Water conservancy project concrete building surface damage identification method and system and electronic equipment

Abstract

The invention relates to the technical field of hydraulic engineering safety inspection, and provides a hydraulic engineering concrete building surface damage identification method and system and electronic equipment. The method comprises the following steps: acquiring a concrete building surface damage image, and performing feature extraction on the concrete building surface damage image by using a damage feature extraction module to obtain concrete building surface damage features; performing feature dimension filtering on the surface damage features of the concrete building by using a discriminant feature index module to obtain discriminant feature data; and using a trained integrated learning classifier to identify the discrimination feature data to obtain a damage category corresponding to the concrete building surface damage image. According to the method provided by the invention, three damages of cracks, fragmentation and holes and a normal type of image can be identified with fewer training samples, and the method has high damage identification performance.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy project safety inspection, and in particular to a method, system and electronic equipment for identifying surface damage of concrete buildings in water conservancy projects. Background Art

[0002] Water conservancy projects have a high level of social and public safety. Once quality problems occur, irreversible consequences may occur, such as cracks, deformation, or collapse in the dam or levee, posing a significant threat to people's lives and property. Therefore, structural safety inspections of water conservancy projects are a key component of their safety and management. Damage assessments of hydraulic concrete structures are a preliminary step in the risk prevention, reinforcement, and repair of water conservancy projects. Promptly identifying damage to hydraulic concrete structures ensures preparedness and ensures the safe operation of water conservancy projects. Water conservancy project safety inspections primarily rely on traditional manual inspections and intelligent inspection technologies. Manual inspections are subject to numerous limitations, such as high cost, low efficiency, and low safety.

[0003] In intelligent inspection technology for surface damage in concrete structures of hydraulic engineering projects, traditional machine learning-based approaches require manual extraction of damage image features, such as texture, edge, color, and shape features. These features are then classified using machine learning classification algorithms, such as support vector machines (SVMs). While traditional machine learning methods offer specificity, controllability, and interpretability, they also suffer from high workload, high subjectivity, and difficulty capturing complex image features. Consequently, these methods suffer from relatively low damage identification accuracy and poor robustness.

[0004] Intelligent damage recognition models based on deep convolutional neural networks (CNNs) do not require manual image feature extraction. Their hierarchical structure can capture complex image information while exhibiting strong robustness and generalization capabilities. Because CNN models contain multiple layers and complex weight structures, they require a large amount of labeled data for model training. With the continuous advancement of deep learning technology, it has been proposed to transfer the knowledge learned by pre-trained models (such as ResNet, VGG, and Xcenpet) on large datasets to the surface damage recognition task of hydraulic concrete structures. Through fine-tuning, these models can be made more adaptable to the target task, avoiding the need to train a new model from scratch and reducing the cost of building labeled datasets and model deployment. Despite this, these methods still require hundreds or even thousands of training images to update the network parameters. The expansion of recognition patterns often means an exponential increase in the amount of labeled data required for training, which results in limited scalability of existing models. To further improve the performance of crack damage recognition, it is proposed to extract crack damage image features using pre-trained models. Cracks and non-cracks are then classified using algorithms such as the K-nearest neighbor algorithm, support vector machines, or decision trees. The accuracy of crack recognition is effectively improved by integrating pre-trained models and traditional machine learning methods, but the recognition mode is relatively simple.

[0005] Existing deep learning-based surface damage identification technologies for hydraulic concrete structures focus on diagnosing crack damage. This single recognition model limits their application scenarios. Although cracks are a common type of surface damage in hydraulic concrete structures, it is undeniable that models with a single functional model cannot provide comprehensive decision-making support for hydraulic structure safety assessments and maintenance management. Summary of the Invention

[0006] In view of the problem that the existing technology of concrete building surface damage identification focuses on the diagnosis of crack damage, resulting in a single identification mode and limited application scenarios, the present invention proposes a method for identifying surface damage of hydraulic engineering concrete buildings. By performing discriminant analysis on the extracted features, discriminative dimensions are selected, the sample dimension is reduced, the types of surface damage identification of hydraulic concrete buildings are increased, and the damage identification efficiency is improved.

[0007] In a first aspect, the present invention provides a method for identifying surface damage of a concrete building in a hydraulic engineering project, comprising:

[0008] Step 1: Acquire a concrete building surface damage image, and use a damage feature extraction module to extract features from the concrete building surface damage image to obtain concrete building surface damage features;

[0009] Step 2: Using a discriminant feature index module to filter the surface damage features of the concrete building to obtain discriminant feature data;

[0010] Step 3: Using the trained ensemble learning classifier to identify the discriminant feature data, and obtain the damage category corresponding to the concrete building surface damage image;

[0011] The damage feature extraction module, the discriminant feature index module and the ensemble learning classifier are trained in advance using sample concrete building surface damage recognition images.

