A method and system for modeling and diagnosing damage evolution in steel structures

Abstract

This invention discloses a method and system for modeling and diagnosing damage evolution in steel structures, relating to the field of steel structure monitoring data analysis technology. The method includes synchronous denoising of multi-source monitoring time series of steel structures to generate steel structure observation feature vectors. Based on these feature vectors, a component damage state boundary mapping is generated, and a monotonically gated damage evolution state space model is trained. The monotonically gated damage evolution state space model is used to perform posterior inversion of steel structure damage to generate component damage confidence intervals, recursively completing the steel structure damage trend early warning determination. This invention forms an end-to-end data processing link from multi-source time series cleaning and feature construction, component-level monotonically gated evolution modeling, to Bayesian posterior inversion and forward recursive early warning, ensuring stable input features while making the damage state bounded, recursive, and capable of outputting confidence intervals, supporting the integrated realization of component-level diagnosis and trend early warning.

Description

Technical Field

[0001] This invention relates to the field of steel structure monitoring data analysis technology, specifically to a method and system for modeling and diagnosing steel structure damage evolution. Background Technology

[0002] As the service life of steel structures such as bridges and large factories increases, steel structure health monitoring is gradually shifting from hardware acquisition to a software algorithm-driven data processing paradigm. Multi-source sensing generates high-dimensional time series, and the computer side performs data cleaning through synchronization alignment, noise reduction, and sliding window segmentation. Subsequently, feature vectors are constructed using power spectral density, time-domain statistics, and modal parameters, and machine learning, state-space modeling, and Bayesian inference are further introduced to achieve damage identification, state estimation, and trend prediction.

[0003] Existing steel structure damage diagnosis algorithms still have key shortcomings. Multi-source time series are often affected by differences in sampling rates, time delay drift, and missing data. If synchronization and quality control are insufficient, the comparability of feature vectors across different windows decreases, leading to fluctuations in diagnostic results due to noise. Many methods tend to focus on damage index regression or classification at a specific moment, lacking recursive damage evolution dynamics, especially lacking constraints on physical consistency such as damage boundaries and monotonically non-decreasing values, which easily leads to damage regression or spurious evolution. Common data-driven models and inference processes are disconnected, making it difficult to simultaneously provide component-level uncertainty quantification and recursive predictions for the future time domain under the same computational kernel. Furthermore, it is difficult to systematically inject component-level differences into the model, enabling the same observation to form differentiated evolution paths on different components. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: existing steel structure damage diagnosis methods have problems such as difficulty in unifying and aligning the timing of multi-source monitoring, resulting in unstable observation features; lack of evolution modeling with boundary and monotonic constraints, leading to damage regression and false evolution; difficulty in outputting component damage confidence intervals and performing future trend recursion and early warning in the inference process; and how to achieve component-level posterior inversion and trend early warning determination under the same monotonic gated state space calculation kernel.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for modeling and diagnosing damage evolution of steel structures, including time-series synchronous denoising processing of multi-source monitoring of steel structures to generate steel structure observation feature vectors.

[0007] Based on the feature vectors observed in the steel structure, a boundary mapping of the component damage state is generated, and a monotonically gated damage evolution state space model is trained.

[0008] A monotonically gated damage evolution state space model is used to perform posterior inversion of steel structure damage to generate component damage confidence intervals, and the steel structure damage trend early warning judgment is recursively completed.

[0009] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis method described in this invention, the multi-source monitoring time-series synchronous denoising process for the steel structure includes: receiving the original time-series data from the steel structure monitoring channels and writing it into a unified time buffer; performing resampling and alignment according to the timestamps of each channel; calculating the cross-correlation alignment offset based on a selected steel structure reference channel and correcting the time delay of each channel; performing outlier removal and missing segment interpolation on the aligned data; performing bandpass filtering on each channel and superimposing wavelet threshold denoising; and outputting the synchronously denoised multi-source monitoring time-series dataset.

[0010] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis method described in this invention, the generation of the steel structure observation feature vector includes: truncating the synchronously denoised multi-source monitoring time-series dataset into a window sequence using a sliding window method; calculating the power spectral density for each window and integrating the spectral energy according to a preset steel structure modal frequency band set to obtain the frequency band energy component; performing time-domain statistical calculations on the same window to obtain the root mean square, kurtosis, and peak factor components; and performing modal parameter identification based on the response correlation matrix on the same window to obtain the modal frequency and damping ratio components. Finally, concatenating the components in channel order and normalizing them yields the steel structure observation feature vector.

