Method and device for identifying dam landslide hazards based on artificial intelligence

Abstract

The present invention relates to an artificial intelligence-based method and device for identifying dam landslide hazards, belonging to the field of dam landslide hazard identification. The method comprises the following steps: collecting dam landslide data from actual dam monitoring points and manually labeling it to obtain a labeled dam landslide dataset; inputting the labeled dam landslide dataset into a trained generative adversarial network based on hierarchical reinforcement to obtain an expanded dam landslide dataset; inputting the expanded dam landslide dataset into a trained neural network based on adaptive convolution kernel optimization to extract features, obtaining extracted dam landslide data features; and inputting the extracted dam landslide data features into a trained support vector machine algorithm based on high-order feature interpolation to obtain dam landslide hazard identification results. This invention improves the accuracy and efficiency of feature extraction, thereby accurately identifying different landslide levels.

Description

Technical Field

[0001] The present invention belongs to the field of dam landslide hazard identification, and in particular relates to a dam landslide hazard identification method and device based on artificial intelligence. Background Art

[0002] With the global climate change and the frequent occurrence of extreme weather events, dam safety has gradually become an important issue in the management of water conservancy infrastructure. Dam landslides are a common and serious type of dam disease. If they are not identified and treated in a timely manner, they may pose a major threat to the surrounding ecological environment, the safety of residents' lives and property, and the stability of the dam itself. Existing methods for identifying dam landslide diseases include: landslide identification methods based on physical constraint machine learning: first collect slope data and save them, embed the physical constraints such as the dynamic equation of erosive landslide fluidity, boundary conditions, initial conditions, etc. into the neural network to learn and obtain a prediction model, and then collect the data of the slope to be tested and use the model to predict its landslide dynamic process, so that the model can better understand the physical meaning of the landslide for accurate prediction. Bedding landslide disaster risk assessment method based on machine learning: obtain regional topographic image data to identify bedding landslide risk points and build a sample library, use machine learning to identify risks in unknown areas to obtain geological disaster data, and solve the problem of low assessment accuracy. A landslide susceptibility assessment method based on CNN-DF was developed. The multicollinearity test and frequency ratio calculation of landslide influencing factors were carried out. The disaster-prone environmental factors were screened to construct similar environments and data sets. A model based on CNN and DF modules was constructed to predict landslide susceptibility and generate a susceptibility map. Finally, the model performance was evaluated.

[0003] In the task of identifying dam landslide hazards, the above technical solutions still have the following problems that need to be further resolved: 1. The identification of dam landslide hazards often relies on limited historical data or a single data source, resulting in insufficient training samples, which affects the generalization ability and accuracy of the model. 2. Traditional feature extraction methods ignore the temporal, multi-scale and complex spatial distribution relationships of dam landslide data, which can easily cause information loss or redundancy. A large amount of manual feature engineering is required when processing high-dimensional, nonlinear and noisy data. 3. The classification accuracy of dam landslide identification is not high, especially it is difficult to distinguish between different landslide grades such as mild and moderate landslides. 4. Some models perform poorly on different data sets. Faced with the complex and changing dam environment, traditional models find it difficult to adapt to new data distributions. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a method and device for identifying dam landslide hazards based on artificial intelligence.

[0005] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0006] The present invention provides an artificial intelligence-based method for identifying dam landslide hazards, comprising the following steps:

[0007] S1. Collect and manually label dam landslide data from actual dam monitoring points to obtain a labeled dam landslide dataset;

[0008] S2. The labeled dam landslide dataset is input into a trained hierarchical reinforcement-based generative adversarial network for data augmentation, thereby obtaining an augmented dam landslide dataset. The hierarchical reinforcement-based generative adversarial network comprises a generator and a discriminator. The generator employs a hierarchical reinforcement learning mechanism to optimize the quality of generated samples through a multi-layer decision-making process.

[0009] S3. Inputting the expanded dam landslide dataset into a trained neural network optimized based on an adaptive convolution kernel for feature extraction to obtain extracted dam landslide data features; the neural network optimized based on the adaptive convolution kernel includes an input layer, a convolution layer, a pooling layer, and a fully connected layer; the convolution layer is residually connected to form residual blocks, and each residual block is skipped, where the skip connection is a conventional residual neural network connection method in the art;

[0010] S4. Input the extracted embankment landslide data features into the trained support vector machine algorithm based on high-order feature interpolation to perform embankment landslide disease identification and classification, and obtain the embankment landslide disease identification results.

[0011] Furthermore, step S1 specifically includes:

[0012] The dam landslide data include geographical location Ra, current temperature Da, humidity value Ia, precipitation around the dam Pa, water level Ha, displacement Va, seepage pressure Sa, crack size Fa on the dam surface, accumulated rainfall La, and historical disease records Ca;

[0013] The labeling categories of embankment landslide data are “normal”, “minor landslide”, “moderate landslide” and “severe landslide”.

[0014] Furthermore, step S2 specifically includes:

[0015] The training process of the generative adversarial network based on hierarchical reinforcement is as follows:

[0016] S21. Initialize the network parameters of the generator and discriminator. The initialization strategy is to use a normal distribution with a mean of 0 and a variance of the identity matrix. The parameter initialization formula is as follows:

[0017] ,

[0018] ,

[0019] in, represents the parameters of the generator, represents the parameters of the discriminator, Indicates that it obeys a specific distribution. represents a normal distribution with mean 0 and variance as the identity matrix, represents the identity matrix;

[0020] S22. Pre-train the discriminator. The loss function formula of the discriminator in the pre-training stage is expressed as follows:

[0021] ,

[0022] in, represents the real dam landslide data, is the loss function of the discriminator, represents the input noise sampled from the noise distribution, represents the noise distribution, Express expectations, represents the discriminator function, represents a generator function, represents the distribution of real dam landslide data, represents the natural logarithm;

[0023] S23. During the adversarial training phase, first, the generator parameters are fixed. The real dam landslide data and the generated dam landslide data output by the generator are input into the discriminator for processing, and the discriminator parameters are updated. The generated dam landslide data output by the generator is obtained by inputting noise sampled from a noise distribution into the generator. Then, the discriminator parameters are fixed, the generator parameters are updated, and the generator adjusts its generation strategy using a hierarchical reinforcement learning strategy. The generator's loss function is expressed as follows:

[0024]

[0025] in, represents the loss function of the generator, represents the weight of the combined embedding loss, Denotes the combined embedding loss; the formula of the combined embedding loss is as follows:

[0026] ,

[0027] ,

[0028] ,

[0029] in, Indicates the number of samples generated in the current batch, represents the feature extraction function of the neural network layer of the generator, Represents the feature extraction function for the The nonlinear transformation function of feature dimensions, Indicates the characteristic dimension of the generated embankment landslide data; represents the first weight parameter, represents the second weight parameter, represents the boundary parameter, which is used to control the balance of losses; represents the L2 norm, Indicates taking the maximum value, Indicates the Real dam landslide data, Indicates the Input noise;

[0030] S24. Each layer of the generator is independently adjusted through reinforcement learning during training. During the training process of the generator, the reward for each adjustment is calculated based on the feedback from the discriminator and the embedding loss, and the formula is as follows:

[0031] ,

[0032] ,

[0033] in, represents the Sigmoid function, represents the hierarchical reinforcement learning rate parameter, Rewards representing adjustments to the generator's neural network layers;

[0034] S25. Repeat steps S22-S24 until the preset stop iteration condition is met, the training is completed, and a trained hierarchical reinforcement-based generative adversarial network is obtained.