[0012] Furthermore, the training process of the damage feature extraction module includes:

[0013] Obtaining surface damage identification images of sample concrete buildings;

[0014] The sample concrete building surface damage recognition image is used to fine-tune the pre-trained lightweight convolutional network, and the classification layer of the fine-tuned pre-trained lightweight convolutional network is removed to obtain the damage feature extraction module.

[0015] Furthermore, the pre-trained lightweight convolutional network adopts a pre-trained RegNetY-800MF model;

[0016] During training, the parameters of the last Bottleneck block layer and the classification layer of the pre-trained RegNetY-800MF model are fine-tuned.

[0017] Furthermore, the training process of the discriminant feature index module includes:

[0018] splicing the sample concrete building surface damage features corresponding to the sample building surface damage identification image to obtain a damage feature matrix;

[0019] Inputting the damage feature matrix into a damage feature discriminant analysis model to perform discriminant analysis to obtain a row sparse projection matrix;

[0020] The row sparse projection matrix is ​​input into the discriminant feature index module, and an index array of concrete building surface damage features is generated by obtaining the subscripts corresponding to non-zero rows in the row sparse projection matrix.

[0021] Furthermore, the use of a discriminant feature index module to filter the surface damage features of the concrete building by feature dimension to obtain discriminant feature data includes:

[0022] Inputting the concrete building surface damage features into the discriminant feature index module;

[0023] Feature dimension filtering is performed based on the index array, dimensions with index subscripts are selected and dimensions without index subscripts are removed to obtain discriminant feature data.

[0024] Furthermore, the damage feature discrimination analysis model is:

[0025]

[0026] Among them, min means minimization, α represents a hyperparameter, W represents the projection matrix, γ represents a hyperparameter, ||W|| 2,1 Indicates L 2,1 norm, μ i represents the mean vector within the i-th damage category, μ j represents the mean vector within the jth class of damage, represents the jth sample of the i-th damage category, n i Indicates the number of samples of the i-th type of damage in the damage category, n j represents the number of samples of the jth type of damage in the damage category, n represents the total number of samples of damage images, and c represents the number of damage categories.

[0027] Furthermore, the iterative reweighting method is used to optimize the damage feature discrimination analysis model to obtain:

[0028]

[0029] Among them, s ij Indicates the last iteration The value of , that is, in the previous iteration, the sample within the class and class mean μ k The residual value, g ij Indicates the last iteration The value in , that is, the mean μ of the i-th class in the previous iteration i and the j-th class mean μ j The residual value is calculated by the following formula:

[0030]

[0031] Furthermore, the base learner of the ensemble learning classifier is a decision tree model, and the ensemble method adopts the bagging method.

[0032] In a second aspect, the present invention provides a surface damage identification system for concrete buildings in water conservancy projects, comprising:

[0033] a feature extraction unit, configured to input a concrete building surface damage image, and extract features from the concrete building surface damage image using a damage feature extraction module to obtain concrete building surface damage features;

[0034] A discriminant feature unit, configured to perform feature dimension filtering on the surface damage features of the concrete building using a discriminant feature index module to obtain discriminant feature data;

[0035] a classification and recognition unit, configured to use a trained ensemble learning classifier to recognize the discriminant feature data and obtain a damage category corresponding to the concrete building surface damage image;

[0036] The damage feature extraction module, the discriminant feature index module and the ensemble learning classifier are trained in advance using sample concrete building surface damage recognition images.

[0037] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described above when executing the program.

[0038] Beneficial effects of the present invention:

[0039] The present invention first uses a small amount of surface damage image data from hydraulic concrete structures to fine-tune the hidden layer parameters of the final module of a lightweight pre-trained model. This is then used to extract damage image features. Secondly, the present invention utilizes an ensemble learning algorithm to robustly classify the damage images. To address the time-consuming nature of ensemble learning algorithms when dealing with high-dimensional samples, the present invention proposes performing discriminant analysis on the extracted feature data. This aims to select discriminative dimensions, reduce sample dimensionality, and improve the efficiency of hydraulic concrete structure surface damage identification.

