A deep learning method for parallel mixed identification of multiple diseases of a bridge
By combining camera and drone photography with deep learning methods, high-dimensional semantic features of bridge defects are extracted and processed in parallel. This solves the problems of difficulty in capturing long-distance dependencies and inference delays in existing technologies, and achieves high efficiency, robustness and generalization ability in bridge defect identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 淳安县公路服务中心
- Filing Date
- 2026-06-15
- Publication Date
- 2026-07-31
AI Technical Summary
Existing convolutional neural networks struggle to capture the long-distance spatial dependencies of bridge defects, recurrent neural networks suffer from high inference latency and cannot be parallelized for acceleration, and multi-label classification methods are prone to gradient conflicts and wasted computational resources.
Image data is acquired by a combination of camera and drone photography. High-dimensional semantic feature maps are extracted through a pre-trained convolutional neural network. The feature sequence is enhanced by row-first slicing and parallel time-sharing modules. Finally, an adaptive attention aggregation module is used to predict the probability distribution of diseases.
It effectively captures the long-distance dependence and local continuity of bridge defects, improves identification robustness, adapts to different lighting and bridge structure types, and adapts to the diversity of actual engineering scenarios.
Smart Images
Figure CN122493340A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and deep learning, and in particular to a deep learning method for parallel hybrid identification of multiple defects in bridges. Background Technology
[0002] As a core component of transportation hubs, the structural integrity of bridges is directly related to public safety. Currently, with the increasingly serious problem of infrastructure aging, defects such as concrete cracking, spalling, steel corrosion, and water seepage / efflorescence are on the rise. Given the limitations of traditional manual inspections, including low efficiency, subjective results, and high personnel safety risks, the industry urgently needs more advanced inspection methods. However, translating relevant theories into practical productivity still faces the following key technological bottlenecks: While existing convolutional neural networks (CNNs) can effectively extract local textures (such as crack edges), they are limited by their receptive field and have difficulty capturing long-distance spatial dependencies of diseases (such as the global direction of penetrating cracks). Traditional Transformer or RNN architectures can model global dependencies, but they often ignore the microscopic local continuity of diseases.
[0003] To capture sequence information, some studies have introduced recurrent neural networks (RNN / GRU); however, the inherent serial computing characteristics of RNNs prevent them from being accelerated by large-scale parallelism using GPUs, resulting in high inference latency and making it difficult to meet the needs of real-time inspection of large-scale bridge groups.
[0004] Bridge defects often present multiple types of concurrent occurrences (such as spalling accompanied by exposed rebar); existing multi-label classification methods often treat labels as independent individuals or use complex graph structures for modeling, which can easily lead to gradient conflicts or waste of computational resources. Summary of the Invention
[0005] This invention provides a deep learning method for parallel hybrid identification of multiple bridge defects, which overcomes the shortcomings of existing technologies such as ignoring the microscopic local continuity of bridge defects, high inference latency, and gradient conflicts or waste of computational resources due to the use of complex graph structure modeling.
[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A deep learning method for parallel hybrid identification of multiple bridge defects includes: S1. Acquire bridge appearance image data through a combination of camera and drone shooting and preprocess the data to obtain preprocessed images; and construct a multi-label image classification dataset based on the preprocessed images. S2. Based on the multi-label image classification dataset, feature extraction is performed using a pre-trained convolutional neural network to obtain a high-dimensional semantic feature map; S3. Perform row-first slicing on the high-dimensional semantic feature map to obtain feature vectors for several rows of spatial locations; flatten and recombine the feature vectors for each row of spatial locations to obtain the input sequence; S4. Based on the input sequence, the enhanced feature sequence is obtained through the parallel time-mapping module; S5. Based on the enhanced feature sequence, the probability distribution of various bridge defects is obtained through the adaptive attention aggregation module.
