Fiber-optic based three-dimensional deformation detection method and system

Abstract

The application discloses a kind of three-dimensional deformation detection method and system based on optical fiber, the method includes: obtaining layout parameter, multi-core optical fiber original measurement sequence and reference point true value sequence, preparation inversion input set;Performing physical consistency joint inversion, real strain tensor field and transfer coefficient field are as common unknowns to be optimized, in single joint loss function including physical compatibility constraint term and dissipation consistency constraint term, by alternately iterative solution, real strain tensor field estimation and transfer coefficient field estimation are generated simultaneously;Performing gated multi-scale feedback compensation, utilize high-frequency abnormal energy to determine gate region, perform local re-weighted inversion, generate closed-loop corrected field;Performing three-dimensional geometry reconstruction, utilize closed-loop corrected transfer coefficient field to identify crack subdomain, perform segmented splicing, generate complete three-dimensional deformation field.The application improves the inversion accuracy and physical authenticity of damage area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural health monitoring, and particularly relates to a three-dimensional deformation detection method and system based on an optical fiber. BACKGROUND

[0002] Large civil engineering structures and aerospace structures inevitably accumulate damage, material degradation and deformation due to loads and environmental erosion during their long-term service. Monitoring the health status of these structures throughout the entire cycle and obtaining the complete three-dimensional deformation field inside the structure are technical prerequisites for realizing structural safety evaluation, damage diagnosis and life prediction. Distributed optical fiber sensing technology, especially multi-core optical fiber sensing, has become a key monitoring method for obtaining structural strain fields due to its advantages of distributed along the way, high precision, anti-electromagnetic interference and weather resistance. How to accurately invert the three-dimensional real deformation field of the host structure from the original data measured by the optical fiber has great engineering application value.

[0003] At present, in the deformation detection technology based on optical fibers, the common data processing flow is a serial mode of calibration first and reconstruction later. At the initial stage of sensor deployment, the strain transfer coefficient (i.e. the ratio of optical fiber strain to structural strain) is calibrated offline once by applying a known reference load or using limited reference point true values (such as strain gauge data), and an initial transfer coefficient field, which is usually assumed to be spatially uniform or segmented uniform, is obtained. In long-term monitoring, the transfer coefficient is regarded as a fixed and known correction parameter. When reconstructing the deformation, the algorithm deducts the temperature disturbance component measured by independent temperature sensors or special temperature cores from the optical fiber measurement data, corrects the measured strain using the fixed transfer coefficient, and reconstructs the corrected optical fiber projection strain into the three-dimensional displacement field or strain tensor field of the structure by geometric integration or least squares method based on the elastic mechanics model.

[0004] However, the above-mentioned existing serial processing flow faces deep technical challenges when dealing with actual scenarios where the structure is damaged. Decoupling the strain transfer coefficient field and the real strain tensor field violates the strong coupling characteristics of the two in the physical damage evolution process, leading to one-way accumulation of errors and ill-posed inversion. Specifically, the calibration first and reconstruction later mode requires that the transfer coefficient k(x) must be accurately known in advance. Once the structure is damaged (such as micro-cracks, interface debonding), k(x) changes from a static parameter to a dynamic unknown quantity evolving with damage. If the initial calibrated k(x) is still used, the error will be irreversibly propagated to the real strain field ε _true(x) reconstruction, resulting in distortion of the reconstruction result. Existing methods usually do not consider the physical irreversibility of damage when trying to update k(x). k(x) as the interface transfer efficiency, its decline is a dissipation process. Existing methods often regard it as a free variable, resulting in the fact that measurement noise easily causes k(x) to appear non-physical rebound in time series (i.e. k _t >k _t-1 ), which is contrary to damage mechanics, resulting in long-term instability of k(x) field estimation. SUMMARY

[0005] The present application aims to provide a fiber-based three-dimensional deformation detection method and system to solve the above problems existing in the prior art.

[0006] Technical solutions, according to one aspect of the present application, a fiber-based three-dimensional deformation detection method, comprising:

[0007] Obtain the layout parameters, the original measurement sequence of the multi-core optical fiber and the true value sequence of the reference points, and prepare the inversion input set;

[0008] Perform physically consistent joint inversion to use the inversion input set to take the real strain tensor field and the transfer coefficient field as common unknowns to be optimized to generate the real strain tensor field estimate and the transfer coefficient field estimate synchronously;

[0009] Perform gated multi-scale feedback compensation to use the real strain tensor field estimate and the transfer coefficient field estimate to generate the closed-loop corrected transfer coefficient field and the closed-loop corrected real strain tensor field;

[0010] Perform three-dimensional geometric reconstruction to use the closed-loop corrected transfer coefficient field, the closed-loop corrected real strain tensor field and the layout parameters to generate the complete three-dimensional deformation field.

[0011] According to another aspect of the present application, a fiber-based three-dimensional deformation detection system, comprising:

[0012] The preparation module is used for obtaining the layout parameters, the original measurement sequence of the multi-core optical fiber and the true value sequence of the reference points, and preparing the inversion input set based thereon;

[0013] The inversion module is used for performing physically consistent joint inversion to use the inversion input set to generate a set of joint inversion estimate fields;

[0014] The compensation module is used for performing gated multi-scale feedback compensation to use the joint inversion estimate fields to generate the closed-loop corrected transfer coefficient field and the closed-loop corrected real strain tensor field;

[0015] and a reconstruction module, configured to perform three-dimensional geometric reconstruction, generate a complete three-dimensional deformation field by using the closed-loop corrected transfer coefficient field, the closed-loop corrected real strain tensor field and the layout parameters.

[0016] By the above technical solution, the application solves the problems of decoupling and one-way accumulation of errors, non-physical rebound and long-term instability, and improves the inversion accuracy and physical authenticity of the damaged area. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 It is a whole flowchart of a three-dimensional deformation detection method based on an optical fiber.

[0018] Figure 2 It is a flowchart of performing a transfer coefficient field sub-problem solution.

[0019] Figure 3 It is a flowchart of performing a gated multi-scale feedback compensation.

[0020] Figure 4 It is a flowchart of performing three-dimensional geometric reconstruction. DETAILED DESCRIPTION

[0021] In order for those skilled in the art to better understand the technical scheme of the present application, the technical method in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] The terms "first", "second", and the like in the specification of the present application and the above drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, or product that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or end.

[0023] In this document, the reference to "embodiments" means that the specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily mutually exclusive or alternative to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] Embodiment one, provide a kind of three-dimensional deformation detection method based on optical fiber, solve the problem that strain transfer coefficient is unknown (by damage caused) and temperature disturbance and real strain are difficult to separate in traditional optical fiber strain monitoring.In a specific embodiment, method can be executed by the computing device configured with processor and memory.

[0025] As Figure 1 Shown, specifically include the following steps:

[0026] Step 1.1, obtain the laying parameter, the original measurement sequence of multicore optical fiber and the reference point true value sequence, and prepare inversion input set based on it.

