Width spread is a critical quality indicator in the hot strip rolling (HSR) manufacturing process. To improve prediction accuracy, a physics-informed machine learning framework with residual learning (PI-MLRL) is proposed, in which a mechanism model, a light gradient boosting machine (LightGBM)-based residual learning module, and a physics-constrained distillation mechanism are integrated. By combining physical consistency with nonlinear fitting capability, an accurate mapping between process variables and width spread is achieved. Experimental results show that the proposed framework outperforms the mechanism model and seven representative data-driven models in terms of mean absolute error, root-mean-square error, and coefficient of determination. Moreover, Shapley additive explanations (SHAP) method is employed for interpretable diagnostics of PI-MLRL predictions, clarifying the effects of key variables on width spread under different operating conditions. Finally, the proposed framework was deployed on a 2160-mm HSR production line, and application results showed that the width spread prediction error was maintained within ± 3 mm, thereby confirming its engineering applicability.

[Objective]The precise estimation of rolling force during the process of cold continuous rolling is of paramount importance for ensuring product quality,enhancing automation levels,improving production efficiency,and optimizing process settings.However,the conventional cold rolling force mechanism model often relies solely on process parameters during the cold-rolling stage.It disregards the genetic effects of the hot rolling process on the material's structure and properties,and it cannot effectively capture the complex,nonlinear impact of cross-process parameters on the rolling force.This results in limited prediction accuracy and generalization ability.This study proposes a cold rolling force prediction model based on Bayesian optimization and an improved light gradient boosting machine(BO-LightGBM),aiming to comprehensively explore the process coupling between hot and cold rolling.The model aims to enhance adaptability and accuracy in predicting force during cold rolling across various steel grades and production scenarios.[Methods]The modeling process involves the development of a multi-source feature system,incorporating 7-dimensional hot rolling parameters(e.g.,finish rolling temperature,coiling temperature,final thickness,etc.),11-dimensional cold rolling parameters(e.g.,strip width,rolling speed,deformation resistance coefficient,etc.),and the predicted output from the traditional mechanism-based model.This comprehensive feature set enables the model to represent a cross-process fusion of variables that collectively influence rolling force.It is acknowledged that there is variability and coupling across different rolling stands in a tandem cold rolling mill.Therefore,a stand-specific modeling strategy is employed.The development of independent prediction models tailored to each stand facilitates the capture of local process characteristics,nonlinear interactions,and contextual dependencies.Furthermore,a Bayesian optimization algorithm is employed to automatically fine-tune the hyperparameters of the BO-LightGBM model for each stand.This approach effectively reduces human intervention,avoids suboptimal manual tuning,and enhances the efficiency and robustness of the learning process.[Results]The impact of incorporating upstream process data was evaluated by training rolling force prediction models on two distinct datasets.The first dataset contained only cold rolling parameters,while the second dataset included both hot and cold rolling parameters.A series of comparative experiments have demonstrated that 1)following the implementation of the hot rolling process parameters,the mean absolute error in rolling force prediction for each stand has been shown to decrease by an average of 1.803 t,whereas the root mean square error has been demonstrated to decrease by an average of 2.573 t,thereby indicating a substantial enhancement in accuracy.2)In a real industrial cold rolling setting system,the model has been found to enhance the rolling force setpoint accuracy for MR T-4CA and MR T-5CA steel grades by 2.024%and 1.962%,respectively,thereby underscoring its practical engineering value and operational significance.[Conclusions]The proposed BO-LightGBM rolling force prediction model demonstrates excellent performance in terms of accuracy,robustness,and generalization.The model effectively incorporates upstream hot rolling data and employs stand-specific learning with automated hyperparameter optimization,thereby capturing the hereditary influence of the hot rolling stage and overcoming the limitations of traditional mechanism models in cross-process modeling.The model offers a promising data-driven solution for intelligent process control in modern steel rolling operations and supports the advancement of smart manufacturing in the metallurgical industry.

