Research Progress on Application of Machine Learning in Continuous Casting

Abstract

Continuous casting is a key core link in steel production with characteristics of strong nonlinearity, multi-parameter coupling and dynamic fluctuations under working conditions. Traditional experience-dependent or mechanism-driven models are no longer suitable for the high-quality and high-efficiency production demands of modern steel industries. Machine learning provides an effective technical path for solving the complex control problems in the continuous casting process through its powerful data mining and pattern recognition capabilities. This paper systematically reviews the research progress of machine learning applications in the field of continuous casting, focusing on three core scenarios: abnormal prediction, quality defect detection and process parameter optimization. It sorts out the evolution from single models to feature optimization and integration, deep learning hybrid models, and mechanism-data dual-driven models. It summarizes the significant achievements of this technology in reducing production risks and improving the stability of cast billet quality, and it analyzes the prominent challenges currently faced such as data distortion and distribution imbalance, insufficient model interpretability and limited cross-scenario generalization ability. Finally, it looks forward to future technological innovation and application expansion directions, providing theoretical support and technical references for the digital and intelligent transformation of the steel industry.

1. Introduction

Continuous casting is a crucial link in the steel production process that connects steelmaking and rolling. Its production efficiency and billet quality directly determine the economic benefits and product competitiveness of steel enterprises [1,2]. However, the continuous casting process involves complex physical and chemical phenomena such as heat transfer, mass transfer, and solidification phase transformation under high-temperature conditions. There is a strong nonlinear coupling relationship between process parameters, such as casting speed, cooling intensity, superheat, and steel liquid composition, and billet quality and production safety [3,4,5]. For instance, an excessively high casting speed can lead to an overly thin initial shell, causing adhesion and steel leakage. Insufficient cooling intensity may exacerbate central segregation. Adjusting a single parameter often leads to conflicts among multiple objectives [6,7,8].

The current continuous casting production is confronted with three core bottlenecks [9,10,11]. First, the risk of sudden abnormal conditions is high. Once incidents such as adhesion and steel leakage or nozzle blockage occur, a single accident can cause equipment damage, production interruption, and significant economic losses. Second, it is difficult to control quality defects. The detection rate of defects such as cracks and inclusions in the cast billet relies on manual inspection, which is lagging and has a high misjudgment rate. Third, the efficiency of process optimization is low. Traditional numerical simulation based on metallurgical mechanisms is difficult to adapt to the dynamic adjustment requirements of multiple steel grades and working conditions. Traditional continuous casting control and optimization mainly rely on two methods [12]. One is the mechanism modeling method, which builds a solidification and heat transfer model of the cast billet based on heat transfer and fluid mechanics and optimizes process parameters through numerical calculation. However, this type of model needs to simplify the coupling effect of multiple physical fields and is highly sensitive to boundary conditions, and the calculation error will increase significantly under complex working conditions. The other is the empirical trial-and-error method, which relies on the experience of engineers in adjusting parameters such as secondary cooling water volume. Although this method is simple to operate, it is highly subjective and unstable, and it is difficult to adapt to the process requirements of steel grade switching. In addition, with the improvement in the automation level of continuous casting, the production line has the ability to collect multi-source data. Faced with the huge amount of data, traditional methods are unable to fully uncover the underlying correlation patterns between the production process and product quality hidden in the data. These bottlenecks have prompted the industry to seek breakthroughs in intelligent technologies, and machine learning can provide a new path to solve the above problems.

Machine learning, as the core branch of artificial intelligence, automatically learns patterns from massive data through algorithms and builds a mapping relationship between process parameters and control targets [13]. Machine learning can provide key support for the intelligent transformation of continuous casting. Continuous casting involves multi-physical field coupling, and the relationship between process parameters and control targets is highly nonlinear. Traditional linear models have limited accuracy, while machine learning, through multi-layer networks or integration strategies, can efficiently fit nonlinear relationships, laying the foundation for precise control [14]. Continuous casting requires real-time adjustment of key parameters to cope with changes in working conditions, and the real-time response characteristics of machine learning are adapted to the dynamic control requirements of continuous casting production. During continuous casting, multiple types of data such as temperature, flow rate, and images are generated. Machine learning can use techniques such as feature fusion to uncover hidden patterns and improve the comprehensiveness of quality control.

In recent years, relevant achievements have been made in the application reviews of machine learning in the field of continuous casting. The review of Zong et al. [15] focuses on a single dimension of slab defects, with defect detection technology as its core. It does not cover full-process scenarios such as anomaly prediction and process parameter optimization, nor does it include an analysis of dataset characteristics. Cemernek et al. [16] adopted algorithm classification and application scenarios as the framework, focusing on the macroscopic combing of algorithm principles and individual cases, without establishing a full-chain analysis system. Moreover, the literature only covers up to 2021 and does not include cutting-edge technologies such as mechanism-data dual-driven models after 2022. Lian et al. [17] reviewed the research achievements of machine learning-assisted analysis of non-metallic inclusion particles in advanced steels, but their study did not involve the application directions such as intelligent monitoring and quality optimization of the continuous casting process. Varfolomeev et al. [18] also only limited themselves to defect statistics and basic algorithms, failing to cover emerging directions such as deep learning hybrid models.

Compared with the above studies, the unique value of this review is reflected in three aspects: First, it constructs a systematic analysis framework including application foundation, core scenarios, dataset and algorithm comparison, challenges and prospects, and adds a chapter on the standardized comparison of datasets and algorithms, filling the gap of a unified evaluation benchmark in this field. Second, it has prominent timeliness, integrating cutting-edge achievements from 2022 to 2025 and incorporating industrial application cases from 2024 to 2025, which is closely aligned with the latest technological developments. Third, it strengthens the orientation towards industrial application, analyzes engineering challenges in combination with workshop working conditions and provides solutions, forming a full-process technical system covering production safety, billet quality and production efficiency.

To ensure the comprehensiveness, timeliness and rigor of the review, this study implemented a multi-dimensional literature retrieval and hierarchical screening process. The retrieval databases include international authoritative platforms (Web of Science Core Collection, Scopus, IEEE Xplore, SpringerLink, Elsevier ScienceDirect) and domestic core resources (CNKI, VIP, Wanfang). Meanwhile, it incorporates steel industry professional journals such as Ironmaking & Steelmaking, Steel Research International, Iron and Steel and China Metallurgy, as well as the proceedings of the IEEE International Conference on Industrial Informatics. The time range of the literature covers articles from the past 30 years, with priority given to screening cutting-edge achievements from 2020 to 2025.

The retrieval adopts a combination strategy of subject terms and keywords. The core keywords are divided into three categories: continuous casting-related terms (e.g., continuous casting, mold, secondary cooling, etc.), machine learning-related terms (e.g., machine learning, deep learning, CNN, etc.), and application scenario-related terms (e.g., anomaly prediction, defect detection, etc.). The target literature is accurately identified through the combination of multi-dimensional keywords.

