Structural state evaluation method based on Bayesian optimization BiLSTM network
By adopting Bayesian method of optimizing BiLSTM network and Wasserstein distance in structural state evaluation, the problem of insufficient objectivity and poor generality of structural state evaluation in the prior art is solved, and a more efficient and accurate structural state evaluation is achieved. .
Patent Information
- Application Number
- CN202411446759.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2025-02-11
AI Technical Summary
The existing structural state evaluation methods have problems such as insufficient objectivity, low real-timeness and low intelligence. Especially in complex civil engineering structures, traditional methods are difficult to effectively utilize massive but poor information-based structural static dynamic data, and require a large number of labels. The sample data is trained, resulting in poor universality.
Using a structural state evaluation method based on Bayesian optimization BiLSTM network, combining Wasserstein distance to measure the difference between the current state of the structure and the reference state, the features are directly extracted from the response data through unsupervised learning, and the neural network is determined using Bayesian optimization. The hyperparameter combination improves the accuracy of the evaluation results and the interpretability of the network.
It improves the accuracy and versatility of structural state evaluation results, can effectively utilize labelless data, reduces dependence on complex models, and enhances the real-time and intelligent evaluation.
Smart Images
Figure CN119415870A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of civil structure status assessment, in particular to a structural status assessment method based on a Bayesian optimization BiLSTM network. Background Art
[0002] Structural status assessment is an important reference for ensuring the safety of operation and maintenance of in-service civil engineering structures. Traditional assessment methods mainly rely on simple theories and expert experience to make decisions, which are not objective enough and have low real-time and intelligent levels. At the same time, actual civil engineering structures often have many components, large geometric dimensions, unclear boundaries and connection conditions, and limited data from on-site inspections, which all bring difficulties to the assessment process. In order to achieve real-time monitoring of structural status, structural health monitoring (SHM) systems have been applied to some important structures (such as large public buildings and long-span bridges) in recent years. The static and dynamic structural data collected by the system can be used to evaluate the structural status, but it also has the characteristics of "massive but poor information". At the same time, considering the monitoring cost, the number and coverage of measurement points are often insufficient, which also increases the difficulty of information acquisition.
[0003] The state assessment methods based on vibration signals mainly include model-based methods and data-driven methods. The former is based on the mathematical and physical models of the structure (such as finite element models) and obtains the changes in the physical parameters of the structure through the optimization solution of inverse problems. However, the establishment of the structural benchmark model is often very difficult; the latter directly establishes the mapping relationship between the physical parameters of the structure and the response signal without the need for a physical model, thus avoiding the process of establishing a complex model. As a representative of data-driven methods, neural networks (NNs) can be used to achieve nonlinear mapping between structural response and parameters, and therefore have been widely used. However, the network training process of NNs belongs to the black box operation mechanism, involving a large number of network parameters that need to be adjusted (such as weights, biases, learning rates, etc.), which need to be determined through experience or trial and error. At the same time, shallow NNs have limited learning ability for complex mapping relationships, and are prone to overfitting in the case of small sample sets, resulting in poor network generalization performance. In contrast, deep NNs can extract data features more effectively through layer-by-layer learning through the combination of multiple hidden layers. They are more efficient than shallow NNs in processing large amounts of data and have better application prospects.
[0004] In recent years, deep learning methods have made great progress and applications in the engineering field, such as Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs) and Recurrent Neural Networks (RNNs) in the field of SHM. CNNs are a type of feedforward deep NNs whose training process includes convolution calculations. They can effectively extract grid features, but the training cost is high and requires a large amount of labeled sample data, which is prone to overfitting; GANs are composed of a generator for obtaining sample distribution and a discriminator for classification. Network training is achieved by introducing a competition mechanism, but there are problems of training instability and mode collapse; RNNs are a type of NNs that can process sequence data. When processing each input data, a hidden state is retained and passed to the next time step as memory information. Therefore, it can be used to extract features related to the structural state from acceleration time series data. In long time series problems, RNNs may have a gradient vanishing problem, which affects the accuracy of the structural evaluation results. To this end, the Long Short-Term Memory (LSTM) network introduces a gating mechanism based on RNNs, which can effectively solve the problem of gradient disappearance or explosion. The hidden layer of the bidirectional LSTM (Bidirectional Long Short-Term Memory, BiLSTM) is composed of a forward LSTM layer and a backward LSTM layer, which can more effectively extract feature information related to the state of the structural system from the structural response signal.
