Intelligent evaluation method and system for structural health state
By combining autoencoder neural networks and Wasserstein distance with hierarchical Bayesian models, the problem of dependence on high-dimensional training data and label data in existing technologies is solved, enabling efficient and reliable continuous assessment of structural health status, which is suitable for long-term monitoring of various engineering structures.
Patent Information
- Application Number
- CN202610503965.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-16
- Publication Date
- 2026-07-14
AI Technical Summary
Existing methods for monitoring structural health status rely on high-dimensional training and label data, resulting in low efficiency and difficulty in continuously evaluating the evolution of structural health status.
By employing an autoencoder neural network combined with the Gaussian cumulative distribution function and Wasserstein distance, structural response data is acquired through a sensor network to construct benchmark and evaluation datasets. Reliability is calculated using a hierarchical Bayesian model to achieve continuous assessment of structural health status.
It enables efficient and reliable continuous assessment of structural health status, reduces data processing burden, and improves assessment accuracy and efficiency, making it suitable for long-term monitoring of various engineering structures.
Smart Images
Figure CN122388412A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of civil engineering, and specifically relates to an intelligent assessment method and system for structural health status. Background Technology
[0002] Structural health monitoring is crucial for structural safety and stability. In the field of structural health monitoring, long-term operational safety assessments typically rely on sensor networks (such as strain gauges and accelerometers) to continuously collect data on structural responses, and then use data processing and feature extraction to monitor or assess the structural health status.
[0003] Currently, commonly used structural health monitoring schemes often employ data-driven approaches, such as machine learning or deep learning. However, these schemes typically use structural health data as the training set, which requires massive amounts of high-dimensional training data. The storage and processing of high-dimensional training data are inefficient. Furthermore, these schemes generally require labeled data for training, which is difficult to obtain. Finally, these schemes cannot continuously assess the evolution of structural health status. Summary of the Invention
[0004] One of the objectives of this invention is to provide an intelligent structural health status assessment method that is capable of continuous assessment of structural health status and is highly reliable and efficient.
[0005] The second objective of this invention is to provide a system for implementing the aforementioned intelligent assessment method for structural health status.
[0006] The intelligent structural health status assessment method provided by this invention includes the following steps:
[0007] S1. Set the baseline state of the target structure;
[0008] S2. Acquire structural response data of various types of the target structure under the baseline state and the state to be evaluated through a sensor network;
[0009] S3. Based on the Gaussian cumulative distribution function, obtain the mean and standard deviation information corresponding to the structural response data information obtained in step S2, so as to compress the structural response data information and construct the benchmark dataset and the evaluation dataset.
[0010] S4. Construct an autoencoder neural network for the target structure based on linear layers and activation function layers;
[0011] S5. Using the benchmark dataset obtained in step S3, train the autoencoder neural network constructed in step S3 to obtain the benchmark state autoencoder neural network, and record the corresponding data information.
[0012] S6. Input the evaluation dataset obtained in step S3 into the baseline state autoencoder neural network obtained in step S4, and record the corresponding data information;
[0013] S7. Based on the data recorded in steps S5 and S6, construct the potential features of the baseline state and the state to be evaluated, calculate the corresponding Wasserstein distance, and construct the corresponding similarity matrix.
[0014] S8. Based on the data recorded in steps S5 and S6, the final reliability of the state to be evaluated is calculated using the hierarchical Bayesian model.
[0015] S9. Use the similarity matrix obtained in step S7 as a quantitative description of the target structure in the state to be evaluated, and use the final reliability obtained in step S8 as a qualitative description of the target structure in the state to be evaluated, thus completing the health status assessment of the target structure in the state to be evaluated.
[0016] Step S1, which sets the baseline state of the target structure, specifically includes the following steps:
[0017] Set the baseline state of the target structure;
[0018] The aforementioned baseline state includes the structural undamaged state, the set operating state, or the state of a set year.
[0019] Step S2, which involves acquiring structural response data of various types of the target structure under both the baseline state and the state to be evaluated through a sensor network, specifically includes the following steps:
[0020] Using a sensor network, we can acquire structural response data for various types of the target structure under both the baseline state and the state to be evaluated.
[0021] The types mentioned include traffic loads and seismic loads;
[0022] The structural response data information includes strain data, acceleration data, and displacement data.
[0023] Step S3, based on the Gaussian cumulative distribution function, obtains the mean and standard deviation information corresponding to the structural response data information acquired in step S2, so as to compress the structural response data information and construct the benchmark dataset and the evaluation dataset. Specifically, this includes the following steps:
[0024] Based on the maximum and minimum values of the obtained structural response data, set the corresponding threshold information;
[0025] For each type of stimulus, calculate the cumulative loading time corresponding to the set threshold;
[0026] Structural response data can be reconstructed using a Gaussian cumulative distribution function. Therefore, the mean value of the corresponding type of excitation is obtained by fitting the data using the following formula. and standard deviation :
[0027] In the formula It is the Gaussian cumulative distribution function; This represents the cumulative loading time. The Gaussian error function; The set threshold;
[0028] The structural response data information is used to obtain the mean value. and standard deviation Compress the representation;
[0029] Construct the corresponding benchmark dataset and evaluation dataset.
[0030] Step S4, which describes constructing the autoencoder neural network based on linear layers and activation function layers to build the target structure, specifically includes the following steps:
[0031] The constructed autoencoder neural network includes an encoder and a decoder;
[0032] The encoder comprises a first coding linear layer, a first coding activation function layer, a second coding linear layer, a second coding activation function layer, a third coding linear layer, and a third coding activation function layer connected in series.
[0033] The encoder's processing procedure is represented as follows:
[0034] In the formula The latent vector output by the encoder; This is the processing function for the third encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the third linear coding layer; This is the bias vector of the third linear coding layer; This is the processing function for the second encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the second linear coding layer; This is the bias vector of the second linear coding layer; This is the processing function of the first encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the first linear coding layer; This is the bias vector of the first linear coding layer;
[0035] The decoder comprises a first decoding linear layer, a first decoding activation function layer, a second decoding linear layer, a second decoding activation function layer, a third decoding linear layer, and a third decoding activation function layer connected in series.