Based on the above reasons, this article reviews the research progress of machine learning applications in the continuous casting field. It summarizes the technical applications and research status of this technology in the three core scenarios. In terms of abnormal prediction, it focuses on typical accidents such as breakout prediction, mold liquid level fluctuations, and nozzle blockages. In terms of quality defect detection, it analyzes the application logic of machine learning in defect identification, quantitative assessment, and cause analysis for key defects such as cracks, center segregation and inclusions. In terms of process parameter optimization, it focuses on core control objects such as casting speed, secondary cooling intensity, and mold operating parameters, and summarizes the target system, data support, and algorithm selection for parameter optimization. It presents the technological evolution from single-parameter control to multi-objective collaborative optimization. It summarizes the evolution characteristics of machine learning from single models to integrated models and then to dual-driven models based on mechanism and data. It also looks forward to future research and application directions, providing comprehensive references for the deep implementation of machine learning in the continuous casting field and the intelligent upgrade of the steel industry.

2. Application of Machine Learning in Continuous Casting Process

2.1. Core Process Steps and Key Control Objectives of Continuous Casting

The continuous casting process consists of five core steps: steel ladle pouring, tundish metallurgy, mold solidification, secondary cooling, and straightening and cutting. Each link achieves the ultimate goal of safety, efficiency, and high quality through the regulation of key parameters [19,20]. However, there are coupling conflicts among the goals. For instance, increasing the casting speed may intensify the liquid level fluctuation. Machine learning provides an effective path for multi-objective collaborative optimization. Ladle pouring is responsible for the transfer of molten steel and temperature maintenance, controlling the temperature drop and secondary oxidation of molten steel. Tundish metallurgy focuses on the buffering of molten steel and the removal of inclusions, ensuring temperature uniformity. Mold solidification is the key to the formation of the initial shell, requiring control of shell thickness and liquid level fluctuation. Secondary cooling achieves shell thickening and complete solidification, preventing crack formation. Straightening and cutting complete the sizing and quality screening of the cast billet, ensuring dimensional accuracy.

Table 1. Process characteristics and control objectives of each phase.

Process Link	Core Function	Key Process Parameters	Control Objectives
Ladle casting	Molten steel transfer and temperature maintenance	Ladle preheating temperature, casting time	Molten steel temperature drop < 5 °C/h, avoiding secondary oxidation
Tundish metallurgy	Molten steel buffering and inclusion removal	Tundish temperature, argon flow rate, residence time	Inclusion removal rate > 80%, temperature uniformity ±3 °C
Mold solidification	Initial slab shell formation	Mold level, cooling water flow rate, vibration parameters	Slab shell thickness ≥ 10 mm (when exiting the mold), mold level fluctuation < ±3 mm
Secondary cooling	Slab shell thickening and complete solidification	Cooling water flow rate of each section, cooling intensity	Avoiding cracks (thermal stress < critical stress), solidification end controlled before straightening
Straightening and cutting	Slab fixed-length cutting and quality screening	Straightening temperature, cutting length	Straightening temperature > 900 °C (for low-carbon steel), fixed-length error < ±5 mm

presents the comparison of process characteristics and control objectives for each step [21,22,23,24].

2.2. Classification of Machine Learning Algorithms Adapted to Continuous Casting Scenario

The machine learning algorithms for the continuous casting scenarios are mainly divided into two categories—traditional machine learning and deep learning—to adapt to different production requirements [16]. The classification of machine learning algorithms suitable for specific continuous casting scenarios is shown in Figure 1.

In traditional machine learning, unsupervised learning uses methods such as K-Means and DBSCAN to cluster and classify operating conditions, and it employs techniques like PCA to reduce and compress high-dimensional data. Supervised learning, based on labeled data, accomplishes tasks such as defect identification and steel leakage prediction. Reinforcement learning optimizes dynamic strategies like liquid level control and drawing speed adjustment through interaction with the production environment. For high-complexity demands, deep learning, including CNN for defect recognition in cast billet images, LSTM and RNN for capturing temporal patterns to support anomaly early warning, and Transformer for mining the coupling correlations of multi-physical fields to provide process optimization decisions, can also optimize production scheduling and identify abnormal equipment conditions [25,26,27,28].

2.3. Complete Process of Data-Driven Modeling for Continuous Casting

The specific workflow for data-driven modeling for continuous casting is shown in Figure 2. Firstly, key information throughout the process is obtained through data collection, covering process parameters (tundish temperature, casting speed, molten steel composition), equipment status (operating temperature, pressure), product quality (defects, dimensions), and environmental data (room temperature and humidity). During this process, it is necessary to ensure the timeliness, completeness, and accuracy of the data. Then, preprocessing is carried out. By cleaning and eliminating outliers, filling in missing values, standardizing/normalizing the data scale, and converting unstructured data into structured data through format conversion, the data quality is improved. Next, feature selection is conducted, using correlation analysis to eliminate low-correlation features, relying on tree models and ANOVA to evaluate feature importance, and using PCA and mutual information to eliminate redundant features to reduce model complexity. Then, models such as SVM and CNN are selected based on the modeling goals, and the training, validation, and test sets are divided. The model structure is established and the loss function and optimizer are set. The parameters are iteratively updated to approach the true values. Then, through methods such as grid search and Bayesian optimization, combined with validation set indicators (accuracy, RMSE), hyperparameter tuning is performed to find the optimal hyperparameters for learning rate, convolution kernel size, etc., to optimize the model performance [29,30,31].

The aforementioned general workflow for data-driven modeling in continuous casting exhibits distinct adaptive characteristics in the core scenarios of Section 3, Section 4 and Section 5. The selection of data types and algorithms is jointly determined by scenario requirements, data characteristics, and industrial constraints. Anomaly prediction (Section 3) centers on time-series data and classification algorithms: since anomalies such as sticker breakout and liquid level fluctuation feature obvious temporal evolution, it is necessary to capture precursor signals like sudden temperature changes and parameter drift through continuous time-series data. Classification algorithms such as SVM and LSTM can establish boundaries between normal and abnormal states, achieve binary classification or risk grading, while filtering noise and meeting real-time response requirements. Quality inspection (Section 4) relies on image data and deep learning networks because images can intuitively present fine-grained features of defects, such as spatial morphology and size. Models like CNN and CNN-LSTM can automatically extract hierarchical features, avoid the limitations of manual feature design, adapt to the rapid processing needs of large volumes of images, and address the pain points of traditional detection. Process parameter optimization (Section 5) employs machine learning models as fast surrogate models, stemming from the extremely low computational efficiency of traditional numerical simulations, which fail to adapt to dynamic adjustment requirements. Trained machine learning models can construct high-precision parameter-to-target mapping relationships and be embedded into optimization loops with millisecond-level response speeds. They also possess lightweight characteristics, adapt to on-site low-computing-power environments, and enable rapid optimization under multi-target coupling conflicts.

