Abstract: Frame structure health detection and fault diagnosis is a very important scientific problem in mechanical engineering, civil engineering and other disciplines. In order to improve the accuracy of frame structure fault diagnosis under noise condition, this paper improves the existing TICNN (convolution neural networks with training interference) model, and obtains a new convolutional neural network model with strong anti-noise ability. In order to verify the superiority of the improved TICNN in anti-noise, the experimental results show that the improved TICNN has the best effect compared with TICNN, 1DCNN(dimensional convolution neural network) and WDCNN(first layer wide convolution kernel deep convolution Neural Network). On this basis, the fault diagnosis experiment of a four story steel structure model is carried out by using the improved TICNN and the two classification method. The results show that the improved TICNN can still get high diagnosis accuracy under noise conditions, which verifies the advantages of the improved TICNN.