Training Dataset Research Articles

Initial state diversification for efficient AlphaZero-style training

AlphaZero has achieved superhuman performance in Go, chess, and shogi with a general reinforcement learning (RL) algorithm. This achievement is remarkable because AlphaZero does not rely on any training dataset of strong players. However, AlphaZero-style training requires substantial computational resources. Gumbel AlphaZero, a recently introduced more efficient version of AlphaZero, reduces the computational cost of AlphaZero training. The goal of this study is to further improve the playing strength of Gumbel AlphaZero under a limited amount of computational resources. We focus on the diversity in training games, inspired by procedural generation and domain randomization in RL studies, and propose a novel method, initial state diversification. This method diversifies the initial states of a self-play game to encourage the RL agent to understand the game in a more general manner through diverse experiences. For example, in shogi, the initial state of each self-play game is diversified by rearranging the pieces under realistic domain constraints. Experiments demonstrated that training with initial state diversification improves the playing strength of Gumbel AlphaZero in shogi, within the same computational budget for training.

Impact of Adverse Weather and Image Distortions on Vision-Based UAV Detection: A Performance Evaluation of Deep Learning Models

Unmanned aerial vehicle (UAV) detection in real-time is a challenging task despite the advances in computer vision and deep learning techniques. The increasing use of UAVs in numerous applications has generated worries about possible risks and misuse. Although vision-based UAV detection methods have been proposed in recent years, a standing open challenge and overlooked issue is that of adverse weather. This work is the first, to the best of our knowledge, to investigate the impact of adverse weather conditions and image distortions on vision-based UAV detection methods. To achieve this, a custom training dataset was curated with images containing a variety of UAVs in diverse complex backgrounds. In addition, this work develops a first-of-its-kind dataset, to the best of our knowledge, with UAV-containing images affected by adverse conditions. Based on the proposed datasets, a comprehensive benchmarking study is conducted to evaluate the impact of adverse weather and image distortions on the performance of popular object detection methods such as YOLOv5, YOLOv8, Faster-RCNN, RetinaNet, and YOLO-NAS. The experimental results reveal the weaknesses of the studied models and the performance degradation due to adverse weather, highlighting avenues for future improvement. The results show that even the best UAV detection model’s performance degrades in mean average precision (mAP) by −50.62 points in torrential rain conditions, by −52.40 points in high noise conditions, and by −77.0 points in high motion blur conditions. To increase the selected models’ resilience, we propose and evaluate a strategy to enhance the training of the selected models by introducing weather effects in the training images. For example, the YOLOv5 model with the proposed enhancement strategy gained +35.4, +39.3, and +44.9 points higher mAP in severe rain, noise, and motion blur conditions respectively. The findings presented in this work highlight the advantages of considering adverse weather conditions during model training and underscore the significance of data enrichment for improving model generalization. The work also accentuates the need for further research into advanced techniques and architectures to ensure more reliable UAV detection under extreme weather conditions and image distortions.

Tra ined Wi thout My C onsent: Detecting Code Inclusion In Language Models Trained on Code

Code auditing ensures that the developed code adheres to standards, regulations, and copyright protection by verifying that it does not contain code from protected sources. The recent advent of Large Language Models (LLMs) as coding assistants in the software development process poses new challenges for code auditing. The dataset for training these models is mainly collected from publicly available sources. This raises the issue of intellectual property infringement as developers’ codes are already included in the dataset. Therefore, auditing code developed using LLMs is challenging, as it is difficult to reliably assert if an LLM used during development has been trained on specific copyrighted codes, given that we do not have access to the training datasets of these models. Given the non-disclosure of the training datasets, traditional approaches such as code clone detection are insufficient for asserting copyright infringement. To address this challenge, we propose a new approach, TraWiC; a model-agnostic and interpretable method based on membership inference for detecting code inclusion in an LLM’s training dataset. We extract syntactic and semantic identifiers unique to each program to train a classifier for detecting code inclusion. In our experiments, we observe that TraWiC is capable of detecting 83.87% of codes that were used to train an LLM. In comparison, the prevalent clone detection tool NiCad is only capable of detecting 47.64%. In addition to its remarkable performance, TraWiC has low resource overhead in contrast to pair-wise clone detection that is conducted during the auditing process of tools like CodeWhisperer reference tracker, across thousands of code snippets.

