Damage identification method and device for cable-stayed bridge based on mass data

Abstract

The invention relates to a damage identification method and device for a cable-stayed bridge based on mass data. The method comprises the following steps: constructing a standard finite element model of the cable-stayed bridge, and collecting structural dynamic response data under various damage working conditions; an improved complete adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm is provided to decompose data, and the noise level is dynamically adjusted through a standard deviation feedback mechanism to obtain an IMF component; the Hurst index is innovatively used for screening meeting Hgt; 0.5, and the reconstruction error MSElt; a target IMF component of 0.0003 is obtained; and constructing and training a one-dimensional convolutional neural network model, inputting a target IMF component, and outputting an identification result of a damage position and degree. The model comprises a leading convolution layer group, an intermediate feature layer group, a depth abstraction layer group and a classification output layer. In addition, the invention also provides a damage identification device, electronic equipment, a computer readable storage medium and a computer program product, so as to realize the method. The damage of the cable-stayed bridge can be effectively identified, and technical support is provided for bridge health monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring, and particularly relates to a method and device for damage identification of cable-stayed bridges based on massive data. Background Art

[0002] With the wide application of the structural health monitoring system for bridge engineering, massive long-term monitoring data has been accumulated for long-span bridges such as cable-stayed bridges. These data contain rich structural state information, providing an important basis for damage identification. However, the monitoring data of cable-stayed bridges has characteristics such as strong noise interference, non-stationarity, and high dimensionality, and traditional damage identification methods face severe challenges.

[0003] Currently, damage identification methods based on machine learning have gradually become a research hotspot, but there are still significant defects in their practical applications: noise sensitivity problem: when traditional one-dimensional convolutional neural networks directly process noisy signals, the model is easily interfered by environmental noise, resulting in insufficient extraction of damage features and a sharp decline in recognition accuracy as the noise intensity increases; existing signal processing methods rely on fixed parameter settings, and modal aliasing and false components are easily generated during the decomposition process, resulting in the loss of key feature information; most studies directly use all intrinsic mode function components as inputs without considering the information effectiveness of different components, and redundant features increase the model complexity and reduce the generalization ability; existing methods are mostly limited to single damage location or quantification, lacking the ability to jointly identify multi-location and multi-degree compound damage, and the quantification error often exceeds 5%, making it difficult to meet the engineering accuracy requirements; traditional noise suppression algorithms use static thresholds and cannot dynamically adjust the noise level according to signal characteristics, resulting in a significant increase in signal reconstruction error in high-noise environments. Although some studies have tried to combine signal decomposition and deep learning, their noise adjustment mechanism is fixed and no feature quality evaluation index is introduced, resulting in insufficient robustness and computational efficiency of the model in the scenario of massive data of cable-stayed bridges. Summary of the Invention

[0004] The present application provides a method and device for damage identification of cable-stayed bridges based on massive data to solve the problems of insufficient noise sensitivity, limited signal decomposition, insufficient feature screening, and lack of dynamic adaptability in the prior art.

[0005] The first aspect of the present application provides a method for damage identification of cable-stayed bridges based on massive data, including the following steps: S1: Construct a benchmark finite element model of the cable-stayed bridge; S2: Collect structural dynamic response data under various damage conditions according to the benchmark finite element model; S3: An improved complete adaptive noise ensemble empirical mode decomposition algorithm ICEEMDAN is proposed to decompose the structural dynamic response data to obtain IMF components, where the decomposition algorithm ICEEMDAN is reflected in dynamically adjusting the noise level through a standard deviation feedback mechanism; S4: Innovatively perform Hurst index screening on the IMF components, and select target IMF components that satisfy the Hurst index H>0.5 and the reconstruction error MSE<0.0003; S5: Construct and train a one-dimensional convolutional neural network model, input the target IMF components into the trained one-dimensional convolutional neural network model, and output the identification results of the damage location and damage degree, where the one-dimensional convolutional neural network model includes a leading convolutional layer group, an intermediate feature layer group, a deep abstraction layer group, and a classification output layer.

[0006] Preferably, the decomposition algorithm ICEEMDAN includes: S3.1.1 Add the initial noise level to the original data to obtain noisy data, and the formula is as follows:

[0007]

[0008] where: x(t) is the original signal; is the noise level; N 0(t) is a Gaussian white noise sequence;

[0009] S3.1.2: Use EMD to decompose the noisy data to obtain the first residual signal, and the formula is as follows:

[0010] R1 = <M(x i )> (2)

[0011] where: <·> is the overall average; M(·) is the local mean of the input signal;

[0012] S3.1.3: Calculate the error between the first and the original data, and the formula is as follows:

[0013]

[0014] where: is the standard deviation of the IMF1 component; σ x is the standard deviation of the original data; E1 is the error;

[0015] S3.1.4: Update the noise level according to the error E1 using a non-linear strategy, and the formula is as follows:

[0016] σ1 = σ0 * exp(β * E1) (6)

[0017] Wherein: is the updated noise level; β is the control coefficient for noise adjustment;

[0018] S3.1.5: Remove the first IMF component from the original signal to obtain the remaining signal, and the formula is as follows:

[0019] x1(t) = x(t) - IMF1(t) (7)

[0020] S3.1.6: Add the updated noise to the remaining signal, and the calculation formula is as follows:

[0021]

[0022] Wherein: N 1(t) is a new set of Gaussian white noise;

[0023] S3.1.7: Perform EMD decomposition on the new noisy signal to extract the second-order intrinsic mode function;

[0024] S3.1.8: Repeat the above calculations to continue adding noise to the remaining signal, perform EMD decomposition, extract the subsequent IMF components, and calculate the error and adjust the noise level after each round of iteration:

[0025] σ n+1 = σ n * exp(β * E n )(9)

[0026] Wherein: is the updated noise level; En is the error; β is the control coefficient for noise adjustment.

[0027] Preferably, the calculation of the Hurst exponent includes: S4.1.1: Preprocess the given time series x(t) to obtain the processed series, and the formula is as follows:

[0028]

[0029] Wherein: is the mean value of the time series;

[0030] S4.1.2: Use the least squares method for nonlinear fitting to obtain the local trend terms of each series, and the formula is as follows:

[0031]

[0032] Wherein a n is the fitting polynomial coefficient;

[0033] S4.1.3: Utilize the local trend term y m (k) Calculate the second-order fluctuation function of the time series, and the function is as follows:

[0034]

[0035] F q (n) ∝ n α (13)

[0036] Where: α is the Hurst exponent;

[0037] S4.1.4: Calculate the slope of the fitting curve to obtain the Hurst exponent, and the formula is as follows:

[0038]

[0039] S4.1.5: By calculating the Hurst exponents of each IMF component, the IMF components can be effectively selected, and the MSE calculation formula is:

[0040]

[0041] Where: x(t) is the original signal; is the reconstructed signal.

