A method for early damage identification of floor decking

Abstract

This invention discloses a method for early damage identification of floor decking. The method includes: image data acquisition and annotation, texture enhancement preprocessing based on multi-scale gradient guidance, generation of damage probability heatmaps fused with physical priors, construction of a dual-stream feature interaction backbone network, design of a multi-scale deformable feature aggregation and damage-perceived channel weighted classifier, and multi-task dynamic weighted loss optimization and model training. By fusing image enhancement and mechanical strain prior information at the input stage, and introducing an adaptive interaction mechanism of global and local features into the network structure, the method effectively improves the identification ability of early damage such as microcracks, local corrosion, and loose connections. Experimental results show that the method of this invention exhibits higher detection accuracy and robustness under complex backgrounds and multiple damage levels, significantly outperforming existing convolutional neural networks and conventional detection models.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring and intelligent detection technology, and in particular to a method for identifying early damage to floor decking. Background Technology

[0002] The existing technologies have the following shortcomings: 1. Existing technologies for floor decking damage detection generally rely on image data captured under fixed lighting or from a single angle, resulting in insufficient feature diversity, poor model adaptability to different working conditions, and difficulty in accurately identifying weak feature damage such as early microcracks. 2. Traditional image enhancement methods often employ histogram equalization or filtering enhancement, which cannot effectively distinguish between background texture and damage features, often leading to noise amplification and increased false detection rates, and limited performance in identifying early damage on complex structural surfaces. 3. Existing deep learning detection networks generally adopt a single-stream feature extraction structure, lacking collaborative modeling of local details and global morphology, and fixed convolutional sampling is difficult to adapt to damage geometry deformation, resulting in insufficient detection capability for small-sized cracks and irregular damage areas. 4. Most existing damage recognition models use a single loss function with fixed weights, ignoring uneven class samples and dynamic changes between tasks, easily leading to overfitting or training bias, and making it difficult to achieve balanced recognition and stable generalization across different damage levels.

[0003] In recent years, general-purpose deep learning models such as Faster R-CNN, YOLO, and U-Net have been attempted for damage detection. While their performance surpasses traditional methods, they possess inherent limitations in the specific scenario described in this invention: Model architecture mismatch: These models are mostly designed for general purposes. Their single-stream backbone and fixed-shape anchors make it difficult to effectively balance the feature extraction of global structural context and local subtle damage, and they are insensitive to irregular, small-sized early damage. Lack of physical guidance: Relying entirely on data-driven learning is like operating a "black box," without incorporating prior physical knowledge such as structural mechanics and strain concentration. This leads to a lack of physical rationality in their decisions, making them prone to false positives or false negatives when damage features are not obvious. Passive reliance on data augmentation: Data augmentation is usually achieved through simple geometric transformations (rotation, scaling) or color dithering, lacking preprocessing mechanisms that target and enhance the damage features themselves (such as edge and texture contrast). The models struggle to learn weak features from low-quality raw images.

[0004] Therefore, proposing an early damage identification method for floor decking to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for early damage identification of floor decking. This method achieves high-precision identification and positioning of floor decking under different damage levels and complex environments, thereby overcoming the problems of low sensitivity, poor robustness and insufficient generalization ability of traditional methods in early micro-damage detection, and providing reliable technical support for structural safety monitoring and maintenance decisions.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for identifying early damage to floor decking includes the following steps: S1. Under various lighting conditions and shooting angles, use an industrial camera to capture images of the floor deck surface; S2. A multi-scale gradient-guided texture enhancement algorithm is used to preprocess the floor deck surface image. By separating the structural gradient component and texture component in the floor deck surface image, the damaged edge features are enhanced in a targeted manner to generate an enhanced image. S3. Based on the mechanical properties of the floor decking, the strain concentration area is calculated through finite element simulation to generate a damage probability heat map. This damage probability heat map is then stitched together with the enhanced image in the channel dimension to form an enhanced input image that integrates physical priors. S4. Construct a labeled dataset containing multiple levels of damage based on multiple enhanced input images, and divide it into training set, validation set and test set; S5. Construct an injury recognition model and input the training set, validation set, and test set into the injury recognition model to achieve model training, validation, and testing; S6. Input the image to be identified into the trained damage recognition model. The model outputs the damage category, location information and confidence score.

[0007] Preferably, in S2, a multi-scale gradient-guided texture enhancement algorithm is used to preprocess the floor deck surface image, specifically including: An enhanced image is generated by linearly combining the surface image of the floor decking, a gradient enhancement term, and a texture suppression term. The gradient enhancement term is obtained by extracting the high-frequency components of the damaged edges using the Laplacian operator and then convolving them with a Gaussian kernel. The texture suppression term is obtained by convolving the original image with a Gaussian kernel. This process enhances the features of the damaged edges and suppresses background texture interference, as shown below:

[0008] in, The image of the acquired floor deck surface, To enhance the image, For the Laplace operator, Standard deviation Gaussian kernel function, Standard deviation Gaussian kernel function, Indicates the convolution operation; To enhance the weights for gradient boosting, This is the texture suppression weight.

[0009] Preferably, in S3, based on the mechanical properties of the floor decking, strain concentration regions are calculated through finite element simulation to generate a damage probability heatmap, specifically including: Based on the finite element simulation results of the displacement field at the bolt connection points, the norm of the mixed partial derivatives of the displacement field is calculated to characterize the degree of strain concentration. A Gaussian weighted function is then used to weight and sum the influence regions of each bolt connection point to generate a damage probability heatmap, which is expressed as follows:

[0010] in, The heatmap representing the probability of damage is in eigenvalues ​​at that location Index for bolt connection points, For the first The weighting coefficient of each bolt connection point For the first Finite element simulation results of the displacement field at each bolt connection point For displacement field The mixed partial derivatives, Represents norm operations, For the first The x-coordinate of each bolt connection point For the first The ordinate of each bolt connection point For the first The radius of influence of each bolt connection point.