Accurate evaluation of mechanical properties is essential for advanced manufacturing and quality control of metallic materials. Although laser ultrasonic techniques provide a non-contact and non-destructive solution for material characterization, robust and physically interpretable prediction of macroscopic mechanical properties from complex ultrasonic signals remains challenging. In this study, a physically interpretable multimodal convolutional neural network (CNN) framework, guided by ultrasonic time-frequency characteristics, is proposed for mechanical property prediction by jointly exploiting laser ultrasonic time-frequency representations and process-related parameters. Specifically, laser ultrasonic signals are transformed into continuous wavelet transform-based time-frequency images and fused with macroscopic process parameters to predict yield strength, tensile strength, and elongation. To explicitly address model interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM)-based visualization and statistical contribution analysis are employed to identify time-frequency features most relevant to accurate prediction. In parallel, the influence of ultrasonic waveform representation strategies is systematically investigated by comparing full-wave inputs with the conservative peak-front half-wave truncation commonly adopted in industrial laser ultrasonic analysis. The results demonstrate that the proposed multimodal framework enables robust and stable prediction of strength-related properties, whereas the predictability of elongation remains inherently limited. Importantly, full-wave ultrasonic representations consistently outperform conservative waveform truncation, providing more comprehensive information and leading to improved prediction accuracy and feature attribution consistency. Overall, this study establishes a practical and interpretable multimodal learning framework for laser ultrasonic-based mechanical property evaluation, offering improved robustness and enhanced physical insight for industrial laser ultrasonic applications.

The deformation and stress behavior during shear-compression bonding of intermediate slabs was investigated, and a systematic evaluation index to quantify bonding quality and process efficiency was established. A finite element simulation model based on plastic deformation theory was developed to analyze the influence of key process parameters. Single-factor analysis and an orthogonal experimental design, combined with range and variance analysis, were employed to evaluate the significance and sensitivity of multiple parameters. A quadratic polynomial regression model was further constructed to describe nonlinear relationships between process variables and performance indices. The results indicate that edge width, overlap amount, reduction, reduction speed, and slab temperature are the dominant factors governing bonding quality. The optimal parameter combination was determined as an edge width of 30 mm, overlap amount of 3 mm, reduction of 20 mm, reduction speed of 60 mm/s, and slab temperature of 1060 °C. The proposed integrated approach provides a reliable basis for process parameter optimization and contributes to enhancing bonding performance and consistency in endless rolling operations.

In hot strip rolling (HSR) production, strip profile quality is critical to product performance and subsequent processing. Compared with finite element and mechanistic models that depend on prior assumptions, machine learning (ML) can efficiently capture the complex nonlinear relationships between process parameters and profile, thereby offering significant advantages in predictive accuracy. However, most existing strip profile prediction approaches rely on single "black-box" ML models and are limited to mean value prediction, which fails to simultaneously capture the overall evolution of the strip full-length profile and its local fluctuations. This limitation primarily arises from the inherent residuals of single ML model predictions, which are often associated with long-term process drifts or short-term disturbances during rolling. Furthermore, the lack of transparent physical interpretability in "black-box" models restricts their trustworthy deployment in industrial practice. To address the above challenges, we propose a novel paradigm for high-precision and interpretable strip profile prediction, which integrates efficient profile feature extraction, light gradient boosting machine (LightGBM)based multi-scale residual correction regression, and shapley additive explanations (SHAP)-driven post-hoc interpretability analysis, thereby achieving a balance among predictive accuracy, robustness, and transparency. This framework has been successfully deployed on a 2250 mm HSR production line. Experimental results demonstrate that the proposed feature extraction method achieves superior reconstruction accuracy compared with both the sixth-order polynomial fitting method and the legendre polynomial fitting method. In industrial applications, the proposed framework demonstrated high prediction accuracy for C40, with more than 93% of the coils showing prediction errors within +/- 4 mu m. Moreover, the SHAP-based feature contribution analysis revealed the key process variables governing strip profile formation, thereby ensuring both transparency and reliability of the proposed prediction paradigm.

