Structural damage identification method and device based on decision fusion, medium and product

Through a decision-making fusion method, combined with the spectral Markov transfer field and ResNet-34 model, the dynamic weight allocation mechanism is solved, and the problems of low accuracy of single sensors and reduced fusion accuracy in traditional methods are improved, and the identification of structural damage is improved. Accuracy and adaptability.

CN119027740APending Publication Date: 2024-11-26CHANGSHA UNIVERSITY OF SCIENCE AND TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411234127.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Traditional structural damage recognition methods rely on a single sensor to cause low accuracy of recognition results, and traditional multi-sensor data fusion methods are difficult to fully utilize the advantages of different sensors, resulting in a decrease in recognition accuracy, especially in complex operating conditions, which is difficult to achieve accurate diagnosis.

Method used

Using a decision fusion-based approach, multiple sample data sets and damage recognition models are constructed, and the acceleration response signal is converted into structural images using the spectral Markov transfer field, and trained using the pre-trained ResNet-34 model, combined with dynamic weights The allocation mechanism performs damage identification to ensure that excellent performing models occupy a more important position in the decision-making process.

Benefits of technology

Improve the accuracy and adaptability of structural damage recognition, better capture the time-dependent and nonlinear characteristics of the signal, process complex spectrum characteristics, enhance adaptability to different application scenarios, solve the problem of low accuracy and low accuracy of a single sensor The problem of degradation of accuracy of traditional fusion methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027740A_ABST
    Figure CN119027740A_ABST
Patent Text Reader

Abstract

The invention discloses a structure damage identification method and device based on decision fusion, a medium and a product, and the method comprises the steps: constructing a plurality of sample data sets, and dividing each sample data set into a training set and a verification set; loading a plurality of damage identification models; training the damage identification model by using the training set to obtain a target model; performing damage identification on the verification set by using the target model; calculating the recognition accuracy of the target model on the verification set according to the damage recognition result; calculating the weight of the target model according to the recognition accuracy; performing damage identification on the to-be-identified structure image by using each target model; and calculating the score of each damage category according to the damage identification result and the weight of each target model, and taking the damage category corresponding to the highest score as the final damage category of the to-be-identified structural image. The method solves the problem that the traditional fusion method is difficult to make full use of the advantages of different sensors so that the recognition precision is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of structural damage identification, and in particular relates to a structural damage identification method, device, storage medium and product based on decision fusion. Background Art

[0002] In the field of modern structural health monitoring (SHM) and fault diagnosis, especially when it comes to the health monitoring of complex structures such as cable-stayed bridges, the collection and analysis of vibration signals is crucial. With the widespread application of multi-sensor technology, how to effectively fuse multi-dimensional sensor data has become a key challenge to improve monitoring accuracy.

[0003] Traditional diagnostic methods usually rely on a single sensor signal. However, due to environmental noise, sensor failure and other reasons, the data from a single sensor may not be reliable enough, which in turn affects the accuracy of the diagnostic results.

[0004] In multi-sensor data fusion, fixed weighted or average voting decision fusion methods are usually used. These fusion methods are difficult to fully utilize the advantages of different sensors, especially when some sensors perform well, these fusion methods cannot effectively highlight their contributions, resulting in a decrease in diagnostic accuracy.

[0005] The above methods often show great limitations when facing complex and changing working environments, and it is difficult to accurately diagnose a variety of complex working conditions. How to effectively fuse data from multiple sensors and accurately identify structural damage remains a challenge. Summary of the invention

[0006] The purpose of the present invention is to provide a structural damage identification method, device, medium and product based on decision fusion to solve at least one of the problems that the traditional method relies on a single sensor, resulting in low accuracy of identification results, and the traditional fusion method is difficult to fully utilize the advantages of different sensors, resulting in reduced identification accuracy.

[0007] The present invention solves the above technical problems through the following technical solutions: a structural damage identification method based on decision fusion, comprising:

[0008] Constructing multiple sample data sets, and dividing each of the sample data sets into a training set and a validation set; wherein each sample data set corresponds to a sensor for collecting an acceleration response signal of a structure, and each sample in each sample data set includes a structural image and its damage category, and the structural image is an image generated according to the acceleration response signal of the structure collected by the corresponding sensor;

[0009] Load multiple damage identification models; each damage identification model corresponds to a sample data set;

[0010] Using the training set of each sample data set to train the corresponding damage recognition model to obtain a target model; using the target model to perform damage recognition on the validation set of the sample data set; and calculating the recognition accuracy of the target model on the validation set of the sample data set based on the damage recognition result;

[0011] The weight of the corresponding target model is calculated according to the recognition accuracy of each target model on the validation set of the corresponding sample data set;

[0012] For a structure image to be identified, using each target model to perform damage identification on the structure image to be identified;

[0013] The score of each damage category is calculated according to the damage recognition result of each target model on the structural image to be identified and the weight of each target model, and the damage category corresponding to the highest score is used as the final damage category of the structural image to be identified.

