Adaptive Sampling Technique Research Articles

An Adaptive Frequency Sampling Technique Based on SMoM and EDM for Wideband Electromagnetic Scattering Data of Multiscale Targets

An Adaptive Frequency Sampling Technique Based on SMoM and EDM for Wideband Electromagnetic Scattering Data of Multiscale Targets

Enhancing local road pavement condition prediction using Bayesian-optimized ensemble machine learning and adaptive synthetic sampling technique

ABSTRACT Pavement condition prediction helps road agencies to schedule maintenance, rehabilitation, and reconstruction, and to allocate limited funds and resources to such activities. Compared to state highways, pavement performance prediction for local roads has received relatively little attention in the literature due to perceptions of low importance, low levels of investment in data collection, poor data quality, and high variation within the data elements. Additionally, local road pavement condition data may suffer from dataset imbalance, often leading to unreliable condition predictions. Hence, this paper introduces a methodology to predict local pavement condition using various single estimator and ensemble machine learning (ML) models along with the adaptive synthetic sampling method. The study develops nine (9) Bayesian-optimised ML models: category boosting (CatBoost), adaptive boosting, decision tree, extra trees, gradient boosting, light gradient-boosting machine, k-nearest neighbour, random forest, and artificial neural network. The ensemble ML and CatBoost were found to exhibit the best model performance, with an average testing accuracy of 0.82, specificity of 0.81, sensitivity of 0.63, and F-measure of 0.61. These results underscore the efficacy of ensemble ML models in pavement condition prediction. The proposed approach can be beneficial to local road agencies in their long-term planning, scheduling, and budgeting.

Advanced multimaterial shape optimization methods as applied to advanced manufacturing of wind turbine generators

AbstractCurrently, many utility‐scale wind turbine generator original equipment manufacturers are dependent on imported rare earth permanent magnets, which are susceptible to market risks from cost instability. To lower the production costs of these generators and stay competitive in the market, several small wind manufacturers are pursuing continuous improvements to both generator design and manufacturing. However, traditional design and manufacturing methods have yielded marginal improvements in wind power performance. This work presents novel methods to redesign a baseline 15‐kW wind turbine generator with reduced rare‐earth permanent magnets by leveraging cutting‐edge three‐dimensional (3D) printed polymer‐bonded permanent magnets and steel. Symmetric, asymmetric, and multimaterial‐magnet parametrization methods are introduced for shape optimization. We extend the symmetric and asymmetric methods to the back iron in the stator to further investigate the impact and opportunities for performance improvements with lesser active materials. We employ a design‐of‐experiments approach with parametric computer‐aided design for shape generation and evaluate different designs by magneto‐thermal modeling and finite‐element analysis. We use adaptive sampling technique to identify better performing designs with lesser magnet mass, higher efficiency, and lower cogging torque when compared with the baseline generator. Asymmetric pole designs resulted in a magnet mass in the range of 4.77–5.37 kg, which was 27%–35% lighter than the baseline generator, suggesting that a new design freedom exists that can be enabled by advanced manufacturing, such as 3D printing. Shaping the back iron in the stator resulted in material savings in electrical steel of up to 14.62 kg, which was 20% lighter than the baseline stator. We conducted a structural analysis to evaluate an optimized asymmetric rotor design from the point of view of mechanical integrity and air‐gap stiffness. The magnetically optimal shape profile was shown as having a positive impact on the radial stiffness, and an optimal solution was discovered to reduce the structural mass by nearly 30 kg, which was 29% lighter than the baseline.

Active Learning and Bayesian Optimization: A Unified Perspective to Learn with a Goal

Science and Engineering applications are typically associated with expensive optimization problem to identify optimal design solutions and states of the system of interest. Bayesian optimization and active learning compute surrogate models through efficient adaptive sampling schemes to assist and accelerate this search task toward a given optimization goal. Both those methodologies are driven by specific infill/learning criteria which quantify the utility with respect to the set goal of evaluating the objective function for unknown combinations of optimization variables. While the two fields have seen an exponential growth in popularity in the past decades, their dualism and synergy have received relatively little attention to date. This paper discusses and formalizes the synergy between Bayesian optimization and active learning as symbiotic adaptive sampling methodologies driven by common principles. In particular, we demonstrate this unified perspective through the formalization of the analogy between the Bayesian infill criteria and active learning criteria as driving principles of both the goal-driven procedures. To support our original perspective, we propose a general classification of adaptive sampling techniques to highlight similarities and differences between the vast families of adaptive sampling, active learning, and Bayesian optimization. Accordingly, the synergy is demonstrated mapping the Bayesian infill criteria with the active learning criteria, and is formalized for searches informed by both a single information source and multiple levels of fidelity. In addition, we provide guidelines to apply those learning criteria investigating the performance of different Bayesian schemes for a variety of benchmark problems to highlight benefits and limitations over mathematical properties that characterize real-world applications.

