A bridge bending detection method based on a lightweight multi-scale sparse gating network

Abstract

The application discloses a bridge bending detection method based on a light-weight multi-scale sparse gating network, and relates to the field of bridge structure health monitoring.The method comprises the following steps: acquiring multi-mode optical fiber speckle images corresponding to different bending states of a bridge, and performing preprocessing to obtain standardized speckle images; based on the light-weight multi-scale sparse gating network, initial feature maps are obtained through initial convolution layers and maximum pooling layers; through a multi-level multi-scale feature fusion module, bending sensitive feature extraction and fusion are performed to obtain multi-scale fusion feature maps; a learnable sparse gating module is used for feature screening to obtain sparse enhanced feature maps; through a global context enhancement module, global dependency is modeled to obtain global enhanced feature maps; and the global enhanced feature maps are input into a double-task prediction module to output corresponding results.The method realizes synchronous, light-weight and high-precision detection of the bending degree and position of the bridge, and effectively decouples the multi-parameter cross-sensitivity problem.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring, and more specifically to a bridge bending detection method based on a lightweight multi-scale sparse gated network. Background Technology

[0002] Bridges are subjected to a combination of vehicle loads, wind loads, temperature changes, and earthquakes during their service life, making them highly susceptible to structural damage such as bending and cracking. Distributed sensing technology based on the speckle effect of multimode fiber has received widespread attention in the field of bridge structural health monitoring in recent years due to its simple structure, low cost, and ease of large-area deployment, becoming an important direction to replace traditional electrical sensors and fiber Bragg gratings (FBGs).

[0003] However, existing speckle detection methods based on multimode fiber mostly invert curvature by calculating the overall similarity of speckle images, discretizing continuously changing physical quantities into a finite number of categories. This results in limited resolution of the degree of curvature and makes it difficult to accurately distinguish minute differences in curvature. At the same time, these methods can usually only output single curvature information and cannot simultaneously identify the specific spatial location where the curvature occurs. When faced with the problem of multiple parameters such as curvature and location being sensitive to each other, it is difficult to achieve effective decoupling, which limits their practical application in multidimensional curvature detection of bridges.

[0004] In recent years, deep learning technology has been introduced into fiber speckle image analysis. However, existing solutions mostly use deep convolutional networks (such as ResNet-50 and VGG-16) for feature extraction. These networks have a huge number of parameters, high computational complexity, and demanding hardware resources, making them difficult to deploy on edge computing devices at bridge sites and unable to meet the operation and maintenance needs of real-time monitoring and rapid response. In addition, these general classification networks lack targeted design for bridge bending features and fail to effectively model multi-scale nonlinear changes in speckle images caused by bending, such as local texture perturbations, mesoscale patch diffusion, and large-scale energy distribution shifts, resulting in insufficient sensitivity to minute bending differences.

[0005] Furthermore, existing deep learning methods typically treat bending detection as a single classification or regression task, failing to simultaneously output both the degree of bending and the location of bending in a single inference. To obtain both parameters at the same time, multiple models or multi-stage processing procedures are often required, which not only increases system complexity but also easily leads to error accumulation. More importantly, the representation of the degree of bending and the location of bending in speckle images differs. The former is more related to statistical features and energy diffusion trends associated with deformation amplitude, while the latter is more related to structural shifts and regional response differences associated with spatial distribution. Existing models lack differentiated decoupled branch designs, resulting in overlapping representations between tasks, which further reduces the accuracy of multi-task recognition.

[0006] Therefore, how to design a bridge bending detection method based on a lightweight multi-scale sparse gating network that can effectively extract and enhance the multi-scale bending-sensitive features of speckle images, while achieving synchronous and decoupled output of bending degree and bending position, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a bridge bending detection method based on a lightweight multi-scale sparse gated network to solve the problems of multi-parameter cross-sensitivity, large computational load of deep convolutional networks making edge deployment difficult, and inability to simultaneously output the dual task results of bending degree and position in existing fiber optic sensing technology, thereby achieving efficient, accurate and real-time monitoring of bridge bending status.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A bridge bending detection method based on lightweight multi-scale sparse gated networks includes the following steps:

[0010] S1. Obtain multimode fiber speckle images corresponding to different bending states of the bridge, and preprocess them to obtain standardized speckle images.

