A fiber-optic sensing system and method for monitoring the health of a bridge structure

Abstract

This invention discloses a fiber optic sensing system and monitoring method for bridge structural health monitoring, comprising the following steps: pre-setting segmented reference fiber optic rings in the sensing optical cable; independently calibrating the strain and temperature coefficients of each segment by switching via optical switches, and dynamically correcting the spatial mapping coordinates to obtain precisely located distributed vibration and strain data; decoupling the data into a temperature field, a quasi-static strain field, and a dynamic vibration signal via a physically guided neural network; extracting time-frequency ridges from the dynamic signal and constructing a variable frequency band filter bank for filtering; windowing the filtered signal in blocks; extracting key modal components near the natural frequencies through physical constraint extension and adaptive noise complete empirical mode decomposition; synthesizing a full-time domain intrinsic mode function set after relaxing and splicing the sub-block boundaries; screening effective components based on the modal prior library and confidence matrix; verifying the structural modal feature vector with theoretical vibration modes; comparing it with the health status reference vector; and determining damage and providing early warning if the distance threshold is exceeded.

Description

Technical Field

[0001] This invention relates to the field of bridge structure technology, and in particular to a fiber optic sensing system and monitoring method for bridge structural health monitoring. Background Technology

[0002] This invention relates to the field of bridge structural health monitoring technology, and in particular to a fiber optic sensing system and monitoring method for bridge structural health monitoring.

[0003] With the continuous construction of transportation infrastructure and the increasing service life of bridges, long-term, real-time, and accurate health monitoring of bridge structures has become a key requirement for ensuring operational safety, extending service life, and reducing life-cycle maintenance costs. Fiber optic sensing technology, with its advantages of resistance to electromagnetic interference, corrosion resistance, high sensitivity, and the ability to perform long-distance distributed measurements, has become an important technical means in the field of bridge structural health monitoring. Distributed fiber optic sensing systems based on Brillouin scattering and Rayleigh scattering can simultaneously acquire temperature, strain, and vibration information along the fiber optic path, providing a technical foundation for comprehensive condition perception of bridge structures.

[0004] However, existing technologies still have significant shortcomings in the following aspects: First, regarding the multi-physics decoupling of sensor signals, Brillouin scattering signals exhibit cross-sensitivity to temperature and strain, while Rayleigh scattering signals are simultaneously affected by vibration, strain, and temperature. Existing coefficient calibration methods often employ whole-segment calibration, which is ill-suited to the spatial inhomogeneities of temperature and strain fields along long-distance bridges. When the ambient temperature changes significantly, fixed strain and temperature coefficients introduce substantial systematic errors. Some existing solutions, such as the patent with publication number CN121144770A, employ Physically Guided Neural Network (PINN) for multi-physics decoupling of fiber optic signals, but their application scenario is overhead line icing prediction, outputting only icing thickness, and does not address the joint decoupling of the temperature field, quasi-static strain field, and dynamic vibration signal within the bridge structure.

[0005] Secondly, in the data processing of long-distance distributed monitoring, to meet the real-time processing needs of massive sensor data, it is usually necessary to divide long-term continuous signals into shorter data blocks for parallel computation. Signal decomposition methods based on ensemble empirical mode decomposition (EEMD) algorithms are extremely sensitive to signal endpoints. Truncation of data blocks can lead to spurious intrinsic mode function components at the boundaries, i.e., endpoint effects. When the data block length is short, the proportion of endpoint effect-contaminated regions in the entire sub-block data increases sharply, severely compressing the effective decomposition result interval. In the prior art, the patent with publication number CN117606719A uses a moving time window to segment the acceleration signal, but it focuses on the clustering and screening problem of modal parameters between different windows, and does not solve the problem of the endpoint effect introduced by segmentation affecting the fidelity of the decomposition result. The patent with publication number CN121117950A uses variational mode decomposition to process the vibration signal of the switch cabinet, but it only processes short-time impact signals and does not address the modal discontinuity and splicing artifacts caused by block decomposition in long-distance distributed optical fiber monitoring.

[0006] Third, in terms of structural damage assessment and localization, existing technologies mostly use single or empirical thresholds for anomaly detection, making it difficult to balance detection sensitivity and false alarm rate. Patent CN117606719A monitors the condition by observing the time-varying curves of natural frequencies, but lacks a health benchmark model based on long-term statistical characteristics, thus failing to quantify the confidence level of the damage. Patent CN121117950A uses a dynamic fault feature vector library for similarity matching, but its application is limited to mechanical fault diagnosis of switchgear, and its correlation with the physical meaning of bridge structural modal parameters is weak, making it difficult to directly transfer to bridge scenarios.