[0042] Preferably, the one-dimensional convolutional neural network model includes: S5.1.1: The input layer receives a 6000-dimensional time series signal; S5.1.2: The first convolutional layer uses 16 convolutional kernels with a size of 8 and a stride of 2; S5.1.3: The second pooling layer uses max pooling with a pooling size of 2; S5.1.4: The deep convolutional layer uses 512 convolutional kernels with a size of 2; S5.1.5: The global average pooling layer is connected to a 512-node fully connected layer; S5.1.6: The Dropout layer sets a dropout rate of 0.3; S5.1.7: The output layer uses the Softmax activation function.

[0043] Preferably, before decomposing the structural dynamic response data by using an improved complete adaptive noise ensemble empirical mode decomposition algorithm ICEEMDAN to obtain IMF components, it further includes: S3.2.1: Add Gaussian white noise to the original stress response signal x(t) to construct a noisy signal:

[0044] x_noise(t) = x(t) + ησ_xξ(t) (16)

[0045] Where, σ_x is the signal standard deviation, η ∈ {5%, 10%, 15%} is the noise intensity coefficient, and ξ(t) is a standard normal distribution random number.

[0046] Preferably, the multiple damage conditions adopt a multi-label classification strategy, where the multi-label classification strategy includes a non-damaged state encoding, single damage conditions, and multiple damage conditions, including: S5.2.1: Encoding the non-damaged state as category 0, corresponding to the non-damaged condition; S5.2.2: Encoding the single damage conditions by grading according to the damage factor: the damage factors of 0.05 / 0.10 / 0.15 at measuring point 3 correspond to categories 1-3 respectively, and the same-level damage factors at measuring point 8 correspond to categories 4-6; S5.2.3: Encoding rule for multiple damage conditions: when there is a combined damage at measuring points 3 and 11, the damage factors of 0.05 / 0.10 / 0.15 correspond to categories 7-9 respectively; S5.2.4: The output layer of the one-dimensional convolutional neural network is set with 10 neurons, corresponding to the probability distributions of damage conditions of categories 0-9 respectively.

[0047] In the second aspect of the present application, an embodiment provides a damage identification device for massive data of a cable-stayed bridge, including: a modeling module for constructing a reference finite element model of the cable-stayed bridge; a data acquisition module for collecting structural dynamic response data under multiple damage conditions according to the reference finite element model; a signal decomposition module that proposes an improved complete ensemble empirical mode decomposition algorithm with adaptive noise ICEEMDAN to decompose the structural dynamic response data to obtain IMF components, where the decomposition algorithm ICEEMDAN is reflected in dynamically adjusting the noise level through a standard deviation feedback mechanism; a feature screening module for innovatively performing Hurst index screening on the IMF components, and selecting target IMF components that satisfy the Hurst index H>0.5 and the reconstruction error MSE<0.0003; a network construction module for constructing and training a one-dimensional convolutional neural network model, inputting the target IMF components into the trained one-dimensional convolutional neural network model, and outputting the identification results of the damage location and damage degree, where the one-dimensional convolutional neural network model includes a leading convolutional layer group, an intermediate feature layer group, a deep abstraction layer group, and a classification output layer.

[0048] In the third aspect of the present application, an embodiment provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the program to implement a damage identification method for a cable-stayed bridge based on massive data as described in the above embodiment.

[0049] In the fourth aspect of the present application, an embodiment provides a computer-readable storage medium, on which a computer program is stored, and the program is executed by a processor to be used to implement a damage identification method and device for a cable-stayed bridge based on massive data as described in the above embodiment.

[0050] In the fifth aspect of the present application, an embodiment provides a computer program product, including a computer program or instruction, to be used to implement a damage identification method for a cable-stayed bridge based on massive data as described in the above embodiment.

[0051] Therefore, this application has the following beneficial effects:

[0052] In the embodiments of this application, the improved ICEEMDAN algorithm is used to dynamically adjust the noise suppression level. By combining the Hurst exponent to screen key feature components, a deeply optimized 1D-CNN model is constructed, achieving high-precision identification of cable-stayed bridge damage and breakthroughs in anti-noise performance. In a low-noise environment, the model can accurately extract damage-sensitive features. Under 15% high-intensity noise interference, the signal reconstruction quality can still be guaranteed through a dynamic compensation mechanism. At the same time, based on the multi-label classification strategy and the fusion of finite element simulation data, the system supports the collaborative identification of cross-measurement point composite damage and realizes the quantitative grading of the damage degree. Therefore, this application solves the problems in the prior art such as strong noise sensitivity, high feature redundancy, insufficient multi-damage identification ability, and low computational efficiency, improves the cable-stayed bridge damage identification accuracy to over 99%, controls the quantitative error within 3%, realizes the full-process optimization from data acquisition, feature extraction to damage diagnosis, and provides a highly reliable intelligent solution for bridge health monitoring.

[0053] Some of the additional aspects and advantages of this application will be given in the following description, some will become obvious from the following description, or be understood through the practice of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The above and / or additional aspects and advantages of this application will become obvious and easy to understand from the following description of the embodiments in conjunction with the drawings, where:

[0055] Figure 1 is a flowchart of a cable-stayed bridge damage identification method based on massive data according to an embodiment of this application;

[0056] Figure 2 is a flowchart of a cable-stayed bridge damage identification method based on massive data according to an embodiment of this application;

[0057] Figure 3 is a schematic diagram of a cable-stayed bridge finite element model according to an embodiment of this application;

[0058] Figure 4 is a schematic diagram of the layout of the damage area, measurement points and stress sensors of the main girder of a cable-stayed bridge according to an embodiment of this application;

[0059] Figure 5 is a schematic diagram of the stress signal response before and after damage at measurement point 3 according to an embodiment of this application;

[0060] Figure 6 is a schematic diagram of the position of the damage area on the element according to an embodiment of this application;

[0061] Figure 7Schematic diagram of reconstruction error provided according to an embodiment of the present application;

[0062] Figure 8 Curves of loss values and accuracies of the training set and the test set provided according to an embodiment of the present application

[0063] Schematic diagram;

[0064] Figure 9 Schematic diagram of the real bridge stress response data provided according to an embodiment of the present application;

[0065] Figure 10 Schematic diagram of the structure of the large - scale data damage identification device for cable - stayed bridges provided according to an embodiment of the present application;

[0066] Figure 11 Schematic diagram of the structure of the electronic device provided according to an embodiment of the present application;

[0067] Figure 12 Introduction diagram of the method and device for damage identification of cable - stayed bridges based on large - scale data provided according to an embodiment of the present application. Detailed implementation manners

[0068] The embodiments of the present application will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present application, but should not be construed as limiting the present application.