本文重点分析了感应加热技术在提升热轧带钢质量与成材率中的核心作用.边部加热技术通过精准补偿边部温降,可显著改善带钢温度均匀性,细化边部晶粒,降低边裂缺陷率,从而提升成材率.本文系统阐述了边部加热器的系统组成、应用效果分析、仿真研究及配套系统,特别是磁-热-结构全耦合模型等仿真技术的应用,可有效缩短工艺设计周期并降低设备调试能耗.此外,控制参数的智能化升级,结合数字孪生与强化学习算法,重塑工艺调控范式.最后,本文展望了边部加热技术的未来发展趋势,包括多物理耦合场求解、控制参数优化、设备选型逻辑等方向.

In the production process of hot strip rolling, the on-load roll gap shape of the downstream stand directly affects the strip profile of the finished product and the strip shape control at the end of rolling, including the original roll shape, thermal expansion shape of the roll and wear roll shape. The roll wear mechanism is complex, which is also the key factor to determine the total amount and profile quality of rolled strip during the roll service cycle of hot rolling. The existing roll wear models are data empirical models, and have not yet established a theoretical model that fully considers the influencing factors. This paper proposes a roll wear calculation model suitable for hot rolling mills. The model is based on the fatigue wear theory in the friction and wear principle, the spherical asperity model and the micro-protrusions distribution model based on the statistical model are applied. The model can characterize the force, friction coefficient and contact state of the roll at various points along the strip width. According to the roll characteristics of the four-high rolling mill, the wear models of work rolls and backup rolls were established respectively. On this basis, the roll wear prediction model was established. The established model can accurately predict the amount of roll wear and the wear shape of the roll at any time and predictive error within 5%, which provides a good support for the shape control and process optimization of the roll during the service period.

Accurately capturing the complex relationships among chemical composition, process parameters, and mechanical properties is essential for quality control in hot strip rolling, yet theoretical models are constrained by idealized assumptions and simplified boundary conditions, making them unsuitable for applications under highly variable industrial conditions. To address these limitations, this study develops an ensemble learning framework guided by the technique for order preference by similarity to ideal solution. Built on an industrial data platform, the framework unifies process and quality data, adaptively combines multiple base learners through performance-based weighting, and uses optimization-driven hyperparameter tuning to achieve high-precision prediction of mechanical properties. Shapley additive explanations analysis is further employed to interpret feature interactions and quantify their contributions to mechanical behavior, providing mechanistic insight that supports alloy design and process optimization. Validation on a domestic hot strip rolling production line demonstrates clear advantages over conventional ensemble methods, particularly for tensile strength prediction, where the model achieves an R 2 of 0.97, an mean absolute error of 11.82 MPa and an root mean square error of 17.15 MPa, confirming its predictive accuracy and industrial applicability.

Wedge shape is a critical indicator of strip profile quality in hot strip rolling, yet its control remains challenging due to inter-stand delays, nonlinear dynamics, and stochastic disturbances. To address these challenges, this study develops an artificial intelligence (AI)-based control framework for high-precision wedge regulation in the finishing stage. A model predictive control (MPC) strategy is formulated with consideration of inter-stand time-delay effects, and the weighting parameters of its cost function are optimized offline using a genetic algorithm (GA) to enable adaptive allocation of control effort among multiple rolling stands under physical constraints. To evaluate control performance under industrial conditions, a support vector regression (SVR) model is constructed as a data-driven virtual plant to predict wedge deviation based on key process parameters. Industrial implementation on a 1580 mm hot strip mill producing silicon steel demonstrates that the proposed strategy improves wedge shape consistency, increasing the proportion of strip length within a wedge deviation of plus or minus 20 mu m by approximately five percentage points and improving the coil qualification rate by 10.56 percentage points. These results confirm that integrating artificial intelligence-based optimization with predictive control provides an effective and industrially viable solution for wedge shape control.

