Microstructure prediction method based on technological parameter optimization of copper rod continuous casting

Abstract

The invention relates to the technical field of five-wheel continuous casting low-oxygen copper, in particular to a copper rod continuous casting process parameter optimization-based microstructure prediction method, which comprises the following steps of: formulating an adjusting factor set based on a comparison result of a crystallization line position of a copper casting blank and a standard crystallization line position; obtaining a microstructure image and inputting the microstructure image into the feature comparison model to obtain an anomaly set; inputting the production elements and the abnormal set into a parameter optimization model, and screening out a plurality of adjustment schemes; setting a reference factor to perform serialized adjustment on the adjustment scheme, and casting a copper casting blank according to the previous adjustment scheme; various passive data of the cast wheel are obtained, and a real-time matrix is obtained; the production elements and the real-time matrix are input into a microscopic image prediction model, and microstructure prediction is carried out; through construction of a feature comparison model and a parameter optimization model and application of a real-time matrix and a microscopic image prediction model, the automation level of the production process is improved, and a scientific basis is provided for standardization of process parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of five-wheel continuous casting of low-oxygen copper, and particularly to a method for predicting the microstructure by optimizing process parameters based on copper rod continuous casting. Background Art

[0002] Copper is widely used in the industrial field. As a key process in copper processing, copper rod continuous casting converts molten copper into copper rods with specific shapes and properties through a five-wheel continuous casting method. During this process, the molten copper undergoes a series of physical changes such as solidification and crystallization in the mold. Process parameters such as temperature, drawing speed, and cooling intensity play a decisive role in the microstructure and final properties of the copper rod. The microstructure is directly related to important characteristics such as the strength, electrical conductivity, and corrosion resistance of the copper rod. Therefore, achieving precise process parameter optimization and microstructure prediction is of great significance for improving the quality of copper rods.

[0003] In the prior art, copper casting means have been continuously developed. In the early traditional process, the adjustment of process parameters mainly relied on manual experience. Operators adjusted parameters such as temperature and speed based on their long-term accumulated experience. With the progress of technology, some enterprises introduced simple sensors to monitor basic parameters such as temperature and speed, which enabled basic data recording during the production process and provided a basis for preliminary parameter regulation. At the same time, some advanced technologies have emerged continuously, such as simulating and predicting the microstructure by building models, providing a reference direction for process optimization at the theoretical level.

[0004] However, these prior arts have many drawbacks. The adjustment based on manual experience is highly subjective and lacks systematicness. The judgments of different operators vary greatly, making it difficult to ensure the consistency and stability of product quality, and being unable to cope with complex and changing production situations. Although simple sensors can obtain basic parameters, they cannot accurately monitor and feedback the key microstructure characteristics that affect the quality of copper rods, resulting in a lack of pertinence in process parameter optimization. The existing microstructure prediction models, due to their failure to closely combine with the real-time state of copper cast billets, such as the position of the crystallization line and actual abnormal situations of the microstructure, make the prediction results deviate from actual production and unable to provide effective support for precise optimization of process parameters, ultimately restricting the improvement of copper rod quality and production efficiency.

[0005] The information disclosed in this background art section is only intended to deepen the understanding of the overall background art of the present invention, and should not be regarded as an admission or any form of suggestion that this information constitutes the prior art known to those skilled in the art. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a method for predicting the microstructure by optimizing the process parameters based on copper rod continuous casting. Through the construction of a feature comparison model and a parameter optimization model, as well as the application of a real-time matrix and a microstructure image prediction model, not only the automation level of the production process is improved, but also a scientific basis is provided for the standardization of process parameters.

[0007] The method for predicting the microstructure by optimizing the process parameters based on copper rod continuous casting of the present invention includes: Based on the comparison result between the crystallization line position of the copper casting blank to be optimized and the standard crystallization line position, determine the influencing factors and formulate a set of adjustment factors; Obtain the microstructure image of the copper casting blank to be optimized and input it into a pre-constructed feature comparison model to obtain an abnormal set; Input the production factors corresponding to the copper casting blank to be optimized and the abnormal set into a pre-constructed parameter optimization model, output multiple preset schemes, and screen out multiple adjustment schemes including adjustment factors; Set the reference factors to serially adjust the adjustment scheme, and cast the copper casting blank according to the front adjustment scheme; Obtain various passive data of the casting wheel during casting to obtain a real-time matrix; Input the production factors obtained in real time and the real-time matrix into a pre-trained microstructure image prediction model for microstructure prediction.

[0008] As a preferred scheme of the present invention, The method for comparing the crystallization line position of the copper casting blank to be optimized with the standard crystallization line position includes: Establish a standard crystallization line morphology database for different copper species; Use the digital image correlation method to correct the distortion of the cross-sectional image of the casting blank to be optimized, and convert the actual crystallization line coordinates to the standard reference system; Calculate the three-dimensional parameter deviation of the crystallization line. As a preferred scheme of the present invention, the influencing factors include at least one of casting temperature, casting speed, carbon coating thickness, blockage of the cooling water nozzle, surface crack of the casting wheel and its turning repair, and scale formation of the casting wheel.

[0009] As a preferred scheme of the present invention, the construction method of the feature comparison model includes: Collect copper casting blank microstructure image samples, perform image preprocessing and zoning processing, and set area labels; Construct a feature comparison model using a regional branch structure, and introduce an attention mechanism into the feature comparison model; Divide the samples into a training set and a test set, use the area label as an auxiliary input, and train the feature comparison model with the training set; The cross-entropy loss function is used to measure the difference between the model prediction results and the true annotations, and the structural parameters of the model are adjusted; The test set is imported into the feature comparison model to evaluate the performance of the feature comparison model, and the feature comparison model is optimized according to the evaluation results.

[0010] As a preferred embodiment of the present invention, the zoning treatment includes: dividing the four sides of the copper casting blank from the outside to the inside into a fine equiaxed crystal zone, a columnar crystal zone, and a coarse equiaxed crystal zone in sequence, and the four sides share a common coarse equiaxed crystal zone.

