The June 2025 (Vol. 33, No. 2) issue of CIT. Journal of Computing and Information Technology brings four papers from the areas of network security, 3D modeling, brain-computer interaction, and recommendation systems.
Accurate prediction of product lifecycle stages is crucial for enhancing inventory turnover and strategic planning in the tobacco industry. This paper proposes an intelligent prediction model that integrates Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU), further optimized by an Improved Grey Wolf Optimizer (IGWO). The model fuses multi-source enterprise data—including sales trends, IoT logistics information, environmental conditions, and inventory records—to dynamically forecast lifecycle stages and remaining durations. The dataset comprises 180,000 labeled samples collected from real-world tobacco enterprise operations, encompassing multi-source variables such as sales volume, inventory changes, logistics routes, and environmental feedback. Experimental evaluations based on this dataset demonstrate that the proposed IGWO-CNN-GRU model achieves a Mean Squared Error (MSE) of 2.13, a Mean Absolute Error (MAE) of 1.17, and an R² of 0.932, significantly outperforming baseline models. In practical deployment simulations, the prediction deviation is limited to ±5 days, improving allocation efficiency and reducing inventory risks. The approach provides a robust and adaptable solution for full-lifecycle management in tobacco supply chains, offering practical value for intelligent production and market deployment strategies.
The Internet of Medical Things (IoMT) consists of interconnected devices and applications that enable real-time collection, transmission, and analysis of medical data for healthcare applications. This study utilizes medical data from the publicly available BCICIV2a dataset rather than data collected directly from individuals or medical institutions. With advancements in neuroinformatics and intelligent computing, the classification of electroencephalography (EEG) signals has become increasingly important, particularly for detecting and predicting epilepsy. However, existing EEG classification methods often suffer from low accuracy, high computational complexity, and slow processing. To address these challenges, this study proposes an EEG classification approach utilizing a Backpropagation Neural Network (BPNN) enhanced with Bayesian optimization. This method enhances the identification and prediction of epileptic seizures by utilizing IoT-enabled EEG data. Performance evaluation on the BCICIV2a dataset demonstrates that the proposed model achieves an accuracy of 93.21%, outperforming conventional techniques. The results indicate that this approach enhances efficiency and accuracy in EEG signal processing, contributing to real-time medical diagnostics. The integration of IoMT with advanced neural networks represents a significant advancement in medical informatics and telemedicine, providing promising directions for future research and clinical applications.
Facial 3D modeling technology is widely used and has become an important research direction in the fields of artificial intelligence and computer vision. However, the modeling accuracy and robustness of existing technologies in dealing with weak texture areas and complex lighting conditions are insufficient, which limits their practical application in production. Therefore, a facial 3D modeling method based on multi-scale feature fusion and lighting robustness optimization was proposed, and a multiscale dense feature network and lighting robustness feature fusion network were constructed. The experimental outcomes indicated that the method exhibited excellent performance on the dataset. Among them, structural similarity reached 0.954, and the average absolute error was the lowest at 0.63 mm. Under dynamic lighting conditions, the feature consistency reached 0.941, and the point cloud error was reduced to 0.85 mm. In addition, tests in security and virtual reality scenarios showed that after using this method, the accuracy increased to 92.8%, the peak signal-tonoise ratio reached 33.0 dB, and the model running efficiency improved to 36 frames per second, verifying the practicality and reliability of the method. The research provides new ideas for developing stable, efficient, and practical facial 3D modeling methods, which is expected to promote the widespread application of related technologies in complex environments.
This research proposes an efficient parallel graph partitioning algorithm for the big data environment, aiming to solve the bottlenecks of traditional clustering techniques in terms of processing speed and scalability. The algorithm adopts a multi-level graph partitioning framework, decomposing the network information processing task into multiple levels, gradually simplifying the graph structure and backtracking refinement, thereby significantly reducing the computational complexity while ensuring the partitioning quality. The algorithm focuses on balancing the node cohesion within partitions and the edge cutting cost of inter-partition communication. By constructing a global objective function, it minimizes the number of edges across partitions and the workload differences among various sub-graphs, thereby achieving a more balanced partitioning result. The research results show that this algorithm achieves a resource utilization rate of 0.95. In the Hadoop cluster environment, 95% of the computing resources are effectively used for actual task processing, which is significantly higher than that of the competing algorithms. The energy efficiency ratio reaches 0.98, indicating that the number of tasks completed per unit of energy consumption is close to the optimal level, which is superior to the 0.78 to 0.67 range of existing methods, reflecting the advantages of this algorithm in green computing. The load imbalance rate is only 0.00395, and the point weight imbalance rate is 0.00141, which are much lower values than those of the comparison algorithm. This indicates that the algorithm achieves a high degree of balance in task allocation and node weight distribution, effectively avoiding resource waste and performance bottlenecks.