A two-stage deep learning architecture for detection global coastal and offshore submesoscale ocean eddy using SDGSAT-1 multispectral imagery

Submesoscale ocean eddies are essential oceanic phenomenon that control and influence the ocean energy cascade. Most existing eddy detection methods rely on low-resolution satellite altimeter data, which fail to capture submesoscale ocean features and oceanographic phenomena in shallow water. Introducing high-resolution multispectral data can alleviate these problems, yet it has been largely overlooked. A generalized and efficient deep learning architecture that combines developments in deep learning with Sustainable Development Goals Science Satellite 1 (SDGSAT-1) multispectral data from earth observations offers a potential pathway for more fine detection of ocean eddies. Considering that oceanic eddy exhibits spatially sparse characteristics on high-resolution remote sensing scenes, the oceanic eddy detection (OED) model suitable for global coastal and offshore regions is divided into two stages: eddy information judgment and eddy position determination. Correspondingly, SDGSAT-1 multispectral data from November 2021 to December 2022 were carried out to construct two submesoscale eddy datasets for training and testing each stage model. The union validation of multiple metrics demonstrates that the proposed OED model and its stage models achieve state-of-the-art (SOTA) performance, especially in optically complex coastal and offshore waters. We applied the model to real-world scenes captured by SDGSAT-1 in 2023, and found that the detected results were mainly located at the water depth below 200 m. The authenticity of the recognition results is validated using sea surface chlorophyll concentration, temperature, and topography data, indicating that the OED model has achieved remarkable effectiveness under various sea conditions. In addition, the temporal distributions and statistical characteristics of detected submesoscale eddies are analyzed over an extended period (November 2021 to November 2023). Finally, HISEA-2, Landsat-9, and Sentinel-2 served as testing grounds to validate the generalization of the proposed methodology, with experimental results demonstrating that the OED model possesses significant developmental potential for multi-source remote sensing data. This paper presents a comprehensive deep learning framework for the global-scale detection of submesoscale eddies and underscores the pivotal role of high-resolution multispectral imagery as an innovative data source for global coastal and offshore eddy identification.

Employing Transfer Learning in Land-use Land-cover for Risk Management

Abstract. In disaster management, land-use land-cover (LULC) maps are vital for real-time situational awareness and coordinated responses. These maps aid in coordinating field operations, guiding rescue teams, and identifying high vulnerability areas. They ensure accurate spatial information sharing and management during disaster response efforts. Deep transfer learning models have emerged as powerful tools for LULC classification, addressing challenges like insufficient training data and complex classification tasks. In this research instead of building networks from scratch, pre-trained networks that used EuroSAT benchmark dataset are employed. Several deep learning models including 1-layer CNN, 4-layers CNN, VGG16 and an improved ResNet-50 network as proposed method are considered and compared in this study. The results were analyzed in both quantitative and qualitative ways. In the quantitative mode, the measurement criteria such as Overall Accuracy (OA), F1Score, Precision and Recall were calculated, and in the qualitative mode, the class diagram was drawn in the feature space to check the separability of the classes. Finally, the results show high overall accuracy score of 95.9% indicating the high potential of our proposed network for ResNet-50. The proposed method has resolved insufficient training dataset by implementing data augmentation that it can be solved the problem of lack dataset.

On the local convergence of ADAM-DPGAN with simultaneous and alternating gradient decent training methods

