Method and system for identifying apparent diseases of hydraulic concrete structure

Abstract

The invention discloses an apparent disease identification method and system for a hydraulic concrete structure, and the method comprises the steps: collecting and marking data, expanding the data, training a feature extraction model, training a feature dimension reduction model, training a classifier model, and the like. And identifying the disease type of the target hydraulic concrete structure through data analysis processing of the feature extraction model, the feature dimension reduction model and the classifier model. According to the method, the generative adversarial network is combined with the nonlinear significance feature enhancement module to increase sample diversity, and richness of model training data and generalization ability of the model are improved; dynamic adjustment is performed according to the characteristics of the current data, so that the feature extraction process is more flexible; key features are reserved during dimension reduction, and the effectiveness of feature representation and the calculation efficiency of the model are improved; the high-order neural network is combined with a local sensitivity adjustment strategy, so that the classifier can identify different disease types more accurately, and misjudgment is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy project inspection and analysis, and particularly to a method and system for identifying apparent diseases of hydraulic concrete structures. Background Technique

[0002] Leakage, corrosion, cracks, spalling, etc. are the most common apparent diseases of hydraulic concrete structures. These diseases can lead to a reduction in the safety performance of concrete structures or even failure. Therefore, timely and accurate identification of concrete apparent diseases has important engineering significance. Traditional methods for identifying concrete diseases mainly rely on manual inspection or simple image analysis techniques, which have problems such as high labor intensity, low detection efficiency, and insufficient identification accuracy.

[0003] In recent years, the technical research in this area in China has been continuously deepening and developing, and many innovative and feasible patent technology solutions have been formed. For example, in the patent application with the application number CN202410715443.2, a method for concrete crack identification and width quantification using laser thermography is proposed. By establishing a concrete detection platform, using an infrared thermal imager to record infrared videos of the wall surface of the concrete to be detected, using a fixed-frequency laser generator to heat the wall surface of the concrete to be detected, obtaining a sequence of thermal images of the concrete surface, and then processing the sequence of thermal images of the concrete surface, a concrete surface crack width-temperature quantification model is established to identify and quantify the defective parts of the concrete. Another example is the patent application with the application number CN202411019923.1, which proposes an automatic pouring control system based on visual recognition monitoring. By processing and analyzing the information parameters of the pouring object and the information parameters of the hopper, structural feature information and concrete quality information are obtained, and by performing pouring analysis on the structural feature information, the pouring state information corresponding to the hopper is obtained. Then, by performing control analysis on the pouring state information, pouring control information is obtained, and then control modifications are made according to the actual pouring control information during pouring. There is also the patent application with the application number CN202410901154.1, which proposes a semantic segmentation method for quantifying the surface bubble defects of fair-faced concrete based on an improved YOLOv5 model. By establishing and labeling a semantic segmentation dataset for the surface bubble defects of fair-faced concrete, these datasets are divided into a training set and a validation set; a deep learning semantic segmentation model for identifying the surface bubble defects of fair-faced concrete structures is constructed; the constructed semantic segmentation model is trained and validated; the validated semantic segmentation model is used for quantitative identification of bubble defects; finally, the maximum diameter and area of the bubble defects in the identified image are extracted. Although these new solutions for the identification and detection methods and systems of hydraulic concrete structures have each made certain progress in terms of technical effects, there are still some relatively obvious technical problems to be solved, mainly reflected in: when using a generative adversarial network to generate in the existing solutions, no optimization is carried out for the specific features of concrete disease images, and the generated samples lack sufficient diversity and prominent key features, resulting in poor data augmentation effects and limiting the generalization ability of the model; when the existing feature extraction neural network processes complex data, it often encounters problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions, resulting in low model training efficiency and easy loss of useful information during feature extraction; the existing dimensionality reduction algorithms cannot effectively distinguish important features from secondary features, easily causing information loss or interference from redundant features, thus affecting the performance and computational efficiency of subsequent classification tasks; when the existing classifier models process concrete disease identification, most of them fail to dynamically adjust the contributions of each part in the loss function and are difficult to flexibly optimize the model according to the error sources, resulting in low classification accuracy, especially when dealing with multi-category diseases, the misjudgment rate is relatively high.These technical problems still need to be studied and new methods and systems designed to solve them in order to improve the effect of identifying apparent diseases in hydraulic concrete structures. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method for identifying apparent diseases in hydraulic concrete structures, comprising the following steps:

[0005] S1, data collection and annotation: Collect surface images of hydraulic concrete structures, and use annotation tools for accurate annotation to form a detailed annotation library for training and validating deep learning models;

[0006] S2, data augmentation: Use a non-linear saliency feature enhancement module to generate training samples;

[0007] S3, feature extraction model training: Input the augmented data into the feature extraction model for training the feature extraction model;

[0008] S4, feature dimensionality reduction model training: Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model;

[0009] S5, classifier model training: Input the data after dimensionality reduction into the classifier for training the classifier model;

[0010] S6, collect surface images of the target hydraulic concrete structure, and identify the disease type of the target hydraulic concrete structure through data analysis and processing by the feature extraction model, feature dimensionality reduction model, and classifier model.

[0011] As a further description of the present invention, step S2 includes:

[0012] S201, initialize the parameters of the generative adversarial network;

[0013] S202, perform non-linear saliency feature extraction;

[0014] S203, the generator generates new images according to the enhanced feature map and existing model parameters;

[0015] S204, the discriminator evaluates the difference between the generated image and the real image;

[0016] S205, adjust the parameters of the generator and discriminator according to the feedback of the discriminator;

[0017] S206, repeat the above steps iteratively until the preset stop iteration condition is met.

[0018] Furthermore, step S2 includes:

[0019] S301, initialize according to the bionic algorithm initialization method;

[0020] S302. For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function;

[0021] S303. According to the fitness of the individuals, select the best-performing individuals from the current population for retention as candidate solutions for the next generation;

[0022] S304. Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange some genes to generate new offspring;

[0023] S305. Repeat the above steps iteratively until the preset iteration stop condition is met.

[0024] Furthermore, the step S4 includes:

[0025] S401. Let the data input to the autoencoder neural network be X r , and the encoder adopts a multi-layer non-linear mapping structure to map the high-dimensional data to an initial low-dimensional feature space;

[0026] S402. After the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space;

[0027] S403. The feature adjustment module recursively optimizes the initially generated low-dimensional features;

[0028] S405. Repeat the above steps iteratively until the preset iteration stop condition is met.

[0029] Furthermore, the step S5 includes:

[0030] S501. Initialize the weights and biases of the high-order neural network;

[0031] S502. Input the data after dimensionality reduction and propagate it layer by layer through the network layers until the output layer;

[0032] S503. Combine local sensitivity to calculate the loss between the output and the actual label;

[0033] S504. Calculate the gradient of the loss function with respect to each parameter, and dynamically adjust the gradient update strategy according to the direction and magnitude of the gradient, as well as the performance of the parameter in the local sensitive area;

[0034] S505. Update the weights and biases in the network according to the adjusted gradient and learning rate;

[0035] S506. Repeat the above steps iteratively until the preset iteration stop condition is met.

[0036] On the other hand, the present invention also provides an apparent disease identification system for hydraulic concrete structures, which specifically includes a data acquisition module, a data augmentation module, a feature extraction model, a feature dimensionality reduction model, and a classifier model, where:

[0037] The data acquisition module is used to capture surface images of hydraulic concrete structures;

[0038] The data augmentation module is used to augment the surface images to generate training samples required for training the feature extraction model, the feature dimensionality reduction model, and the classifier model;

[0039] The feature extraction model is used to perform feature extraction processing on the surface images;

[0040] The feature dimensionality reduction model is used to perform feature dimensionality reduction processing on the data after feature extraction;

[0041] The classifier model is used to identify and classify concrete diseases for the data after feature dimensionality reduction.

[0042] Furthermore, the data acquisition module generates the training samples through a generative adversarial network combined with a non-linear saliency feature enhancement module.

[0043] Furthermore, the feature extraction model adopts a neural network algorithm based on dynamic population evolution optimization as the feature extraction model, automatically adjusts the evolution rules according to the characteristics of training data through a self-correction mechanism, and optimizes the neural network weight configuration through crossover and mutation operations.

[0044] Furthermore, the feature dimensionality reduction model adopts an autoencoder neural network based on feature refinement as the dimensionality reduction model. The autoencoder neural network based on feature refinement consists of an encoder, a decoder, and a feature adjustment module. The encoder is responsible for mapping high-dimensional input data to a low-dimensional feature space, and the decoder is used to reconstruct the reduced-dimensional features back to the original space.

[0045] Furthermore, the classifier model adopts a high-order neural network model as the classifier model, and the high-order neural network is based on an adaptive adjustment mechanism of the gradient direction.