[0005] At present, deep NNs also have limitations in the application of structural status assessment and require adaptive adjustments. For example, network hyperparameters determined through experience or trial and error will cause the trained network to be only applicable to specific target structures and have poor versatility. For this reason, some optimization algorithms such as grid search, random search and genetic algorithms are used to obtain hyperparameter combinations, but they cannot effectively utilize the prior information of parameters, resulting in a waste of computing resources. Secondly, the network training process of supervised learning requires a large amount of labeled sample data, and the test samples and training samples must come from the same sample population. For this reason, it is often necessary to obtain the response data of the structure in different states through finite element models to construct a test sample set for network training. However, the establishment of a structural benchmark model is often very difficult. In contrast, the network training mechanism based on unsupervised learning does not rely on the label of the data, and can directly extract features from the structural response data, which is more suitable for the field of structural status assessment. Summary of the invention
[0006] The present invention proposes a structural state assessment method based on a Bayesian optimized BiLSTM network, which, combined with the Wasserstein distance, can more effectively measure the difference between the current state of the structure and the reference state (non-destructive working condition).
[0007] The present invention adopts the following technical solutions.
[0008] The invention discloses a structural state assessment method based on a Bayesian optimization BiLSTM network. The method aims at different target structures and combines the BiLSTM network to establish a correlation model between response data at different positions of the target structure. The method adopts an unsupervised network learning mechanism and uses Bayesian optimization (BO) to confirm the hyperparameter combination during the network training process, thereby objectifying the hyperparameter determination process of neural networks (NNs). The method increases the network interpretability when establishing the correlation model between response data at different positions of the target structure. The method directly extracts features from the response data using the unsupervised learning network training mechanism to improve the accuracy of the structural state assessment result.
[0009] The method is used as follows:
[0010] Firstly, sensors arranged at different positions of the target structure are selected so that their response data can jointly form a sensor cluster. Reference sensors and adjacent sensors in the sensor cluster are selected, and the response data in the intact state of the structure are standardized and processed with sliding windows.
[0011] Secondly, the preprocessed response data is input into the BiLSTM network for training, combined with BO optimization to determine the hyperparameter combination of the network;
[0012] Finally, the structural response data in the undamaged state is used as the baseline state, and the response data in the current state of the structure are respectively input into the trained BO-BiLSTM network, and the MSE value between the corresponding predicted data and the actual data is calculated. The difference between the MSE values of the current state and the baseline state is quantified by the Wasserstein distance, and the state index is constructed to evaluate the current state of the structure.
[0013] The method comprises the following steps:
[0014] Step S1, selection of sensor cluster and data preprocessing;
[0015] Step S2, constructing a BO-BiLSTM network;
[0016] Step S3: Evaluate the structural status.
[0017] The target structure includes a frame structure or a beam structure.
[0018] In step S1, the method of selecting the sensor cluster is to first select the key position of the target structure to arrange the sensor as the reference sensor, and then arrange the adjacent sensors on the left and right of the reference sensor. When the acquisition time length is T, the response data of the reference sensor collected is Z = {z 1 ,z 2 ,...,z T}; The response data collected from the adjacent sensors is The reference sensor and neighboring sensors together form a sensor cluster.
[0019] In step S1, the method of data preprocessing is: standardizing the response data of the reference sensor and the neighboring sensor respectively, so as to avoid the instability problem caused by amplitude change during network training while retaining the characteristics of the response data; then, performing sliding window processing on the standardized data to adapt to the input requirements of the BiLSTM network.
[0020] The response data in step S1 are response data under the condition of structural integrity.
[0021] The specific method of step S2 is as follows: input the response data after standardization and sliding window processing under the structural non-destructive working condition into the BiLSTM network for training, and establish the association model between the sensor clusters; the input of the network is X, and the output is Z; the dropout layer is added in the training process to avoid overfitting, and the fully connected (FC) layer is introduced to reduce the dimension of the data; according to the situation of the response data, the optimization interval of the hyperparameter is set, and the BO is combined for optimization to determine the hyperparameter combination of the network; in the optimization process, the mean square error (MSE) between the network prediction result and the actual situation is selected as the evaluation function.
[0022] The hyperparameters include initial learning rate, regularization coefficient, and batch size.
[0023] The specific method of step S3 is as follows: input the X data of the current state of the structure and the lossless state into the trained BO-BiLSTM network respectively, and output the prediction of Z; take the Z data of the structure in the lossless state as the reference state, calculate the MSE value between the predicted data and the actual data respectively, and measure the difference between the current sample and the training sample of the lossless working condition; finally, quantify the difference between the MSE value of the current state and the reference state through the Wasserstein distance, so as to construct the state index η, so as to evaluate the current state of the structure.