[0035] Furthermore, step S3 specifically includes:

[0036] The training process of the neural network based on adaptive convolution kernel optimization is as follows:

[0037] S31. In the initial stage of training, the data in the expanded embankment landslide dataset is hierarchically processed by grouping the embankment landslide data into different categories and assigning them to different sub-networks. The formula is as follows:

[0038] ,

[0039] ,

[0040] in, Represents the neural network The dam landslide data output after layer processing, Represents the neural network The weight parameters of the layer, represents the input dam landslide data feature matrix of the neural network, represents the activation function of the neural network, Represents the neural network Layer The weight of the feature, represents the feature dimension of the dam landslide data input to the neural network, The first Features, Represents the neural network The bias of the layer, Represents the Sigmoid activation function; the gradient descent method is used to adaptively update the neural network parameters. The parameter update formula is expressed as follows:

[0041] ,

[0042] ,

[0043] in, represents the learning rate of the neural network, Represents the loss function of the neural network for the The gradient of the layer's weights, Represents the loss function of the neural network for the The gradient of the layer's bias, represents the loss function of the neural network, Indicates parameter update operation;

[0044] S32. During the initialization phase of the adaptive convolution kernel, the convolution layer automatically selects the most suitable convolution kernel size and step size based on the input embankment landslide data. The convolution kernel parameters are optimized using a regularization strategy. The formula is as follows:

[0045] ,

[0046] in, represents the output features of the convolutional layer, represents the adaptive convolution kernel weight, represents the convolution operation, Represents the bias of the convolution layer; the parameters are solved according to the regularization strategy of minimizing the target error, and the adaptive optimization of the initial value of the convolution kernel is completed. The formula is as follows:

[0047] ,

[0048] in, is the parameter update operation, represents the regularization coefficient, represents the Frobenius norm; the gradient descent method is used to update the convolution kernel so that the convolution kernel can adapt to the distribution of embankment landslide data in real time during training. The formula is as follows:

[0049] ,

[0050] in, Represents the loss function of the adaptive convolution kernel weight;

[0051] S33. Using the local attention mechanism, the model focuses on the local areas with key discrimination in the input embankment landslide data. By weighting the features, redundancy and noise interference are suppressed to achieve effective feature extraction. The formula is as follows:

[0052] ,

[0053] in, represents the feature importance score, represents the attention weight, Represents attention bias; based on the Softmax function, the weighting of the local area is realized, and the feature importance score is converted into attention weight. The formula is as follows:

[0054] ,

[0055] in, represents the local attention weight, Indicates the The feature importance score corresponding to each feature, Indicates the total number of samples input to the neural network in the current batch;

[0056] S34. In the multi-scale feature fusion stage, the features extracted by different convolution kernels are weighted and superimposed according to the local attention weights obtained in step S33. The formula is as follows:

[0057] ,

[0058] ,

[0059] ,

[0060] in, represents the fused multi-scale features, Indicates the The weight coefficient of the convolution scale, Indicates the convolution kernel weights, Indicates the number of multi-scales set, Indicates the The importance coefficient of each scale, Indicates the The complexity adjustment coefficient of each scale;

[0061] S35. During the optimization of residual connections, skip connections are built between different layers to alleviate gradient vanishing and enhance the efficiency of deep feature transfer. The convolution kernel weights in the residual connections are updated as follows:

[0062] ,

[0063] in, represents the convolution kernel weight in the residual connection, represents the learning rate of the residual connection, Represents the gradient of the loss function of the neural network with respect to the convolution kernel weights in the residual connection; the input features are integrated into the residual path and nonlinear transformation is implemented, and the residual output is completed according to the jump sum method. The formula is as follows:

[0064] ,

[0065] in, represents the overall output characteristics of the neural network, Represents the bias of the residual connection;

[0066] S36. Repeat iterations S32-S35 until the preset stop iteration condition is met, and a trained neural network based on adaptive convolution kernel optimization is obtained.

[0067] Furthermore, step S4 specifically includes:

[0068] The training process of the support vector machine algorithm based on high-order feature interpolation is as follows:

[0069] S41. Initialize the phase space mapping parameters, define the initial phase space mapping, adapt to different feature extraction characteristics of embankment landslide data through a dynamic adjustment strategy, and define the kernel function of the support vector machine. The formula is as follows:

[0070] ,

[0071] ,

[0072] in, represents the kernel function of the extreme learning machine, Indicates the data points in the input space, Indicates the data points in the input space, represents the width parameter of the kernel function of the extreme learning machine, represents the L2 norm, represents the adaptive adjustment coefficient, is the data sample variance input to the extreme learning machine, representing the global variance; the adaptive adjustment coefficient is automatically adjusted according to the dispersion of the extracted embankment landslide data features, and the formula is as follows:

[0073] ,

[0074] in, represents the adaptive adjustment coefficient, Indicates the number of samples input to the extreme learning machine in the current batch, Indicates the data points in the input space, Represents the mean of all training samples input to the extreme learning machine in the current batch;

[0075] S42. The original feature space is converted into a higher-dimensional feature space by using a high-order feature interpolation method. The interpolation of high-order features is achieved by the following polynomial mapping:

[0076] ,

[0077] in, represents a function that maps the data input to the extreme learning machine into a high-dimensional space, Indicates the data points in the input space; Indicates the highest order of the polynomial;

[0078] S43. Perform dynamic adjustment of the phase space, automatically adjust the structure of the mapping function according to the distribution characteristics of the training data, and then optimize the classification boundary. Dynamically adjust the structure of the mapping function based on the distribution of the training data and use the loss function to optimize the mapping parameters. The formula is as follows:

[0079] ,

[0080] in, represents the loss function of the extreme learning machine, Indicates the The class labels of the data points, represents the kernel function of the extreme learning machine, Represents the width parameter value of the kernel function of the extreme learning machine in the previous iteration, represents the maximum value function;

[0081] S44. After obtaining a suitable high-dimensional feature space, the support vector machine optimization training is performed by adjusting the weight and bias parameters of the support vector machine to minimize the objective function. The formula is as follows:

[0082] ,

[0083] in, is the objective function of the support vector machine, Indicates that the parameters and Implement minimization constraints;

[0084] The constraints of the objective function are: ,in, represents the weight vector of the support vector machine, represents the bias term of the support vector machine; Represents the regularization parameter of the support vector machine, which controls the trade-off between the error term and model complexity; Represents the objective function influencing parameter, which is used to adjust the influence of the kernel function in the optimization process; Indicates the slack variables, Indicates the slack variables;

[0085] S45. The stability of the algorithm is evaluated by cross-validation and stability indicators. The formula is as follows:

[0086] ,

[0087] in, represents the stability index, represents the number of cross validation folds, Represents the cross validation The accuracy of the fold.

