Interval uncertainty damage identification method based on CBAM-Unet neural network

Abstract

The invention discloses an interval uncertainty damage identification method based on a CBAM-Unet neural network, and the method comprises the steps: generating a damage information matrix through employing a four-neighborhood growth method, and obtaining the strain field data of a damage structure through the simulation of finite element software, so as to construct a data set; a convolution attention mechanism is fused into a U-net framework, and the distinction degree of damaged area features is enhanced; performing interval uncertainty analysis on the structural damage based on a Chebyshev polynomial by utilizing a dimension reduction analysis method; a collocation model required by a dimension reduction method is generated through transfer learning, and data dependence and calculation cost are reduced. The method focuses on structural damage identification, fully considers the uncertainty of the structure in the aspects of geometry, external load, material attributes and the like, systematically gives a complete process from data set construction, neural network model construction to uncertainty analysis, and realizes accurate identification of structural damage positioning and quantification under the influence of uncertainty factors.

Description

Technical Field

[0001] This invention relates to structural damage identification technology, and in particular to a method for identifying interval uncertainty damage based on the CBAM-Unet (ConvolutionalBlock Attention Module U-net) neural network. This method transforms the structural damage identification problem into an image segmentation problem, obtains the configuration and degree information of the damage by means of the structural strain field, fully considers the impact of the above-mentioned uncertainty on damage identification, establishes a complete technical process from dataset construction, neural network model building to uncertainty analysis, and realizes accurate identification of structural damage location and quantification under the influence of uncertainty factors. Background Technology

[0002] With technological advancements and changing demands, engineering structures are often becoming larger and more complex. Under harsh working environments, complex operating conditions, and variable external factors, structural damage is inevitable. As damage accumulates, the performance of structural materials gradually decreases, affecting the normal operation of the structure and even leading to instability or failure, creating serious safety hazards and threatening the lives and property of the public. Therefore, to ensure the normal operation of structures and prevent disasters, timely detection of structural damage and assessment of its extent are crucial. Real-time and accurate acquisition of structural safety information, enabling timely repair and maintenance, can greatly reduce the occurrence of accidents and protect the safety of people's property.

[0003] Furthermore, engineering structures often involve numerous uncertainties during fabrication, processing, and actual use, influenced by limitations in manufacturing processes, deviations in technical methods, inherent product defects, and human interference. Considering that these uncertainties can generally be categorized into uncertainties related to material properties, loads, physical dimensions, solution conditions, and reference models, these factors significantly impact the accuracy of damage identification. Therefore, conducting uncertainty analysis closely related to structural damage identification is crucial for improving the reliability of damage identification results, ensuring that damage identification effectively serves structural safety maintenance, and ultimately preventing safety risks arising from misjudgments or omissions in damage assessment. Summary of the Invention

[0004] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies and provide a method for identifying interval uncertainty-based damage using a CBAM-Unet neural network. This invention transforms the structural damage identification problem into an image segmentation problem, utilizes the structural strain field to obtain information on the configuration and extent of damage, and fully considers the impact of the aforementioned uncertainties on damage identification, thereby accurately identifying damage features under the influence of uncertainty factors.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for identifying interval uncertainty-based damage using a CBAM-Unet neural network, comprising:

[0007] S110 uses the four-neighborhood growth method to generate a damage information matrix and obtains strain field data of the damaged structure through finite element software simulation to construct a complete dataset.

[0008] S120 integrates the convolutional attention mechanism into the U-net framework to form the CBAM-Unet neural network. The attention mechanism enhances the discriminativeness of damage region features, effectively achieving high-precision localization and quantitative identification of structural damage from strain field data.

[0009] S130, using transfer learning to generate the collocation model required for dimensionality reduction analysis;

[0010] S140 uses dimensionality reduction analysis to perform interval uncertainty analysis on structural damage based on Chebyshev polynomials.

[0011] A computing device includes: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method.

[0012] A readable storage medium storing program instructions that, when read and executed by a computing device, cause the computing device to perform the method.

[0013] Beneficial effects:

[0014] This method focuses on structural damage identification, fully considering the uncertainties of structures in terms of geometry, external load, and material properties. It systematically presents a complete process from dataset construction and neural network model building to uncertainty analysis, realizing accurate identification of structural damage location and quantification under the influence of uncertain factors. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the design of the interval uncertainty damage identification method based on the CBAM-Unet neural network according to the present invention.

