Higher Accuracy Scores Research Articles

Optimasi Model Semi-Supervised Learning dengan SVM dan Naïve Bayes

<p><em>Determining labels in text mining is the most important thing, apart from using lexicon based, you can also use human understanding to interpret sentences into positive, negative and neutral categories. Sentiment analysis is to measure an opinion taken from tweets to be able to analyze the text. Researchers carry out a learning process and data testing. The focus of this research is to be able to determine the polarization of text into positive, negative and neutral forms using a semi-supervised machine learning model. In the previous data training process, labeling was carried out using human understanding to obtain positive, negative and neutral labels. Next, in the testing process, the data is not labeled, the role of the learning machine is so that the test data gets a label. The semi-supervised learning (SSL) technique is used to label unlabeled data with the algorithm used to process the training data using SVM and NB. The statistical evaluation used is cross validation and to measure the level of accuracy of the two algorithms using a confusion matrix. SVM received a high accuracy score in this study compared to NB, SVM got a high accuracy score in this study compared to NB, SVM got 88.97% accuracy and NB 83.02%</em></p>

DEEP LEARNING-BASED ADAPTIVE ENSEMBLE LEARNING MODEL FOR CLASSIFICATION OF MONKEYPOX DISEASE

Monkeypox a viral disease resembling smallpox often transmitted via animal contact or human-to-human transmission. Symptoms include fever, rash, and respiratory issues. Healthcare experts initially may confuse it with chickenpox or measles due to its rarity, but swollen lymph nodes typically distinguish it. Diagnosis involves tissue sampling and polymerase chain reaction (PCR) testing, although PCR tests have limitations like time consumption and false negatives. Deep learning-based detection offers advantages over PCR, including reduced risk of exposure, quicker results, and improved accuracy. In this study, a novel adaptive ensemble learning (AEL)-based model for monkeypox diagnosis is proposed. This proposed ensemble learning model aims to enhance diagnosis accuracy by combining different deep learning models, leveraging an adaptive approach for model combination. Experimental studies using MSLD and MSID datasets show promising results, with ensemble models achieving high accuracy, precision, recall, and F1 scores. The ResNet101+VGG16 (92.46% accuracy, 92.75% precision, 93.22% recall, and 92.98% F1 score) ensemble model performs best for MSLD, while DenseNet121+Xception (97.58% accuracy, 96.57% precision, 95.74% recall, and 96.14% F1 score) excels for MSID. In addition, the proposed AEL model outperforms previous studies using the same datasets, showcasing its potential for improved monkeypox diagnosis.

Machine learning-driven prediction of EBNA1 inhibitors against Epstein–Barr virus in nasopharyngeal carcinoma

Nasopharyngeal carcinoma (NPC), particularly prevalent in regions such as Malaysia, is a significant health concern often linked to Epstein-Barr virus (EBV) infection. The EBV nuclear antigen 1 (EBNA1), crucial for EBV survival and NPC tumorigenicity, has emerged as a potential therapeutic target for EBV-positive NPC. In this study, we utilized quantitative structure-activity relationship (QSAR) models to predict potential inhibitors of EBNA1. These models were developed based on the molecular fingerprints of known EBNA1 inhibitors, using both classification and regression approaches. Our QSAR classification models demonstrated consistently high precision, recall, F1 score, and accuracy scores across the training set. The top-performing models, constructed using logistic regression algorithms, achieved perfect precision scores of 1.000 in the test set evaluation. These models’ recall, F1 score, and accuracy scores were 0.571, 0.727, and 0.667, respectively. On the other hand, the best-performing model among the regression models was built using the sequential minimal optimization regression algorithm, achieving a correlation coefficient of 0.703. The mean absolute error and root mean square error of this QSAR regression model were 0.173 and 0.217, respectively, whereas the relative absolute error was 0.689. We screened the enamine advanced compound library using this regression model to predict compounds with potential EBNA1 inhibitory effects. This led to the identification of the top 10 compounds with the most promising predicted EBNA1 inhibitory properties.

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.

