Development and Application of Integrated Learning System for Hot-Rolled Steel
1. Research Background
The comprehensive quality level of hot-rolled steel is an important indicator for measuring the overall technological development of the steel industry. With the support of the continuous and huge investment in scientific research and technological development by the state, China has made significant progress in the fields of mechanical property regulation, size and shape control, and surface quality optimization of hot-rolled steel. The hot rolling production technology represented by the "New Generation TMCP" has been successfully developed, improving the strength and toughness of steel. The automatic thickness and width control technology was introduced, absorbed and innovated, ensuring the dimensional accuracy of the products. The developed hot-rolling oxidation control technology has improved the surface quality of steel. However, the above-mentioned progress is the crystallization of technological advancements in the industrial age, featuring distinct independent control characteristics of each working independently. While enhancing a single performance indicator, it is bound to sacrifice other quality indicators. Therefore, how to further improve the comprehensive quality of hot-rolled steel determines whether China can further enhance the competitiveness of its products and achieve efficient production, thus remaining invincible in the global competition for basic raw materials.
The hot rolling process is a typical steel forming process with multi-field coupling. The combined effect of temperature and deformation causes a series of complex physical and metallurgical processes within the rolled piece, including solid solution of elements, precipitation of microalloying elements, recovery and recrystallization softening. These organizational evolution behaviors not only determine the internal organizational structure and mechanical properties of the product, but also determine the deformation resistance of the rolled piece and affect the changes of force-energy parameters during the rolling process. They are the core factors controlling the shape and dimensional accuracy of the product. On the other hand, during the rolling process, the rolled pieces are constantly exposed to high temperatures and air environments, and it is inevitable that severe surface high-temperature oxidation will occur. As the interfacial medium between the roll and the rolled piece, the change in the thickness of the iron oxide scale can alter the interfacial friction coefficient, thereby affecting the rolling force-energy parameters. Meanwhile, the statistical results of industrial production show that over 70% of the surface quality defects of products are caused by improper control of high-temperature oxidation behavior.
In conclusion, the microstructure evolution, surface oxidation behavior and force and energy parameter changes of hot-rolled products present the characteristics of strong coupling and black box state. Only by solving this problem can the coordinated optimization of surface quality, mechanical properties and dimensional accuracy be achieved, and the comprehensive quality of the products be improved. However, the traditional hot-rolling production control technology has been unable to solve the strongly coupled black box problem of this complex nonlinear system.
2. Technical routes and solutions
The project team, through the big data mining of the hot rolling industry, integrated the steel rolling process mechanism and physical metallurgical mechanism, and developed an integrated learning system for hot-rolled steel that comprehensively considers the evolution of microstructure, the evolution of iron oxide scale thickness, interfacial friction and the change of rolling force, as shown in Figure 1. On this basis, by applying the multi-objective optimization theory and method, the comprehensive control of the quality of hot-rolled products was achieved.
Picture
During the hot rolling process, the evolution of the internal microstructure of the rolled piece determines its macroscopic rheological stress. Rolling force, as a key parameter that can be detected in real time and precisely in industrial production, can accurately reflect the changes in microstructure. Taking rheological stress as a bridge and through machine learning of industrial big data of force and energy parameters in hot rolling, the evolution process of austenite recrystallization and grain morphology during the rolling process can be revealed. In addition, during the hot rolling production process, iron oxide scale is generated on the surface of the steel at any time, which acts as a lubricating medium at the interface between the rolls and the rolled piece, affecting their contact state and thereby influencing the changes in the rolling load of the rolled piece. The project team has cracked the strong coupling relationship among microstructure, force-energy load and friction coefficient. Based on the precise prediction of the softening behavior and friction state of the rolled piece, it can accurately predict the changes in rolling force during the hot rolling process, thereby effectively improving the control accuracy of thickness and plate shape.