[0024] The calculation formula of the state index η is:
[0025]
[0026] In the formula, ρh, ρ u Represent the probability distribution of the MSE value of the structure in the intact state and the current state; Π(ρ h ,ρ u ) represents the set of all possible joint distributions;
[0027] In step S3, samples are obtained from the possible joint distribution γ (mse x ,mse y ), ‖mse x -mse y Represents the distance between the two samples, and then we can get the expected value of the distance between the two samples under the corresponding γ inf means taking the lower bound of all possible expected values as the value of the state indicator η.
[0028] The present invention proposes a structural state assessment method based on Bayesian optimization BiLSTM network, which has the following advantages: (1) Bayesian optimization can efficiently confirm the hyperparameter combination of BiLSTM network for different training samples, increase the interpretability of the network, and establish the correlation model between the response data at different positions, thereby improving the accuracy of the structural state assessment results; (2) The network training mechanism of unsupervised learning does not require labeled data, and can directly extract features from the response data, which is more suitable for the field of structural state assessment. (3) Compared with traditional similarity measurement methods (such as KL divergence, JS divergence, etc.), Wasserstein distance can more effectively measure the difference between the current state of the structure and the baseline state. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0030] Attached Figure 1 It is a flow chart of the structural state evaluation method based on the Bayesian optimization BiLSTM network of the present invention. DETAILED DESCRIPTION
[0031] As shown in the figure, a structural state assessment method based on a Bayesian optimized BiLSTM network is disclosed. The method targets different target structures and combines a BiLSTM network to establish a correlation model between response data at different positions of the target structure. An unsupervised network learning mechanism is used. Bayesian optimization (BO) is used to confirm the hyperparameter combination during network training, thereby objectifying the hyperparameter determination process of neural networks (NNs). The network interpretability is increased when establishing a correlation model between response data at different positions of the target structure. The method directly extracts features from the response data using an unsupervised learning network training mechanism to improve the accuracy of the structural state assessment results.
[0032] The method is used as follows:
[0033] Firstly, sensors arranged at different positions of the target structure are selected so that their response data can jointly form a sensor cluster. Reference sensors and adjacent sensors in the sensor cluster are selected, and the response data in the intact state of the structure are standardized and processed with sliding windows.
[0034] Secondly, the preprocessed response data is input into the BiLSTM network for training, combined with BO optimization to determine the hyperparameter combination of the network;
[0035] Finally, the structural response data in the undamaged state is used as the baseline state, and the response data in the current state of the structure are respectively input into the trained BO-BiLSTM network, and the MSE value between the corresponding predicted data and the actual data is calculated. The difference between the MSE values of the current state and the baseline state is quantified by the Wasserstein distance, and the state index is constructed to evaluate the current state of the structure.
[0036] The method comprises the following steps:
[0037] Step S1, selection of sensor cluster and data preprocessing;
[0038] Step S2, constructing a BO-BiLSTM network;
[0039] Step S3: Evaluate the structural status.
[0040] The target structure includes a frame structure or a beam structure.
[0041] In step S1, the method of selecting the sensor cluster is to first select the key position of the target structure to arrange the sensor as the reference sensor, and then arrange the adjacent sensors on the left and right of the reference sensor. When the acquisition time length is T, the response data of the reference sensor collected is Z = {z 1 ,z 2 ,...,zT}; The response data collected from the adjacent sensors is The reference sensor and neighboring sensors together form a sensor cluster.
[0042] In step S1, the method of data preprocessing is: standardizing the response data of the reference sensor and the neighboring sensor respectively, so as to avoid the instability problem caused by amplitude change during network training while retaining the characteristics of the response data; then, performing sliding window processing on the standardized data to adapt to the input requirements of the BiLSTM network.
[0043] The response data in step S1 are response data under the condition of structural integrity.
[0044] The specific method of step S2 is as follows: input the response data after standardization and sliding window processing under the structural non-destructive working condition into the BiLSTM network for training, and establish the association model between the sensor clusters; the input of the network is X, and the output is Z; the dropout layer is added in the training process to avoid overfitting, and the fully connected (FC) layer is introduced to reduce the dimension of the data; according to the situation of the response data, the optimization interval of the hyperparameter is set, and the BO is combined for optimization to determine the hyperparameter combination of the network; in the optimization process, the mean square error (MSE) between the network prediction result and the actual situation is selected as the evaluation function.
