Collapsibility grade automatic discrimination method based on deep learning

Abstract

The invention discloses a deep learning-based collapsibility grade automatic discrimination method. The method comprises the following steps of S1, generating a preprocessed loess geological engineering index data set; s2, obtaining a loess sampling unit graph structure data set with a weighted adjacent matrix; s3, forming a loess sampling unit weighted graph structure data set with node attributes; s4, constructing a multi-target comprehensive fitness function; s5, randomly generating differential flower pollination algorithm population individuals, and performing global search and local search on the differential flower pollination algorithm initial population to obtain an optimal graph attention network hyper-parameter vector and an optimal node feature selection mask vector; and S6, obtaining an optimized graph attention network model, deducing the loess sampling unit weighted graph structure data set with node attributes, and outputting a collapsibility grade prediction result of each sampling unit. According to the method, the influence of misjudgment of different collapsibility levels on engineering safety is dynamically evaluated, so that the algorithm is guided to preferentially improve the recognition accuracy of high-risk levels.

Description

Technical Field

[0001] The present invention relates to the technical field of loess collapsibility, and in particular to a method for automatically distinguishing collapsibility levels based on deep learning. Background Art

[0002] The collapsibility of loess is a geological issue of widespread concern in engineering construction. Its determination is of great significance for foundation treatment, structural design, and construction safety. Accurate assessment of collapsibility is crucial for ensuring infrastructure safety, particularly in Northwest and North China, where loess is widely distributed. However, traditional methods for determining collapsibility rely primarily on manual drilling, laboratory testing, and interpretation based on engineering experience. These processes are cumbersome, time-consuming, and highly subjective, often susceptible to limitations in data sample size, the technical staff's experience level, and test errors, resulting in unstable and inapplicable results.

[0003] At present, some studies have attempted to introduce machine learning models to assist in distinguishing the level of loess collapsibility, such as using decision trees, support vector machines, and random forest methods to model. However, there is a general problem of insufficient consideration of spatial structural information. The formation of loess collapsibility is significantly correlated with regional hydrogeological conditions, spatial location distribution, and stratigraphic genesis. Traditional machine learning methods only treat various geological indicators as independent input variables, ignoring the potential coupling relationship between sampling units in terms of spatial distance, hydrological indicators, and stratigraphic stratification, resulting in poor generalization ability of the model when promoted across regions.

[0004] In addition, existing research mostly relies on empirical settings or exhaustive searches in feature selection and model parameter adjustment, lacks dynamic feedback mechanisms and adaptive tuning strategies, and is prone to falling into local optimality, resulting in insufficient adaptability of the model to complex engineering geological data. Especially in the realistic context of missing values, outliers and indicator redundancy in multiple source indicators, the existing methods are imperfect in the data preprocessing link, which further affects the discrimination performance and risk perception ability of the subsequent classification model.

[0005] Therefore, there is an urgent need for an intelligent discrimination method that can integrate spatial location, hydrological indicators and stratigraphic consistency information, and has the ability of adaptive feature screening and hyperparameter optimization, so as to improve the accuracy and reliability of collapsible grade assessment in loess areas. Summary of the Invention

[0006] One purpose of the present invention is to propose a method for automatic identification of collapsibility levels based on deep learning. The present invention dynamically evaluates the impact of misjudgment of different collapsibility levels on engineering safety, thereby guiding the algorithm to prioritize improving the recognition accuracy of high-risk levels.

[0007] According to an embodiment of the present invention, a method for automatically determining the collapsibility level based on deep learning includes the following steps:

[0008] S1. Collect multi-source loess geological engineering indicator data, construct a multi-source loess geological engineering indicator dataset, perform preprocessing on the multi-source loess geological engineering indicator dataset, and generate a preprocessed loess geological engineering indicator dataset;

[0009] S2. Construct a loess sampling unit graph structure based on the preprocessed loess geological engineering index dataset, generate weighted edge weights, and obtain a loess sampling unit graph structure dataset with a weighted adjacency matrix;

[0010] S3. Constructing a node attribute vector using each indicator in the preprocessed loess geological engineering indicator dataset, and embedding the node attribute vector into a loess sampling unit graph structure dataset with a weighted adjacency matrix to form a loess sampling unit weighted graph structure dataset with node attributes;

[0011] S4. Initialize the graph attention network model based on the weighted graph structure dataset of loess sampling units with node attributes and construct a multi-objective comprehensive fitness function;

[0012] S5. Randomly generate individuals in the differential flower pollination algorithm population, perform global search and local search on the initial population of the differential flower pollination algorithm, and iterate until the preset termination condition is met to obtain the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector;

[0013] S6. The graph attention network model is reconstructed using the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector to obtain the optimized graph attention network model. The weighted graph structure dataset of loess sampling units with node attributes is inferred to output the prediction results of the collapsibility level of each sampling unit.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Collect multi-source loess geological engineering indicator raw data and construct a multi-source loess geological engineering indicator raw data set, where each loess sampling unit corresponds to an engineering indicator vector. The engineering indicator vector contains engineering indicators in multiple dimensions, where each dimension represents a natural moisture content indicator, a natural dry density indicator, a porosity ratio indicator, and a groundwater level depth indicator.

[0016] S12. Perform missing value filling processing on the original dataset of multi-source loess geological engineering indicators, using the nearest neighbor interpolation method to fill in the missing items, and form a multi-source loess geological engineering indicator dataset after missing value filling;

[0017] S13. Performing outlier removal processing on the multi-source loess geological engineering indicator dataset after missing value filling to form a multi-source loess geological engineering indicator dataset after outlier removal;

[0018] S14. Perform numerical normalization on the multi-source loess geological engineering indicator dataset after outliers are removed, and use the linear interval scaling method to standardize each engineering indicator dimension to form the normalized loess geological engineering indicator dataset D norm .