[0027] Specifically, the original measurement sequence raw _measurements Of multicore optical fiber includes the strain reading and corresponding time stamp of the multiple cores along the optical fiber path acquired from the multicore optical fiber measurement device (for example, demodulator based on optical frequency domain reflection OFDR).The reference point true value sequence ref _truth Including the strain or displacement data and time stamp measured by independent reference sensor (for example, fiber grating FBG, strain gauge or displacement meter) laid in the key position of structure.Laying parameter deploy _params It is a set of static data describing the physical laying state of optical fiber sensor, preferably, laying parameter deploy _params At least include: the directional vector of each core in the cross section of optical fiber, the spatial coordinates of fiber path in the coordinate system of structure (i.e. fiber path), the sampling interval of sensor, and the definition of coordinate system adopted.

[0028] After obtaining the above data, time alignment is performed, for example, by clock correction and time interpolation (such as linear interpolation or spline interpolation), the original measurement sequence raw _measurements Of multicore optical fiber and the reference point true value sequence ref _truth Are aligned to the unified time reference, and the time synchronization data package synced _data .

[0029] Based on the fiber path coordinates in laying parameter, the measuring points along the path in the time synchronization data package synced _data Are mapped to the unified spatial grid (for example, one-dimensional fiber path coordinates or three-dimensional structure grid), and the spatialized measurement frame spatial _frame Is formed.In this process, preferably, outlier detection and mutation point identification are performed, for example, the method based on amplitude threshold, the robust statistical method based on median absolute deviation (MAD) or mutation point detection algorithm is used, and the identified abnormal data points are recorded in the valid data mask valid _maskIf a temperature reference channel exists in the system (such as a dedicated temperature sensing fiber or environmental sensor), these data are processed synchronously to obtain the temperature reference data temp. _ref .

[0030] Preparing the inversion input set further includes constructing a measurement equation model and an initial field. In a preferred embodiment, the inversion input set includes: measurement data, a measurement equation model, and an initial field.

[0031] Measurement Equation Model (meas) _model This describes the physical mapping relationship between the projected strain in each core direction, transmitted through the interface, and the superposition of temperature disturbances to form the measured values. Preferably, in constructing the measurement equation model, the... _model At this time, it is necessary to refine the direction vectors of each core, especially considering the effects of fiber path bending and torsion, to obtain a more accurate direction mapping table. Furthermore, this can be achieved by combining the equipment noise nominal value with spatial measurement frames. _frame The statistical characteristics are used to estimate the noise level of each channel and form a channel weight library, which is embedded in the measurement equation model. _model In this context, it is used to weight data consistency items in subsequent joint inversions to improve the stability of the inversion.

[0032] The initial field includes initial guesses for the transmission coefficient field, the true strain tensor field, and the temperature perturbation field. The initial transmission coefficient field k... _init The initial true strain tensor field ε can be set to 1 across the entire field under the initial micro-damage assumption, or set as a piecewise constant according to structural partitions. _init In the neighborhood of the reference point, the truth sequence ref of the reference point can be used. _truth And it was roughly estimated using the least squares method. The initial temperature perturbation field ΔT was obtained. _init If the temperature reference data is temp _ref If available, it can be obtained directly through regression. In an optional, more general implementation, if the temperature reference data temp _ref Missing (i.e., no temperature sensor), initial temperature disturbance field ΔT _init This method can be obtained by extracting and regularizing the low-frequency residuals from spatialized measurement frames, making it independent of predetermined temperature measurements.

[0033] Preparing the input set also includes the reference point truth sequence ref _truth For this processing, robust regression and outlier removal methods are preferably used to eliminate drift and outliers, and the reference point weights (ref) are obtained based on their variance estimation. _weight It is used for anchoring reference points in subsequent inversion.

[0034] Step 1.2, perform physics-consistent joint inversion, with the inversion input set, generate real strain tensor field estimate and transfer coefficient field estimate simultaneously, as co-optimized unknowns. In this embodiment, performing physics-consistent joint inversion further includes generating temperature perturbation field estimate. The detailed implementation of this step will be described in embodiment two.

[0035] Step 1.3, perform gated multi-scale feedback compensation, with the real strain tensor field estimate and the transfer coefficient field estimate, generate closed-loop corrected transfer coefficient field and closed-loop corrected real strain tensor field. Perform targeted closed-loop correction for local errors (especially in fracture regions) that may exist in joint inversion.

[0036] Step 1.4, perform three-dimensional geometric reconstruction, with the closed-loop corrected transfer coefficient field, the closed-loop corrected real strain tensor field, and the layout parameters, generate complete three-dimensional deformation field. Moreover, performing three-dimensional geometric reconstruction further utilizes the temperature perturbation field estimate. Specifically, the temperature perturbation field estimate can be used to separate the strain component caused by temperature from the closed-loop corrected real strain tensor field before geometric reconstruction, to obtain the deformation field caused purely by mechanical factors (such as load, damage), and can also be output as an independent environmental parameter.

[0037] Optionally, a three-dimensional deformation detection method based on optical fiber can also be: obtaining a multi-core optical fiber original measurement sequence, a reference point true value sequence, and layout parameters; based on the sequence and the layout parameters, preparing measurement data, a measurement equation model, and an initial field; based on the measurement data, the measurement equation model, and the initial field, performing physics-consistent joint inversion to simultaneously generate real strain tensor field estimate, transfer coefficient field estimate, and temperature perturbation field estimate; based on the real strain tensor field estimate, the transfer coefficient field estimate, and the temperature perturbation field estimate, performing gated multi-scale feedback compensation to generate closed-loop corrected transfer coefficient field and closed-loop corrected real strain tensor field; based on the closed-loop corrected transfer coefficient field, the closed-loop corrected real strain tensor field, the temperature perturbation field estimate, and the layout parameters, performing three-dimensional geometric reconstruction to generate complete three-dimensional deformation field.

[0038] Embodiment two, provides a detailed, preferred implementation of physics-consistent joint inversion (PI-JI).

[0039] Specifically, performing physics-consistent joint inversion includes: constructing a single joint loss function for simultaneously constraining the real strain tensor field and the transfer coefficient field.

[0040] In this embodiment, the prepared inversion input set (including spatialized measurement frame spatial _frame , measurement equation model meas _model , reference point weight ref _weight, initial transfer coefficient field k _init , initial true strain tensor field ε _init , initial temperature perturbation field ΔT _init , etc.), a unified objective function, i.e., joint loss function L, is constructed to minimize both the data residuals and the degree of violation of physical constraints.

[0041] The joint loss function contains at least data consistency term, physical compatibility constraint term and dissipation consistency constraint term; and by minimizing the joint loss function, the true strain tensor field estimate and the transfer coefficient field estimate are solved.

[0042] In a preferred embodiment, the joint loss function L can be expressed as a weighted sum of multiple penalty terms: L = L _data + L _comp + L _k_smooth + L _diss + L _anchor + L _temp ;

[0043] The specific form can be written as:

[0044] L = ∑ i ||m _i (x)-k(x)·(u i T ε _true (x)u i )-β i ΔT(x)||2 2 +λ _c ||▽×(▽×ε _true T || F 2 +λ _k TV(k(x))+λ _m [k _t (x)-k _t-1 (x)] + 2 +λ _a ∑ x_j∈Ref ||ε _true (x _j )-ε _ref (x _j )||2 2 +λ _T TV(ΔT(x));

[0045] Wherein, ∑ i ||m _i (x)-k(x)·(u i T ε _true (x)u i )-βi ΔT(x)||2 2 is a data consistency term L _data ;

[0046] λ _c ||▽×(▽×ε _true T | F 2 is a physical consistency constraint term L _comp ;

[0047] λ _k TV(k(x)) is a transfer coefficient field smoothing term L _k_smooth ;

[0048] λ _m [k _t (x)-k _t-1 (x)] + 2 is a dissipation consistency constraint term L _diss ;

[0049] λ _a ∑ x_j∈Ref ||ε _true (x _j )-ε _ref (x _j )||2 2 is a reference point anchoring term L _anchor ;

[0050] λ _T TV(ΔT(x)) is a temperature field smoothing term L _temp .