[0011] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis method described in this invention, the step of generating component damage state boundary mapping based on steel structure observation feature vectors includes: establishing unbounded component evolution latent variables for each component of the steel structure and forming a latent variable vector; mapping each unbounded component evolution latent variable to a component damage state using a Sigmoid function, limiting the component damage state to the interval between 0 and 1, and forming a component damage state vector; generating component hierarchical codes based on steel structure component parameters and storing them in association with the steel structure observation feature vectors; initializing the unbounded component evolution latent variables with the steel structure observation feature vectors corresponding to the healthy baseline window, and using the initialization result as the initial state input.

[0012] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis method described in this invention, the training of the monotonic gated damage evolution state space model includes: constructing a state transition operator and an observation operator with the component damage state vector as the state; the state transition operator calculates a non-negative evolution increment for the unbounded latent variables of each component's evolution and multiplies it by the component evolution gating factor; the component evolution gating factor is calculated by the steel structure observation feature vector and the component hierarchical encoding through a gating network, and its value ranges from 0 to 1. When the component evolution gating factor is lower than the steel structure component evolution gating trigger threshold, the evolution increment of the corresponding component is set to zero. The observation operator maps the component damage state vector and the component hierarchical encoding to a predicted observation feature vector. During training, the difference between the measured steel structure observation feature vector and the predicted observation feature vector is used as the observation feature residual, and a gating sparsity constraint of summing the absolute values ​​of the component evolution gating factor and a state change constraint of summing the absolute values ​​of the differences between the component damage state vectors at adjacent time points are superimposed. The parameters of the state transition operator and the observation operator are trained simultaneously through gradient updates.

[0013] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis method described in this invention, the step of generating component damage confidence intervals by performing posterior inversion of steel structure damage using a monotone-gated damage evolution state space model includes: reading the trained monotone-gated damage evolution state space model and performing recursive Bayesian estimation based on particle filtering on the real-time input sequence of steel structure observation feature vectors. At each time step, the component damage state vector from the previous time step is time-advanced using a state transition operator to obtain a priori component damage state set. Then, the corresponding predicted observation feature vector is calculated using an observation operator, and the innovation quantity is calculated together with the measured steel structure observation feature vector. The particle weights are calculated based on the innovation quantity, and the component damage state set is updated to obtain a posterior set. The quantile intervals are calculated for the posterior set according to the steel structure component damage confidence coverage level, and the component damage confidence intervals are output and stored in conjunction with the component number.

[0014] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis method described in this invention, the recursive completion of the steel structure damage trend early warning determination includes: using the posterior set corresponding to the component damage confidence interval as the initial set, calling the state transition operator to perform forward recursion according to the preset steel structure damage trend recursion step size and recursion step number to obtain a predicted set of component damage states in the future time domain. For each component, the proportion of component damage states in the statistical prediction set exceeding the steel structure component damage severity determination threshold is taken as the over-limit probability, and the average increment of component damage states at adjacent recursion times is calculated as the damage evolution rate. When the over-limit probability is not less than the steel structure damage early warning confidence determination threshold or the damage evolution rate is not less than the steel structure damage acceleration determination threshold, the early warning identifier and trigger time of the corresponding component are output.

[0015] As a preferred embodiment of the steel structure damage evolution modeling and diagnosis system of the present invention, it includes a multi-source temporal synchronous feature extraction module, a boundary mapping training evolution module, and a posterior inversion recursive early warning judgment module.

[0016] The multi-source time-series synchronous feature extraction module is used for time-series noise reduction processing of multi-source monitoring of steel structures to generate steel structure observation feature vectors.

[0017] The boundary mapping training and evolution module is used to generate component damage state boundary mappings based on steel structure observation feature vectors and to train a monotonically gated damage evolution state space model.

[0018] The posterior inversion recursive early warning judgment module is used to perform posterior inversion of steel structure damage using a monotonic gated damage evolution state space model to generate component damage confidence intervals and recursively complete the early warning judgment of steel structure damage trend.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for modeling and diagnosing damage evolution in steel structures.

[0020] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for modeling and diagnosing damage evolution in steel structures.

[0021] The beneficial effects of this invention are: By performing resampling alignment, cross-correlation delay correction, anomaly missing data handling, and bandpass and wavelet denoising on the multi-source monitoring time series of steel structures, a consistent time reference and data quality control across channels and windows were achieved. Subsequently, the cleaned time series was transformed into observation feature vectors such as power spectral density band energy, time-domain statistics, and modal parameters. This compressed the original signal into a stable, comparable, and learnable numerical representation, facilitating the next step of training the evolution model with a unified input and reducing noise-induced feature drift.

[0022] By establishing unbounded latent variables for component evolution and forming a boundary mapping of component damage states via a Sigmoid function, the component damage states are computationally constrained to the 0-1 interval and can be continuously updated. Furthermore, by introducing component-level encoding and a gating network, and combining non-negative evolutionary increments with a zero-gating threshold mechanism to train state transition and observation operators, damage evolution is modeled as a component-level, recursive, monotonic, and piecewise sparse state-space process. This provides an inferable dynamic and observational mapping relationship under the same kernel for the next step.

