Elsevier

Computers & Structures

Volume 256, November 2021, 106634
Computers & Structures

Camp chain for inverse problems of structures

https://doi.org/10.1016/j.compstruc.2021.106634Get rights and content

Highlights

  • The CC is proposed for inverse analysis at a low cost of FEM calculations.

  • The CC is a chain of camps with each built only by a few nodes.

  • A camp is a recognition model kernelled by Linear-Gaussian function.

  • Inversion accuracy quickly improves with the chain extension.

  • Superior to chain methods kernelled by MLR, RBF, and ANN over a dam case.

Abstract

Simulation-based inverse analysis plays a vital role in studying the actual working conditions of hydraulic structures, and has been largely implemented by network models in recent years. However, network-based models generally demand expensive sampling of numerical simulations and often rely on different search algorithms. Accordingly, a camp chain (CC) model is proposed to reduce numerical calculations and seek unknowns independently. The CC is comprised of a set of small size networks named camps, and each camp is an independent pattern recognition model from the unknown parameters to the corresponding system responses. An additive model involving the Linear-Gaussian kernel transformation is introduced as the mapping function for straightforward linear identification. Predictions from the previous camp direct the construction of the following one, thereby shaping a chain. Subsequently, the inversion accuracy of the CC improves along with every new camp. The effectiveness of the CC has been validated through the inversion study of thermal conductivities of a concrete dam model. Current tests indicate that the CC estimates objectives with high accuracy and speed when the number of verifications divided by the number of objectives is larger than the golden ratio.

Introduction

Optimisation progress is usually described as a search for the optimal solution in the design space. It is also posed as an inverse problem under certain circumstances. The task of inversion in the hydraulic domain often aims to find specific physical parameters or boundary conditions of the medium corresponding to recorded observations [1]. Numerical modelling methods such as the finite element method (FEM) have been regarded as standard tools for modern engineering structures, and their simulation reliability has been significantly increased with recent advances [2], [3]. For brevity, different kinds of numerical modelling are represented by FEM in the paper. On the assumption that the FEM simulations make rigorous adherence to the practical system state, an inverse analysis is feasible by revisiting the numerical process [4], [5]. Therefore, inverse problems have been implemented more and more by constructing mapping networks between sampled sought-for parameters and corresponding FEM responses at the monitoring points [6], [7], [8], [9], [10]. However, the use of network models is found to be subject to a few limitations.

On the one hand, building network models demands a considerable number of initial network nodes to avoid bad local minima and uninvertible singular matrices. In general, a network node requires a pair of input–output samples, which corresponds to a direct FEM forward calculation. Unfortunately, FEM computation in the engineering domain is typically time-consuming because of the complex mesh and boundary conditions. Even network models of low density cannot avoid unnecessary sampling at points far away from the objectives [11], [12]. In addition, a substantial amount of monitoring data by recent technological advances [13], [14], [15], [16] subsequently increases the requisite number of training samples [17], [18].

On the other hand, a range of network models are subject to one-way computing features, and thereby often resort to different search algorithms to achieve fast inversion. The radial basis function (RBF) network coupled with particle swarm optimization has been used for parameter estimations of high dams and foundations [11], [19], [20]. Classical backpropagation neural networks (NNs) combined with various tools, such as harmony search [4] and Bayesian theory [21], have been utilised to carry out inverse analysis of the physical or mechanical parameters for dam body materials. A Kriging network integrated by a genetic algorithm has also been applied to optimise the hydroelectric flow [22].

In this light, this paper proposes a camp chain (CC) model to solve inverse problems of engineering structures. The CC aims to reduce the total usage of FEM simulations and seeks for the unknowns independently. The CC consists of a chain of lightweight networks named camps, and each camp is a pattern recognition model constructed from only a few pairs of input–output samples. The material property or physical variables generally play the role of the objective parameters to be inversely estimated. The observed or simulated system responses, such as temperature changes and external deformation, are often referred to as verifying references. A pair of sampled objective parameters and the corresponding verifying references constitutes a node of the camp. The initial camp has a comparatively lower predictive accuracy than a sufficiently sampled model. Thereafter, the camp progressively moves towards the target by adding new and discarding old training samples during each update. Simultaneously the camp shrinks itself quickly, and its predictive accuracy is subsequently improved. The moving process of the camp forms a chain structure, and the predictions of the former camp direct the construction of the latter camp.

The mapping function between the inputs and outputs of each camp is simulated by a Linear-Gaussian (L-G) kernel transformation, which combines the linear function and a Gaussian function auxiliary item. This framework increases the R-squared value [23] of the model without taking in more interfering variables and subsequently improves the fitting and predictive accuracy of each camp. The nonlinear relationship between physical parameters and system responses is emulated by the Gaussian function. Moreover, the Gaussian function contributes to smoothing the model and adds stability. In fact, the unsatisfying errors produced by rough modelling with few nodes are compensated for when the camp approaches the objective point and shrinks into its close vicinity; this is because a coarse prediction at this point still attains high accuracy.

Essential techniques and procedures for building the CC are illustrated in detail in Section 2, including the sampling method for the initial camp, the pattern recognition function used in the camps, and the inversion scheme. Section 3 first gives a walkthrough of implementing the CC on an example of estimating the concrete thermal conductivities of a 2-D dam model, through which the inversion effectiveness and accuracy have been validated. Subsequently, the optimal number of camp nodes, the ability to deal with large unknowns, possible ill-conditions, and sensitivity to hyperparameters are discussed. Comparisons to other network-based models, respectively involving multiple linear regression (MLR), RBF and NN are demonstrated in Section 4. Finally, conclusions are given in Section 5.