[0011] Preferably, in S4, the constructed damage identification model includes: (1) A dual-stream feature interaction backbone network is used to process the enhanced input image in parallel, extract global structural features and local detail features respectively, and perform adaptive fusion through a dynamic gating mechanism; (2) Multi-scale deformable feature aggregation module, which is used to perform deformable convolution and feature reweighting operations on the fused features and aggregate multi-scale features in a recursive fusion manner; (3) Damage perception channel weighted classifier, which uses a channel-space joint weighting mechanism to calculate feature response and output the classification probability distribution of damage degree; (4) Calculate the multi-task dynamic weighted loss function, jointly optimize the damage classification loss, damage region regression loss and damage degree consistency loss, and balance different tasks through dynamic weight coefficients; (5) Model iterative training and parameter update: The model parameters are optimized based on the multi-task dynamic weighted loss function.

[0012] Preferably, in the dual-stream feature interaction backbone network, adaptive fusion through a dynamic gating mechanism specifically includes: The gating weights are calculated using a dynamic gating mechanism, and then the global and local feature stream outputs are weighted and fused using these gating weights. Combined with upsampling and downsampling operations, global structural features and local detail features are adaptively fused, as shown below:

[0013] in, Represents the fused feature map. The dynamic gating weights are calculated using the Sigmoid function. This represents the learnable convolutional kernel parameters corresponding to the global feature flow. This represents the output features of the global feature stream. This represents the learnable convolution kernel parameters corresponding to the local feature flow, where σ is the standard deviation of the Gaussian kernel. This represents the output features of the local feature stream. This represents element-wise multiplication. This indicates an upsampling operation.

[0014] Preferably, in the multi-scale deformable feature aggregation module, deformable convolution and feature reweighting operations are performed on the fused features, specifically including: Deformable convolution operations are performed on the feature maps at each scale. The offset convolutional layers predict the sampling point offsets, adjust the convolution sampling positions, and reweight the features using spatial attention weights, as shown below:

[0015] in, Indicates the first Layer input feature map, Indicates the first The output feature map of the layer after deformable convolution and feature reweighting. The regular sampling grid represents the 3×3 convolution kernel. Represents the sampling grid Sampling points in This represents the weight of the convolution kernel at the p-th sampling point; Indicates the first Learnable offsets of layer feature maps This represents the spatial attention weight value at the p-th sampling point.

[0016] Preferably, in the damage-sensing channel-weighted classifier, a channel-spatial joint weighting mechanism is used to calculate the feature response, specifically including: The spatially weighted characteristic response value for each channel is calculated using the following formula:

[0017] in, Indicates the first Spatial weighted characteristic response values ​​of each channel Indicates the first Each channel is located in eigenvalues ​​at that location Indicates the first Each channel is located in Spatial attention weights at the location The height dimension of the input image. The width dimension of the input image; Based on the feature mean of each channel, the output of the Softmax function is adjusted using a temperature coefficient to calculate the channel weights, which are expressed as follows:

[0018] in, Indicates the first The weight of each channel, Indicates the first The characteristic mean of each channel, Indicates the temperature coefficient. This indicates the total number of channels.

[0019] The preferred formula for the multi-task dynamic weighted loss function is as follows:

[0020] in, For multi-task dynamic weighted total loss, The consistency loss weighting coefficient, For regression loss weighting coefficients, For consistent damage levels, The regression loss is for the damaged area.

[0021] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for early damage identification of floor decking as described above.

[0022] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) This invention provides a method for early damage identification of floor decking that integrates image visual features and physical and mechanical priors. It enhances the features of micro-damage by using a multi-scale gradient-guided texture enhancement algorithm, improves the physical rationality of identification by combining the damage probability heat map generated by finite element strain analysis, and constructs a deep network structure with dual-flow feature interaction and multi-scale deformable feature aggregation to achieve high-precision identification and positioning of floor decking under different damage levels and complex environments. This overcomes the problems of low sensitivity, poor robustness and insufficient generalization ability of traditional methods in early micro-damage detection, and provides reliable technical support for structural safety monitoring and maintenance decisions.

[0023] (2) This invention introduces a multi-scale gradient-guided texture enhancement algorithm to decompose the original image into structural and texture components and to perform targeted enhancement of the damage edge features, thereby achieving significant enhancement of early micro-damage features such as fine cracks, local corrosion and loose connections, effectively solving the problems of low damage contrast and indistinct features in complex backgrounds.

[0024] (3) This invention introduces prior physical and mechanical information into the floor decking damage identification process, generates a damage probability heat map of strain concentration area through finite element simulation, and integrates it with texture enhancement image in the channel dimension, so that the identification model has both image appearance features and strain distribution information, which significantly improves the accuracy and physical consistency of early damage location. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating an early damage identification method for floor decking provided by the present invention; Figure 2 This diagram illustrates the effect of the multi-scale gradient-guided texture enhancement of this invention on damage features. Figure 3 This diagram illustrates the improvement in damage localization accuracy based on physical prior thermal mapping according to the present invention. Figure 4 This is a graph showing the detection performance of the multi-scale deformable feature aggregation module of the present invention under different damage sizes; Figure 5 This is a comparison chart of the accuracy of the method of the present invention with other conventional methods in the classification of level 5 injuries; Figure 6 This is a graph showing the detection performance of the method of the present invention compared with other conventional methods at different damage sizes. Detailed Implementation

[0027] 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.