[0014] Furthermore, according to the acceleration response signal of the structure collected by the sensor, a spectral Markov transfer field is used to generate a structural image, including:

[0015] The frequency spectrum signal of the acceleration response signal is calculated using the following formula:

[0016]

[0017] Wherein, s(k) represents the kth component in the spectrum signal, x(i) represents the value of the i-th sampling point in the acceleration response signal, n represents the number of sampling points of the acceleration response signal, abs() represents the modulus value function, and j represents the imaginary unit;

[0018] Performing normalization processing on each component in the spectrum signal to obtain a normalized component of the spectrum signal;

[0019] The value range of the amplitude set is evenly divided to obtain Q sub-intervals; wherein the amplitude set includes the amplitudes of all normalized components of the spectrum signal;

[0020] Each amplitude in the amplitude set is mapped to a subinterval; wherein the specific mapping formula is:

[0021]

[0022] Among them, A k represents the kth amplitude in the amplitude set, q j represents the jth subinterval, argmin j It means finding j that makes the function reach the minimum value;

[0023] Generate a spectral Markov transition field matrix according to the probability of each amplitude being transferred from the current subinterval to other subintervals;

[0024] The spectral Markov transfer field matrix is ​​converted into a two-dimensional image to obtain a structural image.

[0025] Furthermore, the damage identification model selects a pre-trained ResNet-34 model, and the ResNet-34 model includes an input layer, a convolutional layer, a maximum pooling layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, an average pooling layer and a fully connected layer connected in sequence.

[0026] Furthermore, the convolution layer includes 64 convolution kernels of size 7×7 and a step size of 2; the pooling kernel size of the maximum pooling layer is 3×3 and a step size of 2; the first residual layer includes 3 residual blocks, the second residual layer includes 4 residual blocks, the third residual layer includes 6 residual blocks, and the fourth residual layer includes 3 residual blocks, each residual block includes 2 3×3 convolution kernels; the pooling kernel size of the average pooling layer is 7×7.

[0027] Furthermore, the corresponding damage identification model is trained using the training set of each sample data set, including:

[0028] Freeze the convolutional layer, the first residual layer, and the second residual layer of the ResNet-34 model, and use the training set of the sample data set to train the third residual layer, the fourth residual layer, and the fully connected layer of the ResNet-34 model once to obtain a trained ResNet-34 model;

[0029] The training set of the sample data set is used to perform secondary training on the entire ResNet-34 model after the first training to obtain the target model.

[0030] Furthermore, the calculation formula for the weight of each target model is:

[0031]

[0032] Among them, w i represents the weight of the i-th target model, Acc i It represents the recognition accuracy of the i-th target model on the validation set of the i-th sample data set, DI represents the set distribution index, and N represents the number of target models or sample data sets.

[0033] Furthermore, the score for each injury category is calculated as follows:

[0034]

[0035]

[0036] Among them, Sco_g_x t Represents the structure image x to be recognized t The score for the gth injury category, w i represents the weight of the i-th target model, N represents the number of target models, R i (x t ) represents the structural image x that the i-th target model needs to recognize t The damage identification result, P(R i (x t ), g) represents the structural image x that the i-th target model treats to be recognized t The damage identification result is the probability of the g-th damage category, and C represents the number of damage categories.

[0037] Based on the same concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program / instruction stored in the memory, wherein the processor executes the computer program / instruction to implement the structural damage identification method as described above.

[0038] Based on the same concept, the present invention also provides a computer-readable storage medium on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the structural damage identification method as described above is implemented.

[0039] Based on the same concept, the present invention also provides a computer program product, including a computer program / instruction, which implements the structural damage identification method as described above when executed by a processor.

[0040] Beneficial Effects

[0041] Compared with the prior art, the advantages of the present invention are:

[0042] When fusing the damage identification results of multiple target models, the present invention introduces a dynamic weight allocation mechanism, which can adjust the contribution of each target model according to the actual situation, ensuring that the damage identification results with excellent performance occupy a more important position in the decision-making process. While improving the recognition accuracy, the adaptability to different application scenarios is improved, and the problem of low recognition result accuracy due to reliance on a single sensor and the difficulty of traditional fusion methods in fully utilizing the advantages of different sensors, resulting in reduced recognition accuracy, is solved.

