Structural Online Damage Identification and Dynamic Reliability Prediction Method Based on Unscented Kalman Filter Authors: Yan Zhang, Yongbo Zhang, Jinhui Yu, Fei Zhao, Shihao Zhu Published: 2024 ABSTRACT As sensor monitoring technology continues to evolve, structural online monitoring and health management have found numerous applications across various fields. However, challenges remain concerning the real-time diagnosis of structural damage and the accuracy of dynamic reliability predictions. In this paper, a structural online damage identification and dynamic reliability prediction method based on Unscented Kalman Filter (UKF) is presented. Specifically, in the Wiener degradation process with random effects on structural performance, the structural damage identification is initially realized using UKF. Following that, the EM algorithm is employed for estimating the performance model parameters. Eventually, dynamic reliability prediction is realized based on conditional probability. The simulation results indicate that the method effectively estimates the damage state during the structure’s use while providing accurate, real-time, and dynamic reliability predictions for the system. 1. Introduction Online structural monitoring and health management utilize modern information technology to diagnose and detect the damage state of a structure promptly, and to make real-time predictions and assessments of its reliability and safety [1]. “Online” means real-time or near real-time monitoring and evaluation of the health status of the structure, that is, the system can continuously collect data, analyze the data in a timely manner, and provide decision support and emergency response. The development of this technology has taken nearly fifty years; however, several key issues still limit its application, including the reliability and stability of the sensors, the challenges of real-time data acquisition and transmission, and the accuracy of the algorithm software. Currently, the primary methods for dynamic damage identification using sensors include model modification [2,3], dynamic fingerprinting [4,5], genetic algorithms (GAs) [6,7], neural networks [8,9], wavelet analysis [10,11,12], and real-time signal processing methods [13,14,15,16]. The model modification pinpoints damage within the structure by comparing experimental vibration responses with theoretical calculations, utilizing optimized constraints to assess changes in stiffness. The dynamic fingerprinting method utilizes a relevant database to identify the corresponding damage state based on measured kinetic changes. The damage identification method based on GA uses the algorithm to find the relevant kinetic parameters that closely match the measured data to determine the location and extent of damage. Zhang [6] applied an improved generalized genetic algorithm to damage recognition; Ramezani Meysam [7] proposed an improved GA based on finite vibration modes to determine the damage severity and location in units and connections. Damage detection using neural networks establishes the relationship between key structural parameters and damage states to facilitate structural damage identification. Veronika [8] proposed a structural damage identification method of composite rotors based on fully connected neural networks and convolutional neural networks. Zhan [9] proposed a damage identification method in a beam-like structure using a strain FRF-based damage index and an artificial neural network. Wavelet analysis breaks down structural dynamic signals into valuable information for analyzing the state of structural damage. Both the model modification and the dynamic fingerprinting method require testing and collecting substantial information, which limits their applicability in complex structures. GA and neural networks require substantial computational resources, exhibit slow convergence rates, and are