Generative Adversarial Networks (GANs) do not ensure the privacy of the training datasets and may memorize sensitive details. To maintain privacy of data during inference, various privacy-preserving GAN mechanisms have been proposed. Despite the different approaches and their characteristics, advantages, and disadvantages, there is a lack of a systematic review on them. This paper first presents a comprehensive survey on privacy-preserving mechanisms and offers a taxonomy based on their characteristics. The survey reveals that many of these mechanisms modify the GAN learning algorithm to enhance privacy, highlighting the need for theoretical and empirical analysis of the impact of these modifications on GAN convergence. Among the surveyed methods, ADAM-DPGAN is a promising approach that ensures differential privacy in GANs for both the discriminator and the generator networks when using the ADAM optimizer, by introducing appropriate noise based on the global sensitivity of discriminator parameters. Therefore, this paper conducts a theoretical and empirical analysis of the convergence of ADAM-DPGAN. In the presented theoretical analysis, assuming that simultaneous/alternating gradient descent method with ADAM optimizer converges locally to a fixed point and its operator is L-Lipschitz with L < 1, the effect of ADAM-DPGAN-based noise disturbance on local convergence is investigated and an upper bound for the convergence rate is provided. The analysis highlights the significant impact of differential privacy parameters, the number of training iterations, the discriminator’s learning rate, and the ADAM hyper-parameters on the convergence rate. The theoretical analysis is further validated through empirical analysis. Both theoretical and empirical analyses reveal that a stronger privacy guarantee leads to a slower convergence, highlighting the trade-off between privacy and performance. The findings also indicate that there exists an optimal value for the number of training iterations regarding the privacy needs. The optimal settings for each parameter are calculated and outlined in the paper.

LEAVES: An Expandable Light-curve Data Set for Automatic Classification of Variable Stars

With the increasing amount of astronomical observation data, it is an inevitable trend to use artificial intelligence methods for automatic analysis and identification of light curves for full samples. However, data sets covering all known classes of variable stars that meet all research needs are not yet available. There is still a lack of standard training data sets specifically designed for any type of light-curve classification, but existing light-curve training sets or data sets cannot be directly merged into a large collection. Based on the open data sets of the All-Sky Automated Survey for SuperNovae, Gaia, and Zwicky Transient Facility, we construct a compatible light-curve data set named LEAVES for automated recognition of variable stars, which can be used for training and testing new classification algorithms. The data set contains a total of 977,953 variable and 134,592 nonvariable light curves, in which the supported variables are divided into six superclasses and nine subclasses. We validate the compatibility of the data set through experiments and employ it to train a hierarchical random forest classifier, which achieves a weighted average F1-score of 0.95 for seven-class classification and 0.93 for 10-class classification. Experimental results prove that the classifier is more compatible than the classifier established based on a single band and a single survey, and has wider applicability while ensuring classification accuracy, which means it can be directly applied to different data types with only a relatively small loss in performance compared to a dedicated model.

Centerline-guided reinforcement learning model for pancreatic duct identifications.

Pancreatic ductal adenocarcinoma is forecast to become the second most significant cause of cancer mortality as the number of patients with cancer in the main duct of the pancreas grows, and measurement of the pancreatic duct diameter from medical images has been identified as relevant for its early diagnosis. We propose an automated pancreatic duct centerline tracing method from computed tomography (CT) images that is based on deep reinforcement learning, which employs an artificial agent to interact with the environment and calculates rewards by combining the distances from the target and the centerline. A deep neural network is implemented to forecast step-wise values for each potential action. With the help of this mechanism, the agent can probe along the pancreatic duct centerline using the best possible navigational path. To enhance the tracing accuracy, we employ landmark-based registration, which enables the generation of a probability map of the pancreatic duct. Subsequently, we utilize a gradient-based method on the registered data to extract a probability map specifically indicating the centerline of the pancreatic duct. Three datasets with a total of 115 CT images were used to evaluate the proposed method. Using image hold-out from the first two datasets, the method performance was 2.0, 4.0, and 2.1mm measured in terms of the mean detection error, Hausdorff distance (HD), and root mean squared error (RMSE), respectively. Using the first two datasets for training and the third one for testing, the method accuracy was 2.2, 4.9, and 2.6mm measured in terms of the mean detection error, HD, and RMSE, respectively. We present an algorithm for automated pancreatic duct centerline tracing using deep reinforcement learning. We observe that validation on an external dataset confirms the potential for practical utilization of the presented method.

Combining Low-Rank and Deep Plug-and-Play Priors for Snapshot Compressive Imaging.