[0051] wherein the symbols have the following meanings: m _i (x) is the measured strain value of the i-th core at position x; k(x) is the transfer coefficient field to be solved, preferably taking values in the interval [0,1]; ε _true (x) is the real strain tensor field to be solved; u i is the direction unit vector of the i-th core; β i is the temperature sensitivity coefficient of the i-th core; ΔT(x) is the temperature perturbation field to be solved; λ _c , λ _k , λ _m , λ _a , λ _T are the weight hyperparameters of the respective terms; ▽× denotes the curl operation; TV(...) denotes the total variation regularizer term, used to encourage piecewise smooth solutions; [z] + denotes max(z,0), used to impose one-way constraints; k _t (x) and k _t-1(x) represents the transfer coefficient fields of the current and previous time periods, respectively; Ref represents the set of reference points; ε _ref (x _j ) represents the true strain at the reference point; T represents the matrix transpose; ||...|| F This represents the Frobenius norm of the matrix. The physical compatibility constraint term L... _comp It is a constraint imposed on the true strain tensor field and is associated with the curl compatibility of the true strain tensor field (i.e., the Saint-Venant compatibility condition), used to guarantee ε. _true (x) is physically integrable. The dissipative consistency constraint term L... _diss It is a constraint imposed on the transfer coefficient field and is associated with the time monotonically non-increasing property of the transfer coefficient field (k _t (x)<=k _t-1 (x) reflects the irreversible cumulative nature of damage (decreasing k value).

[0052] Before constructing the loss function L, preferably, unit normalization and scale normalization operations are performed to improve the numerical conditions of the optimization problem and ensure convergence. Continuous operators in L (such as ▽× and TV) are discretized (e.g., using finite differences), and a physical regularization operator library, phys, is constructed. _ops This is provided for subsequent solvers to use.

[0053] Since L also includes ε _true Given three sets of unknowns: (x), k(x), and ΔT(x), and the existence of k(x)*ε _true The multiplicative coupling terms of (x) are difficult to solve directly. This application employs an alternating iterative approach to solve for minimizing the numerical value of a single joint loss function.

[0054] Alternating iteration (or block coordinate descent) decomposes the original problem into three subproblems that are executed in a loop:

[0055] The first step involves fixing the transfer coefficient field to perform the solution of the true strain tensor quantum problem, updating the true strain tensor field. In this step, k(x) and ΔT(x) are treated as fixed values. At this point, L with respect to ε _true The part (x) represents a quadratic objective, corresponding to solving a large sparse linear system. Preferably, the system is solved efficiently using a preconditional conjugate gradient (PCG) or multigrid solver. After solving, numerical stabilization operations, such as local residual smoothing or step size back, can be performed to reduce numerical oscillations without violating physical constraints, thus obtaining the updated true strain tensor field ε. _true_iter .

[0056] The second step is to fix the updated true strain tensor field, perform the transfer coefficient field subproblem solving, and update the transfer coefficient field. Then, use the ε updated in the previous step... _true_iterand the current ΔT(x) is treated as a fixed value. At this time, the part in L with respect to k(x) is an optimization problem (usually convex) as Figure 2 The solving process includes:

[0057] Based on the fixed updated real strain tensor field, the intermediate transfer coefficient field is obtained. Preferably, it is realized by calculating the gradient of L to k and performing a proximal gradient update, which can effectively handle the non-smooth TV regular term (L _k_smooth ), to obtain the intermediate transfer coefficient field k _mid .

[0058] Interval projection is performed on the intermediate transfer coefficient field to meet the preset physical value domain constraint. Specifically, the interval projected intermediate transfer coefficient field k _mid_clipped =clip(k _mid ,0,1) operation is performed, that is, all values less than 0 in k _mid are set to 0, and all values greater than 1 are set to 1, to force it to meet the physical value domain [0, 1].

[0059] Monotone projection is performed on the interval projected intermediate transfer coefficient field, and the monotone projection is associated with the last period transfer coefficient field to force to meet the time monotone non-increasing property, to generate the updated transfer coefficient field. Specifically, the transfer coefficient field k _iter =min(k _mid_clipped ,k _prev ) is executed, where k _prev is the transfer coefficient field of the last time period. The monotone projection operation strictly guarantees k _t (x) <= k _t-1 (x), which realizes the dissipation consistency constraint. Further, the transfer coefficient field k _iter can be implemented with anisotropic smoothing (i.e. reducing smoothing along the suspected crack direction, protecting the edge) and small value truncation (eliminating numerical jitter).

[0060] The third is the temperature disturbance field estimation. Fixing the updated real strain tensor field ε _true_iter and the transfer coefficient field k _iter , if there is temperature reference data temp _ref , it can be solved by linear regression or kernel regression. If there is no temperature reference data temp _ref , it is preferably to extract the low frequency component from the data residual and assist with TV regular to estimate ΔT. The slow change (low frequency) of k(x) and ΔT(x) (usually also low frequency) are coupled in the data consistency term, which easily leads to the low frequency disambiguation problem. In this embodiment, this problem can be alleviated by using a damping update or controlling the solving order.

[0061] The real strain tensor field estimate and the transport coefficient field estimate are generated after the alternating iteration converges. Specifically, the above three sub-problems will be executed in a loop until a convergence criterion is satisfied (e.g., the relative drop of the joint loss function L is below a pre-set threshold such as le-5, or the KKT optimality condition residual is up to the mark). After convergence, the solution of the current round is output, i.e., the real strain tensor field estimate _true_hat , the transport coefficient field estimate k _hat , and the temperature perturbation field estimate AT _hat .

[0062] The transport coefficient field estimate k _hat and the real strain tensor field estimate e _true_hat can be evaluated in the vicinity of the converged solution by a diagonal approximation of the curvature matrix (diagonal of the Hessian matrix) or a subsample Monte Carlo method, and their variances or uncertainties at each spatial location are generated as a joint inversion uncertainty weight field _confidence for the subsequent gated multi-scale feedback compensation step.

[0063] Embodiment three elaborates on the preferred implementation of the gated multi-scale feedback compensation GMRC.

[0064] The physics-consistent joint inversion (PI-JI) is a global optimization process, and the introduced smoothness regularizer (e.g., the transport coefficient field smoothness term L _k_smooth ) may cause over-smoothing in the real, physically abrupt regions (e.g., crack tips) while suppressing noise, introducing low-frequency drift or underestimating damage. This embodiment provides a closed-loop feedback mechanism to use high-frequency anomalies as a gating signal, and only in these anomaly sub-domains to conduct targeted local re-estimation of the joint inversion results to correct the above issues.