[0011] S2. Input the standardized speckle image into a lightweight multi-scale sparse gated network, and process it through an initial convolutional layer and a max pooling layer to obtain an initial feature map;

[0012] S3. Through a series of multi-level, multi-scale feature fusion modules, the initial feature map is subjected to bending-sensitive feature extraction and fusion to obtain a multi-scale fused feature map;

[0013] S4. A learnable sparse gating module is used to perform feature filtering on the multi-scale fused feature map to obtain a sparse enhanced feature map.

[0014] S5. Input the sparse enhanced feature map into the global context enhancement module, model the global dependency relationship based on the channel attention mechanism, and obtain the global enhanced feature map;

[0015] S6. Input the global enhanced feature map into the dual-task prediction module. After shared feature extraction, input the curvature prediction branch and the curvature location prediction branch respectively, and output the curvature category result and the curvature location category result of the bridge.

[0016] Preferably, in S2, the initial convolutional layer includes a two-dimensional convolutional layer with a kernel size of 3×3, a stride of 1, a padding of 1, and an output channel number of 32, followed by a batch normalization layer and a ReLU activation function; the max pooling layer is a max pooling layer with a kernel size of 2×2 and a stride of 2.

[0017] Preferably, in S3, each multi-scale feature fusion module uses a parallel multi-branch structure to extract bending-sensitive features under different receptive fields;

[0018] The parallel multi-branch structure includes 1×1 convolutional branches, 3×3 convolutional branches, 5×5 convolutional branches, and dilated convolutional branches;

[0019] The output features of the four branches are concatenated along the channel dimension to obtain multi-scale concatenated features;

[0020] Then, a 1×1 depthwise separable convolution layer is used to uniformly map and fuse the multi-scale spliced ​​features, and the corresponding fused feature map is output.

[0021] Preferably, the 1×1 convolution branch, 3×3 convolution branch, and 5×5 convolution branch are all composed of depth-separable convolutions; the dilated convolution branch is composed of two layers of tandem 3×3 dilated depth-separable convolutions.

[0022] Preferably, in step S4, the feature selection by the learnable sparse gating module includes:

[0023] Multi-scale fusion feature maps Channel compression is performed sequentially through a first 1×1 convolutional layer, a ReLU activation layer, and a second 1×1 convolutional layer, and then a gated weight map G with a preset value range is generated by the Sigmoid activation function.

[0024] Calculate multi-scale fused feature maps Variance statistics in the spatial dimension, based on a preset baseline threshold. and variance response generation dynamic threshold ;

[0025] Compare the gating weight map G with the dynamic threshold Element-by-element comparisons are performed to obtain the binary sparse mask M;

[0026] The binary sparse mask M is combined with the multi-scale fused feature map. Element-wise multiplication yields a sparse enhancement feature map. .

[0027] Preferably, in step S5, the processing procedure of the global context enhancement module includes:

[0028] Sparse augmented feature map of input Each channel is compressed into a global descriptor by an adaptive global average pooling layer, resulting in a channel statistical vector z.

[0029] Then, a 1×1 convolutional layer is used for channel mapping, followed by a Sigmoid activation layer to generate channel enhancement coefficients. ;

[0030] The channel enhancement coefficient With sparse augmented feature maps Element-wise multiplication is performed, and residual connections and sparse enhancement feature maps are used. Add them together to obtain the global enhanced feature map. .

[0031] Preferably, in step S6, the processing procedure of the dual-task prediction module includes:

[0032] Global Enhanced Feature Map The input is a shared classification head, which includes an adaptive global average pooling layer, a flattening layer, and a first fully connected layer, and the output is a shared feature vector. ;

[0033] The shared feature vector Input the curvature prediction branch and the curvature location prediction branch respectively; the curvature prediction branch includes a second fully connected layer, whose output dimension is the number of curvature categories, and outputs the probability distribution of the bridge curvature belonging to each category through the Softmax function; the curvature location prediction branch includes a third fully connected layer, whose output dimension is the number of curvature location categories, and outputs the probability distribution of the bridge curvature location belonging to each category through the Softmax function.

