Abstract:

In the linear control of cantilever construction of long-span continuous rigid frame bridge, it is found that the existing prediction methods have systematic defects in model construction and learning mechanism. The traditional methods have weak nonlinear fitting ability, and the machine learning model is easy to fall into local optimum or lack of generalization performance. In addition, the static modeling paradigm of ' off-line training and fixed parameters ' is widely used, which is difficult to dynamically adapt to the time-varying characteristics and error accumulation law of structural response in the construction process. In order to overcome the above problems, this paper proposes a COA-BP surrogate model that integrates crayfish optimization algorithm ( COA ) and BP neural network, and innovatively introduces a phased incremental learning mechanism. Firstly, based on FEA NX, a refined solid finite element model is established. Combined with the variability of key parameters such as concrete bulk density, elastic modulus and prestressed tension control stress, the Latin hypercube sampling is used to generate the input parameter combination, and the theoretical formwork elevation of each beam section is inversely calculated. The measured elevation is obtained after the completion of on-site construction, and the difference between the two is used as the output target of the model. The COA algorithm is used to optimize the initial weights and thresholds of the BP network, which effectively improves the convergence speed and global search ability of the model. On this basis, a phased learning strategy is designed : block 3 ~ 5 is the static learning stage, and the model initialization is completed by using the difference between the measured and theoretical elevations ; block 6 enters the incremental learning stage. The model guides the elevation adjustment of the formwork based on the prediction results. After the completion of each segment, the elevation difference corresponding to the new measured data is included in the training set to realize the dynamic closed-loop of ' construction, learning and optimization '. Based on the actual continuous rigid frame bridge project, the results show that the maximum error of No.6 block after correction is − 1.8 mm. The prediction error of the subsequent beam section continues to converge, and the prediction curve decreases smoothly, which is significantly better than the prediction effect of the traditional method. The effectiveness of the proposed COA-BP surrogate model in improving the accuracy and adaptability of linear prediction is verified.

Key words: linear prediction, continuous rigid frame bridge, crayfish algorithm, BP neural network, incremental learning, COA-BP model, bridge engineering