Deep Clustering and 1D−CNN for Early Warning of Uniaxial Compression Failure in Yellow Sandstone

The variation characteristics of acoustic emission (AE) parameters can reveal the underlying mechanisms of rock failure. Therefore, establishing a precursor identification model based on AE features is central to monitoring and providing early warning for yellow sandstone failure. Uniaxial compression experiments were conducted on yellow sandstone to investigate its AE characteristics. A deep clustering algorithm combining an autoencoder and K−Means was employed to perform cluster analysis on AE parameter segments covering the entire uniaxial compression process. The analytical results indicate that parameter segments labeled “2” were essentially absent during the initial and middle loading stages but exhibited dense − continuous distribution characteristics immediately prior to the peak stress. Therefore, the emergence of label “2” parameter segments can be regarded as a precursor feature for the failure of yellow sandstone. Based on this precursor feature, we developed a short−term warning model for uniaxial failure of yellow sandstone using a hybrid 1D−CNN and BiLSTM method. The results show that the 1D−CNN+BiLSTM model identified a substantial number of Label “2” AE segments, but only in the “15~30” second window preceding peak stress. The timing of these identifications exhibited an error of merely “2~4” seconds compared to the clustering results. This indicates that the 1D−CNN+BiLSTM model can effectively capture the changing features of AE signal segments, enabling effective early warning for the uniaxial compressive failure of yellow sandstone.