[0065] The gated multi-scale feedback compensation is executed, as shown in FIG. 3, specifically including: Figure 3

[0066] Step S31, multi-scale decomposition is performed based on the real strain tensor field estimate to generate a high-frequency anomaly energy map. Specifically, to separate the local, sharp damage signal (high frequency) from the global, smooth deformation signal (low frequency), multi-scale decomposition is performed. In a preferred implementation, the decomposition can employ a redundant wavelet transform (e.g., a shift-invariant wavelet) or a sparse representation based on a damage feature dictionary. After decomposition, the energy of each high-frequency scale coefficient is aggregated (e.g., summed or L2 norm) to form a high-frequency anomaly energy map E _HF on the spatial location. The high-value region on the high-frequency anomaly energy map E _HF is a strong indicator of the presence of high-frequency abruptness in the real strain tensor field estimate e _true_hat , i.e., cracks or severe damage. ​

[0067] Step S32, determine the gating region according to the high-frequency anomaly energy map. The high-frequency anomaly energy map E _HF is converted into a binary gating mask gate _mask , which is used to identify sub-regions that need to be locally re-optimized subsequently. In one implementation, by setting a gating threshold τ, the regions where E _HF > τ are defined as the initial mask. Preferably, to improve robustness, the threshold τ is not a fixed value, but is adaptively generated by a robust statistical method, for example, estimate the noise baseline based on the median absolute deviation (MAD) of the high-frequency anomaly energy map E _HF , and set τ to be several times (e.g. 3 times MAD) of the baseline. The initial binary mask usually contains noise and is discontinuous, and preferably post-processing is performed on it, such as connected component filtering (to remove fragmented regions with too small area) and morphological operations (such as dilation or closing operation, to fill holes and connect adjacent regions).

[0068] In a more preferred implementation, considering that cracks usually have anisotropic morphology (extending in a certain direction), direction correction is also included. Specifically, the principal strain direction estimated by the true strain tensor field ε _true_hat is used to perform anisotropic expansion or contraction on the gating mask gate _mask , so that the mask is moderately expanded in the suspected crack extension direction and contracted in the transverse direction. Quality control (e.g. checking minimum area, slenderness ratio, boundary adhesion degree, etc.) can be performed on the gating mask gate _mask to remove artifacts and form the final gating region.

[0069] Step S33, perform local re-weighted inversion in the gating region. This includes: for the gating region, construct a regionalized weight map; the regionalized weight map local _weights is a weight field corresponding to the spatial position. In the gating region covered by the gating mask gate _mask , take a set of predetermined values, and take another set of values (e.g. remain unchanged) outside the region. Preferably, the map adopts a kernel function to smoothly transition at the edge of the gating region, to avoid introducing new artificial discontinuities at the boundary of the gating region. The construction of the map can also combine the joint inversion uncertainty weight field _confidence , i.e. give higher re-weighting amplitude in areas with high uncertainty.

[0070] The regionalized weight map is used to locally modify the joint loss function, and the modification includes increasing the weight of the data consistency term, and reducing the regularization weight of the transfer coefficient field; increasing the weight of the data consistency term corresponds to increasing the weight of the data consistency term L _data , in the gating region (high-frequency anomaly area), the measured strain value m_i (x) Reflects the true physical discontinuity and should reduce the smoothing of the model to it. Lower the regularization weight of the transfer coefficient field, corresponding to lower the smoothing term L _k_smooth (i.e. λ _k *TV(k(x))) of the transfer coefficient field k(x) _k Itself is highly varying (from 1 to 0), over-smoothing (large λ _k ) or too strong monotonicity constraint (diffusive uniformity) will hinder the invoker to find the true steep solution. Outside the region covered by the gate _mask mask, the weight remains unchanged. Preferably, at the edge of the gate _mask mask, a kernel function (e.g. Gaussian kernel) is adopted for smooth transition to avoid creating new artificial steps at the region boundary. In addition, the uncertainty weight field _confidence can also be incorporated into the local _weights weighting map to impose stronger re-weighting in regions with high uncertainty.

[0071] With the modified joint loss function, perform the physically consistent joint inversion within the gated region. Specifically, with the true strain tensor field estimate ε _true_hat , the transfer coefficient field estimate k _hat , and the temperature perturbation field estimate ΔT _hat as the initial values for the hot start, within the region covered by the gate _mask mask, perform 1-2 rounds of the alternating iteration of the true strain tensor subproblem -> transfer coefficient field subproblem -> temperature perturbation subproblem with the local _weights weighting map. The result of the local iteration, i.e. the locally updated transfer coefficient field k _local and the locally updated true strain tensor field ε _true_local . In a preferred embodiment, further include a rollback mechanism: if the local iteration leads to the global loss function L _data worsening, or the data consistency term L _local within the gated region does not improve, then judge that this time re-weighting is too aggressive, discard the locally updated transfer coefficient field k _true_local and the locally updated true strain tensor field ε _true_hat , and roll back to the solution before the start (ε _hat , k _local ).

[0072] Step S34, fusion and convergence. The method further comprises the following steps:

[0073] Obtain the locally updated transfer coefficient field and the locally updated true strain tensor field, i.e. k _local and ε_true_local .

[0074] The locally updated fields within the gating region are fused with the jointly inverted estimated fields outside the gating region to generate the closed-loop corrected transfer coefficient field and the closed-loop corrected true strain tensor field. The jointly inverted estimated fields are the estimated transfer coefficient field k _hat and the estimated true strain tensor field ε _true_hat . The fusion operation combines the locally updated transfer coefficient field k _local and the locally updated true strain tensor field ε _true_local within the gating region with the estimated transfer coefficient field k _hat and the estimated true strain tensor field ε _true_hat outside the gating region. Preferably, the fusion is not a simple region replacement, but a kernel-weighted fusion with a consistency correction performed at the boundary of the gating mask gate _mask to ensure the final output fields are still smoothly transitioned in space. The final outputs are the closed-loop corrected transfer coefficient field k _star and the closed-loop corrected true strain tensor field ε _true_star .

[0075] In an alternative embodiment, the GMRC step itself can constitute an outer loop. If the fused result still does not meet the convergence criteria (e.g. the gating region residual is still above a threshold), the construction of the regionalized weight map process can be returned to for a reduced reweighting amplitude and a retry. Optionally, the temperature perturbation field estimate ΔT _hat is not updated in the GMRC, but directly taken from the global estimate.

[0076] Embodiment Four, detailing a preferred implementation of the three-dimensional geometric reconstruction.

[0077] In this embodiment, the three-dimensional geometric reconstruction is performed, as shown in FIG. 4, and specifically includes the following steps: Figure 4

[0078] Step S41, performing geometric integration on the closed-loop corrected true strain tensor field to obtain a three-dimensional displacement field. The true strain tensor field ε _true_star is a strain tensor field and cannot be directly equated to deformation. In order to obtain an intuitive three-dimensional displacement field, it must be geometrically integrated. Specifically, according to the fiber geometry path in the deployment parameter deploy _params , the true strain tensor field ε _true_star ​Decomposition or projection of a 3x3 tensor field into curvature, shear curvature and twist fields along the fiber path. Integration is similar to solving differential equations, which requires setting boundary conditions. Boundary conditions are set according to engineering practice, such as fixed end zero displacement, simply supported boundary or symmetric boundary. Integration will generate unknown integration constants, which correspond to the translation and rotation of rigid body. Preferably, a reference point true value sequence ref _truth is used to eliminate rigid body modes with reference point displacement as reference, calibrate integration constants, and obtain a unique displacement solution. In an alternative embodiment, geometric integration can be achieved in two ways:

[0079] Option 1 (integration along fiber path): For the fiber path, sequential integration and drift correction are used. This method accumulates displacement along the fiber path step by step, and the calculation is fast, but the integration drift may be caused by measurement noise.