Developing and validating machine learning-based prediction models for frailty occurrence in those with chronic obstructive pulmonary disease.

Frailty is a medical syndrome caused by multiple factors, characterized by decreased strength, endurance, and diminished physiological function, resulting in increased susceptibility to dependence and/or death. Patients with chronic obstructive pulmonary disease (COPD) tend to be more vulnerable to frailty due to their physical and psychological burdens. Therefore, the aim of this study was to develop a reliable and accurate vulnerability risk prediction model for frailty in patients with COPD in order to improve the identification and prediction of patient frailty. The specific objectives of this study were to determine the prevalence of frailty in patients with COPD and develop a prediction model and evaluate its predictive power. Clinical information was analyzed using data from the 2018 China Health and Retirement Longitudinal Study (CHARLS) database, and 34 indicators, including behavioral factors, health status, mental health parameters, and various sociodemographic variables, were examined in the study. The adaptive synthetic sampling technique was used for unbalanced data. Three methods, ridge regressor, extreme gradient boosting (XGBoost) classifier, and random forest (RF) regressor, were used to filter predictors. Seven machine learning (ML) techniques including logistic regression (LR), support vector machines (SVM), multilayer perceptron, light gradient-boosting machine, XGBoost, RF, and K-nearest neighbors were used to analyze and determine the optimal model. For customized risk assessment, an online predictive risk modeling website was created, along with Shapley additive explanation (SHAP) interpretations. Depression, smoking, gender, social activities, dyslipidemia, asthma, and residence type (urban vs. rural) were predictors for the development of frailty in patients with COPD. In the test set, the XGBoost model had an area under the curve of 0.942 (95% confidence interval: 0.925-0.959), an accuracy of 0.915, a sensitivity of 0.873, and a specificity of 0.911, indicating that it was the best model. The ML predictive model developed in this study is a useful and easy-to-use instrument for assessing the vulnerability risk of patients with COPD and may aid clinical physicians in screening high-risk patients.

A new modified deep learning technique based on physics-informed neural networks (PINNs) for the shock-induced coupled thermoelasticity analysis in a porous material

In this article, a new modified deep learning (DL) method based on physics-informed neural networks (PINNs) is proposed for analyzing generalized coupled thermoelasticity in a porous material under shock loadings using Lord–Shulman (LS) theory. The PINN-based method demonstrates remarkable capabilities in solving differential equations and identifying unknown parameters. It is employed to solve a system of coupled partial differential equations (PDEs) governing a porous half-space material, considering thermal and strain relaxation coefficients in the LS theory. The optimal structure of the PINN is investigated through sensitivity analyses. Two adaptive sampling techniques, residual-based adaptive refinement (RAR) and residual-based adaptive distribution (RAD), are employed to enhance solution quality within the optimized architecture. The proposed forward PINN utilizes known values of field variables at initial and boundary conditions. The efficiency and effectiveness of the proposed PINN approach are demonstrated through three distinct scenarios. Non-parametric statistical tests and L 2 relative error analysis validate the extraordinary potential of the proposed PINN-based method in accurately capturing the system behavior. The extrapolation results, represented as time history plots, showcase exceptional accuracy in this study, overcoming the limitations of conventional numerical methods in larger temporal domains.