[0069] The following describes a method and device for damage identification of a cable-stayed bridge based on massive data according to an embodiment of the present application. Aiming at the problems of strong noise sensitivity, high feature redundancy, and insufficient multi-damage identification accuracy proposed in the above-mentioned background technology, the present application provides a method for damage identification of a cable-stayed bridge based on massive data. In this method, the noise suppression level is dynamically adjusted through an improved ICEEMDAN algorithm, key feature components are screened by combining the Hurst index, and a deeply optimized 1D-CNN model is constructed to achieve high-precision identification of cable-stayed bridge damage and a breakthrough in anti-noise performance. Through a standard deviation feedback mechanism, the noise compensation factor γ (γ = 0.3) is optimized in real time, and the damage location accuracy of 98.59% can still be maintained under 15% high-intensity noise interference (traditional method ≤ 58.44%), and the reconstruction error is reduced to 0.0002 (compared with 0.0015 of VMD). At the same time, based on the fusion of the multi-label classification strategy and finite element simulation data, the system supports the collaborative identification of cross-measurement point composite damage, with a single damage location error ≤ 0.5 m (measurement point spacing 2 m) and a multi-damage quantification error ≤ 2.94% (traditional method ≥ 8.2%). This method significantly improves the robustness and computational efficiency of damage identification in a complex noise environment through the collaborative optimization of dynamic noise suppression, feature effectiveness evaluation, and lightweight network design, and at the same time solves the problems of strong noise sensitivity, high feature redundancy, insufficient multi-damage identification ability, and large computational resource consumption in traditional methods, providing a highly reliable and low-latency intelligent solution for the health monitoring of cable-stayed bridges, and achieving a breakthrough in the full-process performance from data acquisition, feature extraction to damage diagnosis.

[0070] Specifically, Figure 1 is a schematic flow chart of a method for damage identification of a cable-stayed bridge based on massive data provided by an embodiment of the present application.

[0071] As Figure 1 shown, the method for damage identification of a cable-stayed bridge based on massive data includes the following steps:

[0072] In step S1, a reference finite element model of the cable-stayed bridge is constructed.

[0073] Among them, the reference finite element model is established using ABAQUS software. The main girder adopts C3D8R solid elements (mesh size 0.5 m), the stay cables adopt T3D2 truss elements, and fixed constraint conditions are set at the bottom of the cable tower.

[0074] It can be understood that a refined model containing 500,000 elements (main girder C3D8R / cable stay T3D2) is established through ABAQUS, 10 damage conditions (damage factor 0.05 - 0.15) are set, and the model accuracy is verified (modal frequency error < 1.2%), providing a high-confidence data basis for subsequent simulations.

[0075] It should be noted that the model is double calibrated through measured modal parameters (frequency error < 1.2%, mode shape MAC value > 0.95) and static load tests (displacement deviation < 2 mm) to ensure the engineering applicability of damage simulation. The main beam element size is optimized to 0.5 m to capture local damage effects, the cable prestress application accuracy reaches 0.1 kN, and the tower constraint conditions strictly reproduce the actual bridge bearing form.

[0076] In step S2, structural dynamic response data under various damage conditions are collected according to the benchmark finite element model.

[0077] Among them, the dynamic response data includes the stress time history signals (sampling rate 200 Hz) of 15 measuring points on the main beam and the cable force vibration spectrum data (frequency resolution 0.1 Hz).

[0078] It can be understood that the stress time history (sampled at 200 Hz) covering 15 measuring points on the main beam (spacing 2 m) and the cable force spectrum data are used to generate 4,752,000 groups of samples with 5% - 15% noise to simulate the data characteristics under actual environmental interference.

[0079] It should be noted that the data acquisition module integrates a temperature compensation algorithm to eliminate the influence of environmental temperature drift on the strain signal (temperature drift error < 0.5 με / ℃). The white noise excitation bandwidth covers the main vibration frequency band of the bridge from 0.1 - 50 Hz, and the damage conditions include typical damage types such as mid-span cracks in the main beam and cable end anchorage relaxation, ensuring the completeness and engineering representativeness of the dataset.

[0080] In step S3, an improved complete ensemble empirical mode decomposition algorithm ICEEMDAN is proposed to decompose the structural dynamic response data to obtain IMF components.

[0081] Among them, the decomposition algorithm ICEEMDAN reduces mode mixing by introducing complementary white noise and adds a dynamic noise adjustment mechanism to achieve significant improvements in signal decomposition accuracy and reconstruction effect.

[0082] It can be understood that through dynamic noise adjustment (β iteration formula) and standard deviation feedback, mode mixing is suppressed, and 8 - 10 order IMF components are decomposed, reducing 63% of the false components compared with traditional methods and improving the signal decomposition purity.

[0083] It should be noted that the initial noise intensity β_0 is set to 20% of the standard deviation of the input signal (β_0 = 0.2σ_x), and the noise injection amount is dynamically adjusted through the residual energy ratio E_k during the decomposition process. Compared with the traditional CEEMDAN, the correlation coefficient of IMF components in this method is increased to 0.92 in a 15% noise environment (0.68 for traditional methods), and the proportion of false modal energy is reduced to 3.7% (9.8% for traditional methods).

[0084] In the embodiment of the present application, the decomposition algorithm ICEEMDA includes: S3.1.1 adding the initial noise level to the original data to obtain the noisy data, and the formula is as follows:

[0085]

[0086] Where: x(t) is the original signal; is the noise level; N 0(t) is the Gaussian white noise sequence;

[0087] S3.1.2: Using EMD to decompose the noisy data to obtain the first residual signal, and the formula is as follows:

[0088] R1 = <M(x i )> (2)

[0089] Where: <·> is the overall average value; M(·) is the local mean value of the input signal;

[0090] S3.1.3: Calculate the error between the first and the original data, and the formula is as follows:

[0091]

[0092] Where: is the standard deviation of the IMF1 component; σ x is the standard deviation of the original data; E1 is the error;

[0093] S3.1.4: According to the error E1, use the non-linear strategy to update the noise level, and the formula is as follows:

[0094] σ1 = σ0 * exp(β * E1) (6)

[0095] Where: is the updated noise level; β is the control coefficient for noise adjustment;

[0096] S3.1.5: Remove the first IMF component from the original signal to obtain the remaining signal, and the formula is as follows:

[0097] x1(t) = x(t) - IMF1(t) (7)

[0098] S3.1.6: Add the updated noise to the remaining signal, and the calculation formula is as follows:

[0099]

[0100] Where: N1(t) is a new group of Gaussian white noise;

[0101] S3.1.7: Perform EMD decomposition on the new noisy signal to extract the second-order intrinsic mode function;

[0102] S3.1.8: Repeat the above calculations to continue adding noise to the remaining signal, perform EMD decomposition, extract subsequent IMF components, and calculate the error and adjust the noise level after each iteration:

[0103] σ n+1 = σ n * exp(β * E n ) (9)

[0104] Where: is the updated noise level; En is the error; β is the control coefficient for noise adjustment.

[0105] It can be understood that by decomposing the error feedback to control the noise intensity, this method actively injects high-intensity noise at the initial stage to separate high-frequency interference components (such as environmental vibration noise). As the number of iterations increases, the noise intensity is reduced exponentially to focus on the mid-low frequency damage characteristics. This strategy reduces the modal aliasing degree of the IMF components to 0.12 (0.54 for the traditional method), and the correlation coefficient is increased to 0.92, ensuring the reliability of subsequent feature screening.

[0106] In step S4, perform Hurst exponent screening on the innovations of the IMF components, and select the target IMF components that satisfy the Hurst exponent H > 0.5 and the reconstruction error MSE < 0.0003.