Section snippets

Initial camp

Similar to all other pattern recognition models [24], each camp of the CC attempts to identify the mapping relationship between the input and the output data. The input X1 is extracted from the experienced value range of the unknown parameters. Suppose the numerical simulation highly approximates the practical engineering project, then the corresponding output Y1 at the verification points can be obtained by using the numerical model. In the CC, only the input of the initial camp involves

Problem set up

The Xiluodu double-curved arch dam is located on the upper Yangtze River, China. The simplified 2D model of the intermediate cross-section of the dam for studying heat transfer is used to test the feasibility and validity of the CC inversion [11], [39]. Its crest elevation, top thickness and bottom thickness in the middle are respectively 610 m, 14 m and 60 m [40], [41], [42]. As displayed in Fig. 4, the dam is designed to use 3 different kinds of concrete thereby corresponding to 3 thermal

Comparison

The case study of estimating the thermal conductivities of the concrete dam model will continue to be used in this section, to allow comparison with inverse methods respectively based on networks of MLR, RBF and NN. Also, Latinized stratified sampling will continue to be used as the initial camp sampling method, and y0-yi2<0.005 will be used as the termination criterion.

Conclusion

To avoid unnecessary numerical computation during simulation-based optimisation and inversion, the camp chain (CC) method is proposed. The CC is comprised of a series of networks with a small number of training samples for each, called camps for short, where the predictions of the previous camp constitute part of the training samples of the next camp. One pair of training samples is associated with a network node. Each camp is a plainly equipped pattern recognition model, with the L-G kernel

Data accessibility

The code for complete CC inversion is written in MATLAB and the code for Fig. 1, Fig. 5, Fig. 6, Fig. 7, Fig. 8, Fig. 9, Fig. 10, Fig. 11, Fig. 12, Fig. 13, Fig. 14 was compiled with Plotly and ggplot2 in R, which can be accessed for educational purposes on Mendeley Data [63].

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

A grateful acknowledgement is given to the State Key Laboratory of Hydro-science and Engineering, Tsinghua University. The author would also like to thank Daniel Leditschke for the English language editing.

References (63)

  • S.-H. Chen

    Computational geomechanics and hydraulic structures

    (2019)
  • M. Schäfer

    Computational engineering: Introduction to numerical methods

    (2006)
  • P. Sun et al.

    Parameter sensitivity and inversion analysis of a concrete faced rock-fill dam based on HS-BPNN algorithm

    Sci China Technol Sci

    (2016)
  • C.A. Rogers et al.

    An inverse POD-RBF network approach to parameter estimation in mechanics

    Inverse Prob Sci Eng

    (2012)
  • M. Bonnet et al.

    Inverse problems in elasticity

    Inverse Probl.

    (2005)
  • M.J. Kochenderfer et al.

    Algorithms for optimization

    (2019)
  • L. Guohe

    Global Surrogate Model with Temporal-spatial Basis and Its Application in Parameter Inversion of Dam [Doctoral thesis]

    (2017)
  • S. Kitayama et al.

    Sequential Approximate Optimization using Radial Basis Function network for engineering optimization

    Optimiz Eng.

    (2011)
  • F. Ansari

    Sensing issues in civil structural health monitoring

    (2005)
  • Q. Li et al.

    Smart Monitoring of a Super High Arch Dam during the First Reservoir-Filling Phase

    J Aerosp Eng.

    (2017)
  • L. Clarkson et al.

    Critical review of tailings dam monitoring best practice

    Int J Min Reclam Environ

    (2020)
  • R. Xiao et al.

    Deformation monitoring of reservoir dams using GNSS: An application to south-to-north water diversion project, China

    IEEE Access.

    (2019)
  • S. Chi et al.

    Back analysis of the permeability coefficient of a high core rockfill dam based on a RBF neural network optimized using the PSO algorithm

    Math Probl Eng.

    (2015)
  • X. Wang et al.

    Inverse parametric analysis of seismic permanent deformation for earth-rockfill dams using artificial neural networks

    Math Probl Eng.

    (2012)
  • C. Gu et al.

    Stochastic Inversion Method for Concrete Dams on the Basis of Bayesian Back Analysis Theory

    Adv Civil Eng.

    (2019)
  • S. Biswas et al.

    Hydroelectric Flow Optimization of a Dam: A Kriging-Based Approach

  • N.R. Draper et al.

    Applied regression analysis

    (1998)
  • C.M. Bishop

    Pattern recognition and machine learning

    (2006)
  • E. George et al.

    Statistics for experimenters: design, innovation, and discovery

    (2005)
  • L. Kuipers et al.

    Uniform distribution of sequences

    (2012)
  • R. Rinehart

    The equivalence of definitions of a matric function

    Am Math Month.

    (1955)
  • Cited by (2)

    • Identification of arch dam mechanical parameters based on sensitivity analysis and Hooke–Jeeves algorithm optimization

      2022, Structures
      Citation Excerpt :

      The traditional ratio method experiences difficulty in establishing an explicit relationship between the mechanical parameters to be solved and the observed data; thus, the numerical simulation process must be repeated in the inversion process. The calculation process is considerable, time-consuming, exhibits slow convergence, and easily falls into the local extremum [5–7]. With the development of intelligent algorithms, genetic algorithms, neural networks, and other methods are being widely used in engineering construction and management.

    • Parallel-over-series fusion model for predicting internal temperature of Xiluodu arch dam

      2022, Shuili Fadian Xuebao/Journal of Hydroelectric Engineering
    View full text