Training Phase Research Articles

Machine learning predictive insight of water pollution and groundwater quality in the Eastern Province of Saudi Arabia

This study presents an innovative approach for predicting water and groundwater quality indices (WQI and GWQI) in the Eastern Province of Saudi Arabia, addressing critical challenges of scarcity and pollution in arid regions. Recent literature highlights the increasing attention towards WQI based on water pollution index (WPI) and GWQI as essential tools for simplifying complex hydrogeological data, thereby facilitating effective groundwater management and protection. Unlike previous works, the present research introduces a novel hybrid method that integrates non-parametric kernel Gaussian learning (GPR), adaptive neuro-fuzzy inference system (ANFIS), and decision tree (DT) algorithms. This approach marks the first application of a non-parametric kernel for groundwater quality pollution index prediction in Saudi Arabia, offering a significant advancement in the field. Through laboratory analysis and the combination of various machine learning (ML) techniques, this study enhances prediction capabilities, particularly for unmonitored sites in arid and semi-arid regions. The study’s objectives include feature engineering based on dependency sensitivity analysis to identify the most influential variables affecting WQI and GWQI, and the development of predictive models using ANFIS, GPR, and DT for both indices. Furthermore, it aims to assess the impact of different data portions on WQI and GWQI predictions, exploring data divisions such as (70% / 30%), (60% / 40%), and (80% / 20%) for training and testing phase, respectively. By filling a critical gap in water resource management, this research offers significant implications for the prediction of water quality in regions facing similar environmental challenges. Through its innovative methodology and comprehensive analysis, this study contributes to the broader effort of managing and protecting water resources in arid and semi-arid areas. The result proved that GPR-M1 exhibited exceptional testing phase accuracy with RMSE = 0.0169 for GWQI. Similarly, for WPI, the ANFIS-M1 achieved high testing predictive skills with RMSE = 0.0401. The results emphasize the critical role of data quality and quantity in training for enhancing model robustness and prediction precision in water quality assessment.

Rock dynamic strength prediction in cold regions using optimized hybrid algorithmic models

Predicting the dynamic mechanical characteristics of rocks during freeze–thaw cycles (FTC) is crucial for comprehending the damage process of FTC and averting disasters in rock engineering in cold climates. Nevertheless, the conventional mathematical regression approach has constraints in accurately forecasting the dynamic compressive strength (DCS) of rocks under these circumstances. Hence, this study presents an optimized approach by merging the Coati Optimization Algorithm (COA) with Random Forest (RF) to offer a reliable solution for nondestructive prediction of DCS of rocks in cold locations. Initially, a database of the DCS of rocks after a series of FTC was constructed, and these data were obtained by performing the Split Hopkinson Pressure Bar Test on rocks after FTC. The main influencing factors of the test can be summarized into 10, and PCA was employed to decrease the number of dimensions in the dataset, and the microtests were used to explain the mechanism of the main influencing factors. Additionally, the Backpropagation Neural Network and RF are used to construct the prediction model of DCS of rock, and six optimization techniques were employed for optimizing the hyperparameters of the model. Ultimately, the 12 hybrid prediction models underwent a thorough and unbiased evaluation utilizing a range of evaluation indicators. The outcomes of the research concluded that the COA-RF model is most recommended for application in engineering practice, and it achieved the highest score of 10 in the combined score of the training and testing phases, with the lowest RMSE (4.570,8.769), the lowest MAE (3.155,5.653), the lowest MAPE (0.028,0.050), the highest R2 (0.983,0.94).

Parallel hyperparameter optimization of spiking neural networks

Hyperparameter optimization of spiking neural networks (SNNs) is a difficult task which has not yet been deeply investigated in the literature. In this work, we designed a scalable constrained Bayesian based optimization algorithm that prevents sampling in non-spiking areas of an efficient high dimensional search space. These search spaces contain infeasible solutions that output no or only a few spikes during the training or testing phases, we call such a mode a “silent network”. Finding them is difficult, as many hyperparameters are highly correlated to the architecture and to the dataset. We leverage silent networks by designing a spike-based early stopping criterion to accelerate the optimization process of SNNs trained by spike timing dependent plasticity and surrogate gradient. We parallelized the optimization algorithm asynchronously, and ran large-scale experiments on heterogeneous multi-GPU Petascale architecture. Results show that by considering silent networks, we can design more flexible high-dimensional search spaces while maintaining a good efficacy. The optimization algorithm was able to focus on networks with high performances by preventing costly and worthless computation of silent networks.