Recommendation systems face the challenge of balancing dynamic short-term preferences with stable long-term interests to deliver personalized and timely recommendations. Traditional methods often treat these aspects separately, leading to suboptimal integration and limited adaptability to evolving user behavior. This paper introduces Temporal-Aware Neural Networks (TANR), a novel framework that leverages a time-aware Transformer architecture to dynamically balance short-term and long-term user preferences. The proposed model incorporates a time decay mechanism within the attention layer to adjust the influence of recent and historical interactions, ensuring a balanced representation of user behavior. Additionally, TANR employs a hybrid training framework combining offline pre-training with online incremental updates, enabling real-time adaptation to user behavior shifts. Extensive experiments on the MovieLens-1M and MIND datasets demonstrate that TANR outperforms state-of-the-art models in both short-term engagement metrics (e.g., Hit Rate, NDCG) and long-term user retention. The results highlight the effectiveness of TANR in capturing temporal dynamics and improving recommendation accuracy, offering a robust solution for modern recommendation systems.
This paper presents a novel real-time fire detection framework tailored for IoT devices by integrating the fine-tuned YOLOv10 model with the Accelerator module. Trained on the FireSmokeDataset (Roboflow) and an additional dataset we collected via Roboflow, the system covers fire, smoke, and distracting objects. Optimized for resource-constrained edge devices, the framework demonstrates exceptional performance, achieving high mean average precision (mAP) for fire and smoke detection, with metrics exceeding 84% and a maximum mAP50 of over 91%. We target deployments in residential homes, industrial facilities, and forest monitoring stations. A key contribution of the proposed framework is the construction of a diverse dataset encompassing fire, smoke, and distracting objects - an element often overlooked in existing fire detection datasets. Additionally, fine-tuning the YOLOv10 model components in conjunction with hardware acceleration ensures both prediction accuracy and improved inference response performance. Comprehensive evaluations confirm the system's robustness, scalability, and practicality under various operating conditions. Through experimental analysis, the YOLOv10-S (small) model stands out for its balance between efficiency and resource usage, making it a suitable choice for low-cost real-time applications with resource constraints. By utilizing the Coral Accelerator, the proposed framework reduces inference time by 58% compared to CPU-based implementations, achieving a latency of just 1.7 seconds per frame. The system's lightweight design ensures reliable deployment in remote areas with limited computational resources and unstable network connectivity, maintaining high accuracy while minimizing false alarms.
The September 2025 (Vol. 33, No. 3) issue of CIT. Journal of Computing and Information Technology brings four papers from the areas of graph data processing, computer vision, and business intelligence.
To address the issues of data silos, low detection accuracy, and insufficient generalization ability in traditional methods for power grid intrusion diagnosis, this study proposes the use of federated learning to construct a power grid intrusion diagnosis model and incorporates convolutional neural networks and long short-term memory network optimization models on this basis. The experiment outcomes indicate that in performance analysis, the accuracy of the raised model is 97.3%, the precision is 97.7%, the recall is 90.8%, the F1 value is 91.1%, the loss rate is 0.02, and the communication efficiency is 93.3%. In the case analysis, the error rate of the proposed model in dealing with Dos and Probe attacks does not exceed 1%, the storage value of abnormal intrusion information is 204 MB, the training time is 47.7 s, and the total expenditure required for the model in actual operation is the lowest. In summary, the raised model can substantially enhance the precision and timeliness of power grid intrusion diagnosis, and possesses significant practical utility, which can be widely applied in smart grid security systems.
