Scholarlink

<p>The smart factory is a highly automated and connected production facility that uses advanced technologies and smart manufacturing principles to optimize efficiency, flexibility, and productivity. Smart factories rely on artificial intelligence as their primary technology. Applying machine learning (ML) and deep learning (DL) algorithms has yielded positive outcomes in fault identification. It is anticipated that the number of problems with machinery will rise as the use of smart machinery increases. The number of ML and DL algorithms used to diagnose and detect machinery faults is growing daily. This paper compares and reviews the commonly used ML and DL models used for fault detection. To identify and diagnose faults, we used three publicly accessible fault datasets: Case Western Reserve University (CWRU), Intelligent Maintenance Systems (IMS), and Tennessee Eastman Process Simulation (TEP) datasets. All experiments were conducted using a stratified 80 − 20 train-test split, with 10% of the training set allocated to validation, and a random seed of 42. We also compare these algorithms with recently published models. The results demonstrate that CNN-BiGRU (a deep learning model that incorporates convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU)) performs better than all other models for the CWRU dataset, achieving 99.82%, 99.80%, 99.86%, 99.77%, 99.81%, and 99.99% values for accuracy, precision, recall, f1score, and specificity. For the IMS dataset, ExtraTrees performs better than any other model with 98.03%, 97.33%, 97.76%, 97.95%, 97.85%, and 99.36% values for accuracy, precision, recall, f1score, and specificity. For the TEP dataset, LightGBM performs better than any other model with a value of 92.43%, 92.05%, 93.81%, 92.43%, 92.84%, and 99.55% for accuracy, precision, recall, f1score, and specificity. Statistical analysis, including Friedman and Nemenyi tests, confirmed significant results.</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

Comparative analysis of machine learning and deep learning approaches for fault detection in industrial applications using CWRU, IMS, and TEP datasets

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Mohamed Abdel-Basset,
  • Alaa Elmor,
  • Karam M. Sallam,
  • Ibrahim Alrashdi,
  • Bilal Arain

摘要

The smart factory is a highly automated and connected production facility that uses advanced technologies and smart manufacturing principles to optimize efficiency, flexibility, and productivity. Smart factories rely on artificial intelligence as their primary technology. Applying machine learning (ML) and deep learning (DL) algorithms has yielded positive outcomes in fault identification. It is anticipated that the number of problems with machinery will rise as the use of smart machinery increases. The number of ML and DL algorithms used to diagnose and detect machinery faults is growing daily. This paper compares and reviews the commonly used ML and DL models used for fault detection. To identify and diagnose faults, we used three publicly accessible fault datasets: Case Western Reserve University (CWRU), Intelligent Maintenance Systems (IMS), and Tennessee Eastman Process Simulation (TEP) datasets. All experiments were conducted using a stratified 80 − 20 train-test split, with 10% of the training set allocated to validation, and a random seed of 42. We also compare these algorithms with recently published models. The results demonstrate that CNN-BiGRU (a deep learning model that incorporates convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU)) performs better than all other models for the CWRU dataset, achieving 99.82%, 99.80%, 99.86%, 99.77%, 99.81%, and 99.99% values for accuracy, precision, recall, f1score, and specificity. For the IMS dataset, ExtraTrees performs better than any other model with 98.03%, 97.33%, 97.76%, 97.95%, 97.85%, and 99.36% values for accuracy, precision, recall, f1score, and specificity. For the TEP dataset, LightGBM performs better than any other model with a value of 92.43%, 92.05%, 93.81%, 92.43%, 92.84%, and 99.55% for accuracy, precision, recall, f1score, and specificity. Statistical analysis, including Friedman and Nemenyi tests, confirmed significant results.