[0011] As a preferred embodiment of the present invention, the method for constructing a real-time matrix includes: Defining the matrix dimension structure, dividing the passive data according to a preset time span, and calculating the data mean within each time span. The data means are arranged in chronological order to obtain the real-time matrix.

[0012] As a preferred embodiment of the present invention, the method for serial adjustment includes: Determining the weight of each reference factor, scoring each adjustment plan on the reference factors, calculating the comprehensive score of each adjustment plan according to the weight and the score, and sorting the adjustment plans from high to low according to the comprehensive score.

[0013] As a preferred embodiment of the present invention, the reference factors include at least one of the implementation cost of the adjustment plan, the confidence level of the expected grain refinement effect, the collaborative score of equipment parameter adjustment, and the success rate of historical similar cases.

[0014] As a preferred embodiment of the present invention, before inputting into the microscopic image prediction model, the following processing is performed: Using the principal component analysis method, performing dimensionality reduction processing on the production factor data and the real-time matrix data respectively, and extracting their respective principal components; Splicing the principal components to form a new feature vector as the input of the microscopic image prediction model.

[0015] As a preferred embodiment of the present invention, the passive data includes at least one of the casting wheel temperature, the casting wheel vibration frequency, and the casting wheel surface roughness.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1) By obtaining the crystallization line position and microstructure image of the copper casting blank, and combining the pre-constructed feature comparison model and parameter optimization model, the key factors affecting the quality of the copper rod can be systematically analyzed, and a targeted adjustment plan can be output; compared with the traditional technology that relies on manual experience or simple sensors, the present invention realizes the transformation from subjective judgment to data-driven, significantly improving the accuracy and efficiency of process parameter optimization; by setting reference factors to serially adjust the adjustment plan, the feasibility and optimality of the optimization plan are ensured, thus effectively solving the problem of lack of systematicness and consistency in parameter adjustment in the prior art; 2) By obtaining the passive data of the casting wheel in real time and constructing a real-time matrix, and combining the pre-trained microstructure prediction model, the microstructure of the copper casting blank can be predicted in real time. This function makes up for the deficiency of microstructure monitoring and feedback ability in the prior art, enabling the timely discovery and correction of abnormalities during the production process and avoiding the generation of quality defects; at the same time, by using the principal component analysis method to perform dimensionality reduction processing on the data, the calculation efficiency and prediction accuracy of the model are further improved, providing strong technical support for the quality control of the copper rod; 3) The construction of the feature comparison model and parameter optimization model, as well as the application of the real-time matrix and microstructure prediction model, not only improve the automation level of the production process, but also provide a scientific basis for the standardization of process parameters; the reference factors mentioned in the present invention comprehensively consider the economy and feasibility in actual production, making the optimization plan closer to the production reality and helping the enterprise to achieve high-quality and high-efficiency production of copper rods. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow schematic diagram of the present invention; Figure 2 is a structural schematic diagram of the copper casting blank and the standard crystallization line; Figure 3 is a structural schematic diagram of the copper casting blank, the fine equiaxed crystal zone, the columnar crystal zone and the coarse equiaxed crystal zone; Reference numerals in the drawings: 1, copper casting blank; 2, standard crystallization line; 3, fine equiaxed crystal zone; 4, columnar crystal zone; 5, coarse equiaxed crystal zone. DETAILED DESCRIPTION OF THE INVENTION

[0018] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made in conjunction with the drawings of the specification.

[0019] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention, but the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0020] Secondly, the "embodiment" referred to herein means a specific feature, structure or characteristic that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or alternative embodiment that is mutually exclusive with other embodiments. Embodiment

[0021] Reference Figure 1 , this embodiment provides a method for predicting the microstructure based on the optimization of process parameters for copper rod continuous casting, including: S1 Based on the comparison result between the crystallization line position of the copper casting blank to be optimized and the standard crystallization line position, determine the influencing factors and formulate a set of adjustment factors; S2 Obtain the microstructure image of the copper casting blank to be optimized and input it into a pre-constructed feature comparison model to obtain an abnormal set; S3 Input the production factors corresponding to the copper casting blank to be optimized and the abnormal set into a pre-constructed parameter optimization model, output multiple preset schemes, and screen out multiple adjustment schemes that include adjustment factors; S4 Set reference factors to serially adjust the adjustment schemes, and cast copper casting blanks according to the earlier adjustment schemes; S5 Obtain various passive data of the casting wheel during casting to obtain a real-time matrix; S6 Input the production factors obtained in real time and the real-time matrix into a pre-trained microstructure prediction model for microstructure prediction; Through the above steps, first, by comparing the crystallization line position of the copper casting blank to be optimized with the standard crystallization line position, the influencing factors can be quickly locked, providing a clear direction for subsequent operations and effectively correcting the randomness and blindness of manual experience adjustment; at the same time, the collection of microstructure images and the application of the feature comparison model can accurately identify abnormalities at the microscopic level, making up for the shortcoming that simple sensors cannot reach key microscopic features, echoing the analysis of the crystallization line position, achieving all-round quality insight from macro to micro, and greatly improving the accuracy of judging the root cause of quality problems; the multiple preset schemes output by the parameter optimization model and the screening of multiple adjustment schemes that include adjustment factors not only target macro production parameters but also take into account microscopic quality hazards, greatly enhancing the pertinence and practicality of the schemes, avoiding the blindness of process parameter optimization, and effectively ensuring the quality and production efficiency of copper rods; by setting reference factors to serially adjust the schemes, comprehensively weighing multiple factors such as cost, quality improvement effect, and equipment adaptability, it changes the one-sided situation of single-dimensional consideration in the past. At the same time, the construction of the real-time matrix provides dynamic and vivid data for microstructure prediction, making the prediction results closely connected with actual production and jointly building a real-time feedback mechanism with other links.