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the present invention provides a method for identifying early damage to floor decking, comprising the following steps: S1. Under various lighting conditions and shooting angles, use an industrial camera to capture images of the floor deck surface; S2. A multi-scale gradient-guided texture enhancement algorithm is used to preprocess the floor deck surface image. By separating the structural gradient component and texture component in the floor deck surface image, the damaged edge features are enhanced in a targeted manner to generate an enhanced image. S3. Based on the mechanical properties of the floor decking, the strain concentration area is calculated through finite element simulation to generate a damage probability heat map. This damage probability heat map is then stitched together with the enhanced image in the channel dimension to form an enhanced input image that integrates physical priors. S4. Construct a labeled dataset containing multiple levels of damage based on multiple enhanced input images, and divide it into training set, validation set and test set; S5. Construct an injury recognition model and input the training set, validation set, and test set into the injury recognition model to achieve model training, validation, and testing; S6. Input the image to be identified into the trained damage recognition model. The model outputs the damage category, location information and confidence score.

[0030] Specifically, this invention proposes a method for identifying early damage to floor decking, the main steps of which are as follows: S1, Floor Decking Image Data Acquisition The floor decking image data acquisition process uses a high-resolution industrial camera under various lighting conditions and shooting angles to ensure complete coverage of the floor decking at different stages of use and damage states. The acquisition equipment is fixed on a three-axis stabilized gimbal and systematically scans and captures images of the floor decking surface along a preset path, keeping the distance between the lens and the decking surface constant at 1.5 meters, and setting the resolution of each image to 4096×3072 pixels.

[0031] To comprehensively characterize early damage features, data collection covered key areas such as bolted connection areas, plate-rib junctions, and mid-span regions. Images of each area were collected from three perspectives: frontal view, 45-degree oblique view, and side view. A five-level damage severity classification system was used to construct the training dataset, including no damage, minor damage, moderate damage, severe damage, and critical condition. At least 2000 original images were collected for each category.

[0032] S2, Texture enhancement preprocessing of floor decking image data Damage detection of floor decking faces challenges such as complex background textures and low contrast in early damage areas. Conventional processing methods use histogram equalization to enhance image contrast, but this can easily amplify background noise and weaken the features of fine cracks, leading to a decline in damage detection performance.

[0033] This invention employs a multi-scale gradient-guided texture enhancement algorithm. By separating the structural gradient components and texture components in an image, it performs targeted enhancement of damaged edge features, thereby enhancing the visual salience of the damaged area and suppressing interference from complex background textures. The specific steps are as follows: 1) Define the Gaussian kernel function A Gaussian kernel function is defined based on the standard deviation parameter to control the filtering scale. Multi-scale smoothing filtering is achieved using image convolution operations. Weighted averaging is performed on the spatial locations of the image plane to smooth the image and suppress background noise, as expressed below:

[0034] In the formula, This represents the Gaussian kernel function, used for image smoothing filtering, and characterizes the basis of multi-scale smoothing; The standard deviation of the Gaussian kernel controls the filtering scale; The horizontal coordinate of the image plane represents the spatial location. The vertical coordinate of the image plane represents the spatial location; It is a natural exponential function.

[0035] 2) Calculate the enhanced image An enhanced image is generated by linearly combining the original image, a gradient enhancement term, and a texture suppression term. The gradient enhancement term is obtained by extracting the high-frequency components of the damaged edges using the Laplacian operator and then convolving them with a Gaussian kernel. The texture suppression term is obtained by convolving the original image with a Gaussian kernel. This process enhances the features of the damaged edges and suppresses background texture interference, as shown below:

[0036] In the formula, The original input image is the acquired image of the damage to the floor decking. To enhance the image, a damage image of the enhanced floor decking is generated; This is the Laplacian operator, used to extract high-frequency components from damaged edges; Standard deviation The Gaussian kernel controls the degree of enhancement at the damage edge; Standard deviation The Gaussian kernel controls the degree of background texture suppression; Indicates the convolution operation; For gradient boosting weights, the preferred settings are... ; For texture suppression weights, the preferred setting is... .

[0037] In a specific embodiment, the impact of texture enhancement algorithms on damage features is analyzed to verify the enhancement effect of multi-scale gradient-guided texture enhancement algorithms on various damage features, such as... Figure 2 As shown in the figure, the experiment compares the differences in damage feature response values ​​between the original and enhanced images, focusing on four typical early damage modes: fine cracks, localized corrosion, surface peeling, and loose connections. The feature response value is a dimensionless indicator; a higher value indicates a more significant damage feature. It is evident from the figure that the feature response values ​​of all damage types are significantly improved after enhancement using this technique. In particular, the feature response values ​​of fine cracks and loose connections, the two most easily overlooked early damage types, show the most significant improvement, indicating that the texture enhancement algorithm of this technique can effectively strengthen damage edge features, significantly improve the visual salience of early damage, and provide a more reliable feature basis for subsequent identification. In terms of experimental configuration, the same feature extraction method is used to process both the original and enhanced images to ensure comparability of the results.

[0038] S3. Generation of Damage Region Heatmaps Based on Physical Priors and Construction of Training Dataset In the task of detecting damage to floor decking, the early damage morphology is irregular and the annotation cost is high. Conventional target detection methods rely on a large number of bounding boxes for training, but it is difficult to accurately annotate when the damage morphology is irregular, and the annotation cost is high, resulting in insufficient generalization ability of the detection model.