[0023] By reading the trained monotone-gated damage evolution state-space model and employing particle filtering recursive Bayesian estimation, the posterior distribution update of component damage state is achieved under the driving force of real-time observed feature vectors. Furthermore, the component damage confidence interval is output as a quantile, allowing the diagnostic results to include uncertainty and be stored in conjunction with the component number. Then, the posterior set is forward recursively used to form a future prediction set, and trend warning determination is completed based on the dual criteria of exceedance probability and damage evolution rate. This unifies diagnosis, prediction, and warning within the same inference chain. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is an overall flowchart of a steel structure damage evolution modeling and diagnosis method provided in Embodiment 1 of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for modeling and diagnosing damage evolution in steel structures is provided, comprising: S1: Perform time-series synchronous noise reduction processing on the multi-source monitoring of the steel structure to generate the steel structure observation feature vector.

[0028] The system receives raw time-series data from the steel structure monitoring channels and writes it to a unified time buffer. Resampling and alignment are then performed according to the timestamps of each channel. The cross-correlation alignment offset is calculated using a selected steel structure reference channel as a benchmark, and the time delay of each channel is corrected. Outlier removal and missing segment interpolation are performed on the aligned data. Bandpass filtering is applied to each channel, and wavelet threshold denoising is superimposed, outputting a synchronized and denoised multi-source monitoring time-series dataset. The steel structure reference channel is selected as the monitoring channel with the lowest missing rate and the largest root mean square within the effective window. If there are ties, the channel with the smallest channel number is selected.

[0029] Furthermore, a unified time buffer uses a fixed sampling rate of 200Hz to resample each channel. The fixed sampling rate is set according to the rule that it should not be lower than four times the highest frequency of the preset steel structure mode frequency band set. When the original sampling rate is lower than the fixed sampling rate, anti-aliasing low-pass filtering is performed first, followed by upsampling; when it is higher than the fixed sampling rate, low-pass filtering is performed first, followed by downsampling. The cutoff frequency of the anti-aliasing low-pass filter is set to 0.45 times the fixed sampling rate, i.e., 90Hz, and the filter is a fourth-order Butterworth digital filter.

[0030] The cross-correlation alignment offset is calculated by taking the offset corresponding to the maximum cross-correlation value within a search range of ±1.0 s. When the maximum normalized cross-correlation coefficient is less than 0.60, the window alignment is deemed unreliable and the window is marked as invalid, not proceeding to subsequent feature extraction. The 0.60 threshold is derived from the minimum acceptable correlation constraint for alignment consistency in engineering monitoring and is a fixed value. Invalid windows are not included in the statistics for power spectral density calculation, time-domain statistics calculation, modal parameter identification, and normalization processing, and their feature vectors are either not output or are empty with an invalid flag.

[0031] Outlier removal employs the Hampel filtering criterion. Within a sliding median window of 11 sampling points, if the deviation of a sample from the window's median is greater than three times the median absolute deviation (MAD), it is identified as an outlier and removed. The three-times-MAD threshold is derived from the three-times-difference principle and is a fixed value. For missing segments, the interpolation rule is as follows: cubic spline interpolation is used when the duration of the missing segment does not exceed 2.0 seconds; otherwise, the window containing the missing segment is marked as invalid.

[0032] Furthermore, the passband of the bandpass filter is fixed at 0.5Hz to 50Hz. Its lower limit is used to suppress temperature drift and quasi-static trends, while its upper limit is used to suppress high-frequency noise and cover the preset steel structure mode frequency band set. Wavelet thresholding denoising uses a 5-level decomposition of the db4 wavelet and a soft thresholding method. The threshold is set according to the general thresholding rule as the noise standard deviation multiplied by 3.90, where 3.90 is a fixed coefficient. The noise standard deviation is estimated by dividing the median of the absolute values ​​of the first-level detail coefficients by 0.6745.

[0033] The synchronized and denoised multi-source monitoring time-series dataset is truncated into a window sequence using a sliding window method. For each window, the power spectral density is calculated, and the spectral energy is integrated against a preset set of steel structure mode bands to obtain the band energy components. Time-domain statistics are calculated for the same window to obtain the root mean square, kurtosis, and peak factor components. Modal parameter identification based on the response correlation matrix is ​​performed on the same window to obtain the modal frequency and damping ratio components. The components are then concatenated in channel order and normalized to obtain the steel structure observation feature vector.

[0034] Furthermore, the sliding window length is fixed at 2048 sampling points and the window step size is fixed at 1024 sampling points, corresponding to a 50% overlap rate. Within each window, a Hanning window is first applied to reduce spectral leakage before power spectral density calculation. The window length and overlap rate are preset fixed parameters to ensure a repeatable trade-off between temporal and frequency resolution for the features.

