<p>Transverse thickness difference is a key indicator for evaluating the quality of cold-rolled silicon steel products, jointly determined by hot-rolled silicon steel data and cold rolling process parameters. However, there are “process barriers” and “data islands” between different production lines, resulting in low accuracy and poor interpretability. In addition, the transverse thickness difference mainly relies on manual sampling measurement. The lag and uncertainty of the measurement results lead to a lack of effective online control methods. To overcome this, a cold-rolled silicon steel transverse thickness difference control framework based on interpretable machine learning is proposed. First, a hot–cold rolling cross-process data platform is established to match and integrate multivariate data from different production lines, providing a data foundation. Then, an RUN-HLSSVM-AdaBoost model prediction model combining Runge–Kutta algorithm optimized hybrid kernel least squares support vector machine and AdaBoost ensemble modeling method is established. Afterward, the adaptive bandwidth kernel density estimation improved by local weighting strategy is used to construct a prediction interval, which characterizes the uncertainty of the prediction results. SHAPley Additive exPlanations interpretable method is used to break the “black box” limitation and reveal the influence of hot and cold rolling parameters on the transverse thickness difference, and finally, an online control strategy is proposed. Industrial experiments have verified the effectiveness of the above framework, and the transverse thickness difference has been significantly improved, which provides a new paradigm for solving the problem of online control of transverse thickness difference of cold-rolled silicon steel.</p>