Scholarlink

<p>The vast amount of time-series data generated by industrial equipment operation contains rich state information. These data, characterized by high dimensionality, non-stationarity, and intense noise, reflect the complex dynamic behaviors of equipment. However, the unpredictable nature of anomalies may lead to sudden failures, posing significant threats to production efficiency. Traditional time-series forecasting methods often struggle to capture multi-scale dynamic variations and long-term dependencies when processing high-dimensional, non-stationary, and noisy data in industrial scenarios such as steel continuous casting, resulting in insufficient prediction accuracy that fails to meet the demands of industrial predictive maintenance. Accurately forecasting future equipment states and timely anomaly detection are crucial to achieving industrial predictive maintenance and mitigating or preventing unexpected failure risks. This study innovatively proposes a predictive maintenance framework based on time-series data and a long-term forecasting method tailored for industrial time-series data. The framework consists of two core modules: forecasting and anomaly detection. To effectively capture multi-scale dynamic variations in equipment, the forecasting module first employs a wavelet transform to decompose time-series data into global trends and local fluctuations. Subsequently, an adaptive masked attention mechanism is introduced to fuse multi-scale features, deeply exploring long-term dependencies and latent patterns in the data to enhance long-term forecasting accuracy. The anomaly detection module aims to improve the model’s generalizability in industrial scenarios and facilitate its transferability to anomaly detection tasks. To this end, this study adopts a few-shot fine-tuning strategy to optimize the pre-trained time-series model. Experimental results on steel continuous casting processes demonstrate that the proposed method improves forecasting accuracy by 30.6% compared to traditional approaches. It achieves an F1-score of 0.84 for anomaly detection, fully validating its superiority in industrial predictive maintenance.</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

Wavelet-Inspired Encoder Model Applied in Industrial Predictive Maintenance

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Xueyi Wang,
  • Qi Zhang,
  • Baiyan Li,
  • Sen Wang,
  • Jinsong Bao

摘要

The vast amount of time-series data generated by industrial equipment operation contains rich state information. These data, characterized by high dimensionality, non-stationarity, and intense noise, reflect the complex dynamic behaviors of equipment. However, the unpredictable nature of anomalies may lead to sudden failures, posing significant threats to production efficiency. Traditional time-series forecasting methods often struggle to capture multi-scale dynamic variations and long-term dependencies when processing high-dimensional, non-stationary, and noisy data in industrial scenarios such as steel continuous casting, resulting in insufficient prediction accuracy that fails to meet the demands of industrial predictive maintenance. Accurately forecasting future equipment states and timely anomaly detection are crucial to achieving industrial predictive maintenance and mitigating or preventing unexpected failure risks. This study innovatively proposes a predictive maintenance framework based on time-series data and a long-term forecasting method tailored for industrial time-series data. The framework consists of two core modules: forecasting and anomaly detection. To effectively capture multi-scale dynamic variations in equipment, the forecasting module first employs a wavelet transform to decompose time-series data into global trends and local fluctuations. Subsequently, an adaptive masked attention mechanism is introduced to fuse multi-scale features, deeply exploring long-term dependencies and latent patterns in the data to enhance long-term forecasting accuracy. The anomaly detection module aims to improve the model’s generalizability in industrial scenarios and facilitate its transferability to anomaly detection tasks. To this end, this study adopts a few-shot fine-tuning strategy to optimize the pre-trained time-series model. Experimental results on steel continuous casting processes demonstrate that the proposed method improves forecasting accuracy by 30.6% compared to traditional approaches. It achieves an F1-score of 0.84 for anomaly detection, fully validating its superiority in industrial predictive maintenance.