[0007] In summary, existing technologies in the field of distributed fiber optic bridge monitoring have not effectively solved problems such as high-precision decoupling of multi-physics fields, suppression of endpoint effects in the block processing of massive data, and reliable damage identification based on statistical modeling. Therefore, there is an urgent need for a fiber optic sensing system and monitoring method for bridge structural health monitoring to overcome the shortcomings of the existing technologies. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing a fiber optic sensing system and monitoring method for bridge structural health monitoring.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for monitoring the health of bridge structures includes the following steps: Multiple segmented reference fiber rings of known length are pre-installed in the sensing optical cable of the bridge to be monitored. Specifically, segmented reference fiber rings of 1.0m length are pre-installed at the bridge tower, pier, and mid-span section of the sensing optical cable. The optical switch matrix automatically switches at intervals of 1 to 24 hours. The strain coefficient and temperature coefficient of the sensing segment defined by the adjacent segmented reference fiber rings are independently calibrated using the known length. The change in fiber length is measured using a low coherence optical frequency domain reflection auxiliary system at a sampling rate of 50Hz. The spatial mapping coordinates of the scattered signal are dynamically corrected based on the change in fiber length to obtain distributed vibration and strain sensing data with accurate spatial location. When the cumulative length change exceeds 0.01m, a spatial coordinate translation correction command is sent to the demodulation system. The corrected sensing data (Brillouin gain spectrum and Rayleigh scattering phase) is input into the physical guided neural network decoupler. Its loss function is composed of a mean square error term and a physical residual term based on the coupled mechanical equation and the heat conduction equation, weighted by 0.7 and 0.3 respectively. The decoupled output is the temperature field, quasi-static strain field and dynamic vibration signal. Time-frequency analysis (using short-time Fourier transform with a window length of 1024 points and a step size of 512 points) is performed on the dynamic vibration signal to extract the time-frequency ridge. Based on the time-varying dominant frequency range identified by the time-frequency ridge, a variable frequency band filter bank with a center frequency and bandwidth that adaptively change with time is constructed. The dynamic vibration signal is then filtered using the filter bank. Specifically, the frequency range with an energy proportion greater than 70% is taken as the time-varying dominant frequency. The center frequency and bandwidth of the variable frequency band filter bank are updated every 5 minutes accordingly, and filtering is performed. The filtered signal is divided into sub-data blocks according to a preset overlap duration (60 seconds, with adjacent overlaps of no less than 6 seconds). For each sub-data block, physical constraint waveform extension is first performed, and then adaptive noise complete set empirical mode decomposition is used for parallel computation to obtain preliminary intrinsic mode function components. Subsequently, variational mode extraction is performed on the target frequency band related to the natural frequency of the bridge to finely reconstruct the final intrinsic mode function components. Specifically, the dominant instantaneous frequency is extracted to construct a time-varying autoregressive moving average model and the signal is extended by 5% at both ends. Adaptive noise complete set empirical mode decomposition with white noise added twice is performed on the extended signal. Then, variational mode extraction is performed on the mode components related to the first 10 natural frequencies of the structure. After deleting the first and last 3% intervals of adjacent sub-data blocks, boundary relaxation splicing is performed at the position of the maximum cross-correlation coefficient in the overlapping area to synthesize the intrinsic mode function set in the full time domain. Based on a pre-built prior library of structural modal parameters, effective modal components with clear structural physical significance are screened out through modal confidence matrix mapping. The amplitudes of the effective modal components are arranged along the spatial axis to verify their similarity with the theoretical mode shape. The structural modal feature vectors are then extracted. Specifically, the pre-built prior library of structural modal parameters is retrieved, and the marginal peak frequency of each full-time modal component is calculated. Components with a difference of less than 0.05 Hz from the first 10 natural frequencies in the prior library and a cosine similarity of greater than 0.85 between the measured mode shape and the theoretical mode shape and the amplitude arrangement of the measurement points are identified as effective modal components and their feature vectors are extracted. The structural modal feature vector is compared with the health status reference vector obtained by statistical modeling of historical data under long-term health status. When the similarity distance exceeds a preset threshold, it is determined that the bridge has damage or stiffness change. The damage location is located according to the spatial mapping coordinates and a graded warning is issued. Specifically, the Mahalanobis distance is calculated between the feature vector and the mean reference vector of the Gaussian mixture model generated by the same process from no less than 90 days of historical health data. When the Mahalanobis distance is greater than 3.0, damage is determined and a warning is issued based on the corrected spatial coordinates.

[0010] The above technical solution further includes: Furthermore, a tight-buffered temperature sensing fiber is laid inside the sensing optical cable. The temperature sensing fiber has a gap between the fiber core and the outer sheath, and the core coating thickness is not less than 0.5 mm, so that it only senses changes in ambient temperature and does not respond to external stress. Based on the temperature sensing fiber, the spatial continuous temperature distribution is obtained at a measuring point every 1m along the longitudinal direction of the bridge. The bridge tower area and the main beam area are divided into the first temperature zone and the second temperature zone, respectively. Each zone independently stores the transfer function curves of strain coefficient and temperature coefficient as a function of temperature. The temperature range of the curve is -20℃ to 80℃ and the sampling interval is 1℃. After the demodulator measures the real-time temperature value of a certain measuring point, it first determines the zone to which the measuring point belongs, and then determines the current strain coefficient and temperature coefficient by linear interpolation in the transfer function curve of the corresponding zone based on the temperature value.

[0011] Furthermore, after the physical-guided neural network outputs the temperature field, quasi-static strain field, and dynamic vibration signal, a cross-sensor parameter verification step is also performed: Calculate the sensitivity coefficients of Rayleigh scattering phase to dynamic vibration, Brillouin shift to quasi-static strain, and Rayleigh shift to temperature, respectively. When the ratio of the sensitivity of dynamic vibration to Rayleigh scattering phase to Brillouin frequency shift is greater than 10, the ratio of the sensitivity of quasi-static strain to Brillouin frequency shift to Rayleigh frequency shift is greater than 5, and the ratio of the sensitivity of temperature to Brillouin frequency shift to Rayleigh frequency shift is less than 0.2, a 3x3 dual-frequency sensitivity matrix is ​​constructed using these three ratios. The three signals output by the neural network are used as observation values ​​and input to the extended Kalman filter. The extended Kalman filter sets the state vector to include three components: temperature, quasi-static strain, and dynamic vibration amplitude, and the observation vector to include two components: Brillouin frequency shift and Rayleigh scattering phase. The diagonal elements of the process noise covariance matrix are set to 0.01, and the diagonal elements of the observation noise covariance matrix are set to 0.1. The filtering period is consistent with the sampling period of the fiber optic demodulator. When the innovation sequence of 5 consecutive sampling points exceeds 3 times the standard deviation, the process noise covariance matrix is ​​triggered to adaptively amplify. The amplification factor is 2.0 to cope with the state tracking lag caused by sudden load changes. After state prediction and observation update iteration, the components with an absolute value of the output residual greater than 0.1 at each sampling point are corrected.

[0012] Furthermore, for the three bridge types of beam bridges, cable-stayed bridges, and suspension bridges, the reference passbands are preset to 0.1Hz to 30Hz, 0.05Hz to 50Hz, and 0.03Hz to 80Hz respectively in the system configuration file; When designing a variable frequency band filter bank, first read the bridge type parameters in the system configuration file to select the corresponding reference passband, and then superimpose the frequency range covered by the time-frequency ridge with an energy ratio greater than 70% in the spectrum during short-time Fourier transform onto the reference passband. The superposition rule is to take the smaller value of the lower limit of the reference passband and the lower limit of the recognition interval as the lower limit of the filter bank, and take the larger value of the upper limit of the reference passband and the upper limit of the recognition interval as the upper limit of the filter bank. When the recognition range is empty, only the reference passband is used. When the recognition range exceeds 3 times the reference passband, it is truncated to 3 times the range of the reference passband.

