Objective Currently, the insufficient full-element digitization and the prevalence of “data silos” in fully mechanized mining faces (FMMF) result in poor cutting precision and operational instability of shearers under complex undulating and abrupt geological conditions. These factors pose significant challenges to high-precision proactive trajectory planning.

Methods Guided by the academic philosophy that “mining is data acquisition”, this paper proposes a coal seam data-driven cutting trajectory planning method for FMMFs. It elaborates on three core technologies: high-precision digital coal seam modeling, generative model updating and digital twin construction, and data-driven trajectory planning.To address low modeling precision, an intelligent interpolation-based digital coal seam construction method is introduced. By fusing initial geological exploration data with real-time cutting detection data, this method mitigates local smoothing and overfitting defects, establishing a high-precision digital representation of the working face. Regarding the dynamic evolution of the digital coal seam, a generative updating approach is developed. By establishing a temporal-topological network between real-time detection and historical data, the method enables the progressive evolution of the digital coal seam as the mining face advances. Simultaneously, a digital twin model is constructed to achieve real-time mapping and “virtual-real” synchronization of mining operations. For the trajectory planning challenge, a prediction model based on Causal Convolutional improved Gated Recurrent Units (CC-GRU) is established. Combined with predicted slices from the digital coal seam, multi-step look-ahead rehearsals are conducted to achieve adaptive planning and dynamic error correction.

Results and Conclusions Simulation verification using field data from a specific FMMF demonstrates that the proposed method reduces the planning errors for the roof and floor to 8.2 mm and 7.7 mm, respectively, within 100 cutting cycles. With continuous data accumulation and iterative learning, the error eventually stabilizes at approximately 5.0 mm. The comparative analysis confirms that the data-driven intelligent trajectory planning outperforms traditional manual-teaching “memory cutting” in terms of accuracy and adaptability. This approach effectively addresses the inadequate representation of real-time spatial-temporal evolution, satisfies the precision requirements of intelligent mining, and provides a theoretical foundation for high-precision dynamic modeling and intelligent cutting in FMMFs.