International Journal of Advances in Intelligent Informatics
Not a member yet
235 research outputs found
Sort by
Enhanced intrusion detection in smart grids using extended long short-term memory variants
No full textSmart grid systems, which integrate traditional energy infrastructure with modern communication technologies, face significant cybersecurity challenges due to their dynamic architecture and continuous data exchange. The diversity and interconnectivity of devices increase vulnerability to malicious intrusions, underscoring the need for advanced, scalable detection methods. This study aims to develop an intrusion detection system (IDS) for smart grids by leveraging recent advances in deep learning, specifically enhanced variants of Long Short-Term Memory (LSTM), xLSTM, sLSTM, and mLSTM. These sequence modelling architectures were adapted and fine-tuned within our IDS framework to capture complex spatio-temporal patterns and handle heterogeneous, high-dimensional data effectively. A comprehensive evaluation on two benchmark datasets, NSL-KDD and DNP3, demonstrates the robustness of the proposed approach. On the NSL-KDD, xLSTM, sLSTM, and mLSTM achieved accuracies of 98.16%, 98.55%, and 98.54%, respectively. On the more modern and protocol-specific DNP3 dataset, which represents real-world SCADA-focused attacks, the models maintained their superior performance, with accuracies of 99.50%, 99.33%, and 99.42%, respectively. The high, consistent accuracy across both datasets demonstrates the models' dependability and adaptability for intrusion detection in smart grid infrastructure. The study’s targeted enhancement of LSTM-based architectures contributes a novel and effective approach to protecting critical intelligent systems from emerging cyber threat
Advanced deep learning techniques for sentiment analysis: combining Bi-LSTM, CNN, and attention layers
Full text linkOnline platforms enhance customer engagement and provide businesses with valuable data for predictive analysis, critical for strategic sales forecasting and customer relationship management. This study explores in depth the potential of sentiment analysis (SA) to enhance sales forecasting and customer retention for small and large businesses. We collected a large dataset of product review tweets, representing a rich consumer sentiment source. We developed an artificial neural network based on a dataset of product review tweets that captures both positive and negative sentiments. The core of our model is Bi-LSTM (Bidirectional Long Short-Term Memory) architecture, enhanced by an attention mechanism to capture relationships between words and emphasize key terms. Then, a one-dimensional convolutional neural network with 64 filters of size 3x3 is applied, followed by Average_Max_Pooling to reduce the feature map. Finally, two dense layers classify the sentiment as positive or negative. This research provides significant benefits and contributions to sentiment analysis by accurately identifying consumer sentiment in product review tweets. The proposed model that integrated Bi-LSTM with attention mechanism and CNN detects negative sentiment with a precision of 0.97, recall of 0.98, and F1-score of 0.98, allowing companies to address customer concerns, improving satisfaction and brand loyalty proactively. In addition, the proposed model presents a better sentiment classification on average for both positive and negative sentiments, and accuracy (96%) compared to the other baselines. It ensures high-quality input data by reducing noise and inconsistencies in product review tweets. Moreover, the dataset collected in this study serves as a valuable benchmark for future research in sentiment analysis and predictive analytics
Classification of Bitter gourd leaf disease using deep learning architecture ResNet50
Full text linkThe primary goal of this research is to develop a feasible and efficient method for identifying the disease and to advocate for an appropriate system that provides an early and cost-effective solution to this problem. Due to their superior computational capabilities and accuracy, computer vision and machine learning methods and techniques have garnered significant attention in recent years for classifying various leaf diseases. As a result, Resnet50 and Resnet101 were proposed in this study for the classification of bitter gourd disease. The 2490 images of bitter gourd leaves are classified into three categories: Healthy leaf, Fusarium Wilt leaf, and Yellow Mosaic leaf. The proposed ResNet50 architecture accomplished 98% accuracy with the Adam optimizer. The ResNet101 architecture achieves an average accuracy of 94% with the Adam optimizer. As a result, the proposed model can differentiate between healthy and diseased bitter gourd leaves. This research contributes to the development of methods for detecting bitter melon leaf disease using computer vision and machine learning, achieving high accuracy and supporting automatic disease diagnosis. The results can help farmers quickly and cost-effectively detect diseases early, thereby increasing agricultural productivity
A dual-phase hybrid framework for real-time grayscale image denoising in structured noise
