Xueli Song
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xueli Song has a total of 55 co-authors. Here are the top co-authors by number of shared publications:
- Rongpeng Li
School of Science, Chang’an University, Xi’an 710064, Shaanxi, P. R. China
8 shared publications Yuzhu Xiao
School of Science, Chang’ an University, Xi’ an 710064, Shaanxi, People’s Republic of China
6 shared publications- Yuzhu Xiao
School of Science, Chang’an University, Xi’an 710064, P. R. China
5 shared publications Supei Zheng
School of Science, Chang’ an University, Xi’ an 710064, Shaanxi, People’s Republic of China
4 shared publicationsXinbo Li
4 shared publications- Kaiming Wang
School of Science, Chang’an University, Xi’an 710064, P. R. China
3 shared publications
Latest Publications
- 1 August 2026
- Published by AIP Publishing in Physics of Fluids
- Vol. 38 (8)
Adaptive regularized hybrid physics-informed neural networks for large-gradient problems
To address the issues of insufficient accuracy and susceptibility to numerical oscillations in traditional physics-informed neural networks (PINNs) when solving large-gradient problems of hyperbolic conservation laws, this paper proposes an adaptive regularized hybrid PINN (ARH-PINN) to accurately capture large-gradient fields such as shock waves and suppress numerical oscillations. The ARH-PINNs method introduces a novel neural network architecture that integrates a Fourier embedding layer, a multi-layer perceptron, and a radial basis function layer to effectively capture both the global features and local structures of solutions to hyperbolic conservation laws. To improve the solution accuracy of large-gradient fields such as shock waves and suppress numerical oscillations, we propose an adaptive regularization strategy based on a compression indicator and a sharpness indicator and further conduct a sensitivity analysis of the associated parameters. To balance convergence speed and numerical stability, we incorporate input symmetric normalization and learning rate annealing into the training pipeline. Verified through extensive numerical examples on classical one-dimensional (1D) and two-dimensional (2D) conservation laws, the proposed ARH-PINNs can accurately resolve large-gradient fields such as shock waves, while effectively suppressing non-physical oscillations and retaining low numerical dissipation.
- 1 August 2026
- Published by AIP Publishing in Physics of Fluids
- Vol. 38 (8)
Adaptive regularized hybrid physics-informed neural networks for large-gradient problems
To address the issues of insufficient accuracy and susceptibility to numerical oscillations in traditional physics-informed neural networks (PINNs) when solving large-gradient problems of hyperbolic conservation laws, this paper proposes an adaptive regularized hybrid PINN (ARH-PINN) to accurately capture large-gradient fields such as shock waves and suppress numerical oscillations. The ARH-PINNs method introduces a novel neural network architecture that integrates a Fourier embedding layer, a multi-layer perceptron, and a radial basis function layer to effectively capture both the global features and local structures of solutions to hyperbolic conservation laws. To improve the solution accuracy of large-gradient fields such as shock waves and suppress numerical oscillations, we propose an adaptive regularization strategy based on a compression indicator and a sharpness indicator and further conduct a sensitivity analysis of the associated parameters. To balance convergence speed and numerical stability, we incorporate input symmetric normalization and learning rate annealing into the training pipeline. Verified through extensive numerical examples on classical one-dimensional (1D) and two-dimensional (2D) conservation laws, the proposed ARH-PINNs can accurately resolve large-gradient fields such as shock waves, while effectively suppressing non-physical oscillations and retaining low numerical dissipation.
- 4 February 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (6)
An improved robust sparse Bayesian learning method by matrix balancing for accurate identification of structural damage
Modal sensitivity-based model updating has proven to be to be an effective approach for damage identification. However, the numerical ill-conditioned problem due to the large condition number of the sensitivity matrix and the uncertainties caused by inadequacies in models and noise in measurements greatly affect the performance of existing modal sensitivity-based model updating. In our early work, we have proposed a robust sparse Bayesian learning (RSBL) method, in which the mixture of Gaussians (MoGs) is employed to accurately model the real uncertainties of damage identification.RSBL effectively mitigates the impact of uncertainties on damage identification, but its performance is still limited by the ill-conditioned problem. In this paper, we find theoretically that the sensitivity matrix's condition number governs the accuracy of damage identification, and propose to reduce the sensitivity matrix's condition number by incorporating a matrix balancing in RSBL. As a result, an improved RSBL method based on modal sensitivity using matrix balancing and MoGs is proposed, and it is solved through an iterative Expectation-Maximization algorithm combined with the Laplace approximation. Extensive numerical and experimental studies show that the proposed method significantly reduces the sensitivity matrix's condition number, which in turn remarkably improves the accuracy of damage identification compared to RSBL.