[0046] Advantages of the present invention:

[0047] 1. By combining a generative adversarial network with a non-linear saliency feature enhancement module to generate more disease image samples to address the problem of insufficient actual training data, the non-linear saliency feature enhancement module highlights the key regions in the disease images, enabling the generator to more accurately simulate the disease characteristics and generate training data with stronger diversity and higher quality, thus solving the problems of data scarcity and model overfitting. The combination of the generative adversarial network and the non-linear saliency feature enhancement module makes the generated disease image samples increase sample diversity while maintaining a sense of reality, thereby effectively enhancing the richness of the model training data and the generalization ability of the model.

[0048] 2. Adopt a neural network algorithm based on dynamic population evolution optimization, and automatically adjust the evolution rules according to the characteristics of the training data through a self-correction mechanism, effectively avoiding the problems of gradient disappearance, gradient explosion, and local optimal solutions. This algorithm optimizes the weight configuration of the neural network through crossover and mutation operations, making feature extraction more efficient and enhancing the learning ability of the model. The feature extraction algorithm based on dynamic population evolution optimization can be dynamically adjusted according to the characteristics of the current data, making the feature extraction process more flexible, avoiding problems such as gradient disappearance and local optimal solutions in traditional neural networks, and enhancing the adaptability of the model to complex data.

[0049] 3. Use an autoencoder neural network based on feature refinement for feature dimensionality reduction. The encoder maps high-dimensional features to a low-dimensional space. The feature adjustment module reinforces important features and weakens secondary features through recursive feature adaptive optimization, ensuring the simplicity of the features and the computational efficiency of the model while retaining effective information during the dimensionality reduction process. The use of an autoencoder neural network based on feature refinement retains key features while performing dimensionality reduction, enhancing the effectiveness of feature representation and the computational efficiency of the model, and ensuring the support of the dimensionality reduction process for the classification task.

[0050] 4. The classifier uses a high-order neural network with an adaptive adjustment mechanism based on the gradient direction, combined with a local sensitivity adjustment strategy. By analyzing the error sources, it dynamically adjusts the contributions of each part in the loss function. Through this strategy, the generalization ability and accuracy of the model are improved, ensuring that the classifier has better robustness and flexibility when dealing with different types of diseases. The combination of the high-order neural network and the local sensitivity adjustment strategy makes the classifier show higher classification accuracy and robustness when dealing with the identification of concrete diseases, and can more accurately identify different disease types and reduce misjudgments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flowchart of the method for identifying the apparent diseases of the hydraulic concrete structure in the embodiment of the present invention;

[0052] Figure 2 It is a training flowchart of the neural network algorithm based on dynamic population evolution optimization in the embodiment of the present invention;

[0053] Figure 3 This is the training flowchart of the auto - encoding neural network algorithm based on feature refinement in the embodiments of the present invention. Detailed implementation manners

[0054] The following will describe the embodiments of the present invention in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0055] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", "first", "second", etc. indicate the orientation or position or sequential relationship based on the orientation or position or sequential relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation of the present invention.

[0056] A method for identifying the apparent diseases of hydraulic concrete structures of the invention includes the following steps:

[0057] S1. Data acquisition and annotation: Collect the surface images of the hydraulic concrete structures, and use annotation tools for accurate annotation to form a detailed annotation library for training and validating the deep learning model. Based on the method of the present invention, the system for identifying the apparent diseases of the hydraulic concrete structures of the present invention includes a data acquisition module for capturing the surface images of the hydraulic concrete structures. The data acquisition of the present invention can be based on drones and pipe tunnel inspection robots, etc. These devices are equipped with high - resolution cameras and sensors and are specifically used for capturing the surface images of the hydraulic concrete structures. The collected image data is in RGB format, and each image consists of three parts: the original color image, the image with enhanced contrast, and the grayscale image. The annotation is accurately performed by experienced engineers using annotation tools to form a detailed annotation library for training and validating the deep learning model.

[0058] S2. Data augmentation: Use a non - linear saliency feature enhancement module to generate training samples. Based on the method of the present invention, the system for identifying the apparent diseases of the hydraulic concrete structures of the present invention includes a data augmentation module for augmenting the surface images to generate the training samples required for training the feature extraction model, the feature dimensionality reduction model, and the classifier model.