Snapshot compressive imaging (SCI) is a promising technique that captures a 3-D hyperspectral image (HSI) by a 2-D detector in a compressed manner. The ill-posed inverse process of reconstructing the HSI from their corresponding 2-D measurements is challenging. However, current approaches either neglect the underlying characteristics, such as high spectral correlation, or demand abundant training datasets, resulting in an inadequate balance among performance, generalizability, and interpretability. To address these challenges, in this article, we propose a novel approach called LR2DP that integrates the model-driven low-rank prior and data-driven deep priors for SCI reconstruction. This approach not only captures the spectral correlation and deep spatial features of HSI but also takes advantage of both model-based and learning-based methods without requiring any extra training datasets. Specifically, to preserve the strong spectral correlation of the HSI effectively, we propose that the HSI lies in a low-rank subspace, thereby transforming the problem of reconstructing the HSI into estimating the spectral basis and spatial representation coefficient. Inspired by the mutual promotion of unsupervised deep image prior (DIP) and trained deep denoising prior (DDP), we integrate the unsupervised network and pre-trained deep denoiser into the plug-and-play (PnP) regime to estimate the representation coefficient together, aiming to explore the internal target image prior (learned by DIP) and the external training image prior (depicted by pre-trained DDP) of the HSI. An effective half-quadratic splitting (HQS) technique is employed to optimize the proposed HSI reconstruction model. Extensive experiments on both simulated and real datasets demonstrate the superiority of the proposed method over the state-of-the-art approaches.

VTSIM: Attention‐Based Recurrent Neural Network for Intersection Vehicle Trajectory Simulation

ABSTRACTSimulating vehicle trajectories at intersections is one of the challenging tasks in traffic simulation. Existing methods are often ineffective due to the complexity and diversity of lane topologies at intersections, as well as the numerous interactions affecting vehicle motion. To address this issue, we propose a deep learning based vehicle trajectory simulation method. First, we employ a vectorized representation to uniformly extract features from traffic elements such as pedestrians, vehicles, and lanes. By fusing all factors that influence vehicle motion, this representation makes our method suitable for a variety of intersections. Second, we propose a deep learning model, which has an attention network to dynamically extract features from the surrounding environment of the vehicles. To address the issue of vehicles continuously entering and exiting the simulation scene, we employ an asynchronous recurrent neural network for the extraction of temporal features. Comparative evaluations against existing rule‐based and deep learning‐based methods demonstrate our model's superior simulation accuracy. Furthermore, experimental validation on public datasets demonstrates that our model can simulate vehicle trajectories among the urban intersections with different topologies including those not present in the training dataset.

Utilizing Saliva Metabolomics for Diagnosing Gastric Cancer and Exploring the Changes in Saliva Metabolites After Surgery.

Gastric cancer (GC) is a disease with high prevalence and mortality, but we lack convenient and accurate methods to screen for this disease. Thus, we aimed to search for some salivary biomarkers and explore changes in metabolites in patients' saliva after radical gastrectomy. A total of 152 subjects were divided into three groups (healthy group, GC group, and one-week postoperative group). After simple processing, saliva samples were analyzed by liquid chromatography-mass spectrometry. First, we used total ion chromatography and principal component analysis to determine the metabolite profiles. Next, t-test, partial least squares discriminant analysis, support vector machine, and receiver operating characteristics curve analysis were performed to identify biomarkers. Then, Fisher discriminant analysis and hierarchical clustering analysis were performed to determine the discriminating ability of biomarkers. Finally, we established a generalized linear model to predict GC based on biomarkers, and used bootstrapping for internal validation. Between the healthy and GC groups, we identified four biomarkers: lactic acid, kynurenic acid, 3-hydroxystachydrine, and S-(1,2,2-trichlorovinyl)-L-cysteine. We used stepwise regression to select five metabolites and develop a model with areas under the curve equal to 0.973 in the training dataset and 0.924 in the validation dataset. Between the GC and one-week postoperative groups, we found two differential metabolites: 19-hydroxyprostaglandin E2 and DG (14:0/0:0/18:2n6). Differential metabolites were observed among the three groups. GC could be initially diagnosed on the basis of detection of these biomarkers. Moreover, changes in salivary metabolites in postoperative patients could provide important insights for basic studies.