[0080] Option 2 (based on structure grid): For the structure grid path, Poisson reconstruction is used. This method treats the curvature / twist field as the second derivative of the displacement field, and solves a large Poisson equation (sparse linear equation set) on the entire structure grid. This method has slightly larger calculation amount, but the result is globally consistent and not prone to drift, and is generally the preferred option.

[0081] The output of this step is a three-dimensional displacement field assuming the structure is continuous.

[0082] Step S42, based on the closed-loop corrected transfer coefficient field, identifies the crack subdomain. The three-dimensional displacement field obtained in step S41 assumes that the structure is continuous, which is not true when there are cracks or slips. This step uses the closed-loop corrected transfer coefficient field k _star , k _star is a representation of strain transfer efficiency, and the area with low value (close to 0) physically corresponds to the area where the optical fiber and the structure are detached or the structure is cracked. Therefore, the closed-loop corrected transfer coefficient field k _star can be used to identify the crack subdomain, and generate the crack subdomain mask crack _mask . Identifying the crack subdomain includes the following steps: using a composite criterion that at least relates to the amplitude condition or gradient condition of the closed-loop corrected transfer coefficient field, and the abnormal energy condition of the closed-loop corrected true strain tensor field. Specifically, the composite criterion preferably combines:

[0083] Amplitude condition of k _star : for example, k _star <0.5;

[0084] Gradient condition of k _star : for example, ||grad(k _star )||>threshold (indicating the edge of k value mutation); grad is the gradient operator.

[0085] Abnormal energy condition: for example, using high-frequency abnormal energy map E _HF , or the real strain tensor field ε based on closed-loop correction _true_star The calculated strain energy anomaly. At the same time, multiple criteria (for example, low k value and high energy) are met to determine the crack subdomain, which has higher reliability than single criterion, and can effectively reduce misjudgment.

[0086] Step S43, according to the crack subdomain, perform segmented splicing on the three-dimensional displacement field to generate a complete three-dimensional deformation field. The crack subdomain mask crack _mask Generated in S42 is used to correct the continuous displacement field. Specifically, instead of performing global integration, local geometric integration is performed on both sides of the crack subdomain mask crack _mask Identified crack subdomain, which will produce discontinuous displacement on both sides of the crack, and the difference between the two is the cross-slit displacement jump variable, which has a clear physical meaning, i.e. crack opening or shear slip. According to the contact or gap model, splice the two independent displacement fields to obtain a segmented spliced displacement field. Preferably, in order to make the results more reasonable in vision and mechanics, a transition zone is constructed in the vicinity of the crack, and the segmented spliced displacement field is transitionally smoothed based on energy consistency and smoothness constraints in the transition zone to avoid displacement step artifacts, obtaining a smoothed displacement field. In a more robust implementation, the smoothed displacement field can be subjected to a final closed-loop check, for example, using reference point historical envelope and residual statistics for verification. If the check fails, the criteria can be adjusted and recalculated. The displacement field after the check passes is the final output complete three-dimensional deformation field of the present application, which contains both continuous deformation of the structure and discontinuous jump at the crack.

[0087] Embodiment five provides a fiber-based three-dimensional deformation detection system, which can be a general-purpose or special-purpose computing device, such as an industrial control computer (IPC), a server, a workstation, or an embedded system.

[0088] The system preferably includes one or more processors, memory such as RAM or DRAM, non-volatile storage such as hard disk HDD or solid state disk SSD, and one or more input / output (I / O) interfaces in terms of hardware structure. The I / O interface is used to communicate with external devices, and is preferably connected to at least one or more multi-core fiber measurement devices (such as OFDR or FBG demodulator) for receiving multi-core fiber original measurement sequences, and optional reference sensors (such as strain gauges, displacement meters) for receiving reference point true value sequences. The non-volatile storage stores computer executable instructions, which when loaded into the memory and executed by the processor, cause the processor to implement the following functional modules:

[0089] A preparation module is configured to obtain the deployment parameters, the multi-core fiber raw measurement sequence and the reference point true value sequence, and to prepare the inversion input set based thereon. Specifically, the preparation module performs data acquisition, time synchronization, spatial grid mapping, and outlier detection (e.g., using the median absolute deviation method). Further, the module is configured to construct the inversion input set, which preferably includes the measurement equation model meas _model , the reference point weight ref _weight , and the initial field k _init , ε _init , ΔT _init (k _init is the initial transfer coefficient field, ε _init is the initial true strain tensor field, and ΔT _init is the initial temperature perturbation field). In an optional implementation, the module is configured to extract and regularize the initial temperature field from the low-frequency spatial components of the measurement residuals, if temperature reference data is missing.

[0090] An inversion module is configured to perform a physically consistent joint inversion using the inversion input set, and to generate a set of joint inversion estimated fields. Specifically, the inversion module is configured to construct a single joint loss function L, which includes at least a data consistency term, a physical compatibility constraint term (e.g., based on the Saint-Venant compatibility condition), and a dissipation consistency constraint term (e.g., based on the time monotonic non-increasing property). The inversion module is further configured to minimize the loss function using an alternating iterative approach, including alternatingly solving a true strain tensor sub-problem, a transfer coefficient field sub-problem, and a temperature perturbation field estimation. The inversion module finally outputs a set of joint inversion estimated fields, which preferably includes a true strain tensor field estimate ε _true_hat and a transfer coefficient field estimate k _hat .

[0091] A compensation module is configured to perform a gated multi-scale feedback compensation using the joint inversion estimated fields, and to generate a closed-loop corrected transfer coefficient field and a closed-loop corrected true strain tensor field. Specifically, the compensation module performs a closed-loop correction procedure. The module is configured to perform a multi-scale decomposition (e.g., a redundant wavelet transform) on the joint inversion estimated fields ε _true_hat to generate a high-frequency anomaly energy map E _HF . Based on the map, a gated region gate _mask is determined, and a locally re-weighted inversion is performed within the region. The re-weighting is preferably achieved by modifying the joint loss function, e.g., by increasing the weight of the data consistency term and decreasing the regularization weight of the transfer coefficient field. The module is configured to fuse the locally updated fields k _local , ε _local and the globally estimated fields k _hat , ε _true_hat to generate the closed-loop corrected transfer coefficient field k_star and the real strain tensor field ε _true_star .

[0092] a reconstruction module configured to perform a three-dimensional geometric reconstruction to generate a complete three-dimensional deformation field using the loop-corrected transfer coefficient field, the loop-corrected real strain tensor field and the layout parameters. Specifically, the reconstruction module performs a geometric integration and a patching process. For the loop-corrected real strain tensor field ε _true_star a geometric integration (e.g., a Poisson-type reconstruction) is performed to obtain a three-dimensional displacement field. The reconstruction module is configured to utilize the loop-corrected transfer coefficient field k _star to identify crack sub-domains, e.g., using a correlation of the loop-corrected transfer coefficient field k _star amplitude and anomaly energy. The module performs a piecewise patching of the three-dimensional displacement field according to the crack sub-domains (e.g., calculating the cross-crack displacement jump), generating a final complete three-dimensional deformation field.