[0039] This invention utilizes the mechanical properties of floor decking to construct a strain concentration region prediction model to generate a damage probability heatmap. This heatmap is then stitched together with a texture-enhanced image along the channel dimension to form an enhanced input that incorporates prior physical knowledge. This provides subsequent networks with both image features and damage spatial probability information. The specific steps are as follows: 1) Calculate the damage probability heatmap Based on the finite element simulation results of the displacement field at bolted connection points, the norm of the mixed partial derivatives of the displacement field is calculated to characterize the degree of strain concentration. A Gaussian weighted function is then used to weight and sum the influence regions of each bolted connection point to generate a damage probability heatmap, thus indicating the spatial probability distribution of potential damage, expressed as:

[0040] In the formula, A heatmap representing the probability of damage is used to indicate the spatial probability distribution of possible damage. The heatmap representing the probability of damage is in Eigenvalues ​​at; This is the index for the bolt connection point, with values ​​ranging from 1 to 4. For the first The weighting coefficient for each bolt connection point is set to 0.25 by default. For the first Finite element simulation results of displacement field at a bolt connection point; For displacement field The mixed partial derivatives characterize the degree of strain concentration; This indicates norm operations, used to take the absolute value of mixed partial derivatives; For the first The x-coordinate of each bolt connection point; For the first The ordinate of each bolt connection point; For the first The influence radius of each bolt connection point is taken as 1 / 10 of the plate span.

[0041] 2) Image stitching enhancement with heatmap The texture-enhanced image and the damage probability heatmap are concatenated along the channel dimension to form an enhanced input that incorporates physical priors. This provides subsequent networks with input data that simultaneously contains image features and damage probability information, represented as follows:

[0042] In the formula, This indicates enhanced input, which serves as input for subsequent networks; This indicates a channel splicing operation.

[0043] The data annotation process was completed independently by three professionally trained engineering inspection personnel. They used a pixel-level semantic segmentation annotation method and used annotation tools to accurately delineate the outline of the damaged area. The annotation categories strictly corresponded to the five-level damage classification system. At the same time, they recorded the specific morphological characteristics of the damage, including typical early damage modes such as micro-cracks, local corrosion, surface peeling, and loosening of connections.

[0044] The final dataset contains 10,000 labeled samples, which are divided into training, validation and test sets in a ratio of 7:2:1.

[0045] In one embodiment, the focus is on evaluating the effect of damage probability heatmaps based on physical priors on improving damage localization accuracy. For example... Figure 3 As shown, the experiment uses a grouped bar chart to compare the results of no physical prior, conventional thermograms, and the physical prior thermograms of this invention. Figure 3 The cross-comparison (CCOM) ratio of the proposed method is compared to that of other methods at five different damage levels. The horizontal axis represents the damage level, ranging from no damage to a dangerous state; the vertical axis represents the CCOM ratio, with higher values ​​indicating more accurate damage area localization. Experimental results show that the proposed method achieves the highest CCOM ratio across all damage levels, particularly in the early stages of minor and moderate damage, where the improvement is most significant. This indicates that the proposed method utilizes the strain concentration region prediction model constructed based on the mechanical properties of the floor decking. This model characterizes the strain concentration by calculating the norm of the mixed partial derivative of the displacement field, enabling more accurate prediction of the potential spatial location of damage. In contrast, methods without physical prior knowledge rely entirely on image features, making accurate localization difficult when early damage features are not readily apparent. While conventional thermal mapping methods provide some spatial prior knowledge, they lack consideration for the structural characteristics of the floor decking, resulting in limited localization accuracy. Experimental results demonstrate that combining physical prior knowledge with image features can significantly improve the reliability of early damage localization.

[0046] S4. Construct a damage identification model S401, Dual-stream Feature Interaction Backbone Network In the task of damage detection of floor decking, the image contains global structural features and local detail features. However, conventional convolutional neural networks use a single feature stream, which makes it difficult to capture both types of features at the same time. This results in limited damage detection performance in complex backgrounds and an inability to effectively balance the representation of macroscopic damage morphology and microscopic crack features.

[0047] This invention employs a dual-stream feature interaction backbone network, extracting global structural information and local detail information of an image through parallel and independent global and local feature streams, respectively. A dynamic gating mechanism is then used to adaptively fuse the two feature streams, taking into account both macroscopic damage morphology and microscopic crack features. The specific steps are as follows: 1) Global Feature Flow Construction The input feature map is downsampled four times using a 7×7 convolution operation with a stride of 2, a max-pooling layer, and a sequence of residual blocks to progressively reduce the spatial resolution and extract global structural features from the image. The output is a global feature map with lower spatial resolution and a higher number of channels. The output feature map of the global feature flow is defined as follows: It is used to carry global structural information in the image; in, The height dimension of the input image; The width dimension of the input image; This represents the number of channels in the global feature stream. It represents the real number space, indicating that the elements of the feature map are real values; This indicates that the feature map height is 1 / 16 of the input image height, representing the spatial resolution after four downsampling operations; This indicates that the feature map width is 1 / 16 of the input image width, representing the spatial resolution after four downsampling operations.

[0048] 2) Construction of local feature flows By employing 5×5 dilated convolutions with a stride of 1, depthwise separable convolutions, and residual blocks, local detail features in the image are extracted, outputting a local feature map with high spatial resolution and a low number of channels. The output feature map of the local feature flow is defined as follows: It is used to carry local detail information in the image; in, This represents the number of channels in the local characteristic flow.