[0040] Compared to machine learning-based surface damage identification methods for hydraulic concrete structures, this method eliminates the need for manual extraction of damage image features, reduces the influence of subjective factors, and improves the intelligent level of surface damage identification for hydraulic concrete structures. Compared to deep learning-based surface damage identification methods for hydraulic concrete structures, the method proposed in this paper can identify three types of damage—cracks, fragments, and holes—as well as normal images with fewer training samples, demonstrating superior damage identification performance.

[0041] The proposed method achieved an average accuracy of 96.71%, with a standard deviation of 0.01 across 10 experimental results. The inference time required for 200 test images was only 85.5 milliseconds. In summary, the proposed method exhibits good scalability and high recognition accuracy. This scalability is reflected in the model's flexibility in pattern expansion. Therefore, the proposed method can provide algorithmic support for intelligent surface damage inspection of concrete structures in water conservancy projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic flow chart of a method for identifying surface damage of concrete buildings in water conservancy projects provided by an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of a technical route for identifying surface damage in concrete buildings of a water conservancy project provided by an embodiment of the present invention;

[0044] Figure 3 The module training architecture of the surface damage identification technology for concrete buildings in water conservancy projects provided by the embodiment of the present invention;

[0045] Figure 4 Schematic diagram of the pre-trained RegNetY-800MF network structure provided by an embodiment of the present invention;

[0046] Figure 5 A schematic diagram of the pre-trained ResNet-18 network structure provided in an embodiment of the present invention;

[0047] Figure 6 A schematic diagram of the pre-trained MobileNet-v3-large network structure provided in an embodiment of the present invention;

[0048] Figure 7 A schematic diagram of the pre-trained EfficientNet-B8 network structure provided in an embodiment of the present invention;

[0049] Figure 8 A schematic diagram of damage image feature extraction provided by an embodiment of the present invention;

[0050] Figure 9 A schematic diagram of a discriminative feature indexing module provided in an embodiment of the present invention;

[0051] Figure 10 A schematic diagram of an integrated learning classification module provided in an embodiment of the present invention;

[0052] Figure 11 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention;

[0053] Figure 12 Schematic diagram of a high-fill open channel experimental model for a water conservancy project provided by an embodiment of the present invention;

[0054] Figure 13 A schematic diagram of surface damage of a high fill concrete slope provided by an embodiment of the present invention;

[0055] Figure 14 Schematic diagram of the loss value convergence and accuracy change of the model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] The embodiment of the present invention provides a method for identifying surface damage of concrete buildings in water conservancy projects, such as Figure 1 As shown, the specific steps include:

[0058] Step 1: Acquire a concrete building surface damage image, and use a damage feature extraction module to extract features from the concrete building surface damage image to obtain concrete building surface damage features;

[0059] Step 2: Use the discriminant feature index module to filter the feature dimension of the concrete building surface damage image features to obtain discriminant feature data;

[0060] Step 3: Use the trained ensemble learning classifier to identify the discriminant feature data and obtain the damage category corresponding to the concrete building surface damage image;

[0061] Among them, the damage feature extraction module, the discriminant feature indexing module and the ensemble learning classifier are pre-trained using sample concrete building surface damage recognition images.

[0062] The method provided by an embodiment of the present invention obtains the surface damage image features of the concrete building by inputting the surface damage image of the concrete building to be identified into the surface damage feature extraction model of the hydraulic concrete building. The damage image features are then filtered by the index array in the discriminant feature index module to generate discriminant feature data. Finally, the discriminant feature data is robustly classified using an ensemble learning classifier trained based on an ensemble learning algorithm. The method proposed by the present invention can identify three types of damage, namely cracks, fragmentation and cavities, as well as normal types of images, and has high damage recognition performance.

[0063] In another embodiment of the present invention, the training method and process of the damage feature extraction module, the discriminant feature index module and the ensemble learning classifier are explained. Figure 2 and Figure 3 As shown, the training is divided into three stages:

[0064] 1. Damage feature extraction module

[0065] In the first stage, considering the limited computing resources of the actual application deployment platform, a common lightweight convolutional network is selected for training. The pre-trained models are all trained on the large-scale and diverse ImageNet-1K dataset, which contains 1,000 categories and a total of about 128,000 images, covering common objects and scenes, such as people, buildings, and street scenes. The embodiment of the present invention transfers the parameters learned by the pre-trained lightweight convolutional network on the public dataset to the surface damage recognition of hydraulic concrete buildings, and uses the existing sample concrete building surface damage recognition images to fine-tune the parameters of the final module. The classification layer of the fine-tuned network structure is removed to obtain a damage feature extraction module. At this time, the input of the model is a two-dimensional RGB image, and the output is a one-dimensional hydraulic concrete surface damage feature vector, that is, the sample concrete building surface damage feature, which is spliced ​​to obtain a damage feature matrix.

