Robust identification of impact force acting on composite structures via a single FBG sensor using variational mode decomposition Authors: Jian Huang, Xiaoming Yang, Quanxin Jiang, Li Zhang Published: 2026 ABSTRACT AbstractIdentifying impact force on composite structures is crucial for structural integrity assessment, yet it remains a formidable challenge under operational conditions where impact-induced vibrations are corrupted by ambient noise and other dynamic loads. To address this limitation, this paper develops a robust two-stage identification framework that requires only a single Fiber Bragg Grating (FBG) sensor. The core innovation of the proposed methodology lies in leveraging Variational Mode Decomposition (VMD) to decouple the complex, noisy strain signal into its constituent modal strain responses, effectively isolating the impact’s signature from operational interference. From these purified modal responses, the uncorrected modal constant vector is extracted to serve as a unique fingerprint for impact localization over a grid of predefined potential impact points. The location is accurately determined by evaluating the cosine similarity between this vector and a pre-computed modal matrix of potential impact sites. Subsequently, with the location identified, the impact force’s time history is reconstructed using a sparse deconvolution model, which is efficiently solved by the Alternating Direction Method of Multipliers (ADMM). Experimental validation on a composite plate with simplified airfoil geometry confirms the method’s efficacy. The results demonstrate a localization success rate across the ten potential impact points exceeding 90% even under simulated operational conditions (harmonic and white noise excitation) and accurate force reconstruction, demonstrating the framework’s promising robustness, efficiency, and practical applicability for in-service structural health monitoring (SHM).Keywords: Impact force identification, Composite structure, Variational mode decomposition, FBG, Structural health monitoringSubject terms: Engineering, Mathematics and computing Introduction Composite materials have been widely used in the manufacture of critical structural components within aviation, marine, and energy industries because of their exceptional mechanical properties, including high specific strength and specific stiffness1,2. However, during maintenance or operational phases, these composite structures are susceptible to external impact loads, resulting in internal damage that is undetectable through visual inspection3,4. Conventional approaches employ non-destructive testing (NDT) techniques, such as ultrasonic detection and infrared thermography, to assess impact-induced damage offline. However, these approaches are often time-consuming and financially prohibitive. Consequently, identifying external impact forces and leveraging their characteristics to evaluate structural integrity has emerged as a viable alternative. Two primary methodologies exist for impact force identification (ImFoId) on composite structures: direct measurement and indirect measurement. In practical engineering applications, direct force measurement is frequently infeasible due to challenges associated with sensor placement at targeted locations. Hence, indirect identification of impact loads via structural vibration responses (e.g., strain, acceleration) has emerged as the predominant methodology over the past three decades5,6. Identifying impact force through structural vibration response is a typical inverse problem, which can be solved in both time and frequency domains. The most refined approach for ascertaining impact forces within the time domain involves the utilization of the triangulation technique. This technique employs an intricately arranged network of sensors to capture the Time of Arrival (TOA) delays of elastic waves at designated sensor locations. Upon acquiring the TOA delays, and by integrating them with the relative distances between the sensors, one can efficiently calculate the coordinates of the impact source7–9. Despite the principle of triangulation technique is straightforward and easy to implement, however it demands an exceedingly high sampling rate, significantly elevating the sampling costs of the system10. Moreover, the triangulation technique is predicated upon a foundational understanding of the propagation velocity of elastic waves within the given structure. In the context of composite structures, the inherent anisotropy and wave dispersion present considerable challenges in precisely determining both the speed and trajectory of elastic wave propagation, potentially resulting in erroneous localization results11. In order to overcome the shortcomings of triangulation techniques, ImFoId methods using the complete vibration response of the structure have been developed12,13. The identification of impact force through complete vibration response is an ill-posed inverse problem characterized by a rank-deficient transfer function matrix exhibiting low condition number, rendering it both numerically unstable and computationally prohibitive14. Among conventional inverse solution strategies, pseudoinverse-based methodologies and least squares optimization have demonstrated limited efficacy due to their inability to address the inherent ill-conditioning and noise sensitivity of impact response data15. To address these limitations, regularization frameworks have emerged as dominant approaches for impact force identification (ImFoId), with Tikhonov regularization (l2-norm penalty) and sparse regularization (l1-norm penalty) being prominently applied across time and frequency domains to stabilize the solution process16,17. These techniques effectively mitigate the ill-posedness by introducing