3. Application of Machine Learning in Abnormal Prediction of Continuous Casting

Continuous casting is a crucial step in steel production, and its process is influenced by multiple factors such as fluctuations in molten steel composition, coupling of process parameters, and changes in equipment status. This makes it prone to safety and quality issues such as breakout prediction, surface cracks, and abnormal liquid levels. Traditional anomaly detection methods based on empirical thresholds are unable to capture the hidden abnormal characteristics under multi-variable interaction and have limitations such as lagged prediction and high false alarm rates. Machine learning, with its powerful ability to mine time series features and recognize spatial patterns, can extract the evolution rules of abnormal conditions from massive production data, providing technical support for precise and early prediction of continuous casting anomalies. Systematic research has been conducted in core scenarios such as mold leakage, liquid level fluctuations, and nozzle blockages.

3.1. Breakout Prediction

Sticker breakout is one of the most serious accidents in continuous casting production. In essence, after the mold shell and the copper plate of the mold stick together, they are torn apart during the casting process, resulting in leakage of the molten steel. The core challenge in prediction lies in detecting the weak temperature abnormal signals at the initial stage of the mold shell’s sticking. He et al. [32] analyzed the temperature time series data of the mold thermocouple and found that there is a clear characteristic of temperature rising rapidly and then dropping before the occurrence of sticker breakout. Based on this, a GA-BP neural network model was constructed, and logical rules were introduced to optimize the output threshold of the model to achieve effective leakage steel prediction, effectively solving the problem that the traditional BP network is prone to fall into local optimum. Wang et al. [33] considered the spatial correlation of multiple thermocouples and established a time series feature model of a single thermocouple and a spatial feature model of multiple thermocouples. The single-thermocouple model captures the trend of temperature mutation of a single thermocouple, and the multi-thermocouple model analyzes the temperature gradient difference in adjacent thermocouples. Through weighted fusion, the prediction accuracy is improved, successfully avoiding the false alarm problem caused by a single-thermocouple failure. Duan et al. [34] also used the dynamic time warping algorithm instead of Euclidean distance to measure the similarity of temperature time series, which is conducive to improving the robustness of the model in temperature changes under complex working conditions. Liu et al. [35], based on the cast steel temperature measured by the thermocouple during the mold continuous casting process, established the cast steel temperature and its speed thermal imager and proposed a visual detection method. Figure 3 shows the visualization comparison diagram of breakout detection using this method. Figure 3a is the copper plate temperature thermal image at the current moment, which can intuitively present the current temperature distribution. Figure 3b is the temperature thermal image n seconds ago, serving as a historical benchmark for comparison. The colors in the figure correspond to the real-time temperature distribution of the mold copper plates, with different colors representing different temperature values. This can intuitively reflect the temperature differences among various regions of the copper plates. Figure 3c is the temperature change rate thermal image, which can identify the “V”-shaped abnormal area of breakout prediction. This method also provides a technical means for developing a more visual and intelligent mold monitoring system, helping to improve the accuracy of the breakthrough prediction system.

Machine learning in the field of continuous casting security takes early warning, low false alarm rate and high robustness as its core goals, forming a technical pattern of multi-algorithm collaboration and multi-data fusion. Traditional models such as GA-BP and multi-thermocouple feature fusion models improve the reliability of prediction by mining temporal features and avoiding faults of single sensors. Deep learning models such as ACWGAN-GP and SVM combined with adaptive PCA further enhance the ability to extract weak features and handle data fluctuations. At present, it still faces challenges such as data distortion caused by sensor drift, insufficient generalization ability across steel grades and equipment, and the difficulty in balancing the lead time of prediction and reliability. It is necessary to continuously promote optimization by relying on multi-dimensional modeling, online learning algorithms and standardized anomaly sample libraries.

3.2. Detection of Liquid Level Fluctuation

The stable control of the liquid level in the mold is crucial for the smoothness of continuous casting. Abnormal fluctuations in the liquid level under high casting speed conditions are prone to cause accidents such as slag inclusion and steel breakout. Traditional control methods are difficult to cope with dynamic disturbances under complex working conditions. A deep learning prediction model based on continuous casting big data analyzes and determines the deep correlation between the continuous casting process and the liquid level fluctuations, and it explores the frequency domain characteristics of the liquid level fluctuations and process parameters. This enables high-precision, real-time prediction of instantaneous abnormal liquid level fluctuations in the mold. Using machine learning for liquid level fluctuation prediction can achieve more efficient production process control, reduce production losses caused by abnormal liquid level fluctuations, and improve production efficiency and cast slab quality.

Sun et al. [36] proposed a feedforward neural network that integrates Bayesian optimization and nested cross-validation, using small sample industrial data to accurately predict the liquid level fluctuations of the continuous casting mold. The method overcomes overfitting, outperforming traditional models in terms of accuracy and stability, and reveals the influence mechanism of process parameters on the liquid level. Combined with visual analysis, it provides a stable operation range for on-site use, offering concise and reliable optimization guidance. Figure 4a shows the workflow of BO-NCV-FNN. Figure 4b is the corresponding network topology diagram. The speed of casting, argon gas volume, thickness of protective slag, and distance from the submerged nozzle are selected as the process variables received by the input layer. The entire structure is simple, with few parameters, and is conducive to rapid convergence with small samples. Have represents time-averaged level fluctuation and h represents initial slag layer thickness. Figure 4c is the fitting graph of predicted values and actual values. The left figure is a scatter density diagram, with data points densely clustered near the 45° baseline, and the brighter the color, the higher the fitting degree.

Su et al. [37] proposed a combined algorithm of variational mode decomposition and support vector regression to improve the prediction accuracy of the liquid level in the mold. This algorithm first processes the liquid-level signal through variational mode decomposition and empirical mode decomposition, then uses wavelet threshold denoising to remove noise interference, and finally uses the support vector regression model for prediction. The correlation coefficient of this combined algorithm increased by 3%, the root mean square error decreased by 36.1%, and its anti-noise interference ability was stronger. He et al. [38] aimed to solve the problem of multiple interfering factors and low accuracy in the prediction of the liquid level fluctuation in continuous casting molds, and they adopted genetic algorithms to optimize convolutional neural networks. The study selected 15 key features from a dataset consisting of 138 production parameters, and the prediction success rate of the continuous casting machine liquid level fluctuation was better than models such as random forest, support vector machine, and multi-layer perceptron. Tirian et al. [39] proposed a continuous casting adaptive control system based on neural networks and fuzzy logic. This system combines logical judgment with machine learning algorithms to establish the system and identify liquid level fluctuations and temperature errors. Lei et al. [40] adopted the strategy of decomposition and reconstruction, processed the complex signals collected by the eddy current sensor through EMD, and used it as input for support vector regression (SVR) prediction. The model achieved good prediction accuracy.

