A method for reconstructing structural vibration response based on a regenerative constraint supervised deep learning model

Abstract

The application provides a structure vibration response reconstruction method based on a regenerative constraint supervised deep learning model, first, collecting structure vibration response data and preprocessing; using the preprocessed structure vibration response data to generate training samples; then using the samples to train the regenerative constraint supervised deep learning model, in the model training process, using the loss function Loss of the regenerative constraint to guide the training, Loss includes target loss Loss obj and regenerative constraint Loss regen , the regenerative constraint Loss regen is composed of a priori term Loss prior and a stability term Loss stable , the target loss Loss obj is used to punish the error between the model output and the label, the a priori term Loss prior is used to represent the self attribute of the structure vibration response, and the stability term Loss stable is used to prevent the model from falling into a local optimal solution, thereby improving the response reconstruction accuracy of the supervised deep learning model.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring, and in particular to a method for reconstructing structural vibration response based on a regenerative constraint supervised deep learning model. Background Technology

[0002] Extreme environmental loads, including earthquakes and floods, as well as aging fixtures, can cause varying degrees of damage to engineering structures during use, thereby endangering public safety. Structural health monitoring systems are widely used in the field of civil engineering. By collecting vibration responses of structures and analyzing structural performance or damage characteristics, they provide management and maintenance decisions for structures, and are an important means of ensuring the safety and reliability of buildings.

[0003] The accuracy and completeness of structural vibration response directly impact the decision-making of structural health monitoring systems. Reliable damage identification and condition assessment technologies rely on complete and undamaged monitoring datasets. However, in practical engineering, due to budget constraints or unmeasurable locations, the sensor locations installed on the monitored structure are carefully selected, and their number is far less than the total degrees of freedom of the structure. Furthermore, due to factors such as sensor failure, transmission failure, and component aging, the vibration response measured by the sensors will inevitably be lost or damaged, which seriously affects the performance of structural health monitoring. In addition, a data loss rate of 0.5% to 2.5% has an impact on power spectral density comparable to 10% observation noise. Replacing sensors requires significant human and financial resources, and embedded sensors are often impossible to replace. Therefore, how to reconstruct lost structural vibration responses with high accuracy is an important research direction.

[0004] Supervised deep learning methods are widely used in structural vibration response reconstruction due to their deep representation learning capabilities. Guided by labels, these methods can learn complex nonlinear mappings between inputs and labels, typically achieving expected results under various missing conditions. Among them, the Joint Convolutional-Transformer (JCT) model and Deep Fully Convolutional Neural Network (DFCN) are classic examples. However, existing methods primarily improve reconstruction accuracy by refining the model architecture or directly transferring advanced methods from computer vision. Although there may be similarities between other signals and structural vibration response signals, directly transferring and applying supervised deep learning models inevitably ignores the inherent properties of the structural vibration response, thus reducing reconstruction accuracy. Furthermore, due to the diversity of structures, the inherent properties of structural vibration responses cannot be represented qualitatively or quantitatively. Summary of the Invention

[0005] To address the issue that existing supervised deep learning methods neglect the inherent properties of structural vibration response, leading to reduced reconstruction accuracy, this invention proposes a structural vibration response reconstruction method based on a regenerative constraint-supervised deep learning model. By constructing unified and reasonable artificial constraints, the method guides the model to learn the inherent properties of structural vibration response, thereby improving the interpretability of the model and ultimately enhancing reconstruction accuracy.

[0006] The technical solution of this invention is as follows:

[0007] A method for reconstructing structural vibration response based on a regenerative constraint-supervised deep learning model includes the following steps:

[0008] Step 1: Collect structural vibration response data using sensors deployed on the engineering structure and preprocess the data; generate training samples using the preprocessed structural vibration response data.

[0009] The structural vibration response data is complete data without any missing items;

[0010] The preprocessed structural vibration response data is sliced ​​to obtain Each segment of data, as Labels of each training sample;

[0011] For the Labels of each training sample To pass through the set mask matrix The corresponding training samples are calculated as follows:

[0012]

[0013] In the formula, Indicates the first term containing missing items One input sample; Indicates the first One sensor, Indicates the first in the sample One sampling point; The index set representing the location of missing items. This represents element-wise multiplication;

[0014] Step 2: Train the regenerative constraint-supervised deep learning model:

[0015] The regenerative constraint-supervised deep learning model can be any deep learning model that requires labels to participate in the training process.