[0088] Furthermore, the weight of the combined embedding loss in step S23 is set to 0.3, and the boundary parameter is set to 0.2; the hierarchical reinforcement learning rate parameter in step S24 is set to 0.3; the learning rate of the neural network in step S31 is set to 0.03; the regularization coefficient in step S32 is set to 0.3; the attention bias in step S33 is set to 0.3; the learning rate of the residual connection in step S35 is set to 0.03; the highest order of the polynomial in step S42 is set to 6; the regularization parameter of the support vector machine in step S44 is set to 1; and the number of cross-validation folds in step S45 is set to 10.

[0089] Furthermore, the preset condition for stopping the iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0090] The present invention also provides an artificial intelligence-based dam landslide disease identification device, which executes the artificial intelligence-based dam landslide disease identification method, including:

[0091] Sample collection module: used to collect dam landslide data from actual dam monitoring points and manually annotate them to obtain annotated dam landslide datasets;

[0092] Sample expansion module: This module is used to input the labeled embankment landslide dataset into a trained hierarchical reinforcement-based generative adversarial network for data expansion, thereby obtaining an expanded embankment landslide dataset. The hierarchical reinforcement-based generative adversarial network includes a generator and a discriminator. The generator uses a hierarchical reinforcement learning mechanism to optimize the quality of generated samples through a multi-layer decision-making process.

[0093] Feature extraction module: used to input the expanded dam landslide dataset into a trained neural network optimized based on adaptive convolution kernels for feature extraction, thereby obtaining the extracted dam landslide data features; the neural network optimized based on adaptive convolution kernels includes an input layer, a convolution layer, a pooling layer, and a fully connected layer; the convolution layer is composed of residual blocks, and each residual block is skipped, which is a conventional residual neural network connection method in the field;

[0094] Dam landslide disease identification module: used to input the extracted dam landslide data features into the trained support vector machine algorithm based on high-order feature interpolation to perform dam landslide disease identification and classification, and obtain the dam landslide disease identification results.

[0095] The advantages of the present invention are:

[0096] The present invention adopts a data expansion method based on a hierarchical reinforcement generative adversarial network. The generator uses a reinforcement learning mechanism to optimize the quality of generated samples. The discriminator effectively improves the diversity and authenticity of generated data through adversarial training. The parameters of the generator and discriminator are dynamically optimized, making the generated data more practical, improving the diversity of samples and the generalization ability of the training model. A neural network algorithm based on adaptive convolution kernel optimization is used for feature extraction. By selecting the optimal convolution kernel size and step size, the ability to capture multi-scale features is enhanced. At the same time, a local attention mechanism is used to further focus on the key areas in the embankment landslide data, suppress noise interference, and improve feature extraction. The accuracy and efficiency of the proposed method are improved; a support vector machine algorithm based on high-order feature interpolation is used to map the embankment landslide data to a higher-dimensional feature space, thereby enhancing the model's ability to express nonlinear relationships. At the same time, by dynamically adjusting the phase space mapping structure, the classification boundary is optimized, making the model more accurate in identifying different landslide levels; an enhanced training strategy and model evaluation mechanism are adopted, an adaptive strategy is introduced during the training process, and the feature learning process is optimized by combining embedding loss, edge loss, and residual connection techniques. At the same time, cross-validation and stability indicators are used to evaluate the performance of the model on different data subsets to ensure the effectiveness and stability of the algorithm in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0098] Figure 1 is a flow chart of the steps of the present invention;

[0099] Figure 2 This is a comparison chart of the recognition accuracy of the discriminator of the method of the present invention and the conventional method;

[0100] Figure 3 Comparison of computational complexity between using local attention mechanism and without local attention;

[0101] Figure 4 Comparison of the classification accuracy between the method of the present invention and other methods. DETAILED DESCRIPTION

[0102] The following will clearly and completely describe the technical solutions in the embodiments of the present invention 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. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0103] Example 1

[0104] In this embodiment, Figure 1 As shown, the present invention provides a method for identifying dam landslide hazards based on artificial intelligence, the specific steps of which include:

[0105] S1. Collect and manually label dam landslide data from actual dam monitoring points to obtain a labeled dam landslide dataset;

[0106] Specifically, the dam landslide data includes geographic location Ra, current temperature Da, humidity value Ia, precipitation Pa around the dam, water level Ha, displacement Va, seepage pressure Sa, size of cracks on the dam surface Fa, accumulated rainfall La, and historical disease records Ca; data collection methods mainly include automatic sensor collection and manual calibration. The data of each monitoring point are regularly collected through sensors or remote sensing equipment and uploaded to the cloud storage platform at a high frequency to ensure the integrity and real-time nature of the data; the data storage format uniformly adopts a standardized data table format, including CSV or JSON format.

[0107] It should be noted that this embodiment is only intended to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0108] The labeling categories of embankment landslide data are “normal”, “minor landslide”, “moderate landslide” and “severe landslide”.

[0109] S2. The labeled dam landslide dataset is input into a trained hierarchical reinforcement-based generative adversarial network for data augmentation, thereby obtaining an augmented dam landslide dataset. The hierarchical reinforcement-based generative adversarial network comprises a generator and a discriminator. The generator employs a hierarchical reinforcement learning mechanism to optimize the quality of generated samples through a multi-layer decision-making process.