[0059] It is understandable that in the task of the present invention, the acquisition, annotation, and preprocessing of training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model at the same time. The present invention adopts a generative adversarial network algorithm based on non-linear saliency feature enhancement for sample generation, thereby achieving data augmentation. Traditional generative adversarial networks are mainly used to generate visually realistic images. By adopting a non-linear saliency feature enhancement module, the present invention can more accurately simulate and enhance the key features in the concrete surface disease images, allowing the model to focus on enhancing the regions that have the greatest impact on the classification decision in the images, so as to generate more high-quality and diverse training samples.

[0060] Specifically, the training process of the generative adversarial network algorithm based on non-linear saliency feature enhancement is as follows:

[0061] S201. Initialize the parameters of the generative adversarial network. In one embodiment, the network parameters of the generator and the discriminator are initialized randomly, and the initialized values follow a uniform distribution in the interval (0, 1).

[0062] S202. Perform non-linear saliency feature extraction. In each training batch, use non-linear saliency mapping on the real images to highlight the key regions of the concrete surface diseases, and use this as the input focus of the generator. Specifically, the extraction method of the saliency feature map is expressed as:

[0063]

[0064] where S c represents the saliency feature map; I c is the input original image; Sigc() is the sharpness-based Sigmoid activation function; * represents the convolution operation; W k,c and b k,c are the convolution kernel and the bias respectively; α k，c is the feature fusion weight of the k-th scale; Kc is the number of feature scales; Relu() is the rectified linear unit activation function, which is used to enhance the non-linear expression ability of the model and suppress the influence of negative values.

[0065] In one embodiment, the calculation method of the sharpness-based Sigmoid activation function is expressed as:

[0066]

[0067] where x is the input of the function, K c is the parameter that adjusts the sharpness of the saliency response, and τ c is the threshold, which determines the sensitivity of the saliency activation.

[0068] S203. The generator generates new images based on the enhanced feature map and the existing model parameters. These images should be visually similar to the real apparent disease images but contain random variations to increase data diversity. Specifically, the generator of the present invention uses the saliency feature map and the random noise vector to generate new images through the structure fusion module, which is expressed as:

[0069]

[0070] In the formula, γ c and β c are learned fusion parameters used to balance the influence of the saliency feature and the random noise, and γ c is the first fusion parameter, and β c is the second fusion parameter; Tanh() is the hyperbolic tangent function, which enhances the expressiveness of the noise vector; and respectively represent the weights and biases of the generator; G c () is the generator function; z c is the random noise vector.

[0071] Furthermore, the fusion parameters of the generator balance the generated images based on the saliency feature and the noise vector, and the calculation method is expressed as:

[0072]

[0073] In the formula, and are learning parameters used to map the saliency feature to the fusion weight, and, is the weight of the first fusion parameter, is the bias of the first fusion parameter; Sof() is the Softmax function; Nc is the number of samples input in the current batch; and are the parameters that map the noise vector to the fusion weight, and, is the weight of the second fusion parameter, is the bias of the second fusion parameter; and respectively represent the i-th element of the saliency feature map and the noise vector.

[0074] S204. The discriminator evaluates the difference between the generated image and the real image, especially in the region where the saliency feature is enhanced, and the discrimination method is expressed as:

[0075]

[0076] In the formula, λ c is the dynamically adjusted weight.

[0077] In one embodiment, the dynamically adjusted weights are used to balance the influence of the generated image and the real image during the discrimination process to improve the flexibility and accuracy of discrimination. The calculation method is expressed as:

[0078]

[0079] In the formula, represents the concatenation of vectors; and are learning parameters used to dynamically adjust the weight λ based on the features of the input image c , thereby more finely controlling the response of the discriminator, and, is the first parameter for dynamic adjustment, is the second parameter for dynamic adjustment.

[0080] S205. According to the feedback of the discriminator, adjust the parameters of the generator and the discriminator. The optimization algorithm is implemented through gradient descent. The calculation method of the loss function during the adversarial training process is expressed as:

[0081]

[0082]

[0083] In the formula, is the loss function of the generator, is the loss function of the discriminator; D c () is the discriminator function; δ c is a parameter used to adjust the variability of the generated image; Var() represents the variance of the generated image, which is used to promote image diversity; Hinge() is the hinge loss function, which is used to provide a more stable and robust optimization target; represents the expectation.

[0084] In one embodiment, the calculation method of the parameter used to adjust the variability of the generated image is expressed as:

[0085]

[0086]

[0087] In the formula, σ 2 () represents the variance function, and μ() represents the mean function.