[0066] like Figure 8 As shown in Figure 2, the training process of the damage feature extraction module can be represented by the following process:

[0067] Obtaining surface damage identification images of sample concrete buildings;

[0068] The pre-trained lightweight convolutional network is fine-tuned using sample concrete building surface damage recognition images. The classification layer of the fine-tuned pre-trained lightweight convolutional network is removed to obtain the damage feature extraction module.

[0069] Furthermore, the pre-trained lightweight convolutional network of the embodiment of the present invention adopts the pre-trained RegNetY-800MF model, such as Figure 4 As shown in the red box, the parameters of the last Bottleneck block layer and the classification layer of the pre-trained RegNetY-800MF model are fine-tuned during training.

[0070] Optionally, the pre-trained lightweight convolutional network may also use a pre-trained ResNet-18 model, a pre-trained MobileNet-v3-large model, or a pre-trained EfficientNet-B8 model. Figure 5 As shown in the red box, when using the pre-trained ResNet-18 model, the last residual block and classification layer of the ResNet-18 network structure are fine-tuned; Figure 6 As shown in the red box, when using the pre-trained MobileNet-v3-large model, the last bneck layer and subsequent network layers of the MobileNet-v3-large network structure are fine-tuned; Figure 7As shown in the red box, when using the pre-trained EfficientNet-B8 model, the last Inverted Residual layer and subsequent network layers of the EfficientNet-B8 network structure are fine-tuned.

[0071] 2. Damage feature discriminant analysis model and discriminant feature index module

[0072] In the second stage, the pre-trained lightweight convolutional network is typically trained on a specific dataset. This dataset differs significantly from the surface damage image data of concrete structures in hydraulic engineering projects. As a result, the damage feature extraction module cannot fully fit the data when extracting image features, resulting in noise. Therefore, this paper proposes to perform discriminant analysis on the extracted damage features to reduce sample dimensionality and improve class separability.

[0073] It is understandable that Linear Discriminant Analysis (LDA) is a supervised dimensionality reduction technique. Its core idea is to map high-dimensional data to low-dimensional space, maximize the distance between classes and minimize the distance within classes. Therefore, the samples after dimensionality reduction have good class separability. The inter-class distance and intra-class distance of samples are usually expressed by the inter-class divergence S. b and intra-class divergence S w To measure, the calculation formula is as follows:

[0074]

[0075]

[0076] Among them, μ i represents the mean vector within the i-th damage category, μ j represents the mean vector within the jth class of damage, represents the jth sample of the i-th damage category, n i Indicates the number of samples of the i-th type of damage in the damage category, n j represents the number of samples of the jth type of damage in the damage category, n represents the total number of samples of damage images, and c represents the number of damage categories.

[0077] LDA maximizes the inter-class distance by maximizing the inter-class divergence after projection, which is formally expressed as maxTr(WS b W T ), and at the same time, minimize the intra-class divergence after projection to minimize the distance between samples within the class, formally expressed as minTr(WS w W T ). The learned projection matrix can be used to project new samples from the d-dimensional original space to the l-dimensional subspace, where d>>l.

[0078] The damage feature discrimination analysis model constructed by the present invention is:

[0079]

[0080] Among them, min means minimization, Tr means trace number, W means row sparse projection matrix, S w represents the intra-class dispersion of damage categories, S b represents the inter-class divergence of damage categories, α represents a hyperparameter, γ represents a hyperparameter, ||W|| 2,1 Indicates L 2,1 norm.

[0081]

[0082] Where d represents the dimension of the input damage feature matrix, l represents the number of selected features, and is also the number of projection vectors (the number of columns in the projection matrix), and w ij Represents the elements of the projection matrix.

[0083] Obviously, the first item of the damage feature discriminant analysis model combines the principle of linear discriminant analysis. By minimizing the first item, the intra-class divergence of damage categories can be minimized and the inter-class divergence of damage categories can be maximized. The hyperparameter α is used to adjust the relative importance of the intra-class divergence of damage categories and the inter-class divergence of damage categories. In addition, the present invention introduces L in the second item. 2,1 Norm regularization constraint is used to impose row sparsity on the row sparse projection matrix W. In the process of minimizing the objective function (damage feature discriminant analysis model), when the hyperparameter γ is a large value, ||W|| 2,1 The smaller the value of will be, the more zero rows will appear in the projection matrix W. Because the rows of the projection matrix W correspond to the dimensions of the original data, it is determined that the dimensions corresponding to these zero rows do not play a role in the subspace projection process and can be eliminated.

