Data reconstruction and structural damage identification method and device based on gradient coupling mechanism, equipment, medium and product

Abstract

The invention discloses a gradient coupling mechanism-based data reconstruction and structural damage identification method, apparatus and device, a medium and a product, and relates to the technical field of structural health monitoring. According to the invention, a joint loss function including sensor data reconstruction loss and structural damage identification result loss is constructed, and joint optimization is carried out on a convolutional auto-encoder in a data compression stage and a full-connection neural network in a damage identification stage. And an optimization result is synchronously fed back to the convolutional auto-encoder and the full-connection neural network by adopting a gradient coupling mechanism, so that a mapping relation between existing sensor data and missing sensor data and a mapping relation between data compression characteristics and a damage state are established, and the cooperative training efficiency of sensor data reconstruction and structural damage identification is effectively improved. And meanwhile, the training efficiency and the result precision of double tasks of sensor data reconstruction and structural damage identification are remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of structural health monitoring, and in particular to a method, device, equipment, medium and product for data reconstruction and structural damage identification based on a gradient coupling mechanism. Background Art

[0002] As bridge monitoring systems evolve toward higher density and higher frequency, the large-scale deployment of sensor networks has triggered an explosive growth in data volumes. Modern large bridges often integrate hundreds of multi-type sensor nodes, generating continuous data streams covering multiple physical fields, including strain, vibration, and displacement. This exponential expansion of data volume not only leads to an exponential increase in storage media costs but also poses a severe challenge to data transmission bandwidth and real-time processing capabilities. Traditional compression methods based on signal processing and mathematical operations are no longer able to cope with the continuous onslaught of data.

[0003] The task of structural damage identification presents new technical challenges amidst a surge in data. The complex structural characteristics of ultra-long-span bridges often conceal damage signatures within weak signals. These signatures must not only be accurately extracted from a strong noise background but also meet the timeliness requirements of real-time online diagnosis. Traditional pattern recognition methods rely on manual feature engineering, which inherently limits feature sensitivity and algorithm generalization. These methods struggle to adapt to the rapid iteration of new bridge materials and structural morphologies, resulting in a significant gap between damage identification accuracy and actual engineering needs.

[0004] Artificial intelligence technology offers a new breakthrough in resolving this contradiction. Deep learning-driven compression coding algorithms construct nonlinear mappings through convolutional neural networks, achieving data dimensionality reduction while preserving damage-sensitive features. However, existing models still face dual constraints: on the one hand, the high computational complexity of deep networks conflicts with the limited computing power of the equipment, restricting the training and development of larger models; on the other hand, the abstract features automatically extracted during compression are spatially misaligned with the features required for damage identification, resulting in the implicit loss of key information during the encoding and decoding process, affecting the accuracy of damage identification. Summary of the Invention

[0005] The purpose of this application is to provide a data reconstruction and structural damage identification method, device, equipment, medium and product based on a gradient coupling mechanism, so as to improve the accuracy of damage identification while reducing the computational complexity of the deep model.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a data reconstruction and structural damage identification method based on a gradient coupling mechanism, comprising:

[0008] Constructing a data sample set of a target structure; the data sample set includes a plurality of data samples, and the data samples are structural response data of a plurality of sensors with damage labels;

[0009] Build a dual-task deep neural network model consisting of a convolutional autoencoder and a fully connected neural network;

[0010] Construct a joint loss function that includes sensor data reconstruction loss and structural damage identification result loss;

[0011] According to the data sample set and the joint loss function, a dual-task deep neural network model is trained using a gradient coupling mechanism to obtain a trained dual-task deep neural network model;

[0012] The trained dual-task deep neural network model is used to reconstruct sensor data and identify structural damage of the target structure.

[0013] In a second aspect, the present application provides a data reconstruction and structural damage identification device based on a gradient coupling mechanism, wherein the data reconstruction and structural damage identification device based on the gradient coupling mechanism applies the above-mentioned data reconstruction and structural damage identification method based on the gradient coupling mechanism, and the data reconstruction and structural damage identification device based on the gradient coupling mechanism comprises:

[0014] A sample data set construction module is used to construct a data sample set of a target structure; the data sample set includes a plurality of data samples, and the data samples are structural response data of a plurality of sensors with damage labels;

[0015] A dual-task deep neural network model building module, which is used to build a dual-task deep neural network model that includes a convolutional autoencoder and a fully connected neural network;

[0016] A joint loss function construction module is used to construct a joint loss function that includes sensor data reconstruction loss and structural damage identification result loss;

[0017] A training module is used to train the dual-task deep neural network model using a gradient coupling mechanism according to the data sample set and the joint loss function to obtain a trained dual-task deep neural network model;

[0018] The data reconstruction and structural damage identification module is used to reconstruct sensor data and identify structural damage of the target structure using the trained dual-task deep neural network model.

[0019] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned data reconstruction and structural damage identification method based on the gradient coupling mechanism.

[0020] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned data reconstruction and structural damage identification method based on the gradient coupling mechanism.

[0021] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned data reconstruction and structural damage identification method based on the gradient coupling mechanism.

