Intelligent urban tunnel cross ditch damage assessment method and system

Abstract

The application discloses a kind of intelligent urban tunnel cross ditch damage evaluation method and system, method includes urban tunnel cross ditch image acquisition, cross ditch image dataset optimization, establishes cross ditch damage evaluation model and cross ditch damage evaluation.The application belongs to the field of damage evaluation, specifically refers to a kind of intelligent urban tunnel cross ditch damage evaluation method and system, the scheme is guided model attention minority class by introducing category frequency adjustment factor and designing reward function, improve the evaluation accuracy of small category cross ditch damage;Based on multilevel sampling and adaptive base point adjustment, adapt to different scale damage, capture subtle cracks through small base point, large base point covers serious damage, reduces the rate of missed detection;Through frequency sensitive spatial attention, edge and texture features are strengthened, so that the identification sensitivity of early seepage water stain boundary minor damage is improved;Further improve cross ditch damage evaluation effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of damage assessment, in particular to a kind of intelligent urban tunnel cross ditch damage assessment method and system. BACKGROUND

[0002] Cross ditch damage assessment method is the technical means for judging damage type and degree by collecting urban tunnel cross ditch image, extracting crack, corrosion and other damage features, for identifying drainage structure anomaly.But the general cross ditch damage assessment method exists due to the imbalance of urban tunnel cross ditch image class, leading to the evaluation of cross ditch damage, especially in the case that the rare damage class of corrosion, deformation is relatively less, state prediction is inaccurate, and the evaluation effect is poor;General cross ditch damage assessment method exists for improper evaluation of different scale damage of cross ditch, small scale damage is prone to be missed, large scale damage is prone to be disturbed by background, and then the evaluation effect of cross ditch damage is poor. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present application provides a kind of intelligent urban tunnel cross ditch damage assessment method and system, for the general cross ditch damage assessment method exists due to the imbalance of urban tunnel cross ditch image class, leading to the evaluation of cross ditch damage, especially in the case that the rare damage class of corrosion, deformation is relatively less, state prediction is inaccurate, and the evaluation effect is poor, the present application introduces class frequency adjustment factor by designing reward function, guide model to pay attention to minority class, improve the evaluation accuracy of small class cross ditch damage;For the general cross ditch damage assessment method exists for improper evaluation of different scale damage of cross ditch, small scale damage is prone to be missed, large scale damage is prone to be disturbed by background, and then the evaluation effect of cross ditch damage is poor, the present application is based on multilevel sampling and adaptive base point adjustment, adapts to different scale damage, reduces the missing rate by small base point to capture fine cracks and large base point to cover serious damage;Through frequency sensitive spatial attention, the recognition sensitivity of early leakage water stain boundary micro damage is improved by strengthening edge and texture features;Further improve the evaluation effect of cross ditch damage.

[0004] The technical scheme adopted by the present application is as follows: the present application provides a kind of intelligent urban tunnel cross ditch damage assessment method, which comprises the following steps:

[0005] Step S1: urban tunnel cross ditch image acquisition;

[0006] Step S2: cross ditch image dataset optimization;

[0007] Step S3: establish cross ditch damage assessment model;

[0008] Step S4: cross ditch damage assessment.

[0009] Further, in step S1, the urban tunnel cross-ditch image acquisition is the urban tunnel cross-ditch image, and the image is labeled with damage status labels; thus, a cross-ditch image dataset is obtained.

[0010] Furthermore, in step S2, the optimization of the transverse trench image dataset specifically includes the following steps:

[0011] Step S21: Design the reward function; train a pre-trained model using a convolutional neural network for preliminary classification of images of urban tunnel cross ditches; define the loss function as cross-entropy loss; and design the reward function. Introducing a category frequency adjustment factor , is represented as: Input image; y is the true category label; γ is the probability that the reinforcement learning strategy network selects the cross groove image of category a; γ is the attention coefficient. It is the category weight; It is the frequency adjustment coefficient; couny(·) is the frequency of occurrence of the category in the selected image set;