[0035] The power spectral density was calculated using the Welch method, with a fixed segment length of 512 points and a fixed segment overlap rate of 50%. The obtained power spectrum was then integrated segmentally within a preset steel structure modal frequency band set. The preset steel structure modal frequency band set consisted of four bands: [0.5, 5] Hz, [5, 15] Hz, [15, 30] Hz, and [30, 50] Hz. The integration results for each band were used as the band energy components and arranged in channel order. The highest frequency in the preset steel structure modal frequency band set was 50 Hz, which satisfies a ratio of at least four times the fixed sampling rate of 200 Hz.

[0036] Modal parameter identification based on the response correlation matrix employs the SSI-COV method. A response correlation matrix is ​​constructed for each window, and candidate poles are solved within the model order range of 10 to 40. Pole stability is determined using fixed thresholds: the relative change in modal frequency between adjacent orders does not exceed 1.0%, the relative change in damping ratio does not exceed 5.0%, and the MAC of the corresponding mode shape is not less than 0.95. When all three conditions are met, the modal frequency and damping ratio are output as modal parameter components. The 1.0%, 5.0%, and 0.95 thresholds are derived from engineering constraints in the stability plot screening and are fixed in value. MAC is the ratio of the squares of the normalized inner products of the corresponding mode shape vectors between adjacent orders, ranging from 0 to 1, and is used to measure mode shape consistency.

[0037] Normalization uses the mean and standard deviation of the healthy baseline window set to perform z-score normalization on each feature dimension. The normalized feature values ​​are then truncated to suppress extreme points. The truncation rule is to set values ​​greater than +5 to +5 and values ​​less than -5 to -5. The ±5 threshold is derived from an anomaly constraint of five times the standard deviation after standardization and is fixed in value.

[0038] It should be noted that this step achieves a common time reference for multi-source data by using a fixed sampling rate and cross-correlation alignment. Hampel data and missing data rules are then used to ensure data validity. After noise suppression using bandpass and wavelet methods, a sliding window is used to construct stable samples. Simultaneously, frequency domain energy, time domain statistics, and modal parameters are extracted, and healthy baseline normalization and truncation are performed to obtain repeatable and drift-resistant observation feature vectors, providing consistent input for subsequent damage evolution modeling.

[0039] S2: Generate component damage state boundary mapping based on steel structure observation feature vectors, and train a monotonically gated damage evolution state space model.

[0040] For each component of the steel structure, an unbounded latent variable for component evolution is established and a latent variable vector is formed. Each unbounded latent variable is mapped to a component damage state using a Sigmoid function, limiting the damage state to the interval between 0 and 1, and forming a component damage state vector. Component hierarchical codes are generated based on the steel structure component parameters and stored in association with the steel structure observation feature vector. The unbounded latent variables for component evolution are initialized using the steel structure observation feature vector corresponding to the healthy baseline window, and the initialization result is used as the initial state input.

[0041] Furthermore, the components are numbered. Perform indexing. Values ​​range from 1 to , This represents the total number of steel structure components. Component parameters include component type, cross-sectional geometric parameters, material grade, connection type, and node topology number. Discrete parameters are first individually encoded, and continuous parameters are linearly normalized according to engineering dimensions. Then, a component-level code is output through a component coding network. This network is a two-layer fully connected structure with a fixed output dimension of 16. The component-level code for each component is bound to its component number for storage.

[0042] The healthy baseline window is the first 300 windows deemed valid in S1. The mean of the steel structure observation feature vectors of this set of healthy baseline windows is calculated as the healthy baseline observation vector. The initialization of the unbounded component evolution latent variables is achieved by inputting the healthy baseline observation vector and the corresponding component level encoding into an initialization network and outputting the initialization value. The initialization network is a two-layer fully connected structure, and the last layer does not use an activation function to ensure that the unbounded component evolution latent variables can take any real number.

[0043] Furthermore, the Sigmoid mapping adopts a fixed form and is consistently used in subsequent state transition operator calculations to ensure boundary consistency. The time advancement step size of the unbounded component evolution latent variables is consistent with the window step size of S1, fixed at 5.12 seconds. The unbounded component evolution latent variables are mapped to component damage states in the interval of 0 to 1, which are used as unified inputs for the state transition operator and the observation operator. The formula for calculating the boundary mapping of component damage states is expressed as follows:

[0044] in, Number the component. The window number, with a value of 1~ , Indicates the total number of windows. For the evolution of unbounded components, there are hidden variables. This indicates the damage state of the component.

[0045] In each window Inside, calculate the values ​​of all components. This forms a component damage state vector, which serves as the input for subsequent gating calculations, state transitions, and observation mappings, ensuring that the component damage state output during any training or inference process does not exceed the bounds.