In order to improve the efficiency of mechanical performance testing of seamless steel pipes, enhance product quality, and reduce the cost of new product development, a machine learning-based prediction model for the mechanical properties of hot continuous rolling seamless steel pipes is proposed. Specifically, a Bayesian optimization optimized categorical boosting (CatBoost) algorithm is employed to train on production data of 723 seamless steel pipes from a 460 mm steel plant. Its predictive performance is compared with other machine learning models, including support vector regression, random forest, and extreme gradient boosting. Furthermore, the Shapley additive explanations method is used for post hoc interpretability analysis of the prediction model. Results demonstrate that the BO-CatBoost model achieves superior performance, offering higher prediction accuracy. Interpretability analysis further reveals that elongation is primarily influenced by diameter and the contents of C, V; yield strength is mainly affected by rolling temperature and the contents of MN, V; while rolling temperature, sizing temperature, and thickness are the key factors affecting tensile strength. Finally, a prediction system based on the proposed method is developed and is applied in industrial production.

Precise prediction of cold-state outer diameter (OD) is essential for dimensional quality control in hot-rolled seamless steel tube production, yet remains challenging under multi-grade and multi-specification operation where process regimes change frequently, and prediction errors can become heavy-tailed. This study proposes an interpretable ensemble learning framework for OD prediction based on a stacking architecture, in which an instance-wise dynamic pruning strategy is introduced to deactivate locally uncompetitive base learners during inference using out-of-fold error evidence. Evaluations under progressively reduced training scales demonstrate consistent generalisation gains; at a 90% training size, the pruned ensemble reduces the test root mean square error from 1.1705 mm to 0.9551 mm with a test mean absolute error of 0.7534 mm and a coefficient of determination of 0.9972, indicating effective suppression of extreme deviations. Model transparency is established via Shapley Additive Explanations and partial dependence plots, which reveal a dominant sizing-stage speed-schedule structure and stage-resolved interaction patterns that are operationally meaningful for process monitoring. Validation at plant scale conducted on 15,300 production samples encompassing nine steel grades and five OD specifications demonstrates a stable error envelope centred near zero. Across all grades, the proportion of predictions within a tolerance of +/- 4 mm consistently exceeds 98.4%.

Laser ultrasonic technology was applied to nondestructively investigate microstructural evolution and predict yield strength in heat-treated TA2 titanium. A predictive model was established by linking ultrasonic attenuation coefficients to grain size, and then relating grain size to yield strength via the Hall-Petch relationship. Timefrequency analysis extracted key spectral parameters, including attenuation coefficients and energy distribution, effectively reflecting grain coarsening and phase transformation. Results indicate a strong correlation between ultrasonic attenuation, grain size, and yield strength. Air-cooled samples exhibited decreasing peak frequencies and narrowing energy bandwidths with prolonged annealing, whereas water-quenched samples maintained higher frequencies and broader bandwidths, indicating more stable microstructures. The laser ultrasonic inspection framework was validated under different cooling conditions, confirming its feasibility for laboratory-scale yield strength prediction. This study demonstrates the potential of laser ultrasonics as a nondestructive method for predicting mechanical properties in titanium alloys, using attenuation coefficients as a bridge through grain size to yield strength. Future work may focus on extending this approach to industrial heat treatment monitoring and process optimization, highlighting its promise for smart manufacturing applications.

Equipment wear and assembly clearances can change the longitudinal stiffness of hot strip mills and further affect roll-gap levelling accuracy and asymmetric strip profile control. In this study, the longitudinal stiffness of a 1580 mm four-high hot strip finishing mill was investigated by combining the analytical calculation of the hydraulic press-down system with a three-dimensional mill-strip finite element model. The effects of typical horizontal and vertical gap forms, including work-roll offset, same-side deflection, roll crossing, and unilateral vertical clearance caused by step-pad wear, on total longitudinal stiffness and stiffness difference between the two sides were analysed systematically. The results show that work-roll horizontal offset changes the longitudinal stiffness in a nonlinear manner, whereas work-roll rotation and roll crossing generally reduce the longitudinal stiffness and increase the stiffness asymmetry between the two sides. Unilateral vertical clearance also causes nonlinear variation in both total stiffness and side-to-side stiffness difference. The proposed method was further applied to the stiffness prediction module of the Guangxi BG 1700 mm hot strip mill production line, providing support for equipment maintenance, roll-gap levelling, and stable strip production.