- 4 February 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (6)
An improved robust sparse Bayesian learning method by matrix balancing for accurate identification of structural damage
Modal sensitivity-based model updating has proven to be to be an effective approach for damage identification. However, the numerical ill-conditioned problem due to the large condition number of the sensitivity matrix and the uncertainties caused by inadequacies in models and noise in measurements greatly affect the performance of existing modal sensitivity-based model updating. In our early work, we have proposed a robust sparse Bayesian learning (RSBL) method, in which the mixture of Gaussians (MoGs) is employed to accurately model the real uncertainties of damage identification.RSBL effectively mitigates the impact of uncertainties on damage identification, but its performance is still limited by the ill-conditioned problem. In this paper, we find theoretically that the sensitivity matrix's condition number governs the accuracy of damage identification, and propose to reduce the sensitivity matrix's condition number by incorporating a matrix balancing in RSBL. As a result, an improved RSBL method based on modal sensitivity using matrix balancing and MoGs is proposed, and it is solved through an iterative Expectation-Maximization algorithm combined with the Laplace approximation. Extensive numerical and experimental studies show that the proposed method significantly reduces the sensitivity matrix's condition number, which in turn remarkably improves the accuracy of damage identification compared to RSBL.
- 11 January 2025
- Published by World Scientific Pub Co Pte Ltd in International Journal of Structural Stability and Dynamics
- Vol. 26 (14)
An ARMAX/ARCH Model to Structural Damage Localization
The time series methods derived from the structural dynamic equation have been effectively applied to damage localization. The current technique utilizes an autoregressive moving average with exogenous inputs (ARMAX) model, neglecting the dynamic responses of certain degrees of freedom, and approximating excitation and noise as a moving average (MA) process, which leads to an assumption of constant residual variance. However, this assumption typically reduces the damage localization accuracy by potentially overlooking structural damage information. To address these issues, this study proposes an approach using conditional standard deviation of the ARMAX and autoregressive conditional heteroskedasticity (ARCH) model residuals to further improve damage localization accuracy. Specifically, first, this method derives the ARMAX model from the structural motion equations, with the autoregressive and input term orders fixed. Second, the excitation, noise, and dynamic responses of degrees of freedom in other directions are considered residuals, which are then modeled as the MA process to ensure whiteness and the ARCH model to account for heteroscedasticity, thereby enhancing structural information capture. Finally, the damage feature, defined as the ratio of the conditional standard deviation of the ARCH model residuals in the test state to that in the reference state, is used to identify damage localization. Numerical simulations and experimental results from Los Alamos National Laboratory indicate that the proposed method significantly enhances damage localization accuracy compared to the existing ARMAX model.
- 11 January 2025
- Published by World Scientific Pub Co Pte Ltd in International Journal of Structural Stability and Dynamics
- Vol. 26 (14)
An ARMAX/ARCH Model to Structural Damage Localization
The time series methods derived from the structural dynamic equation have been effectively applied to damage localization. The current technique utilizes an autoregressive moving average with exogenous inputs (ARMAX) model, neglecting the dynamic responses of certain degrees of freedom, and approximating excitation and noise as a moving average (MA) process, which leads to an assumption of constant residual variance. However, this assumption typically reduces the damage localization accuracy by potentially overlooking structural damage information. To address these issues, this study proposes an approach using conditional standard deviation of the ARMAX and autoregressive conditional heteroskedasticity (ARCH) model residuals to further improve damage localization accuracy. Specifically, first, this method derives the ARMAX model from the structural motion equations, with the autoregressive and input term orders fixed. Second, the excitation, noise, and dynamic responses of degrees of freedom in other directions are considered residuals, which are then modeled as the MA process to ensure whiteness and the ARCH model to account for heteroscedasticity, thereby enhancing structural information capture. Finally, the damage feature, defined as the ratio of the conditional standard deviation of the ARCH model residuals in the test state to that in the reference state, is used to identify damage localization. Numerical simulations and experimental results from Los Alamos National Laboratory indicate that the proposed method significantly enhances damage localization accuracy compared to the existing ARMAX model.
