integrating bias-corrected SPPs into the SWAT model significantly increased the RR simulations in the MRB, indicated by higher NSE values [0.72–0.85] compared to uncorrected SPPs [-0.37 to 0.85] at the Kratie station. Besides, the inconsistent performance of bias-corrected products between rainfall analysis and RR modeling was observed, with ADJ_CDR outperforming ADJ_TRMM in the SWAT model. These results highlight the significance of using bias-corrected SPPs in hydrological modeling applications, especially in areas with limited ground-based precipitation data, and highlight the need for further research to refine bias correction methods and address the limitations of the SWAT model.

A Generative Shape Compositional Framework to Synthesize Populations of Virtual Chimeras.

Generating virtual organ populations that capture sufficient variability while remaining plausible is essential to conduct in silico trials (ISTs) of medical devices. However, not all anatomical shapes of interest are always available for each individual in a population. The imaging examinations and modalities used can vary between subjects depending on their individualized clinical pathways. Different imaging modalities may have various fields of view and are sensitive to signals from other tissues/organs, or both. Hence, missing/partially overlapping anatomical information is often available across individuals. We introduce a generative shape model for multipart anatomical structures, learnable from sets of unpaired datasets, i.e., where each substructure in the shape assembly comes from datasets with missing or partially overlapping substructures from disjoint subjects of the same population. The proposed generative model can synthesize complete multipart shape assemblies coined virtual chimeras (VCs). We applied this framework to build VCs from databases of whole-heart shape assemblies that each contribute samples for heart substructures. Specifically, we propose a graph neural network-based generative shape compositional framework, which comprises two components, a part-aware generative shape model that captures the variability in shape observed for each structure of interest in the training population and a spatial composition network that assembles/composes the structures synthesized by the former into multipart shape assemblies (i.e., VCs). We also propose a novel self-supervised learning scheme that enables the spatial composition network to be trained with partially overlapping data and weak labels. We trained and validated our approach using shapes of cardiac structures derived from cardiac magnetic resonance (MR) images in the UK Biobank (UKBB). When trained with complete and partially overlapping data, our approach significantly outperforms a principal component analysis (PCA)-based shape model (trained with complete data) in terms of generalizability and specificity. This demonstrates the superiority of the proposed method, as the synthesized cardiac virtual populations are more plausible and capture a greater degree of shape variability than those generated by the PCA-based shape model.