Engineering reports : open access最新文献

This research has introduced a hybrid model that integrates the long short-term memory (LSTM) and extreme gradient boosting (XGBoost) models to assess students' mental health states, particularly to identify students' levels of stress, mood, and fatigue. The physiological measures measured were heart rate (HR), heart rate variability (HRV), electrodermal activity (EDA), and skin temperature. All measures were recorded using wearable sensors and underwent processing, such as normalization, noise filtering, and feature extraction, to ensure the signal quality was fit for analysis and interpretability. While the LSTM network can accurately represent the temporal dynamics present in the physiological sequences, the XGBoost model is critical in obtaining high accuracy through the classification of features' non-linear interactions and decision boundary optimization. The experimental validation through the technique of fivefold cross-validation shows that the hybrid model performs with high accuracy of 0.98 on average, F1-score of 0.98, and consistently low false-positive and false-negative rates when compared to SVM, Random Forest, and single deep learning model methods that serve as baseline methods. The results assure the framework's reliability, consistency, and clarity in reasoning over different data conditions. This novel method provides a strong platform for the real-time, data-driven monitoring and early detection of psychological distress, thus allowing educators, mental-health professionals, and caregivers to make timely interventions and improve the overall well-being of students.

Anterior Cervical Corpectomy and Fusion (ACCF), which is one of the common surgeries used to treat cervical spine diseases, has been widely applied in clinical practice. The commonly used internal fixation forms in ACCF surgery include the traditional Anterior Vertebral Body Screw-Plate (AVBSP) structure and the Anterior Cervical Pedicle Screw-Plate (APSP) structure, both of which are combined with titanium mesh to achieve support and bone fusion. The purpose was to investigate the effects of different surgical plans on cervical spine biomechanics and the interplay between internal fixation instruments after surgery. In this study, a finite element model of the human lower cervical spine (C3-C7) after ACCF surgery was established. The surgical plan consisted of two internal fixation forms (AVBSP and APSP) and two titanium mesh forms (linear and curved), combined in different ways. The mechanical sensitivity of adjacent intervertebral disc nuclei to different surgical plans was significantly different. The stress concentration areas on the vertebral body entry surface varied with different entry methods, and the stress values were greatly affected by cervical movements. The related instrument studies showed that the choice of anterior fixation method would affect the stress level and distribution of the titanium mesh. Theoretically, the combination of curved titanium mesh and AVBSP is beneficial to reducing the overall stress level of the internal fixation instruments and titanium mesh. The research provides a theoretical basis for the selection of clinical surgical plans. It is advantageous in enhancing postoperative stability of cervical vertebrae while reducing the risk of recurrence or other complications. Clinically, when selecting the excision fusion surgical plan based on the condition of the patient's cervical lesion, consideration should be given to the matching characteristics between internal fixation methods and titanium mesh forms, as well as their effects on the biomechanics of adjacent segments.

The rapid growth of e-commerce and the emergence of BNPL (Buy Now, Pay Later) financial products have significantly increased loan applications. However, the accumulation of losses from customer defaults poses a serious bankruptcy risk for BNPL providers. Unlike most studies that focus on credit scoring for traditional microloans, this research specifically uses BNPL loan data. A major concern in this domain is the imbalanced nature of the data, which can adversely affect model performance. To this end, we compared ensemble learning models in combination with data balancing methods and proposed a novel combination of logistic regression, SMOTE-NC, and LightGBM, which has not been extensively explored in previous studies. Additionally, we introduced two new variables— ‘active internet banking’ and ‘active mobile banking’—to investigate whether the use of digital banking platforms can indicate creditworthiness. Regression analysis confirmed the significance of the new variables, alongside key predictors such as ‘Education’, ‘Collateral type’, ‘Long-term accounts count’, ‘Received loans count’, ‘Active loans count’, and ‘Loan amount’. The proposed method achieved an F1-score of 84.66% for the default class, a 23% improvement over models without balancing techniques. Implementing this model could reduce realized BNPL losses by 26.84%, underscoring its potential to mitigate risks in this sector.