3.3. Detection of SEN Clogging

During the continuous casting process, the clogging of the submerged nozzle (SEN) is one of the key issues that restrict the smooth progress of production. The clogging can lead to uneven steel flow and component segregation in the cast slab, and in severe cases, it can even cause a breakage accident. Traditional judgment methods based on manual observation or empirical thresholds are difficult to detect the latent evolution trend of the clogging in advance. Machine learning, with its ability to dynamically mine time-series data and model nonlinear relationships, can extract blocking features from multiple sources of data such as process parameters and steel composition, and it achieve quantitative prediction and early warning of blocking risks. It has become the core technical direction for nozzle clogging prediction [41].

Wang et al. [42] addressed the issue that the blocking index (QI) is difficult to be monitored in real time, and they proposed a time-series prediction model based on long short-term memory network (LSTM). Twelve time-series parameters such as casting speed, molten steel temperature difference, and argon gas flow rate were used as inputs. It could visualize the distribution differences in the blocking index for different steel grades, providing a quantitative basis for clogging risk classification. Girase et al. [43] constructed a basic indicator system for blocking quantification. By analyzing data such as the steel flow velocity at the nozzle outlet and pressure loss, parameters such as clogging rate and flow uniformity index were proposed, providing clear prediction targets for machine learning models. Diniz et al. [44] proposed a clogging of the submerged entry nozzle detection model based on deep neural network (DNN). Without relying on metallurgical prior knowledge, it directly used eight real-time process parameters such as mold liquid level fluctuation and plug rod position adjustment frequency as inputs. After testing, the model achieved a recall rate of 98.75% and an accuracy rate of 97.8%. This model can quickly identify anomalies in the early stage of blocking and provide time for taking adjustment and intervention measures in a timely manner. Kuthe et al. [45] established a clogging index prediction model by combining the adaptive neuro-fuzzy inference system (ANFIS) with LSTM. Compared with traditional methods, machine learning can capture blocking trends earlier, improving prediction accuracy and efficiency. Figure 5 shows the model prediction response and LSTM forecasting results. Figure 5a presents the castability index (CI) of multiple furnace runs, ANFIS estimation, and LSTM prediction. The colors of green, yellow, and red are used to distinguish normal, warning, and blocking states, visually presenting the model’s ability to predict blocking risks. Figure 5b shows the LSTM prediction results for data within a certain period, reflecting the time series tracking of the blocking trend, and providing a basis for early intervention.

Based on the research work of scholars both at home and abroad, this article lists the typical application scenarios of machine learning in abnormal prediction of continuous casting, as shown in

Table 2. Typical applications of machine learning in abnormal prediction of continuous casting.

Authors	Intelligent Algorithms	Applications	Data Type	Evaluation Metric	Optimization Effect	Reference
He et al.	GA-BPNN	Breakout prediction	Mold level signal (collected by eddy current sensor)	CC, RMSE, MAE, MAPE	Optimizes the model by combining logical rules, improves the reliability of breakout prediction, and provides an effective method for continuous casting breakout early warning	[32]
Duan et al.	DBSCAN-DTW	Breakout prediction	Mold temperature and vibration parameters (collected by thermocouples and vibration sensors)	Accuracy, Recall, F1-Score, False Alarm Rate	Processes data through clustering and dynamic time warping, enhances the reliability of breakout prediction data, and optimizes the model input for prediction	[34]
Sun et al.	FNN	Mold level fluctuation	Molten steel chemical composition and billet size data (collected by composition analyzer and size measuring instrument)	RMSE, R2, MAE	Constructs a new feedforward neural network model, realizes effective prediction of mold level fluctuation, and provides support for stable mold level control	[36]
Su et al.	VMD-SVR	Mold level fluctuation	Mold cooling water flow rate and copper plate temperature data (collected by flow sensors and thermocouples)	Precision, Recall, Accuracy, MSE	Uses VMD to process signals, improves the prediction accuracy of SVR, optimizes the mold level prediction effect, and assists in mold level regulation	[37]
He et al.	GA-CNN	Mold level fluctuation	Continuous casting speed and tundish level time-series data (collected by speed sensors and level sensors)	MAPE, RMSE, CC	Optimizes the CNN structure with GA, enhances the accuracy of mold level fluctuation prediction, and adapts to signal fluctuations under complex working conditions	[38]
Lei et al.	EMD-SVR-GA	Mold level fluctuation	Process parameter time-series data (mold level, width, argon pressure/flow, stopper rod position, etc.)	NRMSE	Integrates the advantages of multiple algorithms, optimizes the mold level prediction model, and improves prediction accuracy and stability	[40]
Wang et al.	LSTM	Nozzle clogging	Process variable data (slide gate opening, mold level, drawing speed, tundish operation status)	Accuracy, Precision, Recall, F1-Score, MCC	Realizes the prediction of the nozzle clogging index (QI) for the next 48 s and achieves visual index distribution	[42]
Diniz	DNN	Nozzle clogging	Molten steel chemical composition data (Al, Si, Ca, Mn, S, etc.) and stopper rod position data	RMSE	Realizes binary classification detection directly through process parameters without metallurgical prior knowledge	[44]
Kuthe et al.	ANFIS-LSTM	Nozzle clogging	Mold copper plate temperature data (collected by thermocouple array)	Model classification performance (distinguishing true/false sticker breakout)	Accurately quantifies the influence weights of components such as Al, Ca, and Si on clogging, and achieves high accuracy in castability prediction for the next 24 min	[45]
Wang et al.	ACWGAN-GP	Breakout prediction	Mold level, drawing speed, tundish stopper control signal	MSE	Integrates generative adversarial network and visual detection, enhances the ability to extract precursor features of breakout, and improves the adaptability of breakout prediction	[46]
Salah et al.	SVM	Breakout prediction	9 types of key parameters including tundish molten steel temperature, mold copper plate temperature, drawing speed, mold level, etc.	Prediction Accuracy, Alarm Rate, MSE, RMSE, R2	Combines adaptive PCA to handle data fluctuations, improves the accuracy of mold level anomaly identification, and assists in the evaluation and prevention of breakout defects	[47]
Shi et al.	LWOA-TSVR	Breakout prediction	Mold level time-series data (collected by sensor)	Precision, Recall, F1-Score	Enhances the generalization ability of the model, improves the stability of breakout prediction, and provides a reliable model for breakout risk prediction	[11]
Wu et al.	SAC	Mold level fluctuation	Molten steel chemical composition (S, Al, Ca), tundish temperature, molten steel stirring time	Correct %, Unf. detected %, False Alarms	Captures the trend of abnormal mold level changes, improves the sensitivity of mold level anomaly detection, and assists in stable mold level control	[48]
Vannucci et al.	NN-FL	Nozzle clogging	Slab surface defect images and process parameter data (collected by visual sensors and process sensors)	Detection Accuracy, IoU, Recall	Optimizes the classification method for non-uniform data and improves the model’s adaptability to sensitive working conditions	[49]

.