[0013] Furthermore, for each intrinsic mode function component obtained from the adaptive noise complete set empirical mode decomposition, after calculating the Pearson correlation coefficient with the corresponding sub-data block, if the Pearson correlation coefficient is less than 0.2, the component is identified as a noise component and directly removed; if the Pearson correlation coefficient is between 0.2 and 0.5, the component is marked as a suspected noise component and temporarily stored; if the Pearson correlation coefficient is greater than 0.5, the component is identified as a signal component and directly retained. Subsequent variational mode extraction is performed only on the signal component and the component marked as suspected noise. After the variational mode extraction is completed, the Pearson correlation coefficient between the reconstructed signal of the suspected noise component and the original sub-data block is calculated again. If it is still less than 0.5, it is finally removed.

[0014] Furthermore, when generating the health status reference vector, fiber optic vibration data were continuously collected for no less than 90 days before the bridge opened to traffic. Each day, from 00:00 to 23:59, 120 60-second signal segments were randomly selected. After each signal segment underwent the aforementioned decoupling, filtering, decomposition, and effective modal component calibration process, three characteristic parameters of each effective modal component were extracted: The center frequency corresponding to the marginal spectral peak, the ratio of the energy integral value within the frequency band to the total energy, and the structural damping ratio calculated by the logarithmic attenuation method; After collecting all feature vectors of the same measurement point and the same order mode, a Gaussian mixture model is fitted to the feature vector set. The number of Gaussian components is 3 to 5. The mean vector of each Gaussian component is taken as the health status reference vector. At the same time, the outer envelope of the ellipsoid formed by three times the standard deviation of each Gaussian component is taken as the Mahalanobis distance threshold confidence boundary.

[0015] A fiber optic sensing system for monitoring the health of bridge structures, comprising: The spatial self-calibration sensing unit is equipped with 1.0m segmented reference fiber rings in the optical cable at the bridge tower, pier and mid-span sections. The optical switch matrix is ​​connected to switch at intervals of 1 to 24 hours to connect the sensing segments between adjacent rings to the demodulator for calibration. A low coherence optical frequency domain reflection module is set up to monitor the fiber length at a sampling rate of 50Hz. When the change exceeds 0.01m, a spatial coordinate correction command is output. The physical-guided neural network decoupling module consists of a 5-layer fully connected network. The loss function is composed of the mean square error of the data and the physical residual constraint with weights of 0.7 and 0.3, respectively. The input is the spatially corrected Brillouin gain spectrum and Rayleigh scattering phase, and the output is the temperature field, quasi-static strain field and dynamic vibration signal. The adaptive variable filter module performs a short-time Fourier transform on the dynamic vibration signal with a window length of 1024 points and a step size of 512 points, extracts the frequency band with an energy ratio greater than 70% as the time-varying dominant frequency, and updates the variable frequency band filter bank coefficients and filters every 5 minutes. The hybrid mode decomposition module divides the filtered signal into blocks of 60 seconds each with an overlap of at least 6 seconds. It extracts the two ends of the extrapolated signal of the dominant frequency time-varying autoregressive moving average model, performs two adaptive noise complete set empirical mode decompositions on the extrapolated signal with white noise, performs variational mode extraction on the first 10 natural frequency related mode components, and performs boundary relaxation splicing at the position of the maximum cross-correlation coefficient in the overlapping area. The modal physics calibration and damage assessment module includes a priori library of structural modal parameters containing the first 10 natural frequencies and mode shapes. Components with a difference of less than 0.05 Hz between the marginal spectrum peak frequency and the prior frequency, and a cosine similarity greater than 0.85 between the measured mode shape and the theoretical mode shape, are marked as valid. After extracting the feature vector, the Mahalanobis distance between the Mahalanobis distance and the reference vector of the mean of the Gaussian mixture model obtained from 90 days of historical data is calculated. When the distance value is greater than 3.0, damage is determined and the location is determined based on the corrected coordinates.

[0016] Furthermore, the spatial self-calibrating sensing unit has a loose-tube temperature sensing fiber installed in the same cable, which only responds to temperature changes. Based on this, the system divides the bridge tower and the main beam into two temperature zones. Each zone stores a coefficient table with a temperature interval of 1°C, and the current coefficient is determined by linear interpolation based on the real-time temperature.

[0017] Furthermore, after obtaining the initial decomposed components, the hybrid mode decomposition module calculates the Pearson correlation coefficient between each component and the corresponding sub-data block, and only retains the components with a correlation coefficient greater than or equal to 0.2 to participate in the variational mode extraction.

[0018] Furthermore, when constructing the reference vector, the modal physical calibration and damage judgment module continuously collects data for 90 days, randomly extracts 120 60-second segments of signal each day, fits a Gaussian mixture model to the energy, center frequency and damping ratio eigenvectors obtained through the full process of each segment, and takes the mean of each component as the reference vector.

[0019] The present invention has the following beneficial effects: This invention employs a strategy that combines physically constrained waveform extension with boundary relaxation splicing. Even under the stringent condition of dividing data into numerous short-duration sub-blocks for parallel processing, it can fundamentally suppress endpoint effects and ensure that the modes spliced ​​across data blocks are continuous in phase and amplitude without artifact-free abrupt changes. This provides an algorithmic guarantee for high-fidelity analysis of long-distance, long-term monitoring data. Attached Figure Description