[0012] Step S22: The loss function L for optimizing the transverse trench image dataset is expressed as: ;in, These are the parameters of the pre-trained model; It is a pre-trained model policy network with parameters of of The output represents the predicted probability of classifying the transverse trench image; It is the input feature of the u-th transverse groove image; These are real labels; It is the initial adjustment coefficient; It is the output of the reinforcement learning strategy, i.e., the image that selects the category; is the output of the pre-trained model; 'a' is the selected category; 'x' is the input transverse groove image. These are adjustment parameters; t is the current training step number;

[0013] Step S23: Reinforcement Learning Fine-tuning; By transforming the task into a reinforcement learning problem, the agent continuously optimizes the selection policy in the transverse groove image selection task, ultimately obtaining an optimized set of transverse groove images; each transverse groove image is selected through a policy network, and a reward value is obtained based on the selection result to further optimize the policy; represented as: ;in, It is the expected reward; It is the expectation of all images x in the image set D; For time step t, the strategy The expectation of the cross-groove image pair log-probability of the selected category a; is an estimated value of the advantage function, representing the pros and cons of the reinforcement learning strategy selecting the category relative to selecting the cross-groove image compared with other categories at time step t; when the loss converges, the cross-groove image dataset optimization is completed, and the final cross-groove image dataset is obtained.

[0014] Further, in step S3, the establishment of the cross-groove damage evaluation model is based on the final cross-groove image dataset, and the establishment of the cross-groove damage evaluation model is realized on the basis of the VGG network architecture; specifically including the following steps:

[0015] Step S31: multi-level sampling; according to the scale of the cross-groove damage, the cross-groove image is divided into different layers of feature maps, and the base points are uniformly distributed; for each layer, the number of negative base points is proportionally distributed; for each feature layer, the number of non-damaged areas is proportionally distributed, the total number of negative base points is m, and the selected number of negative base points is m / 5; the base point size is adaptively adjusted, and the base point size s is adjusted to ; wherein, and are the width and height of the image respectively; is the initial base point size;

[0016] Step S32: simplify channel attention; the input feature map of the cross-groove image is re-calibrated based on the simplified channel attention module, represented as: ; ; B=BN(P); ; S=B×V; A=Sigmoid(S); ; wherein, P is the global average pooling result of the channel attention; GAP is the global average pooling; F is the feature map; is the average value of the cth channel; H and W are the height and width of the channel, and i1 and j1 are the corresponding indexes; is the pixel value; B is the batch normalization result; BN is the batch normalization operation; is the value of the normalized channel; is the channel mean value; is the standard deviation; is the smoothing term; γ and β are scaling and offset parameters respectively; S is the channel weight calculation result; V is the channel weight; Sigmoid(·) is the Sigmoid activation function; is the feature map after channel recalibration;

[0017] Step S33: frequency-sensitive spatial attention; the edge and texture features of the damaged area are strengthened, and spatial recalibration is performed through local-local interaction; represented as: ; ; ; ; ; ; wherein, Low is a low-frequency part feature map; High is a high-frequency part feature map; LPF is a low-pass filter; Sim is a spatial similarity; Z is an activation function ReLU processing result; and b is a bias term; is an original similarity feature map; joint simplified channel attention and frequency-sensitive spatial attention , improve the performance on different cross-gutter images, and obtain a final attention output , which is expressed as: ; wherein, and are attention weights;

[0018] Step S34: constructing a loss function; constructing a contrastive learning loss , introducing a texture adjustment factor , the contrastive learning loss is expressed as: ; wherein, is the number of final cross-gutter image datasets; i and k are image indexes; and are the real categories of the i-th image and the j-th image, respectively; 、 and are the feature vectors of the i-th image, the j-th image and the k-th image, respectively; is a temperature coefficient; i2 and j2 are pixel indexes; and a base point cross-entropy loss is constructed , which is expressed as: ; wherein, is the number of base points; is the target classification score of the i-th base point; is the actual category of the base point; and a final loss function is expressed as: ; wherein, is a loss weight; and LBP(·) is a local binary pattern.