[0084] The traditional intra-class divergence and inter-class divergence calculation formulas have potential square operations. As shown in Equations (5) and (6), they are easily affected by outliers when calculating the intra-class sample distance and inter-class distance.

[0085]

[0086]

[0087] The equivalent form (7) of the damage feature discrimination analysis model (3) can be derived from equations (5) and (6):

[0088]

[0089] Alternatively, observing the damage feature discriminant analysis model (7), in the process of learning the row sparse projection matrix, the square operation in the objective function can easily amplify the influence of outliers in the data. Therefore, the embodiment of the present invention uses the iterative reweighting method to optimize the objective function, and the following is obtained:

[0090]

[0091] Among them, s ij Indicates the last iteration The value of , that is, in the previous iteration, the sample within the class and class mean μ k The residual value, g ij Indicates the last iteration The value in , that is, the mean μ of the i-th class in the previous iteration i and the j-th class mean μ j The residual value is calculated by the following formula:

[0092]

[0093] In order to facilitate the solution of the projection matrix, the form of the above optimization problem is simplified and some terms are combined to obtain:

[0094]

[0095] It is understandable that the projection matrix W can be obtained through the eigenvalue decomposition process of the matrix M. The k eigenvalues ​​and eigenvectors obtained by eigenvalue decomposition of the matrix M are selected, and the eigenvectors corresponding to the first l smallest eigenvalues ​​are selected and arranged to form the projection matrix W. At this time, the projection matrix W has d row vectors and l column vectors.

[0096] The algorithm provided by the above discriminant feature analysis module can be represented by the pseudo code provided in Table 1.

[0097] Table 1 Discriminant analysis algorithm for hydraulic concrete characteristic data

[0098]

[0099] It can be understood from the pseudo code provided in Table 1 above that the discriminant feature analysis module mainly takes the data matrix X (damage feature matrix) as input and the row sparse projection matrix W as output. Based on this sparse matrix, the dimension index of the damage feature matrix can be obtained. Figure 9As shown, the discriminative feature indexing module takes the row sparse projection matrix W output by the discriminative model as input and generates an index array of surface damage features for hydraulic concrete structures by obtaining the subscripts corresponding to the non-zero rows. Based on this data, the dimensions of the newly input feature data are filtered, selecting dimensions with index subscripts while removing those without, to obtain the discriminative feature data. Clearly, the discriminative feature data has fewer dimensions than the feature data output in the previous stage, thus reducing the processing time of the ensemble learning classifier.

[0100] The training process of the above discriminative feature index module can be represented by the following process, including:

[0101] The surface damage features of the sample concrete building corresponding to the surface damage identification image of the sample building are spliced ​​to obtain a damage feature matrix;

[0102] Input the damage feature matrix into the damage feature discriminant analysis model to perform discriminant analysis and obtain a row sparse projection matrix;

[0103] The row sparse projection matrix is ​​input into the discriminant feature index module, and the index array of the surface damage features of the concrete building is generated by obtaining the subscripts corresponding to the non-zero rows in the row sparse projection matrix.

[0104] In another embodiment of the present invention, a discriminant feature index module is used to filter the surface damage features of concrete buildings to obtain discriminant feature data, including:

[0105] Input the surface damage features of concrete buildings into the discriminant feature index module;

[0106] Feature dimension filtering is performed based on the index array, selecting dimensions with index subscripts while removing dimensions without index subscripts to obtain discriminant feature data.

[0107] 3. Ensemble learning classification module

[0108] In another embodiment of the present invention, the base learner of the ensemble learning classifier adopts a decision tree model, and the ensemble method adopts a bagging method.

[0109] It is understandable that existing technologies for identifying surface damage in hydraulic concrete structures primarily rely on single classifiers, such as support vector machines, to classify extracted feature samples, resulting in relatively low classification accuracy. Ensemble learning classification algorithms offer significant advantages over single classification algorithms in terms of prediction accuracy, generalization, and noise immunity. Therefore, the present invention selects an ensemble learning classification algorithm as the classifier for the hydraulic concrete structure surface damage identification model. In this embodiment, the base learner of the ensemble learning classifier is a decision tree model, and the ensemble method employs the bagging method.