[0036] The decoder's processing procedure is represented as follows:
[0037] In the formula This is the output of the decoder; This is the processing function of the third decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the third decoding linear layer; This is the bias vector for the third decoding linear layer; This is the processing function for the second decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the second decoding linear layer; This is the bias vector for the second decoding linear layer; This is the processing function of the first decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the first decoding linear layer; This is the bias vector for the first decoding linear layer.
[0038] Step S5 involves using the benchmark dataset obtained in step S3 to train the autoencoder neural network constructed in step S3, obtaining a benchmark state autoencoder neural network, and recording the corresponding data information. Specifically, this includes the following steps:
[0039] During training, the following formula is used as the loss function:
[0040] In the formula The loss function value of the baseline state autoencoder neural network; Loss is driven by mechanical characteristics; For data loss; This represents the total number of sensors; The representation value of the i-th sensor is obtained by reconstructing it through a baseline state autoencoder neural network; Let be the strain ratio, and , Let be the strain value of the i-th sensor under the k-th loading, and m be the number of loading cycles; This represents the actual data during the k-th load. This is the output of the decoder of the autoencoder neural network under the k-th loading.
[0041] Step S6, which involves inputting the evaluation dataset obtained in step S3 into the baseline state autoencoder neural network obtained in step S4 and recording the corresponding data information, specifically includes the following steps:
[0042] The evaluation dataset obtained in step S3 is input into the baseline state autoencoder neural network obtained in step S4, and the following loss function value is calculated. :
[0043] In the formula This represents the mechanical characteristic-driven loss value under the k-th loading. This represents the data loss value during the k-th load. This represents the loss due to the maximum mean difference. This represents the feature representation of the i-th sample in the baseline operating condition; Let be the feature representation of the j-th sample in the current loading condition; n is the number of feature samples in the baseline condition; m is the number of feature samples in the current loading condition. This is the set positive definite kernel function.
[0044] Step S7, which involves constructing latent features of the baseline state and the state to be evaluated based on the data recorded in steps S5 and S6, calculating the corresponding Wasserstein distance, and constructing the corresponding similarity matrix, specifically includes the following steps:
[0045] Constructing latent features of the baseline state for ,in Let be the latent representation of the i-th sample in the feature space under the baseline state. The number of samples in the baseline state;
[0046] Constructing latent features of the state to be evaluated for ,in Let be the latent representation of the i-th sample in the feature space under the state to be evaluated. The number of samples for the state to be evaluated;
[0047] The latent characteristics of the baseline state are calculated using the following formula. Represented as an empirical probability distribution:
[0048] In the formula The empirical probability distribution of the potential features of the baseline state; for Functions used to convert Represented as a point of unit mass;
[0049] The latent characteristics of the state to be evaluated are calculated using the following formula. Represented as an empirical probability distribution:
[0050] In the formula The empirical probability distribution of the potential characteristics of the state to be evaluated; for Functions used to convert Represented as a point of unit mass;
[0051] The second-order Wasserstein distance is calculated using the following formula to quantify the difference between the baseline state and the state to be evaluated:
[0052] In the formula for and The second-order Wasserstein distance; It is a joint probability distribution; For all those with marginal distributions and The joint distribution set; From Sample points extracted from the sample; From Sample points extracted from the sample; The square of the Euclidean distance; This means finding the joint distribution that minimizes the average transportation cost among all possible joint distributions;
[0053] For the data information monitored by the state to be evaluated, calculate the second-order Wasserstein distance between the latent features of any two states to be evaluated. for ,in Let be the empirical probability distribution corresponding to the latent features of state i. Let be the empirical probability distribution corresponding to the latent features of state j;
[0054] Construct a similarity matrix ,in elements in Represented as ,in The set parameter is used to adjust the mapping scale from distance to similarity.
[0055] Step S8, which calculates the final reliability of the state to be evaluated based on the data recorded in steps S5 and S6 and using a hierarchical Bayesian model, specifically includes the following steps:
[0056] Using the latent features of the baseline state and the corresponding loss function, the baseline state is represented as a baseline state feature vector. , The corresponding mean vector Represented as The corresponding covariance matrix Represented as Where I is the identity matrix; Represents the mathematical expectation operator; Describes the covariance operator. For the specified small regularization term;
[0057] The distance between the sample and the distribution is calculated using the following formula:
[0058] In the formula Let i be the distance between the sample to be evaluated and the baseline state distribution; Let i be the feature vector of the i-th sample to be evaluated;
[0059] The following formula will be used to calculate Standardize:
[0060] In the formula For standardization ; express the median; express Standard deviation;
[0061] Will Mapped to damage probability , represented as ;in, The cumulative distribution function of the standard normal distribution;
[0062] Calculate the safety probability for ;
[0063] Construct a prior distribution based on the baseline state. , represented as ;in, Represents the Beta distribution. Shape parameters of the Beta distribution , Shape parameters of the Beta distribution ;
[0064] For each state, there exists: and ;in, Let y be the expected health indicator for state y. Let y be the expected damage index for state y;
[0065] Post-hoc update Represented as ;in, This represents the shape parameter corresponding to "health" in the prior distribution. The shape parameter corresponding to the "damage" in the prior distribution;
[0066] Adopt update For prior distribution Perform a recursive update;
[0067] The reliability of the state to be evaluated is calculated. for ;in, For the set encoding weights, The output of the encoder is obtained by inputting the evaluation dataset into the baseline state autoencoder neural network; The weights are set for the loss function;
[0068] Reliability of the state to be evaluated Normalization is performed to obtain the final reliability of the state to be evaluated. for ;in For the reliability of the updated prior distribution, and , The set baseline coding weights, The weights of the set baseline loss function, The output of the encoder is obtained by inputting the benchmark dataset into the benchmark state autoencoder neural network.