[0022] According to the specific embodiments provided in this application, this application has the following technical effects.

[0023] The present application provides a method, apparatus, device, medium, and product for data reconstruction and structural damage identification based on a gradient coupling mechanism. The present application first constructs a data sample set for the target structure; builds a dual-task deep neural network model including a convolutional autoencoder and a fully connected neural network; constructs a joint loss function including a sensor data reconstruction loss and a structural damage identification result loss; trains the dual-task deep neural network model using a gradient coupling mechanism based on the data sample set and the joint loss function to obtain a trained dual-task deep neural network model; and uses the trained dual-task deep neural network model to perform sensor data reconstruction and structural damage identification on the target structure. The present application constructs a joint loss function including a sensor data reconstruction loss and a structural damage identification result loss, jointly optimizes the convolutional autoencoder in the data reconstruction stage and the fully connected neural network in the damage identification stage, and uses a gradient coupling mechanism to synchronously feed back the optimization results to the convolutional autoencoder and the fully connected neural network, thereby establishing a mapping relationship between existing sensor data and missing sensor data, and between data compression features and damage status, effectively improving the efficiency of the collaborative training of data reconstruction and damage identification, and significantly improving the accuracy of the results of the dual tasks of data reconstruction and damage identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1A flowchart of a data reconstruction and structural damage identification method based on a gradient coupling mechanism is provided in one embodiment of the present application.

[0026] Figure 2 This is a schematic diagram of a data reconstruction and structural damage identification method based on a gradient coupling mechanism provided in one embodiment of the present application.

[0027] Figure 3 This is a elevational layout diagram of a scaled experimental model of a three-span continuous beam bridge provided in one embodiment of the present application.

[0028] Figure 4 A flowchart of the training process of a dual-task deep neural network model provided in one embodiment of the present application.

[0029] Figure 5 A schematic diagram illustrating the training process of a dual-task deep neural network model according to an embodiment of the present application.

[0030] Figure 6 This is a distribution diagram of additional mass installation points for the structural damage degree identification working condition of a three-span continuous beam scale test model provided in one embodiment of the present application.

[0031] Figure 7 This is a distribution diagram of the additional mass installation points for the spatial positioning condition of structural damage in a scaled experimental model of a three-span continuous beam provided in one embodiment of the present application.

[0032] Figure 8 This is a diagram of data reconstruction results of the data reconstruction and structural damage identification method based on the gradient coupling mechanism of the present application, provided as an example of the present application.

[0033] Figure 9 A comparison result diagram of the data reconstruction and structural damage identification method based on the gradient coupling mechanism of the present application and other methods provided in one embodiment of the present application.

[0034] Figure 10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0036] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0037] In an exemplary embodiment, a data reconstruction and structural damage identification method based on a gradient coupling mechanism is provided, such as Figure 1 and Figure 2 As shown, the process includes the following steps 101 to 105.

[0038] Step 101: construct a data sample set of a target structure; the data sample set includes multiple data samples, and the data samples are structural response data of multiple sensors with damage labels;

[0039] Step 102: Building a dual-task deep neural network model including a convolutional autoencoder and a fully connected neural network;

[0040] Step 103: construct a joint loss function including sensor data reconstruction loss and structural damage identification result loss;

[0041] Step 104: training the dual-task deep neural network model using a gradient coupling mechanism based on the data sample set and the joint loss function to obtain a trained dual-task deep neural network model;

[0042] Step 105 : Use the trained dual-task deep neural network model to reconstruct sensor data and identify structural damage on the target structure.

[0043] The embodiment of the present application establishes a deep association between existing sensor data and missing sensor data, compressed feature signals and damage status through dual-network joint training and dynamic gradient coupling strategy, thereby improving the accuracy of damage identification while ensuring efficient data reconstruction.

[0044] The embodiment of the present application constructs a data sample set of the target structure through sensor signal acquisition and data slicing operations; the data sample combination includes multiple data samples, and the data samples are structural response data of multiple sensors with damage labels.

[0045] In another exemplary embodiment, the above step 101 may be replaced by the following steps 201 to 204 .

[0046] Step 201 simulates a target structure damage condition by attaching a mass block to a preset location on the target structure's surface, and simulates a sensor-missing condition for the target structure by setting the data for specific sensor channels to 0. First, a structural model of the target structure is constructed, and n structural response sensors are arranged on the structural model; the structural response sensors are used to acquire structural response data. Then, damage to the target structure is simulated by attaching a preset mass block to a specific location on the structure's surface, and sensor-missing conditions are simulated by setting the data for specific sensor channels to 0.

[0047] Step 202 collects structural response data from multiple sensors within a preset time period under healthy and damaged structural conditions. This step constructs an n×l dataset based on the structural response data obtained from n sensors under different damage conditions, where l is the total length of all structural response data under different damage conditions. In this embodiment of the present application, structural response data is collected by deploying a multi-sensor network; this structural response data is a continuous time series consisting of data sample points generated by simulating different damage conditions and assigning different structural damage labels.

