Damage identification and localization of pultruded <scp>FRP</scp> composites based on convolutional recurrent neural network and metaheuristic intelligent algorithms

Abstract Fiber‐reinforced polymer (FRP) tendons are preferred in civil engineering for their lightweight properties, high strength, corrosion resistance, and electrical insulation. However, initial defects that arise during material preparation can adversely affect the mechanical performance and service life of structures. Local identification technology is inadequate for FRP products with variable thickness cross‐sections, especially FRP tendons, resulting in low detection efficiency. This article presents an innovative inverse problem‐solving framework aimed at simultaneously identifying the location and severity of defects in FRP tendons through frequency change rates. A convolutional recurrent neural network (CRNN) model was developed to establish the mapping between frequency change rates and associated damage information, including location and severity. The CRNN model's database was generated from finite element models (FEM), which were validated against Euler beam vibration theory, demonstrating an absolute error of less than 1%. The CRNN model was trained using this database and optimized through data matrix reconstruction, model refinement, and dilated convolution, achieving a mean absolute error (Mae) of 0.115% for predicting frequency change rate. This performance significantly surpassed that of CNN (0.318%), MLP (0.274%), and LSTM (0.334%) models. The CRNN model served as the surrogate model for the inverse problem, which was addressed using the Slime Mold Algorithm (SMA) model. The prediction error for frequency change rates with SMA model was under 0.5%, notably better than that of FEM. Consequently, this framework simultaneously identifies defects' location and severity in FRP tendons, offering valuable insights for the engineering applications of various FRP products. Highlights CRNN model achieved 0.115% MAE in frequency rates, 41.6% better than MLP. Optimized SMA identified FRP damage location with 97.8% accuracy. Hammering method effectively excited the first 8 frequencies of FRP tendons. Experimental and theoretical errors in FEM and Euler analysis stayed below 1%.