[0107] Among them, the Hurst exponent is calculated using the detrended fluctuation analysis (DFA). After eliminating the trend term by fitting with a cubic polynomial, a logF(n)-logn linear regression model is established, and 1 / 2 of the slope value is H.

[0108] It can be understood that the Hurst exponent (H > 0.5) screening retains the damage-sensitive components. Combining the constraint of MSE < 0.0003, 90% of the redundant modes are eliminated, the feature dimension is compressed to 3 - 4, 95% of the effective information is retained, and the reconstruction error < 0.03%.

[0109] It should be noted that the Hurst exponent is calculated using detrended fluctuation analysis (DFA), the window length covers 10 - 1000 sampling points, and the trend term is eliminated by fitting with a cubic polynomial. After screening, the damage sensitivity index (DSI) of the target IMF components reaches 8.7, which is 4.3 times higher than that of the un-screened components, effectively enhancing the feature separability.

[0110] In the embodiment of the present application, the Hurst exponent calculation includes: S4.1.1: Preprocess the given time series x(t) to obtain the processed sequence, and the formula is as follows:

[0111]

[0112] Wherein: is the mean of the time series;

[0113] S4.1.2: Use the least squares method for non - linear fitting to obtain the local trend terms of each sequence. The formula is as follows:

[0114]

[0115] Wherein: a n is the fitting polynomial coefficient;

[0116] S4.1.3: Use the local trend term y m (k) to calculate the second - order fluctuation function of the time series. The function is as follows:

[0117]

[0118] F q (n) ∝ n a (13)

[0119] Wherein: a is the Hurst exponent;

[0120] S4.1.4: Calculate the slope of the fitting curve to obtain the Hurst exponent. The formula is as follows:

[0121]

[0122] S4.1.5: By calculating the Hurst exponents of each IMF component, the IMF components can be effectively selected. The MSE calculation formula is:

[0123]

[0124] Wherein: x(t) is the original signal; is the reconstructed signal.

[0125] Among them, the screening conditions are set as follows: the Hurst exponent threshold H > 0.5 (retain the damage - sensitive components with long - term correlation), the reconstructed error threshold MSE < 0.0003 (ensure that the signal reconstruction error is lower than the engineering allowable threshold of 0.03%), and the window length n ∈ [10, 1000] (covering short - term fluctuations and long - term trend analysis).

[0126] It is understandable that the cubic polynomial fitting can eliminate 98.5% of the baseline drift compared with the linear / quadratic fitting, and the window length multi-scale analysis ensures the robustness of the H value calculation (standard deviation < 0.03). The damage sensitivity index (DSI = 8.7) of the screened target IMF components is 4.3 times higher than that of the unscreened components (DSI = 2.0), effectively enhancing the feature separability.

[0127] In step S5, a one-dimensional convolutional neural network model is constructed and trained. The target IMF components are input into the trained one-dimensional convolutional neural network model, and the recognition results of the damage location and damage degree are output.

[0128] Among them, the one-dimensional convolutional neural network model includes: a leading convolutional layer group: two-channel 16 kernels (size 8, stride 2), an intermediate feature layer group: 64 kernels (size 4) + a max pooling layer, a deep abstraction layer group: 256 kernels (size 2) + a global average pooling layer, and a classification output layer: a 10-neuron fully connected layer activated by Softmax.

[0129] It is understandable that the stress time history (sampled at 200 Hz) covering 15 measuring points (spacing 2 m) of the main girder and the cable force spectrum data are used to generate 4,752,000 groups of samples with 5%-15% noise, simulating the data characteristics under actual environmental interference.

[0130] It should be noted that the data acquisition module integrates a temperature compensation algorithm to eliminate the influence of environmental temperature drift on the strain signal (temperature drift error < 0.5 με / °C). The white noise excitation bandwidth covers the main vibration frequency band of the bridge from 0.1 to 50 Hz, and the damage conditions include typical damage types such as mid-span cracks in the main girder and cable end anchorage relaxation, ensuring the completeness and engineering representativeness of the data set.

[0131] In the embodiment of the present application, the one-dimensional convolutional neural network model includes: S5.1.1: The input layer receives a 6000-dimensional time series signal; S5.1.2: The first convolutional layer uses 16 convolutional kernels with a size of 8 and a stride of 2; S5.1.3: The second pooling layer uses max pooling with a pooling size of 2; S5.1.4: The deep convolutional layer uses 512 convolutional kernels with a size of 2; S5.1.5: The global average pooling layer is connected to a 512-node fully connected layer; S5.1.6: The Dropout layer sets a dropout rate of 0.3; S5.1.7: The output layer uses a Softmax activation function.

[0132] Among them, the network optimization strategy includes: a decreasing convolutional kernel size design (8→4→2) to achieve multi-scale feature abstraction, replacing the fully connected layer with a global average pooling layer to reduce the number of parameters by 83%, and a Dropout rate of 0.3 to suppress overfitting (the validation set loss fluctuation < 0.5%).

[0133] It is understandable that the deep small kernel (size 2) focuses on the weak damage signal under high-frequency noise (such as the 5με-level strain change with a damage factor of 0.05), and cooperates with the lightweight design (parameter quantity 1.2M) to achieve real-time diagnosis at the 50ms level (220ms for traditional methods), meeting the embedded deployment requirements for bridge monitoring.

[0134] In the embodiment of the present application, before using an improved complete adaptive noise ensemble empirical mode decomposition algorithm ICEEMDAN to decompose the structural dynamic response data to obtain IMF components, it further includes: S3.2.1: Adding Gaussian white noise to the original stress response signal x(t) to construct a noisy signal:

[0135] x_noise(t) = x(t) + ησ_xξ(t) (16)

[0136] where, σ_x is the signal standard deviation, η ∈ {5%, 10%, 15%} is the noise intensity coefficient, and ξ(t) is a standard normal distribution random number.

[0137] It is understandable that this coding strategy maps multi-dimensional information (position + degree) through single labels, reducing the network parameter quantity by 40% (10 vs 18 output nodes), and supporting the diagnosis of 3 compound damages (such as the combination of measuring points 3 + 8 + 11), with a quantitative error ≤ 2.94% (≥ 8.2% for traditional methods) and a misjudgment rate < 0.5%.

[0138] In the embodiment of the present application, multi-label classification strategy is adopted for damage quantitative analysis, including: S5.2.1: Encoding the non-damaged state as category 0, corresponding to the non-damaged working condition; S5.2.2: Encoding single-damage working conditions according to the damage factor levels: the damage factors of 0.05 / 0.10 / 0.15 at measuring point 3 correspond to categories 1 - 3 respectively, and the same-level damage factors at measuring point 8 correspond to categories 4 - 6; S5.2.3: Encoding rule for multi-damage working conditions: when there is a combined damage at measuring points 3 and 11, the damage factors of 0.05 / 0.10 / 0.15 correspond to categories 7 - 9 respectively; S5.2.4: The output layer of the one-dimensional convolutional neural network is set with 10 neurons, corresponding to the probability distributions of damage working conditions of classes 0 - 9 respectively.