[0043] The present invention uses a spectral Markov transfer field to convert a one-dimensional time series signal into a two-dimensional time-frequency image. The spectral Markov transfer field can better retain the time dependence and nonlinear characteristics of the signal, so that in subsequent image recognition, the model can more comprehensively capture and extract key features; the spectral Markov transfer field is combined with frequency domain analysis to calculate the transition probability in the frequency domain, which enables the spectral Markov transfer field to capture not only the dynamics of the signal in the time domain, but also the changing characteristics of the frequency components, which plays a significant role in processing signals containing periodic or complex spectral characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 is a flow chart of a method for identifying structural damage in an embodiment of the present invention;

[0046] Figure 2 is a scaled model diagram of a steel truss bridge in an embodiment of the present invention; wherein the unit of the dimensions is mm;

[0047] Figure 3 Schematic diagram of the arrangement position of the piezoelectric acceleration sensor on the steel truss bridge in an embodiment of the present invention;

[0048] Figure 4 are acceleration response signals collected by the first sensor, the third sensor, the fifth sensor, and the seventh sensor in the embodiment of the present invention;

[0049] Figure 5 It is a structural image generated by using the spectral Markov transfer field SMTF in an embodiment of the present invention;

[0050] Figure 6 is a structural image generated by using the Markov transfer field MTF in an embodiment of the present invention;

[0051] Figure 7 is a structural image generated by using the Gram angular field GADF in an embodiment of the present invention;

[0052] Figure 8 is a structural image generated by using the recursive graph RT in an embodiment of the present invention;

[0053] Fig. 9 It is a structural image generated by short-time Fourier transform STFT in an embodiment of the present invention;

[0054] Fig.10is an image conversion accuracy histogram of SMTF, MTF, GADF, RT and STFT in an embodiment of the present invention;

[0055] Fig.11 This is a ResNet-34 model architecture diagram in an embodiment of the present invention;

[0056] Fig.12 is a comparison chart of the accuracy of average voting and collaborative scoring decision-making in an embodiment of the present invention;

[0057] Fig.13 is the damage category identification confusion matrix of the first target model, the third target model, the fifth target model and the seventh target model in the embodiment of the present invention;

[0058] Fig.14 It is a comparison chart of the accuracy of a single target model in an embodiment of the present invention and the decision fusion strategy of the present invention. DETAILED DESCRIPTION

[0059] The following is a clear and complete description of the technical solutions in the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The technical solution of the present application is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Example 1

[0062] like Figure 1 As shown, a structural damage identification method based on decision fusion provided by an embodiment of the present invention includes the following steps:

[0063] Step 1: Build multiple sample datasets and divide each sample dataset into a training set and a validation set.

[0064] Each sample data set corresponds to a sensor for collecting acceleration response signals of the structure. Each sample in each sample data set includes a structural image and its damage category. The structural image is an image generated based on the acceleration response signal of the structure collected by the corresponding sensor.

[0065] The present invention takes a steel truss bridge as an example. The scaled model of the steel truss bridge is as follows: Figure 2As shown in the figure, A1 to A14 are the web member numbers. The steel truss bridge includes multiple web members, upper chords and lower chords. Five high-sensitivity piezoelectric acceleration sensors are set on the upper chord of the steel truss bridge, and three high-sensitivity piezoelectric acceleration sensors are set on the lower chord of the steel truss bridge. Figure 3 Eight piezoelectric acceleration sensors are used to fully capture the vibration signal of the steel truss bridge under external loads, that is, the acceleration response signal, as shown in Figure 4 As shown; trolleys with different speeds and different masses are used to simulate the external load. In this embodiment, the model of the piezoelectric acceleration sensor is 1A111E, the sensitivity is 100mV / g, the frequency range is 0.5~7000Hz, and the measurement range is ±50g. In addition, the present invention also uses a multi-channel dynamic data acquisition device to record the acceleration response signals collected by 8 piezoelectric acceleration sensors to ensure that the data of 8 piezoelectric acceleration sensors can be collected at the same time. In this embodiment, the dynamic data acquisition device uses a strain tester with a model of DH3822 and a sampling rate of 500Hz.

[0066] Eight piezoelectric accelerometers collect the acceleration response signals of the steel truss bridge under different web damages. The specific damage categories (or damage conditions) are shown in Table 1.

[0067] Table 1 Different types of damage

[0068] Injury Type Damage location Damage Category S1 healthy Damage Category S2 A3 Damage Category S3 A6 Damage Category S4 A9 Damage Category S5 A11 Damage Category S6 A7,A10 Damage Category S7 A8,A12 Damage Category S8 A10,A13

[0069] For each piezoelectric acceleration sensor, 100 acceleration response signals are collected for each damage category to obtain 800 acceleration response signals; the 800 acceleration response signals are converted into time-frequency diagrams to obtain 800 structural images and their damage categories, and a sample data set is constructed from the 800 structural images and their damage categories. A sample data set is constructed from the acceleration response signals collected by each piezoelectric acceleration sensor. In this embodiment, there are 8 piezoelectric acceleration sensors, so 8 sample data sets are obtained.