prone to getting trapped in local minima. Wavelet analysis offers significant advantages for complex signals, but selecting the appropriate basis function can be challenging. The real-time signal processing method performs structural parameter identification by directly constructing equations from dynamic data in the time domain. This method helps prevent information loss that may occur during intermediate conversion processes, including the least squares estimation [13], the maximum likelihood estimation [14], and the Kalman filter method [15,16]. Kalman filter is a recursive linear minimum variance estimation method that uses the system’s state equations and observed data to optimally estimate the state, enabling physical parameter identification. It can integrate dynamic information and holds significant potential for multi-sensor data fusion. Currently, various filter methods are actively used in the field of structural damage and reliability analysis, including the extended Kalman filter (EKF) [17,18], Unscented Kalman Filter (UKF) [19], particle filter (PF) [20], H∞ filter [21], Moving Horizon Estimation (MHE) [22], and Neural Network-Based Filters [23,24]. The EKF is a nonlinear extension of the Kalman filter that uses a Taylor series expansion to linearize the system’s nonlinear dynamics and observation models. However, this method is primarily applicable to linear or mildly nonlinear structures, and the calculation of the Jacobian matrix can be relatively complex. The UKF addresses nonlinear problems using the “unscented transformation”. Unlike the EKF, the UKF does not require the calculation of the Jacobian matrix, offers better adaptability to nonlinear systems, and can achieve higher estimation accuracy. The PF estimates a system’s state using a set of randomly sampled particles and is capable of handling highly nonlinear systems with non-Gaussian noise in parameter estimation. Cristiani [20] diagnosed and predicted the damage of composite double cantilever beam specimens through particle filtering and surrogate models. The H∞ filter designs the filter based on the criterion of minimizing error energy, making it well suited for handling situations with uncertainty and strong noise. Tang [21] used neural networks and H∞ filters to conduct structural damage detection. MHE is an optimization-based method that estimates the current system state by minimizing an error function within a sliding window. Neural Network-Based Filters, on the other hand, use deep learning models to estimate system states and excel in handling complex, nonlinear, or high-dimensional systems. Pathirage [23] proposed a structural damage identification framework based on autoencoders, which can support deep neural networks. Rautela [24] applied model-assisted convolutional and recurrent neural networks for structural damage detection and localization using ultrasonic-guided waves. However, methods like PF, H∞ filters, MHE, and Neural Network-Based Filters require substantial computing resources and memory, which limits their real-time performance. In contrast, the UKF offers significant advantages in both ensuring calculation accuracy and maintaining real-time performance for online damage identification and reliability prediction. On the other hand, methods based on structural performance degradation analysis are currently the dominant approach for studying the dynamic reliability of systems. These methods primarily include regression analysis [25,26,27], time series analysis, stochastic process analysis [28,29,30], and other analytical techniques [31,32]. Regression analysis uses regression modeling to study the process of structural performance degradation. Time series analysis provides reliability predictions and assessments based directly on degraded data. Stochastic process analysis uses specific stochastic models to represent the problem. Common models include the