[0110] Specifically, the training process of the generative adversarial network based on hierarchical reinforcement is as follows:

[0111] S21. Initialize the network parameters of the generator and discriminator. The generated embankment landslide data must have a certain degree of diversity and realism. The initialization strategy is to use a normal distribution with a mean of 0 and a variance of the unit matrix. The generator can generate a variety of fake data at the beginning and provide it to the discriminator for discrimination, thereby gradually improving the quality of the generated data through adversarial training. For the generator and discriminator, the parameter initialization formula is expressed as follows:

[0112] ,

[0113] ,

[0114] in, represents the parameters of the generator, represents the parameters of the discriminator, Indicates that it obeys a specific distribution. represents a normal distribution with mean 0 and variance as the identity matrix, represents the identity matrix;

[0115] S22. To alleviate the imbalance between the generator and the discriminator in the early stages of training, the discriminator is pre-trained so that it can distinguish between real dam landslide data and generated dam landslide data to a certain extent. In the dam landslide data expansion task, the discriminator needs to be sensitive to different attributes (such as water level, displacement, crack size, etc.). Especially when a dam landslide or disease occurs, the changes in these data attributes before and after the landslide are usually significant. For example, before a landslide occurs, changes in water level, increased precipitation, and crack expansion may be precursors. Pre-training can help the discriminator better identify these precursor features. During the pre-training stage, the discriminator not only learns how to distinguish between real dam landslide data and generated data, but also needs to be able to identify the temporal evolution of attributes such as water level, displacement, and precipitation. Through pre-training, the discriminator can make accurate judgments on key features in the dam landslide data (such as the correlation between accumulated rainfall and disease records) in the early stages, thereby providing more effective feedback to the generator and ensuring that the generated data is more realistic. The loss function formula of the discriminator in the pre-training stage is expressed as follows:

[0116] ,

[0117] in, represents the real dam landslide data, is the loss function of the discriminator, represents the input noise sampled from the noise distribution, represents the noise distribution, Express expectations, represents the discriminator function, represents a generator function, represents the distribution of real dam landslide data, represents the natural logarithm;

[0118] S23. During the adversarial training phase, the generator parameters are first fixed. The real dam landslide data and the generated dam landslide data output by the generator are input into the discriminator for processing, and the discriminator parameters are updated. The generated dam landslide data output by the generator is the data obtained by inputting noise sampled from the noise distribution into the generator. The optimization goal of this process is to maximize the probability that the real dam landslide data is correctly classified while minimizing the probability that the generated dam landslide data is misclassified. The generator must not only learn how to generate basic dam state data (such as temperature, humidity, water level, displacement, etc.) but also understand the inherent correlation between various data attributes. For example, water level and displacement may affect each other, while an increase in precipitation will directly affect the crack size and seepage pressure on the dam surface. In this case, the generator needs to be able to generate dam data with reasonable correlation so that the discriminator cannot distinguish between true and false data.

[0119] Then, the discriminator parameters are fixed and the generator parameters are updated. The generator adjusts its generation strategy through a hierarchical reinforcement learning strategy, so that the discriminator mistakenly identifies the generated embankment landslide data as real embankment landslide data, while minimizing the combined embedding loss to improve the quality and diversity of the generated embankment landslide data. In this process of the embankment landslide data expansion task, the discriminator's goal is to maximize the recognition accuracy of real data, especially in the relationship between the historical damage records of the embankment (such as whether there is a disease) and important features such as displacement and crack size. The discriminator will learn how to accurately identify the complex dependencies between these data through training, thereby helping the generator adjust its generation strategy to make the generated data more realistic. The generator's loss function formula is expressed as follows:

[0120] ,

[0121] in, represents the loss function of the generator, represents the weight of the combined embedding loss, Set to 0.3, Represents the combined embedding loss; in the dam landslide data expansion, the introduction of the combined embedding loss can help the generator maintain the diversity of the generated data, while avoiding excessive concentration or overfitting of the generated dam data. For dam landslide data, features such as water level (Ha), displacement (Va) and crack size (Fa) all have temporal and spatial distribution characteristics. The combined embedding loss can help the generator maintain the diversity and rationality of these data attributes when generating data. Through the combined embedding loss, the generator can learn how to balance the relationship between these attributes, so that the generated data is not only consistent in the feature space, but also adaptable to different actual situations. The formula of the combined embedding loss is expressed as follows:

[0122] ,

[0123] in, Indicates the number of samples generated in the current batch, represents the feature extraction function of the neural network layer of the generator, represents the first weight parameter, represents the second weight parameter, represents the boundary parameter, which is used to control the balance of losses; represents the L2 norm, Indicates taking the maximum value, Indicates the Real dam landslide data, Indicates the Input noise;

[0124] Specifically, the calculation method of the square loss term in the combined embedding loss can be expressed as:

[0125] ,

[0126] in, Represents the feature extraction function for the The nonlinear transformation function of feature dimensions, Indicates the characteristic dimension of the generated embankment landslide data;

[0127] Specifically, the margin loss encourages the distance between the generated features and the real features to be at least , thereby avoiding the model from overfitting to a small batch of dam landslide training data. For dam landslide data, especially the historical damage records of the dam and some extreme values ​​(such as abnormally large displacement or crack size), it is of special importance. The edge loss can ensure that the data generated by the generator avoids extreme values ​​that deviate too much from the true distribution. For example, the historical damage records (Ca) of the dam may have a certain boundary relationship with the precipitation (Pa) or the accumulated rainfall (La). The edge loss can ensure that the generator does not generate too much or too little damage record data, avoid generating unreasonable disease labels, and thus improve the authenticity of the data set. The calculation method of the edge loss term is expressed as:

[0128] ,

[0129] in, is a boundary parameter. Preferably, Set to 0.2;

[0130] S24. Each layer of the generator is independently adjusted through reinforcement learning during training. During the training process of the generator, the reward for each adjustment is calculated based on the feedback from the discriminator and the embedding loss, and the formula is as follows:

[0131] ,

[0132] ,

[0133] in, represents the Sigmoid function, represents the hierarchical reinforcement learning rate parameter, Set to 0.3, Rewards representing adjustments to the generator's neural network layers;

[0134] S25. Repeat steps S22-S24 until a preset stopping condition is met, completing the training and obtaining a trained hierarchical reinforcement-based generative adversarial network. The preset stopping condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000.

[0135] S3. Input the expanded dam landslide dataset into a trained neural network optimized with adaptive convolution kernels for feature extraction, thereby obtaining extracted dam landslide data features. The neural network optimized with adaptive convolution kernels includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. The convolution layers are residually connected to form residual blocks, and each residual block is skip-connected, which is a conventional residual neural network connection method in the field. The residual block is constructed by three convolution layers, each followed by an activation function (e.g., Reinforced Luminance Unit) to extract features and introduce nonlinearity. Each convolution layer processes the input data to extract specific feature information, and the activation function enhances the model's expressiveness. In each residual block, the input data is directly passed to the output of the residual block, rather than being derived solely from the results of the convolution layer. Specifically, after being processed by multiple convolution layers, the input data is added to the original input data to form a "residual" output. In other words, the network not only learns the mapping from input to output, but also learns the difference (residual) between the input and output. The skip connections between multiple residual blocks are implemented so that if there are multiple residual blocks in the network, they are connected via skip connections. For example, the output of the first residual block can be directly connected to the input of the second residual block, rather than simply passing the output of the convolution calculation. The skip connections between each residual block allow information to flow more smoothly from the bottom to the top of the network, thus avoiding the performance degradation that occurs when training deep networks. To address the imbalance caused by different types of features (such as high dimensionality, nonlinearity, and noise) in complex embankment landslide data, a neural network algorithm based on adaptive convolution kernel optimization is used for feature extraction. Regularization strategies are used to select the optimal convolution kernel size and step size to enhance the ability to capture multi-scale features. Furthermore, to address the problem of redundant information interference, a local attention mechanism is used to focus on key areas through feature weighting, suppress noise interference, and improve feature extraction efficiency.