Abstract The geometry of the Circle of Willis poses major challenges for mechanical thrombectomy, where device navigability and effective thrombus removal determine treatment success. This study investigated the performance of venturi-inspired aspiration thrombectomy devices in a simplified cerebral artery segment representative of the middle cerebral artery (MCA), a frequent site of occlusion. Five designs (30°, 45°, 60° venturi, 7/11° taper, and cylindrical control) were assessed using a combined computational–experimental framework. On the computational side, unsteady Reynolds-averaged Navier–Stokes (URANS) simulations were performed in ANSYS Fluent 19.2 with k–ε turbulence closure. Blood–clot interactions were modeled using a Volume of Fluid (VOF) multiphase formulation with Carreau–Yasuda non-Newtonian rheology. In vitro, stereolithography-fabricated prototypes were tested with porcine thrombi in silicone arterial phantoms. CFD predicted extraction times of 2.12 s for the control and 1.64 s for the 45° venturi, with efficiency plateauing beyond 45°. Experimental results confirmed this trend, showing the 45° design as optimal and all venturi devices outperforming the control. Fragmentation analysis revealed a trade-off, with the 60° venturi producing more than twice the fragments of the 30°. These findings demonstrate that venturi taper geometry critically influences aspiration efficiency and fragmentation and establish CFD–experiment integration as a foundation for optimizing next-generation thrombectomy devices.

During hot-strip production, lateral head off-tracking at the finisher exit degrades strip quality and destabilizes line operation, forming a key bottleneck to further improvements in flatness accuracy and unmanned operation. To address the issues of response latency, tuning complexity, and limited adaptability inherent in manual control, this study proposes a deep reinforcement learning (DRL)-based method for head deviation control in finishing rolling. Targeting higher control accuracy and system stability, we first define a process-parameter framework governing off-tracking and select representative variables-strip width and thickness, interstand rolling-force asymmetry, work-roll-bending force, flow-stress coefficient, and historical deviation records. On this basis, a particle swarm optimization-support vector regression (PSO-SVR) model is employed to predict deviation, providing feedforward information to the controller. We then introduce a deep Q-network (DQN), formulated under a Markov Decision Process (MDP), to learn the control policy; closed-loop regulation is achieved by adjusting key actuator setpoints (e.g., preset gap differential). The method was deployed and tested on an industrial 2250 mm hot-strip mill under real production conditions. Across multiple steel grades, it reduced head deviation by about 30% on average, improving flatness quality and operational stability. These results provide a practical basis and engineering reference for intelligent control of strip rolling and indicate strong prospects for broad industrial adoption.

Head-end deviation at the finishing entry is strongly influenced by asymmetric roughing transfer-bar geometry, whereas direct measurements of strip posture and head-end shape are generally unavailable before biting at the first finishing stand (F1). To characterize this incoming asymmetry, the exit centerline curve of the second roughing stand (R2) is used as an upstream measurable descriptor and converted into a transfer-bar state representation. The centerline curve is classified into L-, C-, and S-type dominant camber patterns and parameterized by the signed head-end camber amplitude, camber length, global offset, and global deflection angle. These descriptors are introduced into a three-dimensional finite element model of the entry edger roll-strip-mill roll system. A representative industrial case is used for load-level validation, and the average relative error of the F1 total rolling force is 8.39

Traditional calibration methods for line-scan cameras often impose stringent requirements on the displacement accuracy and manufacturing precision of calibration targets. Line-scan cameras are increasingly used in demanding industrial applications, such as high-speed inline measurement and surface defect inspection, because of their high scanning frequency, continuous imaging capability, and high transmission efficiency. Furthermore, the research on binocular stereocalibration methodologies for line-scan cameras remains relatively limited and operationally complex. In particular, there is still a lack of in-depth theoretical research on viewing-plane coincidence adjustment, which fundamentally restricts the spatial detection accuracy of binocular line-scan systems. To address these challenges, this paper proposes a novel static calibration method for monocular line-scan cameras by introducing a viewing-plane coordinate system into a traditional imaging model. This approach uses a simply designed 2-D calibration target and eliminates the need for complex patterns or precise target translation. Building on this monocular framework, we developed a comprehensive binocular line-scan camera imaging model. A key contribution of this study is the introduction of a viewing-plane coincidence correction model that effectively mitigates the potential parallax issues in binocular systems. Subsequently, a virtual 3-D target-based calibration method for binocular line-scan cameras was proposed to facilitate highly accurate spatial detection. Experimental validation demonstrated the high precision and robustness of the proposed approach. The monocular calibration achieves a root mean square reprojection error of 0.26 pixels, while the binocular system attains a spatial detection accuracy of 0.12 mm. These results confirm the efficacy and potential of the proposed methods for high-precision machine-vision applications.