[0069] Step S9, which uses the similarity matrix obtained in step S7 as a quantitative description of the target structure in the state to be evaluated and the final reliability obtained in step S8 as a qualitative description of the target structure in the state to be evaluated, specifically includes the following steps:
[0070] The similarity matrix obtained in step S7 is used as a quantitative description of the target structure in the state to be evaluated:
[0071] The similarity matrix S is a pairwise similarity matrix between states; where the elements are... This represents the similarity between the i-th state and the j-th state; The closer the value is to 1, the closer the i-th state is to the j-th state. The smaller the value, the further apart the i-th state is from the j-th state; the first state in the similarity matrix S is the reference state; the first row of the similarity matrix S represents the similarity between the reference state and all other states;
[0072] The final reliability obtained in step S8 is used as a qualitative description of the target structure in the state to be evaluated:
[0073] like This indicates that the target structure is in an undegraded state relative to the baseline state;
[0074] like This indicates that the target structure is in a state to be evaluated, and that it has degraded relative to the baseline state.
[0075] This invention also provides a system for implementing the intelligent assessment method for structural health status, comprising a state setting module, a data acquisition module, a data construction module, a network construction module, a benchmark training module, an evaluation training module, a matrix construction module, a reliable calculation module, and a health assessment module; the state setting module, data acquisition module, data construction module, network construction module, benchmark training module, evaluation training module, matrix construction module, reliable calculation module, and health assessment module are connected in series; the state setting module is used to set the benchmark state of the target structure and upload the data information to the data acquisition module; the data acquisition module is used to acquire various types of structural response data information of the target structure under the benchmark state and the state to be evaluated through a sensor network based on the received data information, and upload the data information to the data construction module; the data construction module is used to obtain the mean and standard deviation information corresponding to the acquired structural response data information based on the Gaussian cumulative distribution function, to compress the structural response data information, and to construct a benchmark dataset and an evaluation dataset, and upload the data information to the network construction module; the network construction module is used to construct an autoencoder neural network of the target structure based on linear layers and activation function layers based on the received data information, and upload the data information to the network construction module. The system comprises three modules: a baseline training module and a health assessment module. The baseline training module trains the constructed autoencoder neural network using the received data and the obtained baseline dataset, obtaining a baseline state autoencoder neural network and recording the corresponding data. The assessment training module inputs the obtained assessment dataset into the obtained baseline state autoencoder neural network, records the corresponding data, and uploads the data to the matrix construction module. The matrix construction module constructs latent features of the baseline state and the state to be assessed based on the received and recorded data, calculates the corresponding Wasserstein distance, constructs the corresponding similarity matrix, and uploads the data to the reliability calculation module. The reliability calculation module calculates the final reliability of the state to be assessed based on the received and recorded data using a hierarchical Bayesian model and uploads the data to the health assessment module. The health assessment module uses the obtained similarity matrix as a quantitative description of the target structure in the state to be assessed and the obtained final reliability as a qualitative description of the target structure in the state to be assessed, thus completing the health assessment of the target structure in the state to be assessed.
[0076] The intelligent structural health status assessment method and system provided by this invention, through the setting of a baseline state, the acquisition of load responses in the baseline state and the state to be assessed, the construction of an autoencoder neural network and the collection of training process data, combined with a similarity matrix constructed based on Wasserstein distance and the final reliability of the state to be assessed obtained based on a hierarchical Bayesian model, can not only achieve continuous intelligent health status assessment of the target structure relative to the set baseline state, but also has higher reliability, better accuracy and higher assessment efficiency. Attached Figure Description
[0077] Figure 1 This is a schematic diagram of the method flow of the present invention.
[0078] Figure 2 This is a schematic diagram of the data compression process in the method of the present invention.
[0079] Figure 3 This is a schematic diagram of the structure of the autoencoder neural network in the method of the present invention.
[0080] Figure 4 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation
[0081] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The intelligent assessment method for structural health status disclosed in this invention includes the following steps:
[0082] S1. Set the baseline state of the target structure; specifically including the following steps:
[0083] The baseline state of the target structure is set; if the target is a bridge, the present invention can be used to assess the structural health of the bridge.
[0084] The aforementioned baseline state includes the undamaged state of the structure, the set operating state, or the state of a set year; after the baseline state is set, the evaluation results of the state to be evaluated subsequently obtained are all evaluation results of the state to be evaluated relative to the set baseline state.
[0085] S2. Acquire structural response data of various types of the target structure under both the baseline state and the state to be evaluated using a sensor network; specifically including the following steps:
[0086] Using a sensor network, we can acquire structural response data for various types of the target structure under both the baseline state and the state to be evaluated.
[0087] The types mentioned include traffic loads and seismic loads;
[0088] The structural response data information includes strain data information, acceleration data information, and displacement data information;
[0089] S3. Based on the Gaussian cumulative distribution function, obtain the mean and standard deviation information corresponding to the structural response data information obtained in step S2, so as to compress the structural response data information and construct the benchmark dataset and the evaluation dataset; specifically including the following steps:
[0090] This step characterizes the structural response by measuring the cumulative time it takes for the statistical structural response data to exceed a predefined threshold, transforming the original time-series data into compact and noise-robust features. These features are closely related to the evolution of structural damage, such as... Figure 2 As shown;
[0091] Based on the maximum and minimum values of the obtained structural response data, set the corresponding threshold information;
[0092] For each type of stimulus, calculate the cumulative loading time corresponding to the set threshold; for example... Figure 2 As shown in (a); if four thresholds (Level 1 to Level 4) are defined, then the cumulative loading times corresponding to these thresholds are respectively ;
[0093] The resulting cumulative time histogram provides a highly compressed yet representative description of the actual strain response, such as... Figure 2 As shown in (b), the structural response data can be reconstructed using the Gaussian cumulative distribution function. Therefore, the mean value of the corresponding type of excitation is obtained by fitting the data using the following formula. and standard deviation :
[0094] In the formula It is the Gaussian cumulative distribution function; This represents the cumulative loading time. The Gaussian error function; The set threshold;
[0095] Adding more predefined strain thresholds yields a more detailed cumulative time histogram, allowing the fitted CDF to capture richer and more accurate structural response dynamics. Subsequently, the CDF is converted into a probability density function (PDF) to characterize the actual strain response of each sensor; the probability density function is expressed as... ;
[0096] By employing this data compression technique, a large amount of measured strain data is first segmented, then processed using a sliding window strategy, and finally representative features are extracted, such as... Figure 2 As shown in (c). Through the mean and standard deviation This compressed data is used to characterize the data and thus describe the measurement data from each sensor. This method effectively reduces the amount of data while retaining the critical information needed for structural health monitoring.