controlled bias through norm-based penalty terms, thereby enabling robust estimation of impact load characteristics. However, the implementation of regularization techniques-based ImFoId methods remain significant challenges in real world applications. Firstly, the dimensionality of the transfer matrix between impact forces and structural responses escalates significantly with enlarged structural dimensions, leading to an excessively time-consuming deconvolution process. Secondly, to ameliorate the rank deficiency of the transfer function matrix, regularization techniques necessitate a dense array of sensors on structure, which imposes a significant increase in both the sampling cost and wiring complexity for the ImFoId system. These interrelated challenges underscore the need for hybrid methodologies that balance solution accuracy with computational efficiency, potentially through model order reduction techniques or physics-informed machine learning frameworks18. In consideration of the disparities in vibration modes among various potential impact locations, the discriminating modal constant vector can be served as a signature for impact localization19–21. Once the impact location is determined, the time history of the impact force can be rapidly reconstructed through either basis function fitting or regularization methods22. Generally, this approach only requires merely an accelerometer or a strain gauge to achieve the global ImFoId of the structure, and its computational efficiency is significantly higher than that of the simultaneous impact force identification method that depends on regularization techniques23,24. However, all the existing reports conduct ImFoId on the assumption that the structure is only subjected to impact loads25. When the structure is in an operational state, it is often exposed to different kinds of dynamic loads, which makes the frequency components of its vibration response signals extremely complex. As a result, accurately extracting the discriminating modal constant vector from the structural vibration response under actual operating conditions remains a challenging task. Although the single-sensor approach proposed by Goutaudier et al24. demonstrates high efficiency, its performance under significant operational noise has yet to be verified. Similarly, the method in22 is primarily concerned with efficiency under ideal conditions. To address these shortcomings, the critical challenge addressed in this paper is maintaining identification accuracy when impact signals are contaminated by other dynamic loads. To improve the accuracy of ImFoId results under operational conditions, a robust ImFoId method based on variational mode decomposition (VMD) is proposed in this paper. The core innovation lies in integrating VMD as a physics-guided preprocessing step and the requirement of only a single low sampling rate fiber Bragg grating (FBG) sensor for implementation. By selectively extracting the modal strain responses corresponding to the structure’s natural frequencies, VMD effectively serves as an advanced modal filter that isolates the impact signature from operational interference. The reason VMD is selected is that it outperforms other methods, such as empirical mode decomposition (EMD) and wavelet transform, in decomposing complex structural response signals, especially in the context of structural health monitoring (SHM) tasks26. Specifically, once an impact event is detected, the VMD is employed to determine the modal strain responses of the structure. Subsequently, the uncorrected modal constant vector is extracted from the modal responses via the envelope fitting technique. By evaluating the collinearity between the uncorrected modal constant vector and all the columns of the system modal matrix, the location of the impact among a set of potential locations can be rapidly identified. Based on the localization result, a sparse deconvolution model for reconstructing the time history of the impact force is established and solved using the alternating direction method of multipliers (ADMM). This capability enables the robust application of the modal constant vector method using only a single sensor, representing a significant step toward practical in-service SHM applications. To validate the performance and robustness of the proposed method, some experimental verifications are carried out on a laminated composite plate with a simplified airfoil geometry under simulated operational conditions. The remainder of this paper is organized as follows: Section "Theoretical background" introduces the theoretical background, including the sensing principle of FBG, the forward transfer model and VMD. Section "Impact force identification" investigates the proposed ImFoId framework, and a procedure is described for identifying the impact from the vibration measurements of a single FBG sensor. Section "Experimental validation" presents the experimental validations, which is used as a reference to demonstrate and validate the effectiveness of the proposed method. Finally, Section "Conclusion" concludes the paper. Theoretical background Principle of FBG sensing FBG sensors are widely used in ImFoId in special scenarios because of their inherent explosion-proof, electromagnetic interference resistance and chemical corrosion resistance27. In FBG, light beams of specific wavelengths are reflected, whereas other wavelengths remain essentially unaffected. According to Bragg’s law, the wavelength of the reflected light, λ, can be expressed as follows: where neff is the effective refractive index of the fiber, Λ is the grating period of the FBG. When the FBG sensor is subjected to physical perturbations such as stress or temperature, the Bragg wavelength shift (Δλ) can be expressed in terms of axial strain (ε) and temperature variation (ΔT) as follows: where Pe is the photo-elastic constant of the fiber. If the material of the fiber is fused silica, Pe = 0.22. The symbols α and ξ respectively denote the thermal expansion coefficient and the thermo-optical coefficient of the fiber. If the environmental temperature remains a constant, implying that the temperature variation, ΔT, is equal to zero, Eq. (2) can be reformulated as follows: where αε is defined as the sensitivity of the FBG sensor. In this study, the central wavelength of the employed FBG is 1558nm, and the value of αε is calculated as 1.2pm/με. The forward mathematical model The forward modeling problem in plate structure dynamics usually involves determining the amount of motion at each moment, and its main purpose is the evaluation of the vibration state. As shown in Fig. 1, consider a thin plate-like structure undergoes a transverse impact at some point that are not blocked by the boundary conditions. By using the finite element method to build the dynamic model, the differential equation of motion of the plate can be expressed as: where M ∈ ℝN×N, C ∈ ℝN×N and K ∈ ℝN×N are the mass matrix, damping matrix and stiffness matrix of the discrete system, respectively. , and x(t) are respectively the acceleration, velocity and displacement vectors of all nodes at time t. N is the number of physical degree of freedoms (DOFs). f(t) represents the time history of the impact force and L ∈ ℝN×1 denotes the load influence matrix associated with f(t). The solution of Eq. (1) can be derived through either the method of direct integration or the application of the modal superposition principle. The current study’s explorations are founded upon the latter approach. The modal superposition principle entails the representation of the displacement x(t) as a sum of n modal constituents28: where n represents the total number of mode shapes with n ≤ N. ur(t) and ϕr are the rth order modal coordinate (generalized) and rth mass-normalized displacement mode shape vector corresponding to the modal frequency ωr, respectively. Φ = [ϕ1, ϕ2, ∙∙∙, ϕr, ∙∙∙, ϕN] is the displacement mode matrix and u= [u1(t), u2(t), ∙∙∙, ur(t), ∙∙∙, uN(t)]T is the modal coordinate matrix. Schematic diagram of a plate structure undergoes an impact event. Under the assumption of the Rayleigh damping, i.e., C=αM+βK (α and β are constants to the evaluated experimentally), and then introducing the orthonormal relations ΦTMΦ = I, ΦTKΦ = diag (), the system governing equation can be split into a set of distinct equation as follows: where ζr and ωr are the r-th order modal damping ratio and circular frequency, respectively. fr(t)= ϕr∙Lf(t) is the modal force. Suppose the impact occurs at node q, Eq. (6) can be solved by Duhamel’s integral as follows: where is the r-th damped natural angular frequency, ωr and ζr are the rth natural angular frequency and modal damping ratio of the system, respectively. According to the principle of energy balance and the premise of linear elasticity29,30, the energies of displacement and strain should be in balance. Hence, the modal coordinates for strain response and displacement response are identical, and the strain response can be expressed as: where Ψ = [ψ1, ψ2, ∙∙∙, ψr, ∙∙∙, ψn] is the strain mode matrix and ψr denotes the rth mass-normalized strain mode shape vector. Substituting Eq. (7) into Eq. (8), the strain responses at the measurement point, p, can be expressed as: Because the instantaneous nature of impact force, an ideal impulse model with a force intensity of fq can be employed to approximate the time history of the impact force, thus yielding f(t) = fqδ(t). Based on this assumption, Eq. (9) can be rewritten as follows: where represents the rth modal response. Depending on the location of the impact, q, Eq. (10) indicates that the n vibration modes are not excited in uniform proportions via the ponderations ϕr(q). If the impulse occurs at the modal node of the sth mode, it will not contribute to the response (ϕs(q) = 0). The intensity of the impact, fq, does not alter the proportions in which the modes are excited, it uniformly amplifies each modal ponderation22. Consequently, the vector , referred to as uncorrected modal constant vector (Nm is the number of required modes to make the direction of Γ unique), can be employed as the signature for impact localization. The extraction of Γ for impact localization has been extensively documented in various studies23,24. Nevertheless, these methodologies consistently depend on modal constants derived from the complete impact response. When the structural impact response is contaminated with noise at other frequency components, substantial deviations in the fitting outcomes may ensue. Furthermore, modal truncation errors can also contribute to inconsistencies in the results. Therefore, to improve the accuracy of impact localization, it is feasible to directly extract the modal constant vector from the modal responses. Variational mode decomposition VMD represents a groundbreaking advancement in the realm of signal processing31. Introduced by Dragomiretskiy and Zosso in 2014, VMD is predicated on the principle of decomposing a complex signal into a series of intrinsic mode functions (IMFs), each characterized by specific frequency bands. Unlike traditional methods, VMD achieves this by solving a constrained variational problem, which balances fidelity of the mode reconstruction while maintaining their sparsity and frequency band concentration. In VMD, the equation of kth IMF (gk(t)) is expressed as follows: where bk(t) is the instantaneous amplitude, θk(t) is a non-decreasing phase function and the variations in the center frequency ωk(t)=dθk(t)/dt and bk(t) are much slower than those of θk(t). The center frequency and the