MPE Mathematical Problems in Engineering 1563-5147 1024-123X Hindawi Publishing Corporation 175036 10.1155/2012/175036 175036 Research Article Investigation of a Stochastic Inverse Method to Estimate an External Force: Applications to a Wave-Structure Interaction Han S. L. Kinoshita Takeshi Wang Moran Institute of Industrial Science The University of Tokyo 4-6-1 Komaba Meguro-ku Tokyo 153-8505 Japan u-tokyo.ac.jp 2012 14 05 2012 2012 24 02 2012 09 04 2012 25 04 2012 2012 Copyright © 2012 S. L. Han and Takeshi Kinoshita. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The determination of an external force is a very important task for the purpose of control, monitoring, and analysis of damages on structural system. This paper studies a stochastic inverse method that can be used for determining external forces acting on a nonlinear vibrating system. For the purpose of estimation, a stochastic inverse function is formulated to link an unknown external force to an observable quantity. The external force is then estimated from measurements of dynamic responses through the formulated stochastic inverse model. The applicability of the proposed method was verified with numerical examples and laboratory tests concerning the wave-structure interaction problem. The results showed that the proposed method is reliable to estimate the external force acting on a nonlinear system.

1. Introduction

In the field of engineering, it is often very important to estimate external loads acting on dynamic structural systems for a design and analysis of the structural systems. The proper determination of external loads may contribute to extensive practical applications in control, monitoring, and analyzing engineering systems such as bridges under static or dynamic loading, floating structures on waves, and supply systems such as pipes.

However, for some structural systems, it is sometimes difficult to directly measure an external force for some reasons such as installation difficulty of the measurement devices and large magnitude forces. These difficulties led to a necessity of alternative methods to estimate external forces. Therefore, many methods have been developed by means of the inverse problem formalism, inferring an unobservable value from an observable value, except very few cases where the force can be measured directly.

Stevens  comprehensively reviewed force identification problems that arise both in laboratory experiments and field applications. Moreover, it was explained that there exist difficulties in obtaining a satisfactory result since the problem is generally unstable. General numerical methods, which proceed in a step-by-step fashion, fail to treat such a problem in a reliable and practical way. To tackle this difficulty, Zhu and Law  investigated a method based on the modal superposition and the regularization technique. Doyle  and Inoue et al.  studied a reconstruction technique based on the least squares method using the singular value decomposition (SVD). Gunawan et al.  developed a new method based on the quadratic spline approximation to solve the problem.

The aforementioned studies assume that structural systems of interests are linear. However, in practice, nonlinearities are always present in various forms for mechanical or structural systems. Therefore, it is necessary to consider these nonlinearities for reliable system design and analysis. To cite a few examples of the force identification for nonlinear systems, Ma and Ho  developed an inverse method that comprises the extended Kalman filter and the recursive least-squares estimator. Jang et al. [7, 8] presented an inverse procedure based on a deterministic classical regularization method, the Landweber iteration.

It is worth here noting that the aim of the above methods, regardless of considering a linear or nonlinear system, is to make the solution stable by modifying the original formulation. These methods yield only a single estimate of unknowns without quantifying associated uncertainties in the solution. Thus, these kinds of methods are often referred to as the deterministic inverse method.

Due to increasing demands on robustness and reliability as well as due to the rapid growth of computational capacity, it has been more and more important to solve problems under conditions of uncertainty. As a result, recently, problems considering uncertainties have thus attracted much research attention in diverse fields such as electric impedance tomography  and heat conduction [10, 11]. There are a few cases concerning the force identification with rigorous consideration of uncertainties. Wu and Law  recently developed a moving force identification technique based on a statistical system model. They proved the workability of the technique by applying it to the bridge-vehicle system with the Gaussian system uncertainties arising from road surface roughness. Wu and Law  also proposed a dynamic analysis method on the bridge-vehicle interaction problem with the non-Gaussian system uncertainties.

In this paper, an original method based on a stochastic inversion is developed to determine external forces acting on a nonlinear system from measurements of a system response. The method studied here comprises the following four parts: (i) construction of a mathematical inverse model linking an external force to a motion response, (ii) formulation of a stochastic inverse function that allows to determine an external force, (iii) design of probabilistic model to quantify various uncertainties arising from an insufficiency of information on parameters of interests and measurement errors, and (iv) computation of inverse solutions by designing computational tools with full probabilistic description.