Measurement of intelligent computing via Levenberg Marquardt algorithm (LMA) for accurate prediction of fluid forces in a transient non-Newtonian thermal flow

Predicting precise results for the quantities of interest in time-dependent Computational Fluid Dynamics (CFD) simulations requires a significant investment of computational resources and time. To get around these issues, CFD simulations have been joined with Artificial Neural Networks (ANN). An optimally configured artificial neural network (ANN) is given the training and validation data sets produced by computational fluid dynamics (CFD). The flow around a cylinder, which is a well-known benchmark problem for incompressible flows, has been taken into consideration by the hybrid-CFD system. The mathematical model is based on the non-stationary Navier-Stokes and energy equations with viscosity. The basic architecture of the ANN model consists of 10 hidden layers, three output levels, and five input layers. Fast second-order LMA, a top-tier approach, was used to train the network. Both the Mean Square Error (MSE) and the coefficient of determination (R) provide statistical evidence that the ANN projected values for the drag and lift coefficients and average Nusselt number obtained from the finite element analysis are accurate. This analysis suggests that ANNs have the potential to significantly cut down on the amount of time and energy needed to run time-dependent simulations.

Machine Learning-Based Evaluation of Suicide Risk Assessment in Crisis Counseling Calls.

Counselor assessment of suicide risk is one key component of crisis counseling, and standards require risk assessment in every crisis counseling conversation. Efforts to increase risk assessment frequency are limited by quality improvement tools that rely on human evaluation of conversations, which is labor intensive, slow, and impossible to scale. Advances in machine learning (ML) have made possible the development of tools that can automatically and immediately detect the presence of risk assessment in crisis counseling conversations. To train models, a coding team labeled every statement in 476 crisis counseling calls (193,257 statements) for a core element of risk assessment. The authors then fine-tuned a transformer-based ML model with the labeled data, utilizing separate training, validation, and test data sets. Generally, the evaluated ML model was highly consistent with human raters. For detecting any risk assessment, ML model agreement with human ratings was 98% of human interrater agreement. Across specific labels, average F1 (the harmonic mean of precision and recall) was 0.86 at the call level and 0.66 at the statement level and often varied as a result of a low base rate for some risk labels. ML models can reliably detect the presence of suicide risk assessment in crisis counseling conversations, presenting an opportunity to scale quality improvement efforts.

Unified LSPG Model Reduction Framework and Assessment for Hypersonic Computational Fluid Dynamics

A unified least-squares Petrov–Galerkin (LSPG) framework for projection-based model order reduction featuring three different approximation manifolds [affine manifold, quadratic manifold, and nonlinear manifold built using a deep artificial neural network (ANN)] is presented. Its performance was assessed for a variable-speed version of the double-cone hypersonic benchmark problem. First, a high-dimensional viscous computational fluid dynamics model (HDM) was constructed, verified, and validated. The dimensionality of the HDM was then reduced using LSPG, each of the aforementioned approximation manifolds, and a global right reduced-order basis trained in the range 8≤M∞≤13. Each resulting global projection-based reduced-order model (PROM) was hyper-reduced and transformed into a hyper-reduced PROM (HPROM). The accuracy of each constructed HPROM was assessed for various quantities of interest and contrasted with that of snapshot interpolation. For this purpose, three different error measures were considered and discussed in the context of shock-dominated problems. Wall-clock and CPU time speedup factors are reported. Overall, it was shown that using a relatively small set of training data, all constructed LSPG HPROMs were nonlinearly stable, real-time capable, and highly predictive. The LSPG HPROM constructed using a nonlinear approximation manifold and an ANN was the most computationally efficient.

Deep learning model for automated diagnosis of moyamoya disease based on magnetic resonance angiography

This study explores the potential of the deep learning-based convolutional neural network (CNN) to automatically recognize MMD using MRA images from atherosclerotic disease (ASD) and normal control (NC). In this retrospective study in China, 600 participants (200MMD, 200 ASD and 200 NC) were collected from one institution as an internal dataset for training and 60 from another institution were collected as external testing set for validation. All participants were divided into training (N=450) and validation sets (N=90), internal testing set (N=60), and external testing set (N=60). The input to the CNN models comprised preprocessed MRA images, while the output was a tripartite classification label that identified the patient's diagnostic group. The performances of 3D CNN models were evaluated using a comprehensive set of metrics such as area under the curve (AUC) and accuracy. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the CNN's decision-making process in MMD diagnosis by highlighting key areas. Finally, the diagnostic performances of the CNN models were compared with those of two experienced radiologists. DenseNet-121 exhibited superior discrimination capabilities, achieving a macro-average AUC of 0.977 (95% CI, 0.928-0.995) in the internal test sets and 0.880 (95% CI, 0.786-0.937) in the external validation sets, thus exhibiting comparable diagnostic capabilities to those of human radiologists. In the binary classification where ASD and NC were group together, with MMD as the separate group for targeted detection, DenseNet-121 achieved an accuracy of 0.967 (95% CI, 0.886-0.991). Additionally, the Grad-CAM results for the MMD, with areas of intense redness indicating critical areas identified by the model, reflected decision-making similar to human experts. This study highlights the efficacy of CNN model in the automated diagnosis of MMD on MRA images, easing the workload on radiologists and promising integration into clinical workflows. National Natural Science Foundation of China, Tianjin Science and Technology Project and Beijing Natural Science Foundation.