[0093] In an alternative embodiment, the large-scale matrix operations (e.g., PCG solvers or Poisson reconstruction) involved in one or more of the above modules, in particular the inversion module and the compensation module, can be implemented by dedicated hardware, such as a graphics processing unit (GPU), a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC), to accelerate the computation process.

[0094] Embodiment six provides a specific reproducible numerical calculation case to illustrate in detail how the transfer coefficient field sub-problem solving step realizes the dissipation consistency constraint in the physically consistent joint inversion.

[0095] The scenario of this embodiment is that the system is performing an alternating iteration process, and it is specifically to the transfer coefficient field sub-problem solving step. It is assumed that it is a monitoring process evolving over time, and the current is time step t.

[0096] It is assumed that a simplified 1D spatial grid only contains 5 discrete measurement points (x1, x2, x3, x4, x5). The processor reads the transfer coefficient field of the last period (time step t-1) from the memory, denoted as k _prev . k _prev is the result of the last complete monitoring period (e.g., 1 hour ago) convergence, reflecting the structural damage state at time t-1. It is assumed that k _prev = [1.0, 0.9, 0.9, 1.0, 0.8]; indicating that x1 and x4 are intact (k = 1.0); x2 and x3 have slight damage (k = 0.9); x5 has moderate damage (k = 0.8).

[0097] At time step t, the processor of the inversion module obtains the intermediate transfer coefficient field k _midIt is based on the measurement data m at the current time (t). _i (x) and strain field ε _true_iter The calculated value of k, but for which time constraints have not yet been applied. The processor performs an interval projection on this field, i.e., clip(k). _mid To satisfy the physical range constraint [0,1], the intermediate transfer coefficient field k after interval projection is obtained. _mid_clipped Assumption:

[0098] k _mid_clipped =[1.0,0.85,0.95,0.9,0.7];

[0099] For k _mid_clipped Analyze the physical meaning of vectors:

[0100] x1(k=1.0): Same as at time t-1.

[0101] x2(k=0.85): The decrease in the k value (0.9->0.85) indicates that the damage has worsened. This is physically possible.

[0102] x3 (k=0.95): The k value rebounds (0.9->0.95). This is physically impossible (damage does not heal itself) and is likely a non-physical bounce caused by measurement noise or model error.

[0103] x4(k=0.9): The decrease in the k value (1.0->0.9) indicates the appearance of new damage. This is physically possible.

[0104] x5(k=0.7): The decrease in the k value (0.8->0.7) indicates that the damage has worsened. This is physically possible.

[0105] To resolve the non-physical bounce problem at x3, the processor then executes a monotonic projection step. This step is associated with the previous period's transfer coefficient field k. _prev This is to enforce the property of time monotonically non-increasing. Numerically, this is implemented as a point-by-point minimum operation: the transfer coefficient field k... _iter =min(k _mid_clipped ,k _prev );

[0106] The point-by-point calculation process is as follows:

[0107] k _iter (x1)=min(k _mid_clipped (x1),k _prev (x1))=min(1.0,1.0)=1.0;

[0108] k _iter (x2)=min(k _mid_clipped (x2),k_prev (x2)) = min(0.85, 0.9) = 0.85; accepted physically possible damage aggravation;

[0109] k _iter (x3) = min(k _mid_clipped (x3), k _prev (x3)) = min(0.95, 0.9) = 0.9; rejected physically impossible noise rebound, forced k value to 0.9 at t-1 time;

[0110] k _iter (x4) = min(k _mid_clipped (x4), k _prev (x4)) = min(0.9, 1.0) = 0.9; accepted physically possible new damage;

[0111] k _iter (x5) = min(k _mid_clipped (x5), k _prev (x5)) = min(0.7, 0.8) = 0.7; accepted physically possible damage aggravation.

[0112] The final updated transfer coefficient field generated by the iteration step is: k _iter = [1.0, 0.85, 0.9, 0.9, 0.7].

[0113] As can be seen from the calculation process of the embodiment, the dissipation consistency constraint L _diss and its numerical implementation (monotone projection) explicitly introduce the prior knowledge of damage mechanics (i.e. irreversibility of damage) into the inversion framework. This makes the method automatically filter out non-physical phenomena caused by measurement noise (such as rebound at x3), accurately capture the true damage evolution (such as the decrease at x2, x4, x5), and improve the robustness and physical authenticity of the transfer coefficient field estimation.

[0114] According to an aspect of the present application, in the physically consistent joint inversion process taking the transfer coefficient field as the main unknown, the steps of discretization of the joint loss function and the weight strategy can be specifically:

[0115] spatial _frame , measurement equation model meas _model , reference point weight ref _weight , initial transfer coefficient field k _init , initial real strain tensor field ε _init , initial temperature disturbance field ΔT _init , temperature reference data temp _ref , valid data mask valid_mask Unit normalization and dimensionality reduction, grid construction and index table, get norm _index_cfg .

[0116] Based on norm _index_cfg Discrete compatibility residual operator and total variation regularization operator, and the time index of the dissipation consistent term and the transfer coefficient field of the last cycle (if any) are organized into a sequence to form the physical regularization operator library phys _ops .

[0117] According to the statistical distribution of the reference point weight ref _weight and the residual spectrum of the spatialized measurement frame spatial _frame , initialize the weights of each loss term (data consistency, compatibility, total variation, dissipation consistency, reference point anchoring, temperature prior), get the joint loss function L (including weight table and operator pointer).

[0118] According to an aspect of the present application, in the physical consistency joint inversion process taking the transfer coefficient field as the main unknown, the solution process of the real strain tensor sub-problem (fixed transfer coefficient and temperature) can be specifically:

[0119] Get the joint loss function L, the initial transfer coefficient field k _init or the last round of transfer coefficient field k _iter , the initial temperature disturbance field ΔT _init or the last round of temperature disturbance field ΔT _iter , the spatialized measurement frame spatial _frame , the measurement equation model meas _model , the reference point weight ref _weight , the physical regularization operator library phys _ops , assemble the sparse linear system and boundary constraints about the quadratic form of the real strain tensor field ε _true , call the preconditioned conjugate gradient or multi-grid to get the updated real strain tensor field ε _true_iter .

[0120] Map the reference point weight ref _weight to a sparse penalty matrix or a hard constraint condition, load the compatibility residual operator from the physical regularization operator library phys _ops into the system matrix, re-solve and update the real strain tensor field ε _true_iter , enhance the integrability and reference point fitting degree.

[0121] Numerical stabilization (local residual smoothing, boundary neighborhood shape-preserving filtering, step back) is performed on the real strain tensor field ε _true_iter , which reduces the oscillation without changing the constraint satisfaction degree, and generates the stabilized real strain tensor field ε_true_iter .

[0122] fixed transfer coefficient field and temperature perturbation field;

[0123] Construction A _ε ·ε _true =b _ε (including compatibility and reference point terms);

[0124] Iterative solution: ε _true <-PCG(A _ε ,b _ε ,tol);

[0125] Stabilization: ε _true <-Smooth(ε _true ,boundary _aware ).

[0126] where tol is a threshold value, A _ε and b _ε are the matrix and vector for the real strain tensor subproblem, respectively.