[0097] The structural response data information is used to obtain the mean value. and standard deviation Compress the representation;
[0098] Construct the corresponding benchmark dataset and evaluation dataset;
[0099] S4. Construct an autoencoder neural network for the target structure based on linear layers and activation function layers; specifically including the following steps:
[0100] The network constructed in this invention integrates data-driven learning and prior mechanical knowledge, thereby achieving robust and interpretable reconstruction of sensor data, and identifying changes in structural health status through a uniquely designed autoencoder framework.
[0101] The constructed autoencoder neural network includes an encoder and a decoder;
[0102] The encoder comprises a first linear coding layer, a first coding activation function layer, a second linear coding layer, a second coding activation function layer, a third linear coding layer, and a third coding activation function layer connected in series; the encoder is used to compress input data into a low-dimensional and information-rich feature space;
[0103] The encoder's processing procedure is represented as follows:
[0104] In the formula The latent vector output by the encoder; This is the processing function for the third encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the third linear coding layer; This is the bias vector of the third linear coding layer; This is the processing function for the second encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the second linear coding layer; This is the bias vector of the second linear coding layer; This is the processing function of the first encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the first linear coding layer; This is the bias vector of the first linear coding layer;
[0105] The decoder comprises a first decoding linear layer, a first decoding activation function layer, a second decoding linear layer, a second decoding activation function layer, a third decoding linear layer, and a third decoding activation function layer connected in series; the decoder is used to attempt to reconstruct the original high-dimensional strain time series.
[0106] The decoder's processing procedure is represented as follows:
[0107] In the formula This is the output of the decoder; This is the processing function of the third decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the third decoding linear layer; This is the bias vector for the third decoding linear layer; This is the processing function for the second decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the second decoding linear layer; This is the bias vector for the second decoding linear layer; This is the processing function of the first decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the first decoding linear layer; This is the bias vector of the first decoding linear layer;
[0108] The symmetric structure of the network constructed in this invention helps to efficiently learn and compress the feature space. The vector z output by the network serves as a unique state fingerprint to characterize the structural state evolution process.
[0109] S5. Using the benchmark dataset obtained in step S3, train the autoencoder neural network constructed in step S3 to obtain the benchmark state autoencoder neural network, and record the corresponding data information; specifically including the following steps:
[0110] During training, the following formula is used as the loss function:
[0111] In the formula The loss function value of the baseline state autoencoder neural network; Loss is driven by mechanical characteristics; For data loss; This represents the total number of sensors; The representation value of the i-th sensor is obtained by reconstructing it through a baseline state autoencoder neural network; Let be the strain ratio, and , Let be the strain value of the i-th sensor under the k-th loading, and m be the number of loading cycles; This represents the actual data during the k-th load. This is the output of the decoder of the autoencoder neural network under the k-th loading.
[0112] In the measured strain data, the average strain typically reflects the overall deformation behavior and overall stiffness state of the structure, which is consistent with system-level reliability analysis; the maximum strain, on the other hand, characterizes the extreme structural response and is a damage-sensitive feature that can reflect potential stiffness degradation, support performance deterioration, and problems such as cracks or loose connections. Therefore, the network model proposed in this invention uses a sensor-level weight matrix... To consider the correlation of maximum strain changes among sensors at different locations in the baseline model; this weight matrix It reflects the distribution of mechanical characteristics among the sensors; as the structural state evolves, these state-related correlation patterns will change; by analyzing the changes in the strain response of the multi-sensor system, this change can be identified, enabling the model to learn and identify behaviors that deviate from the baseline structural state more effectively.
[0113] Mechanical feature-driven loss is used to penalize inconsistencies between compressed data from different sensors, thereby enhancing the overall coordination of deformation. This loss embeds mechanical correlation by emphasizing sensor pairs with strong or consistent responses. Under structural integrity, the maximum strain response has high spatial consistency, so this loss is small. As the structure degrades, this consistency is gradually disrupted, and the strain distribution between different sensors becomes uneven, leading to an increase in mechanical feature-driven loss.
[0114] S6. Input the evaluation dataset obtained in step S3 into the baseline state autoencoder neural network obtained in step S4, and record the corresponding data information; specifically including the following steps:
[0115] The evaluation dataset obtained in step S3 is input into the baseline state autoencoder neural network obtained in step S4, and the following loss function value is calculated. :
[0116] In the formula This represents the mechanical characteristic-driven loss value under the k-th loading. This represents the data loss value during the k-th load. This represents the loss due to the maximum mean difference. This represents the feature representation of the i-th sample in the baseline operating condition; Let be the feature representation of the j-th sample in the current loading condition; n is the number of feature samples in the baseline condition; m is the number of feature samples in the current loading condition. The set positive definite kernel function;
[0117] Maximum Mean Discrepancy (MMD) is a statistical measure used to quantify the difference between two probability distributions based on a sample. It is achieved by comparing the distance between the mean embeddings of the two distributions in a regenerating kernel Hilbert space (RKHS). Therefore, MMD can be used to assess the distinguishability of two sets of data under a given kernel mapping. A smaller MMD value indicates that the two sets of data come from similar distributions; a larger MMD value indicates that the two sets of data differ significantly in their underlying distributions. In structural health monitoring, MMD can be used to measure the difference in characteristic distributions between different structural states (e.g., undamaged and damaged states). By quantifying this difference, MMD can serve as a damage-sensitive indicator, capable of identifying structural anomalies and assessing the domain adaptation performance between different structural systems.
[0118] This stage introduces MMD loss to explicitly amplify the difference in probability distributions between the baseline state and other structural states. Since MMD measures the difference between two distributions based on sample representations, a small MMD loss indicates that the current state is statistically similar to the baseline state, and thus its loss distribution is close to the reference model. However, when the structural state changes (e.g., damage occurs), the statistical properties of latent features and reconstruction errors will deviate from the baseline state. In this case, the MMD loss will increase, thus... and Further increase the total reference loss on the basis Therefore, compared to the baseline model, the reconstruction loss distribution will show a significant shift, thereby enhancing the sensitivity to minor structural changes and enabling early damage identification.
