<p>Cold rolling strip production is a multi-stand continuous rolling process, so flatness prediction is a typical spatiotemporal series data prediction problem, which requires considering various complex factors affecting flatness and paying attention to the correlation of its spatiotemporal dimensions. Based on this, a cold rolled strip flatness prediction model is proposed, integrating a Bidirectional Gated Recurrent Unit (BiGRU), an Attention mechanism, and an Inverted Transformer (iTransformer). The model adopts a parallel structure, where one branch utilizes a BiGRU-Attention module designed to capture the spatiotemporal correlations in strip production data, with the BiGRU’s hidden layer dimension set to 128; the other branch employs an iTransformer module with a feature dimension of 256 and 8 attention heads to effectively extract key features and model the relationships between parameters using the self-attention mechanism. The features extracted from both branches are fused into a 128-dimensional vector, which is then passed through a fully connected layer for flatness prediction. The prediction results show that the error indicators MSE, RMSE and MAE of the proposed model are 0.937, 0.968 and 0.774 respectively, and the fitting performance indicator <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11837_2025_7456_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> is 0.974, which are better than the comparison models Random Forest (RF), Deep Neural Network (DNN), Long Short-Term Memory (LSTM), BiGRU, iTransformer, and BiGRU-iTransformer models. Feature ablation experiments show that for flatness, the importance of parameters is ranked as follows: forward tension (Tf), roll gap difference (Gd), work roll bending force (WR), exit thickness (Hb), rolling force (<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11837_2025_7456_Article_IEq2.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="16" /> </InlineMediaObject> <EquationSource Format="TEX">\(F\)</EquationSource> <EquationSource Format="MATHML"><math> <mi>F</mi> </math></EquationSource> </InlineEquation>), intermediate roll bending force (IR), strip yield strength (<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11837_2025_7456_Article_IEq3.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="14" /> </InlineMediaObject> <EquationSource Format="TEX">\(Y\)</EquationSource> <EquationSource Format="MATHML"><math> <mi>Y</mi> </math></EquationSource> </InlineEquation>), entrance thickness (Hf), rolling speed (<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11837_2025_7456_Article_IEq4.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="14" /> </InlineMediaObject> <EquationSource Format="TEX">\(V\)</EquationSource> <EquationSource Format="MATHML"><math> <mi>V</mi> </math></EquationSource> </InlineEquation>), backward tension (Tb), strip width (<InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11837_2025_7456_Article_IEq5.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="17" /> </InlineMediaObject> <EquationSource Format="TEX">\(B\)</EquationSource> <EquationSource Format="MATHML"><math> <mi>B</mi> </math></EquationSource> </InlineEquation>). The experimental results are consistent with expert knowledge and underlying mechanisms, verifying the effectiveness of the attention mechanism while enhancing interpretability and credibility.</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

Prediction of Tandem Cold-Rolled Strip Flatness Based on the BiGRU-Attention-iTransformer Model

  • Ming-hua Liu,
  • Ya-han Li,
  • Da-yuan Wu,
  • Zi-xuan Zhu

摘要

Cold rolling strip production is a multi-stand continuous rolling process, so flatness prediction is a typical spatiotemporal series data prediction problem, which requires considering various complex factors affecting flatness and paying attention to the correlation of its spatiotemporal dimensions. Based on this, a cold rolled strip flatness prediction model is proposed, integrating a Bidirectional Gated Recurrent Unit (BiGRU), an Attention mechanism, and an Inverted Transformer (iTransformer). The model adopts a parallel structure, where one branch utilizes a BiGRU-Attention module designed to capture the spatiotemporal correlations in strip production data, with the BiGRU’s hidden layer dimension set to 128; the other branch employs an iTransformer module with a feature dimension of 256 and 8 attention heads to effectively extract key features and model the relationships between parameters using the self-attention mechanism. The features extracted from both branches are fused into a 128-dimensional vector, which is then passed through a fully connected layer for flatness prediction. The prediction results show that the error indicators MSE, RMSE and MAE of the proposed model are 0.937, 0.968 and 0.774 respectively, and the fitting performance indicator \(R^{2}\) R 2 is 0.974, which are better than the comparison models Random Forest (RF), Deep Neural Network (DNN), Long Short-Term Memory (LSTM), BiGRU, iTransformer, and BiGRU-iTransformer models. Feature ablation experiments show that for flatness, the importance of parameters is ranked as follows: forward tension (Tf), roll gap difference (Gd), work roll bending force (WR), exit thickness (Hb), rolling force ( \(F\) F ), intermediate roll bending force (IR), strip yield strength ( \(Y\) Y ), entrance thickness (Hf), rolling speed ( \(V\) V ), backward tension (Tb), strip width ( \(B\) B ). The experimental results are consistent with expert knowledge and underlying mechanisms, verifying the effectiveness of the attention mechanism while enhancing interpretability and credibility.