| In recent years, the traditional inspection methods of cemented steel structures rely too much on manual observation and empirical judgment, and the results are easily affected by subjective factors of inspection personnel. In order to improve the accuracy and reliability of adhesive layer damage detection of adhesive steel structure, a method of adhesive layer damage identification based on deep learning is proposed in this paper. Firstly, the validity of the finite element model is verified by ultrasonic testing of the steel members, and a large number of ultrasonic time-history response data under damaged conditions are generated on this basis. Then, the ultrasonic time-history response under different working conditions is used as the input of the deep learning model to classify the damage, identify the size and location of defects, and quantitatively identify the degree of aging of the bonding layer of the adhesive steel member respectively. Finally, the accuracy and robustness of BiLSTM, BiGRU and 1D-CNN models in each identification task are compared and analyzed, and a deep learn-based damage identification method for the bonding layer of adhesive steel components is established. The results show that the three models can accurately distinguish damage types with an accuracy of 100%. Under the condition of 20dB strong noise, the accuracy of BiLSTM model can reach 93.33%, 86.48% and 73.31% respectively in two kinds of defect quantitative accuracy requirements and defect location. Under the same noise condition, the accuracy of 1D-CNN model can reach 93.95%, 88.09% and 79.85% respectively under three kinds of aging degree detection accuracy requirements. |