[0046] A state transition operator and an observation operator are constructed, with the component damage state vector as the state. The state transition operator calculates a non-negative evolution increment for the unbounded latent variables of each component's evolution and multiplies it by the component evolution gating factor. The component evolution gating factor is calculated by a gating network from the steel structure observation feature vector and the component hierarchical encoding, and its value ranges from 0 to 1. When the component evolution gating factor is lower than the steel structure component evolution gating trigger threshold, the evolution increment of the corresponding component is set to zero. The observation operator maps the component damage state vector and the component hierarchical encoding to the predicted observation feature vector. During training, the difference between the measured steel structure observation feature vector and the predicted observation feature vector is used as the observation feature residual, and gating sparsity constraints of summing the absolute values ​​of the component evolution gating factor and state change constraints of summing the absolute values ​​of the differences between the component damage state vectors at adjacent time points are superimposed. The parameters of the state transition operator and the observation operator are trained simultaneously through gradient updates.

[0047] Furthermore, the gating network takes the steel structure observation feature vector of the current window and the component hierarchical encoding concatenation vector of the corresponding component as input, outputs a component evolution gating factor, and uses Sigmoid as the output activation to limit it to 0 to 1. The steel structure component evolution gating trigger threshold is fixed at 0.30. The setting rule is that 0.30 comes from the upper quantile value of the gating factor obtained by offline statistics of historical health baseline data plus a margin of 0.10, and a lower limit of 0.30 is set. In this embodiment, it is fixed at 0.30.

[0048] The state transition operator uses a non-negative function output for the evolution increment of the latent variables of unbounded component evolution to ensure monotonicity; the non-negative function is Softplus. When the component evolution gating factor is lower than the steel structure component evolution gating trigger threshold, the evolution increment is zeroed by setting the effective gating value after gating bias to zero, and this zeroing rule remains consistent in training and online inference.

[0049] The observation operator employs a component contribution weighted aggregation structure. First, each component's hierarchical encoding is linearly mapped to a component observation contribution vector. Then, the contributions of each component are normalized and weighted according to the component's damage state to obtain the predicted observation feature vector. The first 80% of the window sequence is used as the training set, and the last 20% as the validation set, arranged chronologically. Training uses a sequence mini-batch approach with a fixed sequence length of 50 windows. The optimizer is Adam with an initial learning rate of 0.001, and updates stop when the residuals of the validation set observation features do not decrease for 10 consecutive rounds.

[0050] The component evolution gating factor is calculated from the steel structure observation feature vector and the component hierarchical code. This factor controls whether the component is allowed to evolve within a given window. The formula for calculating the component evolution gating factor is as follows:

[0051] in, This is the gating factor for component evolution. This represents the characteristic vector for steel structure observation. Encode the component hierarchy. This indicates vector concatenation. and These are the parameters for the gating network.

[0052] For each window With each component calculate It is compared with the steel structure component evolution gating trigger threshold of 0.30 to determine whether the unbounded component evolution latent variable of the component obtains a non-zero evolution increment in subsequent state transitions.

[0053] The evolutionary latent variables of unbounded components are advanced by multiplying the non-negative evolutionary increment with the effective gating value after gating bias, thus simultaneously satisfying the non-negativity of the evolutionary increment and the zeroing of the low gating in numerical computation. The evolutionary latent variables of unbounded components in the next window are represented as follows:

[0054] in, This represents the hidden variable representing the evolution of unbounded components in the next window. The window time step is fixed at 5.12 seconds. To evolve incremental networks. Indicating Evolutionary Incremental Networks The set of trainable parameters is determined during the training phase through gradient updates. After training is completed, The nonnegative evolutionary increments of latent variables in the evolution of unbounded components are solidified and used during online inference. It is a non-negative mapping. The evolution-gated trigger threshold for steel structure components is fixed at 0.30. When hour, The evolution increment of this component in this window is zero. When At that time, proceed according to non-negative evolutionary increments. Then, the new component damage state is obtained through the component damage state boundary mapping formula. This is used for gating and observation prediction calculations in the next window. The evolutionary incremental network uses... The input is a three-layer fully connected structure with a ReLU hidden layer and a linear output layer. The output is followed by Softplus to obtain the non-negative evolution increment.

[0055] The component damage state vector is mapped to the component hierarchical encoding to form a predicted observation feature vector. Component contribution-weighted aggregation is used to maintain realizability and reproducibility. The formula for calculating the predicted observation feature vector is expressed as:

[0056] in, To predict the observed feature vector. To observe the operator parameters. To prevent the constant in the denominator from being zero (which is fixed at 10). 6 ). All are component number indices and have values ​​from 1 to... . This represents the total number of steel structure components.