[0048] Step 203 uses a sliding window to slice the structural response data of multiple sensors within a preset time period under both healthy and damaged conditions, obtaining multiple n×m data samples. This step uses a sliding window method, s sample data points are moved each time, and the data set is sliced to obtain multiple n×m data samples, thereby constructing a data sample set. Here, m is the length of the sliding window, and s is the step size of the sliding window.

[0049] In another exemplary embodiment, the length of the sliding window is an integer power of 2, and the sliding step s is generally less than or equal to the sliding window length.

[0050] In another exemplary embodiment, the specific implementation of step 101 is as follows:

[0051] S1. Build a structural model and arrange acceleration sensors.

[0052] Build an experimental or numerical model of the structure and arrange multiple acceleration sensors on the structure.

[0053] For example, the third-span continuous beam bridge in the embodiment is used as an example for explanation. Figure 3 The following is the elevation layout of the scaled experimental model of a three-span continuous beam bridge. The structural response signal under artificial hammer excitation is collected. Figure 3 The sensor shown collects the vertical acceleration signal of the main beam through Figure 1 and Figure 2 The process shown is used to reconstruct sensor data and identify damage.

[0054] For example, taking a three-span continuous beam bridge as an example, a scaled experimental model of the three-span continuous beam bridge is built, and five acceleration sensors are installed on the bridge.

[0055] S2. Collect acceleration sensor network data under different working conditions.

[0056] Regular artificial hammer excitation is applied to the structure, and the acceleration response of the structure under different damage conditions is collected to obtain structural response data. In the embodiment of the present invention, damage is constructed by adding mass to create multiple damage conditions, including single damage, double damage, and triple damage.

[0057] The data set is constructed from the original structural response data to obtain an n×l matrix, where n is the number of sensors and l is the data length.

[0058] For example, taking a three-span continuous beam bridge as an example, n=5 acceleration sensors are installed on the scaled experimental model of the three-span continuous beam bridge. The sampling frequency of each sensor is 2000 Hz, and 64s of data are collected for each working condition. Therefore, the data sample points collected for each working condition are l=64*2000=128,000, and the data size of each working condition is 5×128,000.

[0059] In step S2, the additional mass installation method includes but is not limited to magnet adsorption, bolt connection, wire bundling and tape entanglement, which do not affect the rigidity of the structure itself and effectively increase the mass of the structural nodes.

[0060] Exemplarily, the damage condition is achieved by fixing points on the surface of the main beam of the structure in a magnetic manner.

[0061] In step S2, the structural damage condition simulates the damage condition of the target structure by attaching a preset mass block at a specific position on the surface of the structure. The damage condition simulation adopts a hierarchically configured additional mass system to implement two types of damage simulation mechanisms: 1) adjusting and controlling the damage area by changing the spatial distribution of the mass block; 2) adjusting and controlling the damage degree by adjusting the mass magnitude of the mass block.

[0062] For example, take a three-span continuous beam bridge as an example. Figure 6 As shown in Figure 1, additional mass blocks of equal mass are installed at different spatial positions on the scaled experimental model to simulate structural damage conditions at different positions.

[0063] For example, take a three-span continuous beam bridge as an example. Figure 7 As shown in the figure, by installing additional mass blocks of different levels at specified points on the scaled experimental model, simulations of structural damage conditions of different degrees can be achieved.

[0064] Exemplarily, taking a three-span continuous beam bridge as an example, as shown in Table 1, single-point damage cases (SDC), double-point damage cases (DDC), and triple-point damage cases (TDC) are manufactured by applying different levels of additional mass at specified points. The additional mass is controlled in a multiple relationship according to the artificial division unit mass M: 2.252 kg of the three-span continuous beam bridge scaled experimental model.

[0065] Table 1 Installation rules for additional mass in damage cases

[0066]

[0067] Note: The data in the table are all multiples of the additional mass compared to the unit mass M = 2.252 kg.

[0068] Exemplarily, taking a three-span continuous beam bridge as an example, as shown in Table 2, single-sensor missing cases (SMC) and double-sensor missing cases (DMC) are manufactured by setting the specified sensor data to 0.

[0069] Table 2 Sensor missing cases

[0070] Working conditions Missing sensor Working conditions Missing sensor Working conditions Missing sensor SMC1 No. 1 DMC1 No. 1, No. 2 DMC4 No. 1, No. 5 SMC2 No. 2 DMC2 No. 1, No. 3 DMC5 No. 2, No. 3 SMC3 No. 3 DMC3 No. 1, No. 4 DMC6 No. 2, No. 4

[0071] S3. Perform data slicing to obtain data samples and construct a data set.

[0072] The sliding window is used to slice the original data set. The window length is w (w << l), and the sliding step is s (s ≤ w), obtaining multiple data samples of size n × m.

[0073] Exemplarily, taking a three-span continuous beam bridge as an example, the sliding window method is used for data slicing. The window length is selected as w = 2048, and the step size s = 256. The window length is selected as an integer power of 2 to facilitate subsequent training and feature extraction of the convolutional autoencoder. After data slicing, a data sample set of size 5 × 2048 with a sample size of 500 is obtained.