[0136] Specifically, the training process of the neural network based on adaptive convolution kernel optimization is as follows:

[0137] S31. In the initial training phase, in the task of extracting features from dam landslide data, dam landslide data involve multiple different types of features. Hierarchical processing is performed on the data in the expanded dam landslide dataset to address the imbalance caused by different types of features (including high-dimensional, nonlinear, and noisy) in the dam landslide data. By grouping the dam landslide data into different categories and assigning them to different subnetworks, the ability to focus on the features of each category of dam landslide data is enhanced, reducing redundant interference caused by feature mixing. The formula is as follows:

[0138] ,

[0139] in, Represents the neural network The dam landslide data output after layer processing, Represents the neural network The weight parameters of the layer, represents the input dam landslide data feature matrix of the neural network, Represents the activation function of the neural network.

[0140] Specifically, in order to enable the processing of each layer of the neural network to adaptively optimize different dam landslide data features, some features change dramatically in a specific period or area, while other features change less. In order to deal with this imbalance, a hierarchical processing method is adopted to group the data according to the different categories of its features and assign them to different sub-networks for processing. For example, for the strong correlation between water level (Ha) and displacement (Va), a sub-network is assigned to the network to specifically process the relationship between these two features, while another sub-network can process the relationship between humidity (Ia) and temperature (Da). This can reduce feature mixing and redundant interference, improve the network's ability to focus on each category of features, and thus improve efficiency and accuracy in feature extraction. In this process, the calculation method of the neural network's activation function is expressed as

[0141] ,

[0142] in, Represents the neural network Layer The weight of the feature, represents the feature dimension of the dam landslide data input to the neural network, The first Features, Represents the neural network The bias of the layer, Represents the Sigmoid activation function. To achieve dynamic optimization of the parameters between the layers of the neural network, the gradient descent method is used to adaptively update the neural network parameters. For dam landslide data, the nonlinear relationship between features is particularly significant. For example, the historical damage record of the dam (Ca) may have a complex nonlinear relationship with the surrounding precipitation (Pa) and cumulative rainfall (La). Through continuous iteration and optimization of the adaptive activation function, the neural network can better capture these nonlinear features, improve the model's expressive power, enable the model to more accurately learn the deep relationship between different features, and improve the effect of feature extraction. In this process, the formula for updating the neural network parameters is expressed as follows:

[0143] ,

[0144] ,

[0145] in, represents the learning rate of the neural network, Represents the loss function of the neural network for the The gradient of the layer's weights, Represents the loss function of the neural network for the The gradient of the layer's bias, represents the loss function of the neural network, represents a parameter update operation; preferably, Set to 0.03.

[0146] S32. During the initialization phase of the adaptive convolution kernel, the convolution layer automatically selects the most appropriate convolution kernel size and step size based on the input dam landslide data to enhance the ability to capture multi-scale information. By adaptively selecting the convolution kernel size and step size, the ability to capture multi-scale information in dam landslide data can be enhanced. Dam landslide data often has multi-scale characteristics. For example, changes in water level (Ha) may be small fluctuations in the short term or gradual changes in the long term, and displacement (Va) may also present different scale patterns from local to global. By adaptively optimizing the convolution kernel, the most appropriate data scale can be automatically selected, avoiding the problem that a fixed convolution kernel cannot adapt to the different scale characteristics of dam data. In addition, by optimizing the convolution kernel parameters through a regularization strategy, it can be ensured that the convolution operation adaptively adapts to the distribution of dam landslide data during the learning process, improving the quality of feature extraction. The formula is as follows:

[0147] ,

[0148] in, represents the output features of the convolutional layer, represents the adaptive convolution kernel weight, represents the convolution operation, Represents the bias of the convolution layer; the parameters are solved according to the regularization strategy of minimizing the target error, and the adaptive optimization of the initial value of the convolution kernel is completed. The formula is as follows:

[0149] ,

[0150] in, is the parameter update operation, represents the regularization coefficient, represents the Frobenius norm; preferably, The convolution kernel is updated by gradient descent method so that it can adapt to the distribution of embankment landslide data in real time during training. The formula is as follows:

[0151] ,

[0152] in, Represents the loss function of the adaptive convolution kernel weight;

[0153] S33. Using a local attention mechanism, the model focuses on key discriminative local areas in the input dam landslide data. By weighting the features, redundancy and noise interference are suppressed to achieve effective feature extraction. In the dam landslide data processing task, features such as water level, temperature, and humidity are often interfered with by external factors, such as sudden weather changes or equipment errors. These redundant or noisy data will affect the model's learning process. Through the local attention mechanism, the neural network can focus on those feature areas that play a key role in the occurrence of dam landslides. For example, before a landslide occurs, certain features such as a sharp increase in precipitation or a rapid rise in water level may be key early warning signals. By weighting these areas, the network can suppress noise in irrelevant areas, thereby improving the accuracy of feature extraction. The formula is as follows:

[0154] ,

[0155] in, represents the feature importance score, represents the attention weight, Indicates attention bias, Set to 0.3; implement weighting of local areas based on the Softmax function, and convert feature importance scores into attention weights. The formula is as follows:

[0156] ,

[0157] in, represents the local attention weight, Indicates the The feature importance score corresponding to each feature, Indicates the total number of samples input to the neural network in the current batch;

[0158] S34. In the multi-scale feature fusion stage, based on the local attention weights obtained in step S33, the features extracted by different convolution kernels are weighted and superimposed to simultaneously obtain local details and global structural information, thereby addressing the problem of uneven distribution of dam landslide data. The occurrence of dam landslides does not depend solely on the change of a single feature, but rather on the combined effect of multiple factors. For example, a sharp change in water level may be accompanied by significant crack expansion and an increase in seepage pressure. Through multi-scale feature fusion, the model can comprehensively consider data changes at different scales, helping the model to understand the occurrence mechanism of dam landslides from both global and local levels, thereby improving the integrity and accuracy of feature extraction. The formula is as follows:

[0159] ,

[0160] in, represents the fused multi-scale features, Indicates the The weight coefficient of the convolution scale, Indicates the convolution kernel weights, Indicates the number of multi-scales set;

[0161] Specifically, in order to make the multi-scale fusion process more adaptive, the importance of each scale is dynamically allocated according to the convolution kernel training situation. The formula is as follows:

[0162] ,

[0163] in, Indicates the The importance coefficient of each scale;

[0164] Specifically, the complexity of the convolution kernel is characterized by the Frobenius norm, and then the importance coefficient of each scale is calculated. The formula is as follows:

[0165] ,

[0166] in, Indicates the The complexity adjustment coefficient of each scale;

[0167] S35. During the optimization of residual connections, skip connections are established between different layers to mitigate gradient vanishing and enhance the efficiency of deep feature transfer. During feature extraction of embankment landslide data, the vanishing gradient problem in deep networks may be encountered, especially when dealing with high-dimensional features with complex dependencies. Residual connections enable neural networks to directly transfer information between different layers, ensuring that key information is not lost during transfer. For example, the long-distance dependency between the embankment's historical damage records (Ca) and its geographic location (Ra) can be effectively handled through deep residual connections. This ensures that information is fully transferred and integrated within the multi-layer network structure, thereby improving the model's depth and breadth when extracting embankment landslide data features. The convolution kernel weights in the residual connections are updated as follows:

[0168] ,

[0169] in, represents the convolution kernel weight in the residual connection, represents the learning rate of the residual connection, Set to 0.03, Represents the gradient of the loss function of the neural network with respect to the convolution kernel weights in the residual connection; the input features are integrated into the residual path and nonlinear transformation is implemented, and the residual output is completed according to the jump sum method. The formula is as follows:

[0170] ,

[0171] in, represents the overall output characteristics of the neural network, Represents the bias of the residual connection;

[0172] S36. Repeat S32-S35 until a preset stopping condition is met, thereby obtaining a trained neural network based on adaptive convolution kernel optimization. The preset stopping condition is when a preset maximum number of iterations is reached. Preferably, the preset maximum number of iterations is set to 1000.

[0173] S4. Input the extracted embankment landslide data features into the trained support vector machine algorithm based on high-order feature interpolation to perform embankment landslide disease identification and classification, and obtain the embankment landslide disease identification results.

[0174] Specifically, the training process of the support vector machine algorithm based on high-order feature interpolation is as follows:

[0175] S41. Initialize the phase space mapping parameters, define the initial phase space mapping, adapt to different feature extraction characteristics of embankment landslide data through a dynamic adjustment strategy, and define the kernel function of the support vector machine. The formula is as follows:

[0176] ,

[0177] ,

[0178] in, represents the kernel function of the extreme learning machine, Indicates the data points in the input space, Indicates the data points in the input space, represents the width parameter of the kernel function of the extreme learning machine, represents the L2 norm, represents the adaptive adjustment coefficient, is the data sample variance input to the extreme learning machine, representing the global variance; the adaptive adjustment coefficient is automatically adjusted according to the dispersion of the extracted embankment landslide data features, and the formula is as follows:

[0179] ,

[0180] in, represents the adaptive adjustment coefficient, Indicates the number of samples input to the extreme learning machine in the current batch, Indicates the data points in the input space, Represents the mean of all training samples input to the extreme learning machine in the current batch;

[0181] S42. The original feature space is converted into a higher-dimensional feature space through high-order feature interpolation. This not only enhances the model's ability to capture nonlinear relationships but also preserves important information from the original feature-extracted embankment landslide data. High-order feature interpolation is achieved through the following polynomial mapping:

[0182] ,

[0183] in, represents a function that maps the data input to the extreme learning machine into a high-dimensional space, Indicates the data points in the input space; represents the highest order of the polynomial, Set to 6; the specific implementation of the function that maps data to high-dimensional space includes a combination of cross terms and high-order terms, and the calculation method is expressed as:

[0184] ,

[0185] Where, For the The first data points, For the The first data points.

[0186] S43. Perform dynamic adjustment of the phase space, automatically adjust the structure of the mapping function according to the distribution characteristics of the training data, and then optimize the classification boundary. Dynamically adjust the structure of the mapping function based on the distribution of the training data and use the loss function to optimize the mapping parameters. The formula is as follows:

[0187] ,

[0188] in, represents the loss function of the extreme learning machine, Indicates the The category labels of each data point are, for example, 0, 1, 2, and 3 correspond to the four categories of "normal", "minor landslide", "moderate landslide", and "severe landslide" respectively; represents the kernel function of the extreme learning machine, Represents the width parameter value of the kernel function of the extreme learning machine in the previous iteration, represents the maximum value function;

[0189] S44. After obtaining a suitable high-dimensional feature space, the support vector machine optimization training is performed by adjusting the weight and bias parameters of the support vector machine to minimize the objective function. The formula is as follows:

[0190] ,

[0191] in, is the objective function of the support vector machine, Indicates that the parameters and Implement minimization constraints;

[0192] The constraints of the objective function are: ,in, represents the weight vector of the support vector machine, represents the bias term of the support vector machine; Represents the regularization parameter of the support vector machine, which controls the trade-off between the error term and model complexity; Represents the objective function influencing parameter, which is used to adjust the influence of the kernel function in the optimization process; Indicates the slack variables, Indicates the slack variables; preferably, Set to 1.

[0193] S45. The stability of the algorithm is evaluated by cross-validation and stability indicators. The formula is as follows:

[0194] ,

[0195] in, represents the stability index, represents the number of cross validation folds, Represents the cross validation The accuracy of the fold is used to evaluate the performance of the model on different subsets of embankment landslide data. Set to 10.

[0196] Example 2

[0197] In this embodiment, Figure 2As shown in the figure, we experimentally compare the generative adversarial network based on hierarchical reinforcement in the method of the present invention with the generative adversarial network based on the existing method. By analyzing the recognition accuracy of the discriminator, we show the improvement trend of the discriminator in recognizing the accuracy of real and generated data. It can be seen from the figure that the present invention has an accelerated improvement effect on the accuracy during the training process.

[0198] Example 3

[0199] In this embodiment, we divide the constructed dataset into four subsets: dataset A, dataset B, dataset C, and dataset D. Figure 3 As shown in the figure, an experimental comparison is carried out before and after the method of the present invention adopts the local attention mechanism. It can be seen from the figure that the computational complexity of the method of the present invention is significantly lower than that of the model without the attention mechanism. The method of the present invention can focus on the features of the key areas and reduce redundant calculations.

[0200] Example 4

[0201] In this embodiment, as shown in Figure 4, we conducted an experimental comparison between the method of the present invention and the support vector machine, random forest and traditional CNN network. It can be seen from the figure that in terms of the accuracy comparison of embankment landslide hazard classification, the performance of the present invention is significantly better than other methods, reflecting a better ability to process complex data.