[0022] In some embodiments of the present invention, a method for comparing the crystallization line position of a copper casting blank to be optimized with the standard crystallization line position includes: Establish a standard crystallization line morphology database for different copper grades; More specifically, referring to Figure 2 , the drawing method of the standard crystallization line includes: The cross-sectional shape of the groove of the casting wheel used in the five-wheel continuous casting is an isosceles trapezoid. After the steel belt wraps the casting wheel, it forms an arc-shaped mold together with the groove. Since the cross-sectional shape of the mold is directly determined by the groove and the wrapping method of the steel belt, when the casting blank solidifies in the mold, its cross-sectional shape will exactly replicate the trapezoidal contour of the groove; Draw straight lines from the two corner points of the top edge of the copper casting blank to the center part of the bottom edge respectively, and draw straight lines from the two corner points of the bottom edge of the copper casting blank to the center part of the top edge respectively. Two of the straight lines starting from the two ends of the same side (i.e., the hypotenuse of the trapezoidal contour) form an intersection point, and then connect the two intersection points to form a connecting line. The four straight lines and the connecting line form the standard crystallization line, and the formed standard crystallization line divides the copper casting blank into four regions. The upper and lower two crystal regions are close to the steel belt and the center of the casting wheel, and the left and right two crystal regions are close to the side of the casting wheel; More specifically, the construction method of the standard crystallization line morphology database includes: for different copper grades, using a synchrotron X-ray device to take cross-sectional images of qualified casting blanks and perform the following processing: removing speckle noise and enhancing the crystallization line edge; based on the crystallization line morphology of qualified samples, using non-uniform rational B-splines (NURBS) to fit the geometric contour of the standard crystallization line and construct the standard crystallization line morphology database: Use the digital image correlation method to correct the distortion of the cross-sectional image of the casting blank to be optimized, and convert the actual crystallization line coordinates to the standard reference system; more specifically, mark the reference point set corresponding to the standard image in the cross-sectional image of the casting blank to be optimized, and use a second-order polynomial transformation model (including translation, rotation, scaling, and shear components) to eliminate the microstructure distortion caused by casting stress, and map the position coordinates of each node of the actual crystallization line to the UVW coordinate system of the standard reference system; Calculate the three-dimensional parameter deviation of the crystallization line; more specifically, the three-dimensional parameter deviation includes: Position deviation: The Euclidean distance ΔD of the center point of the crystallization line = |(x, y) 实测 - (x, y) 标准 |; Morphology deviation: The difference in the radius of curvature along the casting direction ΔR = (R 实测 - R 标准 ) / R 标准 ; By establishing a standard crystallization line database associated with copper species characteristics and high-precision digital image correction technology, and combining with the dynamic analysis of three-dimensional parameter deviations, multi-dimensional collaborative diagnosis of casting defects is achieved. This not only solves the misjudgment problem caused by the isolated comparison of the crystallization line position in traditional methods, but also through the cross-verification of morphology, angle and position deviations, latent process fluctuations and equipment wear effects are identified in advance. The above method makes the mapping relationship between process parameters and defect causes clearer, significantly improving the accuracy of abnormal traceability of billet quality and the dynamic adaptability of process control; Comparing the crystallization line of the copper billet to be optimized with the standard crystallization line can intuitively and accurately reveal the quality deviation of the copper billet during the casting process. Because under ideal conditions, the crystal regions on the left and right sides of the billet should have the same shape and size, and the crystallization lines of the upper and lower crystal regions of the billet should be in the central position. Through this comparison, once it is found that the crystallization line deviates from the standard position, such as the crystallization lines of the upper and lower crystal regions of the billet are not in the central position, or the shapes and sizes of the crystal regions on the left and right sides are different, it can be quickly judged that there are abnormalities in the casting process; this helps to quickly locate influencing factors such as uneven pouring temperature and unstable operation of the casting wheel, providing a key basis for formulating targeted adjustment measures in the follow-up, thereby effectively ensuring the quality of the copper billet and significantly improving production efficiency and product qualification rate.