Statistically learned associations among objects bias attention.

A growing body of research suggests that semantic relationships among objects can influence the control of attention. There is also some evidence that learned associations among objects can bias attention. However, it is unclear whether these findings are due to statistical learning or existing semantic relationships. In the present study, we examined whether statistically learned associations among objects can bias attention in the absence of existing semantic relationships. Participants searched for one of four targets among pairs of novel shapes and identified whether the target was present or absent from the display. In an initial training phase, each target was paired with an associated distractor in a fixed spatial configuration. In a subsequent test phase, each target could be paired with the previously associated distractor or a different distractor. In our first experiment, the previously associated distractor was always presented in the same pair as the target. Participants were faster to respond when this distractor was present on target-present trials. In our second experiment, the previously associated distractor was presented in a different pair than the target in the test phase. In this case, participants were slower to respond when this distractor was present on both target-present and target-absent trials. Together, these findings provide clear evidence that statistically learned associations among objects can bias attention, analogous to the effects of semantic relationships on attention.

Machine learning based prospect targeting: A case of gold occurrence in central parts of Tanzania, East Africa

Soil geochemical analyses from central Tanzania reveal significant gold (Au) values, highlighting the potential for further exploration in the region. This study employs ensemble machine learning models—XGBoost-RF, XGBoost-SVM, and XGBoost-ANN—to enhance predictions of Au distribution. Among these, the XGBoost-ANN model showed the highest accuracy during the training phase, achieving a Mean Absolute Percentage Error (MAPE) of 1.275, a Root Mean Square Error (RMSE) of 0.031, an R² of 0.999, and a Pearson Correlation Coefficient (PCC) of 0.999. However, its performance declined in the testing phase with a MAPE of 0.0668 and an RMSE of 0.2491, indicating reduced predictiveness on new data. Spatial analyses using Global and Local Moran's I tests revealed no significant global spatial autocorrelation but identified localized clusters of high and low Au concentrations. Specific areas showed significant spatial dependence, enhancing our understanding of the complex geospatial distribution of Au. These findings support the combined use of predictive modeling and spatial statistical methods to refine mineral exploration strategies, highlighting the value of advanced analytics in identifying promising exploration targets.

The Cumulative Impacts of Fatigue during Overload Training Can Be Tracked Using Field-Based Monitoring of Running Stride Interval Correlations.

Integrating running gait coordination assessment into athlete monitoring systems could provide unique insight into training tolerance and fatigue-related gait alterations. This study investigated the impact of an overload training intervention and recovery on running gait coordination assessed by field-based self-testing. Fifteen trained distance runners were recruited to perform 1-week of light training (baseline), 2 weeks of heavy training (high intensity, duration, and frequency) designed to overload participants, and a 10-day light taper to allow recovery and adaptation. Field-based running assessments using ankle accelerometry and online short recovery and stress scale (SRSS) surveys were completed daily. Running performance was assessed after each training phase using a maximal effort multi-stage running test-to-exhaustion (RTE). Gait coordination was assessed using detrended fluctuation analysis (DFA) of a stride interval time series. Two participants withdrew during baseline training due to changed personal circumstances. Four participants withdrew during heavy training due to injury. The remaining nine participants completed heavy training and were included in the final analysis. Heavy training reduced DFA values (standardised mean difference (SMD) = -1.44 ± 0.90; p = 0.004), recovery (SMD = -1.83 ± 0.82; p less than 0.001), performance (SMD = -0.36 ± 0.32; p = 0.03), and increased stress (SMD = 1.78 ± 0.94; p = 0.001) compared to baseline. DFA values (p = 0.73), recovery (p = 0.77), and stress (p = 0.73) returned to baseline levels after tapering while performance trended towards improvement from baseline (SMD = 0.28 ± 0.37; p = 0.13). Reduced DFA values were associated with reduced performance (r2 = 0.55) and recovery (r2 = 0.55) and increased stress (r2 = 0.62). Field-based testing of running gait coordination is a promising method of monitoring training tolerance in running athletes during overload training.