- 31 December 2024
- Published by Elsevier in Measurement
- Vol. 245
A fundamental Bayesian two-stage model with the generalized fused lasso penalty for structural damage identification
- 31 December 2024
- Published by Elsevier in Measurement
- Vol. 245
A fundamental Bayesian two-stage model with the generalized fused lasso penalty for structural damage identification
- 4 December 2024
- Published by Elsevier in Engineering Structures
- Vol. 325
Channel mode attention network for structural damage identification
- 4 December 2024
- Published by Elsevier in Engineering Structures
- Vol. 325
Channel mode attention network for structural damage identification
- 1 November 2024
- Published by AIP Publishing in Physics of Fluids
- Vol. 36 (11)
Locally linearized physics-informed neural networks for Riemann problems of hyperbolic conservation laws
In this work, we demonstrate that physics-informed neural networks (PINNs) tend to propagate predicted shock wave information bidirectionally in time, which does not align with the actual evolution direction of solutions to hyperbolic conservation laws. This mismatch results in instability and hinders the reduction of the loss of governing equations, as well as the initial condition loss by the deep neural network. In order to tackle this problem, we simplify the complexity of the problem by constructing equivalent linear transport equations in the region of shock wave generation. The speeds of these linearized waves are governed by the Rankine–Hugoniot relations of conservation laws. This approach is termed the Locally Linearized PINNs method. Specifically, an appropriate shock wave detector is initially designed to identify domains where shock waves occur. Near shock waves, the original nonlinear equations are transformed into their linearized forms, thereby modifying the residual terms of the partial differential equations. Additionally, an equilibrium factor is introduced in fluid compression regions to reduce prediction errors and stabilize the training of deep neural networks. Numerical examples illustrate that Locally Linearized PINNs effectively address the challenge of predicting global solutions in PINNs and significantly improve shock-capturing performance for hyperbolic conservation laws.
- 1 November 2024
- Published by AIP Publishing in Physics of Fluids
- Vol. 36 (11)
Locally linearized physics-informed neural networks for Riemann problems of hyperbolic conservation laws
In this work, we demonstrate that physics-informed neural networks (PINNs) tend to propagate predicted shock wave information bidirectionally in time, which does not align with the actual evolution direction of solutions to hyperbolic conservation laws. This mismatch results in instability and hinders the reduction of the loss of governing equations, as well as the initial condition loss by the deep neural network. In order to tackle this problem, we simplify the complexity of the problem by constructing equivalent linear transport equations in the region of shock wave generation. The speeds of these linearized waves are governed by the Rankine–Hugoniot relations of conservation laws. This approach is termed the Locally Linearized PINNs method. Specifically, an appropriate shock wave detector is initially designed to identify domains where shock waves occur. Near shock waves, the original nonlinear equations are transformed into their linearized forms, thereby modifying the residual terms of the partial differential equations. Additionally, an equilibrium factor is introduced in fluid compression regions to reduce prediction errors and stabilize the training of deep neural networks. Numerical examples illustrate that Locally Linearized PINNs effectively address the challenge of predicting global solutions in PINNs and significantly improve shock-capturing performance for hyperbolic conservation laws.
- Open Access
- 21 April 2024
- Published by MDPI AG in Mathematics
- Vol. 12 (8) , 1256
A Probabilistic Structural Damage Identification Method with a Generic Non-Convex Penalty
Due to the advantage that the non-convex penalty accurately characterizes the sparsity of structural damage, various models based on non-convex penalties have been effectively utilized to the field of structural damage identification. However, these models generally ignore the influence of the uncertainty on the damage identification, which inevitably reduces the accuracy of damage identification. To improve the damage identification accuracy, a probabilistic structural damage identification method with a generic non-convex penalty is proposed, where the uncertainty corresponding to each mode is quantified using the separate Gaussian distribution. The proposed model is estimated via the iteratively reweighted least squares optimization algorithm according to the maximum likelihood principle. The numerical and experimental results illustrate that the proposed method improves the damage identification accuracy by 3.98% and 7.25% compared to the original model, respectively.
- 21 April 2024
- Published by MDPI AG in Mathematics
- Vol. 12 (8) , 1256
Open AccessA Probabilistic Structural Damage Identification Method with a Generic Non-Convex Penalty
Due to the advantage that the non-convex penalty accurately characterizes the sparsity of structural damage, various models based on non-convex penalties have been effectively utilized to the field of structural damage identification. However, these models generally ignore the influence of the uncertainty on the damage identification, which inevitably reduces the accuracy of damage identification. To improve the damage identification accuracy, a probabilistic structural damage identification method with a generic non-convex penalty is proposed, where the uncertainty corresponding to each mode is quantified using the separate Gaussian distribution. The proposed model is estimated via the iteratively reweighted least squares optimization algorithm according to the maximum likelihood principle. The numerical and experimental results illustrate that the proposed method improves the damage identification accuracy by 3.98% and 7.25% compared to the original model, respectively.