[0023] In some embodiments of the present invention, the influencing factors include at least one of pouring temperature, pouring speed, carbon coating thickness, blockage of the cooling water nozzle, surface cracks of the casting wheel, and turning and scaling of the casting wheel; The above influencing factors affect the position of the crystallization line, and the principle is as follows: The pouring temperature has a significant impact on the crystallization driving force and the nucleation and growth process. The higher the temperature, the greater the driving force during crystallization, the faster the grain growth rate, which is beneficial to the formation and growth of columnar crystals. Excessive temperature will remelt the free crystals in the copper liquid, reduce the nucleation particles, and the number of equiaxed crystals will decrease significantly, providing more growth space for columnar crystals, resulting in an increase in the average size of columnar crystals, and the position of the crystallization line may change due to the excessive growth of columnar crystals, such as shifting upward; while the pouring temperature is too low, the crystallization driving force is insufficient, the grain growth is slow, it is not easy to form columnar crystals, the number of nucleation in the copper liquid increases, the number of equiaxed crystals increases, and the formed crystal size is smaller, and the position of the crystallization line may be relatively close to the center or show irregular changes; When the casting speed is fast, the copper liquid flows fast in the five-wheel continuous casting system, and the contact time with the casting wheel and the cooling medium is short, so the heat cannot be fully dissipated. During the rotation of the casting wheel, the solidification process of the copper liquid between the wheels is inconsistent, resulting in the position of the crystallization line shifting in the forward direction of the billet due to the delayed solidification. At the same time, the shape of the crystallization line may be irregular due to the flow scouring effect; when the casting speed is slow, the copper liquid stays in the continuous casting system for a long time, and has sufficient time to dissipate heat and solidify. In the five-wheel continuous casting, the cooling effect of each wheel can be fully exerted, the position of the crystallization line will be relatively forward, closer to the gate position, and because the solidification process is relatively stable, the shape of the crystallization line is relatively regular; The carbon coating plays a role in isolating and regulating heat transfer in five-wheel continuous casting; if the carbon coating is too thick, it will weaken the cooling effect of the casting wheel on the copper liquid, delay the solidification of the copper liquid, and during the continuous casting process, the cooling unevenness at each wheel will be more obvious, causing the position of the crystallization line to shift to the area with thick carbon coating and poor cooling effect. At the same time, abnormal crystallization may occur due to local overheating, making the shape of the crystallization line irregular; if the carbon coating is too thin, the heat exchange between the casting wheel and the copper liquid is too strong, the copper liquid solidifies quickly at the contact point with the casting wheel, and the position of the crystallization line will be close to the casting wheel surface. In addition, the lubrication effect of the thin carbon coating is insufficient, which may cause the copper liquid to flow poorly during the continuous casting process, affecting the uniformity and position stability of the crystallization line; In the five-wheel continuous casting, the cooling water nozzles are distributed near each casting wheel, which plays a key role in cooling the copper liquid. When the nozzles are blocked, the casting wheel in the corresponding area is insufficiently cooled, and the copper liquid solidifies slowly in this area. Since the synergistic effect of cooling of each wheel is destroyed, the crystal line will shift to the area near the insufficiently cooled wheel, and obvious asymmetry and distortion will appear, which seriously affects the quality of the ingot and the normal shape of the crystal line. The cracks on the casting wheel surface will change the contact state between the molten copper and the casting wheel during the continuous casting process. Local heat sinks or thermal resistance will be formed at the cracks, resulting in abnormal solidification behavior of the molten copper near the cracks. The crystal line may bend, fork or interrupt at the cracks, and the overall position will also shift due to the difference in heat dissipation and solidification caused by the cracks. If the wheel is not repaired in time or the repair quality is poor, the surface accuracy of the wheel will decrease and the matching accuracy between the wheels will also be affected. During the continuous casting process, the friction and cooling conditions of the copper liquid are uneven, resulting in unstable crystal line position, which may cause periodic or random deviation, and affect the surface quality and internal structure of the ingot. Casting wheel scaling will form a heat-insulating and barrier layer on the casting wheel surface. In the five-wheel continuous casting, the cooling rate of the copper liquid in the scaling area is significantly slowed down, and the solidification time is prolonged. As the casting wheel rotates, the solidification difference between the scaling area and the normal area continues to accumulate, and the crystallization line will shift to the area with severe scaling. In addition, due to the unevenness of scaling, the crystallization line will have irregular bending and deformation, affecting the overall quality of the ingot and the uniformity of crystallization. In addition to the above influencing factors, the following factors will also affect the position of the crystallization line: the efficiency of the cooling system, the humidity in the workshop, the stability of the air flow, etc.

[0024] In some embodiments of the present invention, the method for constructing a feature comparison model includes: Collecting microscopic structure image samples of copper billets, performing image preprocessing and zoning processing, and setting area labels; In this embodiment, a large number of microscopic structure image samples of copper billets under different process conditions are collected through various channels, covering samples under different casting temperatures, speeds, casting wheel states, etc. Professional image preprocessing software can be used to denoise these sample images, removing the noise generated during image acquisition and improving image clarity; at the same time, enhancement processing is performed, such as enhancing the contrast of the image through methods such as histogram equalization; since the tissue characteristics of different regions may be affected by different process factors, the abnormal manifestations are also different, so the samples can be zoned, and clear area labels are set for each image area.

[0025] Constructing a feature comparison model using a regional branch structure and introducing an attention mechanism into the feature comparison model; In the above steps, since the image samples are zoned, a neural network model with a unified structure can be set as the feature comparison model, which has multiple regional feature processing branches inside, used for differential analysis of the microscopic tissue characteristics of different regions in the image. In order to achieve more accurate microscopic tissue abnormality detection, the model further introduces an attention mechanism, through spatial attention or channel attention methods, to automatically enhance the model's perception ability of regional key features.

[0026] Dividing the samples into a training set and a test set, using the area label as an auxiliary input, and training the feature comparison model using the training set; Specifically, in this step, the preprocessed and zoned samples can be divided into a training set and a test set according to a certain ratio, such as 70% and 30%; in order to reduce the training complexity and guide the feature comparison model to efficiently identify the microscopic tissue characteristics of different regions, the area label can be used to guide the model training; area labels are set for each image area in the image preprocessing stage, and during the training process, the area label is used as an auxiliary input or supervision information to guide the model to learn the feature expression methods of the corresponding regions in each branch structure, improving the convergence speed of model training and the accuracy of regional feature recognition.

[0027] Using a cross-entropy loss function to measure the difference between the model prediction result and the true annotation, and adjusting the structural parameters of the model; More specifically, during the model training process, the cross-entropy loss function calculates in real time the degree of difference between the microstructural features predicted by the model and the true microstructural features manually annotated in advance. Based on this difference value, the model automatically adjusts its structural parameters, such as increasing or decreasing the number of neural network layers, changing the connection weights of nodes, etc., so that the model prediction continuously approaches the real situation and improves the recognition accuracy of the microstructural features; the cross-entropy loss function provides a quantitative basis for model optimization, enabling the model to have a clear direction for adjustment. By continuously adjusting the structural parameters, the model can better adapt to the microstructural features and improve the prediction accuracy.

[0028] Import the test set into the feature comparison model, evaluate the performance of the feature comparison model, and optimize the feature comparison model according to the evaluation results.

[0029] More specifically, input the previously divided test set into the trained feature comparison model. The model performs microstructural feature recognition and prediction on the test set samples. By calculating performance indicators such as accuracy and recall, evaluate the performance of the model. If the performance indicators do not meet the expectations, analyze which microstructural feature recognitions the model has biases in, and accordingly further adjust the structural parameters of the model, or increase the number of training set samples and train again until the model performance reaches an ideal state.