respective bandwidth of each mode can be ascertained through an iterative search for the extremum of the variational mode. In this manner, an optimal decomposition effect can be achieved. The constrained variational equation is formulated in the following form: where M is the number of modes obtained by decomposition, δ(t) represents the impulse function, s(t) is the original signal. i is the imaginary unit and satisfies i2=−1. In order to solve the above constrained variational problem, the quadratic penalty parameter α and Lagrange multiplication operator λ(t) can be introduced into Eq. (12) and an unconstrained problem is established: The key steps involved in VMD are summarized as follows: (1) Set the value of , , and ni to zero. (2) Update the value of gk and ωk through Eq. (14). where , and are the Fourier transform of s(t), and λ(t), respectively, ω is the frequency and ni is the number of iterations. (3) Update the value of λ through gradient descent method: (4) Repeat step (2) and step (3) to continuously update the parameter values of gk, ωk and λ until the iteration termination condition is satisfied. where σ is the termination parameter and it is set to 1 × 10−6 in this paper. The performance of VMD is dependent on key parameters such as the number of modes M and the penalty parameter α which must be carefully selected based on the signal characteristics. Impact force identification Impact localization Under the assumption of linear time-invariant system, the strain response of plate structures can be regarded as the superposition of responses induced by impact force and other external forces. Consequently, the results of ImFoId can be significantly enhanced by isolating impact-induced response components from complete strain signals and subsequently utilizing these decomposed components for ImFoId. As demonstrated in Eq. (10), the frequency components of impact response signals exhibit deterministic characteristics, essentially comprising linear combinations of harmonic components centered at the plate’s natural frequencies. This inherent spectral property enables the decomposition of plate strain response, , into three distinct components through the application of VMD coupled with central frequency selection criteria: where Nm denotes the number of IMFs whose center frequency aligns with the natural frequency of the plate, sk(t) (k = 1, 2, …, M-Nm) represents the IMF of other frequency components and e(t) is the residue. After decomposing the strain signal, a selection process is applied to identify the Nm modal strain responses. This is achieved by comparing the center frequency of each IMF with the known natural frequencies of the structure, obtained from the offline modal analysis. An IMF is identified as the rth modal response if its center frequency falls within a tolerance band (±5%) of the corresponding natural frequency. IMFs that do not meet this criterion are considered components of noise or other excitations and are discarded. This selection rule effectively filters the signal and handles cases where VMD might produce ambiguous modes. The rth modal strain response, , is expressed as follows: As can be seen from Eq. (18), once the system parameters ,, and are determined through experimental modal analysis (EMA) or finite element analysis (FEA) in the off-time preparatory phase, the coefficient of the attenuation oscillation curve, , can be rapidly extracted by fitting the envelope curve32. Then, the uncorrected modal constant, , can be calculated as follows: By arranging the uncorrected modal constant from low to high order, the discriminating uncorrected modal constant vector, , can be obtained. To achieve the proposed efficient impact localization for real-time application, it is necessary to establish a matrix during the off-time preparatory phase, which consists of vibration modes of all potential impact locations. The mathematical description of the said matrix,, is expressed as follows: where qN denotes the number of potential impact locations. Note that if and only if the matrix D satisfies the column full rank condition, the result of impact localization is unique, which is also the criterion for selecting Nm. By comparing Eqs. (19) and (20), it can be found that if the yet-to-be identified impact point coincides with the qm point, then the Γ is collinear with the qmth column of matrix D. However, due to the unavoidable measurement noise and modeling errors, they cannot be completely collinear. Therefore, the impact localization can be realized by evaluating the location of the column vector in D with the highest collinearity to Γ. In this paper, the cosine similarity is employed to calculate the collinearity. Force reconstruction Upon the impact location is determined, the forward transfer matrix between strain response and yet-to-be reconstructed impact force f(t) can be determined, and the matrix product relationship is expressed by discretizing Eq. (10) as follows33: where is the transfer matrix, Δt is the time interval, is the measured strain response vector and denotes the inevitable measurement error that corrupts the real response . Reconstructing the time history of the impact force is an ill-posed inverse problem, even small noise in the measured response may cause large fluctuations in the solution. Hence, regularization techniques, including Tikhonov regularization and sparse regularization have been widely used in the impact force reconstruction. A typical regularization model includes a regularization term, which modeled based on the prior of the unknown force. Because of the time sparsity of impact force, the sparse regularization technique contains l1-norm has been proved to be more suitable for solving the impact time history34. The sparse regularization model