[0019] Further, in step S4, the cross-gutter damage evaluation is that the cross-gutter image of the urban tunnel is collected in real time and input into the established cross-gutter damage evaluation model, and the cross-gutter damage is evaluated based on the label category output by the cross-gutter damage evaluation model.

[0020] The intelligent urban tunnel cross-gutter damage evaluation system provided by the application comprises an urban tunnel cross-gutter image acquisition module, a cross-gutter image dataset optimization module, a cross-gutter damage evaluation model establishment module and a cross-gutter damage evaluation module.

[0021] The urban tunnel cross ditch image acquisition module acquires cross ditch images and labels the images with damage state labels to obtain a cross ditch image dataset;

[0022] The cross ditch image dataset optimization module optimizes the image dataset through reinforcement learning and designs a reward function to balance the category distribution.

[0023] The cross ditch damage evaluation model establishment module establishes a cross ditch damage evaluation model based on a VGG network architecture, combined with multi-level sampling, channel attention, and frequency-sensitive spatial attention.

[0024] The cross ditch damage evaluation module performs real-time cross ditch damage evaluation based on the cross ditch damage evaluation model.

[0025] The above-mentioned scheme has the following beneficial effects:

[0026] (1) For the problem that general cross ditch damage evaluation methods are inaccurate in state prediction and poor in evaluation effect due to the imbalance of urban tunnel cross ditch image categories, which leads to poor evaluation of cross ditch damage, especially in the case of relatively few rare damage categories such as rust and deformation, the present scheme introduces a category frequency adjustment factor by designing a reward function to guide the model to pay attention to minority categories and improve the evaluation accuracy of small category cross ditch damage.

[0027] (2) For the problem that general cross ditch damage evaluation methods are not suitable for evaluating different scale damages of cross ditches, which leads to poor evaluation effect of cross ditch damage, such as easy to miss small scale damage and easy to be disturbed by background for large scale damage, the present scheme is based on multi-level sampling and adaptive base point adjustment to adapt to different scale damages, small base points are used to capture fine cracks and large base points are used to cover serious damage to reduce the miss rate, frequency-sensitive spatial attention is used to strengthen edge and texture features to improve the recognition sensitivity of small damage of early leakage water stains, and thus the evaluation effect of cross ditch damage is improved. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A flowchart of an intelligent urban tunnel cross ditch damage evaluation method provided by the present application is shown.

[0029] Figure 2 A schematic diagram of an intelligent urban tunnel cross ditch damage evaluation system provided by the present application is shown.

[0030] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application; based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of the present application.

[0032] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the system or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0033] Embodiment one, refer to Figure 1 The present application provides a kind of intelligent urban tunnel cross ditch damage assessment method, which comprises the following steps:

[0034] Step S1: urban tunnel cross ditch image acquisition;Collect cross ditch image, and the image is labeled damage state label, and cross ditch image dataset is obtained;

[0035] Step S2: cross ditch image dataset optimization;The image dataset is optimized by reinforcement learning, and the reward function is designed to balance the category distribution;

[0036] Step S3: establish cross ditch damage assessment model;On the basis of VGG network architecture, combined with multi-level sampling, channel attention and frequency sensitive spatial attention, cross ditch damage assessment model is established;

[0037] Step S4: cross ditch damage assessment;Based on cross ditch damage assessment model, real-time cross ditch damage assessment is carried out.

[0038] Embodiment two, refer to Figure 1 The embodiment is based on the above embodiment, in step S1, the urban tunnel cross ditch image acquisition is urban tunnel cross ditch image, and the image is labeled damage state label, and the damage state label category is normal cross ditch, crack cross ditch, blocked cross ditch, broken cross ditch, leakage cross ditch, settlement cross ditch, corrosion cross ditch, deformation cross ditch;And the image is scaled, sheared, and cross ditch image dataset is obtained.