[0070] Convolutional neural networks have demonstrated excellent performance in image classification and recognition, target positioning and detection, and damage state recognition. Deep convolutional networks can automatically realize image feature extraction and classification functions, which continuously improves the accuracy of image classification. However, such methods cannot be directly applied to one-dimensional time series classification. In order to fully apply convolutional neural networks to the structural inspection industry, one-dimensional time series are converted into time-frequency diagrams, that is, two-dimensional images, to realize the classification of one-dimensional time series by deep convolutional neural networks. Time-frequency diagrams can express the dynamic characteristics of signals in the form of images, which are particularly effective for nonlinear and non-stationary signals, and can visualize the autocorrelation of structures at all their time scales in two dimensions.

[0071] Before converting a one-dimensional time series signal (i.e., acceleration response signal) into a two-dimensional time-frequency image (i.e., structural image), the one-dimensional time series signal is preprocessed to eliminate noise and outliers, and the one-dimensional time series signal is standardized. There are many ways to convert a one-dimensional time series signal into a two-dimensional time-frequency image, such as Markov Transfer Field (MTF), Gram Angle Field (GADF), Recursive Graph (RT), Short Time Fourier Transform (STFT), etc. Conventional time series to two-dimensional time-frequency image methods often face the limitation of batch processing when processing time series of different lengths. Because the length of the acceleration response signal is not consistent under different external loads, and the length of the acceleration response signal is also different during each measurement, the signal length becomes an important influencing factor in feature classification.

[0072] In order to solve the above problems, the present invention uses spectral Markov transfer field SMTF to convert the acceleration response signal into a structural image, which specifically includes:

[0073] Step 1.1: Calculate the frequency spectrum of the acceleration response signal. The calculation formula is:

[0074]

[0075] Where s(k) represents the kth component in the spectrum signal, x(i) represents the value of the i-th sampling point in the acceleration response signal, n represents the number or length of sampling points of the acceleration response signal, abs() represents the modulus value function, and j represents the imaginary unit. The kth component s(k) in the spectrum signal is a complex number, and the kth component s(k) represents the amplitude and phase information at frequency k. represents the Fourier kernel function.

[0076] Step 1.2: Perform normalization processing on each component in the spectrum signal to obtain the normalized component of the spectrum signal.

[0077] In this embodiment, a minimum-maximum normalization method is used to normalize each component in the spectrum signal. The specific formula is:

[0078]

[0079] Where u(k) represents the kth normalized component of the spectrum signal, s max Represents the maximum component in the spectrum signal, s min Represents the minimum component in the spectral signal.

[0080] Step 1.3: Evenly divide the value range of the amplitude set to obtain Q sub-intervals; wherein the amplitude set includes the amplitudes of all normalized components of the spectrum signal.

[0081] Since each amplitude in the amplitude set is normalized, the value range of the amplitude set is [0,1]. The value range of the amplitude set is evenly divided, that is, the logarithmic interval [0,1] is evenly divided to obtain Q sub-intervals, each of which can be expressed as q j , j∈[1,Q].

[0082] Step 1.4: Map each amplitude in the amplitude set to a subinterval; the specific mapping formula is:

[0083]

[0084] Among them, A k represents the kth amplitude in the amplitude set; q j represents the jth subinterval; argmin j It means finding j that makes the function reach the minimum value, that is, finding The minimum j. Formula (3) can be used to determine which sub-interval each amplitude in the amplitude set is currently in.

[0085] Step 1.5: Generate a spectral Markov transition field matrix based on the probability of each amplitude transferring from the current subinterval to other subintervals.

[0086] In this embodiment, the spectral Markov transition field matrix of size n×n can be expressed as:

[0087]

[0088]

[0089] Among them, p(A a ∈q i ∣A b ∈q j ) indicates that the amplitude is at q when the frequency k = b j After several state changes, it is in q when k = a i The conditional probability of .

[0090] Step 1.6: Convert the spectral Markov transfer field matrix into a two-dimensional image to obtain a structural image.

[0091] The element value of each position in the spectral Markov transfer field matrix represents the gray value of the corresponding pixel point of the image, thereby generating a two-dimensional image. Before converting the spectral Markov transfer field matrix into a two-dimensional image, the spectral Markov transfer field matrix is ​​first standardized so that the gray value of each pixel point of the image is between 0 and 255, which is convenient for subsequent image processing and pattern recognition. The structural image is saved in PNG image format for subsequent analysis and processing.

[0092] Figures 5 to 9 shows two-dimensional images generated using different methods, Fig.10 The image conversion accuracy of various methods is shown by Fig.10 It can be seen that the image conversion accuracy of the spectral Markov transfer field of the present invention is the highest. The spectral Markov transfer field converts the one-dimensional time series signal to the frequency domain, so that the signal can remain stable within a certain structural natural frequency range and is no longer affected by the sensor measurement duration, thereby ensuring the accuracy and consistency of feature classification. The spectral Markov transfer field technology not only improves the flexibility of signal processing, but also enhances the robustness of the model to different measurement conditions, providing a more stable and reliable approach for signal analysis.