[0007] Furthermore, the specific processing of step S3 is as follows: S31. Obtain the height H of the high-dimensional semantic feature map, and divide the high-dimensional semantic feature map along the height of the high-dimensional semantic feature map to obtain the feature vectors of H row spatial positions; S32. Based on the feature vector of the row space position, expand it using the spatial dimension and the channel dimension to obtain the one-dimensional feature vector of the row space position, expressed as:
[0008] In the formula, For the first A one-dimensional feature vector of each row's spatial location; For the first Feature vectors of each row's spatial location; To expand the operation; The width of the high-dimensional semantic feature map; The number of channels in the high-dimensional semantic feature map; The index of the feature vector at the row space location; It is the set of real numbers; S33. Arrange the one-dimensional feature vectors of the row space positions according to the spatial order of the high-dimensional semantic feature map to obtain the input sequence, expressed as:
[0009] In the formula, The input sequence is .
[0010] Furthermore, the parallel time-mapping module includes a normalization layer, a local mixing branch, a global mixing branch that runs parallel to the local mixing branch, and a fusion layer; The normalization layer obtains a standardized feature sequence based on the input sequence through a layer normalization operation; The local hybrid branch obtains local continuous features through depthwise separable one-dimensional convolution based on the standardized feature sequence; The global hybrid branch obtains global context features based on the standardized feature sequence through a multi-head self-attention mechanism; The fusion layer obtains fused features by fusing local continuous features and global context features; based on the fused features, an enhanced feature sequence is obtained through layer normalization.
[0011] Furthermore, the specific steps of the local hybrid branch are as follows: S411. Based on the standardized feature sequence, the transformed input features are obtained through dimensional transformation; S412. Based on the transformed input features, intermediate features are obtained through depthwise separable one-dimensional convolution, expressed as:
[0012] In the formula, As an intermediate feature; To transform the input features; For depthwise separable one-dimensional convolution; S413. Based on the intermediate features and the standardized feature sequence, local continuous features are obtained through residual connection, expressed as:
[0013] In the formula, It is a locally continuous feature; This is a standardized feature sequence.
[0014] Furthermore, the specific steps of the global hybrid branch are as follows: S421. Map the standardized feature sequences to obtain the query matrix, key matrix, and value matrix; S422. Calculate the scaled dot product attention based on the query matrix, key matrix, and value matrix. The expression is:
[0015] In the formula, To scale the dot product attention; For query matrix; The key matrix; It is a value matrix; The dimension of the query vector for each head; S423. Based on the scaled dot product attention, calculate the attention of each head using a multi-head mechanism, expressed as:
[0016] In the formula, For the first The size of the attention; For the first The query weight matrix for each head; For the first The attention head's key weight matrix; For the first The attention value weight matrix of each head; S424. Concatenate the attention of all heads to obtain the global context feature, expressed as:
[0017] In the formula, For global context features; This is the projected weight matrix for the multi-head attention mechanism; The number of heads.
[0018] Furthermore, the expression for fusing local continuous features and global contextual features is as follows:
[0019] In the formula, Features of fusion; These are learnable fusion weights.
[0020] Furthermore, the specific execution steps of the adaptive attention aggregation module are as follows: S51. Based on the enhanced feature sequence, obtain the global feature vector through global average pooling; S52. Set up independent binary classifiers for K disease categories, and input the global feature vector into each independent binary classifier to obtain the probability value of the existence of each disease category; S53. Based on the probability values of each disease category, obtain the Multi-Hot probability vector to obtain the probability distribution of the disease.
[0021] Beneficial effects: The present invention provides a deep learning method for parallel hybrid identification of multiple bridge defects. After extracting high-dimensional semantic feature maps through a pre-trained convolutional neural network, a row-first slicing strategy is used to vectorize and recombine spatial location information and semantic features into a sequence. This preserves the relative spatial layout of defects in the image, avoids the loss of location information caused by simple global pooling, and is beneficial for distinguishing defects that are spatially similar but of different categories. By enhancing the input sequence through the parallel time-mixing module, long-distance dependencies and local contextual information in the feature sequence can be captured simultaneously, enabling the modeling of the interaction relationship between different disease features, thereby improving the robustness of recognition in complex backgrounds, overlapping diseases, or mutual occlusion. Based on a pre-trained convolutional neural network as the feature extraction backbone, it can inherit the general visual representation learned on large natural image datasets. It is then fine-tuned using a multi-disease dataset of bridges, so that the method has good generalization performance for different lighting, shooting angles and bridge structure types (such as beam bridges, arch bridges and cable-stayed bridges), and can adapt to the diversity of actual engineering scenarios. Attached Figure Description
[0022] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating the deep learning method for parallel hybrid recognition according to the present invention. Figure 2 This is a schematic diagram illustrating the process of converting a two-dimensional feature map into a one-dimensional sequence in an embodiment of the present invention; Figure 3 This is a schematic diagram of the parallel timing mixing module in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the adaptive attention aggregation module in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.