[0030] Widely collect copper casting blank microstructural image samples covering diverse process conditions through multiple channels, and perform preprocessing such as denoising and enhancement and partitioning, providing a rich and high-quality data basis for model training. Reasonably divide the samples into a training set and a test set, use a deep learning framework to build a model and optimize the structural parameters guided by the cross-entropy loss function, enabling the model to effectively learn the microstructural features and improve the recognition accuracy. Finally, use the test set to evaluate and continuously optimize the model to ensure that its performance reaches an ideal state, thus providing a very reliable tool for accurately identifying microstructural anomalies, assisting in process parameter optimization, and ensuring the quality of copper rods.

[0031] In some embodiments of the present invention, referring to Figure 3 , the partitioning process includes: sequentially dividing the four sides of the copper casting blank from the outside to the inside into a fine equiaxed crystal zone, a columnar crystal zone, and a coarse equiaxed crystal zone. The four sides share a common coarse equiaxed crystal zone; During the copper continuous casting process, the molten copper is rapidly cooled by the casting wheel and the steel belt. At the same time, the inner walls of the steel belt and the casting wheel serve as heterogeneous nucleation substrates, promoting a large number of nuclei to form at the edges of the billet. Due to the extremely fast cooling rate of the surface grains of the billet and the lack of sufficient time for growth, a thin equiaxed grain zone is formed at the edges. Along with the release of the latent heat of solidification, the temperature rises near the equiaxed grain zone at the edges, gradually meeting the conditions for columnar crystal growth. Some fine grains begin to grow and transform into columnar crystals. Since the growth rate of the columnar crystals perpendicular to the edge direction is faster, it will inhibit the continued growth of grains in other directions, thereby forming a columnar crystal morphology with a specific orientation. As the solidified layer advances inward, the heat dissipation capacity of the solid phase weakens, and the constitutional supercooling degree in front of the interface increases, meeting the conditions for heterogeneous nucleation. At this time, solute atoms in the liquid phase continuously enrich, and broken dendrites, detached edge grains, copper liquid impurities, etc. become nucleation particles. The central copper liquid begins to nucleate and develop into a coarse equiaxed grain zone. More specifically, a professional metallographic image analysis software, such as MetaMorph, is used. The software is opened and the microscopic structure image of the copper billet is imported. Based on the knowledge of microscopic structure, starting from the edge of the image, specific pixel brightness ranges, contrast thresholds, and crystal morphology feature rules are set to identify the outermost fine equiaxed grain zone. Then, moving inward, by analyzing features such as the long axis direction and packing density of the crystals, the columnar crystal zone is delimited. For the central part of the four side images, based on common coarse equiaxed grain features, such as larger grain sizes and relatively irregular shapes, they are jointly delimited as the coarse equiaxed grain zone to complete the zoning operation. The zoning process orderly divides the complex microscopic structure image. Different crystal zones correspond to different influences of the casting process. This process closely cooperates with the previous sample collection, preprocessing, and subsequent model training, enabling the feature comparison model to accurately learn the microscopic structure features according to the characteristics of each region, greatly improving the model's ability to identify microscopic structure abnormalities. Overall, it provides a detailed and accurate basis for optimizing process parameters from the microscopic level, solves the problem in the prior art that it is difficult to deeply analyze the relationship between microscopic structure and process parameters, and effectively improves the refinement level of copper rod quality control.

[0032] In some embodiments of the present invention, the method for constructing a real-time matrix includes: Defining the matrix dimension structure, dividing the passive data according to a preset time span, and calculating the data mean within each time span. The data means are arranged in chronological order to obtain the real-time matrix. The passive data of the casting wheel refers to the data generated and collected by the relevant states of the casting wheel during the continuous casting of copper rods, which can reflect the real-time operating conditions of the casting wheel and will change with factors such as casting temperature and casting speed. During the process of passive data collection of the casting wheel, a dedicated data segmentation module and calculation module can be developed; a time span is preset in advance, for example, every 10 seconds; the data segmentation module segments the continuously collected passive data of the casting wheel according to the set 10 - second time span, and the calculation module uses the mean - value calculation algorithm to calculate the mean value of each segment of data within this time span; subsequently, the calculated data mean values are arranged in chronological order to finally form a real - time matrix, which is stored in the cache directly connected to the microstructure prediction model for the model to call at any time; The real - time matrix orderly integrates the dynamic operation data of the casting wheel and intuitively presents the change of the casting wheel state over time. This process closely cooperates with the previous monitoring of the influencing factors of the casting wheel state and the subsequent microstructure prediction, provides dynamic and crucial real - time data support for microstructure prediction, and cooperates with other steps to make up for the problem of the disconnection between the microstructure prediction model and the actual production state in the existing technology, greatly improving the accuracy and timeliness of prediction, providing a strong basis for timely and accurate adjustment of process parameters, and helping to improve the stability of the copper rod continuous casting process and product quality.

[0033] In some embodiments of the present invention, the method of serial adjustment includes: Determine the weight of each reference factor. When determining the weight, the following method is adopted: organize a weight evaluation team composed of senior process engineers and production management experts, and use a combination of brainstorming and expert experience method. For different reference factors, such as the potential impact of the adjustment plan on product quality improvement, the expected effect on production efficiency improvement, and the degree of change in equipment maintenance difficulty, etc., conduct multiple rounds of discussions. Experts, relying on their rich industry experience and in - depth understanding of the production process, assign corresponding weights to each reference factor; for example, if it is considered that product quality improvement is the most critical, its weight can be set to a relatively high value; Score each adjustment plan on the reference factors. More specifically, form a scoring group composed of professionals in multiple fields, including experts in quality control, production planning, equipment operation and maintenance, etc.; for each adjustment plan, the group members jointly discuss and formulate a unified and detailed scoring standard; for example, set a scoring range from 1 - 10 points, and score within this range according to the degree of influence of the adjustment plan on each reference factor; those with a high degree of influence are given 8 - 10 points, those with a medium degree are given 4 - 7 points, and those with a low degree of influence are given 1 - 3 points; when scoring, each expert independently scores the adjustment plan from different reference factor dimensions based on their own professional knowledge and experience; Calculate the comprehensive score of each adjustment plan according to the weight and score, and sort the adjustment plans from high to low according to the comprehensive score; The calculation formula for the comprehensive score is: Comprehensive score = ∑(weights of each reference factor × scores of corresponding reference factors); Through the above serialization adjustment method, multiple reference factors are comprehensively considered to sort the adjustment plans, which closely cooperates with the previous steps of determining the adjustment plans, enabling the preferentially adopted plan to achieve a balance in multiple aspects such as cost, quality improvement effect, and equipment adaptability; it avoids the one-sidedness of single-factor decision-making, solves the problem in the prior art that process optimization cannot take into account multiple factors, provides a scientific basis for selecting the optimal adjustment plan, and helps to achieve efficient and economical production.