[0093] Step 2: Load multiple damage identification models.

[0094] The number of damage recognition models is equal to the number of sample data sets, and the number of sample data sets is equal to the number of acceleration sensors. The acceleration response signal collected by an acceleration sensor generates a sample data set, and one sample data set trains a damage recognition model. The key to the success of convolutional neural networks (CNNs) is that they can automatically learn hierarchical representations from raw data, which is particularly beneficial in image recognition because traditional artificial feature capture is difficult to summarize the complexity of visual data. However, deep neural network models often have a gradient vanishing problem, resulting in low recognition accuracy. In order to solve this problem, the damage recognition model of the present invention uses the ResNet-34 model. The ResNet-34 model is essentially composed of multiple residual blocks stacked together. The residual blocks use skip connections (Skip Connections) on the basis of ordinary CNNs to achieve the purpose of enabling the model to learn residuals, solve the gradient vanishing and gradient explosion problems that occur in deep neural networks during training, and enhance the comprehensiveness and effectiveness of the ResNet-34 model in image feature extraction.

[0095] Table 2 Detailed architecture of the ResNet-34 model

[0096]

[0097] In a specific embodiment of the present invention, Fig.11As shown in Table 2, the ResNet-34 model includes an input layer, a convolutional layer, a maximum pooling layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, an average pooling layer, and a fully connected layer, which are connected in sequence.

[0098] Input layer: Use MATLAB to receive a structural image of size 224×224×3, normalize the structural image, and convert the image into a pixel matrix. Each element in the pixel matrix represents the pixel value of the corresponding position in the image. The processed image size is 224×224×3, that is, 224 pixels long and 224 pixels wide. Each pixel contains 3 channels (RGB), that is, the input image depth is 3.

[0099] Convolutional layer: Use 64 convolution kernels of size 7×7, stride 2, border padding 3, and the output feature map size is (112, 112, 64).

[0100] Maximum pooling layer: The pooling kernel size is 3×3, the stride is 2, the border padding is 0, and the size of the feature map after pooling is (56, 56, 64).

[0101] The first residual layer: consists of 3 residual blocks, each residual block contains 2 3×3 convolutional layers, and the output feature map size is (56, 56, 64).

[0102] The second residual layer: consists of 4 residual blocks, each residual block contains 2 3×3 convolutional layers, and the output feature map size is (28, 28, 128).

[0103] The third residual layer: consists of 6 residual blocks, each residual block contains 2 3×3 convolutional layers, and the output feature map size is (14,14,256).

[0104] The fourth residual layer: consists of 3 residual blocks, each residual block contains 2 3×3 convolutional layers, and the output feature map size is (7,7,512).

[0105] Average pooling layer: The pooling kernel size is 7×7 and the output feature map size is (1,1,512).

[0106] Fully connected layer: The feature map output by the average pooling layer is processed through a fully connected layer. According to the number of damage categories, the output dimension of the fully connected layer is changed from the original 1000 to 8. Finally, it is processed through the Softmax activation function to output the damage category result.

[0107] Compared with traditional shallow network models, the ResNet-34 model of the present invention can automatically extract multi-level deep features from images, improve the accuracy and robustness of damage identification, and effectively solve the gradient vanishing problem that occurs in deep network training.

[0108] Step 3: Use the training set of each sample data set to train the corresponding damage identification model to obtain the target model.

[0109] In order to save model training time and achieve better training results, for each damage identification model, the present invention adopts transfer learning technology, first loads the ResNet-34 model pre-trained with a public data set, and then trains the pre-trained ResNet-34 model with the sample data set of the present invention, that is, the source domain data set is a public data set, and the target domain data set is a sample data set constructed by the present invention through a structural field test. In a specific embodiment of the present invention, for each damage identification model, the damage identification model is trained using the training set of the sample data set, including:

[0110] Step 3.1: Freeze the convolutional layer, the first residual layer, and the second residual layer of the ResNet-34 model, and use the training set of the sample data set to train the third residual layer, the fourth residual layer, and the fully connected layer of the ResNet-34 model once to obtain a trained ResNet-34 model.

[0111] The convolutional layer, the first residual layer, and the second residual layer of the ResNet-34 model are used to extract the basic features of the image. Freezing the convolutional layer, the first residual layer, and the second residual layer of the ResNet-34 model can maintain its feature representation capabilities learned during pre-training; at the same time, the parameters of the third residual layer, the fourth residual layer, and the fully connected layer that are closer to the output are retained for training. Through this training, the ResNet-34 model can effectively adapt to the classification task of the sample data set.