The exponential growth of the Internet has led to a dramatic rise in the use of web applications, making them integral to businesses, industries, education, financial institutions, and daily life. However, this widespread rise has introduced significant security issues, exposing web applications to various vulnerabilities capable of compromising the confidentiality, integrity, and availability of sensitive data. Therefore, mitigating these vulnerabilities has become vital to ensuring robust information security. Among the myriad of vulnerabilities, Structured Query Language injection (SQLi) is one of the foremost prevalent types of vulnerabilities affecting web-based apps, essential to detect Structured Query Language (SQL) injection vulnerabilities. In practice, penetration testers utilize tools for automated vulnerability assessment with varying strengths and limitations to evaluate the security of web applications. However, these security scanners have certain flaws, such as failing to scan entire web apps and producing inaccurate test results. Furthermore, significant research has been conducted to quantitatively list the outcomes of web application security scanners to examine their limitations and efficacy. Yet, a standardized methodology or criteria for assessing their performance remains elusive. To overcome these challenges, this paper proposes the SQLi-ScanEval Framework, a standardized SQLi detection system that integrates vulnerability and penetration testing scanners into a standardized framework. The proposed framework provides a standardized evaluation environment, thereby overcoming the drawbacks of individual scanners, including insufficient coverage and erroneous data. The proposed SQLi-ScanEval Framework tested seven prominent SQLi vulnerability scanners including OWASP ZAP, Wapiti, Vega, Acunetix, Invicti, Burp Suite and Arachni, on two prominent vulnerable testing applications i.e., Test PHP and Bricks from OWASP Broken Web Applications (BWA). The framework successfully evaluated the performance of each scanner on the basis of recall, accuracy, and precision. The results showed that Acunetix exhibits the highest accuracy i.e., 90.48% on Bricks and 86.96% on Test PHP, with the lowest false positive rates and a recall of 88.89%. The results also reveal notable variations in scanner performance, with scan times varying from 00:02:13 (OWASP ZAP) to 00:43:33 (Invicti) with the Bricks application. The SQLi-ScanEval results also provide valuable insights with the strengths and shortcomings for each scanner, giving penetration testers a practical roadmap for selecting the best tools. As cyber-attacks keep evolving, this study not only enhances decision-making but also extends SQLi techniques for detection, unlocking the way to more secure web applications.

With the rapid development of enterprise digitalization and the global economic environment, enterprise compliance risks and economic management issues have become increasingly complex. Accurate risk identification and efficient economic decision-making are crucial for the sustainable development of enterprises. To address this challenge, this study proposes an enterprise compliance risk identification and economic management optimization model that combines Dynamic Bayesian Network (DBN) with Artificial Intelligence (AI)-driven Reinforcement Learning (RL). The study utilizes a financial risk insight dataset and begins with data preprocessing, including missing value imputation, feature standardization, and time series alignment. Subsequently, a DBN model is constructed, with risk variable conditional probability distributions learned via the Markov Chain Monte Carlo (MCMC) method. An AI-driven RL algorithm is integrated to optimize network parameters, enabling the capture of dynamic evolution and dependencies among risk factors. Empirical results indicate that: (1) The proposed model achieves an AUC-ROC of 0.981 in risk identification tasks, representing an 8.9% improvement over the best baseline model, with an F1-score of 0.937. (2) AI-assisted auditing significantly enhances operational efficiency: the average working hours per case are reduced by 15.3%, the detection rate of high-risk cases is increased by 15.8%, and customer satisfaction is raised to 4.76 points (out of 5). (3) In terms of economic management optimization, the RL strategy increases the risk-adjusted return on capital (RAROC) to 0.236, maintains it at 0.187 under crisis scenarios, and achieves a compliance cost savings ratio of 0.237. These results verify the effectiveness of the DBN-DRL collaborative framework in balancing risk control and economic benefits. This study provides a data-driven intelligent tool enabling dynamic risk identification and economic decision optimization, offering theoretical foundations and practical references for enterprises to construct efficient and sustainable compliance risk management systems.