[0127] According to one aspect of the present application, in the physical consistent joint inversion process taking the transfer coefficient field as the main unknown, the solving process of the transfer coefficient field subproblem (fixed real strain and temperature) can be specifically:

[0128] Obtain the joint loss function L, the updated real strain tensor field ε _true_iter , the spatialized measurement frame spatial _frame , the measurement equation model meas _model , the physical regularization operator library phys _ops , perform one step of proximal gradient (including total variation proximal) on the transfer coefficient field k(x), and perform interval projection with a value between zero and one to obtain the intermediate transfer coefficient field k _mid .

[0129] Based on the dissipation consistency (monotonic non-increasing over time) constraint, the intermediate transfer coefficient field k _mid is compared with the transfer coefficient field of the last cycle (if it is the first cycle, it is compared with the initial transfer coefficient field k _init ), and is projected to the non-increasing set to obtain the constrained projected transfer coefficient field k _iter .

[0130] Anisotropic smoothing (reducing smoothing along the main crack direction, moderate smoothing in the normal direction) and small value truncation are performed on the constrained projected transfer coefficient field k _iter to avoid block artifacts and numerical jitter to obtain the stabilized transfer coefficient field k _iter .

[0131] Compute gradient: g _k = dL / dk;

[0132] Proximal update: k <- Prox _tV (k - a0·g _k );

[0133] Clip: k <- clip(k, 0, 1);

[0134] Dissipation consistency: k <- min(k, k _prev );

[0135] Anisotropic smoothing: k <- AnisoSmooth(k).

[0136] where a0 is the learning rate or step size. Prox _tV is the algorithm operator.

[0137] According to an aspect of the present application, in the process of estimating the temperature disturbance field in the physical-consistent joint inversion process with the transfer coefficient field as the main unknown, the implementation process can further be:

[0138] If there is a temperature reference channel, obtain the temperature reference data temp _ref , the joint loss function L, the updated real strain tensor field e _true_iter , the updated transfer coefficient field k _iter , construct linear regression or kernel regression with the temperature sensitivity coefficient, estimate the temperature disturbance and smooth it to obtain the temperature disturbance field AT _iter .

[0139] If there is no temperature reference channel, obtain the joint loss function L, the updated real strain tensor field e _true_iter , the updated transfer coefficient field k _iter , the spatialized measurement frame spatial _frame , extract the low-frequency component from the data residual, separate the temperature disturbance with the total variation regularization to obtain the temperature disturbance field AT _iter .

[0140] By damping update and order control, the difficulty of distinguishing temperature disturbance and transfer coefficient field at low frequency is reduced, and the stabilized temperature disturbance field AT _iter is obtained.

[0141] According to an aspect of the present application, in the process of generating uncertainty weight in the physical-consistent joint inversion process with the transfer coefficient field as the main unknown, the process of alternating iteration and uncertainty weight generation can further be:

[0142] Obtain the real strain tensor field e _true_iter , the transfer coefficient field k _iter , and the temperature disturbance field AT_iter , joint loss function L, cycle several rounds in the order of real strain tensor subproblem -> transfer coefficient field subproblem -> temperature perturbation subproblem, convergence is judged by the target drop rate and the optimality condition residual, and the real strain tensor field estimate ε is obtained _true_hat , transfer coefficient field estimate k _hat , temperature perturbation field estimate ΔT _hat .

[0143] Generate joint inversion uncertainty weight field by diagonal approximation of curvature matrix or sub-sampling Monte Carlo evaluation of variance near the converged solution _confidence (position-dependent, used for reweighting).

[0144] According to one aspect of the present application, in the process of multi-scale decomposition and abnormal energy in the gated multi-scale feedback compensation GMRC process, it can also be specifically:

[0145] Get real strain tensor field estimate ε _true_hat , use damage feature dictionary or redundant wavelet for multi-scale decomposition to get scale coefficients, and construct high-frequency energy density to get high-frequency abnormal energy map E _HF .

[0146] For high-frequency abnormal energy map E _HF , implement robust statistics (such as median absolute deviation) to estimate the noise lower limit, generate adaptive threshold and percentile threshold candidates, and get threshold candidate set hf _thresholds .

[0147] Evaluate the predictive return of different thresholds on subsequent overall target improvement, select the optimal threshold and solidify, and get the gated threshold τ.

[0148] According to one aspect of the present application, in the process of generating a gated region in the gated multi-scale feedback compensation GMRC process, it can also be specifically:

[0149] Get high-frequency abnormal energy map E _HF , gated threshold τ, generate an initial binary mask, perform connected domain screening and morphological dilation / contraction, and get the gated mask gate _mask .

[0150] According to the principal strain direction of the real strain tensor field estimate ε _true_hat , calculate the anisotropy weight field to make the mask moderately expand in the suspected crack extension direction and shrink in the transverse direction, and get the direction-corrected gated mask gate _mask_oriented .

[0151] For gated mask gate _mask_orientedPerform quality control (minimum area, elongation ratio, border adhesion degree), remove debris and false positives, and form the final gate mask gate _mask (Covering the abnormal sub-domain).

[0152] According to an aspect of the present application, in the process of regionalized reweighting joint inversion in the GMRC process, the process can also be:

[0153] Obtain the gate mask gate _mask , joint inversion uncertainty weight field _confidence , construct the regionalized weight mapping: increase the data consistency term weight and reduce the total variation regular weight of the transfer coefficient field in the gated region, and smoothly transition with the kernel function at the edge to obtain the regionalized weight mapping local _weights .

[0154] Obtain the true strain tensor field estimate ε _true_hat , the transfer coefficient field estimate k _hat , the temperature disturbance field estimate ΔT _hat , the regionalized weight mapping local _weights , perform one to two rounds of local iteration of the true strain tensor subproblem -> transfer coefficient field subproblem -> temperature disturbance subproblem in the gated region to obtain the locally updated transfer coefficient field k _local and the locally updated true strain tensor field ε _true_local .

[0155] If the local iteration worsens the global objective or the gated region residual is not improved, roll back to the previous solution and reduce the reweighting amplitude until improvement is achieved or early stopping is triggered, obtaining the stabilized k _local , the stabilized ε _true_local .

[0156] According to an aspect of the present application, an optional implementation is provided, which describes the preparation process of the measurement equation model, specifically:

[0157] Obtain the spatialized measurement frame spatial _frame , the deployment parameter deploy _params , refine the mapping of each core direction vector and structure coordinates (considering the effects of fiber bending and twisting), and form the direction mapping table dir _map .

[0158] Based on the direction mapping table dir _map , assemble the observation operator and use a sparse storage format to establish a measurement value model of the projected strain in each core direction after the interface transmission and temperature disturbance superposition, obtaining the measurement equation model meas _model .

[0159] Combining device noise claim with spatialized measurement frame spatial _frame Statistics, estimate each channel noise level and form channel weight library, embedded in measurement equation model meas _model , ensure that data consistency items have reliable basis.

[0160] According to one aspect of the application, an optional embodiment is provided, which describes the preparation process of the reference point weight, specifically:

[0161] Obtain the reference point true value sequence ref _truth , effective data mask valid _mask , perform robust regression and outlier rejection to obtain the reference point clean sequence ref _clean .

[0162] According to the reference point clean sequence ref _clean , drift modeling and long-term stability evaluation are performed, and the variance is mapped to the weight to obtain the reference point weight ref _weight .