Optimal estimation of reliability parameter for inverse Pareto distribution with application to insurance data

PurposeThe purpose of this paper is to develop optimal estimation procedures for the stress-strength reliability (SSR) parameter R = P(X > Y) of an inverse Pareto distribution (IPD).Design/methodology/approachWe estimate the SSR parameter R = P(X > Y) of the IPD under the minimum risk and bounded risk point estimation problems, where X and Y are strength and stress variables, respectively. The total loss function considered is a combination of estimation error (squared error) and cost, utilizing which we minimize the associated risk in order to estimate the reliability parameter. As no fixed-sample technique can be used to solve the proposed point estimation problems, we propose some “cost and time efficient” adaptive sampling techniques (two-stage and purely sequential sampling methods) to tackle them.FindingsWe state important results based on the proposed sampling methodologies. These include estimations of the expected sample size, standard deviation (SD) and mean square error (MSE) of the terminal estimator of reliability parameters. The theoretical values of reliability parameters and the associated sample size and risk functions are well supported by exhaustive simulation analyses. The applicability of our suggested methodology is further corroborated by a real dataset based on insurance claims.Originality/valueThis study will be useful for scenarios where various logistical concerns are involved in the reliability analysis. The methodologies proposed in this study can reduce the number of sampling operations substantially and save time and cost to a great extent.

Machine learning approach to predict postpancreatectomy hemorrhage following pancreaticoduodenectomy: a retrospective study.

Postpancreatectomy hemorrhage (PPH) is a rare yet dreaded complication following pancreaticoduodenectomy (PD). This retrospective study aimed to explore a machine learning (ML) model for predicting PPH in PD patients. A total of 284 patients who underwent open PD at our institute were included in the analysis. To address the issue of imbalanced data, the adaptive synthetic sampling (ADASYN) technique was employed. The best-performing ML model was selected using the PyCaret library in Python and evaluated based on recall, precision, and F1 score metrics. In addition to assessing the model's performance on the test data, bootstrap validation (n = 1000) with the original dataset was conducted. PPH occurred in 11 patients (3.9%), with a median onset time of 22days postoperatively. These minority cases were oversampled to 85 using ADASYN. The extra trees classifier demonstrated superior performance with recall, precision, and F1 score of 0.967, 0.914, and 0.937, respectively. Both validation using the test data and bootstrap resampling consistently demonstrated recall, precision, and F1 score exceeding 0.9. The model identified the peak value of C-reactive protein during the first 7 postoperative days as the most significant feature, followed by the preoperative neutrophil-to-lymphocyte ratio. This study highlights the potential of the ML approach to predict PPH occurrence following PD. Vigilance and early interventions guided by such model predictions could positively impact outcomes for high-risk patients.

An Adaptive Spatio-temporal Global Sampling for Presentation Attack Detection

Without developing dedicated countermeasures, facial biometric systems can be spoofed with printed photos, replay attacks, silicone masks, or even a 3D mask of a targeted person. Thus, the threat of presentation attacks needs to be addressed to strengthen the security of the biometric systems. Since a 2D convolutional neural network (CNN) captures static features from video frames, the camera motion might hinders the performance of modern CNNs for video-based presentation attack detection (PAD). Inspired by the egomotion theory, we introduce an adaptive spatiotemporal global sampling (ASGS) technique to compensate the camera motion and use the resulting estimation to encode the appearance and dynamics of the video sequences into a single RGB image. This is achieved by adaptively splitting the video into small segments and capturing their global motion within each segment. The proposed global motion is estimated based on four key steps: dense sampling, FREAK feature extraction and matching, similarity transformation, and aggregation function. This allows using deep models pre-trained on images for video-based PAD detection. Moreover, the interpretation of ASGS reveals that the most important parts for supporting the decision on PAD are consistent with motion cues associated with the artifacts, i.e., hand movement, material reflection, and expression changes. Extensive experiments on four standard face PAD databases demonstrate its effectiveness and encourage further study in this domain.

Optimal Design of a Surface Permanent Magnet Machine for Electric Power Steering Systems in Electric Vehicle Applications Using a Gaussian Process-Based Approach

The efficient design optimization of electric machines for electric power steering (EPS) applications poses challenges in meeting demanding performance criteria, including high power density, efficiency, and low vibration. Traditional optimization approaches often fail to find a global solution or suffer from excessive computation time. In response to the limitations of traditional approaches, this paper introduces a novel methodology by incorporating a Gaussian process-based adaptive sampling technique into a surrogate-assisted optimization process using a metaheuristic algorithm. Validation on a 72-slot/8-pole interior permanent magnet (IPM) machine demonstrates the superiority of the proposed approach, showcasing improved exploitation–exploration balance, faster convergence, and enhanced repeatability compared to conventional optimization methods. The proposed design process is then applied to two surface PM (SPM) machine configurations with 9-slot/6-pole and 12-slot/10-pole combinations for EPS applications. The results indicate that the 12-slot/10-pole SPM design surpasses the alternative design in torque density, efficiency, cogging torque, torque ripple, and manufacturability.