inverse Gaussian process, Gamma process, and Wiener process. Pan [28] used the Wiener process to model the degradation of product performance parameters; Niu [30] accomplished the performance reliability degradation analysis and lifetime prediction based on the generalized normal gray cloud Bayesian Wiener process. There are also some other methods such as dynamic reliability assessment using fault trees and dynamic Bayesian networks [31] and probability density evolution methods [32]. The current research on system dynamic reliability prediction typically assumes that performance parameters are directly observable. However, in practice, the performance degradation process is generally implicit within the system, making direct monitoring difficult. At the same time, performance degradation is accompanied by uncertainty, as the system is exposed to varying environments, stresses, and other factors during operation. Therefore, to realize the real-time reliability prediction of the implied performance degradation process in complex systems, this paper proposes a structural online damage identification and dynamic reliability prediction method based on UKF. It is important to emphasize that, while UKF has been widely applied to nonlinear systems, this paper’s contribution lies in its tailored adaptation for structural damage identification. We integrate online damage detection with dynamic reliability assessment, specifically designed for systems with complex degradation processes. This approach enables the rapid identification of damage coefficients and provides accurate, real-time, and dynamic reliability assessments. The rest of this article is organized as follows. Section 2 provides a specific description of the research problem and establishes the system equations for the implied performance degradation. Section 3 proposes an online damage identification and dynamic reliability prediction method for structures based on UKF. Section 4 carries out simulation experiments and discussion of the method with an example. Section 5 summarizes this article. 2. Problem Statement 2.1. Identification of Implicit Performance Degeneration Processes For general nonlinear systems, there are (1)Xk=f(Xk−1,Φ,tk−1)+wk−1Yk=h(Xk,tk)+vk where Xk is the state variable, Yk is the observation variable, and Φ is the parameter vector. Additionally, wk and vk correspond to the system noise and measurement noise, respectively. Assuming that the characteristic quantity reflecting the performance degradation is represented by φ, its relationship with the parameter vector can be expressed by (2)Φ=u(φ). In structural systems, state variables such as displacement, velocity, and acceleration can be directly observed. However, the stiffness coefficients of the structure, which serve as implied performance variables, are challenging to measure in real-time. This paper utilizes a Wiener process with random effects on the drift coefficients to model the performance degradation process. Assume that the performance eigenvalue φ satisfies (3)φ(τ)=η0+ητ+σB(τ),η~N(μη,ση2) where η0=φ(0) is known; τ represents time; μη, ση, and σ are unknown parameters; B(τ) is standard Brownian motion. Brownian motion with drift satisfies φ(τ+Δτ)−φ(τ)~N(ηΔτ,σ2Δτ). Discretization has (4)φk=φk−1+ηT+εk−1φ0=η0 where T is the sampling step, εk~N(0,σ2T). Then, the system equations for the final implied performance degradation are (5)Xk=f(Xk−1,u(φ))+wk−1φk=φk−1+ηT+εk−1,η~N(μη,ση2)Yk=h(Xk)+v. 2.2. Dynamical System Equations for Structure Stiffness Degradation The dynamical equation of the structure is (6)Mx¨(t)+Cx˙(t)+Kx(t)=Uf(t) where M, C, and K represent the overall mass, damping, and stiffness matrices of the structure, respectively. U denotes the overall equivalent nodal load action vector. The vectors x¨, x˙, and x represent acceleration, velocity, and displacement at a given time, while f(t) represents the time-dependent