[0112] During training, the optimizer is Adam, the learning rate is set to 0.001, and the loss function is the cross entropy loss function.

[0113] Step 3.2: Use the training set of the sample data set to perform secondary training on the entire ResNet-34 model after the first training to obtain the target model.

[0114] During the secondary training, the convolutional layer, the first residual layer, and the second residual layer of the ResNet-34 model were unfrozen so that the convolutional layer, the first residual layer, and the second residual layer participated in the training to further optimize the overall feature extraction capability. Although the third residual layer, the fourth residual layer, and the fully connected layer were trained during the first training process, these layers continued to be optimized in the secondary training phase to ensure that the overall performance of the model was further improved. At this time, a lower learning rate of 0.0001 was used, and the ReduceLROnPlateau learning rate scheduler was applied, which dynamically adjusted the learning rate according to the performance on the validation set, thereby improving the training effect of the model and preventing overfitting. The two trainings ensured that all layers of the model could effectively adapt to the new classification task, and the model was evaluated on the test set in the final stage to verify its classification performance and robustness in practical applications.

[0115] The present invention adopts transfer learning technology to improve the pre-trained model to achieve accurate classification of structural damage, improve training efficiency, reduce the amount of data required in actual engineering, and avoid the model falling into a local minimum.

[0116] Multiple target models can obtain multiple damage identification results during damage identification. Whether each damage identification result can be fully utilized to achieve collaborative identification of damage types depends on the design of the decision fusion strategy.

[0117] Traditional decision fusion strategies include average voting method and standard weighted voting method. The main problem of the average voting method is that the contribution of better target models is weakened, which may affect the fusion result. The standard weighted voting method is used to improve the overall accuracy of each target model. However, the flexibility of the standard weighted voting method is poor because the weight assigned to each target model is always fixed. In order to solve the problems of traditional decision fusion strategies, the present invention proposes a collaborative scoring fusion strategy, specifically steps 4 to 8.

[0118] Step 4: Use the target model to perform damage identification on the validation set of the sample dataset.

[0119] By using the target model to perform damage identification on each sample in the validation set of the corresponding sample data set, the identification result of each sample can be obtained.

[0120] Step 5: Calculate the recognition accuracy of the target model on the validation set of the sample data set based on the damage recognition results.

[0121] The recognition accuracy of the target model on the validation set of the sample data set is calculated based on the damage category of each sample and the damage category identified by the target model. The specific formula is:

[0122]

[0123] Among them, Acc i Acc represents the recognition accuracy of the i-th target model on the validation set of the i-th sample data set. i Reflects the target model R i The overall performance in the current task is the key basis for subsequent weight allocation.

[0124] Step 6: Calculate the weight of the corresponding target model based on the recognition accuracy of each target model on the validation set of the corresponding sample data set.

[0125] In this embodiment, the calculation formula for the weight of each target model is:

[0126]

[0127] Among them, w i represents the weight of the i-th target model, DI represents the set distribution index, and N represents the number of target models or sample data sets.

[0128] The distribution index DI is an adjustable distribution index that depends on the specific task. By adjusting the value of DI, the sensitivity of weight distribution is affected. A larger DI value gives a target model with a higher accuracy a larger weight, enhancing its influence in the final decision; on the contrary, a smaller DI value makes the weight distribution more even, which helps to reduce the dominant role of a single model in the decision, thereby avoiding the risk of overall decision failure due to poor performance of individual models.

[0129] The value of the distribution index DI is determined by trial calculation, that is, when the distribution index DI takes different values, the accuracy of the final damage category is calculated, and the DI value corresponding to the highest accuracy of the final damage category is selected. In this embodiment, DI is calculated from 1 to 8, and finally DI=3 is determined.

[0130] The multi-sensor decision fusion strategy with dynamic weight allocation can dynamically adjust the weight of each target model according to the performance of each target model on the corresponding validation set, ensuring that the sensor data with excellent performance occupies a more important position in the decision-making process. Compared with the traditional fixed weight or average voting method, the present invention can adaptively enhance the influence of high-quality data while reducing the negative impact of low-quality data, thereby significantly improving the accuracy of recognition.

[0131] Step 7: For the structure image to be identified, use each target model to perform damage identification on the structure image to be identified.

[0132] Step 8: Calculate the score of each damage category based on the damage recognition result of each target model for the structural image to be identified and the weight of each target model, and take the damage category corresponding to the highest score as the final damage category of the structural image to be identified.

[0133] For each damage category g(g∈{1,2,...,C}), calculate the structural image x to be identified t Sco_g_x for this injury category t The score calculation adopts the weighted summation method, which is to sum the damage recognition result of each target model for the structural image to be recognized with the weight of the target model. The specific formula is:

[0134]

[0135]

[0136] Among them, Sco_g_x t Represents the structure image x to be recognized t The score for the gth injury category, w i represents the weight of the i-th target model, R i (x t ) represents the structural image x that the i-th target model needs to recognize t The damage identification result, P(R i (x t ), g) represents the structural image x that the i-th target model treats to be recognized t The damage identification result is the probability of the gth damage category, and C represents the number of damage categories. For Table 1, C is equal to 8.

