Scholarlink

<p>Vibration-based fault diagnosis is essential for maintaining safe and reliable operations in industrial facilities, especially in environments in which highly volatile materials are handled, such as in petrochemical plants. This study evaluates the performance of four deep-learning architectures, namely, the one-dimensional convolutional neural network (1DCNN), wavelet transform-based two-dimensional convolutional neural network (WT–2DCNN), long short-term memory (LSTM), and a hybrid 1DCNN–LSTM model, to classify mechanical and leakage-related faults from vibration signals. All experiments were conducted using data collected from a simulation facility for petrochemical processes, which provides operating conditions that resemble actual field environments more closely than typical laboratory setups. The 1DCNN and WT–2DCNN models effectively extracted local temporal and time–frequency features, whereas the LSTM captured long-range temporal dependencies inherent in vibration sequences. Among the evaluated architectures, the 1DCNN–LSTM model achieved the best overall performance, reaching an accuracy of 99.85% with an inference time of only 0.282&#xa0;s. This superior performance can be attributed to a combination of convolution-based short-term feature extraction with recurrent modeling of temporal evolution, enabling the hybrid model to distinguish between four fault conditions with near-perfect consistency. These results highlight the potential of the 1DCNN–LSTM framework as an efficient and reliable solution for real-time fault diagnosis and predictive maintenance in industrial vibration-monitoring 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

Comparative study of the 1DCNN, WT–2DCNN, LSTM, and 1DCNN-LSTM models to predict machinery faults based on vibrations measured from a simulation facility for petrochemical processes

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Daehyeon Ji,
  • Chang-Ju Park,
  • Hyeon Woo Kim,
  • Yangkyu Park

摘要

Vibration-based fault diagnosis is essential for maintaining safe and reliable operations in industrial facilities, especially in environments in which highly volatile materials are handled, such as in petrochemical plants. This study evaluates the performance of four deep-learning architectures, namely, the one-dimensional convolutional neural network (1DCNN), wavelet transform-based two-dimensional convolutional neural network (WT–2DCNN), long short-term memory (LSTM), and a hybrid 1DCNN–LSTM model, to classify mechanical and leakage-related faults from vibration signals. All experiments were conducted using data collected from a simulation facility for petrochemical processes, which provides operating conditions that resemble actual field environments more closely than typical laboratory setups. The 1DCNN and WT–2DCNN models effectively extracted local temporal and time–frequency features, whereas the LSTM captured long-range temporal dependencies inherent in vibration sequences. Among the evaluated architectures, the 1DCNN–LSTM model achieved the best overall performance, reaching an accuracy of 99.85% with an inference time of only 0.282 s. This superior performance can be attributed to a combination of convolution-based short-term feature extraction with recurrent modeling of temporal evolution, enabling the hybrid model to distinguish between four fault conditions with near-perfect consistency. These results highlight the potential of the 1DCNN–LSTM framework as an efficient and reliable solution for real-time fault diagnosis and predictive maintenance in industrial vibration-monitoring applications.