Abstract:  The accurate prediction of reservoir brittleness is of great significance for underground geotechnical engineering disaster warning and oil and gas extraction. A brittle intelligent prediction method is proposed based on the acoustic emission signals generated during rock compression and fracture. Four types of rocks with the same size but different brittleness were experimentally prepared for indoor uniaxial rock fracturing, and the collected fracture signals were preprocessed to create a sample dataset. To address issues such as insufficient training data and limitations of traditional data augmentation methods, an improved deep convolutional generative adversarial network (DCGAN) is proposed. A deep convolutional attention generative adversarial network model based on spectral normalization (CS-DCGAN) is designed to output high-quality time-frequency images of samples, enrich the original sample dataset, and serve as input for the residual network; Extracting, learning, and iteratively training effective information from images to establish an intelligent brittleness prediction model, and continuously adjusting the hyperparameters of the model to improve its prediction accuracy; Finally, a multi-criteria evaluation is performed. The experimental results show that compared with traditional DCGAN, the improved model generates higher sample quality, with a minimum Frechet inception distance (FID) of 67.96, which can alleviate overfitting and improve the performance of the residual network. The average recognition accuracy for different brittleness can reach up to 94.95%, proving the effectiveness of the proposed method.

Key words: rock brittleness, acoustic emission signal, generative adversarial network, residual network, data augmentation