[0074] Exemplarily, for the data sample set with a sample size of 500, a 3:1:1 data set division is performed, serving as the training set, validation set, and test set respectively, ensuring that the samples between the data sets do not leak.

[0075] In another exemplary embodiment, damage condition simulation uses hierarchically configured additional mass blocks to implement two types of damage simulation mechanisms: 1) adjusting and controlling the damage area by changing the spatial distribution of the mass blocks; 2) adjusting and controlling the damage degree by adjusting the mass magnitude of the mass blocks.

[0076] In another exemplary embodiment, the installation methods of the mass blocks include but are not limited to magnet adsorption, bolt connection, wire bundling and tape entanglement, which do not affect the stiffness of the structure itself and effectively increase the mass of the structural nodes.

[0077] In the above steps 102 to 104, a dual-task collaborative model of sensor data compression and structural damage identification based on a gradient coupling joint mechanism is built, trained, and tested. Figure 4 and Figure 5 shown.

[0078] A convolutional autoencoder and a fully connected neural network are designed separately. Data samples are fed into the convolutional autoencoder, which then outputs two results: one is the encoded features of data compression, and the other is the result of data reconstruction. Simultaneously, the data compression encoded features are fed into the fully connected neural network for further structural damage identification, outputting the probability prediction values for various damage conditions.

[0079] The data reconstruction results and damage probability prediction values are respectively input into the loss function to calculate the loss function value, and the damage function is fused according to the gradient coupling joint mechanism. The coupled gradient is then backpropagated, and the parameters of the convolutional autoencoder and the fully connected neural network are updated according to the backpropagated gradient. When the total loss function approaches 0 and stabilizes, the parameters of the optimal model are retained.

[0080] In the structural damage condition, the dual-task deep neural network model will learn the mapping relationship between the sensor data compression features and the structural damage labels, so as to achieve the purpose of obtaining structural damage prediction results through sensor data compression feature recognition.

[0081] In another exemplary embodiment, the present application can also implement the supplementation of missing sensor data of the convolutional autoencoder.

[0082] In this embodiment, the sensor missing working condition of the target structure is simulated by setting the specific sensor channel data to 0, and the missing sensor data is encoded and decoded using a convolutional autoencoder to obtain reconstructed data. The reconstructed data is compared with the complete data corresponding to the missing sensor data to verify the performance of the convolutional autoencoder in the embodiment of the present application in supplementing the missing sensor data.

[0083] Furthermore, in the sensor-missing condition, the dual-task deep neural network model will learn the mapping relationship between existing sensor data and missing sensor data, so as to achieve the purpose of reconstructing the full sensor data (including missing sensor data) through the existing sensor data.

[0084] In another exemplary embodiment, the convolutional autoencoder includes an encoder and a decoder, wherein the encoder is connected to the decoder, and the encoder is further connected to the fully connected neural network;

[0085] The encoder is used to encode sample data to obtain deep coding features.

[0086] During the training process, the decoder is used to decode the compressed coded features, establish a mapping relationship between the existing sensor data and the full sensor data (including missing sensor data), and obtain reconstructed data of the full sensor data (including missing sensor data). This reconstructed data is used to calculate the sensor data reconstruction loss. This reconstructed data can be used to train the data supplementation function of the convolutional autoencoder and can also be used to train the dual-task deep neural network model.

[0087] Among them, the encoder includes multiple convolution modules connected in sequence; the convolution module includes a convolution layer, a batch normalization layer and a LeakyReLU function; the decoder includes multiple deconvolution modules connected in sequence, and the deconvolution module includes a deconvolution layer, a batch normalization layer and a ReLU function.

[0088] The fully connected neural network is used to perform damage identification on the data compression coding features to obtain damage prediction results;

[0089] During the training process, the fully connected neural network is used to perform damage identification on the compressed coding features, establish a mapping relationship between the compressed coding features and the damage labels, and obtain structural damage identification results. The structural damage identification results are used to calculate the structural damage identification loss.

[0090] Among them, the fully connected neural network includes multiple linear connection layers and residual connection layers connected in sequence.

[0091] The embodiment of the present application establishes a joint loss function based on a convolutional autoencoder and a fully connected neural network, and simultaneously optimizes data reconstruction and damage classification tasks through a gradient coupling mechanism.

[0092] For example, taking a three-span continuous beam bridge as an example, as shown in Table 3, the network layers of the convolutional autoencoder include 10 convolutional layers (Conv2d), batch normalization (BN) processing is performed after each convolution layer, and the activation function selects the LeakyReLU function; the decoder adopts five deconvolution layers (ConvTrans2d), batch normalization processing is performed after each deconvolution layer, and the activation function selects the ReLU function.

[0093] Table 3 Structural parameters of convolutional autoencoder

[0094]

[0095] Note: B = Batch; C = Channel; H = Height; W = Width

[0096] For example, taking a three-span continuous beam bridge as an example, as shown in Table 4, the fully connected neural network includes 5 linear connection layers and 4 residual connection layers. The residual connection layer structure is shown in Table 5.

