Abstract: Underground pipelines are the lifeblood of the city. Long-term disrepair results in the degradation of pipeline service performance, causing a variety of environmental problems. Therefore, the type and quantity of underground pipeline diseases should be detected to provide data supporting for pipeline maintenance. However, the method of manual detection is time-consuming and laborious, and the traditional method of computer detection has poor accuracy and low generalization ability. To solve the problem mentioned above, a classification algorithm for multiple underground pipeline diseases based on a fusion convolutional neural network was proposed. The algorithm combined Inception and Residual network architecture to improve detection accuracy. Compared with the existing detection models, the average accuracy and Macro-F1 score of the proposed model reached 93.15% and 0.932, respectively, and the detection evaluation indexes were the best, which proved that the proposed model had the characteristics of accuracy, comprehensiveness, and low error detection rate. By analyzing the actual detection results of the test set with different lighting, obstacles, and overall and local scenes, the results of the proposed model are accurate, which proves that the proposed model has high robustness and strong generalization ability.