This study experimentally investigates the performance, combustion, and emission characteristics of a Reactivity-Controlled Compression Ignition (RCCI) engine fueled with diesel and the high-viscosity oxygenated alcohol 2-Methyl-1-butanol. Experiments were conducted on a single-cylinder common-rail direct injection engine operating at brake mean effective pressures of 3 and 5 bar and fuel injection pressures of 400, 600, and 800 bar. Diesel was directly injected as the high-reactivity fuel, while 2-Methyl-1-butanol was port-injected to establish RCCI combustion. Fuel blends containing 10%, 20%, and 30% alcohol were evaluated and compared with neat diesel operation. Results indicate that increasing injection pressure improves fuel atomization, advances combustion phasing, and enhances heat release characteristics. At full load and 800 bar injection pressure, the D70MB30 blend achieved the highest brake thermal efficiency of 37.4%, compared to 26.0% for neat diesel. Significant emission reductions were also observed, with NO_x decreasing from 4.5 ppm (diesel) to 3.1 ppm and smoke opacity showing a consistent declining trend due to improved charge homogeneity and oxygen availability. However, higher alcohol content resulted in increased CO and HC emissions at part-load conditions because of low-temperature combustion and evaporative cooling effects. These penalties were substantially mitigated at higher injection pressures. Overall, the D70MB30 blend at 800 bar provided the best trade-off between performance and emissions, demonstrating the potential of 2-Methyl-1-butanol as a sustainable alternative fuel for advanced RCCI engine operation.

With the rapid development of technologies such as cloud computing and the Internet of Things, organizations face the thorny reality that network attacks are becoming increasingly diverse, covert, and intelligent. Traditional signature-based intrusion detection systems (IDSs) struggle to address zero-day attacks and advanced persistent threats (APTs), often resulting in low detection rates and high false-positive rates. To address this, this paper proposes an adaptive network intrusion detection system that integrates random forest (RF) and real-time reputation evaluation. The system first preprocesses and normalizes the original network traffic and behavior logs, and then uses a random forest to perform preliminary multi-category classification. It then introduces a historical behavior risk metric, weighting the error rate of the current detection with the device's historical risk profile using exponential decay. A comprehensive reputation score is generated using a continuously differentiable “four-stage” smoothing function: sigmoid in the low-confidence zone, cosine in the medium-low zone, inverse sigmoid in the medium-high zone, and exponential decay in the extremely high zone. Finally, RRF-IPS's reputation scoring system executes automated policies such as bandwidth throttling, warning notifications, and session isolation or blocking, forming a closed “detect-assess-respond-archive” loop. Experimental results demonstrate that, on CICIDS2017, our system improves accuracy by 0.6% and F1 score by 5.9% compared to state-of-the-art methods.

Malware remains a significant concern for modern digital systems, increasing the need for reliable and scalable detection methods. This work proposes an ensemble method that combines a random forest (RF) with a vision transformer (ViT). The approach exploits complementary feature spaces, including bag-of-words (BoW) and image representations, to enhance multi-class malware classification. We also evaluate traditional machine learning models (Naïve Bayes, Support Vector Machine, and RF) and deep learning (DL) models (ResNet50 and ViT) using the Microsoft Malware and Dike datasets. The proposed ensemble model achieves 99.32% accuracy and 98.11% F1 score on the Malware dataset, outperforming individual models and recent state-of-the-art studies. While ViT captures spatial and sequence dependencies via attention mechanisms, RF captures textual and byte-level frequency patterns. Their combination, through a product rule, enhances robustness and reliability in multi-class cybersecurity tasks.

This paper addresses the issue of alarm-based hybrid fault-tolerant control for Markovian jump systems with uncertainty, in which there may exist actuator and structural failures. Through the integration of passive, observer-based active, and adaptive fault-tolerant control techniques, the effect of actuator partial failure, actuator bias failure, and structural failure can be separately attenuated. Moreover, an alarm-based multi-threshold system is designed to invoke the suitable control strategy. Finally, to prove the practical applicability of the given method, a single-link robot arm system is presented, with results confirming the theoretical findings.

Dynamic pricing has become a core approach for optimizing data value in both personal and enterprise contexts. However, accurately determining data prices across multiple dimensions remains a key research challenge. This study proposes a novel dynamic pricing mechanism based on a multi-dimensional dynamic model to enhance pricing accuracy. The model analyzes pricing factors across various dimensions and dynamically adjusts prices according to real-time data attributes and usage scenarios. Experimental results show that the proposed model achieves a pricing deviation of only 129 yuan from actual values (approximately 2.5%), significantly outperforming the traditional equal pricing model. The proposed model reduces the maximum error by 2.5 and the root mean square error by 2.2 in comparison. In addition, it demonstrates improved computational efficiency, with a runtime reduction of 296.10 milliseconds, and achieves an absolute increase of 14.29 percentage points in the F1 score and 14.90 percentage points in recall rate. These results indicate that the multi-dimensional dynamic pricing model offers superior performance in both pricing precision and operational efficiency. The findings provide valuable insights for developing more accurate and adaptable data pricing strategies in real-world applications.