[0057] In training and online inference, the components are first obtained from the component damage state boundary mapping formula and the unbounded component evolution latent variables of the next window. Then calculate using the formula for predicting observed eigenvectors. and will Compared with the measured Perform residual calculations to drive parameter updates.

[0058] Using the observed feature residuals as the principal term, and superimposed with gating sparsity constraints and state change constraints, joint constraint training of the gating and state sequences is achieved. The training loss is expressed as:

[0059] in, This is due to training losses. The gating sparse constraint weight is fixed at 0.01 in this embodiment. The weight for the state change constraint is fixed at 0.05 in this embodiment. For each component The vector formed by the vector. This represents the total number of windows participating in this loss calculation, i.e., the number of steel structure observation feature vector windows contained in a training sequence sample. In this embodiment of the invention, . Indicates the first The component damage state vector at each window time.

[0060] During the training phase, computation is performed on each sequence batch with a length of 50. Updated with Adam , , , , Parameters. After training, the parameters are fixed, and the calculation of the component evolution gating factor, the unbounded component evolution latent variable of the next window, and the predicted observation feature vector are used as the calculation kernel for S3 posterior inversion and trend recursion.

[0061] It should be noted that this step uses unbounded component evolution latent variables and Sigmoid boundary mapping to stably constrain the component damage state to 0~1, and introduces Softplus non-negative evolution increments and zero-gating bias thresholds during state transitions, making the component damage evolution with the window sequence computationally monotonic and piecewise sparsity. Then, component-level encoding is used to participate in gating and observation mapping, enabling the same observation feature to generate differentiated evolution paths on different components. Compared to schemes that only perform single-time-point identification or unconstrained temporal regression, this invention avoids damage regression and noise-driven spurious evolution, and achieves a unified state-space kernel for evolutionary modeling and diagnostic inversion, providing a consistent and differentiable dynamic basis for S3's posterior filtering and trend recursion.

[0062] S3: Utilize the monotonic gated damage evolution state space model to perform posterior inversion of steel structure damage to generate component damage confidence intervals, and recursively complete the early warning judgment of steel structure damage trends.

[0063] The trained monotone-gated damage evolution state-space model is read, and recursive Bayesian estimation based on particle filtering is performed on the real-time input sequence of steel structure observation feature vectors. At each time step, the unbounded latent variable vector of the component evolution from the previous time step is advanced using the state transition operator, and the prior component damage state set is obtained through component damage state boundary mapping. Then, the corresponding predicted observation feature vector is calculated using the observation operator, and the innovation quantity is calculated with the measured steel structure observation feature vector. The particle weights are calculated based on the innovation quantity, and the component damage state set is updated to obtain the posterior set. The quantile intervals are calculated for the posterior set according to the steel structure component damage confidence coverage level, and the component damage confidence intervals are output and stored in conjunction with the component number.

[0064] Furthermore, the particle filter uses a set of component damage states with a fixed number of 500 particles to represent the posterior distribution. The initial particles are generated by adding a zero-mean Gaussian perturbation to the initial component damage state vector output by S2, with the perturbation standard deviation fixed at 0.02. Each particle contains the component damage states of all components and is bound to the corresponding component number.

[0065] The innovation quantity is defined as the difference vector between the measured and predicted characteristic vectors of the steel structure. Particle weights are updated based on the Gaussian likelihood of the innovation quantity. The standard deviation of the likelihood noise is fixed at 0.10 and derived from the root mean square statistics of the difference vectors between the measured and predicted values ​​on the healthy baseline window. After weight updates, normalization is performed to ensure the sum of all particle weights is 1. The Gaussian likelihood is calculated based on the assumption of independent and homoscedasticity in each dimension. The sum of the squared differences in each dimension of the innovation quantity difference vector is converted to the particle likelihood value using the likelihood noise standard deviation of 0.10, and this likelihood value is used to update the particle weights through multiplication.

[0066] When the number of effective particles is less than 50% of the total number of particles, system resampling is triggered. The resampling process uses system resampling, and the weights are reset to a uniform distribution after resampling. The damage confidence coverage level α for steel structure components is fixed at 0.95. For each component, the posterior particle quantiles at the 2.5% and 97.5% quantiles are used as the upper and lower bounds of the damage confidence interval for that component, and the intervals are stored together with the component number and window number. The number of effective particles is calculated as the reciprocal of the sum of the squares of the particle weights; that is, the squares of each particle weight are summed, and the reciprocal is taken as the number of effective particles, used to characterize the degree of weight degradation.