[0088] S206. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0089] S3. Feature extraction model training: Input the augmented data into the feature extraction model for training the feature extraction model. Based on the method of the present invention, the apparent disease recognition system for hydraulic concrete structures of the present invention includes a feature extraction model and a feature dimensionality reduction model, which are used to perform feature extraction processing on the surface image. Input the augmented data into the feature extraction model for training the feature extraction model. The present invention uses a 6-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions may occur, affecting the stability of training and the performance of the model. The present invention uses a neural network algorithm based on dynamic population evolution optimization as the feature extraction model. In the traditional population evolution algorithm, the evolution of all individuals is based on fixed rules. The present invention uses a self-correction mechanism to automatically adjust the evolution rules according to the characteristics of the current training data, and adjusts the probabilities of crossover and mutation according to the change trend of the loss function, so that the algorithm can more flexibly adapt to different data distributions, improving the generalization ability and training efficiency of the model. Specifically, for the training flow chart of the neural network algorithm based on dynamic population evolution optimization, see Appendix Figure 2 as shown

[0090] S301. According to the initialization method of the bionic algorithm, in the initialization stage, generate an initial population, and each individual represents a configuration of network weights. Specifically, let the population size be N p , initialize the weights and biases for the i-th individual, expressed as:

[0091]

[0092] In the formula, W pi is the weight of the neural network corresponding to the i-th individual, b pi is the bias of the neural network corresponding to the i-th individual, represents the weight matrix of the i-th individual in the initial state; represents the bias of the i-th individual in the initial state; σ 2 represents the initialized variance; represents a normal distribution with a mean of 0 and a variance of σ 2 ; is a normal distribution. Preferably, σ 2 is set to 0.01.

[0093] S302. For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function. Specifically, for the i-th individual, calculate the loss on the training data set using its weights and biases, expressed as:

[0094]

[0095] In the formula, L pi is the loss of the neural network corresponding to the i-th individual; mps represents the number of samples input in the current batch; l() represents the composite loss function, and fsig() represents the neural network model function; represents the feature of the j-th sample; represents the label of the j-th sample.

[0096] In one embodiment, the composite loss function includes a regularization term, which can increase the generalization ability of the model. The calculation method is expressed as:

[0097]

[0098] In the formula, MSE() is the mean square error function, Reg(W pi ) is the regularization term, and λ ps is the regularization parameter.

[0099] Furthermore, the calculation method of the regularization term is expressed as:

[0100]

[0101] In the formula, W pi，k is the k-th weight of the neural network corresponding to the i-th individual.

[0102] S303. Select the best-performing individual from the current population according to the fitness of the individual for retention as the candidate solution for the next generation. Specifically, selection is based on the fitness of the individual, and excellent individuals have a higher probability of being selected. The calculation method of the probability of an individual being selected is expressed as:

[0103]

[0104] In the formula, P select (i) represents the probability that the i-th individual is selected; γ pse is the parameter controlling the selection pressure; L pk is the loss of the neural network corresponding to the k-th individual. Preferably, γ pse is set to 2.

[0105] S304. Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange part of their genes to generate new offspring; the mutation operation randomly changes part of the genes in the individual to increase the diversity of the population. Specifically, the crossover operation randomly selects two individuals for gene exchange, which is expressed as:

[0106] W′ pi =α pcs W p1+(1 - α pcs )W p2

[0107] b′ pi = α pcs b p1 +(1 - α pcs )b p2

[0108] where α pcs is the crossover rate, W p1 is the weight of the neural network corresponding to the first selected individual, b p1 is the bias of the neural network corresponding to the first selected individual, W p2 is the weight of the neural network corresponding to the second selected individual, b p2 is the bias of the neural network corresponding to the second selected individual, W′ pi is the weight of the neural network corresponding to the individual after the crossover operation, b′ pi is the bias of the neural network corresponding to the individual after the crossover operation. Preferably, α pcs is set to 0.3.

[0109] Moreover, the mutation operation performs a small - amplitude random perturbation on the weights of the newly generated individuals, expressed as:

[0110]

[0111] where τ 2 represents the variance of the mutation, W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, is the bias of the neural network corresponding to the individual after the mutation operation. Preferably, τ 2 is set to 0.04.

[0112] S305. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0113] S4. Feature reduction model training. Input the data after feature extraction into the feature reduction model for training the feature reduction model. Based on the method of the present invention, the apparent disease identification system for hydraulic concrete structures of the present invention includes a feature reduction model for performing feature reduction processing on the data after feature extraction.