After the rolling process is completed, hot-rolled steel needs to undergo an accelerated cooling process to control its phase transformation behavior. During this period, the deformed austenite undergoes continuous cooling phase transitions such as ferrite, pearlite, bainite and martensite. The main factors influencing the phase transformation behavior of rolled pieces include: the austenite microstructure state after rolling and the cooling path of rolled pieces. Their combined effect determines the phase transformation products, the proportion of each phase and the degree of grain refinement. Under the premise that the rolling process parameters remain basically unchanged, the cooling path will directly determine the phase composition of the steel and thereby its final mechanical properties. Rapid and accurate acquisition of the continuous cooling transition curve (CCT) is conducive to formulating the correct cooling path and achieving precise regulation of the performance of hot-rolled steel. For this purpose, the project team, based on the establishment of CCT databases for different steel grades and in combination with the principles of physical metallurgy, developed a hereditary machine learning modeling method for dynamic phase transformation, achieving the rapid generation of continuous cooling phase transformation curves for different steel grades.
High-strength steel undergoes complex phase transformation behavior during the cooling stage and is highly sensitive to the cooling path. The traditional modeling method that only relies on static digital data cannot fully reflect the influence of cooling path fluctuations on the microstructure and performance of the product, resulting in a significant deviation between the predicted results and the actual performance. To this end, the project team developed a dynamic deep learning model. By introducing convolutional neural networks, it not only effectively overcame the problem of feature loss in traditional data-driven machine learning models when dealing with unstructured data, but also significantly enhanced the multimodal information integration ability for complex physical phenomena, thereby being able to perceive various complex factors that affect the final microstructure and mechanical properties of steel. This modeling method can automatically learn and extract the influence law of the cooling path on the evolution of microstructure, and thereby accurately perceive the fluctuation of mechanical properties with the change of process parameters.
3. Implement the production line and its effect
3.1 Promotion and Application of 1580mm hot continuous rolling and continuous annealing production lines
For a certain 1580mm hot continuous rolling and continuous annealing production line, taking the steel grade series such as low-alloy high-strength steel and IF steel, which are widely used in large quantities, as the research objects, based on industrial big data, an integrated machine learning system was developed and the autonomous optimization of the key influencing factors of the system was achieved. For the produced Nb, NB-Ti microalloy steels and IF steels, realize the full-process temperature field calculation of microstructure evolution during rolling and continuous annealing, as well as the precise calculation of austenite recrystallization behavior, phase transformation behavior and precipitation behavior. And taking SAPH440, QstE380TM, S420MC, QstE460TM, S500MC, M3A45, and M3A21 as examples, the precise prediction of the tissue evolution process was achieved. For grades such as SAPH440, S420MC, QStE420TM, DC04, DC06, and St13, online prediction of mechanical properties is achieved. Regarding yield strength and tensile strength, over 90% of the predicted strength values of steel coils have an error of ±20MPa compared to the actual values. For elongation, more than 90% of steel coils have an error of ±3% between the predicted elongation value and the actual value.
3.2 Promotion and Application of 2250mm hot Continuous Rolling Production Line
Relying on a certain 2250mm hot continuous rolling production line, by integrating steel rolling technology, big data mining, and artificial intelligence technology, and driven by industrial big data, through physical mechanisms and knowledge learning, an integrated learning system for hot-rolled steel was developed. This system deconstructed the complex relationships between key parameters such as processes, techniques, and equipment and the organizational structure, achieving intelligent process design for hot-rolled products. Through the learning strategy of "preliminary learning - reinforcement learning - optimization learning", and by integrating industrial big data-driven and machine learning algorithms, the precise analysis of physical processes such as recrystallization, precipitation, and surface oxidation in the hot rolling process has been achieved. The online performance prediction of more than 20 steel grades including Q235B, Q420B, 600XL, and 700XL has been realized. The strength prediction accuracy reaches ±6%, and the elongation prediction accuracy is ±4%, significantly reducing the amount of mechanical property tests and enhancing the market response capability. In addition, the dynamic soft measurement of the hot-rolled oxidation behavior was achieved. The prediction accuracy of the oxide scale thickness and structure of the product reached ±2μm and ±10% respectively. On this basis, the flexible control technology of the iron oxide scale structure was developed, and the development of new varieties of 700MPa grade acid-free steel and easy-pickling steel series represented by SPHC was achieved.