[0119] S7. Based on the data recorded in steps S5 and S6, construct the latent features of the baseline state and the state to be evaluated, calculate the corresponding Wasserstein distance, and construct the corresponding similarity matrix; specifically including the following steps:
[0120] Constructing latent features of the baseline state for ,in Let be the latent representation of the i-th sample in the feature space under the baseline state. The number of samples in the baseline state;
[0121] Constructing latent features of the state to be evaluated for ,in Let be the latent representation of the i-th sample in the feature space under the state to be evaluated. The number of samples for the state to be evaluated;
[0122] The latent characteristics of the baseline state are calculated using the following formula. Represented as an empirical probability distribution:
[0123] In the formula The empirical probability distribution of the potential features of the baseline state; for Functions used to convert Represented as a point of unit mass;
[0124] The latent characteristics of the state to be evaluated are calculated using the following formula. Represented as an empirical probability distribution:
[0125] In the formula The empirical probability distribution of the potential characteristics of the state to be evaluated; for Functions used to convert Represented as a point of unit mass;
[0126] The second-order Wasserstein distance is calculated using the following formula to quantify the difference between the baseline state and the state to be evaluated:
[0127] In the formula for and The second-order Wasserstein distance; It is a joint probability distribution; For all those with marginal distributions and The joint distribution set; From Sample points extracted from the sample; From Sample points extracted from the sample; The square of the Euclidean distance; This indicates the search for the joint distribution with the minimum average transportation cost among all possible joint distributions; this index reflects the minimum transportation cost required to transform the current potential distribution into the baseline distribution, thus providing a physically meaningful measure of structural state deviation.
[0128] For the data information monitored by the state to be evaluated, calculate the second-order Wasserstein distance between the latent features of any two states to be evaluated. for ,in Let be the empirical probability distribution corresponding to the latent features of state i. Let be the empirical probability distribution corresponding to the latent features of state j;
[0129] Construct a similarity matrix ,in elements in Represented as ,in This is a set parameter used to adjust the mapping scale from distance to similarity; a larger similarity value indicates that the structural states are closer, while a smaller value indicates that the deviation from the baseline state is more obvious.
[0130] Similarity matrices based on Wasserstein distance can quantitatively characterize the long-term evolution of structural health status; in particular, Wasserstein distance can capture changes in the overall distribution and has good robustness to changes in data support, thus it is suitable for the analysis of latent features and reconstruction loss.
[0131] S8. Based on the data recorded in steps S5 and S6, and using a hierarchical Bayesian model, calculate the final reliability of the state to be evaluated; specifically, this includes the following steps:
[0132] The structural reliability is assessed using a hierarchical Bayesian model. This model combines latent features extracted from monitoring data with the statistical distribution of reconstruction loss to quantify the degree of structural deviation of the state to be evaluated relative to the baseline state.
[0133] Using the latent features of the baseline state and the corresponding loss function, the baseline state is represented as a baseline state feature vector. , The corresponding mean vector Represented as The corresponding covariance matrix Represented as Where I is the identity matrix; Represents the mathematical expectation operator; Describes the covariance operator. A small regularization term is set to avoid singularities in the covariance matrix;
[0134] The distance between the sample and the distribution is calculated using the following formula:
[0135] In the formula Let i be the distance between the sample to be evaluated and the baseline state distribution; Let be the feature vector of the i-th sample to be evaluated; this distance simultaneously considers the feature scale and the correlation between features, thus obtaining a dimensionless multivariate health indicator; when and When the direction is close to and the main change direction of the baseline distribution is consistent with the baseline distribution It will be small, when and Distant distribution (especially in directions with small variance) will be big;
[0136] The following formula will be used to calculate Standardize:
[0137] In the formula For standardization , used to quantify the degree of deviation of sample i from the overall distance distribution; express The median is used as a reference for the center position; express The standard deviation is used for normalization scaling.
[0138] Will Mapped to damage probability , represented as ;in, The cumulative distribution function of the standard normal distribution is used to map standardized outliers to probabilities.
[0139] Calculate the safety probability for ;
[0140] Construct a prior distribution based on the baseline state. , represented as ;in, This represents the Beta distribution (used to describe the prior distribution characteristics of probability or normalization indices). Shape parameters of the Beta distribution (The concentration that corresponds to the prior distribution of "healthy" or safe concentration under the baseline condition). Shape parameters of the Beta distribution (The concentration of "damage" in the prior distribution corresponding to the baseline state);
[0141] For each state, there exists: and ;in, This is the expected health indicator for state y (reflecting the overall health of the structure or system). The expected damage index for state y (reflecting the damage risk of the overall structure or system);
[0142] Post-hoc update Represented as ;in, This represents the shape parameter corresponding to "health" in the prior distribution. The shape parameter corresponding to the "damage" in the prior distribution;
[0143] Adopt update For prior distribution Perform a recursive update;
[0144] The reliability of the state to be evaluated is calculated. for ;in, For the set encoding weights, The output of the encoder is obtained by inputting the evaluation dataset into the baseline state autoencoder neural network; The weights are set for the loss function;
[0145] Reliability of the state to be evaluated Normalization is performed to obtain the final reliability of the state to be evaluated. for ;in For the reliability of the updated prior distribution, and , The set baseline coding weights, The weights of the set baseline loss function, The output of the encoder after the benchmark dataset is input into the benchmark state autoencoder neural network;
[0146] S9. Using the similarity matrix obtained in step S7 as a quantitative description of the target structure in the state to be evaluated, and using the final reliability obtained in step S8 as a qualitative description of the target structure in the state to be evaluated, the health status assessment of the target structure in the state to be evaluated is completed; specifically including the following steps:
[0147] The similarity matrix obtained in step S7 is used as a quantitative description of the target structure in the state to be evaluated:
[0148] The similarity matrix S is a pairwise similarity matrix between states; where the elements are... This represents the similarity between the i-th state and the j-th state; The closer the value is to 1, the closer the i-th state is to the j-th state. The smaller the value, the further apart the i-th state is from the j-th state; the first state in the similarity matrix S is the reference state; the first row of the similarity matrix S represents the similarity between the reference state and all other states;
[0149] The final reliability obtained in step S8 is used as a qualitative description of the target structure in the state to be evaluated:
[0150] like This indicates that the target structure is in an undegraded state relative to the baseline state;
[0151] like This indicates that the target structure is in a state to be evaluated, and that it has degraded relative to the baseline state.