Energy Optimization in Ultrasound Tomography Through Sensor Reduction Supported by Machine Learning Algorithms

This paper focuses on reducing energy consumption in ultrasound tomography by utilizing machine learning techniques. The core idea is to investigate the feasibility of minimizing the number of measurement sensors without sacrificing prediction accuracy. This article evaluates the quality of reconstructions derived from data collected through two or three measurement channels. In subsequent steps, machine learning models are developed to predict the number, location, and size of the objects. A reliable object detection method is introduced, requiring less information than traditional signal analysis from multiple channels. Various machine learning models were tested and compared to validate the approach, with most demonstrating high accuracy or R2 scores in their respective tasks. By reducing the number of sensors, the goal is to lower energy usage while maintaining high precision in localization. This study contributes to the ongoing research on energy efficiency in sensing and localization, especially in environments where resource optimization is crucial, such as remote or resource-limited settings.

Deep learning-powered efficient characterization and quantification of microplastics

Characterizing and quantifying microplastics (MPs) are time-consuming and labor-intensive tasks traditionally. This paper presents an artificial intelligence (AI) framework aiming to automate these tasks by integrating computer vision and deep learning techniques. The approach leverages Fourier Transform Infrared (FTIR) spectra and visual images. Primary novelties of this research involve the development of: (1) an AI framework integrating efforts of data processing, analytics, visualization, and human-computer interaction; (2) a method for transforming FTIR data into contour images; (3) data augmentation strategies for resolving data scarcity and imbalance issues; (4) deep learning models for identifying MPs; (5) computer vision algorithms for quantifying MPs; and (6) an engineer-friendly graphic user interface (GUI) for enhancing data accessibility. The AI framework has been applied to polyethylene, polypropylene, polystyrene, polyamide, ethylene-vinyl acetate, and cellulose acetate. Results confirmed the efficacy of the framework, exhibiting high accuracy scores in classification (98 %), segmentation (99 %), and quantification (96 %) tasks. This research advances the capability of automatic assessment of MPs.

Novel IDS Framework: Integrating GWO Feature Selection with KAN Classification

The past few decades have seen significant lifestyle improvements through cutting-edge innovations like AI and fog computing enabled by the World Wide Web, which are crucial for practical applications involving both symmetrical and asymmetrical information distributions. However, these advancements also come with the risk off requent and dangerous cyber-attacks. To mitigate these unwanted malwares, researchers must develop and implement novel intrusion detection systems (IDSs). Challenges such as inadequate accuracy scores due to numerous unnecessary and ineffective features, poor identification of attacks through specific deep learning classification methods, the costly and inefficient use of labeled training databases, and the excessive time required for system development and evaluation hinder there liability and efficiency of IDSs.This paper proposes a hybrid approach that combines the Grey Wolf Optimizer (GWO) for feature selection and the Kolmogorov-Arnold Network(KAN) algorithm for classification to address these challenges.The GWO is utilized to extract important features through meta-heuristic methods, demonstrating its effectiveness in feature selection. The proposed model achieves a high accuracy score, precision, recall, and F1-score of 92.09%, indicating its capability to accurately identify both normal and intrusive activities. The Matthews Correlation Coefficient (MCC) of 0.9019 further validates the model's robustness.

Optimizing ExoMars Rover Remote Sensing Multispectral Science II: Choosing and Using Multispectral Filters for Dynamic Planetary Surface Exploration With Linear Discriminant Analysis

AbstractIn this paper we address two problems associated with data‐limited dynamic spacecraft exploration: data‐prioritization for transmission, and data‐reduction for interpretation, in the context of ESA ExoMars rover multispectral imaging. We present and explore a strategy for selecting and combining subsets of spectral channels captured from the ExoMars Panoramic Camera, and attempt to seek hematite against a background of phyllosilicates and basalts as a test case scenario, anticipated from orbital studies of the rover landing site. We compute all available dimension reductions on the material reflectance spectra afforded by 4 spectral parameter types, and consider all possible paired combinations of these. We then find the optimal linear combination of each pair whilst evaluating the resultant target‐vs.‐background separation in terms of the Fisher Ratio and classification accuracy, using Linear Discriminant Analysis. We find ∼50,000 spectral parameter combinations with a classification accuracy >95% that use 6‐or‐less filters, and that the highest accuracy score is 99.6% using 6 filters, but that an accuracy of >99% can still be achieved with 2 filters. We find that when the more computationally efficient Fisher Ratio is used to rank the combinations, the highest accuracy is 99.1% using 4 filters, and 95.1% when limited to 2 filters. These findings are applicable to the task of time‐constrained planning of multispectral observations, and to the evaluation and cross‐comparison of multispectral imaging systems at specific material discrimination tasks.

