Scholarlink

<p>Real-time forecasting of multivariate time series parameters in forging processes is essential for precise control, but existing models often struggle with transient dynamics and multivariate interdependencies. This study proposes a hybrid statistical and neural network (HSNN) model that integrates autoregressive integrated moving average (ARIMA) module with hierarchical deep learning blocks to incrementally refine linear trends and nonlinear residuals. The HSNN uniquely combines dual attention mechanisms (feature and temporal) with ARIMA-deep learning residual blocks, dynamically weighting multivariate parameter relationships while progressively correcting errors through residual propagation. Validated on 28,800 industrial samples, the HSNN achieves the mean absolute error (<i>MAE</i>) values as low as 0.0153 for vertical clamping percentage and 0.0458 for forging force, outperforming ten benchmarks by 56.52% ~ 78.94% in <i>MAE</i>. Generalization tests on an external dataset from our previous work confirm a 67.27% reduction in <i>MAE</i> compared to traditional backpropagation networks. This research bridges the gap between statistical efficiency and deep learning adaptability, providing a deployable solution for real-time forging control.</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

A novel hybrid statistical and neural network model for forecasting multivariate time series parameters in forging process

  • 期刊论文

  • 发表日期：

    2025年4月7日

查看全文
  • Ning-Fu Zeng,
  • Yong-Cheng Lin,
  • Miao Wan,
  • Gui-Cheng Wu,
  • Ming-Song Chen,
  • Chao Li

摘要

Real-time forecasting of multivariate time series parameters in forging processes is essential for precise control, but existing models often struggle with transient dynamics and multivariate interdependencies. This study proposes a hybrid statistical and neural network (HSNN) model that integrates autoregressive integrated moving average (ARIMA) module with hierarchical deep learning blocks to incrementally refine linear trends and nonlinear residuals. The HSNN uniquely combines dual attention mechanisms (feature and temporal) with ARIMA-deep learning residual blocks, dynamically weighting multivariate parameter relationships while progressively correcting errors through residual propagation. Validated on 28,800 industrial samples, the HSNN achieves the mean absolute error (MAE) values as low as 0.0153 for vertical clamping percentage and 0.0458 for forging force, outperforming ten benchmarks by 56.52% ~ 78.94% in MAE. Generalization tests on an external dataset from our previous work confirm a 67.27% reduction in MAE compared to traditional backpropagation networks. This research bridges the gap between statistical efficiency and deep learning adaptability, providing a deployable solution for real-time forging control.