[0049] 3) Dynamic gating feature interaction The gating weights are calculated using a dynamic gating mechanism, and then the global and local feature stream outputs are weighted and fused using these gating weights. Combined with upsampling and downsampling operations, global structural features and local detail features are adaptively fused, as shown below:

[0050] In the formula, This represents a fused feature map, used to combine global and local features; This represents the dynamic gating weight, calculated using the Sigmoid function, with a value range of [value range missing]. This is used to control the fusion ratio of global and local features, and is expressed as... ; The learnable convolutional kernel parameters corresponding to the global feature flow are trainable parameters; This represents a convolution operation, used for feature transformation. This represents the output features of the global feature stream; The learnable convolutional kernel parameters corresponding to the local feature flow are trainable parameters; This represents the output features of the local feature stream; This indicates an element-wise multiplication operation, used for feature weighting; This indicates an upsampling operation, which uses bilinear interpolation to enlarge the spatial resolution of the feature map to the target size; This indicates a downsampling operation, which uses bilinear interpolation to reduce the spatial resolution of the feature map to the target size. This represents a 1×1 convolution operation, used to adjust the number of feature channels; This means concatenating the global feature stream output and the downsampled local feature stream output along the channel dimension; This represents the Sigmoid activation function, which maps input values ​​to... Interval.

[0051] S402, Multi-scale Deformable Feature Aggregation Module In the task of detecting damage to floor decking, small cracks are easy to miss. Conventional Feature Pyramid Network (FPN) is insufficient in representing the features of small cracks when fusing features at multiple scales, which can easily lead to a decrease in detection performance.

[0052] This invention employs deformable convolution and feature reweighting mechanisms. Deformable convolution adjusts the sampling point positions to adapt to damage geometry deformation, and spatial attention weights are combined to reweight features to highlight important regions. Finally, a top-down recursive fusion method is used to aggregate features at different scales, outputting enhanced multi-scale damage features. The specific steps are as follows: 1) Deformable convolution and feature reweighting Deformable convolution operations are performed on the feature maps at each scale. Offset convolutional layers predict sampling point offsets, adjusting the convolution sampling positions. Spatial attention weights are then incorporated to reweight the features. These spatial attention weights are calculated using attention convolutional layers and the Softmax function, thereby enhancing the feature representation capability. This is expressed as:

[0053] In the formula, Indicates the first Layer input feature map, Feature layers corresponding to different resolutions; Indicates the first The output feature map of the layer after deformable convolution and feature reweighting; A regular sampling grid representing a 3×3 convolution kernel; Represents the sampling grid Sampling points in; This represents the weight of the convolution kernel at the p-th sampling point; Indicates the first Learnable offsets of layer feature maps, passed through offset convolutional layers The calculation yields the following result, which is expressed as: ; This represents the spatial attention weight value at the p-th sampling point, used to adjust the contribution of each position in the feature map; Indicates the first Spatial attention weights of the layer feature maps are passed through attention convolutional layers. It is calculated using the Softmax function, and the calculation method is expressed as follows: ; This represents a normalized exponential function that converts attention scores into a probability distribution. This indicates an offset convolutional layer used to predict the offset of sampling points; This represents an attention convolutional layer used to generate a spatial attention map.

[0054] 2) Recursive feature fusion A top-down recursive fusion approach is adopted, which upsamples the high-level fused features and concatenates them with the features of the current layer that have undergone deformable convolution and feature reweighting. Multi-scale feature aggregation is then achieved through convolution operations, outputting multi-scale damage features, represented as follows:

[0055]

[0056] In the formula, The highest-level fusion feature is directly taken from... ; This represents the output feature map after the input feature map of the 4th layer undergoes deformable convolution and feature reweighting, serving as the starting point for recursive fusion; Indicates the first Layer fusion characteristics, The values ​​3, 2, and 1 correspond to different resolution levels; Indicates the convolution operation; This indicates the upsampled high-level features. With current layer features Perform splicing along the channel dimension; This indicates an upsampling operation, which uses bilinear interpolation to increase the spatial size of the feature map; Indicates the first The fusion features of the layers are used as high-level features input to the current layer in recursive fusion.

[0057] Furthermore, the final output is multi-scale damage features. ,in The resolution is 1 / 4 of the input image. It is 1 / 8. It is 1 / 16. It is 1 / 32.

[0058] In a specific embodiment, the detection performance of the multi-scale deformable feature aggregation module under different damage sizes was verified, with particular attention paid to its ability to detect small-sized damage. For example... Figure 4As shown, the experiment uses an area map to compare the detection recall rates of conventional feature pyramid networks, ASPP modules, PANet, and the deformable aggregation module of this invention at five damage sizes. The horizontal axis represents the damage size, and the vertical axis represents the detection recall rate; a higher value indicates fewer missed damages. Experimental results show that the method of this invention achieves the highest recall rate across all damage sizes, especially for small damages smaller than two millimeters, where the advantage is most significant. The effectiveness of the deformable feature aggregation module is demonstrated by adjusting the sampling point position through deformable convolution to adapt to damage geometry deformation, combining spatial attention weights to reweight features to highlight important regions, and then using a top-down recursive fusion method to aggregate multi-scale features. The percentage improvement in small target detection marked in the figure intuitively demonstrates the breakthrough progress of this invention in small-sized damage detection. In contrast, conventional feature pyramid networks lack sufficient feature representation for small cracks, easily leading to missed detections; the ASPP module expands the receptive field through multi-scale dilated convolutions, but its adaptability to geometric deformation is limited; while PANet improves the feature fusion path, it lacks specific optimization for irregular damage morphologies.

[0059] S403, Damage Perception Channel Weighted Classifier In the task of detecting damage to floor decking, the feature response of local damage is easily weakened. Conventional classifiers use global pooling, which tends to average the entire feature map, resulting in the feature response of local damage areas being diluted and affecting classification accuracy.