[0020] Figure 1 This is a flowchart of a bridge structural health monitoring method proposed in this invention; Figure 2 This is a system block diagram of an optical fiber sensing system for monitoring the health of bridge structures proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-2 As shown, the present invention is a method for monitoring the health of bridge structures, comprising the following steps: Multiple segmented reference fiber rings of known length are pre-installed in the sensing optical cable of the bridge to be monitored. Specifically, segmented reference fiber rings of 1.0m length are pre-installed at the bridge tower, pier, and mid-span section of the sensing optical cable. The optical switch matrix automatically switches at intervals of 1 to 24 hours. The strain coefficient and temperature coefficient of the sensing segment defined by the adjacent segmented reference fiber rings are independently calibrated using the known length. The change in fiber length is measured using a low coherence optical frequency domain reflection auxiliary system at a sampling rate of 50Hz. The spatial mapping coordinates of the scattered signal are dynamically corrected based on the change in fiber length to obtain distributed vibration and strain sensing data with accurate spatial location. When the cumulative length change exceeds 0.01m, a spatial coordinate translation correction command is sent to the demodulation system. The corrected sensing data (Brillouin gain spectrum and Rayleigh scattering phase) is input into the physical guided neural network decoupler. Its loss function is composed of a mean square error term and a physical residual term based on the coupled mechanical equation and the heat conduction equation, weighted by 0.7 and 0.3 respectively. The decoupled output is the temperature field, quasi-static strain field and dynamic vibration signal. Time-frequency analysis (using short-time Fourier transform with a window length of 1024 points and a step size of 512 points) is performed on the dynamic vibration signal to extract the time-frequency ridge. Based on the time-varying dominant frequency range identified by the time-frequency ridge, a variable frequency band filter bank with a center frequency and bandwidth that adaptively change with time is constructed. The dynamic vibration signal is then filtered using the filter bank. Specifically, the frequency range with an energy proportion greater than 70% is taken as the time-varying dominant frequency. The center frequency and bandwidth of the variable frequency band filter bank are updated every 5 minutes accordingly, and filtering is performed. The filtered signal is divided into sub-data blocks according to a preset overlap duration (60 seconds, with adjacent overlaps of no less than 6 seconds). For each sub-data block, physical constraint waveform extension is first performed, and then adaptive noise complete set empirical mode decomposition is used for parallel computation to obtain preliminary intrinsic mode function components. Subsequently, variational mode extraction is performed on the target frequency band related to the natural frequency of the bridge to finely reconstruct the final intrinsic mode function components. Specifically, the dominant instantaneous frequency is extracted to construct a time-varying autoregressive moving average model and the signal is extended by 5% at both ends. Adaptive noise complete set empirical mode decomposition with white noise added twice is performed on the extended signal. Then, variational mode extraction is performed on the mode components related to the first 10 natural frequencies of the structure. After deleting the first and last 3% intervals of adjacent sub-data blocks, boundary relaxation splicing is performed at the position of the maximum cross-correlation coefficient in the overlapping area to synthesize the intrinsic mode function set in the full time domain. When dividing into sub-data blocks, if the remaining signal duration of the tail segment is less than 60 seconds, the tail segment signal is merged with the second-to-last sub-data block into a sub-data block with a duration between 60 and 120 seconds for processing. After merging, the outer extension of the sub-data block and the absolute length of the deletion interval remain unchanged, that is, the outer extension is still 5% of the total length of the merged sub-data block, and the beginning and end deletion intervals are still 3% of the total length of the merged sub-data block. Based on a pre-built prior library of structural modal parameters, effective modal components with clear structural physical significance are screened out through modal confidence matrix mapping. The amplitudes of the effective modal components are arranged along the spatial axis to verify their similarity with the theoretical mode shape. The structural modal feature vectors are then extracted. Specifically, the pre-built prior library of structural modal parameters is retrieved, and the marginal peak frequency of each full-time modal component is calculated. Components with a difference of less than 0.05 Hz from the first 10 natural frequencies in the prior library and a cosine similarity of greater than 0.85 between the measured mode shape and the theoretical mode shape and the amplitude arrangement of the measurement points are identified as effective modal components and their feature vectors are extracted. The construction process of the prior library of structural modal parameters includes: establishing a three-dimensional finite element model using the bridge as-built drawings, with a mesh size no larger than 2m; inputting the material's elastic modulus, density, and boundary conditions; performing modal analysis to extract the first 10 natural frequencies and corresponding mode shape vectors; conducting environmental excitation modal tests before the bridge opens to traffic, deploying an accelerometer every 10m on the bridge deck to continuously collect 24 hours of environmental vibration data; using the random subspace method to identify the first 10 natural frequencies and mode shapes; when the frequency difference between the finite element calculation results and the modal test results is within 5%, taking the average of the two mode shape vectors as the mode shape stored in the prior library; when the frequency difference exceeds 5%, updating the finite element model based on the modal test results and recalculating; The structural modal feature vector is compared with the health status reference vector obtained by statistical modeling of historical data under long-term health status. When the similarity distance exceeds a preset threshold, it is determined that the bridge has damage or stiffness change. The damage location is located according to the spatial mapping coordinates and a graded warning is issued. Specifically, the Mahalanobis distance is calculated between the feature vector and the mean reference vector of the Gaussian mixture model generated by the same process from no less than 90 days of historical health data. When the Mahalanobis distance is greater than 3.0, damage is determined and a warning is issued based on the corrected spatial coordinates.