[0025] This embodiment provides a deep learning method for parallel hybrid identification of multiple bridge defects, such as... Figure 1 As shown, it includes: S1. Acquire bridge appearance image data through a combination of camera and drone shooting and preprocess the data to obtain preprocessed images; and construct a multi-label image classification dataset based on the preprocessed images. Specifically, the preprocessing process is as follows: S11. Adjust the resolution of the bridge appearance image data to a uniform standard. This yields an image with uniform resolution; In this embodiment, the resolution is uniformly adjusted to 512×512; S12. Standardize the image with uniform resolution to obtain a standardized image; The standardization process involves subtracting the mean of the ImageNet dataset from the image with uniform resolution and then dividing by the standard deviation of the ImageNet dataset to obtain a standardized image. The mean and standard deviation of the ImageNet dataset are parameters of publicly available image datasets, and will not be elaborated here; S13. Based on the standardized image, obtain the flipped image by randomly flipping it horizontally; The random horizontal flipping involves flipping the standardized image horizontally with a preset probability to enhance the model's robustness to the distribution of bridge defects in different directions. S14. Based on the flipped image, obtain the cropped image by random cropping; The random cropping involves randomly selecting a certain proportion of the area in the flipped image, such as within the range of 80% to 100%, cropping the area, and then scaling it to a uniform input size, thereby simulating different shooting angles and local observation scenarios. S15. Based on the cropped image, obtain a jittered image through color jittering; The color jittering involves randomly perturbing the brightness, contrast, saturation, and hue of the cropped image to adapt the model to different lighting conditions and environmental changes.
[0026] S2. Based on the multi-label image classification dataset, feature extraction is performed using a pre-trained convolutional neural network to obtain a high-dimensional semantic feature map; Specifically, the pre-trained convolutional neural network preferably uses CNN as the backbone network, and removes the last fully connected layer and global average pooling layer of CNN, retaining the convolutional layer structure; The specific steps for feature extraction using a pre-trained convolutional neural network are as follows: S21. Perform large-size convolution and downsampling operations on the images in the multi-label image classification dataset to obtain low-level edge and texture features; S22. Based on the low-level edge and texture features, the network depth is gradually increased by multiple sets of residual blocks, and the receptive field is continuously expanded to obtain structural information and features of local regions. S23. Based on structural information and the characteristics of local regions, global semantic information is integrated under a larger receptive field to generate a deep feature map with high semantic expressive power, namely a high-dimensional semantic feature map.
[0027] In this embodiment, the image size is analyzed during the feature extraction process using a pre-trained convolutional neural network as follows: Assuming the input image is downsampled by 32 times (standard stride), if the input is 512×512, the spatial size of the output feature map X is 16×16. As the network depth increases, the number of channels in the last stage reaches 2048, constructing a deep feature representation with a dimension of 16×16×2048. This feature map is located at the top of the network and has a very large receptive field, which can capture the global texture distribution of cracks and the abstract semantic features of peeling disease.
[0028] S3. Perform row-first slicing on the high-dimensional semantic feature map to obtain feature vectors for several rows of spatial locations; flatten and recombine the feature vectors for each row of spatial locations to obtain the input sequence; Preferred, such as Figure 2 As shown, the specific processing of step S3 is as follows: S31. Obtain the height H of the high-dimensional semantic feature map, and divide the high-dimensional semantic feature map along the height of the high-dimensional semantic feature map to obtain the feature vectors of H row spatial positions; In this embodiment, the height of the high-dimensional semantic feature map is 16, ultimately yielding feature vectors for 16 row spatial positions; and let the first... The feature vector of each row spatial location is ; S32. Based on the feature vector of the row space position, expand it using the spatial dimension and the channel dimension to obtain the one-dimensional feature vector of the row space position, expressed as:
[0029] In the formula, For the first A one-dimensional feature vector of each row's spatial location; For the first Feature vectors of each row's spatial location; To expand the operation; The width of the high-dimensional semantic feature map; The number of channels in the high-dimensional semantic feature map; The index of the feature vector at the row space location; It is the set of real numbers; In this embodiment, the feature vector of each row spatial position corresponds to a row of the feature map, and its size is 1×16×2048. After flattening the feature vector of the row spatial position, a one-dimensional feature vector is formed. The flattening operation involves arranging all elements in a two-dimensional vector sequentially into a one-dimensional vector, either in width-first or channel-first order. For example: ; S33. Arrange the one-dimensional feature vectors of the row space positions according to the spatial order of the high-dimensional semantic feature map to obtain the input sequence, expressed as:
[0030] In the formula, The input sequence is .