[0034] Preferably, optimization is also performed using a combined loss function L during model training:

[0035]

[0036] in, Cross-entropy loss for the bending degree recognition task, λ is the cross-entropy loss for the curved location recognition task, where λ is the weighting coefficient.

[0037] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0038] 1. This method employs a series of multi-level, multi-scale feature fusion modules. Through parallel multi-branch structures of 1×1, 3×3, 5×5, and dilated convolutions, it can simultaneously capture multi-scale nonlinear changes in speckle images, such as local texture perturbations caused by bridge curvature, mid-scale patch diffusion, and large-scale energy center shift. Combined with the adaptive selection of key responses by a learnable sparse gating module, it enhances sensitive features related to curvature and location, suppresses environmental noise and redundant background perturbations, and thus has a high sensitivity to discriminate minute curvature differences and speckle feature differences at different spatial locations.

[0039] 2. By using a dual-task prediction module, high-level differentiated representation branches for bending degree and bending location are constructed on the basis of sharing the underlying speckle features. This enables the simultaneous output of the bridge bending degree category and bending location category. Compared with traditional methods that require separate modeling or multi-stage processing, this effectively reduces the cross sensitivity and representation aliasing problem between multiple parameters, and improves the system integration and result consistency of bridge multidimensional bending state detection.

[0040] 3. This method employs depthwise separable convolution to construct multi-scale branches and dilated convolutions, combined with a dynamic threshold filtering mechanism of learnable sparse gating modules. This significantly reduces the number of network parameters and computational complexity while maintaining high feature representation capabilities. Furthermore, it establishes dependencies between distant regions of speckle images through a global context enhancement module, compensating for the insufficiency of local receptive fields. The overall network architecture can be deployed on resource-constrained edge computing devices, enabling rapid inference and real-time monitoring of bridge bending states, thus meeting the engineering requirements for lightweight and low-latency detection in bridge maintenance sites. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0042] Figure 1 A flowchart of a bridge bending detection method based on a lightweight multi-scale sparse gating network provided in an embodiment of the present invention;

[0043] Figure 2 This is a diagram of a lightweight multi-scale sparse gating network framework provided in an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of an optical path system provided in an embodiment of the present invention;

[0045] Figure 4 The loss rate curve of the model training process provided in the embodiments of the present invention;

[0046] Figure 5 The accuracy curve of the model training process provided in the embodiments of the present invention;

[0047] Figure 6 A confusion matrix diagram for classifying the degree of curvature provided in an embodiment of the present invention;

[0048] Figure 7 A confusion matrix diagram for classifying bending positions provided in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] like Figure 1 As shown, this embodiment provides a bridge bending detection method based on a lightweight multi-scale sparse gating network, including the following steps:

[0051] S1. Obtain multimode fiber speckle images corresponding to different bending states of the bridge, and preprocess them to obtain standardized speckle images.

[0052] S2. Input the standardized speckle image into a lightweight multi-scale sparse gated network, and process it through an initial convolutional layer and a max pooling layer to obtain an initial feature map; as shown Figure 2 As shown, the lightweight multi-scale sparse gating network specifically includes: an initial convolutional layer and a max pooling layer connected in sequence, a two-level cascaded bridge curvature feature-guided multi-scale feature fusion module, a learnable sparse gating module, a global context enhancement module, and a dual-task prediction module.

[0053] S3. Through a series of multi-level, multi-scale feature fusion modules, the initial feature map is subjected to bending-sensitive feature extraction and fusion to obtain a multi-scale fused feature map;

[0054] S4. A learnable sparse gating module is used to perform feature filtering on the multi-scale fused feature map to obtain a sparse enhanced feature map.

[0055] S5. Input the sparse enhanced feature map into the global context enhancement module, model the global dependency relationship based on the channel attention mechanism, and obtain the global enhanced feature map;

[0056] S6. Input the global enhanced feature map into the dual-task prediction module. After shared feature extraction, input the curvature prediction branch and the curvature location prediction branch respectively, and output the curvature category result and the curvature location category result of the bridge.