Classification of Air Pollutant Index on Data with Outliers and Imbalance Class Problem

The problem of air pollution has become a global issue that has received attention from various countries. Jakarta, Indonesia's capital city, is unavoidable from the same problem. This study will use four parameters of substances PM10, SO2, CO, O3, and nitrogen dioxide to categorize Jakarta's air quality (NO2). The data used is daily data taken from the Air Quality Monitoring Station in Jakarta throughout 2020. The methods used include SVM, Random Forest, Logistic Regression, KNN, CART, and Stacking Algorithm. At the data preparation stage, we found missing values, outliers, and class imbalance problems. Before applying machine learning methods and evaluating accuracy, we used data pre-processing techniques such as the MissForest method, median substitution, and ADASYN. The results prove that the original dataset has a higher accuracy score (0.882 – 0.977) than the balanced dataset (0.829 – 0.976). According to the evaluation results, the Random Forest method has the highest accuracy score for original and balanced datasets. The overall result is better than the identical research, which produces 96.61% accuracy using a neural network. It shows that preprocessing steps such as missing values handling an imbalanced class handling is essential in classification performance.

IoT Sensor Network Electricity Consumption Behaviour Using ClusterAnalysis Algorithm for Network Environment

The Internet of Things (IoT) sensor network plays a crucial role in monitoring electricity consumption behavior by collecting real-time data from various connected devices. Utilizing cluster analysis algorithms, researchers can effectively segment and analyze consumption patterns within a network environment. By grouping similar usage behaviors, these algorithms reveal insights into energy efficiency, peak usage times, and anomalies in consumption. The paper presents a novel methodology for analyzing and classifying users' electricity consumption behavior, utilizing the Clustering Behavior Analysis Weighted Classification (CBAWC) algorithm. By leveraging cluster analysis techniques and weighted classification, the proposed approach allows for the segmentation of consumers into distinct clusters based on their consumption patterns. Through the application of CBAWC, the study provides insights into the diverse behaviors exhibited by consumers, ranging from moderate to high consumption levels, varied peak hours, and peak days. The classification results demonstrate the effectiveness of the algorithm in accurately assigning users to their respective clusters, enabling stakeholders to better understand consumption trends and tailor energy management strategies accordingly. Through the application of CBAWC, consumers are segmented into distinct clusters based on their consumption patterns. For example, clusters exhibit varying average daily consumption levels, with Cluster 1 consuming around 300 kWh, Cluster 2 consuming approximately 450 kWh, and Cluster 3 consuming about 280 kWh on average. Additionally, clusters display different peak hours per day and peak days per month. The classification results demonstrate the algorithm's effectiveness, with high accuracy scores ranging from 0.86 to 1.00 across different user groups.

Integrated Iot Approaches for Crop Recommendation and Yield-Prediction Using Machine-Learning

In this study, we present an integrated approach utilizing IoT data and machine learning models to enhance precision agriculture. We collected an extensive IoT secondary dataset from an online data repository, including environmental parameters such as temperature, humidity, and soil nutrient levels, from various sensors deployed in agricultural fields. This dataset, consisting of over 1 million data points, provided comprehensive insights into the environmental conditions affecting crop yield. The data were preprocessed and used to develop predictive models for crop yield and recommendations. Our evaluation shows that the LightGBM, Decision Tree, and Random Forest classifiers achieved high accuracy scores of 98.90%, 98.48%, and 99.31%, respectively. The IoT data collection enabled real-time monitoring and accurate data input, significantly improving the models’ performance. These findings demonstrate the potential of combining IoT and machine learning to optimize resource use and improve crop management in smart farming. Future work will focus on expanding the dataset to include more diverse environmental factors and exploring the integration of advanced deep learning techniques for even more accurate predictions.