Design and analysis of thermoelectric energy harvester with 85.6% peak end-to-end efficiency and 38 mV self-startup using 4.13 mΩ/°C indirect temperature dependent MPPT

We present an efficient power management system (PMS) to harvest energy from thermoelectric energy generators (TEGs) via indirect temperature tracking. The PMS includes a power converter with an indirect temperature-dependent (ITD) maximum power point tracking (MPPT) controller and two complementary self-startups. The ITD-MPPT controller consists of an ITD input sampling circuit (ITD-ISC) and an ITD impedance matching circuit (ITD-IMC). The ITD-ISC detects and samples the TEG voltage (VTEG) under dynamically varying temperature gradients. The ITD-ISC removes unnecessary detachments of the converter from multi-mode reconfigurable TEGs to recover the otherwise wasted power using an adaptive input sampling (AIS) technique. The ITD-IMC adjusts the on-time of the power switch to control the input impedance (RIN) of the converter with a relatively high resolution of 4.13 mΩ/°C. This approach provides an ultrawide range source tracking from 2 Ω to 200 Ω. We also present a precision zero current switching (ZCS) controller to manage the on-time of the two power switches simultaneously. Two complementary self-startup components are realized using transformer-based (TSC) and charge-pump-based startup circuits (CSC). Integrated circuits for the PMS are fabricated using 180 nm and 28 nm CMOS processes for high efficiency and low voltage, respectively. The measurements indicate that the proposed PMS achieves an end-to-end efficiency (ηEE) above 80% over a wide source VTEG range (100 mV – 450 mV). The peak ηEE of 85.6% is achieved for VTEG = 200 mV, demonstrating the high-performance MPPT using indirect temperature tracking. TSC and CSC demonstrate successful self-startups at input voltages of 38 mV and 200 mV, respectively, thus providing reliable startup methods for various renewable energy sources.

A novel sampled-data asynchronous CatBoost model with knowledge-based guidance for product quality estimation in the industrial processes

For process industry with asynchronous sampling, delayed laboratory analysis of the product quality restricts operation control and optimization. However, it is challenging to ensure the estimation accuracy due to the fact that online estimation model is disturbed by the dynamic change of production conditions. Therefore, a product quality estimation strategy for asynchronous sampling industrial processes under data-driven and knowledge-based guidance is proposed. First, domain knowledge and process mechanism are used to establish the mechanics models and identify the critical variables. Then, the dimensionality of the samples is reduced to extract effective information which is used to partition data samples. Besides, the data sample is balanced by the adaptive synthetic sampling technique under different production conditions. Finally, a CATBoost estimation model is built for product quality online acquisition to improve interpretability and adaptability. The reliability of the proposed estimation approach is verified by a sodium aluminate solution evaporation application case.

Aerodynamic Uncertainty Quantification of a Low-Pressure Turbine Cascade by an Adaptive Gaussian Process

Stochastic variations of the operation conditions and the resultant variations of the aerodynamic performance in Low-Pressure Turbine (LPT) can often be found. This paper studies the aerodynamic performance impact of the uncertain variations of flow parameters, including inlet total pressure, inlet flow angle, and turbulence intensity for an LPT cascade. Flow simulations by solving the Reynolds-averaged Navier–Stokes equations, the SST turbulence model, and γ−Re˜θt transition model equations are first carried out. Then, a Gaussian process (GP) based on an adaptive sampling technique is introduced. The accuracy of adaptive GP (ADGP) is proven to be high through a function experiment. Using ADGP, the uncertainty propagation models between the performance parameters, including total pressure-loss coefficient, outlet flow angle, Zweifel number, and the uncertain inlet flow parameters, are established. Finally, using the propagation models, uncertainty quantifications of the performance changes are conducted. The results demonstrate that the total pressure-loss coefficient and Zweifel number are sensitive to uncertainties, while the outlet flow angle is almost insensitive. Statistical analysis of the flow field by direct Monte Carlo simulation (MCS) shows that flow transition on the suction side and viscous shear stress are rather sensitive to uncertainties. Moreover, Sobol sensitivity analysis is carried out to specify the influence of each uncertainty on the performance changes in the turbine cascade.

