30 July 2025 in IEEE Access
In modern cloud data centers, accurately predicting job failures before they occur is essential for ensuring system reliability, availability, and efficiency. To address this challenge, researchers have progressively developed machine learning and deep learning techniques that examine cloud logs to identify patterns linked to such failures. To surpass the prediction accuracy of earlier models, this paper presents the Multilayer Multi-Prediction Framework (MMPF), a novel approach specifically designed to enhance failure prediction accuracy. The framework is divided into two distinct layers. The first layer aims to detect job failures using a fine-tuned voting mechanism, while the second layer identifies the specific failure type. In the first layer, several classifiers are integrated to build a robust predictive model, which outperforms conventional methods that rely on a single classifier. By aggregating the predictions from multiple base classifiers and making the final decision based on a weighted average of the predicted probabilities, this approach delivers significantly higher accuracy in failure prediction. Decision Trees, K-Nearest Neighbors, Extreme Gradient Boosting, Adaptive Boosting, and Artificial Neural Networks are the classifiers implemented in this layer. Once the failed jobs are detected, the second layer takes over to determine the failure type employing the Random Forest algorithm. Based on experiments conducted with the Google Cluster 2019 trace dataset, the framework achieved impressive accuracy: 99.83% in detecting failed jobs and 99.97% in identifying the nature of the failure.
1 October 2025 in International Journal of Electrical and Computer Engineering (IJECE)
The rapid growth of industrialization and internet of things (IoT) driven advancements in Industry 5.0 necessitates efficient and user-friendly engineering solutions. Gas leakage incidents in coal mines, chemical enterprises, and households pose significant risks to ecosystems and human safety, emphasizing the need for automated and rapid gas-type detection. Traditional detection methods rely on single-source data and focus on isolated spatial or temporal features, limiting accuracy. This paper proposes a multimodal artificial intelligence (AI) fusion technique combining pre-trained convolutional neural networks (CNNs), such as VGG16, with a deep neural network (DNN) model. The particle swarm optimization (PSO) algorithm optimizes CNN hyperparameters, outperforming traditional trial-and-error methods. The system addresses challenges posed by gases being odorless, colorless, and tasteless, which limit conventional human detection methods. By leveraging sensor fusion, the late fusion technique integrates distinct network architectures for unified gas identification. Experimental results demonstrate 95% accuracy using DNN with gas sensor data, 96% with optimized VGG16 using thermal imaging, and 99.5% through multimodal late fusion. This IoT-enhanced solution outperforms single-sensor approaches, offering a robust and reliable gas leakage detection system suitable for industrial and smart city applications.