Abstract: To address issues of accurately detecting flow parameters due to the complex flow characteristics of gas-solid two-phase flow, a detection method based on the improved MobileViT model was proposed, which combined neural network technology to analyze audio signals generated during the flow process of gas-solid two-phase flow. Lightweight MobileViT network was used as the base model. First, spectral subtraction was applied to denoise the collected audio signals, and their Mel spectrogram was extracted as model input. Next, a bidirectional dilated spatial pyramid pooling module was introduced into the model to extract multi-scale feature information in both horizontal and vertical dimensions. Then, a global-local spatial attention mechanism was incorporated to enhance the model's ability to capture and express the key feature information in feature map. Finally, a detail-enhanced convolution replaced the standard 3×3 convolution in the MobileViT block to extract richer local detail features from the feature map. Results show that the improved model achieves good recognition performance for gas-solid two-phase flow under six different flow conditions, with accuracy of 98.833% and 3.166% improvement over the original model.