Predicting the Early Detection of Breast Cancer Using Hybrid Machine Learning Systems and Thermographic Imaging

ABSTRACTBreast cancer is a leading cause of mortality among women, emphasizing the critical need for precise early detection and prognosis. However, conventional methods often struggle to differentiate precancerous lesions or tailor treatments effectively. Thermal imaging, capturing subtle temperature variations, presents a promising avenue for non‐invasive cancer detection. While some studies explore thermography for breast cancer detection, integrating it with advanced machine learning for early diagnosis and personalized prediction remains relatively unexplored. This study proposes a novel hybrid machine learning system (HMLS) incorporating deep autoencoder techniques for automated early detection and prognostic stratification of breast cancer patients. By exploiting the temporal dynamics of thermographic data, this approach offers a more comprehensive analysis than static single‐frame approaches. Data processing involves splitting the dataset for training and testing. A predominant infrared image was selected, and matrix factorization was applied to capture temperature changes over time. Integration of convex factor analysis and bell‐curve membership function embedding for dimensionality reduction and feature extraction. The autoencoder deep neural network further reduces dimensionality. HMLS model development included feature selection and optimization of survival prediction algorithms through cross‐validation. Model performance was assessed using accuracy and F‐measure metrics. HMLS, integrating clinical data, achieved 81.6% accuracy, surpassing 77.6% using only convex‐NMF. The best classifier attained 83.2% accuracy on test data. This study demonstrates the effectiveness of thermographic imaging and HMLS for accurate early detection and personalized prediction of breast cancer. The proposed framework holds promise for enhancing patient care and potentially reducing mortality rates.

Space Weather Forecasts of Ground Level Space Weather in the UK: Evaluating Performance and Limitations

AbstractGeomagnetically Induced Currents (GICs) are a severe space weather hazard, driven through coupling between the solar wind and magnetosphere. GICs are rarely measured directly, instead the ground magnetic field variability is often used as a proxy. Recently space weather models have been developed to forecast whether the magnetic field variability (R) will exceed specific, extreme thresholds. We test an example machine learning‐based model developed for the northern United Kingdom. We evaluate its performance (discriminative skill and calibration) as a function of magnetospheric state, solar wind input and magnetic local time. We find that the model's performance is highest during active conditions, for example, geomagnetic storms, and lowest during isolated substorms and “quiet” intervals, despite these conditions dominating the training data set. Correspondingly, the performance is high when the solar wind conditions are elevated (i.e., high velocity, large total magnetic field strength, and the interplanetary magnetic field oriented southward), and at a minimum when the north‐south component of the magnetic field is highly variable or around zero. Regarding magnetic local time, performance is highest within the dusk and night sectors, and lowest during the day. The model appears to capture multiple modes of magnetospheric activity, including substorms and viscous interactions, but poorly predicts impulsive phenomena (i.e., storm sudden commencements) and longer timescale coupling processes. Future models of mid‐latitude magnetic field variability will need to effectively use longer time intervals of unpropagated (i.e., observations from L1) solar wind to more completely describe the magnetospheric conditions and response.

