The present invention discloses a data difference-driven self-learning dynamic batch process optimization method including the following steps: collect production process data off line; eliminate singular batches through PCA operation; construct time interval and index variance matrices to carry out PLS operation to generate initial optimization strategies; collect data of new batches; run a recursive algorithm; and update the optimization strategy. The present invention utilizes a perturbation method to establish initial optimization strategies for an optimized variable setting curve. On this basis, self-learning iterative updating is carried out for mean values and standard differences on the basis of differences in data statistics, so that the continuous improvement of optimized indexes is realized, and thereby a new method is provided for batch process optimization strategies for solving actual industrial problems. The present invention is fully based on operational data of a production process, and does not need priori knowledge about a process mechanism and a mechanism model. The present invention is applicable to the dynamic optimization of operation trajectories of batch reactors, batch rectifying towers, batch drying, batch fermentation, batch crystallization and other processes and systems adopting batch operation.