Scholarlink

<p>Early and reliable smoke detection is critical for preventing fire-related disasters, safeguarding human lives, and ensuring infrastructure safety. Conventional sensor-based systems, however, are often constrained by fixed threshold settings, delayed responses, and high false alarm rates, which limit their reliability in dynamic real-world Internet of Things (IoT) environments. To address these limitations, this study presents a comprehensive deep learning-based framework for intelligent smoke detection, integrating a diverse set of neural architectures ranging from conventional models such as Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM), to advanced architectures including Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), Deep LSTM, Deep BiLSTM, Deep GRU, and hybrid models such as CNN-LSTM and Stacked CNN-LSTM. To ensure robust model learning, the dataset was standardized using the StandardScaler and balanced using the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance between smoke and non-smoke samples. Each model was further optimized using the Optuna Bayesian hyperparameter optimization framework, enabling systematic and automated fine-tuning of learning parameters for optimal convergence. The experimental evaluation employed multiple binary classification metrics, including Accuracy, Precision, Recall, F<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1\)</EquationSource> </InlineEquation>-Score, ROC-AUC, and PR-AUC, alongside computational efficiency indicators such as training time, testing time, and peak memory usage. The results demonstrate strong predictive performance across all models, with the Hybrid CNN-LSTM model consistently outperforming others across most predictive evaluation metrics, exhibiting superior generalization capability and convergence behavior. Furthermore, the MLP, LSTM, and CNN models achieved the lowest memory footprints and shortest training and testing durations. These findings highlight the effectiveness of deep learning for intelligent smoke detection systems capable of delivering early, accurate, and reliable fire warnings, offering a scalable solution for modern fire prevention and safety applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An optimized deep learning framework for IoT-based smoke detection with enhanced performance and computational efficiency

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月28日

查看全文
  • Mir Nazish,
  • Abdul Manan,
  • M. Tariq Banday

摘要

Early and reliable smoke detection is critical for preventing fire-related disasters, safeguarding human lives, and ensuring infrastructure safety. Conventional sensor-based systems, however, are often constrained by fixed threshold settings, delayed responses, and high false alarm rates, which limit their reliability in dynamic real-world Internet of Things (IoT) environments. To address these limitations, this study presents a comprehensive deep learning-based framework for intelligent smoke detection, integrating a diverse set of neural architectures ranging from conventional models such as Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM), to advanced architectures including Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), Deep LSTM, Deep BiLSTM, Deep GRU, and hybrid models such as CNN-LSTM and Stacked CNN-LSTM. To ensure robust model learning, the dataset was standardized using the StandardScaler and balanced using the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance between smoke and non-smoke samples. Each model was further optimized using the Optuna Bayesian hyperparameter optimization framework, enabling systematic and automated fine-tuning of learning parameters for optimal convergence. The experimental evaluation employed multiple binary classification metrics, including Accuracy, Precision, Recall, F \(1\) -Score, ROC-AUC, and PR-AUC, alongside computational efficiency indicators such as training time, testing time, and peak memory usage. The results demonstrate strong predictive performance across all models, with the Hybrid CNN-LSTM model consistently outperforming others across most predictive evaluation metrics, exhibiting superior generalization capability and convergence behavior. Furthermore, the MLP, LSTM, and CNN models achieved the lowest memory footprints and shortest training and testing durations. These findings highlight the effectiveness of deep learning for intelligent smoke detection systems capable of delivering early, accurate, and reliable fire warnings, offering a scalable solution for modern fire prevention and safety applications.