Machine learning has enabled accurate prediction of three types of abnormalities in continuous casting: sticker breakout, liquid level fluctuation, and nozzle clogging. However, scenario adaptability and generalization capability remain core challenges. Sticker breakout prediction requires capturing subtle signal mutations of sensors in high-temperature environments. Although existing models such as GA-BP and LSTM can provide early warnings, they lack adaptability in cases of multi-thermocouple failure or steel grade switching and thus need optimization by integrating spatial features and metallurgical mechanisms. Liquid level fluctuation prediction relies on hybrid algorithms like VMD-SVR and GA-CNN to improve accuracy, but training based on a single operating condition makes it difficult to cope with sudden disturbances such as steel grade switching or argon flow rate mutations, leading to potential false alarms. Nozzle clogging prediction uses models such as LSTM and ANFIS-LSTM to capture trends and quantify the influence of compositions, but it is limited by continuous and stable time-series data. Its reliability decreases during equipment shutdowns or data interruptions, and its cross-caster migration capability needs improvement. Overall, anomaly prediction models should continuously optimize signal denoising, multi-source feature fusion, and the integration of metallurgical mechanisms to adapt to dynamically complex industrial operating conditions.

4. Application of Machine Learning in Slab Quality Inspection

The quality inspection of slabs is a crucial part of the intelligent control in continuous casting production. Cracks, segregation and inclusions are several typical defects that affect the quality of the slabs. Their formation mechanisms are complex and are influenced by multiple parameters in a coupled manner [50,51,52,53]. Machine learning, with its powerful nonlinear fitting and data mining capabilities, demonstrates significant advantages in the precise detection and prediction of these three types of defects, providing a new technical path for slab quality control [18,54,55].

4.1. Crack Defect Detection

Among various defects in casting slabs, crack defects are the most common type. Machine learning, with its precise ability to identify the correlation between process parameters and cracks, has become a core technical means for crack risk prediction and identification.

Zou et al. [56] employed a combined algorithm of PCA and DNN to conduct detection research on internal cracks in ML40Cr steel slabs. Through large sample verification, it was shown that the prediction accuracy of this model was significantly superior to traditional algorithms such as BP, and the computational efficiency could meet the requirements of online decision-making, providing an efficient technical path for the control of internal cracks in slabs. Zhang et al. [57] combined RF with K-Means and applied it to the detection of longitudinal cracks on mold surfaces. This method effectively reduced the feature dimensions of thermocouple data, achieving no missed or incorrect detections of longitudinal cracks during testing, significantly improving the reliability of longitudinal crack identification. Liu et al. [58] proposed a combined model of PCA-PSO-XGBoost for longitudinal crack detection. Through feature dimension reduction and model parameter optimization, it not only ensured high accuracy during the testing stage but also significantly reduced the false alarm rate, and it could adapt well to the complex working conditions in continuous casting production. Kong et al. [59], based on the improved BP neural network, conducted detection on various types of internal cracks in molds. After large-scale mold sample verification, the overall recognition accuracy of the model was high, especially for severe cracks, and the detection results were highly consistent with the actual production laws.

Duan et al. [60] focused on longitudinal crack detection on low-carbon steel slabs using PCA-SVM as the core method. After dimension reduction processing of the data collected by the mold thermocouple, the model training effect was excellent, maintaining high accuracy during testing, and could achieve no missed detection of longitudinal cracks under specific splitting factors. As shown in Figure 6, through the linkage of real-time temperature monitoring and the output of the PCA-SVM model, the entire process of longitudinal crack detection is presented. Figure 6a shows the real-time temperature change curve of thermocouples in a certain period, capturing the feature of the longitudinal crack from a decrease to an increase. Figure 6b is the distance from the output sample to the hyper-plane, with positive values indicating the crack period and negative values indicating normal periods. Figure 6c is the decision function value, with +1 indicating a crack period and −1 indicating a normal period. The three fully reproduce the dynamic recognition from signal input to crack determination, verifying that the model has no missed detections and responds promptly.

Research on continuous casting defect detection has evolved from single identification to a full chain covering identification, quantification, cause analysis, and process optimization. Suitable technical paths have been formed for the three core defects: cracks, segregation, and inclusions. Relevant models improve detection accuracy through feature dimensionality reduction, parameter optimization, and multi-source data fusion, and utilize tools such as SHAP values to achieve interpretable analysis of defect causes. However, issues including unbalanced data, insufficient dynamic coupling analysis, and the difficulty in balancing accuracy and real-time performance still hinder industrial implementation.

4.2. Segregation Defect Detection

The segregation defect is mainly manifested as central segregation, which is the result of selection and segregation during the solidification process of the molten steel. It is particularly severe in high-carbon steel and is difficult to completely eliminate through subsequent heat treatment [61,62,63,64,65]. Machine learning has achieved quantitative prediction and risk assessment of segregation defects by establishing a nonlinear model between process parameters and segregation degree [66,67,68].

Nieto et al. [69] analyzed the influence of continuous casting process parameters on central segregation and then used the multivariate adaptive spline data mining technology to establish a prediction model for central segregation of continuous casting slabs. The prediction results were in good agreement with the actual values. Nieto et al. [70] also proposed a hybrid algorithm of particle swarm optimization (PSO) and support vector machine (SVM) for central segregation prediction of continuous casting slabs. This algorithm optimizes the parameters of the SVM kernel function and regularization parameters through PSO, enhancing the model’s fitting ability for strongly coupled process parameters. The model’s predicted values were highly consistent with the actual values, providing an accurate quantitative tool for segregation control. Zou et al. [71] addressed the problem of central carbon segregation detection in continuous cast slabs and proposed a regularized extreme learning machine (R-ELM) model. Through regularization processing, it reduces the overfitting risk of high-dimensional data, and the prediction accuracy within the error range of ±0.03 reached 94%. This provides real-time guidance for process parameter optimization. Figure 7a shows the regression analysis of the predicted values of the R-ELM model and the target values. Figure 7b is the distribution of the prediction errors of the R-ELM model, from which it can be seen that the prediction errors are mainly concentrated within the ±0.03 range and follow a normal distribution, and the 94% accuracy rate confirms the stability of the model. Figure 7c reveals the influence laws of process parameters such as overheat and casting speed on segregation. Darker colors indicate more severe central carbon segregation in the steel slab, while lighter colors represent milder central carbon segregation and more uniform internal composition. Figure 7d compares the predicted values of the central carbon segregation index (CSI) of different casting machines with the experimental values. Figure 7 demonstrates the multi-dimensional verification of the segregation detection ability of the R-ELM model.