Protection of Press Publishers in the Age of Generative AI – In Search of Legal Remedies to Adapt to the Pace of Technology

AbstractHas everything already been said on the protection of press publishers? Surely not, and this is due to the new challenges for the press sector posed by Generative AI. This paper aims to discuss the solutions adopted in EU copyright law to safeguard the legal situation of press publishers in the face of the development of Generative AI. It identifies the shortcomings in existing forms of protection and proposes legal remedies for those shortcomings. It offers a fresh perspective on and insights into the new challenges posed by Generative AI, especially regarding its training phase. It explores the approach taken by EU legislation, and also makes reference to solutions proposed in the United States. The objective of this paper is to strike a balance between the need to ensure proper conditions for innovation and the need to protect human creativity and the dissemination thereof. It seeks to contribute to the ongoing debate on copyright and AI by assessing how the current copyright framework suits the development of Generative AI, identifying gaps in protection and paving the way for necessary legal changes in a context of transparency, compensation for use of protected content for AI training purposes, and preservation of rights.

Application of Target Detection Based on Deep Learning in Intelligent Mineral Identification

In contemporary society, rich in mineral resources, efficiently and accurately identifying and classifying minerals has become a prominent issue. Recent advancements in artificial intelligence, particularly breakthroughs in deep learning, have offered new solutions for intelligent mineral recognition. This paper introduces a deep-learning-based object detection model for intelligent mineral identification, specifically employing the YOLOv8 algorithm. The model was developed with a focus on seven common minerals: biotite, quartz, chalcocite, silicon malachite, malachite, white mica, and pyrite. During the training phase, the model learned to accurately recognize and classify these minerals by analyzing and annotating a large dataset of mineral images. After 258 rounds of training, a stable model was obtained with high performance on key indicators such as Precision, Recall, mAP50, and mAP50–95, with values stable at 0.91766, 0.89827, 0.94300, and 0.91696, respectively. In the testing phase, using samples provided by the Geological and Mineral Museum at the School of Earth Sciences and Engineering, Sun Yat-sen University, the model successfully identified all test samples, with 83% of them having a confidence level exceeding 87%. Despite some potential misclassifications, the results of this study contribute valuable insights and practical experience to the development of intelligent mineral recognition technologies.

Grey-box solution for predicting thermo-mechanical response of rocks

Evaluating the thermo-mechanical response of rocks under high temperature treatments is crucial for various engineering geology projects. Current predictions of rock thermo-mechanical response rely on simplistic mathematical fittings treating temperature as a reduction factor, while existing machine learning algorithms often present practical challenges due to their black-box solutions. In this study, highly practical grey-box solutions, utilizing gene expression programming (GEP) are proposed for forecasting rock strength following high-temperature treatments. The dataset, comprising temperature, rock type, rock density, sample size, crack damage stress, confining pressure, and elastic modulus, serves as input parameters, with rock strength from triaxial compression tests as the output. Three grey-box solutions (mathematical formulations) based on distinct input parameter sets are proposed, all demonstrating excellent accuracy with high R2-values (R2 > 0.95) and low error values across both the training and testing phases. Feature importance analysis highlights crack damage stress, confining pressure, and elastic modulus as statistically significant parameters influencing the strength of rocks subjected to high temperatures. External validation of the proposed models indicates strong generalization capabilities, underscoring their ability to perform well beyond the training data. Furthermore, a monotonicity study demonstrates that the proposed models align with the expected physical processes. The proposed formulations offer valuable field implications, effectively addressing the limitations of labor-intensive and costly laboratory processes for evaluating rock thermo-mechanical responses.

Open-Set Recognition Model for SAR Target Based on Capsule Network with the KLD

Synthetic aperture radar (SAR) automatic target recognition (ATR) technology has seen significant advancements. Despite these advancements, the majority of research still operates under the closed-set assumption, wherein all test samples belong to classes seen during the training phase. In real-world applications, however, it is common to encounter targets not previously seen during training, posing a significant challenge to the existing methods. Ideally, an ATR system should not only accurately identify known target classes but also effectively reject those belonging to unknown classes, giving rise to the concept of open set recognition (OSR). To address this challenge, we propose a novel approach that leverages the unique capabilities of the Capsule Network and the Kullback-Leibler divergence (KLD) to distinguish unknown classes. This method begins by deeply mining the features of SAR targets using the Capsule Network and enhancing the separability between different features through a specially designed loss function. Subsequently, the KLD of features between a testing sample and the center of each known class is calculated. If the testing sample exhibits a significantly larger KLD compared to all known classes, it is classified as an unknown target. The experimental results of the SAR-ACD dataset demonstrate that our method can maintain a correct identification rate of over 95% for known classes while effectively recognizing unknown classes. Compared to existing techniques, our method exhibits significant improvements.