In the first part of this paper, the mathematical formulation on the proposed method is presented. Complete description of the proposed method is illustrated through numerical examples. In addition, mathematical characteristics of the constructed model are also discussed. In the second part of this paper, the proposed method is validated through a practical application to a real problem regarding the wave-structure interaction. Experiments for a ship subjected to wave excitation are performed in the ocean engineering basin. Results of the analysis of the experimental data show that wave excitations can be successfully estimated from measurements of the response alone, using the proposed method.

2. Mathematical Formulation 2.1. Preliminaries

The following nonlinear physical model is first considered: (2.1)mx¨(t)+cx˙(t)+kx(t)+SNL(x˙,x;t)=f(t), where m is the mass, c is the damping, k is the restoring force, and SNL is the nonlinear system including nonlinear damping and restoring forces. Initial conditions for this system are given as (2.2)x(0)=α,x˙(0)=β. Let x1,x2 be basic solutions of the following homogeneous equation associated with (2.1): (2.3)x¨+2ξωnx˙+ωn2x=0, where ξ=c/2ωn and ωn=k/m. Then, the solution of (2.1) is given in the form (2.4)x=i=12wi(t)xi(t), where wi’s are continuous functions that satisfy (2.5)i=12w˙ixi=0. The first and second derivatives of (2.4) are (2.6)x˙=i=12wix˙i,x¨=i=12w˙ix˙i+wix¨i. Substituting (2.6) into (2.1) leads to (2.7)i=12wi(x¨i+2ξωnx˙i+ωn2xi)+i=12w˙ix˙i=f-SNL(x,x˙)m. Equation (2.3) implies that the first part of (2.7) is zero since xi’s are the solutions of (2.3). Then, (2.7) can be rewritten as (2.8)i=12w˙ix˙i=f-SNL(x,x˙)m. Equations (2.5) and (2.8) lead to the following system of equations: (2.9)[x1x2x˙1x˙2](w˙1w˙2)=(0f-SNLm). Therefore, w˙i’s are given by (2.10)(w˙1w˙2)=[x1x2x˙1x˙2]-1(0f-SNLm)=1W[x˙2-x2-x˙1x1](0f-SNLm). The inverse matrix in the above equation always exists because xi’s are linearly independent. Integrating (2.10) gives (2.11)w1(t)=-0tx2(τ)mW[f(τ)-SNL(x,x˙)]dτ+D1,w2(t)=-0tx1(τ)mW[f(τ)-SNL(x,x˙)]dτ+D2, where W is Wronskian defined as W(x1,x2)=|x1x˙2-x2x˙1| and Di’s are arbitrary additive constants. Substituting (2.11) into (2.4) leads to the following equation [14, 15]: (2.12)x=D1x1+D2x2+0tx1(τ)x2(t)-x1(t)x2(τ)mW[f(τ)-SNL(x,x˙)]dτ.

If x1, x2 are chosen by satisfying x1(0)=1, x1˙(0)=0 and x2(0)=0, x2˙(0)=1, then the constants D1, D2 become D1=α, D2=β. The preceding derivation is based on the concept of variation of parameters. The reader can refer to [14, 15] for more detailed information.

In (2.12), f-SNL can be considered as an effective force acting on the nonlinear vibrating system. Part of the importance of this integral formulation lies in its use not only for estimating the external forces f but also for determining the nonlinear system SNL that; characterizes dynamic oscillations. For the present study, the aim is to find the external force f acting on nonlinear system. Thus the nonlinear system SNL is assumed to be known.

2.2. A Stochastic Inverse Model for an Indirect Measurement

If a motion response x is measured for the time duration 0<t<T and the nonlinear system SNL is known, (2.12) can be rewritten [7, 8] as (2.13)g=F(f), where (2.14)g(t)=x-αx1-βx2+0TK(t,τ)SNL(x,x˙)dτ, and the operator is defined by (2.15)F(*)0TK(t,τ){*}dτ, with the kernel (2.16)K(t,τ)=x1(τ)x2(t)-x1(t)x2(τ)mW. The estimation of the external force can be achieved by solving (2.13) using the measured response data. This problem can be interpreted as an inverse problem of inferring the unknown value f from the observable physical quantity g.

Equation (2.13) is classified as the integral equation of the first kind , whose unknown is only in the integrand. According to the inverse problem theory [16, 17], the first-kind integral equation is typically ill posed in the sense of stability. The approximation of the integral equation of the first kind yields an ill conditioned system regardless of the choice of approximate methods such as the quadrature method and the Galerkin method with orthonormal basis functions. This ill-conditioning causes a loss of accuracy for direct factorization solution methods. Consequently, conventional or standard methods in numerical algebra cannot be used in a straightforward manner for solving numerical system of (2.13). Thus, it is necessary to adopt more sophisticated methods for stable solution procedure.

