Inversion of image-only intrinsic parameters for steel fibre concrete under combined rate-temperature conditions: An adaptively enhanced machine learning approach

Concrete not only bears quasi-static loads during the service of engineering structures, but also bears impact or explosion due to accidental accidents, so more and more attention has been paid to the study of concrete deformation characteristics and stress distribution. Obtaining the correct constitutive parameters is crucial for the study of the mechanical behavior of concrete, and the determination of constitutive parameters is essentially an inverse process, which is very challenging. In this paper, third-order Bessel curves are used to construct dynamic constitutive equations for steel-fibre concrete under rate-temperature union conditions, and to establish a database of the constitutive parameters corresponding to the factors influencing the mechanical behaviour of concrete. In order to select highly accurate and adaptive intelligent inversion models, Firstly, the black-winged kite optimisation algorithm (BKA) has been improved by improving the black-winged kite leader condition and integrating the optimal perturbation strategy of Morlet wavelet factor, which proves the superiority of the MBKA algorithm by comparing it with the BKA, Dung Beetle Optimisation Algorithm (DBO), Grey Wolf Optimisation Algorithm (GWO), and Harris Hawk Algorithm (HHO) in searching for the optimal results. Secondly, based on the database, the MBKA-LSSVR model, MBKA-LSSVR-Adaboost model, CNN-GRU model and CNN-LSTM model were built sequentially, respectively. The final results show that the MBKA-LSSVR-Adaboost model has the highest accuracy and the best performance for the parameter Pi (i = 0,1,2,3), and the inverted stress-strain curve proves that the proposed method is effective and accurate in determining the proposed constitutive parameters. •The third-order Bessel curves have a simple equation structure and fit the test data of concrete specimens with a small error, with a sum of squared errors up to order 1e-5.•The MBKA-LSSVR model is considered as a weak classifier and the Adaboost algorithm is introduced to improve the inversion accuracy and performance of MBKA-LSSVR.•Feature vectors are extracted by CNN, and the inversion of parameters is carried out by using GRU\LSTM neural network to learn the dynamic change law of features; compared with CNN, it is able to mine the information embedded in the historical data, which effectively improves the accuracy of prediction.•Modifying the judgement condition of the Blackpool Kite leader and introducing the design Morlet wavele