Coal-rock interface prediction based on multichannel correlated complementary features

The coal-rock interface trajectory is multivariate time-series data, and complex correlations exist among different variables, which makes high-precision prediction challenging. To address this problem, this study proposed a coal-rock interface prediction model based on multichannel correlated complementary features, named SSIC-former, which integrated a Centralized Attention Mechanism (CAM), an Interactive Convolution Block (ICB), and a Sharpness-Aware Minimization (SAM) strategy. First, a sliding window method was used to construct continuous samples from the raw data. Then, an SSIC-former architecture for coal-rock identification was built to extract cross-channel correlation information and local features of the coal-rock interface, and reversible instance normalization was introduced to dynamically eliminate data nonstationarity. The CAM extracted correlated complementary features among multiple channels, while the ICB extracted local features at different scales and enabled dynamic cross-scale interaction, and their outputs were fused through residual connections to enhance feature representation. Finally, during the training stage, the SAM strategy was combined to prevent the model from falling into local optima, and the prediction results were output through a projection layer. Experimental results showed that: ① An SSIC-former-based coal-rock interface prediction model achieved a mean absolute error of 6.37 mm, a mean absolute percentage error of 2.79%, a root mean square error of 8.08 mm, a mean square error of 0.07 mm², and a coefficient of determination of 0.99, with an average inference time of 0.006 6 s per sample. Among Transformer-based models, it had the shortest inference time and met the low-latency requirements of real-time operation of shearers. ② Compared with models based on LSTM, Crossformer, Nonstationary_Transformer, FPPformer, iTransformer, and PatchTST, the SSIC-former-based model outperformed the other models in the first five evaluation metrics mentioned above, indicating that the SSIC-former-based model had high prediction accuracy and strong generalization ability and provided more accurate results for coal-rock interface trajectory prediction.