In this study, we formulate a stochastic inverse function to achieve a stable solution procedure. By considering all variables of interests to be random variables, (2.13) can be recast in the form of a stochastic function. Note that, from now on, the capital letter and lower letter refer to a random variable and its realization, respectively.

For an arbitrary nonempty sample space Ω, let us define the multivariate random variable F:ΩRN, F={F1,F2,,FN}, on the state space (2.17)ΓF=i=1NΓFiRN. Then the random variable G:ΩRM, G={G1,G2,,GM}, can be defined to be (2.18)G=F(F) for all possible subsets in Ω on the state space (2.19)ΓG=i=1MΓGiRM. Therefore, for each realization g={g1,g2,,gM}, of gΓG, the following relationship is given: (2.20)g=F(f). In the preceding definitions, N is the number of unknowns, M is the number of measurements, and R is real Euclidean space.

Equation (2.20) can be interpreted as the stochastic function corresponding to (2.13) because it is a function of random variables. Thus, the problem of finding f from a single realization g is the stochastic inverse problem.

2.3. Probabilistic Modeling of Stochastic Inverse Solution

A single realization g, which is the directly observable value, can be readily calculated through (2.14) when the response x is measured. Using this value gobserved, the unknown f of the stochastic inverse problem (2.20) can be expressed [17, 18]: (2.21)p(fgobserved)=p(gobservedf)p(f)p(gobserved), where p(gobservedf) is the likelihood function, p(f) is the prior probability density function, and p(gobserved) is the normalizing constant.

The probabilistic expression (2.21), which is known as the posterior probability distribution function (PPDF), can be considered as the solution to the stochastic inverse problem (2.20). However, (2.21) does not give any information on the desired value f in its present form. It is necessary to design a probability model according to the proper physical process to have a meaningful probabilistic expression. This modeling enables to quantify uncertainties in the solution associated with insufficient information on the system and measurement errors. From now on, the symbol g is used in place of gobserved:g=gobserved.

For the purpose of modeling, components in the right-hand side of (2.21) are separately considered. At first, it is worth noting that the normalizing constant p(g) is not necessary to be computed for sampling procedures [911, 17, 18] such as Monte Carlo simulation, which is used to estimate the target probability distribution. The PPDF (2.21) can thus be evaluated with the slightly simplified form: (2.22)p(fg)p(gf)p(f).

If a measurement error e in the observed data g is assumed to be independent, identically distributed Gauss random noise with zero mean and variance ve, the likelihood function p(gf) can be modeled in the form (2.23)p(gf)(ve)-M/2exp(-F(f)-g222ve), where ·2 refers to Euclidean norm.

Finally, it needs to consider the prior distribution p(f), reflecting the knowledge on the unknowns before the data g is measured. In this study, a pairwise Markov random field  is adopted for prior distribution modeling. The MRF model for the unknown f is of the form (2.24)p(f)γN/2exp(-12γfTWf), where the matrix WRN×N is determined by (2.25)Wij={nii=j,-1i~j,0otherwise. In (2.25), ni is the number of neighbors for the point i, i~j means that the points i, j are adjacent.

Using the models for the likelihood (2.23) and the prior (2.24), the PPDF (2.21) is given as (2.26)p(fg)(ve)-M/2exp(-F(f)-g222ve)γN/2exp(-12γfTWf). Here, it is worth discussing the role of parameters ve, γ. The scaling parameter γ controls the prior model. The variance ve explains the noise level in the measurement data g. These parameters play an important role in the solution procedure because the PPDF is characterized by these values as in (2.26). If the variance ve of the measurement data and the scaling parameter γ are given, the stochastic inverse solution is readily estimated with the probability model (2.26). However, in practice, it is difficult to know these parameters a priori unless experiments are repeatedly conducted enough to collect the data.

The choice of hyperparameters ve and γ sometimes causes another difficulty in solution procedure. However, this difficulty can be naturally resolved by introducing a hierarchical model for the PPDF [10, 11, 19]: (2.27)p(f,γ,veg)(ve)-M/2exp(-F(f)-g222ve)γN/2exp(-12γfTWf)p(γ)p(ve) with the prior density p(γ)γα1-1e-β1γ and p(ve)(ve)-α2-1e-β2/ve.

2.4. Computation of Stochastic Inverse Solution

Equations (2.26) and (2.27) obviously contain the information on the desired solution, that is, the external force f. To extract useful information, these probabilistic expressions need to be evaluated through a numerical sampling technique such as Markov chain Monte Carlo (MCMC) simulation [20, 21].