[0023] In this embodiment: Spatial self-calibration calibration based on segmented reference fiber rings and low-coherence optical frequency domain reflection: In the sensing optical cable, a 1.0m segmented reference fiber ring is pre-installed at the bridge tower, pier, and mid-span section. Its length and parameters are constant and known. The optical switch matrix automatically switches at 1-24 hour intervals, sequentially connecting the sensing segment between two adjacent segmented reference fiber rings to a distributed optical fiber demodulator for Brillouin and Rayleigh scattering measurements. For the sensing segment defined by adjacent reference rings, the strain coefficient and temperature coefficient of that segment of the sensing fiber are independently calibrated by utilizing the characteristic that the reference ring is unaffected by external strain. The basic relationship between Brillouin frequency shift and strain and temperature is: ,in, The Brillouin frequency shift is the change in frequency. For strain coefficient, For temperature coefficient, refer to the inner ring. =0, and its temperature change is sensed through the tight-buffered temperature sensing optical fiber laid in the same cable. That can be obtained For the sensing segment to be measured between adjacent reference rings, the already measured data is used. And temperature distribution, and then calculate the strain coefficient. To achieve independent scaling of each segment; Meanwhile, the low-coherence optical frequency domain reflection-assisted system measures the length change of the sensing fiber due to thermal expansion and contraction or bending in real time at a sampling rate of 50Hz. When the cumulative change exceeds the threshold of 0.01m, the system sends a spatial coordinate translation correction command to the demodulator to remap the original spatial coordinates. Revised to: This eliminates the systematic error caused by changes in fiber length in subsequent damage location accuracy; Physics-guided neural network multiphysics decoupling: 0 Brillouin gain spectrum after spatial coordinate correction and Rayleigh scattering phase As input, it is fed into a physically guided neural network decoupler. This network employs a 5-layer fully connected structure, and its loss function φ(L)L is a weighted sum of data fitting terms and physical residual constraint terms. ,in, Mean squared error between network output and actual annotation; physical residual term. It is constructed based on the fiber-structure coupling mechanical equation and the heat conduction equation, specifically in the following form: ,in, These represent the quasi-static strain field, temperature field, and dynamic vibration signal output by the network prediction, respectively. It is the frequency shift extracted from the input Brillouin gain spectrum. This represents the phase change during Rayleigh scattering. Based on a known physical model describing the effects of strain, temperature, and vibration on the Rayleigh phase, the network, after training, can synchronously output decoupled temperature fields, quasi-static strain fields, and dynamic vibration signals. To further improve the decoupling accuracy, a cross-validation step is introduced after the neural network output: taking advantage of the fact that dynamic vibration only responds significantly to the Rayleigh scattering phase, while the ratio of the sensitivity of quasi-static strain and temperature to Brillouin and Rayleigh frequency shifts differs (the former ratio is greater than 5, and the latter is less than 0.2), a dual-frequency sensitivity matrix is ​​constructed, and the three physical quantities are jointly optimally estimated by extended Kalman filtering to correct the decoupling results with residuals exceeding 0.1; Adaptive variable band filtering based on instantaneous frequency estimation: Time-frequency analysis is performed on the decoupled dynamic vibration signal v(t) using a short-time Fourier transform (STFT) with a window length of 1024 points and a step size of 512 points. Extracting time-frequency ridges from the time-frequency spectrum |V(t,f)|, that is, finding the instantaneous frequency with the maximum energy at each time t. Then calculate with Centered on the signal energy proportions within different bandwidths, the frequency range with an energy proportion greater than 70% is selected as the time-varying dominant frequency interval. ; Based on the identified time-varying frequency range, the coefficients of the variable-band filter bank are updated every 5 minutes, and the center frequency of the filter bank is set to... Bandwidth set to The filter bank uses a fourth-order Butterworth bandpass filter, whose transfer function squared magnitude is: The original dynamic vibration signal v(t) is subjected to zero-phase filtering using this time-varying filter to obtain the adaptively filtered signal. For different bridge types, it is necessary to superimpose the corresponding reference passband: 0.1Hz~30Hz for beam bridges, 0.05Hz~50Hz for cable-stayed bridges, and 0.03Hz~80Hz for suspension bridges, to ensure that the frequency band retained by the filter does not exceed the reasonable excitation frequency range of the bridge type. Block-based mixed mode decomposition and boundary relaxation concatenation: This involves combining the filtered signal... The data is divided into multiple sub-blocks with each segment lasting 60 seconds and an overlap of at least 6 seconds between adjacent blocks. For each sub-data block, m=1, 2, ..., the dominant instantaneous frequency is first extracted using the Hilbert transform, and a time-varying autoregressive moving average model is established based on this. The original signal length is then extended outwards by 5% from both ends of the data block to obtain the extended signal. ,right Perform adaptive noise-complete ensemble empirical mode decomposition, adding only two orders of white noise for integration, decomposing the signal into: Subsequently, the first 10 natural frequencies obtained from the bridge finite element model were analyzed. Variational mode extraction is performed on the corresponding intrinsic mode function components. Variational mode extraction is achieved by refining the reconstruction through solving the following constrained variational problem. Narrowband mode centered : ,in, By iteratively solving using the alternating direction multiplier method, fine-mode components that precisely correspond to the inherent physical modes of the structure are obtained. After all sub-data blocks are independently decomposed, boundary relaxation and splicing are performed: first, the first and last 3% portions of the decomposition result of each sub-block are deleted to eliminate endpoint effects. Then, in the remaining overlapping areas of adjacent blocks, the same-order mode components are spliced. and Calculate the normalized cross-correlation coefficient: ,Pick As alignment and splicing points, the two modal waveforms are seamlessly connected, ultimately synthesizing a set of intrinsic mode functions in the entire time domain. ; Physical calibration and feature extraction based on a prior library of structural modal parameters: A pre-constructed prior library of structural modal parameters is retrieved, containing the first 10 natural frequencies obtained from finite element modal analysis or environmental excitation modal testing under healthy bridge conditions. and the corresponding mode shape vector ; For each full-time eigenmode function Calculate its marginal spectrum: Extracting the peak frequency of the marginal spectrum If the difference between the natural frequency and a certain prior natural frequency satisfies: Then Initially labeled as candidate components of the nth-order mode, the amplitude of this mode component is then extracted at each spatial measurement point along the distributed optical fiber to form the measured mode shape vector. ,calculate Compared with the a priori theoretical mode shape Cosine similarity: ,when When the energy level is greater than 0.85, the modal component is retained as a valid modal component, and its eigenvector is extracted. The eigenvector is determined by the energy of the modal component. Center frequency It consists of the damping ratio calculated using the half-power bandwidth method; Damage assessment and spatial positioning early warning based on Mahalanobis distance: Under the healthy condition of the bridge before it is completed and opened to traffic, fiber optic vibration data are continuously collected for no less than 90 days. 120 signal segments, each 60 seconds long, are randomly extracted daily. After decoupling, filtering, decomposition, and physical calibration of each signal segment, the feature vectors of the effective modes are extracted. A Gaussian mixture model is fitted to all feature vector samples to obtain the mean of each Gaussian component. Covariance And weights, taking the average of the means of all Gaussian components or the mean of the principal components as the health status reference vector. And store its covariance matrix. During online monitoring, newly acquired signals are processed in the same way to obtain real-time feature vectors. Calculate its Mahalanobis distance relative to the health reference vector: ,when When the value is >3.0, it is determined that there is damage or stiffness change in the bridge structure. At this time, based on the corrected spatial mapping coordinates, the location of the fiber optic measuring point with abnormal modal characteristics is mapped to the actual bridge mileage coordinates to locate the specific location of the damage and issue a graded warning.