To solve the problems that the weights and thresholds of discrete Hopfield neural networks are easy to fall into local optima and have insufficient anti-noise ability in digit recognition, a digit recognition method based on discrete Hopfield neural network is proposed, which is optimized by fish swarm algorithm and called AFSA-HOP integration method. The parameters of the discrete Hopfield neural network are optimized by using AFSA's powerful global search ability, and the recognition accuracy of the Hopfield neural network is taken as the fitness function. This allows the Hopfield neural network to maintain a high associative success rate even under high noise-to-signal ratios. Computer simulation experiments show that while the recognition performance of the traditional Hopfield neural network significantly deteriorates when the noise intensity is 0.2, the AFSA-HOP method maintains a high recognition accuracy even at noise intensities of 0.4 and 0.5, demonstrating superior digital recognition performance. This method provides a robust new approach for digital recognition and could be further extended in future applications by integrating other optimization algorithms.
The March 2025 issue brings four papers from the areas of information security, power load forecasting, time series analysis, and decision support systems.
In modern organizational activities, the increasing complexity and dynamism of strategy management have rendered traditional static analysis and experience- based decision-making methods inadequate for meeting the rapidly changing market demands and intricate internal processes. To address these challenges, this paper proposes an automated machine learning-based data-driven decision support system. The system incorporates a flexible and scalable model that integrates a strategy management automation algorithm, combining Long Short-Term Memory (LSTM) networks and Deep Q-Network (DQN) algorithms, to enhance the scientific and accurate nature of decision-making. The integrated algorithm shows a significantly higher probability of successful decision-making in organizational environments of different scales compared to traditional DQN and random strategies, demonstrating its superiority in complex decision-making scenarios. Key data indicate that the algorithm exhibits strong stability and robustness in terms of error function curves, algorithm performance, and the number of successful decisions, further validating its effectiveness under various interference conditions. While existing research has attempted to apply machine learning to strategy management to some extent, common issues include inadequate handling of time series data, suboptimal strategy optimization, and lack of flexibility in system models. Experimental results show that the algorithm's decision success rate is significantly higher than that of traditional DQN and random strategies across various organizational scales, demonstrating its efficiency and stability in complex decision-making environments. This study not only provides innovative technical means for strategy management but also offers theoretical and practical references for the future development of intelligent decision support systems.
The zero trust systems help to improve the overall security of computer networks, while one of the main challenges of zero trust systems is the construction and optimization of continuous authentication models. Aiming at the shortcomings of the existing authentication performance, this paper proposes a continuous authentication model based on an improved flower pollination algorithm (FPA) and extreme gradient boosting (XGBoost) to improve the accuracy of authentication. The model first uses multiple strategies to optimize FPA to obtain MSFPA; then MSFPA is applied to XGBoost for hyper-parameter tuning to obtain MSFPA-XGBoost; and finally, MSFPA-XGBoost is applied to the user's continuous authentication. Among these approaches, MSFPA incorporates chaotic mapping to enhance the initial population. It also utilizes an adaptive transformation probability strategy to dynamically strike a balance between global and local search. Furthermore, it refines the search equation to optimize both global and local search. Additionally, MSFPA employs a pollen cross-boundary correction strategy and incorporates the concept of cross-mutation to augment the algorithm's exploratory capabilities. The experimental results demonstrate that, in the context of parameter optimization tasks, MSFPA exhibits superior performance compared to other optimization algorithms, specifically in terms of search accuracy and convergence speed. In addition, in terms of continuous identity verification effect, compared with each classification model, MSFPA-XGBoost improves the Accuracy, Recall, Precision, F1-Score, and AUC metrics by an average of 1.84%, 2.49%, 2.33%, 2.39%, and 3.11%, which indicates that the proposed model enhances the accuracy of continuous authentication and is more effectively applicable within zero-trust systems.
The study of stock time series has been an important area of research in economics, finance, and management. Time series feature representation serves as the primary approach for studying time series. The K-line chart is a common representation of stock time series data. This study proposes a segmented representation method KCTEP based on the extreme points of stock K-line portfolio trend for K-line chart data, which is validated in sequence compression rate and distance metric. The experimental results show that the KCTEP method significantly improves the trend description by 8.63% and the compression rate by 2.95% when compared with the uniform extreme point representation method and the piecewise aggregation approximation method (PAA).The results lead to significant enhancement of the trend description effect and reduction of the distance metric error.