Parallel PSO for Efficient Neural Network Training Using GPGPU and Apache Spark in Edge Computing Sets

The training phase of a deep learning neural network (DLNN) is a computationally demanding process, particularly for models comprising multiple layers of intermediate neurons.This paper presents a novel approach to accelerating DLNN training using the particle swarm optimisation (PSO) algorithm, which exploits the GPGPU architecture and the Apache Spark analytics engine for large-scale data processing tasks. PSO is a bio-inspired stochastic optimisation method whose objective is to iteratively enhance the solution to a (usually complex) problem by approximating a given objective. The expensive fitness evaluation and updating of particle positions can be supported more effectively by parallel processing. Nevertheless, the parallelisation of an efficient PSO is not a simple process due to the complexity of the computations performed on the swarm of particles and the iterative execution of the algorithm until a solution close to the objective with minimal error is achieved. In this study, two forms of parallelisation have been developed for the PSO algorithm, both of which are designed for execution in a distributed execution environment. The synchronous parallel PSO implementation guarantees consistency but may result in idle time due to global synchronisation. In contrast, the asynchronous parallel PSO approach reduces the necessity for global synchronization, thereby enhancing execution time and making it more appropriate for large datasets and distributed environments such as Apache Spark. The two variants of PSO have been implemented with the objective of distributing the computational load supported by the algorithm across the different executor nodes of the Spark cluster to effectively achieve coarse-grained parallelism. The result is a significant performance improvement over current sequential variants of PSO.

Customized Convolutional Neural Network for Glaucoma Detection in Retinal Fundus Images

Glaucoma is one of the leading causes of permanent blindness and remains a current challenge in the field of ophthalmology. This research aims to present a comprehensive investigation into the development and evaluation of new technology for glaucoma detection in retinal fundus images. The development and evaluation are presented on a customized architecture, using the Convolutional Neural Network (CNN) method. The proposed CNN architecture is designed to address the complex characteristics of glaucoma changes in the identification process. The research dataset consists of 506 retinal images categorized into 117 glaucoma images, 19 suspected glaucoma images, and 370 healthy images. Through our in-depth exploration, we conducted a careful analysis to uncover patterns and fundamental trends related to glaucoma-related features. During the training phase, the proposed CNN achieved outstanding average accuracy, sensitivity, and specificity values of 92.88%, 94.66%, and 89.31%, respectively. In the unseen test dataset, the model demonstrated competitive performance with an accuracy of 80.87%, sensitivity of 85.65%, and specificity of 71.26%. These findings emphasize the potential of the model as a reliable tool for glaucoma detection. The results indicate that the proposed method utilizing a customized CNN architecture is designed for glaucoma detection in retinal fundus images. The presented output results also hold promise for clinical relevance and can be considered an improvement in the care of retinal fundus patients.

Evaluation of stress distributions in trimaterial bonded joints with nano-resin adhesive using machine learning models

Adhesive bonded joints hold significant importance across various industrial sectors in modern engineering, owing to their lightweight nature and myriad advantages. The rising demand for trimaterial joints underscores their utility and versatility. In these joints, the choice of materials for both adherends greatly influences their strength, structural reliability, and overall characteristics. While numerous researches have extensively analyzed stress distributions, their effects, and behaviors, many have relied on a one-factor-at-a-time approach, focusing solely on individual design variables' effects. However, recognizing the intricate interplay among various material combinations and their collective impact on overall performance, this study employs various types of White-box, Black-box, and Grey-box machine learning algorithms to identify an optimized ML model as well as predict stress distributions for any random combinations of upper and lower adherend materials. Dataset of total 178 random material combinations were utilized for the training phases with 5-fold cross validation and model tuning. However, the decision tree regressor emerged as the optimized model by comparing the quantitative metrics of accuracy benchmark as well as the prediction outcomes obtained through all the machine learning models. The maximum prediction accuracy attained was an impressive 99.97 %, while the minimum recorded was 89.74 %. This research aims to identify tailored machine learning model specifically for trimaterial bonded joints where nano layer of resin is utilized as the adhesive.