The MCMC methods are a class of algorithms for sampling from probability distributions based on constructing a Markov chain, which is characterized by the rule that the next state depends only on the current state. The generated samples from the algorithm are used to approximate the target distribution. Metropolis-Hastings and Gibbs algorithms [22, 23] are most widely used for sampling from multidimensional distributions, especially when the number of dimensions for unknowns is high. A sequence of random samples from the PPDF (2.27) can thus be obtained by the following hybrid algorithms [22, 23].

Initialize f0={f10,f20,,fN0}, γ0 and ve0

For i=0:NMCMC-1.

Sample f1i+1~p(f1f2i,f3i,,fNi,γi,vei)

Sample f2i+1~p(f2f1i+1,f3i,,fNi,γi,vei)

Sample fNi+1~p(fNf1i+1,f2i+1,,fN-1i+1,γi,vei)

Sample u1~U(0,1)

Sample γ*~q(γ*γi)

if u1<min{1,p(γ*fi+1,vei)q(γiγ*)p(γifi+1,vei)q(γ*γi)}

γi+1=γ*

else

γi+1=γi

Sample u2~U(0,1)

Sample ve*~q(ve*vei)

if u2<min{1,p(ve*fi+1,γi)q(veive*)p(veifi+1,γi)q(ve*vei)}

vei+1=ve*

else

vei+1=vei.

In the above algorithm, NMCMC is the total number of samples, U is the uniform distribution, and q(X*Xi) is an easy-to-sample proposal distribution with X=γ or ve.

3. Numerical Investigations

In this section, we want to numerically investigate the proposed method to determine external forces acting on nonlinear system through a particular numerical example.

3.1. Numerical Example

As a numerical example, the following physical model is considered: (3.1)mx¨+cLx˙+cNLx˙|x˙|+kLx+kNLx3=f(t), with the physical values; m=1,cL=1,cNL=1,kL=10, and kNL=10. Comparing (3.1) with (2.1) leads to SNL(x˙,x;t)=cNLx˙|x˙|+kNLx3. For the purpose of testing the proposed method, an external force, which is a repeated nonharmonic forcing, is given by (3.2)f(t)=acos(ωt)+κa23-ϑ24ϑ3cos(2ωt), where the values are κ=0.1,a=2,ϑ=0.7616, and ω=0.9039.

At first, the motion responses were simulated with zero initial conditions x(0)=0 and x˙(0)=0 by solving (3.1) through the numerical integration scheme such as the fourth-order Runge-Kutta method. Figure 1 shows the exact external force in (3.2) and the corresponding responses. The dynamic responses in Figure 1 were assumed as the measured data for the inverse procedure. For the numerical example, the number of measurements M and unknown N was taken to be (201,201).

External force and corresponding responses.

The aim is now to estimate the external force from the measurement of system responses using the proposed stochastic estimation method in Section 2. With the measurements in Figure 1, observable quantities g were computed through (2.14). To examine the effect of the level of noises in data, two noisy data with the noise level of σ=0.0158 and σ=0.0316 were generated by adding independent random noises e~N(0,σ2) to the generated responses. Figure 2 shows the exact and two noisy data for g.

Exact and two noisy data for g.

σ = 0.0158

σ = 0.0316

3.2. Singular Value Analysis

In this subsection, the characteristic of the deterministic inverse model (2.13) is examined by singular value analysis , which provides quantitative measures for assessing the numerical system. For this purpose, the discretized or approximated form of (2.13) can be written as (3.3)g(tj)=iΔτK(tj,τi)f(τi),K(tj,τi)=0whentjτi, or simply (3.4)g=Kf. The matrix K is thus expressed as a lower triangular Toeplitz matrix: (3.5)K=[k1k2k1kNkN-1k1]. By using the singular system [16, 17, 24], the matrix K can be decomposed as (3.6)K=UΣVT=i=1NuiσiviT, where U=(u1,u2,,uN) is an unitary matrix, the matrix Σ is an diagonal matrix with nonnegative real numbers σi, which are known as the singular values of K, and VT is the conjugate transpose of V=(v1,v2,,vN). With this singular system, the inverse solution f can easily be written in the form (3.7)f=i=1NuiTgσivi.

Equation (3.7) is known as the pseudo inverse , which is also referred to as a generalized inverse. The purpose of constructing a pseudo inverse in (3.7) is to obtain a systematic expression that can serve as the inverse in some sense for a wider class of numerical systems than invertible ones. The inverse solution was computed through the pseudo inverse (3.7) with the observable quantity g. The results are shown in Figure 3. It can be seen that the computed solutions are totally unstable and thus useless.