[0024] In one embodiment, a tight-buffered temperature sensing fiber is laid inside the sensing optical cable. The temperature sensing fiber has a gap between the fiber core and the outer sheath and the core coating thickness is not less than 0.5 mm, so that it only senses changes in ambient temperature and does not respond to external stress. Based on the temperature sensing fiber, the spatial continuous temperature distribution is obtained at a measuring point every 1m along the longitudinal direction of the bridge. The bridge tower area and the main beam area are divided into the first temperature zone and the second temperature zone, respectively. Each zone independently stores the transfer function curves of strain coefficient and temperature coefficient as a function of temperature. The temperature range of the curve is -20℃ to 80℃ and the sampling interval is 1℃. After the demodulator measures the real-time temperature value of a certain measuring point, it first determines the zone to which the measuring point belongs, and then determines the current strain coefficient and temperature coefficient by linear interpolation in the transfer function curve of the corresponding zone based on the temperature value.

[0025] In one embodiment, after the physical-guided neural network outputs the temperature field, quasi-static strain field, and dynamic vibration signal, a cross-sensor parameter verification step is also performed: Calculate the sensitivity coefficients of Rayleigh scattering phase to dynamic vibration, Brillouin shift to quasi-static strain, and Rayleigh shift to temperature, respectively. When the ratio of the sensitivity of dynamic vibration to Rayleigh scattering phase to Brillouin frequency shift is greater than 10, the ratio of the sensitivity of quasi-static strain to Brillouin frequency shift to Rayleigh frequency shift is greater than 5, and the ratio of the sensitivity of temperature to Brillouin frequency shift to Rayleigh frequency shift is less than 0.2, a 3x3 dual-frequency sensitivity matrix is ​​constructed using these three ratios. The three signals output by the neural network are used as observation values ​​and input to the extended Kalman filter. The extended Kalman filter sets the state vector to include three components: temperature, quasi-static strain, and dynamic vibration amplitude, and the observation vector to include two components: Brillouin frequency shift and Rayleigh scattering phase. The diagonal elements of the process noise covariance matrix are set to 0.01, and the diagonal elements of the observation noise covariance matrix are set to 0.1. The filtering period is consistent with the sampling period of the fiber optic demodulator. When the innovation sequence of 5 consecutive sampling points exceeds 3 times the standard deviation, the process noise covariance matrix is ​​triggered to adaptively amplify. The amplification factor is 2.0 to cope with the state tracking lag caused by sudden load changes. After state prediction and observation update iteration, the components with an absolute value of the output residual greater than 0.1 at each sampling point are corrected.

[0026] In one embodiment, for the three bridge types of beam bridges, cable-stayed bridges, and suspension bridges, the preset reference passbands in the system configuration file are 0.1Hz to 30Hz, 0.05Hz to 50Hz, and 0.03Hz to 80Hz, respectively. When designing a variable frequency band filter bank, first read the bridge type parameters in the system configuration file to select the corresponding reference passband, and then superimpose the frequency range covered by the time-frequency ridge with an energy ratio greater than 70% in the spectrum during short-time Fourier transform onto the reference passband. The superposition rule is to take the smaller value of the lower limit of the reference passband and the lower limit of the recognition interval as the lower limit of the filter bank, and take the larger value of the upper limit of the reference passband and the upper limit of the recognition interval as the upper limit of the filter bank. When the recognition range is empty, only the reference passband is used. When the recognition range exceeds 3 times the reference passband, it is truncated to 3 times the range of the reference passband.

[0027] In one embodiment, for each intrinsic mode function component obtained from the adaptive noise complete set empirical mode decomposition, after calculating the Pearson correlation coefficient with the corresponding sub-data block, if the Pearson correlation coefficient is less than 0.2, the component is identified as a noise component and directly removed; if the Pearson correlation coefficient is between 0.2 and 0.5, the component is marked as a suspected noise component and temporarily stored; if the Pearson correlation coefficient is greater than 0.5, the component is identified as a signal component and directly retained. Subsequent variational mode extraction is performed only on the signal component and the component marked as suspected noise. After the variational mode extraction is completed, the Pearson correlation coefficient between the reconstructed signal of the suspected noise component and the original sub-data block is calculated again. If it is still less than 0.5, it is finally removed.

[0028] In one embodiment, when generating the health status reference vector, fiber optic vibration data is continuously collected for no less than 90 days before the bridge opens to traffic. Each day, from 00:00 to 23:59, 120 60-second signal segments are randomly distributed. After each signal segment undergoes the decoupling, filtering, decomposition, and effective modal component calibration process, three characteristic parameters of each effective modal component are extracted: The center frequency corresponding to the marginal spectral peak, the ratio of the energy integral value within the frequency band to the total energy, and the structural damping ratio calculated by the logarithmic attenuation method; After collecting all feature vectors of the same measurement point and the same order mode, a Gaussian mixture model is fitted to the feature vector set. The number of Gaussian components is 3 to 5. The mean vector of each Gaussian component is taken as the health status reference vector. At the same time, the outer envelope of the ellipsoid formed by three times the standard deviation of each Gaussian component is taken as the Mahalanobis distance threshold confidence boundary.

[0029] A fiber optic sensing system for monitoring the health of bridge structures, comprising: The spatial self-calibration sensing unit is equipped with 1.0m segmented reference fiber rings in the optical cable at the bridge tower, pier and mid-span sections. The optical switch matrix is ​​connected to switch at intervals of 1 to 24 hours to connect the sensing segments between adjacent rings to the demodulator for calibration. A low coherence optical frequency domain reflection module is set up to monitor the fiber length at a sampling rate of 50Hz. When the change exceeds 0.01m, a spatial coordinate correction command is output. The physical-guided neural network decoupling module consists of a 5-layer fully connected network. The loss function is composed of the mean square error of the data and the physical residual constraint with weights of 0.7 and 0.3, respectively. The input is the spatially corrected Brillouin gain spectrum and Rayleigh scattering phase, and the output is the temperature field, quasi-static strain field and dynamic vibration signal. The adaptive variable filter module performs a short-time Fourier transform on the dynamic vibration signal with a window length of 1024 points and a step size of 512 points, extracts the frequency band with an energy ratio greater than 70% as the time-varying dominant frequency, and updates the variable frequency band filter bank coefficients and filters every 5 minutes. The hybrid mode decomposition module divides the filtered signal into blocks of 60 seconds each with an overlap of at least 6 seconds. It extracts the two ends of the extrapolated signal of the dominant frequency time-varying autoregressive moving average model, performs two adaptive noise complete set empirical mode decompositions on the extrapolated signal with white noise, performs variational mode extraction on the first 10 natural frequency related mode components, and performs boundary relaxation splicing at the position of the maximum cross-correlation coefficient in the overlapping area. The modal physics calibration and damage assessment module includes a priori library of structural modal parameters containing the first 10 natural frequencies and mode shapes. Components with a difference of less than 0.05 Hz between the marginal spectrum peak frequency and the prior frequency, and a cosine similarity greater than 0.85 between the measured mode shape and the theoretical mode shape, are marked as valid. After extracting the feature vector, the Mahalanobis distance between the Mahalanobis distance and the reference vector of the mean of the Gaussian mixture model obtained from 90 days of historical data is calculated. When the distance value is greater than 3.0, damage is determined and the location is determined based on the corrected coordinates.