[0137] The damage category corresponding to the highest score is selected as the final damage category of the structural image to be identified. When there are multiple highest scores, in order to ensure the rigor of the decision, the damage category corresponding to the highest score is randomly selected from the multiple highest scores as the final damage category, avoiding the randomness of the decision from having a significant impact on the recognition stability.

[0138] Fig.12 The accuracy comparison between the collaborative scoring fusion strategy of the present invention and the average voting method is shown in FIG. Fig.12 It can be seen that the accuracy of the collaborative scoring fusion strategy of the present invention is better than that of the average voting method.

[0139] Fig.13 The confusion matrices of the 1st target model (corresponding to the 1st sample data set and the 1st sensor), the 3rd target model (corresponding to the 3rd sample data set and the 3rd sensor), the 5th target model (corresponding to the 5th sample data set and the 5th sensor), and the 7th target model (corresponding to the 7th sample data set and the 7th sensor) in the corresponding validation sets are shown. Fig.13In the figure, the horizontal axis of the confusion matrix represents the true label (i.e., the true damage category), the vertical axis represents the predicted label (i.e., the damage identification result output by the target model), and the value on the diagonal represents the number of samples predicted correctly.

[0140] Fig.14 The figure shows the recognition accuracy of 8 target models on the validation set and the recognition accuracy of the validation set obtained by the collaborative scoring fusion strategy of the present invention. The 8 target models correspond to 8 sensors, namely sensor 1 (the first sensor), sensor 2 (the second sensor), sensor 3 (the third sensor), sensor 4 (the fourth sensor), sensor 5 (the fifth sensor), sensor 6 (the sixth sensor), sensor 7 (the seventh sensor), and sensor 8 (the eighth sensor). Fig.14 It can be seen that the average accuracy of the eight target models is 85%, while the collaborative scoring fusion strategy achieves an accuracy of 98.6%, which greatly improves the accuracy and robustness of damage identification. When one of the target models fails to identify the target model, the collaborative scoring fusion strategy can make up for the error, avoiding the randomness and error of a single identification result and improving the identification accuracy.

[0141] The collaborative scoring fusion strategy of the present invention achieves accurate measurement of the contribution of each target model by introducing a dynamic weight allocation mechanism. Compared with the traditional average voting method, this method not only improves the accuracy of decision-making, but also enhances the adaptability to different application scenarios. For example, when the data quality of certain sensors is poor, the weight of the corresponding target model will be automatically reduced, thereby reducing its negative impact on the final decision. This adaptive capability makes the multi-sensor decision-making strategy of the present invention more robust in practical applications and suitable for various complex scenarios.

[0142] After obtaining the final damage identification result, the target model can be further optimized and calibrated through the feedback mechanism. For example, when the actual damage category is inconsistent with the final damage identification result, the accuracy can be further improved by adjusting the distribution index DI or adjusting the training process of the target model. In addition, the validation set can be continuously updated and expanded according to the needs of the actual application scenario to ensure the rationality and timeliness of the weight distribution.

[0143] Example 2

[0144] An embodiment of the present invention also provides an electronic device, which includes: a memory, a processor, and a computer program / instructions stored in the memory, and the processor executes the computer program / instructions to implement the structural damage identification method in the embodiment of the present application.

[0145] Although not shown, the electronic device includes a processor, which can perform various appropriate operations and processes according to the program and / or data stored in the read-only memory (ROM) or the program and / or data loaded from the storage portion into the random access memory (RAM). The processor can be a multi-core processor, or it can include multiple processors. In some embodiments, the processor can include a general main processor and one or more special coprocessors, such as a central processing unit, a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), etc. In RAM, various programs and data required for device operation are also stored. The processor, ROM and RAM are connected to each other via a bus. The input / output (I / O) interface is also connected to the bus.

[0146] The processor and memory are used together to execute the program / instructions stored in the memory. When the program / instructions are executed by the computer, the methods, steps or functions described in the above embodiments can be implemented.

[0147] Although not shown, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the structural damage identification method in the embodiment of the present application is implemented.

[0148] Storage media in embodiments of the present invention include permanent and non-permanent, removable and non-removable items that can be implemented by any method or technology to store information. Examples of storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0149] Readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0150] Although not shown, an embodiment of the present invention further provides a computer program product, including: a computer program / instruction, which implements the structural damage identification method in the embodiment of the present application when executed by a processor.

[0151] What is disclosed above is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or modifications within the technical scope disclosed in the present invention, which should be covered within the protection scope of the present invention.

