Abstract: This study investigates the correlation between mould level fluctuation and slag inclusion defects to trace defect origins. Fluctuation signals were processed with piecewise aggregate approximation (PAA) and low-pass filtering for smoothing and denoising. 853 time-frequency domain features were extracted. The Kolmogorov-Smirnov test and Fisher's exact test analyzed feature-defect correlations, with Benjamini-Yekutieli method for feature selection. A weighted random forest model evaluated feature mining methods. Experimental results demonstrate that the mined-feature model outperforms 1D-CNN using raw signals with enhanced interpretability. Distribution diagrams enable quantitative analysis of slag inclusion causes, supporting optimized mould level control strategies.