[0031] Specifically, the dimension calculation in step S3 is as follows: Each one-dimensional feature vector has 16 (width) × 2048 (channels) = 32768 elements; these 16 row vectors are arranged in order from top to bottom to form the input sequence; this reconstruction method transforms the vertical positional relationship in space into the temporal step relationship of the sequence, enabling the model to use the sequence model to capture the continuity of the disease between the upper and lower rows.
[0032] S4. Based on the input sequence, the enhanced feature sequence is obtained through the parallel time-mapping module; Preferred, such as Figure 3 As shown, the parallel time-mapping module includes a normalization layer, a local mixing branch, a global mixing branch that runs parallel to the local mixing branch, and a fusion layer; The normalization layer obtains a standardized feature sequence based on the input sequence through a layer normalization operation; The local hybrid branch obtains local continuous features through depthwise separable one-dimensional convolution based on the standardized feature sequence; The global hybrid branch obtains global context features based on the standardized feature sequence through a multi-head self-attention mechanism; The fusion layer obtains fused features by fusing local continuous features and global context features; based on the fused features, an enhanced feature sequence is obtained through layer normalization.
[0033] Preferably, the specific steps of the local hybrid branch are as follows: S411. Based on the standardized feature sequence, the transformed input features are obtained through dimensional transformation; Specifically, the dimensionality transformation is performed on the standardized feature sequence. Adding batch dimension yields multidimensional features. ,in, To standardize the length of the feature sequence, To standardize the number of channels in the feature sequence; based on multidimensional features The dimensionality order is rearranged using PyTorch or TensorFlow frameworks to obtain transformed input features. The PyTorch and TensorFlow frameworks mentioned are existing technologies and will not be described in detail here. S412. Based on the transformed input features, intermediate features are obtained through depthwise separable one-dimensional convolution, expressed as:
[0034] In the formula, As an intermediate feature; To transform the input features; For depthwise separable one-dimensional convolution; Specifically, in the depthwise separable one-dimensional convolution, a single position is calculated in the sequence dimension. The expression for the process of local output features is:
[0035] In the formula, For the output sequence at position eigenvalues; The kernel size; For the convolution kernel at the offset Learnable weights at each location; For the input sequence at position eigenvalues; S413. Based on the intermediate features and the standardized feature sequence, local continuous features are obtained through residual connection, expressed as:
[0036] In the formula, It is a locally continuous feature; This is a standardized feature sequence.
[0037] Preferably, the specific steps of the global hybrid branch are as follows: S421. Map the standardized feature sequences to obtain the query matrix, key matrix, and value matrix; Specifically, the expression for the mapping is:
[0038] In the formula, For standardized feature sequences; To generate the query matrix The learnable weight matrix; To generate the key matrix The learnable weight matrix; To generate a value matrix The learnable weight matrix; S422. Calculate the scaled dot product attention based on the query matrix, key matrix, and value matrix. The expression is:
[0039] In the formula, To scale the dot product attention; For query matrix; The key matrix; It is a value matrix; The dimension of the query vector for each head; S423. Based on the scaled dot product attention, calculate the attention of each head using a multi-head mechanism, expressed as:
[0040] In the formula, For the first The size of the attention; For the first The query weight matrix for each head; For the first The attention head's key weight matrix; For the first The attention value weight matrix of each head; S424. Concatenate the attention of all heads to obtain the global context feature, expressed as:
[0041] In the formula, For global context features; This represents the projected weight matrix for the multi-head attention mechanism; The number of heads.