[0045] The hyperparameters include initial learning rate, regularization coefficient, and batch size.
[0046] The specific method of step S3 is as follows: input the X data of the current state of the structure and the lossless state into the trained BO-BiLSTM network respectively, and output the prediction of Z; take the Z data of the structure in the lossless state as the reference state, calculate the MSE value between the predicted data and the actual data respectively, and measure the difference between the current sample and the training sample of the lossless working condition; finally, quantify the difference between the MSE value of the current state and the reference state through the Wasserstein distance, so as to construct the state index η, so as to evaluate the current state of the structure.
[0047] The calculation formula of the state index η is:
[0048]
[0049] In the formula, ρ h , u Represent the probability distribution of the MSE value of the structure in the intact state and the current state; Π(ρ h ,ρ u ) represents the set of all possible joint distributions;
[0050] In step S3, samples are obtained from the possible joint distribution γ (mse x ,mse y ), ‖mse x -mse y Represents the distance between the two samples, and then we can get the expected value of the distance between the two samples under the corresponding γ inf means taking the lower bound of all possible expected values as the value of the state indicator η.
[0051] Example:
[0052] This example proposes a structural state assessment method based on the Bayesian optimization BiLSTM network. The proposed method targets different target structures, such as frame structures or beam structures, and combines the BiLSTM network to establish a correlation model between the response data at different positions of the target structure. It also adopts an unsupervised network learning mechanism and uses Bayesian optimization (BO) to confirm the hyperparameter combination during the network training process, thereby objectifying the hyperparameter determination process of neural networks (NNs).
[0053] The technical solution used in this example includes three modules: sensor cluster selection and data preprocessing, BO-BiLSTM network construction and structural status assessment. The specific process is as follows Figure 1 As shown, the following steps are included:
[0054] Step 1: Selection of sensor cluster and data preprocessing.
[0055] First, the key position of the target structure is selected to arrange the sensor as the reference sensor. When the acquisition time is T, its response data is Z = {z 1 ,z 2 ,...,z T}. The neighboring sensors are arranged on the left and right of the reference sensor, and the response data collected is The reference sensor and the neighboring sensors together form a sensor cluster. The response data are standardized separately to avoid the instability caused by amplitude changes during network training while retaining the characteristics of the response data. Subsequently, the standardized data is processed by sliding window to meet the input requirements of the BiLSTM network.
[0056] Step 2: Construction of BO-BiLSTM network.
[0057] The response data after normalization and sliding window processing under the condition of structural non-destructive working is input into the BiLSTM network for training to establish the association model between sensor clusters. The input of the network is X and the output is Z. The dropout layer is added to the training process to avoid overfitting, and the fully connected (FC) layer is introduced to reduce the dimension of the data. According to the response data, the optimization interval of the hyperparameters (such as initial learning rate, regularization coefficient, batch size, etc.) is set, and the BO is combined for optimization to determine the hyperparameter combination of the network. In the optimization process, the mean squared error (MSE) between the network prediction result and the actual situation is selected as the evaluation function.
[0058] Step 3: Structural condition assessment.
[0059] The X data of the current state of the structure and the lossless state are respectively input into the trained BO-BiLSTM network, and the prediction of Z is output. The Z data of the structure in the lossless state is used as the reference state, and the MSE values between the predicted data and the actual data are calculated to measure the difference between the current sample and the training sample (lossless working condition). Finally, the difference between the MSE values of the current state and the reference state is quantified by the Wasserstein distance, so as to construct the state index η to evaluate the current state of the structure.
[0060]
[0061] In the formula, ρh, ρ u Represent the probability distribution of the MSE value of the structure in the intact state and the current state; Π(ρ h ,ρ u ) represents the set of all possible joint distributions; samples are obtained by sampling from possible joint distributions γ (mse x ,mse y ), ‖mse x -mse y ‖ represents the distance between the two samples, and then we can get the expected value of the distance between the two samples under the corresponding γ. inf means taking the lower bound of all possible expected values as the value of the state indicator η.
[0062] The usage process of this example is as follows: First, sensors arranged at different positions of the target structure are selected, and their response data together constitute a sensor cluster. Reference sensors and adjacent sensors in the sensor cluster are selected, and the response data of the structure in the intact state are standardized and processed with sliding windows; secondly, the preprocessed response data is input into the BiLSTM network for training, and the BO optimization is combined to determine the hyperparameter combination of the network; finally, the structural response data in the intact state is used as the baseline state, and the response data in the current state of the structure are respectively input into the trained BO-BiLSTM network, and the MSE value between the corresponding predicted data and the actual data is calculated. Then, the difference between the MSE values of the current state and the baseline state is quantified through the Wasserstein distance, and a state indicator is constructed to evaluate the current state of the structure.