Apple varieties and growth prediction with time series classification based on deep learning to impact the harvesting decisions

Apples are among the most popular fruits globally due to their health and nutritional benefits for humans. Artificial intelligence in agriculture has advanced, but vision, which improves machine efficiency, speed, and production, still needs to be improved. Managing apple development from planting to harvest affects productivity, quality, and economics. In this study, by establishing a vision system platform with a range of camera types that conforms with orchard standard specifications for data gathering, this work provides two new apple collections: Orchard Fuji Growth Stages (OFGS) and Orchard Apple Varieties (OAV), with preliminary benchmark assessments. Secondly, this research proposes the orchard apple vision transformer method (POA-VT), incorporating novel regularization techniques (CRT) that assist us in boosting efficiency and optimizing the loss functions. The highest accuracy scores are 91.56 % for OFGS and 94.20 % for OAV. Thirdly, an ablation study will be conducted to demonstrate the importance of CRT to the proposed method. Fourthly, the CRT outperforms the baselines by comparing it with the standard regularization functions. Finally, time series analyses predict the ‘Fuji’ growth stage, with the outstanding training and validation RMSE being 19.29 and 19.26, respectively. The proposed method offers high efficiency via multiple tasks and improves the automation of apple systems. It is highly flexible in handling various tasks related to apple fruits. Furthermore, it can integrate with real-time systems, such as UAVs and sorting systems. This research benefits the growth of apple’s robotic vision, development policies, time-sensitive harvesting schedules, and decision-making.

Classification of Dopamine D2 receptor ligands using RDKit Molecular descriptors and Machine Learning Algorithms

Identifying and classifying dopamine D2 receptor agonists and antagonists is essential for the drug discovery and development. In this study, we employed machine learning algorithms, namely, XGBoost, LGBM, ExtraTree, and AdaBoost Classifier, in combination with RDKit molecular descriptors, to classify dopamine D2 receptor ligands. The dataset consisted of 195 molecules, comprising 69 dopamine agonists and 126 dopamine antagonists. The models were trained using 75% of the dataset and evaluated on the remaining 25%. The classifiers demonstrated high accuracy and F1 scores, with the AdaBoost Classifier achieving the highest accuracy of 92%. Receiver operating characteristic (ROC) analysis further confirmed the robustness of the model, as indicated by the area under the curve (AUC) values. The AUC values for the AdaBoost, Extra Tree, LGBM, and XGB classifiers were 0.92, 0.90, 0.87, and 0.89, respectively. Feature selection analysis revealed the important molecular descriptors that significantly contribute to the classification models. The ExtraTree classifier selected the highest number of descriptors (167), while the intersection of the selected descriptors among all models indicated 24 common features that crucial for classification. Classification of external compounds using the developed models revealed that sinedabet was classified as a dopamine D2 receptor antagonist, while lisuride, ropinirole, and quinpirole were classified as dopamine D2 receptor agonists.

Parametric optimization and comparative study of machine learning and deep learning algorithms for breast cancer diagnosis.

Breast Cancer is the leading form of cancer found in women and a major cause of increased mortality rates among them. However, manual diagnosis of the disease is time-consuming and often limited by the availability of screening systems. Thus, there is a pressing need for an automatic diagnosis system that can quickly detect cancer in its early stages. Data mining and machine learning techniques have emerged as valuable tools in developing such a system. In this study we investigated the performance of several machine learning models on the Wisconsin Breast Cancer (original) dataset with a particular emphasis on finding which models perform the best for breast cancer diagnosis. The study also explores the contrast between the proposed ANN methodology and conventional machine learning techniques. The comparison between the methods employed in the current study and those utilized in earlier research on the Wisconsin Breast Cancer dataset is also compared. The findings of this study are in line with those of previous studies which also highlighted the efficacy of SVM, Decision Tree, CART, ANN, and ELM ANN for breast cancer detection. Several classifiers achieved high accuracy, precision and F1 scores for benign and malignant tumours, respectively. It is also found that models with hyperparameter adjustment performed better than those without and boosting methods like as XGBoost, Adaboost, and Gradient Boost consistently performed well across benign and malignant tumours. The study emphasizes the significance of hyperparameter tuning and the efficacy of boosting algorithms in addressing the complexity and nonlinearity of data. Using the Wisconsin Breast Cancer (original) dataset, a detailed summary of the current status of research on breast cancer diagnosis is provided.