[0057] This method extracts multi-scale variation features such as local texture, mesoscale patches, and large-scale energy distribution caused by bending in speckle images through a multi-scale feature fusion module guided by bridge bending features. It then uses a learnable sparse gating module to adaptively filter key responses to suppress redundant noise. A global context enhancement module establishes long-distance dependencies, and finally a dual-task prediction module synchronously outputs the degree of bending and the category of bending location. This effectively reduces the cross sensitivity of multiple parameters, improves the detection accuracy and stability of the multi-dimensional bending state of the bridge, and maintains a lightweight structure to meet the needs of real-time on-site monitoring.

[0058] The following provides further explanation of each step in the above method and its related technical features;

[0059] In this embodiment, S1, multimode fiber speckle images corresponding to different bending states of the bridge are acquired and preprocessed to obtain standardized speckle images.

[0060] In this embodiment, the data acquisition of the bridge's bending state is carried out using the following method: Figure 3 The optical path system shown is as follows: the beam emitted by the semiconductor laser is collimated into parallel light by a collimating lens and injected into the incident end of the multimode fiber through a single-mode fiber; the two ends of the multimode fiber are fixed by clamps, and its middle section is wound around cylinders of different diameters to form a preset bending radius (no bending, 1.25 cm, 2.5 cm, 3.0 cm, 4.3 cm), while the bending point position can be moved along the fiber axis (0 m, 1 m, 2 m, 3 m, 4 m); the speckle light field carrying bending information output from the fiber output end is directly projected onto the photosensitive surface of the CCD / CMOS image sensor, and the sensor converts the light intensity distribution into a digital image signal and transmits it to the computer. This optical path structure is simple, does not require a complex interferometer, and can stably acquire high-resolution speckle images under different degrees and positions of bending.

[0061] During preprocessing, the region of interest in the color speckle image is first cropped and uniformly scaled to 256×256 pixels, then converted into a single-channel grayscale image. Subsequently, the pixel values ​​are linearly mapped from [0,255] to the [-1,1] interval, and finally converted into PyTorch tensor format. This normalization process can eliminate the illumination differences between different acquisition batches and enhance the convergence stability of model training.

[0062] In this embodiment S2, the standardized speckle image is input into a lightweight multi-scale sparse gating network, and the initial feature map is obtained by processing it through an initial convolutional layer and a max pooling layer.

[0063] The initial convolutional layer has the following parameters: 3×3 kernel size, stride 1, padding 1, and 32 output channels, followed by a batch normalization layer and a ReLU activation function; the max pooling layer uses a 2×2 kernel size and a stride 2, which halves the feature map space size from 256×256 to 128×128; after this processing, the initial feature map size is 128×128×32.

[0064] This design reduces the computational load of subsequent modules while preserving effective spatial information. Actual measurements show that the total number of network parameters can be controlled within 1.5M, meeting the real-time inference requirements at the edge.

[0065] In this embodiment S3, the initial feature map is subjected to bending-sensitive feature extraction and fusion through a series of multi-level multi-scale feature fusion modules to obtain a multi-scale fused feature map; wherein, each level of multi-scale feature fusion module adopts a parallel multi-branch structure to extract bending-sensitive features under different receptive fields.

[0066] The parallel multi-branch structure includes 1×1 convolutional branches, 3×3 convolutional branches, 5×5 convolutional branches, and dilated convolutional branches;

[0067] The output features of the four branches are concatenated along the channel dimension to obtain multi-scale concatenated features;

[0068] Then, a 1×1 depthwise separable convolution layer is used to uniformly map and fuse the multi-scale spliced ​​features, and the corresponding fused feature map is output.

[0069] Furthermore, the 1×1 convolution branch, 3×3 convolution branch, and 5×5 convolution branch are all composed of depth-separable convolutions; the dilated convolution branch is composed of two layers of cascaded 3×3 dilated depth-separable convolutions.

[0070] This embodiment sets up a two-stage cascaded multi-scale feature fusion module: the first stage maps the input channels from 32 to 64, and the second stage maps them from 64 to 128, with both stages maintaining a spatial size of 128×128. In each stage, the number of output channels for the four parallel branches are 16 or 32, respectively. The dilated convolution branch consists of two layers of 3×3 depth-separable convolutions cascaded together, with a dilation rate of 3 and a padding rate of 3. This structure expands the receptive field to 13×13 without increasing the number of parameters, effectively capturing the speckle energy center shift caused by curvature. After the outputs of the four branches are concatenated in the channel dimension, they are fused by a 1×1 depth-separable convolution to obtain the multi-scale fused feature map of this stage. After two stages of processing, the network can simultaneously retain fine-grained texture and global energy distribution information.