Claims

1. A structural damage identification method based on decision fusion, It is characterized in that The identification method comprises: Constructing multiple sample data sets, and dividing each of the sample data sets into a training set and a validation set; wherein each sample data set corresponds to a sensor for collecting an acceleration response signal of a structure, and each sample in each sample data set includes a structural image and its damage category, and the structural image is an image generated according to the acceleration response signal of the structure collected by the corresponding sensor; Load multiple damage identification models; each damage identification model corresponds to a sample data set; Using the training set of each sample data set to train the corresponding damage recognition model to obtain a target model; using the target model to perform damage recognition on the validation set of the sample data set; and calculating the recognition accuracy of the target model on the validation set of the sample data set based on the damage recognition result; The weight of the corresponding target model is calculated according to the recognition accuracy of each target model on the validation set of the corresponding sample data set; For a structure image to be identified, using each target model to perform damage identification on the structure image to be identified; The score of each damage category is calculated according to the damage recognition result of each target model on the structural image to be identified and the weight of each target model, and the damage category corresponding to the highest score is used as the final damage category of the structural image to be identified.

2. The structural damage identification method according to claim 1, It is characterized in that According to the acceleration response signal of the structure collected by the sensor, the spectral Markov transfer field is used to generate the structure image, including: The frequency spectrum signal of the acceleration response signal is calculated using the following formula: Where, s(k) represents the kth component in the spectrum signal, x(i) represents the value of the i-th sampling point in the acceleration response signal, n represents the number of sampling points of the acceleration response signal, abs( ) represents the modulus value function, and j represents the imaginary unit; Performing normalization processing on each component in the spectrum signal to obtain a normalized component of the spectrum signal; The value range of the amplitude set is evenly divided to obtain Q sub-intervals; wherein the amplitude set includes the amplitudes of all normalized components of the spectrum signal; Each amplitude in the amplitude set is mapped to a subinterval; wherein the specific mapping formula is: Among them, A k represents the kth amplitude in the amplitude set, q j represents the jth subinterval, argmin j It means finding j that makes the function reach the minimum value; Generate a spectral Markov transition field matrix according to the probability of each amplitude being transferred from the current subinterval to other subintervals; The spectral Markov transfer field matrix is ​​converted into a two-dimensional image to obtain a structural image.

3. The structural damage identification method according to claim 1, It is characterized in that The damage identification model selects a pre-trained ResNet-34 model, and the ResNet-34 model includes an input layer, a convolution layer, a maximum pooling layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, an average pooling layer and a fully connected layer connected in sequence.

4. The structural damage identification method according to claim 3, It is characterized in that The convolution layer includes 64 convolution kernels of size 7×7 and a step size of 2; the pooling kernel size of the maximum pooling layer is 3×3 and a step size of 2; the first residual layer includes 3 residual blocks, the second residual layer includes 4 residual blocks, the third residual layer includes 6 residual blocks, and the fourth residual layer includes 3 residual blocks, each residual block includes 2 3×3 convolution kernels; the pooling kernel size of the average pooling layer is 7×7.

5. The structural damage identification method according to claim 3, It is characterized in that The corresponding damage identification model is trained using the training set of each sample data set, including: Freeze the convolutional layer, the first residual layer, and the second residual layer of the ResNet-34 model, and use the training set of the sample data set to train the third residual layer, the fourth residual layer, and the fully connected layer of the ResNet-34 model once to obtain a trained ResNet-34 model; The training set of the sample data set is used to perform secondary training on the entire ResNet-34 model after the first training to obtain the target model.

6. The structural damage identification method according to claim 1, It is characterized in that The calculation formula for the weight of each target model is: Among them, w i represents the weight of the i-th target model, Acc i It represents the recognition accuracy of the i-th target model on the validation set of the i-th sample data set, DI represents the set distribution index, and N represents the number of target models or sample data sets.

7. The structural damage identification method according to any one of claims 1 to 6, It is characterized in that The score for each impairment category is calculated as: Among them, Sco - g - x t Represents the structure image x to be recognized t The score for the gth injury category, w i represents the weight of the i-th target model, N represents the number of target models, R i (x t ) represents the structural image x that the i-th target model needs to recognize t The damage identification result, P(R i (x t ), g) represents the structural image x that the i-th target model treats to be recognized t The damage identification result is the probability of the g-th damage category, and C represents the number of damage categories.

8. An electronic device comprising a memory, a processor and a computer program / instructions stored on the memory, It is characterized in that The processor executes the computer program / instructions to implement the structural damage identification method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, It is characterized in that When the computer program / instructions are executed by a processor, the structural damage identification method according to any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, It is characterized in that When the computer program / instructions are executed by a processor, the structural damage identification method according to any one of claims 1 to 7 is implemented.