[0097] Table 4 Structural parameters of fully connected neural network

[0098] serial number Layer type (dimensionality change) Neuron dropout rate 1 Linear concatenation layer (4096 → 2048) 0.3 2 Residual connection layer (2048→2048) -- 3 Linear concatenation layer (2048→1024) 0.3 4 Residual connection layer (1024→1024) -- 5 Linear connection layer (1024→512) 0.3 6 Residual connection layer (512→512) -- 7 Linear connection layer (512→256) 0.3 8 Residual connection layer (256→256) -- 9 Linear concatenation layer (256 → output dimension) 0.3

[0099] Table 5 Structure of the residual connection layer

[0100] serial number Layer Type Activation Function describe 1 Linear connection layer -- Maintain input data dimensions 2 Batch Normalization Layer -- Adjusting data distribution 3 LeakyReLU activation function α＝0.2 Activate output data 4 Linear connection layer -- Maintain input data dimensions 5 Batch Normalization Layer -- Adjusting data distribution 6 Residual output layer PReLU Activation after element-wise addition

[0101] In another exemplary embodiment, the above-mentioned joint loss function is:

[0102] L total =α·L cls +(1-α)·L comp ;

[0103] Among them, L total is the joint loss function, L cls is the loss of structural damage identification results, L comp is the sensor data reconstruction loss, and α is the gradient coupling coefficient.

[0104] In this embodiment of the present application, the loss function used to calculate the sensor data reconstruction loss is the Mean Squared Error (MSE) loss function, and the loss function used to calculate the structural damage identification result loss is the Cross Entropy loss function. That is, the MSE loss function is used to calculate the loss between the data reconstruction result and the original input data, while the Cross Entropy loss function is used to calculate the loss between the predicted damage condition probability value and the actual value.

[0105] During the model training process, the mean square error loss function and cross entropy loss function are used to calculate the loss of the convolutional neural network data compression and reconstruction results and the fully connected neural network damage identification results, respectively.

[0106] In another exemplary embodiment, the above step 104 may be replaced by the following steps 301 - 309 .

[0107] Step 301: Divide the data sample set into a training set, a validation set, and a test set. In this embodiment of the present application, the data sample set is divided into a training set, a validation set, and a test set in a ratio of 3:1:1. The samples in the training set, the validation set, and the test set are strictly isolated, and there is no information leakage between the data sets.

[0108] Step 302: Initialize the gradient coupling coefficient, the minimum value of the gradient coupling coefficient, the maximum value of the gradient coupling coefficient, and the gradient coupling coefficient update step size.

[0109] Step 303: Initialize the initial value of the first iteration number to 1, and initialize the initial value of the second iteration number to 1.

[0110] In step 304, the sample data in the training set is input into the dual-task deep neural network model of the t-th iteration, and the joint loss function is used to calculate the joint loss function value of the t-th iteration; t is the value of the first iteration number.

[0111] Step 305, determine whether the value of the second number of iterations reaches the preset number of iterations, obtain a first judgment result, and determine whether the value of the first number of iterations reaches the maximum number of iterations threshold or the joint loss function value of the tth iteration is not greater than the preset loss function threshold, and obtain a second judgment result.

[0112] Step 306: If both the first judgment result and the second judgment result indicate no, then based on the joint loss function value of the t-th iteration, the parameters of the dual-task deep neural network model of the t-th iteration are updated by back propagation to obtain the dual-task deep neural network model of the t+1-th iteration, the value of the second iteration number is increased by 1, and the value of the first iteration number is increased by 1, and the process returns to the step of “inputting the sample data in the training set into the dual-task deep neural network model of the t-th iteration and calculating the joint loss function value of the t-th iteration”;

[0113] Step 307: If the first judgment result indicates yes and the second judgment result indicates no, then based on the joint loss function value of the t-th iteration, the parameters of the dual-task deep neural network model of the t-th iteration are updated by back propagation to obtain the dual-task deep neural network model of the t+1-th iteration. According to the verification set, the gradient coupling coefficient is updated by a dynamic adjustment strategy, the value of the second iteration number is set to 1, and the value of the first iteration number is increased by 1, and the step of "inputting the sample data in the training set into the dual-task deep neural network model of the t-th iteration and calculating the joint loss function value of the t-th iteration" is returned.

[0114] Step 308: If the second judgment result indicates yes, output the dual-task deep neural network model of the tth iteration as the trained dual-task deep neural network model.

[0115] Step 309: Test the trained dual-task deep neural network model using the test set to obtain a trained dual-task deep neural network model.

[0116] In another exemplary embodiment, a classification performance evaluation window with a length of 10 is established, and the F1 score of the validation set is calculated every 5 training cycles and the window sequence is recorded.

[0117] Before training, the hyperparameters are first set. In the embodiment of the present application, the batch size batch_size = 256, the learning rate learning_rate = 0.001, and the optimizer selects the Adam optimizer, which uses a momentum optimization algorithm and has the advantage of adaptively adjusting the learning rate.