[0042] Preferably, the expression for fusing local continuous features and global contextual features is as follows:
[0043] In the formula, Features of fusion; These are learnable fusion weights.
[0044] Specifically, based on the fused features, and through layer normalization, the expression for the enhanced feature sequence is obtained as follows:
[0045] In the formula, To enhance the feature sequence; This is a layer normalization operation.
[0046] S5. Based on the enhanced feature sequence, the probability distribution of various bridge defects is obtained through the adaptive attention aggregation module.
[0047] Preferably, the specific execution steps of the adaptive attention aggregation module are as follows: S51. Based on the enhanced feature sequence, obtain the global feature vector through global average pooling; Specifically, such as Figure 4 As shown, the specific steps to obtain the global feature vector are as follows: S511, Obtain the enhanced feature sequence; S512, Introduce a learnable global query vector. It is used to assess the importance of sequence features; S513. Obtain the unnormalized attention score by calculating the similarity between the global query vector and each feature in the enhanced feature sequence. ,in, For the first Unnormalized attention score, To enhance the first feature sequence One characteristic: S514. Normalize all scores using the Softmax function to obtain the attention weights. ,in, For the first Each attention weight reflects the importance of each position in the sequence to the final classification task. S515. Calculate the global feature vector by weighting the sequence features according to the attention weights. .
[0048] S52. Set up independent binary classifiers for K disease categories, and input the global feature vector into each independent binary classifier to obtain the probability value of the existence of each disease category; In this embodiment, the types of defects include cracks, peeling, corrosion, efflorescence, and exposed rebar; K is set to 5. Specifically, the expression for the probability value of each disease category is as follows:
[0049] In the formula, For the first Disease probability value; Use the Sigmoid activation function; For the first Weight vectors for disease types; For the first Bias terms for disease types; For transpose; An index for diseases; This is the global feature vector; S53. Based on the probability values of each disease category, obtain the Multi-Hot probability vector to obtain the probability distribution of the disease.
[0050] In this embodiment, the binary cross-entropy loss function is used to calculate the loss for each category, and the sum is used as the total loss for backpropagation; the expression for the total damage is:
[0051] In the formula, Total loss; The number of samples; For sample index; The number of diseases; Dynamic category weights; The raw logits output by the model; For the first The sample at the th The actual Multi-Hot annotation value on the disease type; The expression for the dynamic category weight is:
[0052] In the formula, Not including the first Number of samples for each type of disease; For including the first Number of samples for each type of disease.
[0053] The present invention has the following beneficial effects: This invention provides a deep learning method for parallel hybrid identification of multiple bridge defects. After extracting high-dimensional semantic feature maps through a pre-trained convolutional neural network, a row-first slicing strategy is used to vectorize and recombine spatial location information and semantic features into a sequence. This preserves the relative spatial layout of defects in the image, avoids the loss of location information caused by simple global pooling, and is beneficial for distinguishing defects that are spatially close but of different categories. By enhancing the input sequence through the parallel time-mixing module, long-distance dependencies and local contextual information in the feature sequence can be captured simultaneously, enabling the modeling of the interaction relationship between different disease features, thereby improving the robustness of recognition in complex backgrounds, overlapping diseases, or mutual occlusion. Based on a pre-trained convolutional neural network as the feature extraction backbone, it can inherit the general visual representation learned on large natural image datasets. It is then fine-tuned using a multi-disease dataset of bridges, so that the method has good generalization performance for different lighting, shooting angles and bridge structure types (such as beam bridges, arch bridges and cable-stayed bridges), and can adapt to the diversity of actual engineering scenarios.
[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A deep learning method for parallel hybrid identification of multiple bridge defects, characterized in that, include: S1. By combining camera and drone photography, obtain bridge appearance image data and perform preprocessing to obtain preprocessed images; Based on the preprocessed images, a multi-label image classification dataset is constructed; S2. Based on the multi-label image classification dataset, feature extraction is performed using a pre-trained convolutional neural network to obtain a high-dimensional semantic feature map; S3. Perform row-first slicing on the high-dimensional semantic feature map to obtain feature vectors for several rows of spatial locations; flatten and recombine the feature vectors for each row of spatial locations to obtain the input sequence; S4. Based on the input sequence, the enhanced feature sequence is obtained through the parallel time-mapping module; S5. Based on the enhanced feature sequence, the probability distribution of various bridge defects is obtained through the adaptive attention aggregation module.