Adaptive device sampling and deadline determination for cloud-based heterogeneous federated learning

As a new approach to machine learning, Federated learning enables distributned traiing on edge devices and aggregates local models into a global model. The edge devices that participate in federated learning are highly heterogeneous in terms of computing power, device state, and data distribution, making it challenging to converge models efficiently. In this paper, we propose FedState, which is an adaptive device sampling and deadline determination technique for cloud-based heterogeneous federated learning. Specifically, we consider the cloud as a central server that orchestrates federated learning on a large pool of edge devices. To improve the efficiency of model convergence in heterogeneous federated learning, our approach adaptively samples devices to join each round of training and determines the deadline for result submission based on device state. We analyze existing device usage traces to build device state models in different scenarios and design a dynamic importance measurement mechanism based on device availability, data utility, and computing power. We also propose a deadline determination module that dynamically sets the deadline according to the availability of all sampled devices, local training time, and communication time, enabling more clients to submit local models more efficiently. Due to the variability of device state, we design an experience-driven algorithm based on Deep Reinforcement Learning (DRL) that can dynamically adjust our sampling and deadline policies according to the current environment state. We demonstrate the effectiveness of our approach through a series of experiments with the FMNIST dataset and show that our method outperforms current state-of-the-art approaches in terms of model accuracy and convergence speed.

Utilizing soft labels of Convolutional Neural networks for predicting quantum transmission probability: An application to high-dimensional data prediction

We propose a method that utilizes Convolutional Neural Network (CNN) models for predicting continuous physical observables. While CNN models are typically used for solving classification problems, our method utilizes the soft labels of CNN models to predict high-dimensional data. In this study, we successfully apply this method to predict the transmission probability data of a quantum-dot chain system as a function of energy. To optimize the model performance, we explore three different sampling schemes to construct the training datasets. Furthermore, we compare the performance of CNN models with that of Feed-Forward Neural Network models. Our results clearly demonstrate the potential of using soft labels in CNNs for high-dimensional data prediction and highlight the advantages of adaptive sampling technique to improve model’s performance.

TurboGNN: Improving the End-to-End Performance for Sampling-Based GNN Training on GPUs

<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Graph Neural Networks</i> (GNN) have evolved as powerful models for graph representation learning. Sampling-based training methods have been introduced to train large graphs without compromising accuracy. However, it is challenging for the existing GNN systems to effectively utilize multi-core accelerators, especially GPUs, due to a large number of atomic operations and unbalanced workload originating from the serial execution of multiple GNN processing stages. In this paper, we propose a combination of optimization techniques to accelerate the end-to-end performance of the sampling-based GNN training process. Specifically, we propose an adaptive share memory-based sampling technique and a degree-guided thread block scheduling strategy to optimize the graph sampling. Further, based on the observations of resource demand in different training stages, we propose an asynchronous pipeline-based scheduling method, which accelerates the GNN training by decoupling different training stages into a pipeline and therefore improves the GPU resource utilization significantly. The experimental results show that compared with the existing work, the proposed methods can achieve up to 5.6X performance speedup in the end-to-end performance.

Adaptive symplectic model order reduction of parametric particle-based Vlasov–Poisson equation

High-resolution simulations of particle-based kinetic plasma models typically require a high number of particles and thus often become computationally intractable. This is exacerbated in multi-query simulations, where the problem depends on a set of parameters. In this work, we derive reduced order models for the semi-discrete Hamiltonian system resulting from a geometric particle-in-cell approximation of the parametric Vlasov–Poisson equations. Since the problem’s nondissipative and highly nonlinear nature makes it reducible only locally in time, we adopt a nonlinear reduced basis approach where the reduced phase space evolves in time. This strategy allows a significant reduction in the number of simulated particles, but the evaluation of the nonlinear operators associated with the Vlasov–Poisson coupling remains computationally expensive. We propose a novel reduction of the nonlinear terms that combines adaptive parameter sampling and hyper-reduction techniques to address this. The proposed approach allows decoupling the operations having a cost dependent on the number of particles from those that depend on the instances of the required parameters. In particular, in each time step, the electric potential is approximated via dynamic mode decomposition (DMD) and the particle-to-grid map via a discrete empirical interpolation method (DEIM). These approximations are constructed from data obtained from a past temporal window at a few selected values of the parameters to guarantee a computationally efficient adaptation. The resulting DMD-DEIM reduced dynamical system retains the Hamiltonian structure of the full model, provides good approximations of the solution, and can be solved at a reduced computational cost.

