To address the issue that the single-modal characteristics of acoustic emission signals in rock fracture research are not comprehensive, and the existing methods have insufficient feature mining and utilization, a non-steady-state extension stage identification method based on dual-modal acoustic emission data fusion was proposed. Through uniaxial compression experiments on siltstone, acoustic emission signal data were collected. Acoustic emission parameters were extracted and time-frequency diagrams based on wavelet packet transform were introduced to construct a dual-modal dataset combining acoustic emission parameters and time-frequency diagrams. The InceptionV4 network and 1D convolutional neural network-long short-term memory(1D CNN-LSTM) network were used to form a feature extraction module for dual-modal feature extraction. To fully integrate dual-modal features, a dual-path feature fusion module was designed, including a bimodal tensor fusion module, a cross-modal attention module, and a gating mechanism. This module, along with the feature extraction module and a classification prediction module based on a multi-layer perceptron, constituted the non-steady-state extended stage identification network. The experimental results show that the accuracy and F₁ value of this recognition network reach 92.5% and 92.7% respectively, which are 2.33% and 2.72% higher than the optimal single-modal method (1D CNN-LSTM), and 1% and 0.97% higher than the optimal dual-modal fusion method. This result indicates that the proposed method can fully utilize and integrate the features of dual-modal data, achieving balanced optimization of the accuracy and reliability of non-steady-state extended stage identification, and providing a new idea for intelligent early warning of rock instability.