[0071] In this embodiment S4, a learnable sparse gating module is used to perform feature filtering on the multi-scale fused feature map to obtain a sparse enhanced feature map; wherein, the feature filtering performed by the learnable sparse gating module includes:

[0072] Multi-scale fusion feature maps Channel compression is performed sequentially through a first 1×1 convolutional layer, a ReLU activation layer, and a second 1×1 convolutional layer, and then a gated weight map G with a preset value range is generated by the Sigmoid activation function.

[0073] Calculate multi-scale fused feature maps Variance statistics in the spatial dimension, based on a preset baseline threshold. and variance response generation dynamic threshold ;

[0074] Compare the gating weight map G with the dynamic threshold Element-by-element comparisons are performed to obtain the binary sparse mask M;

[0075] The binary sparse mask M is combined with the multi-scale fused feature map. Element-wise multiplication yields a sparse enhancement feature map. .

[0076] In this embodiment, the channel compression path of the learnable sparse gating module is as follows: the first 1×1 convolution compresses 128 channels to 16, ReLU is activated, the second 1×1 convolution maps 16 channels to 1, and the gating weight map G is output via Sigmoid; during dynamic threshold generation, the input feature map is first calculated. The variance statistic in the spatial dimension is then weighted with a preset base threshold of 0.3 to obtain the dynamic threshold. , where σ is the spatial variance and α is the empirical coefficient 0.1;

[0077] Further, G and The binary mask M is obtained by element-wise comparison, and M is compared with... After element-wise multiplication, approximately 40% to 60% of the responses in the feature map are retained, while the rest are suppressed. This sparse filtering can reduce the computational cost of the subsequent global context enhancement module and has no negative impact on the accuracy of curvature classification.

[0078] In embodiment S5, the sparse enhanced feature map is input to the global context enhancement module, and global dependencies are modeled based on the channel attention mechanism to obtain the global enhanced feature map; wherein, the processing procedure of the global context enhancement module includes:

[0079] Sparse augmented feature map of input Each channel is compressed into a global descriptor by an adaptive global average pooling layer, resulting in a channel statistical vector z.

[0080] Then, a 1×1 convolutional layer is used for channel mapping, followed by a Sigmoid activation layer to generate channel enhancement coefficients. ;

[0081] The channel enhancement coefficient With sparse augmented feature maps Element-wise multiplication is performed, and residual connections and sparse enhancement feature maps are used. Add them together to obtain the global enhanced feature map. .

[0082] In this embodiment, the global context enhancement module first processes the input feature map. Adaptive global average pooling is performed to obtain a 128-dimensional channel statistical vector z; then, 128-dimensional channel enhancement coefficients are generated through 1×1 convolution and sigmoid. ;Will After broadcasting spatial dimensions and Multiply element by element, then multiply by the original. Adding the residuals, we get This residual design ensures smooth gradient backpropagation and avoids gradient vanishing in deep networks. After processing by this module, the feature responses of different channels are recalibrated: the energy diffusion channel, which is strongly correlated with the curvature, is enhanced, while the channel, which is correlated with background noise, is suppressed, thereby improving the signal-to-noise ratio of the input features of the dual-task branch.

[0083] In embodiment S6, the global enhanced feature map is input to the dual-task prediction module. After shared feature extraction, it is input to the curvature prediction branch and the curvature location prediction branch respectively, and the curvature category result and the curvature location category result of the bridge are output. The processing of the dual-task prediction module includes:

[0084] Global Enhanced Feature Map The input is a shared classification head, which includes an adaptive global average pooling layer, a flattening layer, and a first fully connected layer, and the output is a shared feature vector. ;

[0085] The shared feature vector Input the curvature prediction branch and the curvature location prediction branch respectively; the curvature prediction branch includes a second fully connected layer, whose output dimension is the number of curvature categories, and outputs the probability distribution of the bridge curvature belonging to each category through the Softmax function; the curvature location prediction branch includes a third fully connected layer, whose output dimension is the number of curvature location categories, and outputs the probability distribution of the bridge curvature location belonging to each category through the Softmax function.