Numerical solutions via the pseudo inverse.

σ = 0.0158

σ = 0.0316

The cause and the degree of instability in the solution can be understood by analyzing the behavior of singular systems, more specifically, each component in the right-hand side of (3.7). Figure 4 shows behaviors of |uiTg|, σi, and the ratio |uiTg|/σi. It can be found that the coefficients |uiTg| become larger than most of singular values σi. Thus, the inverse solution f, which is the sum of the ratio uiTg/σi, is dominated by the small singular values. This explains the reason for significant sign changes in Figure 3. The instability, frequent sign changes, may be accelerated and amplified with the increasing amount of noise in data. This can be easily checked by comparing Figures 3(a) and 3(b). The magnitude of the solution in Figure 3(b) is larger than the one in Figure 3(a).

Picard plots for the singular system with two noisy data in Figure 2.

σ = 0.0158

σ = 0.0316

In summary, for a stable solution, the approximated system should satisfy the condition that coefficients |uiTg| should decay to zero faster than singular values σi. This condition is known as the discrete Picard condition . The figure describing the behavior of singular values is also known as the Picard plots. Since the deterministic inverse model (2.13) fails to satisfy this condition, it is ill posed in the sense of stability. Therefore, the conventional or standard methods in numerical algebra cannot be used to solve the inverse model (2.13).

3.3. Solution to Stochastic Inverse Model

As pointed out, the inverse model (2.13) loses its stability when it is approximated. Thus, the usual numerical scheme cannot be used in straightforward manner. For stable solution procedure, the stochastic inverse method described in Section 2 is followed here by transforming the original deterministic inverse model into stochastic state space.

Accordingly, the inverse solution, unknown external force f, has the form as the PPDF p(f,γ,veg) in (2.27). To extract necessary information from the PPDF, the MCMC algorithm described in Section 2.4 is used. For the algorithm, the values γ,ve, and f were initialized as γ0=1, ve0=0.1, and zero vectors for f0, respectively. The pairs of parameters (α1,β1) and (α2,β2) in (2.27) determine the initial shape of prior densities p(γ) and p(ve). In the MCMC algorithm, these prior densities are greatly refined through the conditioning on the data regardless of the initial choice. In this study, the pairs of parameters (α1,β1) and (α2,β2) were taken to be (1,10) and (1,10) to impose nonnegativity. The total number of samples NMCMC for the algorithm was taken to be 50,000, and the last 25,000 realizations were used to estimate statistics such as mean and standard deviation for external force f since it allows the beginning of the Markov chain. This beginning of chain is often called the burn-in. The scale parameter σX for a proposal distribution q(X*Xi)exp(-(Xi-X*)2/2σX2) can affect the efficiency of the Markov chain. In this study, σX=0.05 was used to produce a sample with reasonably low autocorrelation.

Numerical results are shown in Figure 5. The upper and lower dotted lines denote the 95% credible interval. The credible interval quantifies the degree of uncertainties in the solution. Comparing Figures 5(a) and 5(b) implies that the degree of uncertainties depends on the noise level in the measurement data. The credible interval in Figure 5(a) is narrower than the one in Figure 5(b). It can also be seen that the posterior mean E{f} is fairly accurate compared with the exact solution, regardless of the level of noise.

Numerical solutions via the proposed method.

σ = 0.0158

σ = 0.0316

Figure 6 shows examples of trace plot with respect to the number of sampling for the case σ=0.0158. These points f40 and f140 are highlighted with arrows in Figure 5(a). One of the features that can be observed in Markov chain is that it takes a while to properly sample the target distribution. The evolution of components may be a good tool to check if the burn-in begins. The curve in Figure 6 clearly shows the burn-in phase. This implies that the chain has reached its stationary state and thus works properly.

Examples of trace plots for the MCMC simulation and their posterior distributions.

4. A Particular Application: Estimation of Wave Exciting Forces Acting on a Ship

The results in the previous section were fairly good. However, it has a limitation that numerical simulations were conducted based on the synthesized data. In this section, to ensure the practicability, the proposed method is also applied to the real practical problem regarding the problem of estimating wave exciting forces on a ship in roll motion. The estimation of wave exciting forces is one of the most important issues in offshore structure design since waves are the most important environmental phenomena that affect the system stability of offshore structures.

