Abstract: In order to solve the problem of rapid and real-time identification of tunnel lining diseases, deep learning method is adopted to compile a set of real-time identification and detection software for diseases, and to realize the alarm and counting functions of the first occurrence of new diseases. For the tunnel lining disease target detection frame instability phenomenon based on deep learning, the normalized product correlation threshold judgment method was introduced, by comparing the similarity of the two adjacent detection frames before and after the frame to discriminate whether it was a new disease. Through the experiment, we find the experimental value of the normalized product correlation threshold that can meet both the sensitivity and correctness of the new disease alarm, and the analysis of the possible influencing factors are discussed. An application example of the experimental conditions of this software is given.