[0067] Using the posterior set from which the component damage confidence interval is generated as the initial set, the state transition operator is invoked to perform forward recursion according to the preset steel structure damage trend recursion step size and number of recursion steps, obtaining the component damage state prediction set for the future time domain. For each component, the proportion of component damage states in the statistical prediction set that exceed the steel structure component damage severity judgment threshold is taken as the over-limit probability, and the average increment of the component damage state at adjacent recursion times is calculated as the damage evolution rate. When the over-limit probability is not less than the steel structure damage early warning confidence judgment threshold or the damage evolution rate is not less than the steel structure damage acceleration judgment threshold, the corresponding component's early warning identifier and trigger time are output.

[0068] Furthermore, the recursive step size for the steel structure damage trend is set to 5.12 seconds, consistent with the S1 window step size, and the number of recursive steps for the steel structure damage trend is fixed at 60 steps to cover the prediction time domain of approximately 307 seconds in the future. The forward recursion calls the state transition operator to advance for each posterior particle, obtaining the corresponding set of component damage state predictions while keeping the particle weights unchanged.

[0069] The threshold for judging the severity of damage to steel structure components is fixed at 0.70. The rule is to classify the component damage state from 0 to 1 into engineering grades: 0.00~0.30 (minor), 0.30~0.70 (developing), and 0.70~1.00 (severe), and select the lower limit of the severe grade as the threshold. The probability of exceeding the limit is the proportion of particles with a component damage state greater than 0.70 in the prediction set to the total number of particles. The damage evolution rate is the average of the component damage state increment of the same particle in adjacent recursive steps divided by 5.12 seconds and then applied to all particles.

[0070] Furthermore, the confidence threshold for steel structure damage early warning is fixed at 0.80, derived from adding a 0.05 margin to the upper bound of the exceedance probability obtained by forward recursion of the healthy baseline window and setting a lower limit of 0.80 to suppress occasional noise triggering. The acceleration threshold for steel structure damage is fixed at 0.002 per second, derived from adding three times the standard deviation to the mean damage evolution rate obtained from the healthy baseline window statistics and setting a lower limit not less than 0.002. When any component meets the condition that the exceedance probability is not less than 0.80 or the damage evolution rate is not less than 0.002 per second, the early warning identifier of that component, the recursion step number that first meets the condition, and the corresponding prediction timestamp are output.

[0071] It should be noted that this step uses particle filtering to perform recursive Bayesian estimation on the monotonic gated state-space model, obtaining a component-level posterior set and outputting confidence intervals using quantiles. Then, the posterior set is forward recursively used to form a prediction set. Early warning is triggered based on both the probability of exceeding limits and the evolution rate. The threshold is statistically fixed from the healthy baseline, ensuring that real-time inference and trend determination can be repeatedly executed under the same dynamic kernel.

[0072] Example 2, an embodiment of the present invention, provides a steel structure damage evolution modeling and diagnosis system, including a multi-source temporal synchronous feature extraction module, a boundary mapping training evolution module, and a posterior inversion recursive early warning judgment module.

[0073] The multi-source time-series synchronous feature extraction module is used for time-series synchronous noise reduction of multi-source monitoring of steel structures to generate steel structure observation feature vectors.

[0074] The boundary mapping training evolution module is used to generate component damage state boundary mappings based on steel structure observation feature vectors and to train a monotonically gated damage evolution state space model.

[0075] The posterior inversion recursive early warning judgment module is used to perform posterior inversion of steel structure damage using a monotonic gated damage evolution state space model to generate component damage confidence intervals and recursively complete the early warning judgment of steel structure damage trend.

Claims

1. A method for modeling and diagnosing damage evolution in steel structures, characterized in that, include: Denoising processing is performed on the time-series synchronous monitoring of steel structures from multiple sources to generate steel structure observation feature vectors; Based on the observation feature vectors of steel structures, a component damage state boundary mapping is generated, and a monotonic gated damage evolution state space model is trained. A monotonically gated damage evolution state space model is used to perform posterior inversion of steel structure damage to generate component damage confidence intervals, and the steel structure damage trend early warning judgment is recursively completed.

2. The method for modeling and diagnosing steel structure damage evolution as described in claim 1, characterized in that: The time-series synchronous noise reduction processing for multi-source monitoring of steel structures includes... Receive the raw time-series data from the steel structure monitoring channels and write it into a unified time buffer, then perform resampling and alignment according to the timestamp of each channel; The cross-correlation alignment offset is calculated and the time delay of each channel is corrected based on the selected steel structure reference channel; Perform outlier removal and missing segment interpolation on the aligned data; Bandpass filtering is performed on each channel and wavelet threshold denoising is superimposed to output the multi-source monitoring time series dataset after synchronous denoising.