[0034] In some embodiments of the present invention, the reference factors include at least one of the implementation cost of the adjustment plan, the confidence level of the expected grain refinement effect, the synergy score of equipment parameter adjustment, and the success rate of historical similar cases; More specifically, the implementation cost of the adjustment plan is obtained as follows: form a professional cost accounting team, list in detail the various expenses involved in the adjustment plan, accurately calculate the procurement, installation and commissioning costs of new equipment, and the removal and disposal costs of old equipment for equipment transformation. For raw materials, analyze the cost differences brought about by fluctuations in procurement prices and changes in consumption rates after parameter adjustment. For example, when adjusting the casting temperature, it is necessary to consider the increase or decrease in the loss of molten copper due to temperature change, and comprehensively obtain the implementation cost; The confidence level of the expected grain refinement effect is obtained as follows: Process experts collect data on similar process adjustment cases within the enterprise and the industry, and combine the copper rod continuous casting theory knowledge and simulation analysis software to evaluate the expected grain refinement effect of the adjustment plan. The considered factors include the impact of parameter changes such as temperature and speed on the crystallization process, judge the possibility of the plan achieving the ideal grain refinement effect, and give a confidence level evaluation; The synergy score of equipment parameter adjustment is obtained as follows: Organize an equipment engineer team, and based on the equipment operation principle and past maintenance experience, discuss different parameter adjustment combinations, and evaluate the synergy impact of the adjustment plan on the overall operation of the equipment from aspects such as equipment operation stability, energy consumption, and maintenance difficulty, and quantify the scoring; for example, for the combination of adjusting the casting speed and the casting wheel rotation speed, observe the changes in equipment vibration and motor load, and give a synergy score; The success rate of historical similar cases is obtained as follows: With the help of the enterprise production management database, screen out historical cases that are similar to the current adjustment plan in terms of process parameters, equipment conditions, etc.; count the number of successful cases, calculate its proportion in the total cases, and obtain the success rate of historical similar cases; In addition to the above factors, the following can also be used as reference factors for evaluating the adjustment plan: the impact on the production environment, the convenience of process adjustment, the adaptability to subsequent processing of products, etc.

[0035] In some embodiments of the present invention, before inputting into the microscopic image prediction model, the following processing is performed: Using the principal component analysis method, dimensionality reduction processing is respectively carried out on the production factor data and the real-time matrix data to extract their respective principal components; more specifically, professional data analysis software such as the Scikit-learn library in Python is used; First, various production factor data, such as: casting temperature, casting speed, carbon coating thickness, cooling water flow rate, casting wheel rotation speed, copper liquid purity, etc. are organized into a standard data set format; Assuming that the initial production factor data contains 10 feature dimensions, using the principal component analysis (PCA) module in the Scikit-learn library, through analysis and empirical judgment, 3 principal components are set to be retained because these 3 principal components can explain approximately 85% of the data variance and represent most of the key information; For example, the first principal component comprehensively reflects the co-variation relationship between the casting temperature and the casting speed. When the casting temperature increases, in order to ensure the quality of the copper rod, the casting speed often needs to be adjusted accordingly, and this principal component reflects this linkage effect between the two; The second principal component mainly reflects the correlation between the carbon coating thickness and the cooling water flow rate. The carbon coating thickness affects the heat transfer between the casting wheel and the copper liquid, and the cooling water flow rate determines the rate of heat removal, and the two jointly act on the solidification process of the copper liquid; The third principal component covers a combination of some secondary but still influential factors such as the casting wheel rotation speed and the copper liquid purity. The casting wheel rotation speed affects the forming speed of the copper rod, and the copper liquid purity affects its fluidity and solidification characteristics, and they jointly reflect the comprehensive effect on the production process in this principal component; Through PCA processing, these 3 key principal components are extracted from the originally complex 10-dimensional data, denoted as PC1_production, PC2_production, and PC3_production respectively; For the real-time matrix data, the PCA module of the Scikit-learn library is also used to convert the format of the real-time matrix data to adapt to the requirements of the PCA algorithm. Assuming that the real-time matrix data is originally a 50×10 matrix (50 time points, each time point has 10 parameter data related to the casting wheel, such as casting wheel temperature, vibration frequency, surface roughness, casting wheel torque, casting wheel eccentricity, etc.); According to the data characteristics and experience, 2 principal components are set to be retained, and they can explain approximately 90% of the data changes; For example, the first principal component reflects the main change trend of the casting wheel temperature and the vibration frequency in the time series. The fluctuation of the casting wheel temperature is often accompanied by the change of the vibration frequency, and this principal component captures this dynamic correlation; The second principal component reflects the comprehensive evolution of some parameters such as the surface roughness of the casting wheel, the casting wheel torque, and the casting wheel eccentricity over time. The surface roughness affects the friction force, which in turn affects the casting wheel torque, and the casting wheel eccentricity is also related to the torque and surface wear, and they jointly reflect the comprehensive effect on the real-time state of the casting wheel in this principal component; According to the set number of principal components, the dimensionality reduction of the real-time matrix data is carried out to obtain these 2 principal components, denoted as PC1_real-time and PC2_real-time; The principal components are spliced together to form a new feature vector as the input of the microscopic image prediction model. For example, the 3 principal components obtained by dimensionality reduction of production factor data and the 2 principal components obtained by dimensionality reduction of real-time matrix data are spliced together in sequence through a data processing tool (such as the NumPy library in Python) to finally form a new feature vector [PC1_production, PC2_production, PC3_production, PC1_real-time, PC2_real-time].