[0139] Among them, the basis for setting the noise parameters is: the value of η covers the statistical range of the actual bridge environmental noise (mean 10.3% ± 4.7%), and σ_x is dynamically calculated through a sliding window (length 1s, step size 0.1s) to achieve adaptive adjustment of the noise intensity.

[0140] It is understandable that the dynamic standard deviation calculation adds weak noise (η = 5%) to the high-energy section (σ_x > 10 MPa) to avoid masking of damage features; strong noise (η = 15%) is added to the low-energy section (σ_x < 5 MPa) to simulate the signal attenuation condition at the far end of the sensor. This strategy controls the damage signal attenuation rate within 3% (the traditional fixed threshold method results in 8 - 12% attenuation), and the noise residue ≤ 12% (the traditional method ≥ 20%).

[0141] The embodiment of the present application provides a damage identification method for cable-stayed bridges based on massive data, which can significantly improve the damage identification accuracy and calculation efficiency in a complex noise environment. Through a dynamic noise suppression mechanism, damage-sensitive feature screening, lightweight convolutional network design, and multi-damage collaborative diagnosis optimization, this method realizes enhanced anti-interference ability and breakthroughs in the overall process performance of damage identification. By using an improved ICEEMDAN algorithm, a dynamic update model of the noise level parameter β_k (β_{k + 1} = β_k·exp(-γE_k)) is established, and combined with a standard deviation feedback mechanism, the damage location accuracy remains 98.59% under 15% Gaussian noise interference (the traditional method ≤ 58.44%). The IMF components are screened using a double threshold of the Hurst index (H > 0.5) and the reconstruction error (MSE < 0.0003), eliminating 90% of the redundant modes, with a feature dimension compression ratio of 8:1 and retaining more than 95% of the effective damage information. A 1D-CNN model with a depthwise separable convolution architecture is constructed, with the number of parameters reduced to 1.2M (5.7M for traditional CNN), supporting real-time diagnosis at the 50ms level, with a single damage location error ≤ 0.5m (the measurement point spacing is 2m) and a composite damage quantification error ≤ 2.94%. Thus, problems such as strong noise sensitivity, high feature redundancy, insufficient multi-damage identification, and large resource consumption are solved.

[0142] The damage identification method for cable-stayed bridges based on massive data will be elaborated through a specific embodiment below, as Figure 2 shown, including:[[]]

[0143] Step 1: Construct a reference finite element model of the cable-stayed bridge.

[0144] As Figure 3 shown, the reference finite element model is established using ABAQUS software. The main girder uses C3D8R solid elements (mesh size 0.5m), the stay cables use T3D2 truss elements, and fixed constraint conditions are set at the bottom of the cable tower. Figure 4 As Figure 5 shown, the damage areas cover the mid-span of the main girder (measurement point 3), the 1 / 4 span (measurement point 8), and the cable end anchorage area (measurement point 11), with a measurement point spacing of 2m.

[0145] Specifically, taking a single-pylon single-span cable-stayed bridge as an example, the pylon is of A-shaped vase type, and the stay cables are parallel wire cables, with 11 pairs on each side. The ABAQUS software is used to establish the benchmark finite element model of the cable-stayed bridge. The main girder and the pylon adopt three-dimensional eight-node linear hexahedron elements; the prestressed steel bars in the main girder and the stay cables adopt three-dimensional two-node truss elements. The bottom of the pylon is completely fixed, and vertical and lateral constraints are applied at both ends of the cable-stayed bridge; the main girder and the main pylon are connected by tie; the stay cables, prestressed steel bars are connected to the main girder and the pylon by embedded region connection. The benchmark finite element model is established by the model updating method. The environmental excitation with the same time as the analysis step increment is applied to the main girder, and the transient dynamic analysis is carried out.

[0146] Step 2: Collect the structural dynamic response data under various damage conditions according to the benchmark finite element model.

[0147] As Figure 6 shown, the data acquisition module includes a distributed fiber optic sensor array (sampling rate 200 Hz), covering 15 measuring points (spacing 2 m) on the main girder and the cable anchorage ends. The dynamic response data includes: stress time history signal: recording the dynamic strain response of each measuring point on the main girder (accuracy ±0.5 με); cable force vibration spectrum: analyzing the frequency shift of the cable force (resolution 0.1 Hz); environmental noise data: collecting 5%-15% Gaussian white noise to simulate the actual monitoring environmental interference. As Figure 7 shown, at measuring point 3 under the damage factor of 0.15, the stress peak value drops from 15.2 MPa to 12.9 MPa (a decrease of 14.8%), and the damage characteristics are significant. 4,752,000 groups of samples are generated by finite element simulation, covering 10 types of conditions including single damage, compound damage and non-damaged state. The data division is shown in Table 1.

[0148] Table 1 Condition Results

[0149]

[0150]

[0151] Among them, the damage types and damage factors of each condition are shown in Table 2.

[0152] Table 2 Condition Damage Types and Damage Factors

[0153]

[0154] Specifically, for the 0.10 damage factor condition of the main girder measurement point 3 (mid-span): Signal acquisition: The fiber optic sensor records the stress time history signal (duration 30 s, sampling rate 200 Hz), and 10% Gaussian noise is added; Feature extraction: The cable force vibration spectrum shows that the first-order frequency drops from 2.35 Hz to 2.28 Hz (offset 0.07 Hz); Dataset construction: 120,000 groups of samples for this condition are generated, including variants with noise intensities of 5%-15%.

[0155] Step 3: Use the improved ICEEMDAN algorithm to decompose the dynamic response data to obtain IMF components.

[0156] As Figure 8 shown, the improved ICEEMDAN decomposition process includes: adding the initial noise level to the original data to obtain the noisy data formula as follows:

[0157]

[0158] where: x(t) is the original signal; is the noise level; N0(t) is the Gaussian white noise sequence; Use EMD to decompose the noisy data to obtain the first residual signal:

[0159] R1 = <M(x i )

[0160] The modal component formula is as follows:

[0161] IMF1(t) = x(t) - R1.

[0162] where: <·> is the overall average value; M(·) is the local mean value of the input signal; Calculate the error between the first IMF1(t) and the original data x(t), and the formula is as follows:

[0163]

[0164] where: is the standard deviation of the IMF1 component; σ x is the standard deviation of the original data; E1 is the error; According to the error E1, use the nonlinear strategy to update the noise level formula as follows: ò1 = ò0 * exp(β * E1)

[0166] where: is the updated noise level; β is the control coefficient for noise adjustment; Remove the first IMF component from the original signal to obtain the remaining signal formula as follows:

[0167] x1(t) = x(t) - IMF1(t)

[0168] Add the updated noise to the remaining signal x1(t). The formula is as follows:

[0169]

[0170] Where: N1(t) is a new set of Gaussian white noise; for the new noisy signal Perform EMD decomposition, extract the second-order intrinsic mode function; repeat the above calculation to continue adding noise to the remaining signal, perform EMD decomposition, extract the subsequent IMF components, and calculate the error and adjust the noise level after each round of iteration: o` n+1 = o` n * exp(β * E n )

[0172] Where: is the updated noise level; En is the error; β is the control coefficient for noise adjustment; subtract the latest IMF component from the signal after each decomposition, and add the updated noise to the remaining signal. Decomposition and noise adjustment are continuously performed during the iteration until the error En is lower than the preset error tolerance; the results of decomposing the cable-stayed bridge monitoring data by the ICEEMDAN algorithm. By analyzing different IMF components, the key components containing damage information can be effectively screened out for subsequent CNN model training. The reconstruction errors of the IMF components are shown in Table 3.