for solving Eq. (21) is expressed as follows: where μ>0 represents the regularization parameter. Then, the Lagrange multipliers ρ and γ are introduced to obtain the augmented Lagrange function: Based on Eq. (23), the minimization problem can be solved using ADMM to obtain the optimal solutions for f35. The key steps of ADMM are summarized as follows: (1) Set the value of , , γ1, and to zero. (2) Update the value of f and f* through Eq. (24). where E is the identity matrix and is the number of iterations. (3) Update the value of γ through gradient descent method: (4) Repeat step (2) and step (3) to continuously update the parameters f, f* and γ until the iteration termination condition is satisfied. where c is the tolerance and it is set to 1 × 10−6 in this paper. The selection of the regularization parameter μ and the ADMM penalty parameter ρ is crucial for the reconstruction performance. μ controls the sparsity of the solution, while ρ affects the convergence rate. In this study, these parameters were determined empirically through a preliminary analysis to balance solution accuracy and sparsity, with typical values of μ ranging from 1 to 10 and ρ set to 1. Overall flowchart of proposed ImFoId methodology The overall flowchart of proposed ImFoId methodology is summarized in Fig. 2. As can be seen from the picture, the proposed ImFoId methodology can be divided into off-line preparation phase and on-line identification phase. In the off-line preparation phase, the modal parameters of the target structure are firstly obtained by EMA or FEA, and the number of required modes is then determined by satisfying the column full rank condition of matrix D. Note that the measurement point should avoid the mode node of the first Nm modes. The acquisition and processing of impact response data are carried out in the on-line identification phase. Once the impact response data is acquired, the first Nm order modal strain response is obtained by VMD, and the uncorrected mode constant vector is then extracted by fitting the envelope curve. As the uncorrected mode constant vector Γ is obtained, impact localization can be realized by evaluating the location of the column vector in D with the highest collinearity to Γ. Following the results of impact localization and modal analysis, the transfer matrix H is then determined and the sparse regularization model can be established. Through solving the inverse problem by Eqs. (23)-(26), the time history of the impact force can be reconstructed. Overall flowchart of the proposed ImFoId methodology. While the online identification relies on a single FBG, the method’s efficacy is predicated on a comprehensive offline calibration phase. This calibration burden is a key consideration for practical implementation. For larger or more complex structures, this would necessitate more extensive experimental modal analysis or a high-fidelity finite element model. Therefore, the proposed framework is most readily applicable to critical structural components where such detailed characterization is feasible and warranted as part of the design or maintenance cycle. Experimental validation The ImFoId methodology presented in previous section can theoretically work utilizing a single FBG sensor with low sampling rate. In this section, a laminated composite plate with simplified airfoil geometry is employed to examine the feasibility of the proposed ImFoId methodology. In the experiments, impacts were applied with different heads of a hand-held impact hammer COINV INV9311 (rubber and nylon). The impact load history data and strain response data were simultaneously recorded by NI DAQ card and Gator High-Speed FBG interrogator (1–19.2kHz), respectively. The sampling frequency was set as 5 kHz, and the duration of the measurement was maintained at 1 second. The datum and proposed ImFoId methodology were run on a PC with an AMD Ryzen 5 5500 Processor. Set-up The experimental set-up consists in a laminated composite plate mounted on a specialized clamp that simulates a cantilever boundary condition, as depicted in Fig. 3. Ten points are selected as candidate impact locations, as shown in Fig. 3a. The material of employed plate is T300 carbon fiber with a layup sequence of [0/90]5s. The dimensions of the plate are 500 mm in length and 3 mm in thickness. The width of the plate is 150 mm at the clamped boundary and 100 mm at the free boundary. The FBG sensor is glued at the coordinate (50mm, 150 mm) during the whole tests (see Fig. 3a). To emphasize the efficacy of the proposed ImFoId approach, experiments were carried out under zero initial conditions and external excitation (white noise and harmonic force) conditions. An exciter was utilized to produce diverse external loads, with its apex connected to the plate at the coordinate (200mm, 100 mm). The physical map of the experimental set-up is presented in Fig. 3b. Experimental set-up of the laminated composite plate (a) schematic diagram of the candidate impact points and the measurement point and (b) photo of the experimental set-up. In the off-time preparatory phase, an EMA of the plate was performed prior to the impact tests with the software LMS Test Lab. A tap testing technique was employed to identify 6 displacement modes(mass-normalized) of 10 candidate impact points within the frequency band 50–1200 Hz. The results of the EMA are summarized in Fig. 4. As can be seen from the picture, mode shape #4 shows that the structure was not exactly symmetric. Hence, based on the column full rank condition of mode matrix D, the number of required vibration modes Nm is determined as 3. Moreover, the 6 strain modes(mass-normalized) of the measurement point was also measured through strain mode test, as presented in Fig. 5. Modal properties of the laminated composite