[0152] The evaluation method proposed in this invention is applicable to various structural and engineering applications where sensors are installed, including bridge structures (such as highway bridges, railway bridges, cable-stayed bridges, suspension bridges, box girder bridges, etc.), track systems (such as railway tracks, high-speed railway tracks and sleeper / track slab systems), high-rise buildings and structures (such as steel structures, concrete structures, multi-story and high-rise buildings), underground engineering and tunnels (such as highway tunnels, subway tunnels, underground passages and support structures), and other civil engineering infrastructure (such as dams, revetments, retaining walls, etc.).
[0153] The method of this invention can be widely applied to scenarios such as structural health monitoring and early warning of damage. Through steps such as latent feature analysis, Wasserstein distance, and Bayesian update, it can achieve quantitative assessment and degradation determination of the target structure's state.
[0154] This invention employs a single baseline state (damage-free / reference state) for training. During the inference phase, the model is not trained further; instead, feature extraction and loss assessment are directly performed on subsequent states, achieving completely label-free long-term health assessment. Traditional methods typically rely on multi-state labeled data (supervised learning) or require continuous model updates (online learning), while this invention significantly reduces reliance on labeled data, supports long-term continuous monitoring, and is suitable for "damage-free labeling" scenarios in practical engineering. Based on the traditional autoencoder, a weight matrix based on the maximum strain correlation of multiple sensors is introduced. And a mechanical consistency constraint loss was constructed. Existing methods are solely data-driven and do not explicitly consider the spatial consistency of structural mechanics. In contrast, this invention improves model interpretability by introducing physical meaning (cooperative structural deformation) and enhances sensitivity to structural stiffness degradation. It introduces MMD loss during the inference phase (used only during inference) to compare the potential distribution differences between the current state and the baseline state. This technique amplifies minute structural changes, improves early damage identification capabilities, and expands from "point error" to "distribution difference." By constructing a Beta prior and Beta-Binomial updates, and recursively applying this between states (time evolution), it achieves long-term reliability evolution tracking, integrates uncertainty, and provides probabilistic health indicator assessment results.
[0155] This invention constructs an autoencoder model based on a baseline state and combines it with the maximum mean difference distribution distance metric method, enabling continuous monitoring and evolution assessment of structural health status without the need for damaged label data or multi-state training data. This overcomes the problem that existing technologies rely on a large amount of labeled data and are difficult to apply to long-term unsupervised monitoring in actual engineering.
[0156] This invention introduces a loss function driven by mechanical features and constructs a weight matrix based on the maximum strain information of multiple sensors to constrain the sensor cooperative deformation relationship during the model training process. This makes the model learning results conform to the evolution law of structural mechanics, thereby significantly improving the physical interpretability and recognition accuracy of structural damage identification results and overcoming the problems of lack of mechanical constraints and weak interpretability in existing data-driven methods.
[0157] This application employs a statistical feature compression method to transform high-frequency time-series monitoring data into low-dimensional statistical distribution parameters. This significantly reduces the scale of data processing and computational burden. Furthermore, by using a joint discrimination mechanism of reconstruction error and distribution difference, it improves the model's robustness to environmental noise and its sensitivity to detecting minor structural damage. This overcomes the shortcomings of traditional methods, such as insensitivity to early damage and susceptibility to noise interference.
[0158] like Figure 4The diagram shows the functional modules of the system of this invention: The system for implementing the intelligent assessment method for structural health status disclosed in this invention includes a state setting module, a data acquisition module, a data construction module, a network construction module, a benchmark training module, an evaluation training module, a matrix construction module, a reliable calculation module, and a health assessment module; these modules are connected in series. The state setting module sets the benchmark state of the target structure and uploads the data information to the data acquisition module. The data acquisition module, based on the received data information, acquires structural response data of various types of the target structure under the benchmark state and the state to be evaluated through a sensor network, and uploads the data information to the data construction module. The data construction module, based on the received data information and a Gaussian cumulative distribution function, obtains the mean and standard deviation information corresponding to the acquired structural response data information, compresses the structural response data information, constructs a benchmark dataset and an evaluation dataset, and uploads the data information to the network construction module. The network construction module, based on the received data information and a linear layer and activation function layer, constructs an autoencoder neural network for the target structure. The network transmits data to the benchmark training module. The benchmark training module trains the constructed autoencoder neural network using the received benchmark dataset to obtain a benchmark state autoencoder neural network, records the corresponding data, and uploads it to the evaluation training module. The evaluation training module inputs the obtained evaluation dataset into the obtained benchmark state autoencoder neural network, records the corresponding data, and uploads it to the matrix construction module. The matrix construction module constructs latent features of the benchmark state and the state to be evaluated based on the received and recorded data, calculates the corresponding Wasserstein distance, constructs the corresponding similarity matrix, and uploads the data to the reliability calculation module. The reliability calculation module calculates the final reliability of the state to be evaluated based on the received and recorded data using a hierarchical Bayesian model and uploads the data to the health assessment module. The health assessment module uses the obtained similarity matrix as a quantitative description of the target structure in the state to be evaluated and the obtained final reliability as a qualitative description of the target structure in the state to be evaluated, thus completing the health assessment of the target structure in the state to be evaluated.