[0118] In an embodiment of the present application, based on the training process of the dual-task collaborative model of the convolutional neural network and the fully connected neural network, a three-stage adjustment strategy of the gradient coupling coefficient is introduced during the training process, and the gradient coupling coefficient is adjusted in real time according to the classification performance of the model on the validation set, giving priority to ensuring the optimization of the damage identification task; in the warm-up stage, the proportion of damage identification learning of the fully connected neural network structure is gradually increased, and in the emergency braking stage, the proportion of damage identification learning of the fully connected neural network structure is reduced by exponential decay, and the gradient coupling coefficient is kept unchanged in the maintenance stage; based on the dynamically optimized dual-task collaborative model, real-time processing and calculation of unknown working condition data are performed, and high-fidelity data reconstruction and accurate prediction of damage status are simultaneously achieved.

[0119] In another exemplary embodiment, a three-stage dynamic adjustment strategy for the gradient coupling coefficient is as follows:

[0120] a) Emergency braking stage: When all F1 scores in the evaluation window are ≥ 0.95, execute α new =α0·exp(-0.1(t-t0)) exponential decay update method; where α new is the updated gradient coupling coefficient, α0 is the initial value of the gradient coupling coefficient when the emergency braking stage is triggered, t is the value of the first iteration number, the current training round, t0 is the initial iteration number that satisfies the F1 score of the sample data of each batch is greater than the preset score threshold, that is, the initial iteration round when the emergency braking stage is triggered.

[0121] b) Warm-up phase: When the F1 score in less than half of the evaluation window is ≥ 0.95, execute α new =min(α old +η,α max ) is a step-by-step incremental update method; where α new is the updated gradient coupling coefficient, α old is the gradient coupling coefficient before updating, η is the gradient coupling coefficient update step size, αmax is the maximum value of the gradient coupling coefficient.

[0122] c) Maintenance stage: When conditions a) and b) are not met, the current α value remains unchanged.

[0123] In the subsequent five training cycles, the joint loss function value and coupling gradient are calculated based on the updated α value, and the convolutional autoencoder and fully connected neural network parameters are updated by backpropagation.

[0124] For example, the calculation method of the above F1 score is:

[0125] In the structural damage identification problem, there are two types of definitions for a specific target damage condition:

[0126] Positive: The structure has damage corresponding to the target damage condition description;

[0127] Negative: The structure is healthy (no damage) or has damage corresponding to other damage conditions.

[0128] In the problem of structural damage identification, the relationship between damage prediction results and actual damage conditions can be divided into four categories:

[0129] TP (True Positive): There is damage, and the model correctly identifies it as damage (correct identification);

[0130] FP (False Positive): There is actually no damage, but the model mistakenly judges it as damage (false positive);

[0131] TN (True Negative): There is actually no damage, and the model correctly judges it as no damage (correct identification);

[0132] FN (False Negative): There is damage, but the model mistakenly judges it as no damage (missing report);

[0133] Based on the above definitions, we can calculate the precision (P) and recall (R):

[0134] Precision (P): The proportion of samples predicted to be damaged that are actually damaged.

[0135]

[0136] Recall (R): The proportion of samples that are correctly predicted to be damaged among samples that are actually damaged.

[0137]

[0138] F1 is the harmonic mean of P and R (balancing the two), and the formula for calculating F1 is as follows:

[0139]

[0140] To identify multiple damage conditions, F1 needs to be calculated for each damage condition separately and then aggregated through macro-average (Macro-F1). The calculation steps are as follows:

[0141] (1) Calculate F1 for each damage condition (the current damage condition is recorded as the positive class, and other damage conditions are recorded as the negative class).

[0142] (2) Take the arithmetic mean of F1 for all damage conditions.

[0143] The calculation formula is as follows:

[0144]

[0145] Among them, Macro-F1 is the F1 score, and its value range is (0,1]. The closer it is to 0, the worse the performance of the model damage identification result is, and the closer it is to 1, the better the performance of the model damage identification result is.

[0146] In another exemplary embodiment, the specific implementation method of the above-mentioned step 105 is: collecting structural response data of the target structure under unknown working conditions, obtaining a corresponding data test set, and using the trained dual-task deep neural network model to obtain full sensor data reconstruction results (including missing sensor data) and damage identification results.

[0147] In the embodiments of the present application, the training set, validation set, and test set can all use structural acceleration signals collected from actual laboratory model structures. When certain damage conditions of the actual structure are not convenient to construct, finite element simulation can be used for simulation.

[0148] like Figure 8 As shown, the sensor data reconstruction results of the data reconstruction and structural damage identification method based on the gradient coupling mechanism of this application are Figure 8 Schematic diagram of sensor data reconstruction results, Figure 8 (a) is the time domain comparison diagram of the original and reconstructed signals. Figure 8 (b) is a schematic diagram of the linear fitting results of the original-reconstructed signal. Figure 8 The example in Table 1 is the missing sensor SM1 condition, that is, after the data of sensor No. 1 is artificially set to 0, the full sensor data (including the missing sensor No. 1) is reconstructed based on the data of the remaining four sensors.

[0149] In the embodiment of the present application, for the nine working conditions where sensor data is missing, two data indicators are used to evaluate the results, namely the Normalized Root Mean Square Error (NRMSE) and the Coefficient of Determination (R 2 ) is used to evaluate the error between the reconstructed signal and the original signal. The calculation formula is as follows:

[0150]

[0151] where y i and are real data and predicted data respectively, y max and y min are the maximum and minimum values of the real data, is the mean of the real data.