2. The deep learning method for parallel hybrid identification of multiple bridge defects according to claim 1, characterized in that, The specific processing of step S3 is as follows: S31. Obtain the height H of the high-dimensional semantic feature map, and divide the high-dimensional semantic feature map along the height of the high-dimensional semantic feature map to obtain the feature vectors of H row spatial positions; S32. Based on the feature vector of the row space position, expand it using the spatial dimension and the channel dimension to obtain the one-dimensional feature vector of the row space position, expressed as: In the formula, For the first A one-dimensional feature vector of each row's spatial location; For the first Feature vectors of each row's spatial location; To expand the operation; The width of the high-dimensional semantic feature map; The number of channels in the high-dimensional semantic feature map; The index of the feature vector at the row space location; It is the set of real numbers; S33. Arrange the one-dimensional feature vectors of the row space positions according to the spatial order of the high-dimensional semantic feature map to obtain the input sequence, expressed as: In the formula, The input sequence is .
3. The deep learning method for parallel hybrid identification of multiple bridge defects according to claim 1, characterized in that, The parallel time-mapping module includes a normalization layer, a local mixing branch, a global mixing branch that runs parallel to the local mixing branch, and a fusion layer. The normalization layer obtains a standardized feature sequence based on the input sequence through a layer normalization operation; The local hybrid branch obtains local continuous features through depthwise separable one-dimensional convolution based on the standardized feature sequence; The global hybrid branch obtains global context features based on the standardized feature sequence through a multi-head self-attention mechanism; The fusion layer obtains fused features by fusing local continuous features and global contextual features; Based on the fusion features, an enhanced feature sequence is obtained through layer normalization.
4. The deep learning method for parallel hybrid identification of multiple bridge defects according to claim 3, characterized in that, The specific steps of the local hybrid branch are as follows: S411. Based on the standardized feature sequence, the transformed input features are obtained through dimensional transformation; S412. Based on the transformed input features, intermediate features are obtained through depthwise separable one-dimensional convolution, expressed as: In the formula, As an intermediate feature; To transform the input features; For depthwise separable one-dimensional convolution; S413. Based on the intermediate features and the standardized feature sequence, local continuous features are obtained through residual connection, expressed as: In the formula, It is a locally continuous feature; This is a standardized feature sequence.
5. The deep learning method for parallel hybrid identification of multiple bridge defects according to claim 3, characterized in that, The specific steps of the global hybrid branch are as follows: S421. Map the standardized feature sequences to obtain the query matrix, key matrix, and value matrix; S422. Calculate the scaled dot product attention based on the query matrix, key matrix, and value matrix. The expression is: In the formula, To scale the dot product attention; For query matrix; The key matrix; It is a value matrix; The dimension of the query vector for each head; S423. Based on the scaled dot product attention, calculate the attention of each head using a multi-head mechanism, expressed as: In the formula, For the first The size of the attention; For the first The query weight matrix for each head; For the first The attention head's key weight matrix; For the first The attention value weight matrix of each head; S424. Concatenate the attention of all heads to obtain the global context feature, expressed as: In the formula, For global context features; This represents the projected weight matrix for the multi-head attention mechanism; The number of heads.
6. The deep learning method for parallel hybrid identification of multiple bridge defects according to claim 5, characterized in that, The expression for fusing local continuous features and global context features is as follows: In the formula, Features of fusion; These are learnable fusion weights.
7. The deep learning method for parallel hybrid identification of multiple bridge defects according to claim 3, characterized in that, The specific execution steps of the adaptive attention aggregation module are as follows: S51. Based on the enhanced feature sequence, obtain the global feature vector through global average pooling; S52. Set up independent binary classifiers for K disease categories, and input the global feature vector into each independent binary classifier to obtain the probability value of the existence of each disease category; S53. Based on the probability values of each disease category, obtain the Multi-Hot probability vector to obtain the probability distribution of the disease.