[0114] The data after feature extraction is input into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses an autoencoder neural network based on feature refinement as the dimensionality reduction model. The autoencoder neural network based on feature refinement consists of three parts: an encoder, a decoder, and a feature adjustment module. The encoder is responsible for mapping the high-dimensional input data into a low-dimensional feature space, and the decoder is used to reconstruct the dimensionality-reduced features back to the original space to ensure the reversibility of the dimensionality reduction process. The feature adjustment module dynamically adjusts the dimensionality-reduced feature space through recursive feature adaptive optimization, so that important features can be strengthened and secondary features are gradually weakened, so that the dimensionality-reduced feature representation has better simplicity while retaining data information.

[0115] Specifically, for the training flowchart of the autoencoder neural network algorithm based on feature refinement, see the appendix Figure 3 as shown.

[0116] S401. Let the data input into the autoencoder neural network be X r , and the encoder adopts a multi-layer non-linear mapping structure to map the high-dimensional data into the initial low-dimensional feature space, which is expressed as:

[0117] Z r = Sig enc (W r X r + b r )

[0118] In the formula, Z r represents the initial low-dimensional feature, W r is the weight matrix of the encoder, b r is the bias vector of the encoder, and Sig enc () is the multi-layer Sigmoid activation function of the encoder.

[0119] S402. After the low-dimensional features are generated, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space. When the module is initialized, the same initial weight is assigned to all features for subsequent adjustment according to the feature contribution. The calculation method of the initial weight matrix is expressed as:

[0120] A r = dig(α r )

[0121] In the formula, A r is the initial weight matrix; α r is the feature weight vector, and each element α r in α r，i is initialized to the same value, indicating that all features have the same importance in the initial stage; diag() is a function for extracting the diagonal elements of the matrix.

[0122] Furthermore, the adjusted feature can be expressed as:

[0123] Z′ r = A r Z r

[0124] In the formula, Z′ r is the feature representation after feature weight adjustment.

[0125] S403. The feature adjustment module recursively optimizes the preliminarily generated low-dimensional features. In each round of iteration, the module adjusts the weights of each feature according to the performance of the previous round of features, gradually enhancing the features with important influences and gradually weakening redundant or noisy features. Let the weight update rule in the t-th round of iteration be as follows:

[0126]

[0127] In the formula, represents the feature weight in the (t + 1)-th round of iteration, represents the feature weight in the t-th round of iteration, η r is the learning rate of the autoencoder neural network, L r () is the loss function of the autoencoder neural network, Y r is the label data, represents the gradient of the loss function with respect to the feature weight. Preferably, the loss function of the autoencoder neural network is the reconstruction error loss function. Preferably, η r is set to 0.01.

[0128] S404. To ensure the effectiveness of the dimensionality reduction process, the decoder remaps the low-dimensional features back to the high-dimensional space to ensure that important information is not lost during the dimensionality reduction process. The reconstruction process of the decoder is expressed as:

[0129]

[0130] In the formula, X′ r is the reconstructed high-dimensional data, W′ r is the weight matrix of the decoder, b′ r is the bias vector of the decoder, Sig dec () is the multi-layer Sigmoid activation function of the decoder.

[0131] S405. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0132] S5. Classifier model training: Input the data after dimensionality reduction into the classifier to train the classifier model. Based on the method of the present invention, the apparent disease identification system for hydraulic concrete structures of the present invention includes a classifier model for identifying and classifying concrete diseases for the data after feature dimensionality reduction.

[0133] Input the data after dimensionality reduction into the classifier to train the classifier model. The present invention uses a high-order neural network model as the classifier model, and the high-order neural network is based on an adaptive adjustment mechanism in the gradient direction. The present invention adopts a local sensitivity adjustment strategy for the loss function. The loss function not only considers the overall error but also analyzes the source of the error, and adopts different adjustment strategies for errors from different sources, so that the model has better generalization ability and accuracy.

[0134] Specifically, the training process of the high-order neural network classification algorithm is as follows:

[0135] S501. Initialize the weights and biases of the high-order neural network. The initialization method is expressed as:

[0136]

[0137] In the formula, θ (0) represents the initial value of the network parameters; σ init is the initialization standard deviation; represents a Gaussian distribution with a mean of 0 and a covariance of the identity matrix. Preferably, σ init is set to 0.01.