4.3. Inclusion Detection

Inclusions in cast slabs mainly originate from insufficient steel liquid refining and slag entrainment in the mold, etc. Their quantity and distribution directly affect the toughness and fatigue performance of the steel [72,73,74,75,76]. Machine learning, by integrating multi-source process data, has achieved precise identification and quantitative prediction of inclusion defects [77,78].

Kuthe et al. [45] proposed a combined algorithm of ANFIS and LSTM. Through ANFIS, the influence of Al, Ca, Si, and S element components on the formation of inclusions could be visually analyzed. LSTM was used to predict the pourability in the next 24 min. This provided a basis for process adjustment for inclusion control. Zhang et al. [79] proposed an optimized random forest algorithm. By integrating multiple dimensions such as the fluctuation of the mold level and argon gas flow rate, as well as training based on 7300 samples, a high accuracy rate for inclusion detection was achieved, providing an effective means for controlling the cleanliness of the molten steel. Ji et al. [80] constructed a Bayesian optimization BO-XGBoost model to predict the inclusion defect in the cast slab. After selecting the key features through coefficient screening, the model performed well in the scenario of data imbalance and provided clear guidance for process parameter optimization. Zhou et al. [81] used a CNN-LSTM hybrid model to detect inclusions in the cord steel, integrating time-varying parameters such as casting speed and mold level, and achieving precise defect identification through a single-variable and multi-variable fusion strategy.

Based on the research work of scholars both at home and abroad, this article lists the typical application scenarios of machine learning in the quality inspection of cast slabs as shown in

Table 3. Typical application of machine learning in slab quality inspection.

Authors	Intelligent Algorithms	Applications	Optimization Effect	Reference
Zou et al.	PCA-DNN	Crack defect detection	The accuracy is superior to traditional algorithms, with high computational efficiency, supporting online decision-making	[56]
Zhang et al.	RF-K-Means	Crack defect detection	Reduces feature dimensions and achieves zero missed alarms and false alarms for longitudinal cracks	[57]
Liu et al.	PCA-PSO-XGBoost	Crack defect detection	High accuracy and low false alarm rate, adapting to complex production conditions	[58]
Kong et al.	Improved BPNN	Crack defect detection	High overall recognition rate, excellent identification of severe cracks, consistent with production rules	[59]
Duan et al.	PCA-SVM	Crack defect detection	High-accuracy detection of longitudinal cracks with no missed alarms under specific conditions	[60]
Nieto et al.	PSO-SVM	Segregation detection	The goodness of fit R2 reaches 0.98, realizing quantitative prediction of segregation degree	[70]
Zou et al.	R-ELM	Segregation detection	The prediction hit rate within the error range of ±0.03 reaches 94%, adapting to multi-steel grade detection	[71]
Kuthe et al.	ANFIS-LSTM	Inclusion detection	Visualizes the mechanism of component influence, realizes performance prediction, and provides a basis for process adjustment	[45]
Zhang et al.	ORF	Inclusion detection	Integrates multi-dimensional parameters, improves detection accuracy, and assists in molten steel cleanliness control	[79]
Ji et al.	BO-XGBoost	Inclusion detection	Selects key features, adapts to data imbalance scenarios, and guides process parameter optimization	[80]
Zhou et al.	CNN-LSTM	Inclusion detection	Fuses time-varying parameters and multi-variable strategies to improve the recall rate of defect identification	[81]
Li et al.	BP	Segregation detection	Establishes a multi-dimensional feature dataset, realizes predictive analysis of quality defects, and provides data support for continuous casting process optimization	[67]
Zhou et al.	CNN	Crack defect detection	Classifies useful features from low-level to high-level; this method is simple, effective, and robust for classifying surface defects of steel plates	[82]

.

Machine learning has achieved precise identification and quantification of three core defects: cracks, segregation, and inclusions. However, the complexity of the defects and the characteristics of the data hinder industrial implementation. In crack detection, models such as PCA-DNN and RF-K-Means achieve no-miss detection through feature dimension reduction and ensemble learning. Yet, limited by data resolution and thermocouple layout, they are inadequate in identifying micro-cracks, and their adaptability across different steel grades needs to be improved. Segregation prediction establishes the mapping relationship between process parameters and segregation degree via models such as PSO-SVM and R-ELM, with improved error control capability. Nevertheless, static data struggle to depict the dynamic solidification process, leading to insufficient response to sudden operating conditions such as superheat fluctuations and abrupt changes in cooling intensity. Inclusion detection is significantly affected by composition fluctuations and sensor noise. While models like ORF and CNN-LSTM integrate multi-dimensional parameters, they fail to fully consider changes in feature distribution caused by steel grade differences and thus need to enhance robustness through multi-source feature fusion. In the future, it is necessary to strengthen the integration of data-driven and mechanism-constrained methods, optimize data augmentation techniques for small-sample defects, and improve defect recognition capability under complex operating conditions.

5. Application of Machine Learning in Optimization of Continuous Casting Process Parameters

There exists a complex nonlinear coupling relationship between continuous casting process parameters, slab quality, production efficiency, and equipment safety. Traditional empirical or mechanistic models are unable to meet the requirements of multi-objective collaborative optimization [29,44,83]. Machine learning, by uncovering the underlying patterns in multi-source production data and accurately capturing the dynamic correlations of parameters, has been systematically applied in the optimization of core parameters such as casting speed, cooling intensity, and mold operation, promoting the optimization model to evolve from single-parameter regulation to multi-objective collaboration [84,85].

5.1. Casting Speed Optimization

As a core parameter influencing production efficiency and slab quality, the optimization objective of casting speed focuses on maximizing the casting speed and reducing defect risks while ensuring the integrity of the slab shell and the stability of the liquid level. Miriyala et al. [86] proposed the TRANSFORM-ANN model, which constructs a multi-objective optimization framework using the secondary cooling water velocity as the input, achieving a coordinated optimization of maximizing the casting speed, optimizing the exit temperature of the ingot, and minimizing the bulging volume. This provides an alternative model support for online parameter adjustment. Grešovnik et al. [87,88], based on ANN, integrated the chemical composition of molten steel, the size of the ingot, and the data of the cooling water flow, as well as optimized to obtain the optimal combination of casting speed and secondary cooling intensity for the corresponding steel type. Compared with traditional empirical optimization, the internal quality qualification rate of the ingot significantly improved. Lu et al. [89] proposed a real-time prediction model composed of a convolutional neural network autoencoder and a multi-layer perceptron mixer to address the problems of slow calculation and difficulty in real-time adjustment of the 3D temperature field in traditional continuous casting. They also combined the Bayesian optimization algorithm to optimize the process parameters. Eventually, they achieved millisecond-level three-dimensional temperature field prediction and could quickly adjust the parameters to restore the metallurgical length to a reasonable range when the casting speed was increased and caused abnormal metallurgical length. Figure 8 is the framework diagram of the dual-module collaborative real-time prediction model. The upper left corner of the figure presents the simulated results of the slab’s temperature field, where different colors represent the numerical values of the 3D temperature field in continuous casting, serving as a visual indicator of the temperature distribution. After the 3D temperature field data is preprocessed, it is compressed into latent codes by CNN and reconstructed for verification. The parameter encoder module on the right maps the 9-dimensional process parameters to match the latent codes. After the latent codes of the two modules are fused, the output of the decoder is the 3D temperature field prediction result. Here, CWFR means cooling water flow rate. This framework achieves millisecond-level prediction, balancing accuracy and efficiency, and is suitable for industrial real-time control.