external excitation force. Transforming Equation (6) into the form of a matrix, we have (7)x˙x¨=0I−M\K−M\Cxx˙+0M\Uf(t). Given X=xx˙T, A=0I−M\K−M\C, B=0M\U, it can be obtained (8)X˙=AX+Bf(t). Discretization has (9)Xk=Fk−1Xk−1+Uk−1fk−1 where (10)Fk−1=eA(Δt) (11)Uk−1=∫tk−1tkeA(tk−τ)dτB=[I−e−A(Δt)]A−1B. Xk of Equation (9) contains the displacement vector and the velocity vector. If the stiffness of the structure is selected as the characteristic parameter for damage performance, a damage coefficient is applied to the stiffness matrix of each element to quantify the extent of damage. Let the element stiffness matrix be denoted as Ke, with the corresponding stiffness damage coefficient represented by xk. The variation of the stiffness matrix with time t is (12)Ke(t)=Kexk(t). Setting the variable XK=[xk1,⋯,xkn] and increasing its dimension to include it in the state vector, we have (13)XkXKk=Fk−100IXk−1XKk−1+Uk−10fk−1. In dynamics, measurement equations primarily utilize displacement, velocity, or acceleration as measurement signals. When acceleration sensors are employed, the following measurement equation is (14)Zk=x¨k=M−1Uf(t)−Cx˙(t)−Kx(t) In light of the performance degradation process discussed in Section 2.1, the structural system equations analyzed in this paper are (15)XkXKk=Fk−100IXk−1XKk−1+Uk−10fk−1+wk (16)xkk=xkk−1+ηΔt+εk−1 (17)Zk=M−1Uf(t)−Cx˙(t)−Kx(t)+vk where wk and vk are system noise and measurement noise, η~N(μη,ση2),ε~N(0,σ2). 3. Methodology 3.1. Damage Parameter Estimation by UKF The equations for identifying the structural damage parameters are (18)XkXKk=Fk−100IXk−1XKk−1+Uk−10fk−1+wk (19)Yk=[−M\K−M\C]Xk+M\Ukfk+vk (20)E(wk)=0,E(wiwjT)=Qkδij;E(vk)=0,E(vivjT)=Rkδij;E(wivjT)=0 where (18) is the state equation and (19) is the measurement equation; wk and vk are zero-mean, uncorrelated white noise processes with known covariance matrices Qk and Rk, respectively; δij is the Kronecker-δ function. The process of damage parameter estimation using UKF [33] is as follows: Set Xa=[X,XK]T. The filter is initialized with the predicted mean and covariance of the state. (21)X^a,0+=E(Xa,0) (22)P0+=E(Xa,0−X^a,0+)(Xa,0−X^a,0+)T where Xa,0, X^a,0+, and P0+ denote the initial state vector, the estimate of the initial state vector, and the covariance of the initial state vector, respectively. The equations for updating time. To propagate from step (k−1) to step k, the 2n sampling sigma points are calculated by the unscented transformation rule, as follows:(23)X^a,k−1(i)=X^a,k−1++X˜a(i),i=1,⋯,2n (24)X˜a(i)=(nPk−1+)iT,i=1,⋯,n (25)X˜a(n+i)=−(nPk−1+)iT,i=1,⋯,n where X^a,k−1+ and Pk−1+ denote the posteriori mean and covariance of step (k−1), and X^a,k−1(i) denotes the i-th sigma point. Convert the sigma points to X^a,k(i) by Equation (18), as follows:(26)X^a,k(i)=X^(i)(k)XK^(i)(k)=F^(i)(k−1)00IX^(i)(k−1)XK^(i)(k−1)+U^(i)(k−1)0f(k−1). The prior state estimate X^a,k− at step k is obtained by weighted averaging: (27)X^a,k−=12n∑i=12nX^a,k(i). Calculate the prior covariance matrix Pk− of the state prediction by (28)Pk−=12n∑i=12n(X^a,k(i)−X^a,k−)(X^a,k(i)−X^a,k−)T+Qk−1. The equations for updating measurement. Generate new sigma points using X^a,k− and Pk− by (29)X^a,k(i)=X^a,k−+X˜a(i),i=1,⋯,2n (30)X˜a(i)=(nPk−)iT,i=1,⋯,n (31)X˜a(n+i)=−(nPk−)iT,i=1,⋯,n where X^a,k(i) denotes the i-th sigma point at step k. Convert the sigma points to Y^k(i) by Equation (19), as follows:(32)Y^k(i)=[−M\Kk−M\Ck]X^a,k(i)+M\Ukfk,i=1,⋯,2n. The measurement prediction Y^k and the covariance Py are calculated based on the statistics of the transformed sigma points, as follows:(33)Y^k=12n∑i=02nY^k(i) (34)Py=12n∑i=02n(Y^k(i)−Y^k)(Y^k(i)−Y^k)T+Rk. The cross-correlation covariance, P^xy, is calculated using (35)P^xy=12n∑i=02n(X^a,k(i)−X^a,k−)(Y⌢k(i)−Y^k)T. The Kalman gain matrix Dk is approximated from the cross-correlation and measurement covariances, as follows:(36)Dk=PxyPy−1. Calculate the posteriori estimated mean X^a,k+ and covariance Pk+ of step k, completing the measurement update, as follows:(37)X^a,k+=X^a,k−+Dk(Yk−Y^k) (38)Pk+=Pk−−DkPyDkT. Through the above process, the state estimates at each step can be derived, enabling the parameter identification. 