A data assimilation pressure field measurement method for linear turbine cascades based on compressed sensing

One of the most dominant measuring techniques for linear turbine cascades is to obtain the high spatial resolution pressure field by a discrete point array. In this work, a compressed sensing (CS) based data assimilation methodology is proposed, by which a set of optimal sparse measuring points can be derived through an optimization procedure. Combined with numerical simulations and the data-driven modal decomposition, the high spatial resolution pressure distribution can be reconstructed accurately with sparse random sampling. To this end, detailed comparative research is conducted. First, the impacts of different sparse bases, including the discrete Fourier transform (DFT), the discrete cosine transform (DCT), and the proper orthogonal decomposition (POD) matrices, and the compression ratio on the reconstruction performance are compared and analyzed systematically under different incidence angles and cascade exit isentropic Mach numbers. Results reveal that a CS approach on POD subspace (CS-POD) performs remarkably better than the CS-DFT and CS-DCT in capturing the spatially continuous pressure distribution, even with a small number of measuring points. Furthermore, effects of the order of truncated POD modes and the number of training dataset required to conduct POD on the error are also investigated that exhibits a downward trend with the rise in these two elements. To overcome the deficiency of randomly selected sparse observation sites with this methodology and the resulting high measuring cost under different conditions, a vectorized CS-POD (Vec-CS-POD) model is constructed to obtain one set of measuring distribution that could satisfy the multi-conditional measurements simultaneously, and its reliability and robustness are validated through cascade experiments. With the aid of the Vec-CS-POD based data fusion framework, spatially resolved end wall pressure fields can be acquired by only a few measuring ports, the number of which can be reduced by 77% compared to the conventional uniform arrangement. The generalization ability of the proposed framework is also validated and evaluated; thus, it exhibits broad potential in other flow field measurements.

Deep Learning and Statistical Operations Based features extraction for Skin Cancer Detection

Skin cancer is considered as one of the most serious types of cancer that leads to death worldwide. The number of deaths that caused by skin cancer can be reduced if it is diagnosed at early stages. Skin cancer is usually diagnosed using visual inspection, but it is less accurate. Using deep learning-based methods have been proposed to assist the doctors to diagnose the skin cancers at early accurately. The investigation was achieved on 3600 images collected from kaggle. Two deep learning algorithms used in the study: Vgg-19 and Alexnet to extract the layers 6 and 7. Then, the two layers will be merged to generate the statistical operations like: median, lowest, highest and joined of two layers. The following classifiers were used in the investigation: K-Nearest Neighbor, Random Forest, Naïve Bayes and Decision Tree. However, the study considered the following measures: accuracy, precision, recall, and F-Measure. Three training datasets sizes will be used to investigate the influence on classification accuracy. The results of all datasets were slightly similar. This approves that the extracted features and the statistical operations have an influence on the classification accuracy. The results show that Alexnet performs high accuracy and consumes less time that required for training the model compared to vgg-19. The results of classifiers showed that Random Forest scored high classification accuracy (85.6) compared to other ML classifiers

Surrogate modeling of multi-dimensional premixed and non-premixed combustion using pseudo-time stepping physics-informed neural networks

Physics-informed neural networks (PINNs) have emerged as a promising alternative to conventional computational fluid dynamics (CFD) approaches for solving and modeling multi-dimensional flow fields. They offer instant inference speed and cost-effectiveness without the need for training datasets. However, compared to common data-driven methods, purely learning the physical constraints of partial differential equations and boundary conditions is much more challenging and prone to convergence issues leading to incorrect local optima. This training robustness issue significantly increases the difficulty of fine-tuning PINNs and limits their widespread adoption. In this work, we present improvements to the prior field-resolving surrogate modeling framework for combustion systems based on PINNs. First, inspired by the time-stepping schemes used in CFD numerical methods, we introduce a pseudo-time stepping loss aggregation algorithm to enhance the convergence robustness of the PINNs training process. This new pseudo-time stepping PINNs (PTS-PINNs) method is then tested in non-reactive convection–diffusion problem, and the results demonstrated its good convergence capability for multi-species transport problems. Second, the effectiveness of the PTS-PINNs method was verified in the case of methane–air premixed combustion, and the results show that the L2 norm relative error of all variables can be reduced within 5%. Finally, we also extend the capability of the PTS-PINNs method to address a more complex methane–air non-premixed combustion problem. The results indicate that the PTS-PINNs method can still achieve commendable accuracy by reducing the relative error to within 10%. Overall, the PTS-PINNs method demonstrates the ability to rapidly and accurately identify the convergence direction of the model, surpassing traditional PINNs methods in this regard.