[0152] As shown in Table 6, the data reconstruction results of each sensor missing working condition of the data reconstruction and structural damage identification method based on the gradient coupling mechanism of the present application.

[0153] Table 6 Data reconstruction results for sensor missing condition

[0154]

[0155]

[0156] like Figure 9 As shown in the figure, the structural damage identification method based on the gradient coupling mechanism of this application is compared with the results of the other four comparison models. Figure 9 (a)-(e) in the figure are radar charts comparing the performance of the two-stage training model (Two-Stage), the end-to-end encoder-classifier neural network (E2E-Encoder-FCNN), the end-to-end encoder-classifier neural network (pre-trained encoder) (PT-Encoder-FCNN), the fixed gradient coupling mechanism model (Static Coupled), and the structural damage identification method based on the gradient coupling mechanism of this application. Under the same conditions, the method proposed in this application achieves the best balanced performance in five dimensions: data reconstruction correlation, data compression and reconstruction normalized root mean square error, structural damage identification accuracy, training time, and training iteration rounds.

[0157] The method proposed in this application can collaboratively process the dual tasks of sensor data reconstruction and structural damage identification. The convolutional autoencoder, which falls under the category of unsupervised learning, can extract deep features from the raw data. It also accelerates the training of a fully connected neural network through a combined mechanism of damage label and gradient coupling. The fully connected neural network, which falls under the category of supervised learning, can achieve high-precision and efficient damage identification. Compared to other damage identification methods, this method offers advantages such as low training overhead, strong task collaboration, fast computational speed, and high recognition accuracy.

[0158] Based on the same inventive concept, embodiments of the present application also provide a device for data reconstruction and structural damage identification based on a gradient coupling mechanism, for implementing the aforementioned method for data reconstruction and structural damage identification based on a gradient coupling mechanism. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for data reconstruction and structural damage identification based on a gradient coupling mechanism provided below can be found in the aforementioned limitations of the method for data reconstruction and structural damage identification based on a gradient coupling mechanism, and will not be further elaborated here.

[0159] In an exemplary embodiment, a data reconstruction and structural damage identification device based on a gradient coupling mechanism is provided, comprising:

[0160] A sample data set construction module is used to construct a data sample set of a target structure; the data sample set includes a plurality of data samples, and the data samples are structural response data of a plurality of sensors with damage labels;

[0161] A dual-task deep neural network model building module, which is used to build a dual-task deep neural network model that includes a convolutional autoencoder and a fully connected neural network;

[0162] A joint loss function construction module is used to construct a joint loss function that includes sensor data reconstruction loss and structural damage identification result loss;

[0163] A training module is used to train the dual-task deep neural network model using a gradient coupling mechanism according to the data sample set and the joint loss function to obtain a trained dual-task deep neural network model;

[0164] The data reconstruction and structural damage identification module is used to reconstruct sensor data and identify structural damage of the target structure using the trained dual-task deep neural network model.

[0165] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a data reconstruction and structural damage identification method based on a gradient coupling mechanism is implemented.

[0166] Those skilled in the art will understand that Figure 10 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application and does not constitute a limitation on the computer device to which the solution of the present application is applied. A specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the above-mentioned method embodiments when executing the computer program.

[0167] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0168] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0169] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0170] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0171] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0172] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A data reconstruction and structural damage identification method based on gradient coupling mechanism, characterized in that: include: Constructing a data sample set of a target structure; the data sample set includes a plurality of data samples, and the data samples are structural response data of a plurality of sensors with damage labels; Build a dual-task deep neural network model consisting of a convolutional autoencoder and a fully connected neural network; Construct a joint loss function that includes sensor data reconstruction loss and structural damage identification result loss; According to the data sample set and the joint loss function, a dual-task deep neural network model is trained using a gradient coupling mechanism to obtain a trained dual-task deep neural network model; The trained dual-task deep neural network model is used to reconstruct sensor data and identify structural damage of the target structure.

2. The data reconstruction and structural damage identification method based on the gradient coupling mechanism according to claim 1 is characterized in that: The data sample set for constructing the target structure specifically includes: The damage condition of the target structure is simulated by setting additional mass blocks at preset positions on the surface of the target structure; Simulating the healthy working condition of the target structure without setting additional mass blocks on the surface of the target structure; Collect structural response data from multiple sensors within a preset time period under different working conditions; different working conditions include healthy working conditions and different damaged working conditions; A sliding window is used to slice the structural response data of multiple sensors within a preset time period under each working condition to obtain multiple n×m data samples; where n is the number of sensors, m is the length of the sliding window, the value of m is an integer power of 2, and the step size of the sliding window is less than or equal to the length of the sliding window.

3. The data reconstruction and structural damage identification method based on the gradient coupling mechanism according to claim 2 is characterized in that: The damage condition of the target structure is simulated by setting additional mass blocks at preset positions on the surface of the target structure, including: The damage location is simulated by adjusting the spatial distribution of the additional mass blocks; The damage degree is simulated by changing the mass magnitude of the additional mass block.