[0030] In one embodiment, the spatial self-calibrating sensing unit has a loose-tube temperature sensing fiber installed in the same cable, which only responds to temperature changes. Based on this, the system divides the bridge tower and the main beam into two temperature zones. Each zone stores a coefficient table with a temperature interval of 1°C, and the current coefficient is determined by linear interpolation based on the real-time temperature.

[0031] In one embodiment, after obtaining the initial decomposed components, the hybrid mode decomposition module calculates the Pearson correlation coefficient between each component and the corresponding sub-data block, and only retains the components with a correlation coefficient greater than or equal to 0.2 to participate in the variational mode extraction.

[0032] In one embodiment, when constructing the reference vector, the modal physical calibration and damage judgment module continuously collects data for 90 days, randomly extracts 120 60-second segments of signal each day, fits a Gaussian mixture model to the energy, center frequency and damping ratio eigenvectors obtained through the full process of each segment, and takes the mean of each component as the reference vector.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the health of bridge structures, characterized in that, Includes the following steps: S1. Multiple segmented reference fiber rings of known length are pre-set in the sensing optical cable of the bridge to be monitored. The reference fiber rings are automatically switched by an optical switch matrix. The strain coefficient and temperature coefficient of the sensing segment defined by the adjacent segmented reference fiber rings are independently calibrated using the known length. The change in fiber length is measured using a low coherence optical frequency domain reflection auxiliary system. The spatial mapping coordinates of the scattered signal are dynamically corrected based on the change in fiber length to obtain distributed vibration and strain sensing data with accurate spatial location. S2. Correction instruction: Correct the spatial mapping coordinates of the scattered signal according to the instruction to obtain sensing data with corrected spatial mapping coordinates. Input the sensing data into the physical guidance neural network. The loss function of the physical guidance neural network consists of a data mean square error term and a residual constraint term based on the physical equation. The physical guidance neural network outputs the temperature field, quasi-static strain field and dynamic vibration signal. S3. Perform time-frequency analysis on the dynamic vibration signal to extract the time-frequency ridge line, construct a variable frequency band filter bank based on the time-varying dominant frequency range identified by the time-frequency ridge line, and use the filter bank to filter the dynamic vibration signal; S4. Divide the filtered signal into multiple sub-data blocks according to a preset overlap duration. For each sub-data block, first perform physical constraint waveform extension, then use adaptive noise complete set empirical mode decomposition for parallel computation to obtain preliminary intrinsic mode function components. Subsequently, perform variational mode extraction on the target frequency band related to the bridge's natural frequency to reconstruct the intrinsic mode function components. After deleting the first and last 3% intervals of adjacent sub-data blocks, perform boundary relaxation splicing at the position of the maximum cross-correlation coefficient in the overlapping area to synthesize the full-time domain intrinsic mode function set. S5. Based on the pre-built prior library of structural modal parameters, effective modal components with clear structural physical meaning are screened out by mapping the modal confidence matrix, and the amplitudes of the effective modal components are arranged along the spatial axis to verify similarity with the theoretical mode shape, thereby extracting the structural modal feature vector; S6. The similarity between the structural modal feature vector and the health status reference vector obtained by statistical modeling of historical data under long-term health status is calculated. When the similarity distance exceeds the preset threshold, it is determined that the bridge has damage or stiffness change, and the damage location is located according to the spatial mapping coordinates for graded early warning.

2. The method for monitoring the health of bridge structures according to claim 1, characterized in that, A tight-buffered temperature sensing fiber is laid inside the sensing optical cable. The temperature sensing fiber has a gap between the fiber core and the outer tube, and the core coating thickness is not less than 0.5 mm. Based on the temperature sensing fiber, the spatial continuous temperature distribution is obtained at a measuring point every 1m along the longitudinal direction of the bridge. The bridge tower area and the main beam area are divided into the first temperature zone and the second temperature zone, respectively. Each zone independently stores the transfer function curves of strain coefficient and temperature coefficient as a function of temperature. The temperature range of the curve is -20℃ to 80℃ and the sampling interval is 1℃. After the demodulator measures the real-time temperature value of a certain measuring point, it first determines the zone to which the measuring point belongs, and then determines the current strain coefficient and temperature coefficient by linear interpolation in the transfer function curve of the corresponding zone based on the temperature value.

3. The method for monitoring the health of bridge structures according to claim 1, characterized in that, After the physical-guided neural network outputs the temperature field, quasi-static strain field, and dynamic vibration signal, a cross-sensor parameter verification step is also performed: Calculate the sensitivity coefficients of Rayleigh scattering phase to dynamic vibration, Brillouin shift to quasi-static strain, and Rayleigh shift to temperature, respectively. When the ratio of the sensitivity of dynamic vibration to Rayleigh scattering phase to Brillouin frequency shift is greater than 10, the ratio of the sensitivity of quasi-static strain to Brillouin frequency shift to Rayleigh frequency shift is greater than 5, and the ratio of the sensitivity of temperature to Brillouin frequency shift to Rayleigh frequency shift is less than 0.2, a 3x3 dual-frequency sensitivity matrix is ​​constructed using these three ratios. The three signals output by the neural network are used as observation values ​​and input to the extended Kalman filter. The extended Kalman filter is configured such that the state vector contains three components: temperature, quasi-static strain, and dynamic vibration amplitude, and the observation vector contains two components: Brillouin frequency shift and Rayleigh scattering phase. After state prediction and observation update iteration, the components with an absolute value of the output residual greater than 0.1 at each sampling point are corrected.