FCGNet: Foreground and Class Guided Network for human parsing

Understanding the inherent hierarchical human structure is key to human parsing. To capture the human-specific characteristic, it is necessary to focus on the spatial and class information corresponding to the foreground (i.e., human) in an image. Inspired by these insights, we introduce two supervision signals, spatial foreground information and existent class information in the image. By utilizing foreground information as guidance, the network is guided to generate a human-focused feature map and capture the pixel-wise hierarchical characteristics by computing correlations between pixels. Furthermore, we guide the network to consider class information in the image at the feature level and capture the class-wise relationship by calculating correlations between channels. Moreover, during the training phase, we prevent the network from misclassifying pixels into confusing classes by providing the existent class information in the image to the network at the prediction level. Our model achieves state-of-the-art performance with significantly reduced parameters and Multiply-Accumulate Operations (MACs) in three public benchmarks.

Relationships among fall cone flow index and consistency identifiers of fine-grained soils with emphasis on toughness limit

It is well known that plasticity characteristics are one of the most important factors controlling the engineering behavior of fine-grained soils. In this regard, Atterberg limits are used for evaluation of plasticity and strength behavior of soils. Of all the plasticity identifiers, fall cone flow index, which is the slope of the flow curve is a descriptor of plasticity and undrained shear strength characteristics of cohesive soils. On the other hand, toughness limit is also efficient in prediction of many index properties of soils including plasticity characteristics, compression index as well as residual friction angle. This study is an attempt to establish relationships among fall cone flow index and consistency identifiers of fine-grained soils. Emphasis is given on toughness limit, which is believed to be a practical and efficient approach in assessment of fall cone flow index. Regression-based equations establishing relationships between fall cone flow index, plasticity index, plasticity ratio, toughness limit and liquid limit were presented. Besides, the efficiency of artificial neural networks in prediction of toughness limit and fall cone flow index was investigated. It should be noted that established relationships using regression equations are based on regional data, which come up with coefficient of determination (R2) values ranging between 0.706 and 0.978. On the other hand, ANN models were used to model global data, R2 values were obtained to be between 0.586-0.972 and 0.522-0.889 for training and testing phases, respectively. Relationships concerning local and global data were presented, along with analysis of effectiveness of application of relationships established for analysis of global data.

Examination of Reflective Practice Levels in Tailored Structured Observation Journals

Reflective practice (RP) is crucial in the professional development of teachers. This process enables them to critically analyse their teaching practices leading to informed decisions that help them progress as practitioners. However, the level of reflection may vary from one teacher to another. The analysis of documents such as classroom observation grids and reflective journals can reveal the different levels of reflection. Both tools are essential to trigger reflection in student teachers and stimulate them to seek ways to improve their teaching practice during and after the training phase. Following a qualitative research design that employed document analysis as a research instrument, this study examined the different levels of RP in tailored structured observation journals. The participants of the study were seventeen fourth- and fifth-year English as Foreign Language (EFL) preservice teachers at École Normale Supérieure de Constantine (ENSC), Algeria who filled in structured observation journals during the observation phase of the practical training. Following the framework of Larrivee (2008) where three levels of reflection were identified: surface, pedagogical, and critical, the structured observation journals were analysed. The results of the study showed that surface reflection was the most documented level with 183 instances (n=183) related to the different teaching methods and strategies followed by pedagogical reflection (n=99) and critical reflection (n=11). The study revealed that the structured observation journals assisted preservice teachers in being more reflective while observing lessons. However, they still needed to enhance their critical reflection skills to improve their reflective practice.

Building reliable AI for quantifying uncertainty in particulate matter predictions with deep learning

Predicting the concentrations of particulate matter (PM) has recently become crucial because of its significant impact on air pollution. As the adoption of artificial intelligence (AI) in this domain increases, ensuring the reliability of AI-driven predictions has become paramount. Although many previous studies have harnessed the power of AI for PM prediction, the inherent uncertainty in the results, which is influenced by spatial data imbalances and varying meteorological factors, poses a challenge. Addressing this uncertainty is central to building reliable AI systems. Therefore, we employed the Monte Carlo dropout (MCDO) approach, which is a technique that integrates dropout layers in neural networks during both the training and inference phases, to estimate the prediction uncertainty. Our objective was to address the challenge of prediction uncertainty in PM concentrations, with the ultimate aim of enhancing the reliability and trustworthiness of AI-driven predictions in air quality forecasting. By applying the MCDO approach to grids formed from multidimensional arrays of air quality and weather data, we obtained a 95% confidence interval for the prediction results, thereby demonstrating the trustworthiness of our model. Our evaluation revealed an R-square greater than 0.97 for PM10 and PM2.5, showcasing the robust and reliable predictive performance of the model. This study highlights not only the accuracy of this approach but also the critical role of quantifying uncertainty in building reliable AI systems. Our findings mark a significant step towards a holistic approach to predictive modeling in which reliability and uncertainty quantification are at the forefront.