With the continuous improvement of the national economy, the development of power enterprises is gradually accelerating, and the popularity of smart grids is also increasing. The power grid data center contains a large amount of user data, and analyzing this data can help power companies predict the load of power plants, thereby improving the resource utilization efficiency of power enterprises. However, current load forecasting models still suffer from information leakage and inaccurate predictions during data transmission, storage, and analysis processes. To solve the above problems, this study uses federated learning technology to optimize the long short-term memory network algorithm and analyzes power grid data and load forecasting based on the optimized algorithm. This study first conducted comparative experiments on the optimized algorithm and found that the prediction accuracy of the optimized algorithm reached 94.5%, with a prediction time of only 1.2ms. The analysis of the data using a load forecasting model based on this algorithm showed that the data security of the model has been improved by 23.4%. After using this model, the power company's electricity resource utilization rate increased by 31.8% and operating costs decreased by 27.5%. The proposed power grid data analysis and load forecasting model can ensure the privacy of power grid data and improve prediction accuracy, thereby improving the power grid operation efficiency of power enterprises and optimizing enterprise resource allocation.
The June 2023 (Vol. 31, No. 2) issue of CIT. Journal of Computing and Information Technology brings four papers from the areas of fault diagnostics, network security, data processing, and computer vision.
Printed Circuit Board (PCB) defect detection is crucial for ensuring the quality and reliability of electronic devices. The study proposes an enhanced YOLOv5s model for PCB defect detection, which combines Coordinate Attention (CA), Convolutional Block Attention Module (CBAM), and Inception-style convolutions (IO). This model aims to improve the detection accuracy of small defects while reducing computational complexity. Experiments on the PCB defect dataset demonstrate that the proposed CA-CBAM-IOYOLOv5s model achieves higher accuracy (97.8%), recall (98.6%), and F1 score (98.3%) compared to the basic YOLOv5s and other state-of-the-art models. The model also shows excellent performance in detecting various types of PCB defects, with an average detection accuracy of 98.45% and an average detection time of 0.114 seconds. These results indicate that the proposed model provides a promising solution for efficient and accurate PCB defect detection in industrial applications.
In the information age, patents are an important carrier of scientific research achievements. How to protect patents and effectively transmit them has become an important development measure in the current information age. A big data-related patent retrieval technology based on filtering rules is proposed to address issues such as poor keyword and phrase retrieval positioning in patent analysis. This new technology combines multiple filtering and retrieval methods and builds a data storage and transmission system. The model achieved the best performance when the threshold was set to 100. The frequency of using the training set before and after keyword filtering increased by 10 and the frequency of using the test set before and after keyword filtering increased by 16. The Euclidean distance of the research method decreased by 0.883 compared to other methods. The mean value increased by 0.1611 compared to other methods. The cosine value increased by 0.4300 compared to other methods. Therefore, the new method has a better filtering effect on patent keywords compared to other methods. This has a good guiding effect on the retrieval of big data-related patents.
The operation of virtual power plants in the electricity market requires handling complex resource scheduling and market trading decision-making problems. The research aims to enhance the participation efficiency and responsiveness of virtual power plants in the electricity market and solve practical operational challenges by improving market trading strategies. Therefore, a resource grading model based on improved support vector machine was developed. The model is optimized using adaptive synthetic sampling, principal component analysis, and deep clustering algorithms. In addition, an improved long short-term memory network is utilized to achieve ultra short-term load forecasting. The results showed that the recall rate and F1 mean of the resource grading model based on the improved support vector machine algorithm were as high as 81.07% and 85.41%, respectively. The average prediction error of the improved long short-term memory neural network algorithm is 0.35%, and the maximum error is only 0.62%. In the basic scenario, the maximum deviation between the declared amount of backup auxiliary services based on load adjustable capacity prediction and the actual amount is only 88.62 kW. The method proposed by the research institute has significant advantages in improving the efficiency and responsiveness of virtual power plant market participation, which is conducive to promoting the overall economic benefits of virtual power plants in the electricity market.
The prediction of power output from photovoltaic generation clusters is crucial for optimizing the dispatch of regional photovoltaic generation. Enhancing the accuracy of power prediction for photovoltaic power plant clusters requires the segmentation of distributed photovoltaic systems into clusters. This paper proposes a method for partitioning distributed photovoltaic clusters using a multiobjective genetic algorithm NSGA2, with spatial distance modularity and electricity similarity as optimization objectives to determine the optimal cluster partitioning scheme. The numerical examples and experimental results of the case analysis demonstrate a significant improvement in the convergence speed of the prediction system when employing the clustering partitioning method. This cluster segmentation algorithm significantly reduces the complexity and investment cost of the prediction system.