4. The method for monitoring the health of bridge structures according to claim 1, characterized in that, For beam bridges, cable-stayed bridges, and suspension bridges, the reference passbands are preset to 0.1Hz to 30Hz, 0.05Hz to 50Hz, and 0.03Hz to 80Hz respectively in the system configuration file; When designing a variable frequency band filter bank, first read the bridge type parameters in the system configuration file to select the corresponding reference passband, and then superimpose the frequency range covered by the time-frequency ridge with an energy ratio greater than 70% in the spectrum during short-time Fourier transform onto the reference passband. The superposition rule is to take the smaller value of the lower limit of the reference passband and the lower limit of the recognition interval as the lower limit of the filter bank, and take the larger value of the upper limit of the reference passband and the upper limit of the recognition interval as the upper limit of the filter bank. When the recognition range is empty, only the reference passband is used. When the recognition range exceeds 3 times the reference passband, it is truncated to 3 times the range of the reference passband.

5. The method for monitoring the health of bridge structures according to claim 1, characterized in that, For each intrinsic mode function component obtained from the adaptive noise complete set empirical mode decomposition, after calculating the Pearson correlation coefficient with the corresponding sub-data block, if the Pearson correlation coefficient is less than 0.2, the component is identified as a noise component and directly removed. If the Pearson correlation coefficient is between 0.2 and 0.5, the component is marked as a suspected noise component and temporarily stored. If the Pearson correlation coefficient is greater than 0.5, the component is identified as a signal component and directly retained. Variational mode extraction is performed only on the signal component and the component marked as suspected noise. After the variational mode extraction is completed, the Pearson correlation coefficient between the reconstructed signal of the suspected noise component and the original sub-data block is calculated again. If it is still less than 0.5, it is finally removed.

6. The method for monitoring the health of bridge structures according to claim 1, characterized in that, When generating the health status reference vector, fiber optic vibration data were continuously collected for no less than 90 days before the bridge opened to traffic. Each day, 120 60-second segments of signal were randomly selected. After decoupling, filtering, decomposition, and effective modal component calibration, three characteristic parameters of each effective modal component were extracted: The center frequency corresponding to the marginal spectral peak, the ratio of the energy integral value within the frequency band to the total energy, and the structural damping ratio calculated by the logarithmic attenuation method; After collecting all feature vectors of the same measurement point and the same order mode, a Gaussian mixture model is fitted to the feature vector set. The mean vector of each Gaussian component is taken as the health status reference vector, and the outer envelope of the ellipsoid formed by three times the standard deviation of each Gaussian component is taken as the Mahalanobis distance threshold confidence boundary.

7. A fiber optic sensing system for monitoring the structural health of bridges, characterized in that, include: The spatial self-calibration sensing unit is equipped with 1.0m segmented reference fiber rings in the optical cable at the bridge tower, pier and mid-span sections. The optical switch matrix is ​​connected to switch at intervals of 1 to 24 hours to connect the sensing segments between adjacent rings to the demodulator for calibration. A low coherence optical frequency domain reflection module is set up to monitor the fiber length at a sampling rate of 50Hz. When the change exceeds 0.01m, a spatial coordinate correction command is output. The physical-guided neural network decoupling module consists of a 5-layer fully connected network. The loss function is composed of the mean square error of the data and the physical residual constraint with weights of 0.7 and 0.3, respectively. The input is the spatially corrected Brillouin gain spectrum and Rayleigh scattering phase, and the output is the temperature field, quasi-static strain field and dynamic vibration signal. The adaptive variable filter module performs a short-time Fourier transform on the dynamic vibration signal with a window length of 1024 points and a step size of 512 points, extracts the frequency band with an energy ratio greater than 70% as the time-varying dominant frequency, and updates the variable frequency band filter bank coefficients and filters every 5 minutes. The hybrid mode decomposition module divides the filtered signal into blocks of 60 seconds each with an overlap of at least 6 seconds. It extracts the two ends of the extrapolated signal of the dominant frequency time-varying autoregressive moving average model, performs two adaptive noise complete set empirical mode decompositions on the extrapolated signal with white noise, performs variational mode extraction on the first 10 natural frequency related mode components, and performs boundary relaxation splicing at the position of the maximum cross-correlation coefficient in the overlapping area. The modal physics calibration and damage assessment module includes a priori library of structural modal parameters containing the first 10 natural frequencies and mode shapes. Components with a difference of less than 0.05 Hz between the marginal spectrum peak frequency and the prior frequency, and a cosine similarity greater than 0.85 between the measured mode shape and the theoretical mode shape, are marked as valid. After extracting the feature vector, the Mahalanobis distance between the Mahalanobis distance and the reference vector of the mean of the Gaussian mixture model obtained from 90 days of historical data is calculated. When the distance value is greater than 3.0, damage is determined and the location is determined based on the corrected coordinates.

8. The fiber optic sensing system for bridge structural health monitoring according to claim 7, characterized in that, The spatial self-calibrating sensing unit has a loose-tube temperature sensing fiber installed in the same cable, which only responds to temperature changes. Based on this, the system divides the bridge tower and the main beam into two temperature zones. Each zone stores a coefficient table with a temperature interval of 1°C, and the current coefficient is determined by linear interpolation based on the real-time temperature.

9. The fiber optic sensing system for bridge structural health monitoring according to claim 7, characterized in that, After obtaining the initial decomposed components, the hybrid mode decomposition module calculates the Pearson correlation coefficient between each component and the corresponding sub-data block, and only retains the components with a correlation coefficient greater than or equal to 0.2 to participate in the variational mode extraction.

10. The fiber optic sensing system for bridge structural health monitoring according to claim 7, characterized in that, When constructing the reference vector, the modal physical calibration and damage judgment module continuously collects data for 90 days, randomly extracts 120 60-second segments of signal each day, and fits a Gaussian mixture model to the energy, center frequency and damping ratio eigenvectors obtained by the full-process processing of each segment, and takes the mean of each component as the reference vector.

Images