[0110] Alternatively, as Figure 10 As shown in FIG, the integrated learning classification module takes the feature vector output by the discriminant analysis module as input, outputs the corresponding predicted label, and finally inputs the labeled damage image into the storage medium.

[0111] The embodiment of the present invention further provides a surface damage identification system for concrete buildings in water conservancy projects, comprising:

[0112] The feature extraction unit is used to input the concrete building surface damage image, and use the damage feature extraction module to extract features from the concrete building surface damage image to obtain the concrete building surface damage features;

[0113] A discriminant feature unit is used to filter the surface damage features of the concrete building using a discriminant feature index module to obtain discriminant feature data;

[0114] A classification and recognition unit is used to use the trained ensemble learning classifier to identify the discriminant feature data and obtain the damage category corresponding to the concrete building surface damage image;

[0115] Among them, the damage feature extraction module, the discriminant feature indexing module and the ensemble learning classifier are pre-trained using sample concrete building surface damage recognition images.

[0116] like Figure 11 As shown, an embodiment of the present invention further provides an electronic device, including: a processor (processor) 1101, a communication interface (Communications Interface) 1102, a memory (memory) 1103 and a communication bus 1104, wherein the processor 1101, the communication interface 1102, and the memory 1103 communicate with each other via the communication bus 1104. The processor 1101 can call the logic instructions in the memory 1103 to execute the method in the above embodiment, which includes:

[0117] Step 1: Obtain a concrete building surface damage image, use the damage feature extraction module to extract features from the concrete building surface damage image, and obtain the concrete building surface damage features; Step 2: Use the discriminant feature indexing module to filter the feature dimension of the concrete building surface damage features to obtain discriminant feature data; Step 3: Use the trained ensemble learning classifier to identify the discriminant feature data to obtain the damage category corresponding to the concrete building surface damage image.

[0118] In addition, when the logic instructions in the above-mentioned memory 1103 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0119] To verify the effectiveness of the present invention, this example will conduct an experimental verification. The computer configuration used in the experiment is: Matlab 2022b, Windows 10, Intel (R) Core (TM) i7-7700HQ CPU and 8.00GB RAM. The pre-trained model training device configuration information is: NVIDIA V100 GPU, 32GB VRAM. It should be noted that the present invention does not rely on specific computer configuration conditions.

[0120] Data collection:

[0121] The characteristics of the surface damage image of hydraulic concrete structures are strong interference from background pixels and irregular shapes of damaged areas. Therefore, it is challenging to directly classify the images. In this embodiment, an open channel experimental model of a high fill channel was built. Figure 12 Based on the experimental model, this embodiment collected a set of high fill water-facing slope damage images. The dataset contains four categories: cracks, broken, holes, and normal, as shown in Figure 2. Figure 13 As shown. Each category has 100 images with a length and width of 150 pixels. The cracks are darker than the background area and have certain linear features, but the curvature, length and width of the lines are different. The fragmentation is composed of multiple cracks, which are specifically manifested as intersecting cracks. The width is wider than the cracks and the shape varies. The holes are in the shape of holes or holes in the cement panel, which are generally 2.5 to 10 mm wide and about 10 to 50 mm deep. The data set constructed in this embodiment simulates the phenomenon that the surface of water conservancy engineering buildings has aged and become rough due to long-term water erosion and sunlight exposure. In addition, the weeds and other debris that can be seen everywhere at the image acquisition site further increase the difficulty of identification.

[0122] Benchmark experiments:

[0123] In order to verify the applicability of the pre-trained model on the surface damage dataset of hydraulic concrete buildings, this embodiment carried out a benchmark experiment on the high fill concrete slope damage image dataset. In the benchmark experiment part, the hidden layer parameters are not fine-tuned, but the 1000 output units of the network structure classification layer are replaced with 4 output units. 40% of the samples are randomly selected as the training set, 10% of the samples are used as the validation set, and the remaining 50% of the samples are used as the test set, and then the classification layer parameters are retrained using the training set. The validation set is used to evaluate the status of the model, and the test set is used to test the generalization ability of the model. This embodiment uses the Adaptive Moment Estimation optimization algorithm to train the classification layer parameters, and the learning rate range is [10 -3 , 10 -1 After 10 random experiments, the mean and standard deviation of the accuracy, as well as the inference time required to infer the 200 test set images, are calculated. Figure 14 The results show the convergence of loss value and change of accuracy of ResNet-18, MobileNet-v3-large, EfficientNet-B8 and RegNetY-800MF models on the hydraulic concrete damage dataset as the number of training epochs increases.