[0039] Embodiment three, refer to Figure 1The embodiment is based on the above embodiment, in step S2, the cross-gully image dataset optimization is to optimize the class imbalance cross-gully image dataset, the reward function is designed by reinforcement learning, and the image set is adjusted to balance the class distribution; specifically including the following steps:

[0040] Step S21: design the reward function; use a convolutional neural network to train a pre-trained model for preliminary classification of urban tunnel cross-gully images; define the loss function as cross-entropy loss; and design the reward function , considering the class imbalance in urban tunnel cross-gully images, especially in the minority classes of corrosion and deformation, to enhance the reward function to better guide the model to focus on these classes, and introduce a class frequency adjustment factor , so that the model is more inclined to select those classes that are less in the current selected images, denoted as: ; y is the real class label; is the probability of the reinforcement learning strategy network selecting the cross-gully image of class a; γ is the attention coefficient; is the class weight; the class weight of the minority class is higher than that of the majority class; is the frequency adjustment coefficient; couny(·) is the frequency of the class in the selected image set;

[0041] Step S22: the loss function L of the cross-gully image dataset optimization is represented as: ; wherein, is the pre-trained model parameter; is the output of the pre-trained model strategy network when the parameter is , representing the predicted probability of cross-gully image classification; is the input feature of the u-th cross-gully image; is the real label; is the initial adjustment coefficient; is the reinforcement learning strategy output, i.e. the selected image of the class; is the pre-trained model output; a is the selected class; x is the input cross-gully image; is the adjustment parameter; t is the current training step;

[0042] Step S23: reinforcement learning fine-tuning; by converting the task into a reinforcement learning problem, the agent continuously optimizes the selection strategy in the cross-gully image selection task, and finally obtains the optimized cross-gully image set; each cross-gully image is selected by the strategy network, and the reward value is obtained according to the selection result to further optimize the strategy; denoted as: ; wherein, is the expected reward; ​is the expectation for all images x of the image set D; is the policy selecting the expected value of the cross-gutter image pair log probability of the category a; is the estimated value of the advantage function, indicating the pros and cons of the reinforcement learning policy selecting the category relative to selecting the cross-gutter image compared to other categories at time step t; when the loss converges, the final cross-gutter image dataset is optimized, and the final cross-gutter image dataset is obtained.

[0043] By performing the above operations, for the general cross-gutter damage evaluation method, due to the imbalance of the city tunnel cross-gutter image categories, the state prediction is inaccurate and the evaluation effect is poor for the evaluation of cross-gutter damage, especially for the rare damage categories of rust and deformation, which are relatively small. The present scheme introduces a category frequency adjustment factor by designing a reward function to guide the model to pay attention to the minority class and improve the evaluation accuracy of small category cross-gutter damage.

[0044] Embodiment four, see Figure 1 This embodiment is based on the above-mentioned embodiment. In step S3, the cross-gutter damage evaluation model is established based on the final cross-gutter image dataset, and the cross-gutter damage evaluation model is established based on the VGG network architecture; specifically including the following steps:

[0045] Step S31: multi-level sampling; according to the scale of the cross-gutter damage, the cross-gutter image is divided into different layers of feature maps, the low layer corresponds to small scale damage detection, and the high layer corresponds to large scale damage detection, and the base points are uniformly distributed to realize the targeted capture of different size damages; for each layer, the number of negative base points is proportionally distributed; for each feature layer, the number of non-damage areas is proportionally distributed, the total number of negative base points is m, and the number of selected negative base points is m / 5, to avoid the flooding of non-damage area features to small scale damage; the base point size is adaptively adjusted to optimize the detection effect of different damages; the base point size s is adjusted to ; wherein, and are the width and height of the image respectively; is the initial base point size; through multi-level sampling, the missed detection of small scale damage is reduced: small base points adapt to small size damages such as fine cracks and rust spots, avoiding feature dilution caused by large base points; improve the recognition stability of large scale damage: large base points cover the complete area of large damage and large size damage, reducing the influence of background interference on feature extraction; balance the detection priority of different size damages: by layering and distributing base points, ensure that small scale early damage and large scale serious damage obtain balanced feature weights in model training, avoid neglecting small scale damage due to more prominent visual features of large scale damage;