[0173] Table 3 Reconstruction Errors of IMF Components

[0174]

[0175]

[0176] Specifically, decompose the noisy signal (σ_x = 15.2 MPa) at measuring point 3: Initial noise: β0 = 0.2 × 15.2 = 3.04; Iteration process: First iteration: E1 = 0.15 → β1 = 3.04 × exp(-0.3 × 0.15) = 2.89 Third iteration: E3 = 0.03 → β3 = 2.41 × exp(-0.3 × 0.03) = 2.38. Output result: 9-order IMF components are obtained by decomposition, and the correlation coefficients of IMF3 - IMF5 are > 0.90, and the proportion of spurious mode energy is 3.5%.

[0177] Step Four: Perform Hurst index screening on the innovations of the IMF components to select the target components.

[0178] As Figure 9 shown, to further verify the noise robustness of the method proposed in this paper, as shown in Table 4

[0179] Table 4 Noise Level

[0180]

[0181] Gaussian white noise with levels of 5%, 10%, and 15% respectively is added during the extraction of the main girder stress data. The calculation formula for the noisy stress response is as follows:

[0182] s k = s k,u + s k,u σχ

[0183] Where: s k is the stress response signal with added noise; s k,u is the stress response signal without noise; σ is the standard normal distribution with a mean of 0 and a variance of 1; χ is the intensity of the Gaussian white noise level. By calculating the Hurst exponent of each IMF component, the screening criterion is based on the Hurst exponent threshold. The formula derivation process includes: preprocessing the given time series to obtain the processed series, and its formula is as follows:

[0184]

[0185] Where: is the mean of the time series, and x(t) is the time series; the local trend term of each series is obtained by non-linear fitting using the least squares method, and its formula is as follows:

[0186]

[0187] Where: a n is the fitting polynomial coefficient; the second-order fluctuation function of the time series is calculated using the local trend term, and its function is as follows:

[0188]

[0189] F q (n) ∝ n a

[0190] Where: α is the Hurst exponent, and y m (k) is the local trend term; by changing the sub-interval length of formula (10), the fluctuation function is derived from formulas (11) to (13), and the slope of the fitting curve is calculated to obtain the Hurst exponent. The exponent formula is as follows:

[0191]

[0192] Among them, the smaller the value of a, the greater the signal fluctuation and the rougher the signal. When 0 < a < 0.5, the time series is negatively correlated and the noise proportion is large; when a = 0.5, the signal is uncorrelated; when 0.5 < α < 1, the time series is positively correlated and the proportion of useful information is large. Therefore, by calculating the Hurst exponent of each IMF component, the IMF components can be effectively selected; the components with 0.5 < α < 1 are used to quantify the reconstruction error using the mean square error (MSE). And the optimal unique component is selected. The MSE calculation formula is as follows:

[0193]

[0194] where: x(t) is the original signal; is the reconstructed signal;

[0195] Among them, the model performances are compared as shown in Tables 5 and 6. The lower the MSE, the greater the contribution of the IMF component to the signal reconstruction and the better the quality, which is the target feature.

[0196] Table 5 Comparison of model performances

[0197]

[0198] Table 6 Model scores

[0199]

[0200] Specifically, the quantitative damage analysis of the IMF3 component (H = 0.68, MSE = 0.00018) is shown in Table 7

[0201] Table 7 Quantitative damage analysis

[0202]

[0203] The specific analysis includes: Hurst calculation: when the window n = 500, logF(n) = 1.82, logn = 2.70 → slope = 0.67 → H = 0.335; trend elimination: the cubic polynomial fitting R 2 = 0.992, and the baseline drift is reduced by 98.7%; screening result: retain IMF3 (H = 0.68 > 0.5, MSE = 0.00018 < 0.0003), and its DSI = 9.1, which is 4.6 times higher than that of the un-screened component.

[0204] Step 5: Construct a one-dimensional convolutional neural network model to output the damage recognition result.

[0205] Among them, the architecture of the target model is shown in Table 8,

[0206] Table 8 Architecture of the target model

[0207]

[0208] To verify the effectiveness of the model in actual engineering, this study selected a cable-stayed bridge in the operation period for on-site testing. As shown in Table 9, the training conditions covered the key vulnerable areas (measurement points 3 and 8) of the main girder and typical damage degrees (0.05 - 0.15) to ensure the sensitivity of the model to actual damage modes.

[0209] Table 9 Test Results of the Cable-Stayed Bridge in the [Period]

[0210]

[0211]

[0212] Distributed optical fiber sensors (sampling rate 200 Hz) were arranged on the main girder of the actual bridge to collect stress time history data for 300 s during the operation period. After the measured data were decomposed by ICEEMDAN and screened by the Hurst index, they were input into the trained 1D-CNN model for diagnosis. As shown in Table 10, all four groups of measured signals were determined to be undamaged (label 0), which was consistent with the conclusion of manual detection.

[0213] Table 10 Results of Measured Signals

[0214]

[0215] Specifically, the sample with a damage factor of 0.15 at measurement point 3 was diagnosed: input data: 6000-dimensional time series signal of the IMF3 component; network inference: the first-layer convolution output a 16-channel feature map (size 2996×16); after global average pooling, a 512-dimensional feature vector was generated; the Softmax layer output the probability distribution: the probability of class 3 (measurement point 3 - 0.15 damage) was 99.3%; result verification: the error of the actual damage position was 0.32 m (<0.5 m), and the calculated value of the damage factor was 0.147 (error 1.93%).