Machine learning-based process design of a novel sustainable cooling system

An exponential rise in cooling requirements occurred in the last few years because of rising global surface temperatures, population growth, faster urbanization, and income growth. Developing countries are facing major issues because of the larger impact of these cooling drivers. Conventional vapor compression systems are energy-demanding and involve dangerous chemicals. The current paper proposed an innovative indirect evaporative cooling system with high energy performance, less emissions, and chemical-free operation. To map the full-scale performance, a prototype was developed,and tested in a variety of outside air conditions. Then artificial neural network (ANN)-based machine learning model was developed incorporating important input parameters including outdoor air temperature, air flow rate ratio, working air temperature, and working air wet bulb temperature to predict the supply air temperature. The ANN model having nine neurons in the hidden layer exhibits excellent modeling performance with a coefficient of determination (R2) value of ∼ 1 and root mean square error of 0.046 °C, 0.06 °C, and 0.06 °C in the training, testing, and validation phases respectively. The variable significance analysis carried out by one factor at a time (OFAT) technique reveals that working air inlet temperature is the most important parameter to predict supply temperature with a significance factor of 33 %. According to the combined experimental and ML model, the proposed system generated 130 W of cooling capacity and dropped the temperature by more than 20 °C at 48 °C of outdoor air. The corresponding coefficient of performance achieved (just for cooling) was 32. It is also shown that the enhanced IEC operates steadily in ambient temperatures ranging from 30 to 48 °C and maintains supply air temperatures within the comfort zone of ASHRAE-55 and ISO7730.

Data-based models to investigate protective piles effects on the scour depth about oblong-shaped bridge pier

The primary objective of this study is to examine the suitability of employing a machine learning models (MLMs) comprising Support Vector Machine (SVM), Gene Expression Programming (GEP), and Artificial Neural Network (ANN) algorithms to emulate the influence of geometric and geotechnical parameters of the piles on the scour depth reduction about the oblong-shaped bridge pier. A dataset consisting of 90 laboratory observations was utilized, allocating 70 % of the data for the training and 30 % for the testing the MLMs. An oblong-shaped bridge pier of diameter D, was positioned and subjected to a rectangular flume featuring an erodible bed comprising sand sediment particles of D50 = 1.7 mm under clear water and subcritical flow conditions. The spacing between the piles, LD, the angle of pile orientation with respect to the flow, θ, and the Froude number, Fr, were identified as independent parameters through the dimensional analysis and Buckingham-Π theory. The MLMs’ performance were assessed employing root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), and developed discrepancy ratio (DDR). The outcomes, affirming the predictive capacity of all MLMs included, revealed that the GEP model employing a three-gene configuration exhibited superior accuracy compared to the other two MLMs. Specifically, the performance indicators (RMSE, MAE, R2, DDR) during the training and testing phases were (0.0433, 0.033, 0.981, 6.8) and (0.054, 0.045, 0.958, 4.81) respectively. The sensitivity analysis elucidated the hierarchy of influential variables as follows: Fr, LD and θ.

Identification of Carbon Stars in LAMOST DR9 Based on Deep Learning

Carbon stars play a crucial role in astronomical research and are significant for understanding stellar evolution, measuring cosmic distances, and studying galaxy kinematics. In recent years, identifying carbon stars using machine learning methods and traditional line-index methods has become a research hotspot, but there are still limitations regarding accuracy and automation. In this study, we propose to build a five-class model to identify carbon stars using spectral data from LAMOST DR9. The model achieved 99.45% precision and 91.21% recall on the carbon star testing set. We conducted independent tests using a sample of 1333 known carbon stars that were not used in the training and testing phases, and our model ultimately identified 1199 carbon stars. On this basis, we used this model to screen 11,226,252 spectra of LAMOST DR9 and identified 4383 carbon stars, including 1197 newly discovered carbon stars. To gain a more comprehensive understanding of the characteristics of the 4383 carbon stars obtained, further visual inspection of these spectra was performed to provide more detailed carbon star subtypes.