Journal of Intelligent Systems

Dobrády, Zoltán; Nagy, Szilvia; Hidvégi, Timót

doi: 10.1515/jisys-2025-0015pmid: N/A

AbstractThe Modbus TCP/IP protocol, widely adopted in industrial communications, lacks essential security features, making it vulnerable to cyberattacks such as TCP Connection Exhaustion. This paper proposes a machine learning-based detection framework using the Random Forest (RF) algorithm to identify malicious traffic in Operational Technology (OT) networks. A simulated testbed was created using virtual machines to emulate Modbus server-client communication under normal and attack conditions. Our model achieved F1-score of 99.83 %, precision of 99.9 %, and recall of 99.7 %, clearly demonstrating its accuracy and robustness. These results validate the proposed approach as a lightweight, real-time, and effective intrusion detection system suitable for resource-constrained industrial environments.

Fan, Yilin; Xu, Jiahao; Chen, Bingchi

doi: 10.1515/jisys-2025-0059pmid: N/A

AbstractThis paper proposes a computerized text analysis framework to examine the evolution of China’s industrial internet policies concerning SMEs. Analyzing a corpus of 310 policy documents concerning SMEs in industrial internet domain, we first investigated their external structure. Thematic sequence evolution analysis subsequently identified five key topics. Building on a pivotal policy issued in 2017, the evolution of China’s industrial internet policies concerning SMEs was categorized into two stages. The paper further explored the characteristics of the intergovernmental cooperation network, and examined the evolutionary paths of policy themes using a Sankey diagram. The findings indicate that the five topics in stage 2 are essentially the continuation and further development of the policy themes in stage 1, demonstrating strong policy coherence in industrial internet domain. Theoretically, this study addresses the IAD’s challenges in exploring the evolution of action situations and the diverse processes of structural changes over time, while overcoming the ACF’s limitations in interpreting the interactions of common belief systems and interests.

Douider, Meriem; Amrani, Ibrahim; Abik, Mounia

doi: 10.1515/jisys-2024-0531pmid: N/A

AbstractThe widespread use of machine learning algorithms in dataset modeling requires a thorough understanding of the various tools likely to improve the modeling quality. Any machine learning algorithm has two types of parameters: model parameters and hyperparameters. Parameters are adjusted in the learning process, while hyperparameters are defined a priori. Hyperparameter tuning is an essential element in the modeling process due to its effect on the quality of results. This study focuses on enhancing the modeling accuracy of mosquito species distribution in Morocco by optimizing the hyperparameters of the employed algorithms. Three tuning methods were selected for this purpose: Genetic Algorithms, Bayesian Optimization, and Particle Swarm Optimization. The experimental results confirmed the effect of hyperparameter tuning on the modeling quality, with accuracy improvements ranging from 0.02 to 0.067. In addition, Genetic Algorithms and Bayesian Optimization proved more effective than Particle Swarm Optimization. The hyperparameter tuning process has optimized the modeling quality, which can only enhance the explanation of mosquito distribution.

Abed, Sa’ed; Rashid, Adnan; Hasan, Osman

doi: 10.1515/jisys-2025-0101pmid: N/A

AbstractPhysical Human-robot Interaction (pHRI) involves physical collaboration between humans and robots to perform tasks safely and efficiently in shared environments. Traditionally, pHRI behavior has been analyzed using analytical or simulation-based methods, which may suffer from human error and incomplete coverage of system behavior. Formal methods, such as Interactive Theorem Proving (ITP), offer a mathematically rigorous alternative by constructing logical models of system dynamics and verifying their properties through deduction and proof. This paper proposes an ITP-based approach for the formal analysis of pHRI dynamics using the HOL Light theorem prover. We formalize the one-dimensional admittance control equations, derive their Laplace-domain representations, and conduct a formal stability analysis of the corresponding open and closed-loop controllers. To complement the formal verification, the verified open-loop admittance controller is implemented and simulated in MATLAB, where the step-response and root-locus analyses confirm the formally proven stability guarantees. The results demonstrate that the proposed framework provides both mathematical rigor and practical consistency for ensuring safe and reliable human–robot interaction.