Physics-informed neural networks with residual/gradient-based adaptive sampling methods for solving partial differential equations with sharp solutions

We consider solving the forward and inverse partial differential equations (PDEs) which have sharp solutions with physics-informed neural networks (PINNs) in this work. In particular, to better capture the sharpness of the solution, we propose the adaptive sampling methods (ASMs) based on the residual and the gradient of the solution. We first present a residual only-based ASM denoted by ASM I. In this approach, we first train the neural network using a small number of residual points and divide the computational domain into a certain number of sub-domains, then we add new residual points in the sub-domain which has the largest mean absolute value of the residual, and those points which have the largest absolute values of the residual in this sub-domain as new residual points. We further develop a second type of ASM (denoted by ASM II) based on both the residual and the gradient of the solution due to the fact that only the residual may not be able to efficiently capture the sharpness of the solution. The procedure of ASM II is almost the same as that of ASM I, and we add new residual points which have not only large residuals but also large gradients. To demonstrate the effectiveness of the present methods, we use both ASM I and ASM II to solve a number of PDEs, including the Burger equation, the compressible Euler equation, the Poisson equation over an L-shape domain as well as the high-dimensional Poisson equation. It has been shown from the numerical results that the sharp solutions can be well approximated by using either ASM I or ASM II, and both methods deliver much more accurate solutions than the original PINNs with the same number of residual points. Moreover, the ASM II algorithm has better performance in terms of accuracy, efficiency, and stability compared with the ASM I algorithm. This means that the gradient of the solution improves the stability and efficiency of the adaptive sampling procedure as well as the accuracy of the solution. Furthermore, we also employ the similar adaptive sampling technique for the data points of boundary conditions (BCs) if the sharpness of the solution is near the boundary. The result of the L-shape Poisson problem indicates that the present method can significantly improve the efficiency, stability, and accuracy.

Machine Learning-Based Adaptive Synthetic Sampling Technique for Intrusion Detection

Traditional firewalls and data encryption techniques can no longer match the demands of current IoT network security due to the rising amount and variety of network threats. In order to manage IoT network risks, intrusion detection solutions have been advised. Even though machine learning (ML) helps the widely used intrusion detection techniques currently in use, these algorithms struggle with low detection rates and the requirement for extensive feature engineering. The deep learning model for IoT network intrusion detection is a method for traffic anomaly detection that is suggested by this study. To extract the sequence properties of data flow through a CNN, it combines an attention mechanism with a Long Short Term Memory (LSTM) network. This method uses adaptive synthetic sampling (ADASYN) to increase the size of minority-class samples. The proposed models demonstrated acceptable precision and recall for each class when used for binary-class classification, proving their stability and capacity to identify all classes correctly. The MLP classifier’s accuracy, precision, recall, and F1 value were 87%, 89%, 87%, and 89%, respectively, with an AUC score of 0.88. Overall, the proposed models performed well. The attack and all-class models exhibited good AUCs and macro metrics, the same as the proposed MLP classifier, which had an F1 score of 83% and an AUC score of 0.94. Additionally, it trained the MLP classifier and integrated the ADAM optimizer and category cross-entropy loss function for all-class classification. With an AUC value of 94%, it possessed 84% accuracy, 87% precision, 84% recall, and an 83% F1 score. A further indication of the hybrid model’s ability to combine the benefits of both models to improve overall performance was that it regularly outperformed the MLP model. This model’s accuracy and F1 score are better than those of earlier comparable algorithms, according to experimental results using the publicly accessible benchmark dataset for network intrusion detection (NSL–KDD).

A scalable sparse neural network framework for rare cell type annotation of single-cell transcriptome data

Automatic cell type annotation methods are increasingly used in single-cell RNA sequencing (scRNA-seq) analysis due to their fast and precise advantages. However, current methods often fail to account for the imbalance of scRNA-seq datasets and ignore information from smaller populations, leading to significant biological analysis errors. Here, we introduce scBalance, an integrated sparse neural network framework that incorporates adaptive weight sampling and dropout techniques for auto-annotation tasks. Using 20 scRNA-seq datasets with varying scales and degrees of imbalance, we demonstrate that scBalance outperforms current methods in both intra- and inter-dataset annotation tasks. Additionally, scBalance displays impressive scalability in identifying rare cell types in million-level datasets, as shown in the bronchoalveolar cell landscape. scBalance is also significantly faster than commonly used tools and comes in a user-friendly format, making it a superior tool for scRNA-seq analysis on the Python-based platform.