[0016] Figure 2 This invention relates to the wing geometry model used in the calculation examples;

[0017] Figure 3 This is a graph showing the uncertainty interval results obtained from damage uncertainty interval analysis of the wing model;

[0018] Figure 4 This is a graph showing the results of the uncertainty interval analysis of the damaged area. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] like Figure 1 As shown, this invention proposes a method for identifying interval uncertainty damage based on a CBAM-Unet neural network, comprising the following steps:

[0021] S110 uses the four-neighborhood growth method to generate a damage information matrix, and obtains strain field data of the damaged structure from the damage information matrix through finite element software simulation to construct a complete dataset.

[0022] S120 integrates the convolutional attention mechanism into the U-net framework to form the CBAM-Unet neural network. The attention mechanism enhances the discriminativeness of damage region features, effectively achieving high-precision localization and quantitative identification of structural damage from strain field data.

[0023] S130 utilizes transfer learning to generate the collocation model required for dimensionality reduction analysis, and leverages knowledge transfer to reduce data dependence, significantly reducing the amount of data and computational cost in structural damage uncertainty analysis.

[0024] S140 uses dimensionality reduction analysis to perform interval uncertainty analysis on structural damage based on Chebyshev polynomials.

[0025] S110 uses the four-neighborhood growth method to generate a damage information matrix, and obtains strain field data of the damaged structure from the damage information matrix through finite element software simulation to construct a complete dataset. Specifically, this includes: establishing label data using the four-neighborhood growth method; the label data is a damage information matrix used to calibrate the damage state. The core component of this damage information matrix is ​​the stiffness reduction coefficient α corresponding to each element in the finite element model, where the value of α is limited to the interval [0,1]: when α=0, it means that the corresponding element has completely lost its load-bearing capacity; when α=1, it indicates that the element has not experienced any damage. The index position of each stiffness reduction coefficient in the damage information matrix forms a strict spatial mapping relationship with the geometric coordinates of the elements in the finite element model.

[0026] To ensure that the spatial distribution of the damaged area satisfies both randomness and continuity, a four-neighborhood growth algorithm is used to generate an information matrix containing irregular damaged areas. In S110, generating the damage information matrix using the four-neighborhood growth method can include: assuming the target structure is discretized using finite element analysis to form an N×N element mesh, denoted as Ω, whose elements are the coordinate set of all elements in the mesh {(x,y)|x,y∈{1,2,…,N}}. When the four-neighborhood growth method is initiated, an initial seed element p0(i,j) is randomly determined within Ω and included in the damaged area R; simultaneously, the four neighboring elements of p0 (i.e., (i-1,j), (i+1,j), (i,j-1), (i,j+1)) are assigned to the candidate damaged element set Γ. Then, the iterative growth stage begins: a unit p is randomly selected from the candidate set Γ. k Incorporate it into the damaged area R, and simultaneously add p k Unoccupied cells in the four neighboring regions are added to Γ. Through such iterative cycles, a damage information matrix that conforms to the irregular shape requirement is finally generated.

[0027] In engineering, the strain field of a structure is usually obtained using optical methods. However, in practical applications, it is difficult to obtain a sufficient number of accurately labeled strain field images. Therefore, when constructing relevant datasets, data generated by finite element simulation is often used as a substitute. In S110, obtaining strain field data of a damaged structure from a damage information matrix through finite element software simulation includes: inputting the damage information matrix into the finite element analysis software for solving, which yields strain contour maps of the damaged structure in the x-axis direction, y-axis direction, and 45° oblique direction. After normalizing and preprocessing these strain contour map data, the three contour maps are fused along the strain direction dimension to form a three-dimensional matrix, which serves as the input data for the CBAM-Unet neural network.

[0028] S120 integrates a convolutional attention mechanism into the U-net framework, forming the CBAM-Unet neural network. This attention mechanism enhances the discriminative power of damage region features, effectively enabling high-precision localization and quantitative identification of structural damage from strain field data, including:

[0029] The Convolutional Concentration Attention (CBAM) mechanism is incorporated into the U-net neural network. CBAM, as a hybrid module fusing channel and spatial attention, adaptively enhances key information in the feature map through a sequential action mechanism: first, the channel attention module... For input features Average and max pooling are performed in the spatial dimension to generate two one-dimensional vectors, which are then input into a two-layer MLP. Channel weights are obtained after sigmoid activation; finally, the spatial attention module... The features are averaged and max-pooled along the channel dimension, the results are concatenated and passed through a 7×7 convolutional layer, and spatial weights are generated by a sigmoid function. These two processes operate sequentially and feature weighting is achieved through element-wise multiplication. The overall relationship can be expressed as:

[0030] ,

[0031] In the formula, This represents the final output feature. This indicates element-wise multiplication.

