K-means-SMOTE -ALA-XGBoost: Diagnosis of faults in hot-rolled strip steel plates

In the strip hot rolling process, controlling the steel plate convexity plays a vital role in ensuring product quality. Advanced machine learning (ML) approaches have shown substantial potential for the recognition of convexity-related defects in hot-rolled strip production data. Their advantage lies in the ability to represent complex nonlinear dynamics and intricate inter-variable dependencies, which are generally beyond the analytical capacity of conventional diagnostic techniques. Current machine learning models have not yet adequately addressed the class imbalance issue in hot-rolled strip steel data, often focusing more on learning from normal data. This limits the accuracy of diagnostics. To overcome this limitation, this paper proposes an intelligent diagnostic algorithm named K-means clustering-Synthetic Minority Oversampling Technique-Artificial Lemming Algorithm-EXtreme Gradient Boosting (K-means-SMOTE-ALA-XGBoost). Multiclass imbalance within the dataset is handled by implementing the K-means clustering-Synthetic Minority Oversampling Technique (K-means-SMOTE) approach. Through further analysis using Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP), the diagnostic framework for convexity in hot-rolled strip steel plates was examined, which also aids in exploring its influencing factors. The model’s effectiveness was evaluated by applying it to both the hot-rolling production dataset and the UCI dataset. On the hot-rolling data, it achieved significant diagnostic performance, reaching a Kappa of 0.9885, an F1 score of 0.9923, a Precision of 0.9925, and, notably, both Recall and Specificity of 0.9923. Subsequently, the K-means-SMOTE-ALA-XGBoost (K-means clustering-Synthetic Minority Oversampling Technique-Artificial Lemming Algorithm-EXtreme Gradient Boosting) model proposed in this study was compared with advanced machine learning, neural network, and mathematical models developed in recent years through experimental evaluations. The results demonstrate that the proposed model outperforms others across all evaluation metrics, showing significant advantages in diagnosing the convexity of hot-rolled strip steel plates. The study illustrates that this model outperforms traditional methods in detecting the convexity of hot-rolled strip steel plates, thereby validating its efficacy as a diagnostic tool and facilitating more intelligent control over production operations.