[0016] Using the training samples and their labels obtained in step 1, the regenerative constraint-supervised deep learning model is trained. During the model training process, a loss function based on the regenerative constraint is used. Instructional training:

[0017]

[0018] in Loss to the target For regeneration constraints, It is the regularization coefficient set to limit the strength of the regeneration constraint;

[0019] The target loss is used to penalize the error between the model output and the label;

[0020] The regeneration constraint is determined by prior terms. and stable terms composition:

[0021]

[0022] in: The prior term is defined as:

[0023]

[0024] In the formula, Represents a matrix of all 1s; Indicates the randomly selected first... The mask matrix corresponding to each input sample With mask matrix The missing rates are the same;

[0025] The stable term is defined as:

[0026]

[0027] Step 3: Reconstruct the response using the trained regenerative constraint-supervised deep learning model:

[0028] The missing structural vibration response data actually collected by sensors on the structure is input into a trained regenerative constraint-supervised deep learning model to achieve complete reconstruction of the structural vibration response data.

[0029] Furthermore, the preprocessing process described in step 1 includes downsampling and normalization.

[0030] Furthermore, the target loss adopts the Frobenius norm:

[0031]

[0032] In the formula, This indicates the number of samples in the same batch. Indicates the first in the same batch One sample, This represents a regenerative constraint-supervised deep learning model; express The output of the model after inputting the regenerative constraint-supervised deep learning model.

[0033] Furthermore, in step 2, when training the regenerative constraint-supervised deep learning model, the network training optimizer uses the Adam optimizer, and the learning rate decay strategy uses a step-by-step decay strategy.

[0034] Furthermore, the present invention also proposes an electronic device and a readable storage medium:

[0035] An electronic device includes a processor and a memory, the memory being used to store one or more programs;

[0036] The above method is implemented when the one or more programs are executed by the processor.

[0037] A readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0038] Beneficial effects

[0039] The structural vibration response reconstruction method based on a regenerative constraint-supervised deep learning model proposed in this invention considers the inherent properties of the structural vibration response and solves the problem of low reconstruction accuracy of supervised deep learning models, such as... Figure 2 As shown, target loss Prior terms used to penalize the error between the model output and the label. Used to represent the inherent properties of structural vibration response, stability term This is used to prevent the model from getting stuck in local optima, thereby improving the response reconstruction accuracy of supervised deep learning models.

[0040] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0042] Figure 1 This is a schematic diagram of a structural vibration response reconstruction method based on a regenerative constraint supervised deep learning model according to the present invention;

[0043] Figure 2 This is an exploded view of the regeneration constraint of the present invention;

[0044] Figure 3 This is a schematic diagram of the Guangzhou Tower structural health monitoring system shown in Embodiment 1 of the present invention;

[0045] Figure 4 This is a comparison diagram of the structural vibration response reconstruction results of the JCT model under regenerative constraint shown in Embodiment 1 of the present invention;

[0046] Figure 5 This is a schematic diagram of the structural health monitoring system for the Hardanger Bridge in Norway, as shown in Embodiment 2 of the present invention;

[0047] Figure 6 This is a comparison diagram of the structural vibration response reconstruction results of the DFCN model under regenerative constraint as shown in Embodiment 2 of the present invention. Detailed Implementation

[0048] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0049] Example 1:

[0050] This embodiment uses the Canton Tower as the data collection object.

[0051] The Canton Tower is a 610-meter-tall skyscraper, such as... Figure 3 As shown in (a) of the image. Its main structure adopts a tube-in-tube design, comprising 46 steel ring beams, 24 steel-concrete composite columns, and reinforced concrete inner and outer tubes. The Guangzhou Tower's structural health monitoring system contains over 600 sensors, including 20 single-axis accelerometers arranged at specific heights along the major and minor axes, such as... Figure 3 As shown in (b) of the figure. The system, developed by a team from Hong Kong Polytechnic University, can accurately record the acceleration response under daily environmental loads, and the measurement data between 18:00 on January 19, 2010 and 18:00 on January 20, 2010 were made public.

[0052] In this embodiment, the structural vibration response reconstruction method based on a regenerative constraint-supervised deep learning model includes the following steps:

[0053] Step 1: Acceleration response sample data collection and preprocessing:

[0054] Data was collected using a complete single-axis accelerometer from the structural health monitoring system of the Guangzhou Tower. A fourth-order Butterworth low-pass filter with a cutoff frequency of 5Hz was used to filter out unavoidable high-frequency noise interference during actual measurement and transmission. The data was then downsampled at a sampling rate of 10Hz. Since some raw data may be too large or too small, it can cause gradient explosion or vanishing during backpropagation. Data normalization can effectively alleviate this problem and make the optimization algorithm converge faster during gradient descent. In this embodiment, data normalization scales the data to the range [-1, 1], defined as:

[0055]

[0056] in This represents the maximum absolute value in the sampled data. Indicates the first The first single-axis accelerometer The original data of each sampling point This represents the corresponding normalized data.