To improve the accuracy of axial wall thickness prediction and reduce reliance on manual measurements in hot-rolled steel tube production, a prediction method based on particle swarm optimisation (PSO) and a one-dimensional convolutional neural network (1D-CNN) was developed. A dataset was constructed from 131 industrial samples collected from the production line. An input variable set was established to characterise entry wall thickness, geometric and overall deformation indicators, pass schedule profile features, speed schedule profile and thermal conditions. PSO was then employed to optimise the key hyperparameters of the 1D-CNN. The proposed model was compared with several machine learning models. The results showed that the PSO-1D-CNN model achieved the best predictive performance, with a test root mean square error of 0.0281, a mean absolute error of 0.0217, and a coefficient of determination R 2 of 0.913. Further interpretability analysis using SHapley Additive exPlanations revealed that the centroid of the reduction distribution, the number of stands, the diameter-to-wall ratio, the tube temperature at the sizing exit, and the radial compression ratio were the most influential variables affecting the predictions. Finally, the proposed model was integrated into an online system. This system enables single tube wall thickness prediction, sawing parameter calculation, and batch visualisation for process adjustment and sawing decisions.

Steel sheets and strips are widely used in various industrial fields due to their excellent mechanical properties. However, harsh operating conditions, including strong vibrations, high temperatures, and rapid material movement, significantly degrade measurement precision, thereby limiting flatness evaluation accuracy and affecting product quality. To address this issue, a novel real-time vibration-suppressed flatness detection model for steel sheets and strips is developed in this study. First, a dimensionality-reduction filtering method is introduced by integrating band-pass threshold filtering and radius filtering within the framework of dimensionality reduction. Subsequently, a dynamic iterative compensation model based on a first-order Bezier curve is employed to achieve real-time suppression of vibration interference. Experimental validation, with vibration signals incorporated into the strip model, demonstrates that the proposed vibration-suppression model achieves a mean square error of only 1.680 & times; 10-4 relative to the design reference at a sampling frequency of 1 kHz, enabling adaptation to strip speeds up to 20 m/s in industrial production lines. These results confirm the model's high computational accuracy, stability, and real-time performance. In conclusion, the proposed model provides an effective and high performance solution for real-time flatness quality detection and classification in steel sheet and strip production.

In the pursuit of intelligent manufacturing goals, industrial big data technology has emerged as a key enabler in advancing the steel industry. Traditional rolling force (RF) models typically rely on data from individual cold rolling production lines, leading to lower accuracy and limited interpretability. To overcome this, an industrial data platform has been developed, offering a complete and reliable dataset to enhance the performance of RF prediction models. A data-driven machine learning framework is proposed, employing an improved sparrow search algorithm to optimise the weighting parameters of the broad learning system. The Shapley additive explanations method is further applied to elucidate the contributions of multivariate features from hot and cold rolling, thereby enhancing the interpretability of RF predictions. The performance of the proposed framework was validated on the production line of a leading steel plant, demonstrating significant advantages over existing state-of-the-art models. Furthermore, this study demonstrates and extensively elaborates on the significant impact of hot rolling parameters in enhancing the predictive accuracy of cold RF models. Industrial application validation demonstrates that the proposed framework accurately predicts the RF at the head of cold-rolled strip, enabling feedforward compensation for bending force and effectively improving flatness defects, further confirming the method's efficacy.