Deep learning-powered segmentation and classification of diabetic retinopathy for enhanced diagnostic precision

This study addresses the critical challenge of diabetic retinopathy (DR), a severe complication of diabetes that potentially leads to blindness. We introduce a novel approach to DR detection using transfer learning, leveraging a single fundus photograph to automatically identify the disease’s stage. DR progresses through four stages, posing challenges for early detection, with existing methods often inefficient and prone to disagreements among clinicians. The proposed approach demonstrated in the APTOS 2019 Blindness Detection Competition employs convolutional neural networks (CNNs) and achieved a high quadratic weighted kappa score of 0.92546, highlighting its effectiveness in automatic DR detection and emphasizing the need for timely intervention. This paper first reviews related work, spanning classical computer vision methods to deep learning approaches, with a focus on CNNs. Transfer learning with CNN architectures is explored, showcasing promising results from various studies. Identifying two critical gaps in existing literature, the research emphasizes the need for comprehensive exploration into integrating pre-trained large language models (LLMs) with segmented image inputs for generating test/treatment recommendations. In addition, understanding the dynamic interactions among integrated components, including lesion segmentations, disease classification, and LLMs within web applications, remains essential. The objectives of the study include developing a comprehensive DR detection methodology, exploring and implementing model integration, evaluating performance through competition ranking, contributing significantly to DR detection methodologies, and identifying research gaps. The study encompasses revolutionizing DR detection by integrating cutting-edge technologies, focusing on transfer learning and various model integrations within web applications. The methodology covers data pre-processing, augmentation, segmentation using U-Net neural network architecture, and a detailed training process. The U-Net model demonstrates efficient segmentation of retinal structures with high accuracy and an impressive frames-per-second rate. The results highlight the model’s effectiveness in segmenting blood vessels, hard exudates, soft exudates, hemorrhages, microaneurysms, and the optical disc, with high Jaccard, F1, recall, precision, and accuracy scores. These findings underscore the model’s potential to enhance diagnostic capabilities in retinal pathology assessment, promising improved patient outcomes through timely diagnosis and intervention in combating DR.

Self-supervised learning of wrist-worn daily living accelerometer data improves the automated detection of gait in older adults

Progressive gait impairment is common among aging adults. Remote phenotyping of gait during daily living has the potential to quantify gait alterations and evaluate the effects of interventions that may prevent disability in the aging population. Here, we developed ElderNet, a self-supervised learning model for gait detection from wrist-worn accelerometer data. Validation involved two diverse cohorts, including over 1000 participants without gait labels, as well as 83 participants with labeled data: older adults with Parkinson's disease, proximal femoral fracture, chronic obstructive pulmonary disease, congestive heart failure, and healthy adults. ElderNet presented high accuracy (96.43 ± 2.27), specificity (98.87 ± 2.15), recall (82.32 ± 11.37), precision (86.69 ± 17.61), and F1 score (82.92 ± 13.39). The suggested method yielded superior performance compared to two state-of-the-art gait detection algorithms, with improved accuracy and F1 score (p < 0.05). In an initial evaluation of construct validity, ElderNet identified differences in estimated daily walking durations across cohorts with different clinical characteristics, such as mobility disability (p < 0.001) and parkinsonism (p < 0.001). The proposed self-supervised method has the potential to serve as a valuable tool for remote phenotyping of gait function during daily living in aging adults, even among those with gait impairments.

Identifying Credit Card Fraud with Machine Learning: Evaluation of Algorithms and Oversampling Techniques