[0216] In summary, the present invention realizes a set of intelligent methods for identifying cable-stayed bridge damage with high precision and strong robustness. By means of the improved ICEEMDAN algorithm, dynamic noise suppression is achieved. Combining the Hurst exponent feature screening mechanism with the lightweight one-dimensional convolutional neural network, the identification bottleneck of traditional methods in complex noise environments is overcome. Through the standard deviation feedback mechanism, the noise injection intensity is dynamically adjusted (β_{k + 1}=β_k·exp(-γE_k)). Under 15% Gaussian noise interference, high-purity IMF components can still be extracted, the modal aliasing degree is reduced to 0.12 (0.54 for traditional methods), and the IMF correlation coefficient is increased to 0.92. The double-threshold screening mechanism of the Hurst exponent (H>0.5) and the reconstruction error (MSE<0.0003) is adopted, the feature dimension compression ratio reaches 8:1, 95% of the effective damage information is retained, and the damage sensitivity index (DSI) is increased by 4.3 times. A 1D-CNN model with a depthwise separable convolution architecture is constructed, the number of parameters is compressed to 1.2M (5.7M for traditional CNN), the inference speed reaches 50 ms / time, it supports deployment on embedded devices, the single damage location error ≤ 0.5 m, and the compound damage quantification error ≤ 2.94%. Based on the noise intensity η (5%-15%), the compensation factor γ (0.1-0.5) is dynamically adjusted, and the network weights are updated through backpropagation to ensure the robustness of the system in complex environments such as temperature drift (error < 0.5 με / °C) and vibration frequency band offset (0.1-50 Hz). This solution provides a full-link solution from signal processing to intelligent diagnosis for the health monitoring of cable-stayed bridges, significantly improving the structural safety warning ability and operation and maintenance efficiency, and solving the technical bottlenecks of traditional methods in terms of noise sensitivity, feature redundancy, multi-damage analysis, and real-time performance.

[0217] Next, a device for identifying cable-stayed bridge mass data damage according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0218] Figure 10 It is a block diagram of a device for identifying cable-stayed bridge mass data damage according to an embodiment of the present application.

[0219] As Figure 10 shown, the device 10 for identifying cable-stayed bridge mass data damage includes: a modeling module 100, a data acquisition module 200, a signal decomposition module 300, a feature screening module 400, and a network construction module 500.

[0220] Among them, the modeling module 100 is used to construct a benchmark finite element model of the cable-stayed bridge; the data acquisition module 200 is used to collect structural dynamic response data under various damage conditions according to the benchmark finite element model; the signal decomposition module 300 is used to propose an improved complete adaptive noise ensemble empirical mode decomposition algorithm ICEEMDAN to decompose the structural dynamic response data to obtain IMF components. Among them, the decomposition algorithm ICEEMDAN is reflected in dynamically adjusting the noise level through a standard deviation feedback mechanism; the feature screening module 400 is used to innovatively screen the IMF components by the Hurst index, and select the target IMF components that satisfy the Hurst index H>0.5 and the reconstruction error MSE<0.0003; the network construction module 500 is used to construct and train a one-dimensional convolutional neural network model, input the target IMF components into the trained one-dimensional convolutional neural network model, and output the recognition results of the damage location and damage degree. Among them, the one-dimensional convolutional neural network model includes: a leading convolutional layer group, an intermediate feature layer group, a deep abstraction layer group, and a classification output layer.

[0221] It should be noted that the foregoing explanation of the embodiments of the cable-stayed bridge damage identification method based on massive data also applies to the cable-stayed bridge massive data damage identification device of this embodiment, and will not be repeated here.

[0222] The massive data damage identification device for cable-stayed bridges proposed according to the embodiments of the present application realizes high-precision real-time identification of cable-stayed bridge damage in complex noise environments through dynamic noise suppression, damage-sensitive feature screening, and lightweight convolutional network design. Based on the intelligent analysis of structural dynamic response data, the device overcomes the technical bottlenecks of modal aliasing, feature redundancy, and low computational efficiency of traditional methods under noise interference. By integrating and improving the ICEEMDAN algorithm, the noise injection intensity is dynamically adjusted through a standard deviation feedback mechanism (β_{k+1} = β_k·exp(-γE_k), γ = 0.3). High-purity IMF components can still be extracted under 15% Gaussian noise, the modal aliasing degree is reduced to 0.12 (traditional method ≥ 0.54), and the correlation coefficient of IMF components is increased to 0.92. Adopting a double-threshold screening mechanism of Hurst index (H > 0.5) and reconstruction error (MSE < 0.0003), the feature dimension compression ratio reaches 8:1, 95% of the effective damage information is retained, and the damage sensitivity index (DSI) is increased by 4.3 times. A 1D-CNN model with a depthwise separable convolution architecture is constructed, and the number of parameters is compressed to 1.2M (5.7M for traditional CNN), the inference speed reaches 50 ms / time, and it supports deployment on embedded devices. The single damage location error ≤ 0.5 m (measurement point spacing 2 m), and the compound damage quantification error ≤ 2.94%. Based on the real-time noise intensity η (5% - 15%), the compensation factor γ (0.1 - 0.5) is dynamically adjusted, and the network weights are updated through backpropagation to ensure the robustness of the system in complex environments such as temperature drift (error < 0.5 με / °C) and vibration frequency band offset (0.1 - 50 Hz).

[0223] Figure 11 The structural schematic diagram of the electronic device provided by the embodiment of the present application. The electronic device may include:

[0224] A memory 1101, a processor 1102, and a computer program stored on the memory 1101 and executable on the processor 1102.

[0225] When the processor 1102 executes the program, it implements the cable-stayed bridge damage identification method based on massive data provided in the above embodiments.

[0226] Further, the electronic device further includes:

[0227] A communication interface 1103 for communication between the memory 1101 and the processor 1102.

[0228] The memory 1101 is used to store a computer program executable on the processor 1102.

[0229] The memory 1101 may include a high-speed RAM (Random Access Memory) memory, and may also include a non-volatile memory, such as at least one disk memory.

[0230] If the memory 1101, the processor 1102, and the communication interface 1103 are implemented independently, the communication interface 1103, the memory 1101, and the processor 1102 can be interconnected through a bus and communicate with each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component) bus, an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of simplicity of representation, Figure 8 only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0231] Optionally, in a specific implementation, if the memory 1101, the processor 1102, and the communication interface 1103 are integrated on a single chip, the memory 1101, the processor 1102, and the communication interface 1103 can communicate with each other through an internal interface.

[0232] The processor 1102 may be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0233] The embodiments of the present application also provide a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned method for damage identification of a cable-stayed bridge based on massive data is implemented.

[0234] In addition, the embodiments of the present application also provide a computer program product, including a computer program or instruction, and when the computer program or instruction is executed, the above-mentioned method for damage identification of a cable-stayed bridge based on massive data is implemented.

[0235] In the description of this specification, the descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0236] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of these features. In the description of this application, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0237] Any process or method description shown in a flowchart or described in other ways herein may be understood to represent a module, segment, or portion of code including one or more executable instructions for implementing a customized logic function or process. And the scope of the preferred embodiments of this application includes additional implementations, where the functions may be executed in a substantially simultaneous manner or in a reverse order according to the involved functions, rather than in the order shown or discussed, which should be understood by those skilled in the art to which the embodiments of this application pertain.

[0238] It should be understood that each part of this application can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one of the following well-known technologies in the art or a combination thereof can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0239] Those of ordinary skill in the art of this technology can understand that all or part of the steps carried by the method of implementing the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0240] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A damage identification method for cable-stayed bridges based on massive data, characterized in that, The content of the method includes: S1: Construct a benchmark finite element model of the cable-stayed bridge; S2: Collect structural dynamic response data under various damage conditions according to the benchmark finite element model; S3: An improved complete adaptive noise ensemble empirical mode decomposition algorithm ICEEMDAN is proposed to decompose the structural dynamic response data to obtain IMF components. Among them, the decomposition algorithm ICEEMDAN dynamically adjusts the noise level through a standard deviation feedback mechanism; S4: Innovatively perform Hurst index screening on the IMF components, and select the target IMF components that satisfy the Hurst index H>0.5 and the reconstruction error MSE<0.0003; S5: Construct and train a one-dimensional convolutional neural network model, input the target IMF components into the trained one-dimensional convolutional neural network model, and output the recognition results of the damage location and damage degree. Among them, the one-dimensional convolutional neural network model includes a leading convolutional layer group, an intermediate feature layer group, a deep abstraction layer group, and a classification output layer.