4.1. Motion Equation of a Ship in Roll Motion

The mathematical model of ship rolling in waves (Figure 7) is generally expressed as the following nonlinear differential equation:(4.1)(I+A44)ϕ¨+B(ϕ˙)+C44ϕ=M4(t), where I is the actual mass moment of inertia, A44 is the added mass moment of inertia, B is the nonlinear roll damping function, C44 is the coefficient of restoring moment, ϕ is the roll response, and M4 represents a wave exciting moment acting on a ship. The sum of I and A44 is often referred to as the virtual mass.

Roll motion of a ship in waves.

Equation (4.1) looks similar to the general nonlinear physical model (2.1). However, the mechanism behind this is somewhat complicated because the interaction between the ship hull and the surrounding water should be taken into account. This interaction causes the mechanism for damping and restoring forces in ship motions. Figure 8 illustrates an analogy of the mechanical system and the motion of a ship in water.

Analogy of (a) the general mechanical system and (b) the ship motion in water.

More specifically, forces acting on ships can be decomposed into hydrostatic and hydrodynamic forces. Hydrostatic force is the force induced by hydrostatic pressure (the static buoyancy force) exerted by a fluid at equilibrium due to the force of gravity. This hydrostatic force is related to the restoring mechanism C44 of a ship in water.

Hydrodynamic forces can be considered as the forces resulted from the pressure field disturbed by a moving body. Hydrodynamic forces are divided into three main components. The first is the Froude-Krylov force due to the pressure field in the incident wave. The second is the diffraction force due to the scattering of the incident wave field. The last is the radiation force due to the waves generated by the oscillatory motions of the body. Among them the radiation force is related to the added mass A44 and damping mechanism B of a ship motion. The wave exciting force M4 is concerned with the Froude-Krylov force and the diffraction force. For more detailed information, the reader can refer to .

In this section, the attention is focused on the estimation of the wave excitation M4. For an application, all physical values are separately identified a priori except the wave excitation.

4.2. Experimental Setup

For the purpose of practical application, laboratory tests regarding ship rolling motions in waves were conducted in the Ocean Engineering Basin at The University of Tokyo. The basin, which is often called the towing tank, is designed to perform tests related to various kinds of floating structures. Figure 9 shows the test model and its body plan illustrating the shape of hull form. Table 1 summarizes particulars of the test model.

Particulars of the model.

 Length  Lpp 2.500 m Breadth B 0.387 m Draft D 0.132 m Displacement volume ∇ 0.110 m3 GM 0.074 m Resonance frequency ωn 6.905 rad/s

Illustration of (a) the test model for experimental study and (b) its body plan.

Figure 10 shows an overview of experimental setups. Regular wave fields with a fixed frequency were generated by using the plunge type wave-maker. Then motion responses were measured with the potentiometer attached to the center of gravity of the test model. The incident wave was also recorded by far and near wave probes from the test model. During experiments, the test model was located in the middle of the basin to minimize the effects of reflected waves. For simplicity, other motions of the test model were restricted except the roll motion.

Overview of experimental setup.

Figure 11 shows a typical example of the measured data. We used the stationary region of the data highlighted with red box in Figure 11 to estimate wave exciting moment. For demonstration purposes, three frequency regions were chosen-low, high, and near resonance frequency.

Typical example of time history of data for a test model: wave elevation measured at far wave-probe (a) and at near wave-probe (b) and the corresponding responses (c).

4.3. Analysis of the Experimental Data

As a first attempt, the proposed method was applied to the roll response data when the frequency ω=6.465rad/s in Figure 12. The data was measured with the sampling period Δt=0.01 s. Using this data, the observable quantity g in Figure 13 was calculated through (2.14). This quantity leads to the construction of the PPDF p(M4,γ,veg) for wave exciting moment M4 as in (2.27).

The measured roll response data with ω=6.465rad/s.

Observable quantity g for the response data with ω=6.465rad/s.

In the same way used in the previous section, the information on unknown wave exciting moment M4 was extracted via MCMC algorithm. Without loss of generality, the values γ,ve, and M4 were initialized as γ0=1, ve0=0.1, and zero vectors for M40. The pairs of parameters (α1,β1) and (α2,β2) were also taken to be (1,10) and (1,10).

Figures 14 and 15 show, respectively, the solution obtained by the pseudo inverse (3.7) and the proposed stochastic inverse method. It can be seen that the stochastic inverse solution is stable while the pseudo inverse solution loses its stability and accuracy. Moreover, the result in Figure 14 shows a clear periodicity as expected from the periodic excitation. It is also observed that the stochastic inverse solution provides a credible interval explaining uncertainties in unknown quantity M4.