[0138] S502. Input the data after dimensionality reduction and propagate it layer by layer through the network layer until the output layer. The output of each layer is calculated based on the output of the previous layer and the parameters of the current layer. The process of forward propagation is expressed as:

[0139] a (l) = Sig (l) (W (l) h (l-1) + b (l) )

[0140] In the formula, a (l) is the activation value of the l-th layer; Sig (l) is the activation function of the l-th layer, specifically using the Sigmoid activation function; W (l) and b (l) are the weights and biases of the l-th layer respectively; σ init (l-1) is the output of the previous layer. For the input layer, h (0) is the input feature.

[0141] S503. Calculate the loss between the output combined with local sensitivity and the actual label, and the calculation method is expressed as:

[0142]

[0143] In the formula, is the total loss; L() is the cross-entropy loss function; y i and are the true label and the predicted label of sample i respectively; S() represents the corresponding local sensitivity loss contribution; λ sens is the regularization parameter that controls the influence of the local sensitivity loss. Preferably, λ sens is set to 0.1.

[0144] S504. Calculate the gradient of the loss function with respect to each parameter. According to the direction and magnitude of the gradient, and the performance of the parameter in the locally sensitive area, dynamically adjust the gradient update strategy. The calculation method of the gradient is expressed as:

[0145]

[0146] In the formula, ΔW (l) is the update amount of the weight W (l) ; η fc is the learning rate; is the partial derivative of the loss function with respect to the weight; μ fc is the momentum factor to accelerate the learning process. Preferably, η fc is set to 0.1, and μ fc is set to 3.

[0147] Furthermore, the calculation method of the gradient is expressed as:

[0148]

[0149] Furthermore, the calculation method of the momentum factor μ is expressed as:

[0150]

[0151] In the formula, is the initial value of the momentum factor; β vk is the adjustment coefficient; t is the number of training iterations. Preferably, is set to 0.2, and β vk is set to 3.

[0152] Furthermore, the calculation of the local sensitivity loss S() is expressed as:

[0153]

[0154] where Ne(i) represents the set of neighboring samples of the i-th sample; w ij is the similarity between sample i and sample j; (y i -y j ) 2 is the square of the label difference.

[0155] S505. Update the weights and biases in the network according to the adjusted gradient and learning rate, and the update method is expressed as:

[0156]

[0157] where is the updated weight.

[0158] S506. Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0159] S6. Collect the surface image of the target hydraulic concrete structure, and identify the disease type of the target hydraulic concrete structure through the data analysis and processing of the feature extraction model, feature dimensionality reduction model and classifier model.

[0160] Based on the above method and system of the present invention, the present invention combines a generative adversarial network with a non-linear saliency feature enhancement module, so that the generated disease image samples increase sample diversity while maintaining a sense of reality, thereby effectively improving the richness of model training data and the generalization ability of the model; the feature extraction algorithm based on dynamic population evolution optimization can be dynamically adjusted according to the characteristics of the current data, making the feature extraction process more flexible, avoiding problems such as gradient disappearance and local optimal solutions in traditional neural networks, and improving the adaptability of the model to complex data; the auto-encoder neural network based on feature refinement retains key features while reducing dimensions, improving the effectiveness of feature representation and the computational efficiency of the model, and ensuring the support of the dimensionality reduction process for the classification task; the high-order neural network combines a local sensitivity adjustment strategy, making the classifier show higher classification accuracy and robustness when dealing with concrete disease identification, being able to more accurately identify different disease types and reducing misjudgment.

[0161] The above is only an illustration of the preferred embodiments of the present invention, but it should not be construed as a limitation of the claims. The present invention is not limited to the above embodiments, and its specific structure is allowed to change. In short, all changes made within the protection scope of the independent claims of the present invention are within the protection scope of the present invention.

Claims

1. A method for identifying apparent diseases of hydraulic concrete structures, characterized in that It includes the following steps: S1. Data collection and annotation: Collect the surface images of hydraulic concrete structures, and use annotation tools for precise annotation to form a detailed annotation library for training and validating deep learning models; S2. Data augmentation: Adopt a non-linear saliency feature enhancement module to generate training samples; S3. Feature extraction model training: Input the augmented data into the feature extraction model for training the feature extraction model; S4. Feature dimensionality reduction model training: Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model; S5. Classifier model training: Input the data after dimensionality reduction into the classifier for training the classifier model; S6. Collect the surface images of the target hydraulic concrete structure, and identify the disease types of the target hydraulic concrete structure through the data analysis and processing of the feature extraction model, feature dimensionality reduction model, and classifier model.