[0124] Table 2 Mean and standard deviation of classification accuracy of ResNet, MobileNet, EfficientNet and RegNetY series models on the test set, as well as inference time on the test set

[0125]

[0126]

[0127] Table 2 shows that among the ResNet models, the less complex ResNet-18 model achieved the best fine-tuning results, achieving an average recognition accuracy of 87.15%. Among the MobileNet models, the MobileNet-v3-large model achieved the highest average recognition accuracy and the smallest standard deviation of accuracy across multiple experiments. This demonstrates that the MobileNet-v3-large model is the most stable and more suitable for the hydraulic concrete damage image dataset compared to other versions. Among the EfficientNet models, EfficientNet-B8 performed the best, achieving an average recognition accuracy of 83.85%. Therefore, this model was selected as the representative EfficientNet model. Among the RegNetY models, the RegNetY-800MF model performed particularly well, achieving a recognition accuracy of 90.75%. The RegNetY-12GF and RegNetY-16GF models are closest to RegNetY-800MF, with accuracy differences of approximately 1% to 2%. However, these models have significantly higher computational complexity than RegNetY-800MF, so RegNetY-800MF was selected as the representative method. Based on the above experiments, it's clear that a more complex network structure or a larger number of parameters doesn't necessarily translate into better performance. Instead, it's crucial to consider the model's performance on the target dataset to determine the most reliable model. To further explore the effectiveness of fine-tuning the pre-trained model, we used the model selected in the benchmark experiment to conduct further fine-tuning experiments, aiming to further improve the performance of hydraulic concrete surface damage identification.

[0128] like Figure 14 As shown, (a) is ResNet-18, (b) is MobileNet-v3-large, (c) is EfficientNet-B8, and (d) is RegNetY-800MF. Due to the lack of samples in the training data set, the fitting deviation of the loss values ​​of the model on the validation set and the training set is large, but they all tend to converge. As the number of epoch iterations increases, the accuracy on the training set tends to 100%, while the recognition accuracy on the validation set generally stabilizes between 80% and 90%. In order to avoid overfitting on the training set, this embodiment sets the number of network training rounds based on the convergence of the validation set accuracy. It can be seen from the results of the benchmark experiment that the performance of the model at this time is far from meeting the requirements of practical applications, so the performance of the model needs to be further improved.

[0129] Advanced experiments:

[0130] In order to verify the effectiveness of the technical route proposed in this study, we conducted a surface damage analysis on the hydraulic concrete building dataset. Figure 3Experiments were conducted using the presented technical route. We randomly selected 40% of the images from each class to form the training set, 10% of the images to form the validation set, and the remaining 50% of the images to form the test set. While fine-tuning the pre-trained model using the training set, we obtained feature samples corresponding to each training image, which were used to train the discriminant analysis model and the ensemble learning classifier. After completing the preparation of the surface damage recognition model for hydraulic concrete buildings, we tested the model using the test set images, and this process was performed independently 10 times. The hyperparameter settings of the pre-trained model adopted the benchmark experiment settings, while the hyperparameter α range of the discriminant feature selection model was set to [10 -5 ,10 5 ], the range of γ is set to [10 -2 ,10 2 In addition, we show the mean and standard deviation of the classification accuracy corresponding to the first 100 features. Ensemble learning uses the Bagging algorithm (Bootstrap Aggregating), which uses decision trees as base learners and enhances prediction capabilities by iteratively training multiple base learners.

[0131] Table 3 Mean and standard deviation of accuracy of advanced experiments, as well as inference time on the test set

[0132]

[0133]