[0046] Step S32: Simplified channel attention; the input feature map of the cross-groove image is recalibrated by the simplified channel attention module, the weight of each channel is learned, and the important channel features are focused on; denoted as: ; ; B = BN(P); ; S = B x V; A = Sigmoid(S); ; wherein P is the global average pooling result of the channel attention; GAP is the global average pooling; F is the feature map; is the average value of the cth channel; H and W are the height and width of the channel, and i1 and j1 are the corresponding indexes; is the pixel value; B is the batch normalization result; BN is the batch normalization operation; is the value of the normalized channel; is the channel mean value; is the standard deviation; is the smoothing term; γ and β are the scaling and offset parameters, respectively; S is the channel weight calculation result; V is the channel weight; Sigmoid(·) is the Sigmoid activation function; is the feature map after channel recalibration; the channel attention mechanism focuses on the key visual features of different damage types, thereby improving the recognition sensitivity to subtle or hidden damage;

[0047] Step S33: Frequency-sensitive spatial attention; by strengthening the edge and texture features of the damage area, the spatial positioning ability of the model to small cracks and early rust subtle damage is improved, and spatial recalibration is performed through local-local interaction; denoted as: ; ; ; ; ; ; wherein Low is the low-frequency part feature map; High is the high-frequency part feature map; LPF is the low-pass filter; Sim is the spatial similarity; Z is the activation function ReLU processing result; b is the bias term; is the original similarity feature map; the simplified channel attention and the frequency-sensitive spatial attention are combined to improve the performance on different cross-groove images, and the final attention output is obtained, denoted as: ; wherein and are attention weights;

[0048] Step S34: constructing a loss function; optimizing the feature representation of cross-groove damage to make similar damage features more similar and different damage features more distinct, while avoiding overfitting to irrelevant textures; constructing a contrastive learning loss , for optimizing the intra-class similarity and inter-class difference in the feature space, introducing a texture adjustment factor Suppressing the interference of concrete background texture, the contrastive learning loss is expressed as: ; wherein, is the number of the final cross-groove image dataset; i and k are image indices; and are the true classes of the ith image and the jth image, respectively; , and are the feature vectors of the ith image, the jth image, and the kth image, respectively; is the temperature coefficient; i2 and j2 are pixel indices; constructing a base point cross-entropy loss , for optimizing the classification accuracy of damage classes and focusing on damage prediction in the base point area, expressed as: ; wherein, is the number of base points; is the target classification score of the ith base point; is the actual class of the base point; the final loss function is expressed as: ; wherein, is the loss weight; LBP(·) is the local binary pattern.

[0049] By performing the above operations, the problem of poor cross-groove damage evaluation effect due to improper evaluation of different scale damages of cross-grooves in general cross-groove damage evaluation methods, easy to miss small scale damage and easy to be disturbed by background for large scale damage, is solved. Based on multi-level sampling and adaptive base point adjustment, the scheme adapts to different scale damages, reduces the miss rate by capturing fine cracks with small base points and covering severe damage with large base points, enhances the recognition sensitivity of small damage in the boundary of early leakage water stains by strengthening edge and texture features through frequency sensitive spatial attention, and further improves the cross-groove damage evaluation effect.

[0050] Embodiment five, see Figure 1 This embodiment is based on the above-mentioned embodiments. In step S4, the cross-groove damage evaluation is real-time collection of city tunnel cross-groove images and input into the established cross-groove damage evaluation model. The label categories output by the cross-groove damage evaluation model are used to evaluate the cross-groove damage. If the label categories output by the cross-groove damage evaluation model are any one of crack cross-groove, blocked cross-groove, damaged cross-groove, leakage cross-groove, settlement cross-groove, corrosion cross-groove, and deformation cross-groove, an early warning is given to the management personnel.