[0202] Example 5

[0203] This embodiment provides an artificial intelligence-based dam landslide hazard identification device, which executes the artificial intelligence-based dam landslide hazard identification method described in Example 1, including:

[0204] Sample collection module: used to collect dam landslide data from actual dam monitoring points and manually annotate them to obtain annotated dam landslide datasets;

[0205] Sample expansion module: This module is used to input the labeled embankment landslide dataset into a trained hierarchical reinforcement-based generative adversarial network for data expansion, thereby obtaining an expanded embankment landslide dataset. The hierarchical reinforcement-based generative adversarial network includes a generator and a discriminator. The generator uses a hierarchical reinforcement learning mechanism to optimize the quality of generated samples through a multi-layer decision-making process.

[0206] Feature extraction module: used to input the expanded dam landslide dataset into a trained neural network optimized based on adaptive convolution kernels for feature extraction, thereby obtaining the extracted dam landslide data features; the neural network optimized based on adaptive convolution kernels includes an input layer, a convolution layer, a pooling layer, and a fully connected layer; the convolution layer is composed of residual blocks, and each residual block is skipped, which is a conventional residual neural network connection method in the field;

[0207] Dam landslide disease identification module: used to input the extracted dam landslide data features into the trained support vector machine algorithm based on high-order feature interpolation to perform dam landslide disease identification and classification, and obtain the dam landslide disease identification results.

[0208] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for identifying dam landslide hazards based on artificial intelligence, characterized in that: The following steps are involved: S1. Collect and manually label dam landslide data from actual dam monitoring points to obtain a labeled dam landslide dataset; S2. The labeled dam landslide dataset is input into a trained hierarchical reinforcement-based generative adversarial network for data augmentation, thereby obtaining an augmented dam landslide dataset. The hierarchical reinforcement-based generative adversarial network comprises a generator and a discriminator. The generator employs a hierarchical reinforcement learning mechanism to optimize the quality of generated samples through a multi-layer decision-making process. The training process of the hierarchical reinforcement-based generative adversarial network is as follows: S21. Initialize the network parameters of the generator and discriminator using a normal distribution with a mean of 0 and a variance of the identity matrix. S22. Pre-train the discriminator; S23. During the adversarial training phase, first, the generator parameters are fixed. The real dam landslide data and the generated dam landslide data output by the generator are input into the discriminator for processing, and the discriminator parameters are updated. The generated dam landslide data output by the generator is obtained by inputting noise sampled from a noise distribution into the generator. Then, the discriminator parameters are fixed, the generator parameters are updated, and the generator adjusts its generation strategy using a hierarchical reinforcement learning strategy. The generator's loss function is expressed as follows: in, represents the loss function of the generator, represents the weight of the combined embedding loss, Denotes the combined embedding loss; the formula of the combined embedding loss is as follows: ， ， ， in, Indicates the number of samples generated in the current batch, represents the feature extraction function of the neural network layer of the generator, Represents the feature extraction function for the The nonlinear transformation function of feature dimensions, Indicates the characteristic dimension of the generated embankment landslide data; represents the first weight parameter, represents the second weight parameter, represents the boundary parameter, which is used to control the balance of losses; represents the L2 norm, Indicates taking the maximum value, Indicates the Real dam landslide data, Indicates the Input noise; S24. Each layer of the generator is independently adjusted through reinforcement learning during training. During the training process of the generator, the reward for each adjustment is calculated based on the feedback from the discriminator and the embedding loss, and the formula is as follows: ， ， in, represents the Sigmoid function, represents the hierarchical reinforcement learning rate parameter, Rewards representing adjustments to the generator's neural network layers; S25. Repeat steps S22-S24 until the preset stop iteration condition is met, the training is completed, and a trained generative adversarial network based on hierarchical reinforcement is obtained; S3. Inputting the expanded dam landslide dataset into a trained neural network optimized based on an adaptive convolution kernel for feature extraction to obtain extracted dam landslide data features; the neural network optimized based on the adaptive convolution kernel includes an input layer, a convolution layer, a pooling layer, and a fully connected layer; the convolution layer is residually connected to form residual blocks, and each residual block is skipped, where the skip connection is a conventional residual neural network connection method in the art; S4. Input the extracted embankment landslide data features into the trained support vector machine algorithm based on high-order feature interpolation to perform embankment landslide disease identification and classification, and obtain the embankment landslide disease identification results.

2. The artificial intelligence-based dam landslide hazard identification method according to claim 1 is characterized in that: Step S1 specifically includes: The dam landslide data include geographical location Ra, current temperature Da, humidity value Ia, precipitation around the dam Pa, water level Ha, displacement Va, seepage pressure Sa, crack size Fa on the dam surface, accumulated rainfall La, and historical disease records Ca; The labeling categories of embankment landslide data are "normal", "minor landslide", "moderate landslide" and "severe landslide".

3. The artificial intelligence-based dam landslide hazard identification method according to claim 2, characterized in that: The parameter initialization formula in step S21 is expressed as follows: ， ， in, represents the parameters of the generator, represents the parameters of the discriminator, Indicates that it obeys a specific distribution. represents a normal distribution with mean 0 and variance as the identity matrix, represents the identity matrix; The loss function formula of the discriminator in the pre-training stage in step S22 is expressed as follows: ， in, represents the real dam landslide data, is the loss function of the discriminator, represents the input noise sampled from the noise distribution, represents the noise distribution, Express expectations, represents the discriminator function, represents a generator function, represents the distribution of real dam landslide data, Represents the natural logarithm.