[0057] Then, the normalized output data from the complete single-axis accelerometer is sliced. The data slicing process divides the entire time-series signal into smaller samples, enabling the supervised deep learning model to effectively extract its local and global features. After slicing, the following results are obtained: The labels of each sample are determined, and then for each label, a mask matrix is ​​used. The corresponding samples were calculated:

[0058]

[0059] In the formula, It is the first one used for supervising deep learning models The labels of each input sample, i.e., the complete acceleration response data. Indicates the first term containing missing items One input sample; Indicates the number of single-axis accelerometers. This indicates the number of sampling points for each sample; The index set representing the location of missing items. This indicates element-wise multiplication.

[0060] This embodiment uses 20 single-axis accelerometers, with each sample including 1024 sampling points and a mask matrix. To simulate a 30% sensor failure scenario, the data from sensors 3, 6, 8, 13, 15, and 16 were randomly set to 0. All processed samples were then divided into training and validation sets in an 8:2 ratio for model training and validation.

[0061] Step 2: Train the regenerative constraint-supervised deep learning model:

[0062] The regenerative constraint-supervised deep learning model can be any deep learning model that requires labels during training; in this embodiment, the JCT model is used.

[0063] During model training, the loss function is the core of this invention. This invention employs a loss function with regenerative constraints, including the target loss. and regeneration constraints . This represents an existing loss function used to penalize the error between the model output and the label. Taking the Frobenius norm as an example, it is defined as:

[0064]

[0065] In the formula, This indicates the number of samples in the same batch. Indicates the first in the same batch One sample, This represents a regenerative constraint-supervised deep learning model; express The output of the model after inputting the regenerative constraint-supervised deep learning model.

[0066] Regeneration constraints , by prior terms and stable terms composition.

[0067] The prior term is defined as:

[0068]

[0069] In the formula, Represents a matrix of all 1s; Indicates the randomly selected first... The mask matrix corresponding to each input sample With mask matrix The missing rates are the same.

[0070] The stable term is defined as:

[0071]

[0072] From the perspective of optimization theory, the expected value of the loss is estimated using the Monte Carlo method, defined as...

[0073]

[0074] in, This is the set constraint strength. It is transformed using the generalized Lagrange multiplier method.

[0075]

[0076] in, This is the regularization coefficient set to limit the strength of the regeneration constraint. Due to the limitations of the KKT conditions... Finally, by minimizing the upper bound, we obtain the final loss function for the regeneration constraint, defined as:

[0077]

[0078] The established training dataset was input into the regenerative constraint-supervised deep learning model in batches. The Adam optimizer was used as the network training optimizer, and a step-by-step decay strategy was adopted for the learning rate decay. The regenerative constraint-supervised deep learning model was trained, and the results were validated using a relative error evaluation index. To demonstrate the improvement in the accuracy of structural vibration response reconstruction brought about by this invention, the JCT model was trained under the guidance of the loss function of the final regenerative constraint effect.

[0079] Step 3: Reconstruct the response using the trained model:

[0080] In practical applications, the missing structural vibration response data is input into a trained regenerative constraint-supervised deep learning model to achieve complete reconstruction of the structural vibration response data.

[0081] In this embodiment, samples are randomly selected from the validation set, and the reconstruction result of the 13th sensor is used for presentation. From Figure 4 As can be seen from (a) and (b) in the text, only in The model trained under guidance does not perform well in reconstructing local details, while Figure 4 (c) and (d) show that in The reconstruction results of the model trained under the guidance of regeneration constraints are closer to the true values, indicating that the reconstruction accuracy of the model trained under the guidance of regeneration constraints is higher.

[0082] Example 2:

[0083] This embodiment uses the Hardanger Bridge in Norway as the data collection object.

[0084] The Hardanger Bridge, located in southwestern inland Norway, is a landmark project in the country's suspension bridge system. With a main span of 1308 meters, it boasts the longest span among Norwegian bridges. Figure 5 As shown in (a) above. Since its opening to traffic, the bridge has been equipped with a long-term structural health monitoring system, continuously collecting multi-dimensional data including wind load parameters and three-dimensional acceleration response. This invention selects vertical data from 16 sets of triaxial acceleration sensor arrays deployed at key locations on the bridge deck for analysis, covering the period from January 3rd to 31st, 2018, with a sampling frequency of 10 Hz.

[0085] In this embodiment, the structural vibration response reconstruction method based on a regenerative constraint-supervised deep learning model includes the following steps:

[0086] Step 1: Acceleration response sample data collection and preprocessing:

[0087] Data was collected using a complete array of 16 triaxial accelerometers in the structural health monitoring system. The data was then normalized and scaled to the range [-1, 1], defined as follows:

[0088]

[0089] in This represents the maximum absolute value in the sampled data. Indicates the first The first triaxial accelerometer sensor The original data of each sampling point This represents the corresponding normalized data.