Claims
1. A method for intelligent assessment of structural health status, comprising the following steps: S1. Set the baseline state of the target structure; S2. Acquire structural response data of various types of the target structure under the baseline state and the state to be evaluated through a sensor network; S3. Based on the Gaussian cumulative distribution function, obtain the mean and standard deviation information corresponding to the structural response data information obtained in step S2, so as to compress the structural response data information and construct the benchmark dataset and the evaluation dataset. S4. Construct an autoencoder neural network for the target structure based on linear layers and activation function layers; S5. Using the benchmark dataset obtained in step S3, train the autoencoder neural network constructed in step S3 to obtain the benchmark state autoencoder neural network, and record the corresponding data information. S6. Input the evaluation dataset obtained in step S3 into the baseline state autoencoder neural network obtained in step S4, and record the corresponding data information; S7. Based on the data recorded in steps S5 and S6, construct the potential features of the baseline state and the state to be evaluated, calculate the corresponding Wasserstein distance, and construct the corresponding similarity matrix. S8. Based on the data recorded in steps S5 and S6, the final reliability of the state to be evaluated is calculated using the hierarchical Bayesian model. S9. Use the similarity matrix obtained in step S7 as a quantitative description of the target structure in the state to be evaluated, and use the final reliability obtained in step S8 as a qualitative description of the target structure in the state to be evaluated, thus completing the health status assessment of the target structure in the state to be evaluated.
2. The intelligent structural health status assessment method according to claim 1, characterized in that... Step S1, which sets the baseline state of the target structure, specifically includes the following steps: Set the baseline state of the target structure; The aforementioned baseline state includes the structural undamaged state, the set operating state, or the state of a set year; Step S2, which involves acquiring structural response data of various types of the target structure under both the baseline state and the state to be evaluated through a sensor network, specifically includes the following steps: Using a sensor network, we can acquire structural response data for various types of the target structure under both the baseline state and the state to be evaluated. The types mentioned include traffic loads and seismic loads; The structural response data information includes strain data, acceleration data, and displacement data.
3. The intelligent structural health status assessment method according to claim 2, characterized in that... Step S3, based on the Gaussian cumulative distribution function, obtains the mean and standard deviation information corresponding to the structural response data information acquired in step S2, so as to compress the structural response data information and construct the benchmark dataset and the evaluation dataset. Specifically, this includes the following steps: Based on the maximum and minimum values of the obtained structural response data, set the corresponding threshold information; For each type of stimulus, calculate the cumulative loading time corresponding to the set threshold; Structural response data can be reconstructed using a Gaussian cumulative distribution function. Therefore, the mean value of the corresponding type of excitation is obtained by fitting the data using the following formula. and standard deviation : In the formula It is the Gaussian cumulative distribution function; This represents the cumulative loading time. The Gaussian error function; The set threshold; The structural response data information is used to obtain the mean value. and standard deviation Compress the representation; Construct the corresponding benchmark dataset and evaluation dataset.
4. The intelligent structural health status assessment method according to claim 3, characterized in that... Step S4, which describes constructing the autoencoder neural network based on linear layers and activation function layers to build the target structure, specifically includes the following steps: The constructed autoencoder neural network includes an encoder and a decoder; The encoder comprises a first coding linear layer, a first coding activation function layer, a second coding linear layer, a second coding activation function layer, a third coding linear layer, and a third coding activation function layer connected in series. The encoder's processing procedure is represented as follows: In the formula The latent vector output by the encoder; This is the processing function for the third encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the third linear coding layer; This is the bias vector of the third linear coding layer; This is the processing function for the second encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the second linear coding layer; This is the bias vector of the second linear coding layer; This is the processing function of the first encoding activation function layer, and Use the GELU activation function; This is the weight matrix of the first linear coding layer; This is the bias vector of the first linear coding layer; The decoder comprises a first decoding linear layer, a first decoding activation function layer, a second decoding linear layer, a second decoding activation function layer, a third decoding linear layer, and a third decoding activation function layer connected in series. The decoder's processing procedure is represented as follows: In the formula This is the output of the decoder; This is the processing function of the third decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the third decoding linear layer; This is the bias vector for the third decoding linear layer; This is the processing function for the second decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the second decoding linear layer; This is the bias vector for the second decoding linear layer; This is the processing function of the first decoding activation function layer, and Use the GELU activation function; This is the weight matrix of the first decoding linear layer; This is the bias vector for the first decoding linear layer.
5. The intelligent structural health status assessment method according to claim 4, characterized in that... Step S5 involves using the benchmark dataset obtained in step S3 to train the autoencoder neural network constructed in step S3, obtaining a benchmark state autoencoder neural network, and recording the corresponding data information. Specifically, this includes the following steps: During training, the following formula is used as the loss function: In the formula The loss function value of the baseline state autoencoder neural network; Loss is driven by mechanical characteristics; For data loss; This represents the total number of sensors; The representation value of the i-th sensor is obtained by reconstructing it through a baseline state autoencoder neural network; Let be the strain ratio, and , Let be the strain value of the i-th sensor under the k-th loading, and m be the number of loading cycles; This represents the actual data during the k-th load. This is the output of the decoder of the autoencoder neural network under the k-th loading.
6. The intelligent structural health status assessment method according to claim 5, characterized in that... Step S6, which involves inputting the evaluation dataset obtained in step S3 into the baseline state autoencoder neural network obtained in step S4 and recording the corresponding data information, specifically includes the following steps: The evaluation dataset obtained in step S3 is input into the baseline state autoencoder neural network obtained in step S4, and the following loss function value is calculated. : In the formula This represents the mechanical characteristic-driven loss value under the k-th loading. This represents the data loss value during the k-th load. This represents the loss due to the maximum mean difference. This represents the feature representation of the i-th sample in the baseline operating condition; Let be the feature representation of the j-th sample in the current loading condition; n is the number of feature samples in the baseline condition; m is the number of feature samples in the current loading condition. This is the set positive definite kernel function.