[0134] The first set of experimental results in Table 3 shows that after fine-tuning the hidden layer parameters, the test accuracy of the ResNet, MobileNet-v3-large, and RegNetY models all improved by approximately 0.1% to 2% compared to the baseline. However, the accuracy of the EfficientNet-B8 model decreased by approximately 2.5%, likely due to... Overall, the fine-tuning experiments on the hydraulic concrete damage dataset demonstrated promising results. Incorporating an ensemble learning classification algorithm, the recognition accuracy of the ResNet-18 model improved by 2.27%, MobileNet-v3-large by 4.42%, EfficientNet-B8 by 10.27%, and the corresponding RegNetY-800MF model significantly increased by approximately 3.25%. Furthermore, the standard deviation of the classification accuracy across multiple experiments significantly improved, demonstrating that the ensemble learning algorithm significantly improved classification stability. Combining the ensemble learning classifier with a discriminative feature selection method further improved the classification accuracy by approximately 0.71% to 2.38%, while significantly reducing the time required to process 200 test images. Because the RegNetY-800MF network structure inherently includes an attention mechanism, which somewhat mitigates the influence of irrelevant features, the integration of the discriminative feature selection model resulted in a modest but not significant improvement in classification performance. However, the discriminative feature selection method significantly reduced the ensemble learning algorithm's inference time, from 425ms to 85.5ms, significantly enhancing the method's ability to process images in batches.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for identifying surface damage of concrete buildings in water conservancy projects, characterized in that: include: Step 1: Acquire a concrete building surface damage image, and use a damage feature extraction module to extract features from the concrete building surface damage image to obtain concrete building surface damage features; Step 2: Using a discriminant feature index module to filter the surface damage features of the concrete building to obtain discriminant feature data; Step 3: using the trained ensemble learning classifier to identify the discriminant feature data, and obtaining the damage category corresponding to the concrete building surface damage image; The damage feature extraction module, the discriminant feature index module and the integrated learning classifier are trained in advance using sample concrete building surface damage recognition images.

2. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 1, characterized in that: The training process of the damage feature extraction module includes: Acquire surface damage identification images of sample concrete buildings; The pre-trained lightweight convolutional network is fine-tuned using the sample concrete building surface damage recognition image, and the classification layer of the pre-trained lightweight convolutional network after fine-tuning is removed to obtain the damage feature extraction module.

3. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 2, characterized in that: The pre-trained lightweight convolutional network adopts a pre-trained RegNetY-800MF model; During training, the parameters of the last Bottleneck block layer and the classification layer of the pre-trained RegNetY-800MF model are fine-tuned.

4. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 1, characterized in that: The training process of the discriminant feature index module includes: The sample concrete building surface damage features corresponding to the sample building surface damage identification image are spliced ​​to obtain a damage feature matrix; Inputting the damage feature matrix into a damage feature discriminant analysis model for discriminant analysis to obtain a row sparse projection matrix; The row sparse projection matrix is ​​input into the discriminant feature index module, and an index array of surface damage features of concrete buildings is generated by obtaining the subscripts corresponding to non-zero rows in the row sparse projection matrix.

5. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 4, characterized in that: The method of using a discriminant feature index module to filter the surface damage features of the concrete building to obtain discriminant feature data includes: Inputting the concrete building surface damage features into the discriminant feature index module; Based on the index array, feature dimension filtering is performed, dimensions with index subscripts are selected and dimensions without index subscripts are removed to obtain discriminant feature data.

6. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 4, characterized in that: The damage feature discrimination analysis model is: Among them, min means minimization, α represents a hyperparameter, W represents the projection matrix, γ represents a hyperparameter, ||W|| 2,1 Indicates L 2,1 Norm, μ i represents the mean vector within the i-th damage category, μ j represents the mean vector within the jth class of damage category, represents the jth sample of the i-th damage category, n i represents the number of samples of the i-th damage in the damage category, n j represents the number of samples of the jth type of damage in the damage category, n represents the total number of samples of damage images, and c represents the number of damage categories.

7. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 6, characterized in that: The iterative reweighting method is used to optimize the damage feature discrimination analysis model, and the following is obtained: Among them, s ij Indicates the last iteration The value of is the number of samples within the class in the previous iteration. and class mean μ k The residual value, g ij Indicates the last iteration The value in , that is, the mean μ of the i-th class in the previous iteration i and the j-th class mean μ j The residual value is calculated by the following formula:

8. A method for identifying surface damage of a concrete building of a hydraulic engineering project according to claim 1, characterized in that: The base learner of the ensemble learning classifier is a decision tree model, and the ensemble method adopts the bagging method.

9. A surface damage identification system for concrete buildings in water conservancy projects, characterized in that: include: A feature extraction unit is used to input a concrete building surface damage image, and use a damage feature extraction module to extract features from the concrete building surface damage image to obtain concrete building surface damage features; A discriminant feature unit, used to filter the surface damage features of the concrete building using a discriminant feature index module to obtain discriminant feature data; A classification and recognition unit, used to recognize the discriminant feature data using a trained ensemble learning classifier to obtain a damage category corresponding to the concrete building surface damage image; The damage feature extraction module, the discriminant feature index module and the integrated learning classifier are trained in advance using sample concrete building surface damage recognition images.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

Images