5.2. Cooling Parameters Optimization

The optimization of cooling parameters aims to achieve uniform thickening of the slab shell and avoid thermal stress cracks. The key is to precisely control the cooling water volume and cooling intensity in different zones to match the solidification process of the cast slab. Liu et al. [58] used the PSO-XGBoost algorithm to analyze 120 temperature data collected by thermocouples, optimize the water distribution in each area of the mold for cooling, and effectively alleviate longitudinal crack defects, and the model test accuracy reached 95.8%. Grešovnik et al. [87,88] optimized the casting speed while integrating cooling water flow data synchronously. They developed a synergistic optimization scheme combining secondary cooling intensity and casting speed, which further improved the internal quality of the slab.

5.3. Mold Operation and Molten Steel Composition Optimization

The optimization of the operation parameters of the mold focuses on goals such as stable liquid level and adaptation of vibration parameters, while the optimization of the steel liquid composition is dedicated to reducing risks such as nozzle blockage, providing a guarantee for the smooth operation of the process. Logunova [90] embedded the adaptive fuzzy decision tree into the expert system to assist in analyzing the factors causing casting defects and provide a basis for process adjustment. Matsko [91] developed an adaptive fuzzy decision tree model, embedded the machine learning optimized parameters into the expert system to achieve real-time parameter adaptive adjustment in the continuous casting process, improving the response speed and regulation accuracy of parameters. Kuthe et al. [45] proposed the ANFIS-LSTM combination algorithm, visually analyzed the influence laws of elements such as Al and Ca on nozzle blockage, optimized the control range of the molten steel composition, and improved the prediction accuracy of the nozzle’s pourability performance.

In addition, Ji et al. [80] used SHAP values for analysis, quantitatively inferring the impact of a certain feature on the process production under the condition of keeping other features unchanged. This method has good practicality in data mining and can provide new ideas for the process regulation and optimization of various steel types. Researchers can adjust appropriate process parameters based on the analysis results to improve product quality. Takalo et al. [92], in the defect prediction of ingots, used SHAP values to visualize the GBDT model after training to find the potential correlation between process input parameters and surface defects.

Machine learning has driven the optimization of process parameters from single-parameter regulation to multi-objective collaborative evolution, achieving remarkable results in core aspects such as casting speed, secondary cooling parameters, and mold operating parameters. However, challenges remain in dynamic adaptation and industrial implementation. Casting speed optimization balances production efficiency and billet quality through models such as ANN and TRANSFORM-ANN. Yet, it is constrained by multi-objective coupling conflicts (e.g., the contradiction between casting speed, liquid level stability, and billet shell thickness) and lacks real-time response capabilities, with offline optimization being the primary approach in industrial applications. Secondary cooling parameter optimization leverages PSO-XGBoost and deep learning hybrid models to achieve precise zonal control of cooling water volume, alleviating crack defects. Nevertheless, during steel grade switching and billet size adjustment, parameter adaptability needs to be improved through online learning. Mold operating parameter optimization enhances control accuracy via LSTM and reinforcement learning, but it is affected by sensor response delays and data noise, leaving room for improvement in the real-time performance and stability of adjustments. Overall, it is necessary to strengthen the model’s real-time learning capability and multi-objective conflict coordination mechanism, reduce hardware reliance through lightweight design, and facilitate the transition from offline optimization to online adaptive regulation.

6. Dataset and Algorithm Comparative Analysis

6.1. Dataset Characteristics

Datasets in the continuous casting field often exhibit characteristics of small sample sizes, high dimensions, or large sample sizes but with a single scenario. Su et al.’s [37] VMD-SVR model dataset contains no more than 5000 sets of billet liquid level data, covering only the stable operation of a single curved continuous casting machine. He et al.’s [38] GA-CNN model uses 57,616 sets of data from a double-strand continuous casting machine but focuses on a single scenario of liquid level fluctuations and does not involve steel grade switching. Kuthe et al.’s [45] ANFIS-LSTM model dataset includes 150 heats of SAE1055 steel data, although it covers multiple dimensions such as steel liquid composition, it does not include comparison data from different types of casting machines; thus, its representativeness is limited. Zhang et al.’s [56] dataset for internal crack prediction of steel contains 1600 samples (including 19 industrial parameters). Zou et al.’s [71] R-ELM model dataset contains 2476 sets of data related to the central carbon segregation of steel, with an increased sample size but still limited to a specific steel grade.

The scarcity of abnormal samples in continuous casting leads to a common problem of class imbalance. He et al. [38] selected 15 key features through random forest and set a Dropout layer to alleviate overfitting, but they did not provide variance analysis of cross-validation. Zhang et al.’s [56] PCA-DNN model is for the 4-level classification of internal cracks. In the study, the proportion of severe crack samples was appropriately increased to alleviate the imbalance, and the results showed that the stability was better than that of the model using the untreated dataset. Zhou et al.’s [75] anomaly detection dataset contains 268 heats of data, among which 64.5% are samples with inclusion defects, but there are only 95 non-defective samples, still showing a slight imbalance.

The transfer across steel grades and equipment leads to a significant decline in model performance. Kuthe et al.’s [45] model achieved a prediction accuracy of 95% on SAE1055 steel, but the accuracy dropped to 78% on low alloy steel due to the change in the sensitive threshold of Al and Ca components. Kong et al.’s [59] BPNN model had a prediction accuracy of 86.85% on the No. 1 slab continuous casting machine of a certain steel plant, but it was not verified on other specification casting machines. Lu et al.’s [88] ReP model’s training dataset covered casting speeds of 0.75–1.65 m/min. When the speed exceeded this range (such as above 1.65 m/min), the temperature prediction error significantly increased, and it could not adapt to extreme conditions.