Full text linkImage denoising is a substantial section in the preprocessing stage, especially in medical images. This study proposed a hybrid denoising model for salt-and-pepper removal in grayscale images. The framework uses a U-Net convolutional neural network, modified to perform preliminary denoising, and the Alternating Direction Method (ADM) to refine the structure iteratively. A corrupted pixel location is first determined using an adaptive thresholding scheme. The model is trained with a composite loss function that combines pixel-wise reconstruction accuracy (MSE) and perceptual similarity, as measured by the Structural Similarity Index (SSIM). Tests conducted on benchmarks (e.g., Kodak24, Set14, DIV2K, and TID2013) show that the proposed method surpasses traditional filters and state-of-the-art deep learning models, e.g., FFDNet and DnCNN. The quantitative results are Peak Signal-to-Noise Ratio (PSNR) 32.45 dB, SSIM 0.92 against 30 percent salt-and-pepper noise, and the average speed of inference is 6.2 ms, showing improvements over baseline approaches in performance and appearance. The main innovation is combining a noise-aware adaptive detection step with a specially designed U-Net framework and ADM-sided refinement, achieving better edge preservation and robustness to noise at any level. The framework displays a high potential for use in medical imaging, document recovery, and real-time surveillance
A deep learning ensemble framework for robust classification of lung ultrasound patterns: covid-19, pneumonia, and normal
Full text linkTo advance the automated interpretation of lung ultrasound (LUS) data, multiple deep learning (DL) models have been introduced to identify LUS patterns for differentiating COVID-19, Pneumonia, and Normal cases. While these models have generally yielded promising outcomes, they have encountered challenges in accurately classifying each pattern across diverse cases. Therefore, this study introduces an ensemble framework that leverages multiple classification models, optimizing their contributions to the final prediction through a majority voting mechanism. After training seven different classification models, the three models with the highest accuracies were selected. The ensemble incorporates these top-performing models: EfficientNetV2-B0, EfficientNetV2-B2, and EfficientNetV2-B3, and utilizes this framework to classify patterns in LUS images. Compared to individual model performance, the ensemble approach significantly enhances classification accuracy, achieving an accuracy of 99.25% and an F1-score of 99%. In contrast, the standalone models attained accuracies of 97.8%, 97.6%, and 98.1%, with F1-score of approximately 98%. This research highlights the potential of ensemble learning for improving the accuracy and robustness of automated LUS analysis, offering a practical and scalable solution for real-world medical diagnostics. By combining the strengths of multiple models, the proposed framework paves the way for more reliable and efficient tools to assist clinicians in diagnosing lung diseases
An enhanced pivot-based neural machine translation for low-resource languages
Full text linkThis study examines the efficacy of employing Indonesian as an intermediary language to improve the quality of translations from Javanese to Madurese through a pivot-based approach utilizing neural machine translation (NMT). The principal objective of this research is to enhance translation precision and uniformity among these low-resource languages, hence advancing machine translation models for underrepresented languages. The data collecting approach entailed extracting parallel texts from internet sources, followed by pre-processing through tokenization, normalization, and stop-word elimination algorithms. The prepared datasets were utilized to train and assess the NMT models. An intermediary phase utilizing Indonesian is implemented in the translation process to enhance the accuracy and consistency of translations between Javanese and Madurese. Parallel text corpora were created by collecting and preprocessing data, thereafter, utilized to train and assess the NMT models. The pivot-based strategy regularly surpassed direct translation regarding BLEU scores for all n-grams (BLEU-1 to BLEU-4). The enhanced BLEU ratings signify increased precision in vocabulary selection, preservation of context, and overall comprehensibility. This study significantly enhances the current literature in machine translation and computational linguistics, especially for low-resource languages, by illustrating the practical effectiveness of a pivot-based method for augmenting translation precision. The method's dependability and efficacy in producing genuine translations were proved through numerous studies. The pivot-based technique enhances translation quality, although it possesses limitations, including the risk of error propagation and bias originating from the pivot language. Further research is necessary to examine the integration of named entity recognition (NER) to improve accuracy and optimize the intermediate translation process. This project advances the domains of machine translation and the preservation of low-resource languages, with practical implications for multilingual communities, language education resources, and cultural conservation
Optimization hybrid weighted switching filtering (OHWSF) using SVD and SVD++ for addressing data sparsity