[0060] This invention employs a channel-spatial joint weighting mechanism. First, spatial attention weights are used to calculate spatially weighted feature responses to enhance the feature contribution of the lesion region. Then, channel weights are calculated based on the channel feature mean to strengthen important channels. Finally, the weighted feature responses are input into a fully connected layer for lesion classification prediction. The specific steps are as follows: 1) Calculate the spatially weighted characteristic response For each channel's feature map, a spatially weighted average is calculated. The spatial attention weights are obtained through a 3×3 convolution and a sigmoid function, used to enhance the feature response of the damaged region and suppress background interference. The resulting spatially weighted feature response value for each channel is expressed as follows:

[0061] In the formula, Indicates the first Spatial weighted characteristic response values ​​of each channel; Indicates the first Feature maps of each channel; This represents the row coordinate index of the feature map, with values ​​ranging from 1 to... ; This is the column coordinate index of the feature map, with values ​​ranging from 1 to... ; Indicates the first Each channel is located in Eigenvalues ​​at; Indicates the first Each channel is located in Spatial attention weights at the location are used to enhance the feature response of the damaged area, through the first Spatial attention weight map of each channel get; Indicates the first The spatial attention weight map for each channel is calculated as follows: ; This represents a 3×3 convolution operation.

[0062] 2) Calculate channel weights Based on the characteristic mean of each channel, the output of the Softmax function is adjusted by a temperature coefficient to calculate channel weights, thereby strengthening the contribution of important channels and weakening the influence of secondary channels, as expressed as:

[0063] In the formula, Indicates the first The weight of each channel; Indicates the first The characteristic mean of each channel; Indicates the temperature coefficient, preferably set To control the sharpness of the weight distribution; This indicates the total number of channels.

[0064] 3) Classification prediction The spatially weighted feature responses of each channel are multiplied by their respective channel weights to form a feature vector. This vector is then used to perform the final classification prediction through a fully connected layer and a softmax function, outputting a classification probability distribution, as follows:

[0065] In the formula, This represents the predicted classification probability distribution; This represents the classification weight matrix, which consists of trainable parameters. This represents a column vector composed of the weighted feature responses of each channel; Indicates the first The weight of each channel; Indicates the first Spatial weighted characteristic response values ​​of each channel; This represents the bias term, which is a trainable parameter.

[0066] S404, Calculate the loss function In the task of detecting damage to floor decking, the distribution of damage categories is unbalanced. For example, there are fewer samples of severe damage and more samples of minor damage. A single cross-entropy loss function will cause the optimization process to be biased towards the majority class, making it difficult to effectively learn the damage features of the minority class. At the same time, the conventional fixed-weight multi-task loss cannot adapt to the dynamic changes in the contribution of different tasks during training, affecting the overall performance of the model.

[0067] This invention constructs a multi-task dynamic weighted loss function to jointly optimize three sub-tasks: damage classification, damage region regression, and damage degree consistency. It balances the learning difficulty between categories and the optimization direction between tasks by dynamically adjusting the loss weights of each task based on the number of class samples and the training progress. The specific steps are as follows: 1) Damage Classification Loss Calculation The weighted cross-entropy loss function is used to calculate the damage classification loss, where the class weights are calculated based on the total number of samples in the training set and the number of samples in each class. This is used to balance the differences in the number of samples between different classes and to suppress the majority class from dominating the optimization direction. It is expressed as:

[0068] In the formula, This represents the damage classification loss, which measures the error in the model's prediction of damage categories. This is a category index with values ​​ranging from 1 to 5, corresponding to five levels of severity of damage to the floor decking. The first one-hot encoded vector of the real label The value of the class can be 0 or 1, indicating whether the sample belongs to the ca-th class; The model predicts the probability that a sample belongs to the ca-th class, and the result is output through the Softmax function, with a value ranging from 0 to 1. ; Let be the weight coefficient for the ca-th class, used to balance the differences in sample size between different classes. The calculation method is expressed as follows: ; The total number of samples in the training set represents the scale of the data. This represents the number of samples of the ca-th class in the training set, reflecting the class distribution. It is a logarithmic function, with the default base being the natural constant.

[0069] 2) Calculation of regression loss in the damaged area The overlap between the predicted and actual binary damage masks is calculated using the intersection-over-union (IoU) function and used as the regression loss for the damage region, directly optimizing the localization accuracy of the damage region. This is expressed as:

[0070] In the formula, This represents the regression loss for the damaged area, used to measure the consistency between the predicted damaged area and the actual damaged area. The intersection-union ratio function calculates the ratio of the area of ​​the intersection to the area of ​​the union of two regions. This is a binary mask for the damage predicted by the model, with a value of 0 or 1, representing whether each pixel is judged as a damage. The binary mask for the actual damaged area serves as a monitoring signal.

[0071] 3) Calculation of Consistency Loss in Damage Degree The L2 norm squared of the average and maximum values ​​of the feature vectors from the previous layer of the classifier on the batch samples is used as the damage consistency loss, thereby strengthening the cohesion of similar damage features, expressed as:

[0072] In the formula, It represents the loss of consistency in damage level and is used to enhance the cohesion of similar damage characteristics; This is the average feature vector of the previous layer of the classifier on the batch samples, representing the average feature of the degree of damage; The maximum value of the feature vector of the previous layer of the classifier on the batch samples represents a significant feature that characterizes the degree of damage. This represents the square of the L2 norm, used to measure the difference between two eigenvectors.

[0073] 4) Integration of total loss function The damage classification loss, regression loss, and consistency loss are linearly combined, where the regression loss weight increases with the number of training iterations, and the consistency loss weight is linked to the classification accuracy on the validation set. This forms a multi-task dynamically weighted total loss, which serves as the final optimization objective for model training, expressed as:

[0074] In the formula, The total loss is dynamically weighted across multiple tasks and used as the final optimization objective for model training. The consistency loss weight coefficient is linked to model validation performance and is calculated as follows: ; The regression loss weights increase with the training process, and are calculated as follows: ; The regression weight growth rate control parameter is preferably set to 0.01 to adjust the weight increase curve; The current training iteration number represents the training progress. It is a natural constant; To verify the classification accuracy of the validation set and reflect the model's generalization ability in real time.