Wave exciting moment obtained by pseudo inverse for the response data with ω=6.465rad/s.

Wave exciting moment obtained by stochastic inversion for the response data with ω=6.465rad/s.

For the purpose of validation, roll response was resimulated using the posterior mean of the identified wave exciting moment since the exact solution for M4 cannot be known a priori for the real practical problems unlike numerical simulations. Figure 16 shows the comparison between the re-simulated and the measured responses. The result shows that the re-simulated response is well coincident with the measured one. This proves the validity of the estimated wave exciting moment M4 for this case.

Comparison of the re-simulated response with the measured one for the response data with ω=6.465rad/s.

The same procedure was also applied to the other response data when ω=3.768rad/s and ω=8.796rad/s. All results are shown in Figures 17 and 18, respectively. In both cases, it can be seen that the re-simulated responses using the estimated wave exciting moment well agree with those of measurements.

Numerical results for the response data when ω=3.768rad/s: (a) measured response data, (b) observable value g, (c) wave exciting moment obtained by the proposed method, and (d) comparison of the re-simulated response with the measured one.

Numerical results for the response data when ω=8.796rad/s: (a) measured response data, (b) observable value g, (c) wave exciting moment obtained by the proposed method, and (d) comparison of the re-simulated response with the measured one.

When calculating the wave exciting force acting on a ship, the strip theory  is widely used in the field of naval architecture and ocean engineering. The strip theory is a popular approximation tool to calculate hydrodynamic forces for ships or marine vehicle. The key idea behind this theory is that hydrodynamic forces can be estimated based on a stripwise computation by slicing the entire body of a ship in the longitudinal direction. Resultant wave force is then expressed as the sum of loads that is calculated in each segment on the sliced body.

To ensure the validity of the present method, a qualitative comparison is also made in Figure 19 for wave exciting moments estimated by three different sources: the proposed method, the strip theory, and the experiments. The horizontal axis represents the ratio of the length of test model and wave length. The vertical axis represents nondimensional wave exciting moment. As can be seen in Figure 19, all the results well agree with each other. However, it can be found that there exist slight differences between the proposed method and that of the strip theory when λ/L is small or large. This may be mainly because of the linear assumption in the strip theory.

Qualitative comparison of wave exciting moment.

4.4. Irregular Wave Excitation

It is worth emphasizing that the proposed method is not limited to the case of the stable sinusoidal responses. It is also applicable to the estimation for any form of wave excitation, provided that the motion response is properly recorded. To test this, we also conducted another experiment regarding the ship motion subjected to the wave excitation induced by irregular waves.

Figure 20 shows a time history of measurements. It can be observed that the response is not simple harmonic since the incident wave is irregular. This roll response was then used, according to the proposed stochastic inverse procedure, to estimate the irregular wave excitation. Figure 21 shows the estimated results. Finally, using this estimated wave excitation, the motion was re-simulated numerically in the same manner described in the previous examples. The result in Figure 22 well coincides with the measurement. This proves the applicability of the proposed method to other types of waveforms.

Time history of measurements for the case of irregular wave.

Estimated wave exciting moment through the proposed method.

Comparison of the re-simulated response with the measured one.

5. Conclusions

A new method has been proposed to estimate the external forces acting on a nonlinear vibrating system, based on formulating the stochastic inverse model from the measured roll response data. The method has been evaluated through an analysis of some digitally simulated data for a numerical example. It has been shown that the proposed method can be used to not only identify the external force but also detect uncertainties in solution arising from measurement error or insufficient information on the system.

To ensure applicability, the wave-structure interaction problem regarding the roll motion of a ship has also been treated as a practical application. To this end, a series of experiments were conducted in the ocean engineering basin. The analysis results demonstrate that the proposed method has been successfully applied to estimate the various types of wave excitation. The estimated results were qualitatively and quantitatively good in all cases.

Acknowledgments

The authors would like to thank the Editor and anonymous reviewers for their valuable comments, suggestions, and constructive criticism, which were very helpful in improving the paper substantially.