[0036] Principal component analysis reduces the dimensionality to simplify data complexity and retains key information. Splicing the principal components integrates the features of multi-source data. In cooperation with the subsequent microscopic image prediction model, the model can make predictions based on concise and comprehensive data, avoiding problems such as inaccurate model predictions and low efficiency caused by high data dimensions and a large amount of redundant information, greatly improving the accuracy and efficiency of microscopic structure prediction, and providing strong support for accurately optimizing process parameters.

[0037] In some embodiments of the present invention, the passive data includes at least one of: the temperature of the casting wheel, the vibration frequency of the casting wheel, and the surface roughness of the casting wheel; More specifically, the acquisition method of the casting wheel temperature is: arranging an infrared temperature sensor array uniformly in the axial and circumferential directions of the casting wheel; The acquisition method of the casting wheel vibration frequency is: installing a three-axis vibration sensor at the bearing of the casting wheel; The acquisition method of the surface roughness of the casting wheel is: scanning the surface topography of the casting wheel in real time through a laser profiler to generate surface roughness parameters; The temperature, vibration frequency, and surface roughness of the casting wheel are monitored respectively through multiple sensors to comprehensively obtain the real-time state information of the casting wheel. These data are interrelated and can comprehensively reflect the working conditions of the casting wheel during continuous casting. Combined with the overall process parameter optimization and microscopic structure prediction process, it provides a comprehensive and accurate real-time data basis of the casting wheel for subsequent microscopic structure analysis, avoiding problems such as inaccurate process adjustment caused by incomplete monitoring of the casting wheel state, and effectively improving the stability of the copper rod continuous casting process and product quality; In addition to the above passive data, other characteristics of the casting wheel can also reflect the state of the casting wheel, such as: the thermal expansion amount of the casting wheel, the material characteristic data of the casting wheel, etc.

[0038] In some embodiments of the present invention, for step S3, the construction method of the parameter optimization model includes: Collect a large amount of copper rod continuous casting production data under different process conditions, including but not limited to casting temperature, casting speed, carbon coating thickness, casting wheel related data (temperature, vibration frequency, surface roughness, etc.) and corresponding copper rod microstructure feature data, product quality data, etc. Use data cleaning algorithms to remove outliers and missing values to ensure the accuracy and completeness of the data, and divide the sorted data into a certain proportion, such as 70% as a training set and 30% as a test set; According to the characteristics of the data and the characteristics of the problem, select the appropriate machine learning model, such as decision tree, random forest or neural network model; take the neural network model as an example, use deep learning frameworks such as TensorFlow or PyTorch to build it; determine the network structure, including the number of input layer nodes (corresponding to the number of input data features), the number of hidden layers and nodes, and the number of output layer nodes (corresponding to the number of optimized process parameters); Input the training set data into the built model and set the training parameters, such as learning rate, number of iterations, etc. During the training process, the model calculates the difference between the predicted result and the true value based on the loss function (such as the mean square error loss function for regression problems), and continuously adjusts the model parameters through the back propagation algorithm to gradually reduce the loss function value and continuously improve the model performance; Input the test set data into the trained model and calculate the model's performance indicators, such as accuracy, mean square error, etc.; based on the evaluation results, if the model performance does not meet expectations, adjust the model structure (such as increasing or decreasing the number of hidden layer nodes), reselect the model, or increase the amount of training data, and perform training and evaluation again until the model performance meets the requirements.

[0039] In some embodiments of the present invention, a method for formulating a set of adjustment factors for the determined influencing factors in step S1 includes: Evaluate the feasibility of the identified influencing factors one by one under the current production conditions. For equipment parameters such as casting temperature and casting speed, determine whether the control accuracy and adjustment range of the existing equipment can meet the adjustment requirements; for material properties, consider whether adjustments can be made under the current raw material supply channels and cost constraints; For factors that cannot be adjusted directly, find alternative factors; for example, if the casting temperature cannot be accurately controlled due to old equipment, consider upgrading the equipment locally, or indirectly control the solidification process of the copper liquid in the mold by adjusting other related parameters such as casting speed and copper liquid flow rate; The screened feasible influencing factors and the formulated alternative factors are summarized and organized to form a set of adjustment factors.