Dayssa, Abay Hailemariam; Gondere, Mesay Samuel

doi: 10.1515/jisys-2025-0247pmid: N/A

AbstractAccurate classification of child delivery mode is crucial for improving maternal and neonatal health. In developing countries like Ethiopia, clinical assessments alone often result in misguided medical interventions. Even though machine learning in healthcare has brought promises, the various algorithms along with different real-world datasets perform differently. Hence, the objective of this study was to develop a machine learning model for predicting child delivery mode based on real data. The study followed experimental and exploratory research design utilizing 1,072 antenatal records from Arba Minch General Hospital and Birbir Health Center, Ethiopia. 16 attributes were considered including the outcome, mode of delivery. Predictors included sociodemographic and clinical variables such as age, weight, blood pressure, previous CS, and fetal presentation. Five machine learning algorithms including Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting, and CatBoost were trained and evaluated using hold-out validation. Additionally, a recent deep learning model, TabPFN, and Long Short-Term Memory were examined to expand the exploration. The results showed that RF (93.1 %) achieved the highest accuracy. TabPFN scored the second best accuracy score (92.5 %), demonstrating its potential on smaller tabular data. LSTM performed better than SVM and LR highlighting the consideration of the inherent temporal characteristics of the data.

Samaras, Agorastos-Dimitrios; Moustakidis, Serafeim; Apostolopoulos, Ioannis D.; Papageorgiou, Elpiniki; Papathanasiou, Nikolaos D.; Apostolopoulos, Dimitris J.; Papandrianos, Nikolaos

doi: 10.1515/jisys-2025-0064pmid: N/A

AbstractThis study addresses the diagnostic challenges associated with Non-Small Cell Lung Cancer (NSCLC), the most prevalent form of lung cancer often diagnosed at advanced stages. It aims to develop a computer-aided classification model exclusively utilizing medical images from Positron Emission Tomography (PET) and Computed Tomography (CT) scans. The model identifies benign/malignant Solitary Pulmonary Nodules (SPN) related to NSCLC. A dataset comprising of 456 patients, in total, was curated, featuring 48.68 % benign cases. To achieve its objective, four well-established Deep Learning (DL) algorithms were employed. The dataset was split into three different groups of images, each used for a particular task; training, testing and validation of the model. Notably, the study extends beyond predictive accuracy by delving into the prediction process of the best-performing model, thereby enhancing the explainability of the typically opaque Artificial Intelligence (AI) models. This explainability aspect aims to foster trust and confidence in the model’s outcomes, allowing users to comprehend the decision-making process. The results indicate that the YOLOv8 algorithm emerged as the most accurate classification model, achieving a maximum accuracy of 91.3 % and a maximum True Positive Rate (TPR) of 93.62 %. This study’s importance lies in underscoring the potential of DL approaches in improving NSCLC diagnosis while providing a transparent and understandable classification mechanism. It offers a novel way of explaining classification results from YOLOv8 model and it demonstrates both the effectiveness of DL-assisted predictions in characterizing SPNs and the added value of interpretability, thereby offering a holistic perspective on the model’s capabilities.

Lin, Kaisi; Zhang, Lu

doi: 10.1515/jisys-2025-0066pmid: N/A

AbstractTraffic sign recognition plays a critical role in the development of autonomous driving systems and intelligent transportation networks. However, detecting small traffic signs in real-world scenarios, particularly those captured from vehicle-mounted cameras, remains a significant challenge due to their diminutive size, low resolution, and environmental noise. To address these challenges, we propose an innovative multi-strategy enhancement framework for YOLOv7, designed specifically to improve small target detection. The framework integrates several novel techniques: the SE attention mechanism is incorporated into the ELAN module of the backbone network to enhance feature discriminability; DySample replaces traditional upsampling methods in the head network to refine feature reconstruction; NWD loss is introduced as a superior alternative to CIoU loss, improving the localization accuracy of small objects; and PConv convolution is applied to reduce model parameters without sacrificing performance. Experimental results on the CCTSDB-2021 dataset demonstrate the effectiveness of these enhancements, with [email protected] and [email protected]:0.95 improving by 6.6  and 15.2 %, respectively, compared to the original YOLOv7 model. The proposed algorithm outperforms YOLOv7 by 10.9 % in [email protected] and by 11.89 % in [email protected]:0.95 on the TT100K dataset. Moreover, the optimized model achieves real-time inference at 83 FPS on the CCTSDB-2021 dataset, while reducing the number of parameters by 1.5 million, making it highly efficient for practical deployment in autonomous vehicles. These improvements not only enhance detection accuracy but also meet the real-time processing requirements of intelligent transportation systems.