[0032] The CBAM-Unet model is constructed by embedding CBAM modules within the various layers of the U-net encoder. This model employs a symmetrical architecture for the encoder and decoder units: the encoder consists of four layers, each comprising two sets of "convolution operation-batch normalization-ReLU activation" structures and a 2×2 max pooling layer. The downsampled features are then fed into the CBAM module for weighted operations, enabling the learning of weights in both the channel and spatial dimensions. The decoder, also symmetrically positioned with the encoder, consists of four layers. Each layer first uses a convolutional layer to compress the number of feature channels, then upsamples using a 2×2 transposed convolution. Subsequently, skip connections are used to introduce features corresponding to those in the encoder, and finally, a "convolution-batch normalization-ReLU" structure is used to optimize the fused features.

[0033] S130, using transfer learning to generate the collocation model required for dimensionality reduction analysis, including: dimensionality reduction analysis is used in step S140, which can significantly reduce the number of collocations required in uncertainty analysis; and transfer learning is used in step S130, which can further reduce the amount of data required and the computational cost.

[0034] The specific steps of transfer learning are as follows: Select the midpoint of the uncertain variable. The generated source dataset is used to train the CBAM-Unet network, and the resulting neural network model is used as a pre-trained model. Then, all parameters of the encoder part in the pre-trained model are frozen, and the parameters at the collocation points are used to generate the target dataset. The number of parameters in the target dataset is only one-tenth that of the source dataset. The pre-trained model is then transferred and trained using the target dataset, thus quickly obtaining a high-accuracy collocation model.

[0035] S140 uses dimensionality reduction analysis to perform interval uncertainty analysis of structural damage based on Chebyshev polynomials, including:

[0036] Due to factors such as manufacturing processes, experimental conditions, and human interference, uncertainties exist in geometric and material properties in actual engineering projects, which in turn affect the accuracy of damage identification results. Uncertainty variables There is a certain functional relationship between the damage parameters and the damage parameters. However, this functional relationship is often nonlinear and not explicit. Therefore, it can be fitted using Chebyshev polynomials.

[0037] First, the uncertain variable is linearly transformed:

[0038] ,

[0039] In the formula, This represents the uncertainty variable after normalization. Represents uncertain variables. This represents the lower bound of the interval for uncertain variables. Denote the upper bound of the interval. Use Chebyshev polynomials to... By approximating, we have:

[0040] ,

[0041] In the formula, for Vychebyshev polynomials It is a polynomial function that approximates the given value. It is the order of a single Chebyshev polynomial. It is the order of every dimensional Chebyshev polynomial. for The number of times 0 appears in the text. for The Vychebyshev polynomial coefficients can be obtained using the Gauss-Chebyshev interpolation integral formula:

[0042] ,

[0043] In the formula, The number of interpolation points in each dimension. For interpolation points, These are the interpolation point numbers for each uncertain variable; the interpolation points for each dimension are... Zeros of Chebyshev polynomials:

[0044] .

[0045] As can be seen, when directly using Chebyshev fitting, the number of collocation points increases exponentially with the increase of the number of bits. Therefore, dimensionality reduction is used to reduce the number of collocation points required for the approximate function.

[0046] Will By performing dimensional decomposition and ignoring residual terms, we obtain:

[0047] ,

[0048] In the formula, yes The s-dimensional approximation function, It is a constant. for Dimensional uncertainty variables The influence component.

[0049] If take That is, only a one-dimensional approximation function is considered. The midpoint of the uncertain variable is... As a reference point, we have:

[0050] ,

[0051] but The one-dimensional approximation function can be expressed as:

[0052] ,

[0053] in, The reference only considers the i-th dimension. , It is a constant term;

[0054] Similarly, The two-dimensional approximation function can be expressed as:

[0055] .