[0051] Embodiment six, refer to Figure 2 , based on the above embodiment, the application provides an intelligent urban tunnel cross ditch damage evaluation system, which comprises an urban tunnel cross ditch image acquisition module, a cross ditch image data set optimization module, a cross ditch damage evaluation model establishment module and a cross ditch damage evaluation module;

[0052] The urban tunnel cross ditch image acquisition module acquires cross ditch images and labels the images with damage state labels to obtain a cross ditch image data set.

[0053] The cross ditch image data set optimization module optimizes the image data set through reinforcement learning, and designs a reward function to balance the category distribution.

[0054] The cross ditch damage evaluation model establishment module establishes a cross ditch damage evaluation model based on the VGG network architecture, combined with multi-level sampling, channel attention and frequency sensitive spatial attention.

[0055] The cross ditch damage evaluation module performs real-time cross ditch damage evaluation based on the cross ditch damage evaluation model.

[0056] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0057] Although the embodiments of the application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the application.

[0058] The above describes the application and its embodiments, which are not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if those skilled in the art are inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical scheme can be designed, which should belong to the protection scope of the application.

Claims

1. An intelligent method for assessing damage to cross trenches in urban tunnels, characterized in that: The method includes the following steps: Step S1: Image acquisition of cross trenches in urban tunnels; acquire images of cross trenches in tunnels and label the images with damage status tags to obtain a cross trench image dataset; Step S2: Optimize the transverse trench image dataset; optimize the image dataset through reinforcement learning and design a reward function to balance the class distribution; Step S3: Establish a cross trench damage assessment model; Based on the VGG network architecture, establish a cross trench damage assessment model by combining multi-level sampling, channel attention, and frequency-sensitive spatial attention; Step S4: Cross-cut trench damage assessment; Real-time cross-cut trench damage assessment based on the cross-cut trench damage assessment model; Step S2 includes designing the reward function. Introducing a category frequency adjustment factor , is represented as: Where y is the true category label; γ is the probability that the reinforcement learning strategy network selects the cross groove image of category a; γ is the attention coefficient. It is the category weight; It is the frequency adjustment coefficient; couny(·) is the frequency of occurrence of the category in the selected image set; Step S3 includes frequency-sensitive spatial attention design, which enhances the edge and texture features of the damaged area and performs spatial recalibration through local-local interactions; represented as: ; ; ; ; ; Where Low represents the low-frequency feature map; High represents the high-frequency feature map; and LPF is a low-pass filter. This is the feature map after channel recalibration; Sim is the spatial similarity; Z is the result of ReLU activation function processing; b is the bias term; It is the original similarity feature map; joint simplified channel attention Frequency-sensitive spatial attention To obtain the final attention output , is represented as: ;in, and It is attention weight; Step S4 involves multi-level sampling, which divides the transverse trench image into feature maps of different layers based on the scale of the damage, and evenly distributes base points. For each layer, the number of negative base points is proportionally allocated; for each feature layer, the number of undamaged regions is proportionally allocated, with a total of m negative base points and a selection quantity of m / 5. The base point size is adaptively adjusted, with the base point size s adjusted to... ;in, and These are the width and height of the image, respectively. It is the initial base point size.

2. The intelligent urban tunnel cross-sectional ditch damage assessment method according to claim 1, characterized in that: In step S2, the optimization of the transverse trench image dataset specifically includes the following steps: Step S21: Design the reward function; train a pre-trained model using a convolutional neural network for preliminary classification of images of urban tunnel cross ditches; define the loss function as cross-entropy loss; Step S22: The loss function L for optimizing the transverse trench image dataset is expressed as: ;in, These are the parameters of the pre-trained model; It is a pre-trained model policy network with parameters of of The output represents the predicted probability of classifying the transverse trench image; It is the input feature of the u-th transverse groove image; These are real labels; It is the initial adjustment coefficient; It is the output of the reinforcement learning strategy, i.e., the image that selects the category; is the output of the pre-trained model; 'a' is the selected category; 'x' is the input transverse groove image. These are adjustment parameters; t is the current training step number; Step S23: Fine-tuning reinforcement learning.