[0090] Then, the normalized output data from the complete single-axis accelerometer is sliced. The data slicing process divides the entire time-series signal into smaller samples, enabling the supervised deep learning model to effectively extract its local and global features. After slicing, the following results are obtained: The labels of each sample are determined, and then for each label, a mask matrix is ​​used. The corresponding samples were calculated:

[0091]

[0092] In the formula, It is the first one used for supervising deep learning models The labels of each input sample, i.e., the complete acceleration response data. Indicates the first term containing missing items One input sample; Indicates the number of single-axis accelerometers. This indicates the number of sampling points for each sample; The index set representing the location of missing items. This indicates element-wise multiplication.

[0093] In this embodiment, a mask matrix is ​​used. To simulate a 25% sensor failure condition, the data of sensors 3, 6, 10, and 15 are randomly set to 0.

[0094] Step 2: Train the regenerative constraint-supervised deep learning model:

[0095] The regenerative constraint-supervised deep learning model can be any deep learning model that requires labels during training; in this embodiment, the DFCN model is used.

[0096] The established training dataset is input into the regenerative constraint-supervised deep learning model in batches. The Adam optimizer is used as the network training optimizer, and a step-by-step decay strategy is adopted for the learning rate decay. The regenerative constraint-supervised deep learning model is trained, and the relative error evaluation index is used for verification. To demonstrate the improvement of the accuracy of structural vibration response reconstruction by this invention, the DFCN model is trained under the guidance of the loss function of the final regenerative constraint effect.

[0097] Step 3: Reconstruct the response using the trained model:

[0098] In this embodiment, samples are randomly selected from the validation set, and the reconstruction result of the 15th sensor is used for presentation. From Figure 6 As can be seen from (a) and (b) in the text, only in The reconstruction results of the DFCN model under guidance have significant errors, while Figure 6 From (c) and (d) we can see The reconstruction results of the DFCN model under the guidance of regeneration constraints are closer to the true values, indicating that the reconstruction accuracy of the DFCN model under the guidance of regeneration constraints is higher.

[0099] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for reconstructing structural vibration response based on a regenerative constraint-supervised deep learning model, characterized in that: Includes the following steps: Step 1: Collect structural vibration response data using sensors deployed on the engineering structure and perform preprocessing; Training samples are generated using preprocessed structural vibration response data. The structural vibration response data is complete data without any missing items; The preprocessed structural vibration response data is sliced ​​to obtain Each segment of data, as Labels of each training sample; For the Labels of each training sample To pass through the set mask matrix The corresponding training samples are calculated as follows: In the formula, Indicates the first term containing missing items One input sample; Indicates the first One sensor, Indicates the first in the sample One sampling point; The index set representing the location of missing items. This represents element-wise multiplication; Step 2: Train the regenerative constraint-supervised deep learning model: The regenerative constraint-supervised deep learning model can be any deep learning model that requires labels to participate in the training process. Using the training samples and their labels obtained in step 1, the regenerative constraint-supervised deep learning model is trained. During the model training process, a loss function based on the regenerative constraint is used. Instructional training: in Loss to the target For regeneration constraints, It is the regularization coefficient set to limit the strength of the regeneration constraint; The target loss is used to penalize the error between the model output and the label; The regeneration constraint is determined by prior terms. and stable terms composition: in: The prior term is defined as: In the formula, Represents a matrix of all 1s; Indicates the randomly selected first... The mask matrix corresponding to each input sample With mask matrix The missing rates are the same; The stable term is defined as: Step 3: Reconstruct the response using the trained regenerative constraint-supervised deep learning model: The missing structural vibration response data actually collected by sensors on the structure is input into a trained regenerative constraint-supervised deep learning model to achieve complete reconstruction of the structural vibration response data.

2. The structural vibration response reconstruction method based on a regenerative constraint supervised deep learning model according to claim 1, characterized in that: The preprocessing process described in step 1 includes downsampling and normalization.

3. The structural vibration response reconstruction method based on a regenerative constraint supervised deep learning model according to claim 1, characterized in that: The target loss is expressed using the Frobenius norm: In the formula, This indicates the number of samples in the same batch. Indicates the first in the same batch One sample, This represents a regenerative constraint-supervised deep learning model; express The output of the model after inputting the regenerative constraint-supervised deep learning model.

4. The structural vibration response reconstruction method based on a regenerative constraint supervised deep learning model according to claim 1, characterized in that: In step 2, when training the regenerative constraint-supervised deep learning model, the Adam optimizer is used as the network training optimizer, and the learning rate decay strategy is a step-by-step decay strategy.

5. An electronic device, comprising a processor and a memory, wherein the memory is used to store one or more programs; characterized in that: When the processor executes the one or more programs, it implements the method of any one of claims 1 to 4.

6. A readable storage medium storing a computer program, characterized in that: When a computer program is executed by a processor, it implements the method described in any one of claims 1 to 4.

Images