[0075] S405, Model Iterative Training and Parameter Update The model iterative training process adopts a phased optimization strategy. First, the Xavier initialization method is used to initialize the network parameters, and then the stochastic gradient descent algorithm with momentum is used to update the parameters.

[0076] The learning rate was set to 0.01 and the momentum coefficient to 0.9 at the beginning of training. After every 50 training cycles, the learning rate was reduced to 0.5 times the original value. Each training batch contains 32 images, and augmented samples are generated in real time using data augmentation techniques, including random rotation, brightness adjustment, and contrast transformation.

[0077] During parameter updates, the multi-task dynamic weighted total loss for each batch is calculated, the gradients of each layer of the network are calculated using the backpropagation algorithm, and the gradient norm is limited to within 1.0 using gradient pruning techniques to prevent gradient explosion.

[0078] During training, the model continuously monitors the classification accuracy and crossover ratio of damaged regions on the validation set. When the validation set performance no longer improves within 20 consecutive training cycles, the current optimal model parameters are automatically saved.

[0079] The training process is set to a maximum iteration period of 500. When the maximum iteration period is reached or the performance of the validation set does not improve for 50 consecutive periods, the training process is terminated and the optimal model parameters are loaded for subsequent recognition tasks.

[0080] S5. Early Damage Identification of Floor Decking After the model is trained, the image of the floor decking to be detected is input into the complete damage identification process.

[0081] First, the input image is preprocessed with texture enhancement, and the damaged edge features are enhanced and the background texture interference is suppressed by a multi-scale gradient guided algorithm. Then, the enhanced image is stitched together with the damage probability heatmap based on physical priors along the channel dimension to form an enhanced input that incorporates physical knowledge; The enhanced input is fed into a dual-stream feature interaction backbone network, which extracts global structural features and local detail features respectively, and achieves adaptive feature fusion through a dynamic gating mechanism; The fused features are processed by a multi-scale deformable feature aggregation module, which uses deformable convolution to adjust the sampling point position to adapt to the damage geometry deformation, and aggregates damage features of different scales through recursive fusion. Finally, the damage-aware channel-weighted classifier calculates the feature response based on the channel-space joint weighting mechanism and outputs the probability distribution of five levels of damage through a fully connected layer.

[0082] The identification system determines the final identification result based on the damage category corresponding to the highest probability, and outputs the location information of the damaged area and the confidence score, providing a comprehensive and reliable technical basis for engineering maintenance decisions.

[0083] In a specific embodiment, the performance of different damage identification methods in distinguishing five damage severity categories of floor decking was evaluated. For example... Figure 5 As shown in the figure, the experiment compares the classification accuracy of the method of this invention with five conventional techniques, including conventional convolutional neural networks, U-shaped networks, residual networks, fast region convolutional neural networks, and single-shot multi-frame detectors, across five damage levels: no damage, minor damage, moderate damage, severe damage, and dangerous conditions, using a bar chart. The figure shows that the method of this invention maintains the highest classification accuracy across all five damage levels, especially in the difficult-to-identify minor and moderate damage categories. This demonstrates that the proposed dual-stream feature interaction backbone network and damage-aware channel weighted classifier can effectively capture subtle feature differences between different damage levels, while conventional methods are relatively weaker in early damage identification, highlighting the significant advantage of the method of this invention in fine-grained damage classification.

[0084] In specific embodiments, the detection stability of different methods when faced with damage of different sizes is analyzed. For example... Figure 6As shown, the experiment uses a line graph to illustrate the detection rate changes of each method across five damage size ranges. The damage size gradually increases from microcracks to severe damage, with the detection rate ranging from 0.5 to 1, representing the probability of successfully detecting the damage. Experimental results show that the detection rate of all methods increases with increasing damage size, but the method of this invention performs particularly well in the most challenging microcrack detection, significantly outperforming other comparative methods. Experimental results demonstrate that the design of the multi-scale deformable feature aggregation module in this invention, which adaptively adjusts the sampling position through deformable convolution, can effectively capture the features of small-sized cracks. As the damage size increases, the performance gap between methods gradually narrows, but the method of this invention maintains its leading position across all size ranges, proving its robustness and reliability in full-size damage detection.

[0085] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for early damage identification of floor decking as described above.

[0086] This invention proposes a dual-stream feature interaction backbone network and a multi-scale deformable feature aggregation module. Through dynamic gating fusion of global and local features and geometric adaptive sampling mechanism of deformable convolution, it takes into account the characterization ability of macroscopic structure and micro-crack, and significantly improves the detection accuracy of complex damage morphology and the identification stability of small cracks.

[0087] This invention designs a multi-task dynamic weighted loss function that combines damage classification, region regression, and consistency constraint tasks. By dynamically adjusting the weights of each subtask to balance class imbalances and training process differences, the model achieves better convergence and generalization performance in multi-damage level identification.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0089] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying early damage to floor decking, characterized in that, Includes the following steps: S1. Under various lighting conditions and shooting angles, use an industrial camera to capture images of the floor deck surface; S2. A multi-scale gradient-guided texture enhancement algorithm is used to preprocess the floor deck surface image. By separating the structural gradient component and texture component in the floor deck surface image, the damaged edge features are enhanced in a targeted manner to generate an enhanced image. S3. Based on the mechanical properties of the floor decking, the strain concentration area is calculated by finite element simulation to generate a damage probability heat map. The damage probability heat map is then stitched together with the enhanced image in the channel dimension to form an enhanced input image that integrates physical priors. S4. Construct a labeled dataset containing multiple levels of damage based on multiple enhanced input images, and divide it into training set, validation set and test set; S5. Construct an injury recognition model and input the training set, validation set, and test set into the injury recognition model to achieve model training, validation, and testing; S6. Input the image to be identified into the trained damage recognition model. The model outputs the damage category, location information and confidence score.