Stevens K. K. Force identification problems (an overview) Proceedings of the Proceedings of the SEM Conference on Experimental Mechanics June 1987 Houston, Tex, USA 14 19 Zhu X. Q. Law S. S. Moving load identification on multi-span continuous bridges with elastic bearings Mechanical Systems and Signal Processing 2006 20 7 1759 1782 2-s2.0-33646696622 10.1016/j.ymssp.2005.06.004 Doyle J. F. Force identification from dynamic responses of a bimaterial beam Experimental Mechanics 1993 33 1 64 69 2-s2.0-0027556774 10.1007/BF02322553 Inoue H. Ikeda N. Kishimoto K. Shibuya T. Koizumi T. Inverse analysis of the magnitude and direction of impact force JSME International Journal, Series A 1995 38 1 84 91 2-s2.0-0028977155 Gunawan F. E. Homma H. Morisawa Y. Impact-force estimation by quadratic spline approximation Journal of Solid Mechanics and Materials Engineering 2008 2 1092 1103 Ma C. K. Ho C. C. An inverse method for the estimation of input forces acting on non-linear structural systems Journal of Sound and Vibration 2004 275 3–5 953 971 2-s2.0-3042656975 10.1016/S0022-460X(03)00797-1 Jang T. S. Baek H. Hana S. L. Kinoshita T. Indirect measurement of the impulsive load to a nonlinear system from dynamic responses: inverse problem formulation Mechanical Systems and Signal Processing 2010 24 6 1665 1681 2-s2.0-79952313303 10.1016/j.ymssp.2010.01.003 Jang T. S. Baek H. Choi H. S. Lee S. G. A new method for measuring nonharmonic periodic excitation forces in nonlinear damped systems Mechanical Systems and Signal Processing 2011 25 6 2219 2228 2-s2.0-79957480481 10.1016/j.ymssp.2011.01.012 Kaipio J. P. Kolehmainen V. Somersalo E. Vauhkonen M. Statistical inversion and Monte Carlo sampling methods in electrical impedance tomography Inverse Problems 2000 16 5 1487 1522 10.1088/0266-5611/16/5/321 1800606 ZBL1044.78513 Wang J. Zabaras N. A Bayesian inference approach to the inverse heat conduction problem International Journal of Heat and Mass Transfer 2004 47 17-18 3927 3941 2-s2.0-2942753787 10.1016/j.ijheatmasstransfer.2004.02.028 Ma X. Zabaras N. An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation method Inverse Problems 2009 25 3 10.1088/0266-5611/25/3/035013 035013 2480183 ZBL1161.62011 Wu S. Q. Law S. S. Moving force identification based on stochastic finite element model Engineering Structures 2010 32 4 1016 1027 2-s2.0-77649189814 10.1016/j.engstruct.2009.12.028 Wu S. Q. Law S. S. Dynamic analysis of bridge with non-Gaussian uncertainties under a moving vehicle Probabilistic Engineering Mechanics 2011 26 2 281 293 2-s2.0-79952196008 10.1016/j.probengmech.2010.08.004 Hildebrand F. B. Advanced Calculus for Applications 1976 Prentice Hall Murdock J. A. Perturbations 1991 New York, NY, USA John Wiley & Sons xviii+509 A Wiley-Interscience Publication 1107517 Kirsch A. An Introduction to the Mathematical Theory of Inverse Problems 1996 120 New York, NY, USA Springer x+282 Applied Mathematical Sciences 1479408 Kaipio J. Somersalo E. Statistical and Computational Inverse Problems 2005 160 New York, NY, USA Springer xvi+339 Applied Mathematical Sciences 2102218 Besag J. Green P. Higdon D. Mengersen K. Bayesian computation and stochastic systems Statistical Science 1995 10 1 3 41 1349818 10.1214/ss/1177010123 ZBL0955.62552 Jin B. Zou J. Hierarchical Bayesian inference for ill-posed problems via variational method Journal of Computational Physics 2010 229 19 7317 7343 10.1016/j.jcp.2010.06.016 2677781 ZBL1198.65189 Andrieu C. De Freitas N. Doucet A. Jordan M. I. An introduction to MCMC for machine learning Machine Learning 2003 50 1-2 5 43 2-s2.0-0037262814 10.1023/A:1020281327116 Gilks W. R. Richardson S. Spiegelhalter D. Markov Chain Monte Carlo in Practice 1996 London, UK Chapman & Hall xviii+486 Interdisciplinary Statistics 1397966 Chib S. Greenberg E. Understanding the metropolis-hastings algorithm The American Statistician 1995 49 327 335 Casella G. George E. I. Explaining the Gibbs sampler The American Statistician 1992 46 3 167 174 10.2307/2685208 1183069 Hansen P. C. Rank-Deficient and Discrete Ill-Posed Problems 1998 Philadelphia, Pa, USA Society for Industrial and Applied Mathematics (SIAM) xvi+247 SIAM Monographs on Mathematical Modeling and Computation 1486577 Beck R. F. Cummins W. E. Dalzel J. F. Mandel P. Webster W. C. Motions in waves Principles of Naval Architecture 1989 3 1 190