The physical loss of a credit card or the theft of sensitive credit card data is referred to as credit card fraud. For detection, a variety of machine learning methods can be applied. This study presents several algorithms for identifying transactions as authentic or fraudulent. The study used the credit card fraud detection dataset. The SMOTE approach was utilized for oversampling due to the extremely unbalanced nature of the dataset. Additionally, a feature selection process was carried out, and the dataset was divided into training and test sets. The experiment employed eight machine learning models, including Random Forest, AdaBoost, Support Vector Classifier, Extreme Gradient Boost, and Logistic Regression algorithms. The findings indicate that few algorithms have a high degree of accuracy when detecting credit card fraud. The Random Forest model can be applied to find further anomalies. This approach has proven to be effective in accurately detecting fraudulent activity and reducing the number of false positives and negatives. The results of the experiment also highlight the importance of feature selection in improving the performance of the models by the highest accuracy score with 96%, precision of 100, Recall of 91, and F1Scroe of 95. By using a combination of different machine learning algorithms, financial institutions can enhance their fraud detection systems and better protect their customers from potential financial losses. The findings from this study demonstrate the potential for significant advancements in the field of fraud detection using machine learning techniques.

Self-Supervised Representation Learning and Temporal-Spectral Feature Fusion for Bed Occupancy Detection

In automated sleep monitoring systems, bed occupancy detection is the foundation or the first step before other downstream tasks, such as inferring sleep activities and vital signs. The existing methods do not generalize well to real-world environments due to single environment settings and rely on threshold-based approaches. Manually selecting thresholds requires observing a large amount of data and may not yield optimal results. In contrast, acquiring extensive labeled sensory data poses significant challenges regarding cost and time. Hence, developing models capable of generalizing across diverse environments with limited data is imperative. This paper introduces SeismoDot, which consists of a self-supervised learning module and a spectral-temporal feature fusion module for bed occupancy detection. Unlike conventional methods that require separate pre-training and fine-tuning, our self-supervised learning module is co-optimized with the primary target task, which directs learned representations toward a task-relevant embedding space while expanding the feature space. The proposed feature fusion module enables the simultaneous exploitation of temporal and spectral features, enhancing the diversity of information from both domains. By combining these techniques, SeismoDot expands the diversity of embedding space for both the temporal and spectral domains to enhance its generalizability across different environments. SeismoDot not only achieves high accuracy (98.49%) and F1 scores (98.08%) across 13 diverse environments, but it also maintains high performance (97.01% accuracy and 96.54% F1 score) even when trained with just 20% (4 days) of the total data. This demonstrates its exceptional ability to generalize across various environmental settings, even with limited data availability.

Performance Analysis of EMG Signal Classification Methods for Hand Gesture Recognition in Stroke Rehabilitation

This study evaluates the performance of different classification methods in classifying healthy individuals and stroke patients. The hand gesture variations of the subjects were also analyzed based on electromyography (EMG) signals. Several classification methods were tested in this analysis to find out which method had the most suitable performance. The results showed that Decision Tree and Naive Bayes classifiers achieved the highest performance in classifying EMG signals from healthy individuals and stroke patients, with both methods showing high accuracy, precision, recall, and F1 score. Specifically, Decision Tree excelled in overall accuracy and recall, while Naive Bayes showed superior precision. For hand gesture recognition, SVM, KNN, and Random Forest classifiers showed similarly high performance, achieving accuracy, precision, recall, and F1 score above 82%. Naive Bayes also performed well, especially in precision, while Decision Tree performed poorly compared to other methods. This insight can form the basis for the development of more effective and personalized rehabilitation systems for stroke patients, by utilizing reliable and accurate EMG signal classification

Risk early warning model of tunnel engineering based on computer vision and CNN

Abstract Safety management in tunnel construction is important to ensure the safety of construction personnel and equipment. This paper proposes a tunnel engineering risk early warning model based on computer vision and convolutional neural network (CNN), aiming at realizing real-time monitoring and risk early warning of the construction site using modern image processing technology and deep learning algorithm. Many image data of tunnel engineering construction sites were collected, and the data were preprocessed and marked. Then, the paper designs a depth CNN model for feature extraction and risk identification of construction site images. Through the performance evaluation of the model on the training set and test set, the effectiveness and feasibility of the model in risk early warning are verified. The experimental results show that the proposed model achieves high accuracy, recall, and F1 score on the test set, which is superior to the traditional method based on image features and artificial rules. At the same time, we analyzed the prediction results of the model on the actual construction site images. We found that the model can effectively identify different risk factors and has strong generalization ability and practicability. To sum up, this study provides a new technical means and solution for the safety management of tunnel engineering, which has important theoretical and practical significance.