Abstract: Real time recognition of early fire breaking out in coal preparation plant can greatly ensure the safety of personnel and properties. However, the recognition methods currently applied suffer the drawbacks of being high in misrecognition rate and slow in real-time response, making it difficult, therefore, to meet the actual needs. To deal with the unsatisfactory situation, an improved MobileNetV3-based real-time recognition model is proposed. The attention Module CBSAM (Convolutional Block Second-order Attention Module) is integrated into the MobileNetV3 network to enhance the module′s representation ability, so as to tackle the problems currently encountered, such as delay in recognition in industrial environment, slow inference speed, and difficulty in making real-time fire recognition. The paper outlines the current state of the research work on network structure, lightweight network, and attention mechanism in the field of computer vision. On the basis of the CBSAM module′s average and maximum pooling features, the 2nd-order pooling feature is introduced in the module to improve its feature representation ability, which makes it possible to generate the channel attention images by channel attention module, and spatial attention images by the spatial attention module. For verifying the effectiveness of the model for real-time recognition of fire in coal preparation plant, the effectiveness of the CBSAM module is verified through ablation experiment using the Fire Dunning Database plus the data actually collected in plant. Through comparative experiments, the generalization and effectiveness of CBSAM module and MobileNetV3 network are verified. Results show that integration of CBSAM module with MobileNetV3 network can realize organic fusion of lightweight network and high-efficiency feature extraction, and improve the network′s classification effect; the CBSAM-MobileNetV3 has a high recognition accuracy on the plant-collected dataset; and the integration of CBSAM and MobileNetV3 can lead to reduced interference of redundant information, much better classification effect than those of other lightweight network architectures, and noticeably higher accuracy and applicability in recognition of fire in coal preparation.