[0040] In some embodiments of the present invention, the method for constructing the microscopic image prediction model in step S6 includes: Collect a large number of historical samples covering microscopic images of copper cast billets under different process conditions, their corresponding process parameters, and the passive data of the casting wheel. The microscopic images are obtained by shooting with a professional metallurgical microscope. Invite experts in the fields of materials science and casting to carefully annotate the microscopic images. Mark different microscopic tissue regions in the images, such as equiaxed crystal regions, columnar crystal regions, etc., and mark possible defects, such as pores, cracks, etc. Organize the process parameters and the passive data of the casting wheel corresponding to each microscopic image, and establish a detailed data index for subsequent model training. According to the data characteristics and the microscopic tissue prediction target, select the convolutional neural network (CNN) as the basic model architecture because of its significant advantages in image feature extraction. Use deep learning frameworks, such as TensorFlow or PyTorch, to build the model. Determine the structure of the input layer to match the dimension of the processed data. Build multiple convolutional layers and pooling layers in the middle. The convolutional layers use convolutional kernels of different sizes to perform local feature extraction on the image features and other data features in the input data. The pooling layers are used to reduce the dimension of the feature map, reduce the computational amount, and retain key features at the same time. After the convolutional and pooling layers, add fully connected layers to fuse and map the extracted features. The number of nodes in the final output layer is determined according to the dimension of the microscopic image prediction result. For example, if the output is a microscopic tissue image matrix of a specific size, set the number of nodes in the output layer accordingly. Divide the prepared data into a training set, a validation set, and a test set according to a certain ratio. For example, 70% is used as the training set, 15% is used as the validation set, and 15% is used as the test set. Input the training set data into the built model and set training parameters, such as learning rate, number of iterations, batch size, etc. During the training process, the model adjusts the parameters of the model through the backpropagation algorithm according to the difference between the prediction result and the labeled true value, so that the value of the loss function gradually decreases. During the training process, monitor the performance of the model on the validation set in real time. By comparing the prediction result on the validation set with the true value, observe whether the model has overfitting phenomenon. If the performance index of the model on the validation set no longer improves or even decreases, adjust the training strategy in time, such as reducing the learning rate, reducing the number of iterations, or increasing the regularization term, etc. Input the test set data into the trained model, and calculate various performance metrics of the model, such as accuracy, recall rate, mean squared error, etc., to evaluate the prediction accuracy and generalization ability of the model; according to the evaluation results, if the model performance does not meet the expectations, comprehensively analyze the reasons. It may be that the model architecture is unreasonable, for example, the number and number of nodes of the convolutional layer or fully connected layer are set improperly; it may also be that the training data is insufficient and cannot cover all possible process conditions; or the training parameters are set unreasonably, affecting the convergence effect of the model; for different reasons, take corresponding optimization measures, such as adjusting the model architecture, trying to increase or decrease the number of convolutional layers and fully connected layers, and changing the number of nodes; collect more representative historical data to expand the training set; readjust the training parameters, find the optimal parameter combination through multiple experiments, and then train and evaluate again, repeating the cycle until the model performance meets the requirements of microstructure prediction.

[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A prediction method for the microstructure by optimizing process parameters based on copper rod continuous casting, characterized in that, Including: Based on the comparison result between the crystallization line position of the copper casting billet to be optimized and the standard crystallization line position, determine the influencing factors and formulate a set of adjustment factors; Obtain the microscopic structure image of the copper casting billet to be optimized and input it into a pre-constructed feature comparison model to obtain an abnormal set; Input the production factors corresponding to the copper casting billet to be optimized and the abnormal set into a pre-constructed parameter optimization model, output multiple preset schemes, and screen out multiple adjustment schemes including the adjustment factors; Set reference factors to serially adjust the adjustment schemes, and cast copper casting billets according to the earlier adjustment schemes; Obtain various passive data of the casting wheel during casting to obtain a real-time matrix; Input the real-time obtained production factors and the real-time matrix into a pre-trained microscopic image prediction model for microscopic structure prediction.

2. The microscopic structure prediction method for process parameter optimization based on copper rod continuous casting according to claim 1, characterized in that The comparison method between the crystallization line position of the copper casting billet to be optimized and the standard crystallization line position includes: Establish a database of standard crystallization line morphologies for different copper grades; Use the digital image correlation method to correct the distortion of the cross-sectional image of the casting billet to be optimized, and convert the actual crystallization line coordinates to the standard reference system; Calculate the three-dimensional parameter deviation of the crystallization line.

3. The microstructure prediction method for optimizing process parameters based on copper rod continuous casting according to claim 1, characterized in that, The influencing factors include at least one of casting temperature, casting speed, carbon coating thickness, blockage of the cooling water nozzle, surface cracks of the casting wheel and its lathe repair, and fouling of the casting wheel.

4. The microstructure prediction method for optimizing process parameters based on copper rod continuous casting according to claim 1, characterized in that, The construction method of the feature comparison model includes: Collect microscopic structure image samples of copper casting billets, perform image preprocessing and zoning processing, and set region labels; Construct a feature comparison model using a region branch structure, and introduce an attention mechanism into the feature comparison model; Divide the samples into a training set and a test set, use the region labels as auxiliary inputs, and train the feature comparison model using the training set; Adopt a cross-entropy loss function to measure the difference between the model prediction result and the true annotation, and adjust the structural parameters of the model; Import the test set into the feature comparison model, evaluate the performance of the feature comparison model, and optimize the feature comparison model according to the evaluation results.

5. The microstructure prediction method for process parameter optimization based on copper rod continuous casting according to claim 4, characterized in that The zoning processing includes: dividing the four sides of the copper casting billet into a fine equiaxed crystal zone, a columnar crystal zone, and a coarse equiaxed crystal zone in sequence from the outside to the inside, and the four sides share one coarse equiaxed crystal zone.

6. The microstructure prediction method for optimizing process parameters based on copper rod continuous casting according to claim 1, wherein, The construction method of the real-time matrix includes: Define the matrix dimension structure, divide the passive data according to a preset time span, calculate the data mean within each time span, and arrange the data means in chronological order to obtain a real-time matrix.

7. The microstructure prediction method for optimizing process parameters based on copper rod continuous casting according to claim 1, characterized in that, The method of serial adjustment includes: Determine the weight of each reference factor, score each adjustment scheme on the reference factor, calculate the comprehensive score of each adjustment scheme according to the weight and score, and sort the adjustment schemes from high to low according to the comprehensive score.

8. The method for predicting the microstructure by optimizing the process parameters based on copper rod continuous casting according to claim 7, characterized in that, The reference factors include at least one of the implementation cost of the adjustment scheme, the confidence level of the expected grain refinement effect, the coordination score of equipment parameter adjustment, and the success rate of historical similar cases.

9. The microstructure prediction method for optimizing process parameters based on copper rod continuous casting according to claim 1, characterized in that Before inputting into the microscopic image prediction model, the following processing is performed: Using the principal component analysis method, perform dimensionality reduction processing on the production factor data and the real-time matrix data respectively, and extract their respective principal components; Stitch the principal components together to form a new feature vector as the input of the microscopic image prediction model.

10. The method for predicting the microstructure by optimizing the process parameters based on copper rod continuous casting according to claim 1, characterized in that, The passive data includes at least one of the casting wheel temperature, the casting wheel vibration frequency, and the casting wheel surface roughness.

Images