[0056] Example:

[0057] To gain a fuller understanding of the features of this invention and its applicability to practical engineering, this invention addresses, for example... Figure 2 Damage identification was performed on the wing model shown. The wing model consists of two skin panels and several ribs. A shell element mesh was used to discretize the entire model, with the thickness of both the skin and ribs set to 3 mm. Material parameters were uniformly set to an elastic modulus of 200 GPa and a Poisson's ratio of 0.3. The skin was meshed using a 32×128 structured finite element mesh. Load and constraint conditions were set as follows: the model was subjected to a uniformly distributed force (500 N) perpendicular to the upper skin, and a completely fixed constraint was applied at the wing root.

[0058] Considering that most actual damage occurs in the skin region, irregular damage morphologies are generated using the four-neighborhood growing method, and structural response data under damaged conditions are obtained through finite element simulation. Based on the above model parameters and damage generation rules, a pre-training dataset containing 2000 samples is constructed in batches. Damages in this dataset are all distributed in the skin region, with random locations and shapes. The number of damages is controlled to 1-3, the damage severity is between 5% and 25%, and the area of ​​a single damage is 50-300 units. This dataset is divided into training and testing sets in a 4:1 ratio, with 1600 samples used for model training and 400 samples used for model testing and performance verification, resulting in a pre-trained model.

[0059] The uncertainty variables set in this example are the material's elastic modulus, Poisson's ratio, and shell structure thickness. At a 2% uncertainty level, collocation data were generated using the method described above, and then 19 collocation models were trained through transfer learning.

[0060] Subsequently, a damage uncertainty interval analysis was conducted on the wing model. Four typical operating conditions are presented, and their uncertainty interval results are as follows: Figure 3 As shown. By Figure 3 It is evident that the upper and lower bounds of the damage interval effectively encompass the actual damage. This method can identify not only the uncertainty of damage degree but also the uncertainty of damage size, such as the uncertainty interval of the damage region. Figure 4 As shown, the range of the damaged area can well encompass the actual damaged area. This indicates that the proposed method can effectively cope with the interference of uncertain factors and better meet the needs of structural damage identification in uncertain scenarios in engineering.

[0061] The present invention also provides a computing device, comprising: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method described thereon.

[0062] The present invention also provides a readable storage medium storing program instructions, which, when read and executed by a computing device, cause the computing device to perform the method described thereon.

[0063] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of explaining or limiting the subject matter of the invention.

Claims

1. A method for identifying interval uncertainty-based damage based on a CBAM-Unet neural network, characterized in that, include: S110 uses the four-neighborhood growth method to generate a damage information matrix and obtains strain field data of the damaged structure through finite element software simulation to construct a complete dataset. S120 integrates the convolutional attention mechanism into the U-net framework to form the CBAM-Unet neural network. The attention mechanism enhances the discriminativeness of damage region features, effectively achieving high-precision localization and quantitative identification of structural damage from strain field data. S130, using transfer learning to generate the collocation model required for dimensionality reduction analysis; S140 uses dimensionality reduction analysis to perform interval uncertainty analysis on structural damage based on Chebyshev polynomials.

2. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, S110 includes: establishing label data using the four-neighborhood growth method. The label data is a damage information matrix used to calibrate the damage state. The core component of this damage information matrix is ​​the stiffness reduction coefficient α corresponding to each element in the finite element model. The value range of α is limited to the interval [0, 1]. When α=0, it means that the corresponding element has completely lost its load-bearing capacity. When α=1, it means that the element has not suffered any damage. The index position of each stiffness reduction coefficient in the matrix forms a strict spatial mapping relationship with the geometric coordinates of the element in the finite element model.

3. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, In S110, the damage information matrix is ​​generated using the four-neighborhood growth method, which includes: assuming the target structure is discretized by finite element method to form an N×N element mesh, denoted as Ω, whose elements are the coordinate set of all elements in the mesh {(x,y)|x,y∈{1,2,…,N}}; when the four-neighborhood growth method is started, an initial seed element p0(i,j) is randomly determined in Ω and included in the damage region R; at the same time, the four neighboring elements of p0 (i.e., (i-1,j), (i+1,j), (i,j-1), (i,j+1)) are included in the candidate damage element set Γ; then the iterative growth stage begins: a unit P is randomly selected from the candidate set Γ. k Incorporate it into the damaged area R, and simultaneously add p k Unoccupied cells in the four neighboring regions are added to Γ; through such iterative cycles, a damage information matrix that meets the requirements of an irregular shape is finally generated.

4. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, In S110, the strain field data of the damaged structure obtained from the damage information matrix through finite element software simulation includes: inputting the damage information matrix into the finite element analysis software for solution to obtain strain contour maps of the damaged structure in the x-axis direction, y-axis direction and 45° oblique direction; after normalizing and preprocessing these strain contour map data, the three contour maps are fused according to the strain direction dimension to form a three-dimensional matrix, which is used as the input data of the CBAM-Unet neural network.

5. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, S120 includes: Integrating the Convolutional Attention Mechanism (CBAM) into the U-net neural network: The CBAM adaptively enhances key information in the feature map through a serial action mechanism: first, through the channel attention module... For input features Average and max pooling are performed in the spatial dimension to generate two one-dimensional vectors, which are then input into a two-layer MLP. Channel weights are obtained after sigmoid activation; finally, the spatial attention module... The features are averaged and max-pooled along the channel dimension, the results are concatenated and passed through a 7×7 convolutional layer, and spatial weights are generated by a sigmoid function. The two processes are sequentially combined and feature weighting is achieved through element-wise multiplication. The overall relationship is expressed as follows: ， In the formula, This represents the final output feature. This indicates element-wise multiplication.

6. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, The CBAM-Unet neural network adopts an architecture design in which the encoder and decoder units are symmetrically distributed: the encoder part is divided into four layers, each layer consists of two sets of "convolution operation-batch normalization processing-ReLU activation" structure and a 2×2 max pooling layer. Then, the downsampled features are fed into the CBAM module to perform weighted operations, thereby realizing the learning of weights in the channel dimension and spatial dimension. The decoder and encoder are symmetrically arranged and also consist of four layers. Each layer first uses a convolutional layer to compress the number of feature channels, then uses a 2×2 transposed convolution to perform upsampling, and then introduces the corresponding features from the encoder through skip links. Finally, the fused features are optimized using a "convolution operation - batch normalization - ReLU activation" structure.

7. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, S140 includes: uncertain variables There is a certain functional relationship between the damage parameters and the damage parameters. We fit it using Chebyshev polynomials: First, the uncertain variable is linearly transformed: ， In the formula, This represents the uncertainty variable after normalization. Represents uncertain variables. This represents the lower bound of the interval for uncertain variables. Denote the upper bound of the interval; use Chebyshev polynomials to represent the upper bound of the interval. By approximating, we obtain: ， In the formula, for Vychebyshev polynomials It is a polynomial function that approximates the given value. It is the order of a single Chebyshev polynomial. It is the order of every dimensional Chebyshev polynomial. for The number of times 0 appears in the text. for The Vychebyshev polynomial coefficients are obtained using the Gauss-Chebyshev interpolation integral formula: ， In the formula, The number of interpolation points in each dimension. For interpolation points, These are the interpolation point numbers for each uncertain variable; the interpolation points for each dimension are... Zeros of Chebyshev polynomials: ， Directly using Chebyshev fitting results in an exponential increase in the number of collocation points with increasing digits. Therefore, dimensionality reduction is employed to reduce the number of collocation points required for the approximate function: Performing dimensional decomposition and ignoring residual terms, we obtain: ， In the formula, yes The s-dimensional approximation function, It is a constant. for Dimensional uncertainty variables The influence component; Pick That is, considering only a one-dimensional approximation function, the midpoint of the uncertain variable is... As a reference point, we have: ， but The one-dimensional approximation function is expressed as: ， in, The reference only considers the i-th dimension. , It is a constant term; Similarly, The two-dimensional approximation function is expressed as: 。 8. The interval uncertainty damage identification method based on CBAM-Unet neural network according to claim 1, characterized in that, S130 includes: selecting the midpoint of the uncertain variable. The generated source dataset is used to train the CBAM-Unet network, and the resulting neural network model is used as a pre-trained model. Then, all parameters of the encoder part in the pre-trained model are frozen, and the parameters at the collocation points are used to generate the target dataset. The number of parameters in the target dataset is only one-tenth of that in the source dataset. The target dataset is used to perform transfer training on the pre-trained model to obtain a collocation model with higher accuracy.

9. A computing device, characterized in that, include: At least one processor and a memory storing program instructions; When the program instructions are read and executed by the processor, the computing device performs the method as described in any one of claims 1-8.

10. A readable storage medium storing program instructions, characterized in that, When the program instructions are read and executed by the computing device, the computing device performs the method as described in any one of claims 1-8.

Images