Claims
1. A structural state evaluation method based on Bayesian optimization BiLSTM network, characterized by: The method aims at different target structures, combines the BiLSTM network to establish the association model between the response data at different positions of the target structure, and adopts an unsupervised network learning mechanism. In the network training process, Bayesian optimization is used to confirm the hyperparameter combination, thereby objectifying the hyperparameter determination process of the neural network NNs, and increasing the network interpretability when establishing the association model between the response data at different positions of the target structure. The method directly extracts features from the response data with the unsupervised learning network training mechanism to improve the accuracy of the structural state assessment results.
2. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 1 is characterized in that: The method is used as follows: Firstly, sensors arranged at different positions of the target structure are selected so that their response data can jointly form a sensor cluster. Reference sensors and adjacent sensors in the sensor cluster are selected, and the response data in the intact state of the structure are standardized and processed with sliding windows. Secondly, the preprocessed response data is input into the BiLSTM network for training, combined with BO optimization to determine the hyperparameter combination of the network; Finally, the structural response data in the undamaged state is used as the baseline state, and the response data in the current state of the structure are respectively input into the trained BO-BiLSTM network, and the MSE value between the corresponding predicted data and the actual data is calculated. The difference between the MSE values of the current state and the baseline state is quantified by the Wasserstein distance, and the state index is constructed to evaluate the current state of the structure. The method The steps include: Step S1, selection of sensor cluster and data preprocessing; Step S2, constructing a BO-BiLSTM network; Step S3: Evaluate the structural status.
3. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 2 is characterized in that: The target structure includes a frame structure or a beam structure.
4. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 2 is characterized in that: In step S1, the method of selecting the sensor cluster is to first select the key position of the target structure to arrange the sensor as the reference sensor, and then arrange the adjacent sensors on the left and right of the reference sensor. When the acquisition time length is T, the response data of the reference sensor collected is Z = {z1, z2, ..., z T }; The response data of the adjacent sensors collected is The reference sensor and neighboring sensors together form a sensor cluster.
5. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 4 is characterized in that: In step S1, the method of data preprocessing is: standardizing the response data of the reference sensor and the neighboring sensor respectively, so as to avoid the instability problem caused by amplitude change during network training while retaining the characteristics of the response data; then, performing sliding window processing on the standardized data to adapt to the input requirements of the BiLSTM network.
6. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 5 is characterized in that: The response data in step S1 are response data under the condition of structural integrity.
7. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 6 is characterized in that: The specific method of step S2 is as follows: input the response data after standardization and sliding window processing under the structural non-destructive working condition into the BiLSTM network for training, and establish the association model between the sensor clusters; the input of the network is X, and the output is Z; the dropout layer is added in the training process to avoid overfitting, and the fully connected layer is introduced to reduce the dimension of the data; according to the situation of the response data, the optimization interval of the hyperparameter is set, and the BO is combined for optimization to determine the hyperparameter combination of the network; in the optimization process, the mean square error between the network prediction result and the actual situation is selected as the evaluation function.
8. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 7 is characterized in that: The hyperparameters include initial learning rate, regularization coefficient, and batch size.
9. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 6 is characterized in that: The specific method of step S3 is: input the X data of the current state of the structure and the lossless state into the trained BO-BiLSTM network respectively, and output the prediction of Z; Taking the Z data of the structure in the damage-free state as the reference state, the MSE values between the predicted data and the actual data are calculated respectively to measure the difference between the current sample and the training sample of the damage-free condition; finally, the difference between the MSE values of the current state and the reference state is quantified by the Wasserstein distance, so as to construct the state index η to evaluate the current state of the structure.
10. The structural state assessment method based on Bayesian optimization BiLSTM network according to claim 9 is characterized in that: The calculation formula of the state index η is: Formula 1; In the formula, ρ h , u Represent the probability distribution of the MSE value of the structure in the intact state and the current state; Π(ρ h ,ρ u ) represents the set of all possible joint distributions; In step S3, samples are obtained from the possible joint distribution γ (mse x ,mse y ),||mse x -mse y || represents the distance between the two samples, and then we can get the expected value of the distance between the two samples under the corresponding γ. inf means taking the lower bound of all possible expected values as the value of the state indicator η.