7. The intelligent structural health status assessment method according to claim 6, characterized in that... Step S7, which involves constructing latent features of the baseline state and the state to be evaluated based on the data recorded in steps S5 and S6, calculating the corresponding Wasserstein distance, and constructing the corresponding similarity matrix, specifically includes the following steps: Constructing latent features of the baseline state for ,in Let be the latent representation of the i-th sample in the feature space under the baseline state. The number of samples in the baseline state; Constructing latent features of the state to be evaluated for ,in Let be the latent representation of the i-th sample in the feature space under the state to be evaluated. The number of samples for the state to be evaluated; The latent characteristics of the baseline state are calculated using the following formula. Represented as an empirical probability distribution: In the formula The empirical probability distribution of the potential features of the baseline state; for Functions used to convert Represented as a point of unit mass; The latent characteristics of the state to be evaluated are calculated using the following formula. Represented as an empirical probability distribution: In the formula The empirical probability distribution of the potential characteristics of the state to be evaluated; for Functions used to convert Represented as a point of unit mass; The second-order Wasserstein distance is calculated using the following formula to quantify the difference between the baseline state and the state to be evaluated: In the formula for and The second-order Wasserstein distance; It is a joint probability distribution; For all those with marginal distributions and The joint distribution set; From Sample points extracted from the sample; From Sample points extracted from the sample; The square of the Euclidean distance; This means finding the joint distribution that minimizes the average transportation cost among all possible joint distributions; For the data information monitored by the state to be evaluated, calculate the second-order Wasserstein distance between the latent features of any two states to be evaluated. for ,in Let be the empirical probability distribution corresponding to the latent features of state i. Let be the empirical probability distribution corresponding to the latent features of state j; Construct a similarity matrix ,in elements in Represented as ,in The set parameter is used to adjust the mapping scale from distance to similarity.
8. The intelligent structural health status assessment method according to claim 7, characterized in that... Step S8, which calculates the final reliability of the state to be evaluated based on the data recorded in steps S5 and S6 and using a hierarchical Bayesian model, specifically includes the following steps: Using the latent features of the baseline state and the corresponding loss function, the baseline state is represented as a baseline state feature vector. , The corresponding mean vector Represented as The corresponding covariance matrix Represented as Where I is the identity matrix; Represents the mathematical expectation operator; Describes the covariance operator. For the specified small regularization term; The distance between the sample and the distribution is calculated using the following formula: In the formula Let i be the distance between the sample to be evaluated and the baseline state distribution; Let i be the feature vector of the i-th sample to be evaluated; The following formula will be used to calculate Standardize: In the formula For standardization ; express the median; express Standard deviation; Will Mapped to damage probability , represented as ;in, The cumulative distribution function of the standard normal distribution; Calculate the safety probability for ; Construct a prior distribution based on the baseline state. , represented as ;in, Represents the Beta distribution. Shape parameters of the Beta distribution , Shape parameters of the Beta distribution ; For each state, there exists: and ;in, Let y be the expected health indicator for state y. Let y be the expected damage index for state y; Post-hoc update Represented as ;in, Let be the shape parameter corresponding to "health" in the prior distribution. The shape parameter corresponding to "damage" in the prior distribution; Adopt update For prior distribution Perform a recursive update; The reliability of the state to be evaluated is calculated. for ;in, For the set encoding weights, The output of the encoder is obtained by inputting the evaluation dataset into the baseline state autoencoder neural network; The weights are set for the loss function; Reliability of the state to be evaluated Normalization is performed to obtain the final reliability of the state to be evaluated. for ;in For the reliability of the updated prior distribution, and , The set baseline coding weights, The weights of the set baseline loss function, The output of the encoder is obtained by inputting the benchmark dataset into the benchmark state autoencoder neural network.
9. The intelligent structural health status assessment method according to claim 8, characterized in that... Step S9, which uses the similarity matrix obtained in step S7 as a quantitative description of the target structure in the state to be evaluated and the final reliability obtained in step S8 as a qualitative description of the target structure in the state to be evaluated, specifically includes the following steps: The similarity matrix obtained in step S7 is used as a quantitative description of the target structure in the state to be evaluated: The similarity matrix S is a pairwise similarity matrix between states; where the elements are... This represents the similarity between the i-th state and the j-th state; The closer the value is to 1, the closer the i-th state is to the j-th state. The smaller the value, the further apart the i-th state is from the j-th state; the first state in the similarity matrix S is the reference state; the first row of the similarity matrix S represents the similarity between the reference state and all other states; The final reliability obtained in step S8 is used as a qualitative description of the target structure in the state to be evaluated: like This indicates that the target structure is in an undegraded state relative to the baseline state; like This indicates that the target structure is in a state to be evaluated, and that it has degraded relative to the baseline state.
10. A system for implementing the intelligent structural health status assessment method according to any one of claims 1 to 9, characterized in that... It includes a state setting module, a data acquisition module, a data construction module, a network construction module, a benchmark training module, an evaluation training module, a matrix construction module, a reliable calculation module, and a health assessment module; these modules are connected in series. The state setting module sets the benchmark state of the target structure and uploads the data to the data acquisition module. The data acquisition module, based on the received data, uses a sensor network to acquire structural response data of various types of the target structure under the benchmark state and the state to be evaluated, and uploads the data to the data construction module. The data construction module is used to obtain the mean and standard deviation of the acquired structural response data based on the received data information and the Gaussian cumulative distribution function, so as to compress the structural response data information, construct the benchmark dataset and the evaluation dataset, and upload the data information to the network construction module. The network construction module is used to construct an autoencoder neural network of the target structure based on the received data information, using linear layers and activation function layers, and upload the data information to the benchmark training module; The benchmark training module is used to train the constructed autoencoder neural network based on the received data information and the obtained benchmark dataset, to obtain the benchmark state autoencoder neural network, record the corresponding data information, and upload the data information to the evaluation training module. The evaluation training module is used to input the obtained evaluation dataset into the obtained baseline state autoencoder neural network based on the received data information, record the corresponding data information, and upload the data information to the matrix construction module. The matrix construction module is used to construct the latent features of the baseline state and the state to be evaluated based on the received data information and the recorded data information, calculate the corresponding Wasserstein distance, construct the corresponding similarity matrix, and upload the data information to the reliable calculation module. The reliability calculation module is used to calculate the final reliability of the state to be evaluated based on the received data and the recorded data, using a hierarchical Bayesian model, and then upload the data to the health assessment module. The health assessment module is used to evaluate the health status of the target structure in the state to be evaluated by using the obtained similarity matrix as a quantitative description of the target structure in the state to be evaluated and the obtained final reliability as a qualitative description of the target structure in the state to be evaluated.