- 26 October 2023
- Published by World Scientific Pub Co Pte Ltd in International Journal of Structural Stability and Dynamics
- Vol. 24 (13)
A Structural Damage Localization Method Based on Empirical Probability Mass Function of ARMAX Model Residual and Kullback–Leibler Divergence
Damage localization is very significant in engineering applications. The existing method based on the chi-square distribution of an autoregressive moving average with exogenous inputs (ARMAX) model residual is not applicable for these realistic excitations except Gaussian excitation. To solve the above problem, this paper presents a structural damage localization method based on the empirical probability mass function (EPMF) of the ARMAX model residual and Kullback–Leibler (KL) divergence. In detail, we employ empirical data analysis (EDA) approach to estimate the EPMF of the ARMAX model residual of the data generated by the arbitrary excitation because EDA does not need any a priori knowledge about the model residual. Moreover, the KL divergence is introduced to measure the dissimilarity of the EPMFs in undamaged and damaged states to prove that our method is effective for arbitrary excitation. Finally, the semi-parametric extreme value theory is used to estimate the reliable threshold for localizing the damage. Numerical simulated and experimental results illustrate that the proposed method localizes the damage under different excitations, respectively.
- 26 October 2023
- Published by World Scientific Pub Co Pte Ltd in International Journal of Structural Stability and Dynamics
- Vol. 24 (13)
A Structural Damage Localization Method Based on Empirical Probability Mass Function of ARMAX Model Residual and Kullback–Leibler Divergence
Damage localization is very significant in engineering applications. The existing method based on the chi-square distribution of an autoregressive moving average with exogenous inputs (ARMAX) model residual is not applicable for these realistic excitations except Gaussian excitation. To solve the above problem, this paper presents a structural damage localization method based on the empirical probability mass function (EPMF) of the ARMAX model residual and Kullback–Leibler (KL) divergence. In detail, we employ empirical data analysis (EDA) approach to estimate the EPMF of the ARMAX model residual of the data generated by the arbitrary excitation because EDA does not need any a priori knowledge about the model residual. Moreover, the KL divergence is introduced to measure the dissimilarity of the EPMFs in undamaged and damaged states to prove that our method is effective for arbitrary excitation. Finally, the semi-parametric extreme value theory is used to estimate the reliable threshold for localizing the damage. Numerical simulated and experimental results illustrate that the proposed method localizes the damage under different excitations, respectively.
- 5 October 2023
- Published by Wiley in Mathematical Methods in the Applied Sciences
- Vol. 47 (2) , 1034-1043
Simultaneous uniqueness in determining the space‐dependent coefficient and source for a time‐fractional diffusion equation
This article concerns the uniqueness of an inverse problem of simultaneously identifying the space‐dependent coefficient and source in a one‐dimensional time‐fractional diffusion equation with derivative order and the zero Neumann boundary value. By additional boundary measurements, we first obtain the uniqueness of the coefficient from the Laplace transform and a transformation formula. Then, we further show the uniqueness of the source through the asymptotic behavior of solutions to the corresponding forward problem. The result shows that the uniqueness of the simultaneous identification can be obtained under the condition that the prior information only on one set of parameters in the model is given other than that of two sets.
- 5 October 2023
- Published by Wiley in Mathematical Methods in the Applied Sciences
- Vol. 47 (2) , 1034-1043
Simultaneous uniqueness in determining the space‐dependent coefficient and source for a time‐fractional diffusion equation
This article concerns the uniqueness of an inverse problem of simultaneously identifying the space‐dependent coefficient and source in a one‐dimensional time‐fractional diffusion equation with derivative order and the zero Neumann boundary value. By additional boundary measurements, we first obtain the uniqueness of the coefficient from the Laplace transform and a transformation formula. Then, we further show the uniqueness of the source through the asymptotic behavior of solutions to the corresponding forward problem. The result shows that the uniqueness of the simultaneous identification can be obtained under the condition that the prior information only on one set of parameters in the model is given other than that of two sets.