Full text linkRecommender systems are crucial for filtering vast amounts of digital content and providing personalized recommendations; however, their effectiveness is often hindered by data sparsity, where limited user-item interactions lead to reduced prediction accuracy. This study introduces a novel hybrid model, Optimization Hybrid Weighted Switching Filtering (OHWSF), to overcome this challenge by integrating two complementary strategies: Hybrid Weighted Filtering (HWF), which linearly combines predictions from SVD and SVD++ using a weighting parameter (α), and Hybrid Switching Filtering (HSF), which dynamically selects predictions based on a threshold rating (θ). The OHWSF framework introduces a tunable optimization mechanism governed by the parameter σ₁ to adaptively balance weighting and switching decisions based on actual rating deviations. Unlike existing static or manually tuned hybrid methods, the proposed model combines dynamic switching with weight optimization to minimize prediction error effectively. Extensive experiments on four benchmark datasets (ML-100K, ML-1M, Amazon Cell Phones Reviews, and GoodBooks-10K) demonstrate that OHWSF consistently outperforms traditional collaborative filtering (UBCF, IBCF), matrix factorization techniques (SVD, SVD++), and standalone hybrid models across all evaluation metrics (MAE, MSE, RMSE). The model achieves optimal performance within the range of α = 0.6–0.9 and θ = 1.0–1.5, demonstrating robustness across varying sparsity levels. Notably, OHWSF achieves up to 742.16% MAE improvement over the UBCF model, with significantly reduced training time compared to SVD++. These findings confirm that OHWSF significantly improves prediction accuracy, scalability, and adaptability in sparse data environments. This research contributes a flexible, interpretable, and efficient hybrid recommendation framework suitable for real-world applications
Traffic light optimization (TLO) using reinforcement learning for automated transport systems
Full text linkCurrent traffic light systems follow predefined timing sequences, causing the light to turn green even when no cars are waiting, while the side road with waiting vehicles may still face a red light. Reinforcement learning can help by training an intelligent model to analyze real-time traffic conditions and dynamically adjust signal lights based on actual demand and necessity. If the traffic light becomes intelligent and autonomous then it can significantly reduce the time wasted everyday commuting due to previously determined traffic light timing sequences. In our previous work, we used fuzzy logic to control the traffic light where the time was fixed but in this paper, the waiting time becomes a variable that changes depending on other road variables like vehicles, pedestrians, and times. Moreover, we trained an agent in this work using reinforcement learning to optimize the traffic flow in junctions with traffic lights. The trained agent worked using the greedy method to improve traffic flow to maximize the rewards by changing the signals appropriately. We have two states and there are only two actions to take for the agent. The results of the training of the model are promising. In normal situations, the average waiting time was 9.16 seconds. After applying our fuzzy rules, the average waiting time was reduced to 0.26 seconds, and after applying reinforcement learning, it was 0.12 seconds in a simulator. The average waiting time was reduced by 97~98%. These models have the potential to improve real-world traffic efficiency by approximately 67~68%
Enhanced mixup for improved time series analysis
Full text linkTime series data analysis is crucial for real-world applications. While deep learning has advanced in this field, it still faces challenges, such as limited or poor-quality data. In areas like computer vision, data augmentation has been widely used and highly effective in addressing similar issues. However, these techniques are not as commonly explored or applied in the time series domain. This paper addresses the gap by evaluating basic data augmentation techniques using MLP, CNN, and Transformer architectures, prioritized for their alignment with state-of-the-art trends in time series analysis rather than traditional RNN-based methods. The goal is to expand the use of data augmentation in time series analysis. The paper proposed EMixup, which adapts the Mixup method from image processing to time series data. This adaptation involves mixing samples while aiming to maintain the data's temporal structure and integrating target contributions into the loss function. Empirical studies show that EMixup improves the performance of time series models across various architectures (improving 23/24 forecasting cases and 12/24 classification cases). It demonstrates broad applicability and strong results in tasks like forecasting and classification, highlighting its potential utility across diverse time series applications
Finding a suitable chest x-ray image size for the process of Machine learning to build a model for predicting Pneumonia
Full text linkThis study focused on algorithm performance and training/testing time, evaluating the most suitable chest X-ray image size for machine learning models to predict pneumonia infection. The neural network algorithm achieved an accuracy rate of 87.00% across different image sizes. While larger images generally yield better results, there is a decline in performance beyond a certain size. Lowering the image resolution to 32x32 pixels significantly reduces performance to 83.00% likely due to the loss of diagnostic features. Furthermore, this study emphasizes the relationship between image size and processing time, empirically revealing that both increasing and decreasing image size beyond the optimal point results in increased training and testing time. The performance was noted with 299x299 pixel images completing the process in seconds. Our results indicate a balance between efficiency, as larger images slightly improved accuracy but slowed down speed, while smaller images negatively impacted precision and effectiveness. These findings assist in optimizing chest X-ray image sizes for pneumonia prediction models by weighing diagnostic accuracy against computational resources