4. The artificial intelligence-based dam landslide hazard identification method according to claim 3 is characterized in that: Step S3 specifically includes: The training process of the neural network based on adaptive convolution kernel optimization is as follows: S31. In the initial stage of training, the data in the expanded embankment landslide dataset is hierarchically processed by grouping the embankment landslide data into different categories and assigning them to different sub-networks. The formula is as follows: ， ， in, Represents the neural network The dam landslide data output after layer processing, Represents the neural network The weight parameters of the layer, represents the input dam landslide data feature matrix of the neural network, represents the activation function of the neural network, Represents the neural network Layer The weight of the feature, represents the feature dimension of the dam landslide data input to the neural network, The first Features, Represents the neural network The bias of the layer, Represents the Sigmoid activation function; the gradient descent method is used to adaptively update the neural network parameters. The parameter update formula is expressed as follows: ， ， in, represents the learning rate of the neural network, Represents the loss function of the neural network for the The gradient of the layer's weights, Represents the loss function of the neural network for the The gradient of the layer's bias, represents the loss function of the neural network, Indicates parameter update operation; S32. During the initialization phase of the adaptive convolution kernel, the convolution layer automatically selects the most suitable convolution kernel size and step size based on the input embankment landslide data. The convolution kernel parameters are optimized using a regularization strategy. The formula is as follows: ， in, represents the output features of the convolutional layer, represents the adaptive convolution kernel weight, represents the convolution operation, Represents the bias of the convolution layer; the parameters are solved according to the regularization strategy of minimizing the target error, and the adaptive optimization of the initial value of the convolution kernel is completed. The formula is as follows: ， in, is the parameter update operation, represents the regularization coefficient, represents the Frobenius norm; the gradient descent method is used to update the convolution kernel so that the convolution kernel can adapt to the distribution of embankment landslide data in real time during training. The formula is as follows: ， in, Represents the loss function of the adaptive convolution kernel weight; S33. Using the local attention mechanism, the model focuses on the local areas with key discrimination in the input embankment landslide data. By weighting the features, redundancy and noise interference are suppressed to achieve effective feature extraction. The formula is as follows: ， in, represents the feature importance score, represents the attention weight, Represents attention bias; based on the Softmax function, the weighting of the local area is realized, and the feature importance score is converted into attention weight. The formula is as follows: ， in, represents the local attention weight, Indicates the The feature importance score corresponding to each feature, Indicates the total number of samples input to the neural network in the current batch; S34. In the multi-scale feature fusion stage, the features extracted by different convolution kernels are weighted and superimposed according to the local attention weights obtained in step S33. The formula is as follows: ， ， ， in, represents the fused multi-scale features, Indicates the The weight coefficient of the convolution scale, Indicates the convolution kernel weights, Indicates the number of multi-scales set, Indicates the The importance coefficient of each scale, Indicates the The complexity adjustment coefficient of each scale; S35. During the optimization of residual connections, skip connections are built between different layers to alleviate gradient vanishing and enhance the efficiency of deep feature transfer. The convolution kernel weights in the residual connections are updated as follows: ， in, represents the convolution kernel weight in the residual connection, represents the learning rate of the residual connection, Represents the gradient of the loss function of the neural network with respect to the convolution kernel weights in the residual connection; the input features are integrated into the residual path and nonlinear transformation is implemented, and the residual output is completed according to the jump sum method. The formula is as follows: ， in, represents the overall output characteristics of the neural network, Represents the bias of the residual connection; S36. Repeat iterative steps S32-S35 until the preset stop iteration condition is met to obtain a trained neural network based on adaptive convolution kernel optimization.

5. The artificial intelligence-based dam landslide hazard identification method according to claim 4 is characterized in that: Step S4 specifically includes: The training process of the support vector machine algorithm based on high-order feature interpolation is as follows: S41. Initialize the phase space mapping parameters, define the initial phase space mapping, adapt to different feature extraction characteristics of embankment landslide data through a dynamic adjustment strategy, and define the kernel function of the support vector machine. The formula is as follows: ， ， in, represents the kernel function of the extreme learning machine, Indicates the data points in the input space, Indicates the data points in the input space, represents the width parameter of the kernel function of the extreme learning machine, represents the L2 norm, represents the adaptive adjustment coefficient, is the data sample variance input to the extreme learning machine, representing the global variance; the adaptive adjustment coefficient is automatically adjusted according to the dispersion of the extracted embankment landslide data features, and the formula is as follows: ， in, represents the adaptive adjustment coefficient, Indicates the number of samples input to the extreme learning machine in the current batch, Indicates the data points in the input space, Represents the mean of all training samples input to the extreme learning machine in the current batch; S42. The original feature space is converted into a higher-dimensional feature space by using a high-order feature interpolation method. The interpolation of high-order features is achieved by the following polynomial mapping: ， in, represents a function that maps the data input to the extreme learning machine into a high-dimensional space, Indicates the data points in the input space; Indicates the highest order of the polynomial; S43. Perform dynamic adjustment of the phase space, automatically adjust the structure of the mapping function according to the distribution characteristics of the training data, and then optimize the classification boundary. Dynamically adjust the structure of the mapping function based on the distribution of the training data and use the loss function to optimize the mapping parameters. The formula is as follows: ， in, represents the loss function of the extreme learning machine, Indicates the The class labels of the data points, represents the kernel function of the extreme learning machine, Represents the width parameter value of the kernel function of the extreme learning machine in the previous iteration, represents the maximum value function; S44. After obtaining a suitable high-dimensional feature space, the support vector machine optimization training is performed by adjusting the weight and bias parameters of the support vector machine to minimize the objective function. The formula is as follows: ， in, is the objective function of the support vector machine, Indicates that the parameters and Implement minimization constraints; The constraints of the objective function are: ,in, represents the weight vector of the support vector machine, represents the bias term of the support vector machine; Represents the regularization parameter of the support vector machine, which controls the trade-off between the error term and model complexity; Represents the objective function influencing parameter, which is used to adjust the influence of the kernel function in the optimization process; Indicates the slack variables, Indicates the slack variables; S45. The stability of the algorithm is evaluated by cross-validation and stability indicators. The formula is as follows: ， in, represents the stability index, represents the number of cross validation folds, Represents the cross validation The accuracy of the fold.

6. The artificial intelligence-based dam landslide hazard identification method according to claim 5, characterized in that: The weight of the combined embedding loss in step S23 is set to 0.3, and the boundary parameter is set to 0.2; the hierarchical reinforcement learning rate parameter in step S24 is set to 0.3; the learning rate of the neural network in step S31 is set to 0.03; the regularization coefficient in step S32 is set to 0.3; the attention bias in step S33 is set to 0.3; the learning rate of the residual connection in step S35 is set to 0.03; the highest order of the polynomial in step S42 is set to 6; the regularization parameter of the support vector machine in step S44 is set to 1; and the number of cross-validation folds in step S45 is set to 10.

7. The artificial intelligence-based dam landslide hazard identification method according to claim 6, characterized in that: The preset condition for stopping iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

8. An artificial intelligence-based dam landslide hazard identification device, which executes the artificial intelligence-based dam landslide hazard identification method according to claim 1, characterized in that: include: Sample collection module: used to collect dam landslide data from actual dam monitoring points and manually annotate them to obtain annotated dam landslide datasets; Sample expansion module: used to input the labeled dam landslide dataset into the trained hierarchical reinforcement-based generative adversarial network for data expansion, thereby obtaining the expanded dam landslide dataset; the hierarchical reinforcement-based generative adversarial network includes a generator and a discriminator; The generator uses a hierarchical reinforcement learning mechanism to optimize the quality of generated samples through a multi-layer decision-making process; Feature extraction module: used to input the expanded dam landslide data set into the trained neural network based on adaptive convolution kernel optimization for feature extraction, and obtain the extracted dam landslide data features; The neural network based on adaptive convolution kernel optimization includes an input layer, a convolution layer, a pooling layer, and a fully connected layer; the convolution layer is composed of residual blocks, and each residual block is skipped, which is a conventional residual neural network connection method in the field; Dam landslide disease identification module: used to input the extracted dam landslide data features into the trained support vector machine algorithm based on high-order feature interpolation to perform dam landslide disease identification and classification, and obtain the dam landslide disease identification results.

Images