The precise monitoring of Volatile Organic Compounds (VOCs) is critical for ecological security and human health. The traditional electrical VOCs sensors are susceptibility to electromagnetic interference, chemical poisoning, and high cross-sensitivity. To address the limitations of traditional electrical sensors, this paper proposes an intelligent optical fiber speckle detection method enhanced by deep learning. By utilizing a multimode fiber (MMF) to con-struct a misalignment spliced structure, multiple modes are excited to significantly enhance the interaction between the evanescent field and the target gas. An improved dual-branch ResNet-18 deep residual network is introduced to extract and decouple features from complex speckle pattern images, aiming to facilitate simultaneous qualitative gas classification and quantitative regression prediction. Experimental results demonstrate that within a VOCs con-centration range of 0-900 ppm, the classification branch achieves an average recognition accuracy of 95%. Moreover, the regression branch exhibits superior linear correlation and prediction precision across the entire range. This study provides a robust and innovative technical solution for high-sensitivity VOCs monitoring in complex industrial environments.