4. The data reconstruction and structural damage identification method based on the gradient coupling mechanism according to claim 1 is characterized in that: The joint loss function is: L total =α·L cls +(1-a)·L comp ; Among them, L total is the joint loss function, L cls is the loss of structural damage identification results, L comp is the sensor data reconstruction loss, and α is the gradient coupling coefficient.

5. The data reconstruction and structural damage identification method based on the gradient coupling mechanism according to claim 1 is characterized in that: According to the data sample set and the joint loss function, the dual-task deep neural network model is trained using a gradient coupling mechanism to obtain a trained dual-task deep neural network model, specifically including: Divide the data sample set into training set, validation set and test set; Initialize the gradient coupling coefficient, the minimum value of the gradient coupling coefficient, the maximum value of the gradient coupling coefficient and the update step size of the gradient coupling coefficient; Initialize the first iteration number to 1, and initialize the second iteration number to 1; Input the sample data in the training set into the dual-task deep neural network model of the t-th iteration, and use the joint loss function to calculate the joint loss function value of the t-th iteration; t is the value of the first iteration; Determine whether the value of the second iteration number reaches the preset iteration number to obtain a first judgment result, and determine whether the value of the first iteration number reaches the maximum iteration number threshold or the joint loss function value of the t-th iteration is not greater than the preset loss function threshold to obtain a second judgment result; If both the first judgment result and the second judgment result indicate no, then based on the joint loss function value of the t-th iteration, the parameters of the dual-task deep neural network model of the t-th iteration are updated by back propagation to obtain the dual-task deep neural network model of the t+1-th iteration, the value of the second iteration number is increased by 1, and the value of the first iteration number is increased by 1, and the process returns to the step of "inputting the sample data in the training set into the dual-task deep neural network model of the t-th iteration and calculating the joint loss function value of the t-th iteration"; If the first judgment result indicates yes and the second judgment result indicates no, then based on the joint loss function value of the t-th iteration, the parameters of the dual-task deep neural network model of the t-th iteration are updated by back propagation to obtain the dual-task deep neural network model of the t+1-th iteration. According to the validation set, the gradient coupling coefficient is updated by a dynamic adjustment strategy, the value of the second iteration number is set to 1, and the value of the first iteration number is increased by 1, and the process returns to the step of "inputting the sample data in the training set into the dual-task deep neural network model of the t-th iteration and calculating the joint loss function value of the t-th iteration". If the second judgment result indicates yes, outputting the dual-task deep neural network model of the tth iteration as the trained dual-task deep neural network model; The trained dual-task deep neural network model is tested using the test set to obtain a trained dual-task deep neural network model.

6. The data reconstruction and structural damage identification method based on the gradient coupling mechanism according to claim 5 is characterized in that: According to the test set, the gradient coupling coefficient is updated using a dynamic adjustment strategy, specifically including: Input the sample data of the preset batch in the test set into the dual-task deep neural network model of the t+1th iteration, and calculate the F1 score of the sample data of each batch; When the F1 score of each batch of sample data is greater than the preset score threshold, the gradient coupling coefficient is updated using the following formula: a new =α0·exp(-0.1(t-t0)); Among them, α new is the updated gradient coupling coefficient, α0 is the initial value of the gradient coupling coefficient, and t0 is the initial number of iterations required to ensure that the F1 score of each batch of sample data is greater than the preset score threshold; When the number of batches whose F1 scores are all greater than the preset score threshold is within a preset range, the gradient coupling coefficient is updated using the following formula; a new =min(a old +h,a max ); Among them, α new is the updated gradient coupling coefficient, α old is the gradient coupling coefficient before updating, η is the gradient coupling coefficient update step size, α max is the maximum value of the gradient coupling coefficient.

7. A data reconstruction and structural damage identification device based on a gradient coupling mechanism, characterized in that: The data reconstruction and structural damage identification device based on the gradient coupling mechanism applies the data reconstruction and structural damage identification method based on the gradient coupling mechanism according to any one of claims 1 to 6, and the data reconstruction and structural damage identification device based on the gradient coupling mechanism comprises: A sample data set construction module is used to construct a data sample set of a target structure; the data sample set includes a plurality of data samples, and the data samples are structural response data of a plurality of sensors with damage labels; A dual-task deep neural network model building module, which is used to build a dual-task deep neural network model that includes a convolutional autoencoder and a fully connected neural network; A joint loss function construction module is used to construct a joint loss function that includes sensor data reconstruction loss and structural damage identification result loss; A training module is used to train the dual-task deep neural network model using a gradient coupling mechanism according to the data sample set and the joint loss function to obtain a trained dual-task deep neural network model; The data reconstruction and structural damage identification module is used to reconstruct sensor data and identify structural damage of the target structure using the trained dual-task deep neural network model.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the data reconstruction and structural damage identification method based on the gradient coupling mechanism according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the data reconstruction and structural damage identification method based on the gradient coupling mechanism described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the data reconstruction and structural damage identification method based on the gradient coupling mechanism described in any one of claims 1 to 6 is implemented.

Images