Zhao, Qin; Liu, Jiuhong; Ma, Xian; Liu, Tongbo; Du, Yingzhen

doi: 10.1515/jisys-2025-0076pmid: N/A

AbstractDeep learning techniques have been widely used in various fields. However, they face significant security challenges due to the existence of adversarial examples. Traditional black-box adversarial attack methods mainly rely on swarm intelligence optimization algorithms to identify optimal perturbation pixels, which requires intensive computational resources. In some typical applications such as medical image recognition, public datasets are often used to train deep learning models. It is worth noting that such dataset inherently contains some basis features for deep learning models to learn discriminative representations. And these features can serve as critical cues for constructing adversarial samples. Inspired by this observation, a novel adversarial attack method was proposed. First, some sensitive locations are identified within the dataset without querying the target model. Moreover, the adversarial attack samples are constructed based on these locations. Different from white-box and black-box attack, dataset characteristics are utilized to construct adversarial attack samples. The proposed method investigates naturally occurring vulnerabilities in the data, offering new insights for enhancing data augmentation techniques and attack strategies, while also providing a promising direction for improving model robustness. Experimental results demonstrate that this method can achieve attack effectiveness comparable to the Particle Swarm Optimization (PSO) algorithm.

Alex, Benoi; Yedurkar, Dhanalekshmi Prasad; Al-Turjman, Fadi; Stephan, Thompson; Hamid, Yasir; Shah, Mohd Asif

doi: 10.1515/jisys-2024-0108pmid: N/A

AbstractThis study aims to develop an automatic system for detecting toxic comments in online environments, particularly on social networking platforms. The focus is on efficiently identifying and categorizing toxicity in user-generated comments to address the growing issue of harmful online content. A novel hybrid model, Text-BGRU-CNN, combining Bidirectional Gated Recurrent Unit (BGRU) and Text Convolutional Neural Network (Text-CNN), is introduced for multilabel toxicity detection. This model uses pre-trained word embeddings for word vector generation and a range of filters to extract local features and long-term dependencies in text. It incorporates a fully connected layer, a normalization layer, and an output layer for multilabel category prediction. The proposed hybrid model demonstrates superior classification accuracy in experimental trials. It was tested on a dataset divided into training and testing sets, enhanced by significant pre-processing. Structural modifications, including increasing dense units and filters, were evaluated. The final model, combining GRUs with a single CNN layer, achieved an accuracy of 0.9944 in classifying toxic comments. The study evidences the efficacy of a hybrid GRU and single-layer CNN model in online toxic comment classification. Results suggest that simpler model architectures, supplemented by extensive pre-processing, yield high accuracy and efficient training. The findings underscore the importance of further research to understand biases in trained classifiers and suggest exploring alternative methods for representing sequences and independent training of sparsely labeled classes in future work.

Uma Mahesh, Rajanahalli Nataraj; Chandrashekar, Puttaswamy

doi: 10.1515/jisys-2024-0393pmid: N/A

AbstractThe binary classification of three-dimensional (3-D) objects for phase-only digital holographic information is performed using the various deep learning network models such as ResNet50, ResNet101, ResNet152, ResNet18, ResNet34, EfficientNetB0, DenseNet121, DenseNet169, Neural Architectural Search (NAS) Network, and InceptionV3. The four 3-D objects considered to perform binary classification are ‘triangle-square’, ‘circle-square’, ‘square-triangle’, and ‘triangle-circle’. The 3-D object ‘triangle-square’ has been considered for Class 1 and the remaining three 3-D objects have been considered for Class 2. The digital holograms of 3-D objects have been formed using the phase-shifting digital holographic (PSDH) technique and numerically reconstructed to obtain phase images. The phase image dataset consisting of 2,880 images was trained using all the various deep learning network models to obtain the results. The results such as loss/accuracy, loss/positive predictive value (PPV), and loss/sensitivity curves on the training/validation sets, error matrix, and performance metrics namely accuracy, PPV, sensitivity, F1-score, Matthews correlation coefficient (MCC), cohen_kappa_score (CKS), balanced_accuracy_score (BAS), jaccard_score (JS), log_loss (LL), hinge_loss (HL), and brier_score_loss (BSL) are shown for the binary classification task. Finally, the results such as receiver operating characteristic (ROC), and PPV-sensitivity curve are also shown to justify the performance of the work. The results obtained from the deep residual network models i.e. ResNet50, ResNet101, ResNet152, ResNet18, and ResNet34 were compared with other deep learning network models such as EfficientNetB0, DenseNet121, DenseNet169, NAS Network, and InceptionV3.