基于可微分优化的结构性态大规模智能反演方法OA

With the surge in demand for intelligent construction and operation and maintenance,numerous inverse problems have emerged,such as structural behavior inverse deduction using big data from inspection and monitoring.To address the inefficiencies and dimensionality curse of traditional inversion methods like the modified finite element method,this paper proposed an intelligent inversion method for engineering structures based on differentiable optimization.First,a forward neural network surrogate model for structural computation was trained to establish an efficient mapping relationship between state parameters and structural responses.Then,using measured data as the target,the end-to-end characteristics and automatic differentiation technology of the surrogate model were utilized to internalize the state parameters into the neural network's training parameters.A nested gradient descent algorithm was then used to perform inverse optimization,achieving efficient search and rapid convergence in the high-dimensional parameter space.A case study was conducted on the inversion of a self-built house's uneven settlement field.The results show that this method can successfully integrate data such as crack distribution and damage phenomena to invert the spatial distribution of absolute settlement values of the 52 bottom-level columns.Furthermore,the structural deformation morphology after substituting back into the finite element method is consistent with the on-site inspection and monitoring data,verifying its accuracy.A multi-initial-value inversion strategy was further introduced to verify the applicability of the proposed method to multiple solutions to inversion problems.This method improves the efficiency of a single inversion by more than 4 000 times compared to the modified finite element method,while traditional techniques even struggle to converge.Furthermore,the multi-solution inversion can obtain the distribution of the remaining bearing capacity of all bottom columns,providing crucial information support for reinforcement decisions.