2. The method for damage identification of a cable-stayed bridge based on massive data according to claim 1, wherein The decomposition algorithm ICEEMDAN includes: S3.1.1 Add the initial noise level to the original data to obtain noisy data, and the formula is as follows: x0(t) = x(t) + ò0·N0(t) (1) Where: x(t) is the original signal; ò0 is the noise level; N0(t) is the Gaussian white noise sequence; S3.1.2: Use EMD to decompose the noisy data to obtain the first residual signal, and the formula is as follows: R1 = <M(x i )> (2) Where: <·> is the overall average value; M(·) is the local mean value of the input signal; S3.1.3: Calculate the error between the first and the original data, and the formula is as follows: Wherein: is the standard deviation of the IMF1 component; σ x is the standard deviation of the original data; E1 is the error; S3.1.4: According to the error E1, use a non-linear strategy to update the noise level, and the formula is as follows: σ1 = σ0*exp(β*E1) (6) Where: ò1 is the updated noise level; β is the control coefficient for noise adjustment; S3.1.5: Remove the first IMF component from the original signal to obtain the remaining signal, and the formula is as follows: x1(t) = x(t) - IMF1(t) (7) S3.1.6: Add the updated noise to the remaining signal, and the calculation formula is as follows: Where: N1(t) is a new set of Gaussian white noise; S3.1.7: Perform EMD decomposition on the new noisy signal to extract the second-order intrinsic mode function; S3.1.8: Repeat the above calculations to continue adding noise to the remaining signal, perform EMD decomposition, extract the subsequent IMF components, and calculate the error and adjust the noise level after each round of iteration: σ n+1 = σ n * exp(β * E n ) (9) Where: ò n+1 is the updated noise level; En is the error; β is the control coefficient for noise adjustment.

3. The method for damage identification of a cable-stayed bridge based on mass data according to claim 1, characterized in that, The Hurst index calculation includes: S4.1.1: Preprocess the given time series x(t) to obtain the processed series, and the formula is as follows: Wherein: is the mean of the time series; S4.1.2: Use the least squares method for non-linear fitting to obtain the local trend terms of each series, and the formula is as follows: where: a n is the fitting polynomial coefficient; S4.1.3: Utilize the local trend term y m (k) Calculate the second-order volatility function of the time series, and the function is as follows: F q (n) is proportional to n a (13) Where: α is the Hurst index; S4.1.4: Calculate the slope of the fitting curve to obtain the Hurst index, and the formula is as follows: S4.1.5: The IMF components can be effectively selected by calculating the Hurst exponents of each IMF component. The MSE calculation formula is as follows: Where: x(t) is the original signal; is the reconstructed signal.

4. The method for damage identification of a cable-stayed bridge based on massive data according to claim 1, wherein The one-dimensional convolutional neural network model includes: S5.1.1: The input layer receives a 6000-dimensional time series signal; S5.1.2: The first convolutional layer uses 16 convolutional kernels with a size of 8 and a stride of 2; S5.1.3: The second pooling layer uses max pooling with a pooling size of 2; S5.1.4: The deep convolutional layer uses 512 convolutional kernels with a size of 2; S5.1.5: The global average pooling layer is connected to a 512-node fully connected layer; S5.1.6: The Dropout layer sets a dropout rate of 0.3; S5.1.7: The output layer uses the Softmax activation function.

5. The method for damage identification of a cable-stayed bridge based on massive data according to claim 1, characterized in that, Before decomposing the structural dynamic response data by using an improved complete ensemble empirical mode decomposition algorithm with adaptive noise (ICEEMDAN) to obtain IMF components, it also includes: S3.2.1: Adding Gaussian white noise to the original stress response signal x(t) to construct a noisy signal: x_noise(t) = x(t) + ησ_xξ(t) (16) Where, σ_x is the signal standard deviation, η ∈ {5%, 10%, 15%} is the noise intensity coefficient, and ξ(t) is a standard normal distribution random number.

6. The method for damage identification of a cable-stayed bridge based on massive data according to claim 1, wherein The multiple damage conditions adopt a multi-label classification strategy. Among them, the multi-label classification strategy includes non-destructive state coding, single damage conditions, and multi-damage conditions, including: S5.2.1: Coding the non-destructive state as category 0, corresponding to the non-damage condition; S5.2.2: The single damage conditions are coded by grading according to the damage factor: the damage factors of 0.05 / 0.10 / 0.15 at measurement point 3 correspond to categories 1-3 respectively, and the same-level damage factors at measurement point 8 correspond to categories 4-6; S5.2.3: Coding rule for multi-damage conditions: When there is a combined damage at measurement points 3 and 11, the damage factors of 0.05 / 0.10 / 0.15 correspond to categories 7-9 respectively; S5.2.4: The output layer of the one-dimensional convolutional neural network is set with 10 neurons, corresponding to the probability distributions of damage conditions of classes 0-9 respectively.

7. An apparatus for damage identification of a cable-stayed bridge based on massive data, characterized in that, It includes: A modeling module, S1: Constructing a benchmark finite element model of a cable-stayed bridge; A data acquisition module, S2: Collecting structural dynamic response data under multiple damage conditions according to the benchmark finite element model; A signal decomposition module, S3: An improved complete ensemble empirical mode decomposition algorithm with adaptive noise (ICEEMDAN) is proposed to decompose the structural dynamic response data to obtain IMF components. Among them, the decomposition algorithm ICEEMDAN is reflected in dynamically adjusting the noise level through a standard deviation feedback mechanism; A feature screening module, S4: Innovatively screening the Hurst exponents of the IMF components, and selecting target IMF components that satisfy the Hurst exponent H > 0.5 and the reconstruction error MSE < 0.0003; Network construction module, S5: Construct and train a one-dimensional convolutional neural network model, input the target IMF component into the trained one-dimensional convolutional neural network model, and output the recognition results of the damage location and damage degree. Among them, the one-dimensional convolutional neural network model includes a leading convolutional layer group, an intermediate feature layer group, a deep abstraction layer group, and a classification output layer.

8. An electronic device, characterized in that, Comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the program to implement the cable-stayed bridge damage recognition method based on massive data according to any one of claims 1-6.

9. A computer-readable storage medium storing a computer program or instructions thereon, characterized in that, When the computer program or instruction is executed, it implements the cable-stayed bridge damage recognition method based on massive data according to any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instruction is executed, it implements the cable-stayed bridge damage recognition method based on massive data according to any one of claims 1-6.

Images