plate within 50–1200 Hz (displacement mode shape of ten candidate impact points, natural frequency and damping ratio). Six strain modes (mass-normalized) of the measurement point. Impact force identification under the absence of other load types The experimental verification of the proposed ImFoId method was firstly conducted under the absence of other load types to evaluate the accuracy of the theoretical framework. The occurrence of impact events was introduced to the plate by employing the impact hammer equipped with a nylon tip to any of the ten candidate impact points. Specifically, for the VMD processing of this section, the number of modes was set to M = 6 (a value larger than the number of expected modes in the frequency band of interest), the penalty parameter α was set to 8000, and the Lagrange multiplier was initialized to zero. In addition, the initial frequencies were set to the natural frequencies obtained during the off-line preparation phase. In this subsection, point #9 is randomly selected to illustrate the details of ImFoId. The strain response of the measuring point (sampling length 5000, duration 1 second), the time-frequency spectrum of the strain response and the modal strain response obtained by VMD are presented in Fig. 6. In the time-frequency spectrum shown in Fig. 6b, it can be clearly seen that the impact force excites the multi-order modal response of the plate, and the low-order modal response exists throughout the sampling window. Using the basic steps of impact force localization shown in Fig. 2, the first three modal strain responses of the plate are extracted by using VMD to decompose the original strain response signal. Subsequently, a signal of 0.3 seconds was intercepted for envelope curve fitting, and the fitting results are shown in Fig. 6c. Finally, according to the Eq. (19), the modal constant vector composed of the first three uncorrected modal constants is calculated. By assessing the collinearity between the modal constant vector and the column vectors of the system modal matrix, the impact location results are obtained. Extraction of uncorrected modal constants through VMD and envelope curve fitting (a) strain response of measurement point (b) time-frequency spectrum and (c) results of envelope curve fitting. Following the above steps, the tests of impact localization were carried out on the remaining 9 points of the plate, and the localization results of ten potential impact points on the plate were finally obtained, as shown in Fig. 7. It can be seen from Fig. 7 that the index of impact location, cosine similarity, is the highest in the real impact localization, indicating that the impact sources on the ten potential impact points are accurately located. To further demonstrate the performance of the proposed method, the repeatability tests and the localization tests under the excitation of the rubber hammer are also carried out. The test results are summarized in Table 1. It can be seen from Table 1 that under the excitation of nylon tip, the average success rate of impact localization of ten potential impact points is 100 %, while under the excitation of rubber tip, the average success rate is 80 %. This is because the impact force generated by the nylon tip has a shorter duration, which is more in line with the assumption that the impact force in Eq. (10) is an ideal impulse. Nevertheless, the success rate of the forty tests across ten potential impact points reached 90%, which demonstrates that the proposed method exhibits a high level of accuracy. Moreover, the average time consumption for the forty tests was approximately 1 second, indicating that the proposed method possesses excellent computational efficiency. Result of impact localization. Table 1Results of impact localization and mean values of error indices in the absence of other load types.Hammer tipGroupLoc. (/10)Success (%)RE (%)PRE (%)Nylon110/1010044.8817.29210/1010046.6523.44Rubber18/108015.088.9328/108016.167.72Tol36/409030.6914.35 Results of impact localization and mean values of error indices in the absence of other load types. The results of reconstructing the time history of impacts through sparse regularization technique, as described by Eq. (22), are presented in Fig. 8. In this study, the first four orders of modal parameters of the plate were employed to establish the transfer matrix H, with the dimensions of 1000 ×1000, as these modes were observed to dominate the impact response under ideal conditions and provided a well-conditioned matrix. To evaluate the performance of reconstruction of impact time history, two different error indices are employed, namely the relative error (RE) and the peak relative error (PRE)35. As can be seen from Table 1 and Fig. 8, the time history of the reconstructed impact force aligns closely with the real impact force, indicating that the reconstruction results are highly accurate. Moreover, the evolution of the residual term (i.e., ) along iteration number for reconstructing the impact force by ADMM under different hammer tips is shown in Fig. 9. It can be seen from the picture that the iterations of two reconstruction processes to reach the relative tolerance c = 1×10−6 are 82 and 2610, respectively. The corresponding computation times are respectively 0.14 seconds and 0.31 seconds, which demonstrates a high reconstruction efficiency. Force reconstruction of the impact force applied at point #9 (a) impact generated by a nylon tip and (b) impact generated by a rubber tip. The evolution of the residual for reconstructing the impact force acting on the 9th point of the CFRP plate by ADMM (a) impact generated by a nylon tip and (b) impact generated by a rubber tip. Impact force identification under operational conditions Most of the existing literatures