6.2. Algorithm Comparison

To address the lack of structured and condition-normalized comparisons in algorithms, a unified benchmark of “sample size ≥ 1000 sets, containing ≥ 10 core features, and abnormal sample proportion 1–10%” is adopted here to standardize the comparison from three dimensions: data conditions, task difficulty, and evaluation metrics. Under this benchmark, the performance differences between traditional models and deep learning models are clearly visible. Su et al.’s [36] VMD-SVR model achieved a liquid level prediction effect of R = 0.9992 and RMSE = 0.6910 at only 1/8 of the computational cost of deep learning. He et al.’s [38] GA-CNN model demonstrated the advantage of deep learning in real-time scenarios with high accuracy of R² = 0.98 and MAE = 0.1093 and a prediction delay of ≤100 ms. Zhang et al.’s [56] PCA-DNN model achieved a verification accuracy of 92.2%. Lu et al.’s [89] ReP model’s MAPE for predicting the 3D temperature field was as low as 0.49%, far exceeding the traditional DNS model.

At the same time, application scenarios can be classified into simple, medium, and complex based on task difficulty. In simple tasks such as single-parameter prediction, the performance of traditional models is close to that of deep learning models. For instance, the VMD-SVR model proposed by Su et al. [37] and the GA-CNN model by He et al. [38] have a difference in liquid level prediction accuracy of less than 3%. However, in complex tasks such as multi-objective optimization and 3D field prediction, deep learning models have significant advantages. The ANFIS-LSTM model by Kuthe et al. [45] can predict nozzle clogging with a lead time of 24 min. The ReP model by Lu et al. [88], combined with Bayesian optimization, can adaptively adjust secondary cooling parameters in just 5.2 s. These efficiencies far exceed those of traditional computational models. In terms of evaluation metrics, for anomaly prediction, recall rate and false alarm rate are the key focuses. The GA-CNN model by He et al. [38] has a false alarm rate of no more than 1%. The multi-scale CNN-LSTM model by Zhou et al. [75] has an anomaly detection recall rate of 93.06%. In quality inspection, F1 score is used to balance precision and recall. The PCA-SVM model by Duan et al. [60] has a F1 score of 0.95 for longitudinal crack detection, and the BO-XGBoost model by Ji et al. [80] has an AUC of 0.811 for slag inclusion defect prediction. For process optimization, the AMR index by Sun et al. [36] and the MAE value by Lu et al. [89] are used to quantitatively evaluate the practical application value of the model.

Based on comparisons under unified conditions, it is clear that different algorithms have distinct trade-offs in terms of performance, cost, and applicable scenarios. This trade-off characteristic determines that algorithm selection must precisely match industrial demands. Tree-based models and traditional linear models have inherent advantages in small sample sizes, cross-scenario transferability, and interpretability. The RF-K-Means model by Zhang et al. [57] has no false negatives in longitudinal crack detection, and its computational cost is 60% lower than that of CNN, making it particularly suitable for the transformation of old equipment. The XGBoost model by Ji et al. [80] can quantify the impact of features such as casting speed and slab width on slag inclusion defects through SHAP values, meeting the requirements of quality traceability. The ELM model by Zou et al. [71] can predict central carbon segregation in just 0.02 s, making it suitable for online rapid response scenarios. However, the high accuracy of deep learning models often relies on high hardware investment. The GA-CNN model by He et al. [38] requires two Tesla V100 GPUs (NVIDIA, Santa Clara, CA, USA) and a sampling frequency of 100 Hz. In contrast, the VMD-SVR model by Su et al. [37] can run on a CPU with a sampling frequency of only 20 Hz, reducing hardware costs by 70%, which is more in line with the actual conditions of small and medium-sized steel mills. Nevertheless, some deep learning models have achieved a balance between performance and cost through optimization. For example, the ReP model by Lu et al. [89], although requiring GPU training, can run on a regular PC during deployment, with a single prediction taking only 0.12 s. Based on task types, for anomaly prediction scenarios, VMD-SVR [37] is suitable for low computational power requirements, while GA-CNN [38] is preferred for real-time control. In quality defect detection, PCA-SVM [60] is suitable for offline scenarios, while PCA-DNN [56] is better for online detection. For process parameter optimization, PSO-SVM [70] is efficient for single-objective tasks, while ReP + BO [88] performs better for multi-objective tasks. For composition and segregation prediction, ELM [71] can meet the requirements of rapid response, while R-ELM [71] or ANFIS-LSTM [45] models are needed for high-precision scenarios. This structured analysis based on task characteristics provides a clear basis for algorithm selection in different industrial scenarios, avoiding the one-sided judgment of accuracy above all else.

7. Summary and Outlook

Machine learning has formed a complete technical system in the continuous casting field. Through data purification and model application, it has achieved remarkable results in core scenarios such as anomaly prediction and quality inspection, outperforming traditional methods and laying the foundation for large-scale application. However, there is still a gap between technological achievements and industrial reality, with core challenges stemming from the inherent constraints of production scenarios. The harsh environment in continuous casting workshops leads to significant data distortion. Long-term operation of sensors can cause drift, and equipment upgrades can cause sudden changes in data distribution, increasing model operation and maintenance costs. Sensor noise can also generate invalid samples, which require special preprocessing to ensure training effectiveness. The principle of prioritizing high-quality production results in imbalanced data categories. Abnormal working conditions are rare to begin with, and the shutdown operations during severe anomalies further lead to a scarcity of complete abnormal samples, making it difficult to support the model’s full learning of fault features. Additionally, most deep models cannot fully correlate with metallurgical mechanisms, failing to meet the demands of process adjustments. Moreover, the contradiction between the low computing power of old equipment and the high-frequency real-time control requirements of production also restricts the implementation of high-precision models.

To address these challenges bound to industrial constraints, existing research has developed some practical solutions. In terms of data processing, high temporal resolution data collection combined with denoising methods can be used to deal with sensor drift. At the same time, based on the actual data scale, undersampling or oversampling techniques can be flexibly adopted to optimize data category balance. To enhance model interpretability, the key is to combine metallurgical mechanism analysis with feature quantification and evaluation. First, core process parameters are selected based on solid metallurgical mechanism knowledge, and then the influence weights of each parameter are determined through feature importance analysis, ultimately achieving dual verification of model prediction results and industrial mechanism validation. Model deployment optimization requires precise matching with on-site computing power conditions. In low computing power scenarios, traditional lightweight models can be prioritized to control costs, while in medium to high computing power scenarios, techniques such as model compression and parallel computing can be used to improve response speed. Thus, the best balance between model performance and deployment costs can be found.

In summary, the application value of machine learning in the continuous casting field has been verified, but scene adaptability remains the core bottleneck. In the future, efforts should focus on dual-driven modeling by mechanism and data, promoting multi-modal data fusion, and integrating metallurgical knowledge into model construction to enhance interpretability and generalization ability. For small sample and imbalanced data issues, processing methods suitable for industrial scenarios should be developed. For real-time requirements, lightweight online learning algorithms should be developed to reduce hardware dependence. At the same time, a full-process intelligent decision-making system should be constructed. Customized models should be developed based on steel type characteristics, and cross-domain technologies should be integrated to expand application boundaries. This will drive continuous casting production from merely meeting standards to achieving high quality and efficiency, providing support for the intelligent development of the steel industry.