3. The intelligent urban tunnel cross-sectional ditch damage assessment method according to claim 2, characterized in that: In step S23, the reinforcement learning fine-tuning involves transforming the task into a reinforcement learning problem. The agent continuously optimizes its selection strategy in the transverse groove image selection task, ultimately obtaining an optimized set of transverse groove images. Each transverse groove image is selected through a policy network, and a reward value is obtained based on the selection result to further optimize the policy. This is represented as: ;in, It is the expected reward; It is the expectation of all images x in the image set D; For time step t, the strategy The expected log probability of the cross groove image of category a; It is an estimate of the advantage function, representing the superiority of the reinforcement learning policy in selecting a class relative to selecting other classes of transverse groove images at time step t; when the loss converges, the optimization of the transverse groove image dataset is completed, and the final transverse groove image dataset is obtained.

4. The intelligent urban tunnel cross-cutting ditch damage assessment method according to claim 3, characterized in that: In step S3, the establishment of the transverse trench damage assessment model is based on the final transverse trench image dataset and is implemented on the basis of the VGG network architecture; specifically, it includes the following steps: Step S31: Multi-level sampling; Step S32: Simplify channel attention; perform channel recalibration on the input feature map of the transverse groove image based on the simplified channel attention module, as shown below: ; B = BN(P); S = B × V; A = Sigmoid(S); Where P is the global average pooling result of channel attention; GAP is the global average pooling; and F is the feature map. It is the average value of the c-th channel; H and W are the height and width of the channel, and i1 and j1 are the corresponding indices; B represents the pixel value; B is the batch normalization result; BN is the batch normalization operation. It is the value of the standardized channel; It is the channel mean; It is the standard deviation; This is the smoothing term; γ and β are the scaling and offset parameters, respectively; S is the calculated channel weight; V is the channel weight; Sigmoid(·) is the Sigmoid activation function; Step S33: Frequency-sensitive spatial attention design; Step S34: Construct the loss function.

5. The intelligent urban tunnel cross-sectional ditch damage assessment method according to claim 4, characterized in that: In step S3, the construction of the loss function is to construct the contrastive learning loss. Introducing texture adjustment factor The contrastive learning loss is expressed as: ;in, This represents the number of images in the final transverse trench image dataset; i and k are the image indices. and These are the true categories of the i-th and j-th images, respectively; , and These are the feature vectors of the i-th image, the j-th image, and the k-th image, respectively. i2 is the temperature coefficient; i2 and j2 are pixel indices; construct the base point cross-entropy loss. , is represented as: ;in, It is the number of base points; It is the target classification score of the i-th base point; The actual class of the base point; the final loss function Represented as: ;in, It is the loss weight; LBP(·) is the local binary mode.

6. The intelligent urban tunnel cross-cutting ditch damage assessment method according to claim 5, characterized in that: In step S4, the cross-ditch damage assessment involves real-time acquisition of cross-ditch images of urban tunnels and inputting them into the established cross-ditch damage assessment model. The damage to the cross-ditch is then assessed based on the label categories output by the cross-ditch damage assessment model.

7. An intelligent urban tunnel cross-ditch damage assessment system, used to implement the intelligent urban tunnel cross-ditch damage assessment method as described in any one of claims 1-6, characterized in that: It includes a module for acquiring images of cross trenches in urban tunnels, a module for optimizing cross trench image datasets, a module for establishing a cross trench damage assessment model, and a module for assessing cross trench damage. The urban tunnel cross-ditch image acquisition module acquires images of tunnel cross-ditches and labels the images with damage status tags to obtain a cross-ditch image dataset. The transverse trench image dataset optimization module optimizes the image dataset through reinforcement learning and designs a reward function to balance the class distribution. The trench damage assessment model building module establishes a trench damage assessment model based on the VGG network architecture, combining multi-level sampling, channel attention, and frequency-sensitive spatial attention. The cross trench damage assessment module performs real-time cross trench damage assessment based on the cross trench damage assessment model.

Images