have primarily focused on the ImFoId in an idealized environment, whereas investigations into ImFoId under operational conditions remain scarce. This study incorporates the operational conditions by employing an exciter to simulate external excitation, and examines the robustness of proposed method in identifying the impact force under operational conditions. The external excitation signal utilized in this study is a composite of a sinusoidal wave and Gaussian white noise, with point #2 randomly selected as the research object. Prior to testing, a sinusoidal wave with an amplitude of 1.3 mm and a frequency of 150 Hz was mixed with Gaussian white noise at a signal-to-noise ratio of 10 dB to serve as the excitation signal, which was then input into the control system of the signal generator. Subsequently, a power amplifier was employed to amplify the excitation signal, enabling the exciter to function effectively. Fig. 10 presents the external excitation signal (Fig. 10a), the strain response of the measuring point (Fig. 10b), and the time-frequency spectrum (Fig. 10c) of the strain response. As can be seen from Fig. 10c, even under external excitation, harmonic components centered around the natural frequency of the plate are still clearly discernible in the vibration response, and these harmonic components will be extracted as the basis for impact location. Strain response of plate under combined action of multiple types of external loads (a) external excitation signal composed of sinusoidal wave (1.3mm - 150 Hz) and white noise (SNR = 10dB) (b) strain response at the measurement point and (c) time-frequency spectrum. Figure 11 shows the results of impact localization of the plate under the combined action of multiple types of external loads. Since the impact point is close to the second-order modal node of the plate, the second-order strain modal response signal is not successfully extracted from the original signal (see Fig. 11a). Fortunately, the first-order, third-order and fourth-order modal strain responses were successfully extracted. Since the column vectors composed of the first-order, third-order, and fourth-order vibration modes still have the uniqueness of direction, the uncorrected modal constants extracted from these three strain modal responses can still be used for impact localization. The result of cosine similarity assessment is shown in Fig. 11b. As can be seen from the picture, the cosine similarity at point 2 is much larger than that at other locations, indicating that the impact force applied to point 2 is accurately located. Table 2 and Fig. 12 further summarize the results of impact localization for ten potential impact points on the plate under the excitation of sinusoidal waves with varying amplitudes and frequencies. Throughout the entire test, the signal-to-noise ratio (SNR) of Gaussian white noise was fixed at 10 dB. As shown in Table 2, the success rate across a total of 120 tests on ten potential impact points reached 90.83%, which demonstrates that the proposed impact force localization method exhibits strong robustness. Results of impact localization of point #2 subjected to multiple types of external loads (a) decomposed modal strain responses and the fitting results of envelope curves and (b) results of impact localization. Table 2Results of Impact localization and mean values of error indices in operational conditions.Parameters of sinusoidal waveGroupLoc. (/10)Success (%)RE (%)PRE (%)0.9mm - 10Hz19/109041.9921.7729/109043.0824.531.4mm - 10Hz18/108051.4626.1927/107052.5219.860.5mm - 50Hz110/1010046.0621.14210/1010046.8018.061.0mm - 50Hz19/109041.2530.8929/109036.9520.870.7mm - 150Hz110/1010037.3519.9729/109036.9420.131.3mm - 150Hz19/109038.9121.51210/1010040.6822.61Tol109/12090.8342.8322.29 Results of Impact localization and mean values of error indices in operational conditions. Results of Impact localization under operational conditions. To quantitatively demonstrate the necessity of the VMD stage, a baseline comparison was conducted. The localization procedure was repeated on the same noisy signals from the operational condition tests, but without applying VMD. The uncorrected modal constants were extracted by directly fitting the raw signal using band-pass filters centered at the natural frequencies. The resulting localization success rate dropped to 42.61%, with many cases showing ambiguous results where the cosine similarity for the true location was not uniquely the highest. This comparison confirms that the VMD-based modal filtering is the key component enabling the high accuracy and robustness of the proposed framework. Figure 13 shows the reconstruction results of the impact force acting on point 2 under different parameter settings of external loads. In this study, the first six orders of modal parameters of the plate were employed to establish the transfer matrix H, with the dimensions of 1000 ×1000. Under operational conditions, higher-order modes can be excited and contribute to the overall response. Therefore, the first six modes were used to construct the transfer matrix H to capture a fuller dynamic range of the structure, improving the model’s ability to distinguish the impact response from the external excitation noise. Due to the existence of external excitation, it can be seen that the RE and PRE of the reconstruction results are higher than those in the ideal case. In addition, observing Table 2 and Fig. 13, it can be seen that with the increase of the amplitude of sinusoidal wave, the RE index and PRE index of the reconstruction results also increase. But on the whole, the accuracy of the reconstruction results is acceptable in engineering. The evolutions of the residual for reconstructing the impact force are presented in Fig. 14. As can be seen from