Abstract: Gas leak recognition is essential for industrial safety and environmental protection， particularly in the petrochemical and energy industries， where the accidental release of flammable or toxic gases can lead to major incidents and environmental damage. Although traditional point sensors offer high sensitivity， their limited spatial coverage， susceptibility to environmental disturbances， and non-negligible warning latency render them inadequate for rapid leak recognition in complex and dynamic scenes. In contrast， mid-infrared multispectral imaging systems （MIR-MSI） exploit the characteristic absorption features of gas molecules in the mid-infrared band to enable non-contact， wide-field visual monitoring of gas distributions， providing an alternative modality for industrial gas recognition. However， their practical deployment is constrained by the high cost of MIR-MSI data acquisition and the scarcity of labeled samples， which limit the training and generalization of deep learning models. Moreover， the high spectral dimensionality and temporal resolution impose a substantial computational burden， and edge devices in industrial sites often have limited processing power， hindering real-time inference and early warning capabilities. To address these issues， this study presents a physics-guided MIR multispectral video generation and recognition framework. Under data-limited conditions， a physically consistent gas video generation method was developed based on an improved Gaussian diffusion plume model， to represent continuous leakage as the spatiotemporal evolution of a sequence of puffs. Wind advection， turbulent perturbation， and wavelength-dependent absorption were incorporated to synthesize realistic multispectral gas plume videos. In addition， to satisfy the real-time and computational constraints of industrial applications， a tensor-sparse low-rank decomposition-based gas plume recognition algorithm was developed. By integrating spatiotemporal smoothness constraints， the method effectively separates sparse dynamic plumes from the low-rank background， achieving efficient and robust gas recognition and visualization. The proposed framework provides technical support for the intelligent and real-time implementation of mid-infrared multispectral gas recognition in industrial environments.