[0086] Furthermore, this also includes using a combined loss function L for optimization during model training:

[0087]

[0088] in, Cross-entropy loss for the bending degree recognition task, λ is the cross-entropy loss for the curved location recognition task, where λ is the weighting coefficient.

[0089] During training, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio; for samples without curvature, their position label is set to -1. Automatic masking is applied; the combined loss function has λ=0.5, and the optimization process uses the AdamW optimizer with an initial learning rate of 0.001 and a weight decay coefficient of 0.5. The Reduce LR On Plateau strategy is used to dynamically adjust the learning rate. This strategy uses the accuracy of the curvature recognition on the validation set as the monitoring indicator. When this indicator does not improve for 5 consecutive training cycles, the learning rate is reduced by a scaling factor of 0.5. This training strategy ensures the balance between shared representation and differentiated modeling between the two tasks, effectively avoiding mutual interference between tasks.

[0090] In actual bridge monitoring, pre-laid multimode optical fibers are coupled to key parts of the bridge structure. Stable optical signals are injected into the optical fibers through semiconductor lasers, and the CCD / CMOS image sensor at the output end of the optical fibers acquires speckle images in real time under the current state. The acquired raw speckle images are processed through the same preprocessing steps as in the training phase to form standardized speckle images.

[0091] Subsequently, the standardized speckle image is fed into a pre-trained lightweight multi-scale sparse gating network. The model sequentially passes through initial convolution and pooling, two-level multi-scale feature fusion, learnable sparse gating filtering, and global context enhancement. Finally, the dual-task prediction module outputs two detection results at once: the degree of curvature category, such as 2.5 cm or no curvature, and the curvature location category, such as 2 m or 3 m. The entire inference process can be completed within seconds on an edge computing device without relying on cloud computing power, realizing real-time, synchronous, and multi-dimensional detection of the bridge's curvature status, and providing bridge maintenance personnel with an immediate basis for structural health status assessment.

[0092] Furthermore, to verify the effectiveness of the method proposed in this invention, model training and testing were performed on the aforementioned dataset of multiple multimode fiber speckle images; such as Figure 4 As shown, the model's loss curve decreases smoothly during training, and the validation set loss remains highly consistent with the training set loss, indicating that the designed lightweight multi-scale sparse gating network has good convergence characteristics on the bridge speckle image dataset and can stably learn the nonlinear mapping from speckle pattern to bending state; Figure 5As shown, the accuracy of the model in recognizing the degree of curvature steadily increases during training, and the accuracy of the validation set tends to converge and remain stable in the later stage of training. This indicates that, under the role of feature extraction and screening mechanisms, the network can fully capture the discriminative differences of speckle images under different degrees of curvature and has strong generalization ability, thus providing a guarantee for high-precision recognition in actual bridge monitoring scenarios.

[0093] like Figure 6 As shown, all non-zero elements in the bending degree classification confusion matrix are concentrated on the diagonal. All five categories (no bending, 1.25 cm, 2.5 cm, 3 cm, and 4.3 cm) achieved completely correct classification with zero misclassification, demonstrating the superior performance of the proposed method in distinguishing minute differences in bending amplitude. This method can meet the needs of fine identification of different bending levels in bridge structural health monitoring. Figure 7 As shown, apart from a very small number of samples being misclassified at individual locations in the confusion matrix of bending positions, the categories of the remaining locations are all correctly identified, and the overall recognition effect is ideal. The confusion between a small number of adjacent locations reflects that the speckle response has a certain similarity in the spatial neighborhood, but it does not affect the overall detection accuracy, further verifying the engineering applicability of this method in decoupling and synchronous output of multidimensional bending parameters.