3.2. EM Algorithm Based on the Wiener degradation process with random effects, the damage coefficient xk satisfies (39)xk(t)=η0+ηt+σB(t),η~N(μη,ση2). where η is the drift coefficient representing the degradation of structural stiffness damage performance, which follows a normal distribution with mean μη and variance ση2; σ is the diffusion coefficient; and B(t) denotes the standard Brownian motion. Discretization has (40)xkk=xkk−1+ηΔt+εk−1η~N(μη,ση2),εk−1~N(0,σ2). The parameters to be estimated in this degradation model include μη, ση2, and σ2. The parameters include hidden variables, which cannot be directly estimated using the maximum likelihood method. The Expectation–Maximization (EM) algorithm is commonly used to solve the parameter estimation problem involving hidden variables in probabilistic models. Let Θ=(μη,ση2,σ2)T denote the parameter vector; the EM algorithm [34] is required, as the model contains hidden variables. Set Z to be the overall degradation dataset, with a total of n data sets, each with m time points. Let Θk=(μηk,(ση2)k,(σ2)k)T be the estimate of the parameters at step k. Given Θk, there is (41)P(Zi|ηi,Θk)=∏j=1Zif(Δxkij|Θk). Given that ηi~N(μη,ση2), the posterior estimate of ηi follows normal distribution, provided that Zi and Θk are known. Let its mean be μik and its variance be (σi2)k. According to Bayes’ theorem, we can derive (42)P(ηi|Zi,Θk)∝P(Zi|ηi,Θk)P(ηi|Θk)∝exp−∑j=1m(Δxkij−ηiΔtj)22(σ2)kΔtjexp−(ηi−μηk)22(ση2)k∝exp−ηi−(∑j=1mΔxkij)(ση2)k+μηk(σ2)k/(∑j=1mΔtj)(ση2)k+(σ2)k22(σ2)k(ση2)k/(∑j=1mΔtj)(ση2)k+(σ2)k. The posterior estimate of ηi follows a normal distribution and can be obtained as (43)P(ηi|Zi,Θk)=12π(σi2)kexp−(ηi−μik)22(σi2)k where (44)μik=(∑j=1mΔxkij)(ση2)k+μηk(σ2)k(∑j=1mΔtj)(ση2)k+(σ2)k (45)(σi2)k=2(σ2)k(ση2)k(∑j=1mΔtj)(ση2)k+(σ2)k. Under the conditions of knowing Zi and Θk, the EM algorithm first obtains (46)Q(Θ|Y,Θk)=−12∑i=1n[(m+1)ln2π+∑j=1mlnΔtj+mlnσ2+lnση2+∑j=1m(Δxkij)2−2μikΔxkijΔtj+(Δtj)2((μik)2+(σi2)k)σ2Δtj+(μik)2+(σi2)k−2μikμη+μη2ση2]. At the M-step, setting ∂Q∂Θ=0, the latest estimated value of parameter Θ is determined, and the final solution is (47)μηk+1=1n∑i=1nμik (48)(ση2)k+1=1n∑i=1n(μik)2+(σi2)k−2μikμηk+1+(μηk+1)2 (49)(σ2)k+1=1nm∑i=1n∑j=1m(Δxkij)2−2μikΔxkijΔtj+(Δtj)2((μik)2+(σi2)k)Δtj. The EM algorithm operates through two fundamental steps, the E-step and the M-step, and iterates multiple times until a predefined convergence criterion is met, ultimately providing the maximum likelihood estimate of the parameters. The iteration terminates when the difference between the parameter vectors Θk and Θk+1 is less than a predefined threshold. 3.3. Dynamic Reliability Prediction Method By combining the concepts of “Soft Failure” and “First Hitting Time”, for the Wiener process with random effects, the probability density function (PDF) and reliability function (RF) have been derived from Ref. [35]:(50)fT(t)=w2πt3(ση2t+σ2)exp−(w−μηt)22t(ση2t+σ2) (51)RT(t)=Φw−μηtση2t2+σ2t−exp(2μηwσ2+2ση2w2σ4)×Φ−2ση2wt+σ2(μηt+w)σ2ση2t2+σ2t where w denotes a given safety threshold value; Φ(·) denotes the standard normal cumulative distribution function. The expectation of remaining useful life (RUL) can be expressed, as follows:(52)E(T)=2wσηDμη2ση where D(z)=exp(−z2)∫0zexp(x2)dx represents the Dawson integral concerning the real number z. ti represents the i-th time of sampling when the device is functioning properly. The real-time reliability at moment t is expressed using conditional reliability [36]:(53)R(t|ti)=R(t)R(ti). Let xi represent the observed value of performance degradation at moment ti, and let w denote the set predefined threshold. There is (54)R(t|ti)=P(x(t)