3. The method for modeling and diagnosing steel structure damage evolution as described in claim 2, characterized in that: The generated steel structure observation feature vector includes, The multi-source monitoring time-series dataset after synchronous denoising is truncated into a window sequence by sliding window. For each window, the power spectral density is calculated and the frequency band energy component is obtained by integrating the spectral energy according to the preset steel structure mode frequency band set. The root mean square, kurtosis, and peak factor components are obtained by performing time-domain statistics calculations on the same window. Modal frequency and damping ratio components are obtained by performing modal parameter identification based on the response correlation matrix on the same window; The components are spliced ​​together in channel order and normalized to obtain the steel structure observation feature vector.

4. The method for modeling and diagnosing steel structure damage evolution as described in claim 3, characterized in that: The generation of component damage state boundary mapping based on steel structure observation feature vectors includes... For each component of the steel structure, establish unbounded component evolution latent variables and form a latent variable vector; Each unbounded component evolution latent variable is mapped to the component damage state via Sigmoid, so that the component damage state is limited to the interval between 0 and 1 and forms a component damage state vector. Generate component hierarchical codes based on steel structure component parameters and store them in association with the steel structure observation feature vectors; The evolutionary latent variables of unbounded components are initialized with the steel structure observation feature vector corresponding to the healthy baseline window, and the initialization results are used as the initial state input.

5. The method for modeling and diagnosing steel structure damage evolution as described in claim 4, characterized in that: The trained monotonically gated damage evolution state-space model includes... A state transition operator and an observation operator are constructed with the component damage state vector as the state. The state transition operator calculates the non-negative evolution increment for the unbounded latent variables of component evolution for each component and multiplies it with the component evolution gating factor. The component evolution gating factor is calculated by the steel structure observation feature vector and the component hierarchical encoding through a gating network, and takes a value from 0 to 1. When the component evolution gating factor is lower than the steel structure component evolution gating trigger threshold, the evolution increment of the corresponding component is set to zero; The observation operator maps the component damage state vector and the component hierarchical encoding to a predicted observation feature vector; During training, the difference between the measured steel structure observation feature vector and the predicted observation feature vector is used as the observation feature residual. The gating sparsity constraint of summing the absolute values ​​of the component evolution gating factor and the state change constraint of summing the absolute values ​​of the differences between the component damage state vectors at adjacent time points are superimposed. The parameters of the state transition operator and the observation operator are trained simultaneously through gradient updates.

6. The method for modeling and diagnosing steel structure damage evolution as described in claim 5, characterized in that: The method of using a monotonically gated damage evolution state-space model to perform posterior inversion of steel structure damage to generate component damage confidence intervals includes: Read the trained monotonic gated damage evolution state space model and perform recursive Bayesian estimation based on particle filtering on the real-time input steel structure observation feature vector sequence; At each time step, the state transition operator is used to advance the component damage state vector of the previous time step to obtain the a priori component damage state set. Then, the observation operator is used to calculate the corresponding predicted observation feature vector and the innovation quantity is calculated together with the measured steel structure observation feature vector. The posterior set is obtained by calculating particle weights based on the innovation quantity and updating the component damage state set. The quantile intervals are calculated for the posterior set based on the confidence coverage level of damage to steel structure components. The component damage confidence intervals are then output and stored in conjunction with the component numbers.

7. The method for modeling and diagnosing steel structure damage evolution as described in claim 6, characterized in that: The recursive completion of the steel structure damage trend early warning determination includes: Using the posterior set corresponding to the component damage confidence interval as the initial set, the state transition operator is called to perform forward recursion according to the preset steel structure damage trend recursion step size and recursion step number to obtain the component damage state prediction set in the future time domain. The proportion of component damage states exceeding the threshold for judging the severity of steel structure component damage in the statistical prediction set for each component is taken as the over-limit probability, and the average increment of component damage states at adjacent recursive time points is taken as the damage evolution rate. When the probability of exceeding the limit is not less than the confidence threshold for steel structure damage warning or the damage evolution rate is not less than the acceleration threshold for steel structure damage, the warning identifier and trigger time of the corresponding component are output.

8. A steel structure damage evolution modeling and diagnosis system, employing the steel structure damage evolution modeling and diagnosis method as described in any one of claims 1 to 7, characterized in that: It includes a multi-source temporal synchronous feature extraction module, a boundary mapping training and evolution module, and a posterior inversion recursive early warning judgment module; The multi-source time-series synchronous feature extraction module is used to perform time-series synchronous noise reduction processing on multi-source monitoring of steel structures to generate steel structure observation feature vectors. The boundary mapping training and evolution module is used to generate component damage state boundary mappings based on steel structure observation feature vectors and to train a monotonically gated damage evolution state space model. The posterior inversion recursive early warning judgment module is used to perform posterior inversion of steel structure damage using a monotonic gated damage evolution state space model to generate component damage confidence intervals and recursively complete the early warning judgment of steel structure damage trend.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the steel structure damage evolution modeling and diagnosis method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the steel structure damage evolution modeling and diagnosis method as described in any one of claims 1 to 7.

Images