[0094] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0095] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bridge bending detection method based on lightweight multi-scale sparse gating networks, characterized in that, Includes the following steps: S1. Obtain multimode fiber speckle images corresponding to different bending states of the bridge, and preprocess them to obtain standardized speckle images. S2. Input the standardized speckle image into a lightweight multi-scale sparse gated network, and process it through an initial convolutional layer and a max pooling layer to obtain an initial feature map; S3. Through a series of multi-level, multi-scale feature fusion modules, the initial feature map is subjected to bending-sensitive feature extraction and fusion to obtain a multi-scale fused feature map; S4. A learnable sparse gating module is used to perform feature filtering on the multi-scale fused feature map to obtain a sparse enhanced feature map; wherein, the feature filtering performed by the learnable sparse gating module includes: Multi-scale fusion feature maps Channel compression is performed sequentially through a first 1×1 convolutional layer, a ReLU activation layer, and a second 1×1 convolutional layer. Then, a gated weight map G with a preset value range is generated using a Sigmoid activation function. Multi-scale fusion feature maps are then calculated. Variance statistics in the spatial dimension, based on a preset baseline threshold. and variance response generation dynamic threshold The gating weight map G is compared with the dynamic threshold. Element-by-element comparisons are performed to obtain a binary sparse mask M; the binary sparse mask M is then fused with a multi-scale feature map. Element-wise multiplication yields a sparse enhancement feature map. ; S5. Input the sparse enhanced feature map into the global context enhancement module, model the global dependency relationship based on the channel attention mechanism, and obtain the global enhanced feature map; S6. Input the global enhanced feature map into the dual-task prediction module. After shared feature extraction, input the curvature prediction branch and the curvature location prediction branch respectively, and output the curvature category result and the curvature location category result of the bridge.

2. The bridge bending detection method based on a lightweight multi-scale sparse gating network according to claim 1, characterized in that, In S2, the initial convolutional layer includes a two-dimensional convolutional layer with a kernel size of 3×3, a stride of 1, a padding of 1, and 32 output channels, followed by a batch normalization layer and a ReLU activation function; the max pooling layer is a max pooling layer with a kernel size of 2×2 and a stride of 2.

3. The bridge bending detection method based on a lightweight multi-scale sparse gating network according to claim 1, characterized in that, In S3, each multi-scale feature fusion module uses a parallel multi-branch structure to extract bending-sensitive features under different receptive fields. The parallel multi-branch structure includes 1×1 convolutional branches, 3×3 convolutional branches, 5×5 convolutional branches, and dilated convolutional branches; The output features of the four branches are concatenated along the channel dimension to obtain multi-scale concatenated features; Then, a 1×1 depthwise separable convolution layer is used to uniformly map and fuse the multi-scale spliced ​​features, and the corresponding fused feature map is output.

4. The bridge bending detection method based on a lightweight multi-scale sparse gating network according to claim 3, characterized in that, The 1×1 convolution branch, 3×3 convolution branch, and 5×5 convolution branch are all composed of depth-separable convolutions; the dilated convolution branch is composed of two layers of cascaded 3×3 dilated depth-separable convolutions.

5. The bridge bending detection method based on a lightweight multi-scale sparse gating network according to claim 1, characterized in that, In S5, the processing procedure of the global context enhancement module includes: Sparse augmented feature map of input Each channel is compressed into a global descriptor by an adaptive global average pooling layer, resulting in a channel statistical vector z. Then, a 1×1 convolutional layer is used for channel mapping, followed by a Sigmoid activation layer to generate channel enhancement coefficients. ; The channel enhancement coefficient With sparse augmented feature maps Element-wise multiplication is performed, and residual connections and sparse enhancement feature maps are used. Add them together to obtain the global enhanced feature map. .

6. The bridge bending detection method based on a lightweight multi-scale sparse gating network according to claim 1, characterized in that, In step S6, the processing procedure of the dual-task prediction module includes: Global Enhanced Feature Map The input is a shared classification head, which includes an adaptive global average pooling layer, a flattening layer, and a first fully connected layer, and the output is a shared feature vector. ; The shared feature vector Input the curvature prediction branch and the curvature location prediction branch respectively; the curvature prediction branch includes a second fully connected layer, whose output dimension is the number of curvature categories, and outputs the probability distribution of the bridge curvature belonging to each category through the Softmax function; the curvature location prediction branch includes a third fully connected layer, whose output dimension is the number of curvature location categories, and outputs the probability distribution of the bridge curvature location belonging to each category through the Softmax function.

7. The bridge bending detection method based on a lightweight multi-scale sparse gating network according to claim 1, characterized in that, This also includes using a combined loss function L for optimization during model training: ； in, Cross-entropy loss for the bending degree recognition task, λ is the cross-entropy loss for the curved location recognition task, where λ is the weighting coefficient.

Images