the picture, the durations of the two iterations are respectively 0.11 and 0.16 seconds, indicating the remarkable efficiency of the reconstruction process. Results of time history reconstruction of impact force applied at point #2 under different operational conditions (a) 0.7mm-150Hz and (b) 1.3mm-150Hz. The evolution of the residual for reconstructing the impact force applied at point #2 under different operational conditions (a) 0.7mm-150Hz and (b) 1.3mm-150Hz. Discussions The purpose of this subsection is to propose improvements of the ImFoId method presented in this paper. As can be observed from Table 1, there is a positive correlation between the hardness of the hammer and the success rate of impact localization on ten potential impact points. Specifically, the harder the hammer, the higher the success rate. To elaborate on this phenomenon more comprehensively, the spectrums of impact forces produced by different types of hammer tips are plotted, as depicted in Fig. 15. From Fig. 15, it is evident that the amplitude of the ideal impact force remains constant across the entire frequency domain. In contrast, the amplitudes of the impact forces generated by the nylon tip and the rubber tip decline as the frequency rises. Nevertheless, when compared to the rubber tip, the amplitude of the impact force generated by the nylon tip decays at a slower pace in the low-frequency band, being more similar to the ideal impact force. Consequently, the success rate of impact localization under the excitation of the nylon hammer is higher. The modeling error stemming from assuming the real impact force as an ideal impulse is one of the primary causes of inaccurate localization results. Spectrums of different impact forces. Figure 16 illustrates the results of impact localization at point #2 using four uncorrected modal constants (the first, third, fourth, and fifth orders). A comparison with Fig. 11 reveals that increasing the number of uncorrected modal constants enhances the contrast in cosine similarity between point #2 and other potential impact points, leading to more distinct impact localization results. Consequently, irrespective of the hammer tip used to generate the impact force, appropriately augmenting the number of uncalibrated modal constants improves the accuracy of impact localization. Nevertheless, extracting higher-order uncorrected modal constants typically necessitates a higher sampling rate, which may substantially increase computational and sampling costs. Results of impact localization of point #2 using four uncorrected modal constants (a) impact generated by nylon tip and (b) impact generated by rubber tip. Conclusion This paper presents and validates a robust impact identification methodology for composite structures, uniquely requiring only a single FBG sensor and demonstrating efficacy under simulated operational conditions. The core innovation is a two-stage framework: initially, variational mode decomposition (VMD) acts as a modal filter to isolate impact-induced strain responses from noisy operational signals. From these purified responses, an "uncorrected modal constant vector" is extracted as a unique signature for localization through collinearity assessment. Subsequently, a sparse regularization model, efficiently solved by ADMM, reconstructs the impact force’s time history based on the identified location. Experimental validation on a composite airfoil plate demonstrated the framework’s potential for robust performance, achieving over 90% localization success across ten potential impact points even amidst significant external harmonic and stochastic excitations. While reconstruction accuracy showed some sensitivity to external noise intensity, the PRE of the reconstructed force was captured with an average precision of 22.29%, which is generally considered engineering-acceptable. It should be acknowledged that the validation was conducted on a single composite plate with a fixed sensor location and a specific set of simulated operational loads, while a detailed sensitivity analysis of key parameters is not carried out in this study, we recognize its importance for further validation. Future work should focus on extending this framework to more complex structures, considering various sensor placements, and testing against a wider range of operational noise profiles, and conducting such sensitivity analyses to fully establish its generalizability and limitations. Acknowledgements The authors are pleased to acknowledge the financial support of the Natural Science Foundation of Hubei Province (Grant No. 2025AFC005) and Jingmen Science and Technology Burea (Grant NO. 2024ZDYF004, 2025ZDYF006). Supports from the Fundamental Research Funds of Jingchu University of Technology (Grant NO. YY202444, YY202442) are also appreciated. Author contributions Jian Huang: Writing – original draft, Validation, Software, Formal analysis. Xiaomingyang: Visualization, Software, Resources, Methodology.Quanxin Jiang: Supervision,Funding acquisition. Li Zhang: Writing – review & editing, Validation, Methodology, Funding acquisition, Formal analysis, Conceptualization. Funding The authors are pleased to acknowledge the financial support of the Natural Science Foundation of Hubei Province (Grant No. 2025AFC005) and Jingmen Science and Technology Burea (Grant NO. 2024ZDYF004, 2025ZDYF006). Supports from the Fundamental Research Funds of Jingchu University of Technology (Grant NO. YY202444, YY202442) are also appreciated. Data availability The datasets used and analyzed during the current study available from the corresponding author on reasonable request. Declaration Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and analyzed during the current study available from the corresponding author on reasonable request.