2. The method for identifying the apparent diseases of hydraulic concrete structures according to claim 1, wherein: The step S2 includes: S201. Initialize the parameters of the generative adversarial network; S202. Perform non-linear saliency feature extraction; S203. The generator generates new images according to the enhanced feature mapping and existing model parameters; S204. The discriminator evaluates the difference between the generated image and the real image; S205. According to the feedback of the discriminator, adjust the parameters of the generator and the discriminator; S206. Repeat the above steps iteratively until the preset stop iteration condition is met.

3. The method for identifying the apparent diseases of hydraulic concrete structures according to claim 1, characterized in that: The step S2 includes: S301. Initialize according to the bionic algorithm; S302. For each individual in the population, use its corresponding neural network configuration to process the input training data, calculate the output of the model, and evaluate its performance according to a predetermined loss function; S303. According to the fitness of the individual, select the best-performing individual from the current population for retention as a candidate solution for the next generation; S304. Generate new individuals through crossover and mutation operations. The crossover operation allows two excellent individuals to exchange part of their genes to generate new offspring; S305. Repeat the above steps iteratively until the preset stop iteration condition is met.

4. The method for identifying the apparent diseases of hydraulic concrete structures according to claim 1, wherein: The step S4 includes: S401. Let the data input to the autoencoder neural network be X r , and the encoder adopts a multi-layer non-linear mapping structure to map the high-dimensional data to the initial low-dimensional feature space; S402. After generating low-dimensional features, the feature adjustment module automatically generates feature weights according to the feature importance in the current feature space; S403. The feature adjustment module recursively optimizes the initially generated low-dimensional features; S405. Repeat the above steps iteratively until the preset stop iteration condition is met.

5. The method for identifying the apparent diseases of hydraulic concrete structures according to claim 1, characterized in that: The step S5 includes: S501. Initialize the weights and biases of the high-order neural network; S502. Input the data after dimensionality reduction and propagate it layer by layer through the network layer until the output layer; S503. Combine local sensitivity to calculate the loss between the output and the actual label; S504. Calculate the gradient of the loss function with respect to each parameter, and dynamically adjust the gradient update strategy according to the direction and magnitude of the gradient, as well as the performance of the parameter in the local sensitive area; S505. Update the weights and biases in the network according to the adjusted gradient and learning rate; S506. Repeat the above steps iteratively until the preset stop iteration condition is met.

6. A system for identifying apparent diseases of hydraulic concrete structures, characterized in that: It includes a data acquisition module, a data augmentation module, a feature extraction model, a feature dimensionality reduction model, and a classifier model; The data acquisition module is used to capture the surface images of hydraulic concrete structures; The data augmentation module is used to augment the surface images to generate training samples required for training the feature extraction model, the feature dimensionality reduction model, and the classifier model; The feature extraction model is used to perform feature extraction processing on the surface images; The feature dimensionality reduction model is used to perform feature dimensionality reduction processing on the data after feature extraction; The classifier model is used to perform concrete disease identification and classification on the data after feature dimensionality reduction.

7. The apparent disease identification system for hydraulic concrete structures according to claim 6, wherein: The data acquisition module generates the training samples through a generative adversarial network combined with a non-linear saliency feature enhancement module.

8. The apparent disease identification system for hydraulic concrete structures according to claim 6, wherein: The feature extraction model adopts a neural network algorithm based on dynamic population evolution optimization as the feature extraction model, automatically adjusts the evolution rules according to the characteristics of the training data through a self-correction mechanism, and optimizes the neural network weight configuration through crossover and mutation operations.

9. The apparent disease identification system for hydraulic concrete structures according to claim 6, wherein: The feature dimensionality reduction model adopts an autoencoder neural network based on feature refinement as the dimensionality reduction model. The autoencoder neural network based on feature refinement consists of three parts: an encoder, a decoder, and a feature adjustment module. The encoder is responsible for mapping high-dimensional input data to a low-dimensional feature space, and the decoder is used to reconstruct the reduced-dimensional features back to the original space.

10. The apparent disease identification system for hydraulic concrete structures according to claim 6, wherein: The classifier model adopts a high-order neural network model as the classifier model, and the high-order neural network is based on an adaptive adjustment mechanism in the gradient direction.

Images