2. The method for early damage identification of floor decking according to claim 1, characterized in that, In step S2, a multi-scale gradient-guided texture enhancement algorithm is used to preprocess the surface image of the floor decking, specifically including: An enhanced image is generated by linearly combining the surface image of the floor decking, a gradient enhancement term, and a texture suppression term. The gradient enhancement term is obtained by extracting the high-frequency components of the damaged edges using the Laplacian operator and then convolving them with a Gaussian kernel. The texture suppression term is obtained by convolving the original image with a Gaussian kernel. This process enhances the features of the damaged edges and suppresses background texture interference, as shown below: in, The image of the acquired floor deck surface, To enhance the image, For the Laplace operator, Standard deviation Gaussian kernel function, Standard deviation Gaussian kernel function, Indicates the convolution operation; To enhance the weights for gradient boosting, This is the texture suppression weight.

3. The method for early damage identification of floor decking according to claim 1, characterized in that, In step S3, based on the mechanical properties of the floor decking, strain concentration regions are calculated through finite element simulation to generate a damage probability heatmap, specifically including: Based on the finite element simulation results of the displacement field at the bolt connection points, the norm of the mixed partial derivatives of the displacement field is calculated to characterize the degree of strain concentration. A Gaussian weighted function is then used to weight and sum the influence regions of each bolt connection point to generate a damage probability heatmap, which is expressed as follows: in, The heatmap representing the probability of damage is in eigenvalues ​​at that location Index for bolt connection points, For the first The weighting coefficient of each bolt connection point For the first Finite element simulation results of the displacement field at each bolt connection point For displacement field The mixed partial derivatives, Represents norm operations, For the first The x-coordinate of each bolt connection point For the first The ordinate of each bolt connection point For the first The radius of influence of each bolt connection point.

4. The method for early damage identification of floor decking according to claim 1, characterized in that, In S5, the constructed damage identification model includes: (1) A dual-stream feature interaction backbone network is used to process the enhanced input image in parallel, extract global structural features and local detail features respectively, and perform adaptive fusion through a dynamic gating mechanism; (2) Multi-scale deformable feature aggregation module, which is used to perform deformable convolution and feature reweighting operations on the fused features and aggregate multi-scale features in a recursive fusion manner; (3) Damage perception channel weighted classifier, which uses a channel-space joint weighting mechanism to calculate feature response and output the classification probability distribution of damage degree; (4) Calculate the multi-task dynamic weighted loss function, jointly optimize the damage classification loss, damage region regression loss and damage degree consistency loss, and balance different tasks through dynamic weight coefficients; (5) Model iterative training and parameter update: The model parameters are optimized based on the multi-task dynamic weighted loss function.

5. The method for early damage identification of floor decking according to claim 4, characterized in that, The dual-stream feature interaction backbone network, which performs adaptive fusion through a dynamic gating mechanism, specifically includes: The gating weights are calculated using a dynamic gating mechanism, and then the global and local feature stream outputs are weighted and fused using these gating weights. Combined with upsampling and downsampling operations, global structural features and local detail features are adaptively fused, as shown below: in, Represents the fused feature map. The dynamic gating weights are calculated using the Sigmoid function. This represents the learnable convolutional kernel parameters corresponding to the global feature flow. This represents the output features of the global feature stream. The learnable convolutional kernel parameters represent the local feature flow. , This represents the output features of the local feature stream. This represents element-wise multiplication. This indicates an upsampling operation.

6. The method for early damage identification of floor decking according to claim 4, characterized in that, In the multi-scale deformable feature aggregation module, deformable convolution and feature reweighting operations are performed on the fused features, specifically including: Deformable convolution operations are performed on the feature maps at each scale. The offset convolutional layers predict the sampling point offsets, adjust the convolution sampling positions, and reweight the features using spatial attention weights, as shown below: in, Indicates the first Layer input feature map, Indicates the first The output feature map of the layer after deformable convolution and feature reweighting. The regular sampling grid represents the 3×3 convolution kernel. Represents the sampling grid Sampling points in This represents the weight of the convolution kernel at the p-th sampling point; Indicates the first Learnable offsets of layer feature maps This represents the spatial attention weight value at the p-th sampling point.

7. The method for early damage identification of floor decking according to claim 4, characterized in that, The damage-sensing channel-weighted classifier employs a channel-space joint weighting mechanism to calculate feature responses, specifically including: The spatially weighted characteristic response value for each channel is calculated using the following formula: in, Indicates the first Spatial weighted characteristic response values ​​of each channel Indicates the first Each channel is located in eigenvalues ​​at that location Indicates the first Each channel is located in Spatial attention weights at the location, The height dimension of the input image. The width dimension of the input image; Based on the feature mean of each channel, the output of the Softmax function is adjusted using a temperature coefficient to calculate the channel weights, which are expressed as follows: in, Indicates the first The weight of each channel, Indicates the first The characteristic mean of each channel, Indicates the temperature coefficient. This indicates the total number of channels.

8. The method for early damage identification of floor decking according to claim 4, characterized in that, The formula for the calculated multi-task dynamic weighted loss function is as follows: in, For multi-task dynamic weighted total loss, Classify the loss as damage, The consistency loss weighting coefficient, For regression loss weighting coefficients, For consistent damage level loss, The regression loss is for the damaged area.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for early damage identification of floor decking as described in any one of claims 1 to 8.

Images