[0163] The reference point weight ref _weight is registered into the reference point anchoring module, and the reference point weight ref _weight is obtained.

[0164] According to one aspect of the application, an optional embodiment is provided, which describes the preparation process of the initial value construction, specifically:

[0165] Obtain the deployment parameter deploy _params , give the initial transfer coefficient field k _init (uniform or partition constant) to the component partition, and perform gradual transition at the boundary.

[0166] In the reference point neighborhood, the initial true strain tensor field ε _init is constructed by least squares, and the noise is reduced by slight smoothing.

[0167] If there is temperature reference data temp _ref , the initial temperature disturbance field ΔT _init is directly regressed; if there is no temperature reference, the initial temperature disturbance field ΔT _init is extracted and regularized from the residual low frequency of the spatialized measurement frame spatial _frame .

[0168] The physical consistency joint inversion (PI-JI) framework proposed in the application discards the serial process, and constructs a single joint loss function L to solve the transfer coefficient field k(x) and the true strain tensor field ε _true(x) is simultaneously treated as a common unknown to be optimized. An alternating iterative approach is used to solve the problem, ensuring that the transfer coefficient field k(x) and the true strain tensor field ε are both solved. _true (x) involves bidirectional and mutual constraints during the solution process, resolving the issue of ε caused by static errors in k(x). _true (x) The problem of permanent distortion is solved by addressing decoupling and the problem of one-way error accumulation.

[0169] The method in this application explicitly introduces a dissipative consistency constraint term L into the joint loss function L. _diss Furthermore, it is related to the monotonically non-increasing property of time. In solving the field problem of the transfer coefficient, it is achieved through the monotonically projective step (k _iter =min(k _mid_clipped ,k _prev The algorithmically forces the transfer coefficient field k(x) to be solved to satisfy k _t ≤k _t-1 The physical constraints directly suppress non-physical rebound caused by measurement noise, ensuring the physical accuracy and long-term stability of damage estimation, and solving the problems of non-physical rebound and long-term instability.

[0170] The gated multi-scale feedback compensation (GMRC) method and crack identification proposed in this application further utilize the above-mentioned results to perform refined solutions for high-damage areas, providing a complete and high-precision three-dimensional deformation field.

[0171] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A three-dimensional deformation detection method based on optical fiber, characterized in that, include: Obtain deployment parameters, original measurement sequences of multi-core optical fibers, and true value sequences of reference points to prepare an inversion input set; Perform physical consistency joint inversion, using the inversion input set, treat the true strain tensor field and the transfer coefficient field as common unknowns to be optimized, and simultaneously generate the true strain tensor field estimate and the transfer coefficient field estimate; Perform gated multi-scale feedback compensation, and use the real strain tensor field estimation and the transfer coefficient field estimation to generate the closed-loop corrected transfer coefficient field and the closed-loop corrected real strain tensor field. Perform three-dimensional geometric reconstruction, and generate a complete three-dimensional deformation field using the closed-loop corrected transfer coefficient field, the closed-loop corrected real strain tensor field, and the layout parameters. in, The inversion input set includes: measurement data, measurement equation model, and initial field; Performing a joint physical consistency inversion further includes: generating an estimate of the temperature perturbation field; Furthermore, three-dimensional geometric reconstruction is performed to further utilize temperature perturbation field estimation; Perform joint physical consistency inversion, including: A single joint loss function is constructed to simultaneously constrain the true strain tensor field and the transfer coefficient field; The joint loss function must include at least a data consistency term, a physical compatibility constraint term, and a dissipation consistency constraint term; By minimizing the joint loss function, the estimation of the true strain tensor field and the transmission coefficient field are obtained.

2. The method according to claim 1, characterized in that, The physical compatibility constraint is a constraint imposed on the real strain tensor field and is related to the curl compatibility of the real strain tensor field. The dissipative consistency constraint is a constraint imposed on the transmission coefficient field and is associated with the time monotonically non-increasing property of the transmission coefficient field.

3. The method according to claim 1, characterized in that, Performing physical consistency joint inversion includes: using an alternating iterative approach as a numerical solution process to minimize a single joint loss function; Alternating iterations include: With the transfer coefficient field fixed, perform the solution of the real strain tensor quantum problem and update the real strain tensor field; Fix the updated true strain tensor field, perform the subproblem solving of the transfer coefficient field, and update the transfer coefficient field; The true strain tensor field estimate and the transfer coefficient field estimate are generated after the alternating iteration converges.

4. The method according to claim 3, characterized in that, Solving the field subproblem involving the transfer coefficients includes: Based on the fixed updated real strain tensor field, the intermediate transfer coefficient field is obtained; Perform interval projection on the intermediate transfer coefficient field to satisfy the preset physical range constraints; Perform monotonic projection on the intermediate transfer coefficient field after interval projection. The monotonic projection is associated with the transfer coefficient field of the previous period to force the time monotonic non-increasing property to be satisfied, and generate an updated transfer coefficient field.

5. The method according to claim 1, characterized in that, Perform gated multi-scale feedback compensation, including: Multi-scale decomposition is performed based on the estimation of the real strain tensor field to generate a high-frequency anomaly energy map. The gating region is determined based on the high-frequency anomaly energy map; Perform a local reweighted inversion within the gated area.

6. The method according to claim 5, characterized in that, Performing a local reweighted inversion includes: For the gated region, construct a regionalized weight mapping; Regionalized weight mapping is used to locally modify the joint loss function, including increasing the weight of the data consistency term and decreasing the regularization weight of the transmission coefficient field; A modified joint loss function is used to perform physical consistency joint inversion within the gated region.

7. The method according to claim 1, characterized in that, Performing 3D geometric reconstruction includes: Perform geometric integration on the true strain tensor field after closed-loop correction to obtain the three-dimensional displacement field; Based on the transfer coefficient field after closed-loop correction, the crack subdomain is identified; The three-dimensional displacement field is segmented and spliced ​​based on the crack subdomain to generate a complete three-dimensional deformation field.

8. A three-dimensional deformation detection system based on optical fiber, characterized in that, include: The preparation module is used to acquire deployment parameters, original measurement sequences of multi-core optical fibers, and true value sequences of reference points, and to prepare an inversion input set based on these. The inversion module is used to perform physically consistent joint inversion, generating a joint inversion estimate field using the inversion input set; The compensation module is used to perform gated multi-scale feedback compensation, and uses the joint inversion estimation field to generate the closed-loop corrected transfer coefficient field and the closed-loop corrected true strain tensor field. The reconstruction module is used to perform three-dimensional geometric reconstruction, using the closed-loop corrected transfer coefficient field, the closed-loop corrected true strain tensor field and the layout parameters to generate a complete three-dimensional deformation field. in, The inversion input set includes: measurement data, measurement equation model, and initial field; Performing a joint physical consistency inversion further includes: generating an estimate of the temperature perturbation field; Furthermore, three-dimensional geometric reconstruction is performed to further utilize temperature perturbation field estimation; Perform joint physical consistency inversion, including: A single joint loss function is constructed to simultaneously constrain the true strain tensor field and the transfer coefficient field; The joint loss function must include at least a data consistency term, a physical compatibility constraint term, and a dissipation consistency constraint term; By minimizing the joint loss function, the estimation of the true strain tensor field and the transmission coefficient field are obtained.

Images