3.3 Promotion and application of 5500mm wide and thick plate production lines
Facing the changes in production and consumption structure brought about by the new industrialization, the characteristics of the wide and thick plate production process, such as complex variety structure and numerous small-batch orders, have become more prominent. The large amount of leftover billets generated has caused huge economic losses to enterprises. Moreover, the excessive steel grades have complicated the steelmaking process, seriously affecting the continuous improvement of production efficiency and product quality. In response to the above-mentioned challenges, the project team, relying on the 5500mm wide and thick plate production line, developed an integrated learning system for hot-rolled steel for typical wide and thick plate production lines. It can conduct self-learning of the model based on the actual production status, process and environmental changes, and achieve automatic optimization of model parameters. The high-precision prediction of the mechanical properties of multiple grades of products, including A, AH32, AH36, DH36, and Q355MD, has been successfully achieved. On this basis, a flexible design object library for the residual billet process was established. The evaluation function for the optimal organizational structure and performance indicators of the residual billet production was proposed. An intelligent matching optimization algorithm was developed, and a predictive model of "composition - process - structure - performance" of the residual billet was established. Based on the comprehensive consideration of the combined effects of strengthening mechanisms such as fine grains, precipitation, dislocations and phase transformation, a flexible design method for the rolling process was proposed to achieve the flexible design of cross-thickness and cross-strength processes for three major series of steel grades, namely C-Mn, pipeline and low alloy.
3.4 Promotion and application of 5000mm wide and thick plate production lines
In recent years, steel for offshore wind power has emerged as a strong force, with a sharp increase in demand. According to research by authoritative institutions, each megawatt of offshore wind power uses approximately 200 tons of medium and heavy plates. During the "14th Five-Year Plan" period, the newly installed capacity of offshore wind power will exceed 44GW, and it is estimated that the demand for medium and heavy plates will be no less than 8.8 million tons. Among them, there is a huge demand for steel plates with large thickness (60-150mm) for offshore wind turbine towers and pipe piles. In terms of composition design, conventional extra-thick plates generally adopt high C, high Mn+Cr, Ni, Cu and other precious metals, as well as a small amount of Nb, V, Ti grain refinement elements. The mechanical properties of extra-thick steel plates are mainly guaranteed through solid solution strengthening and grain refinement of alloying elements. The original design not only has a high cost, but also due to the severe central segregation of alloying elements in the extra-thick continuous casting slab, the performance of the core of the extra-thick plate has decreased significantly, especially the low-temperature impact toughness of the core is difficult to meet the performance requirements of the extra-thick plate. The project team, in response to the 355-460 mpa grade wind power steel, combined with the characteristics of the 5000mm wide and thick plate production line, adopted the ultra-low C and N composition route to redesign the composition system of the extra-thick steel plate, increasing the niobium content in the steel, reducing the content of carbon, manganese and other precious metal alloys, and improving the central segregation of the extra-thick plate billet. On this basis, combined with the integrated machine learning system for hot-rolled steel, by calculating the temperature and microstructure distribution in the thickness gradient direction of thick plates, the rolling process design of 60-150mm thick wind power steel is realized. The results of batch production show that the qualification rate of mechanical properties of steel plates with a strength grade of 355MPa is 100%, and the qualification rate of one-time impact performance at the core reaches more than 98%.
4 Conclusion
The project team comprehensively utilized the principles of machine learning, deep learning and physical metallurgy, and integrated data with multi-modal structures to innovatively propose an integrated machine learning system for steel, which has been successfully applied to multiple production lines in China. This system realizes the high-precision prediction of the microstructure evolution, surface oxidation behavior and rolling force of the rolled piece. Meanwhile, it also provides strong support for the optimization of process parameters, the development of new steel grades and new processes, thereby enhancing the overall production efficiency.
15358968703