[0019] Optionally, the S2 includes the following steps:

[0020] S21. Each loess sampling unit in the normalized loess geological engineering index dataset is treated as a graph node to construct a node set V, where each node in the node set corresponds to a loess sampling unit, and the node attributes consist of the normalized engineering index vector of the loess sampling unit;

[0021] S22. By comparing the spatial position coordinates of the i-th node and the j-th node in the three-dimensional space Calculate the spatial Euclidean distance between any two loess sampling units Corresponding to the longitude, latitude and depth components respectively;

[0022] S23. Calculate the similarity of hydrogeological indicators between any two loess sampling units by calculating the absolute difference between the i-th node and the j-th node in the groundwater depth indicator dimension and taking the complementary value.

[0023] S24. Calculate the stratification consistency index between any two loess sampling units by calculating the normalized value difference between the i-th node and the j-th node in the stratification age indicator dimension and multiplying the normalized value difference by the exponential function value after the adjustment coefficient.

[0024] S25. Based on the spatial Euclidean distance, hydrogeological index similarity and stratification consistency index, the weighted edge weight w between any i-th node and j-th node is calculated by edge weight fusion. i,j ;

[0025] S26. Construct an edge set between each node in the graph structure G = (V, E, W). The edge set consists of all node pairs whose weighted edge weights are greater than the edge weight threshold. The edge weight threshold is used to eliminate low-similarity node pairs to form a sparse graph structure, and the weighted adjacency matrix A is generated by filling the weighted edge weights. Each element of the weighted adjacency matrix indicates whether there is an edge connection between two nodes and its corresponding edge weight value. Output the loess sampling unit graph structure dataset G = (V, A) with the weighted adjacency matrix.

[0026] Optionally, S3 includes the following steps:

[0027] S31. Construct a node attribute vector for each loess sampling unit in the normalized loess geological engineering indicator dataset, and arrange each node attribute vector in sequence to form a node attribute matrix X, where each row in the node attribute matrix represents a loess sampling unit node, and each column represents an engineering indicator dimension;

[0028] S32. Combine the node attribute matrix and the weighted adjacency matrix to construct a weighted graph structure dataset G = (V, A, X) with node attributes;

[0029] S33. Perform consistency dimension verification processing on the weighted graph structure dataset G=(V, A, X) with node attributes, and output the formed loess sampling unit weighted graph structure dataset G=(V, A, X) with node attributes.

[0030] Optionally, the S4 includes the following steps:

[0031] S41. Initialize the graph attention network model based on the loess sampling unit weighted graph structure dataset with node attributes, and set the graph attention network hyperparameter vector, including the number of layers L of the graph attention network model and the number of attention heads per layer H. l , hidden unit dimension d l , nonlinear activation function σ l , Dropout ratio p l , initial learning rate η0, construct the initial graph attention network model structure set Θ0;

[0032] S42. Divide the dataset G into training set G train , validation set G val and the cross-interval test set G cross , cross-interval test set G cross Used to measure the generalization ability of the graph attention network model between different loess profiles; train the initial graph attention network model defined by the initial graph attention network model structure set, and record the key indicators of each stage, including the macro-average F1 value of the collapsibility grade category on the validation set val,m , macro-average F1 value on the cross-section test set F1 cross , the number of iterations and total number of model parameters required for training to converge on the validation set;

[0033] S43. Set a dynamic risk sensitivity weight for each collapsibility category The dynamic risk sensitivity weight is updated in each generation of evolution according to the ratio of the number of misjudgments to the number of accurate judgments of the collapsibility grade category in the previous generation;

[0034] S44. Multiply the macro-average F1 value of each collapsible grade category on the validation set by the corresponding dynamic risk sensitivity weight and then sum them up. Average them by the number of collapsible grade categories to obtain the adaptive macro-average F1 weighted index.

[0035] S45. Calculate the cross-profile generalization ability retention factor R gen ,The cross-section generalization ability preservation factor is used to measure the performance preservation of the graph attention network model when migrating between sections;

[0036] S46. Calculate the network deployment complexity index C deploy ,The network deployment complexity index is used to quantify the resource overhead of the graph attention network model during deployment;

[0037] S47. A dynamic feedback multi-objective fitness function is constructed by weighting the adaptive macro-average F1 weighted index, the cross-profile generalization capability preservation factor, and the network deployment complexity index according to the weighted coefficients.

[0038] Optionally, the S5 includes the following steps:

[0039] S51. Randomly generate the initial population of the differential flower pollination algorithm in the search space. Let the population size be P. Then each individual in the population Expressed as a joint vector form in, Represents the node feature selection mask vector of the i-th individual;

[0040] S52. Perform global search and local search on the initial population of the differential flower pollination algorithm. The global search and local search are respectively a differential vector mutation operation and a pollen tossing operation. The differential vector mutation operation is used to perform breadth exploration in the graph attention network hyperparameter vector space, and the pollen tossing operation is used to perform local search near the optimal solution.

[0041] S53. Selecting mask vector for node features The binary probability resampling mechanism is used, combined with the mask granularity control factor δ to perform the following update:

[0042]

[0043] Among them, rand() represents a randomly generated uniform variable, and the current dimension validity score is obtained by aggregating and normalizing the attention weights of each dimension in the graph attention network model;

[0044] S54. For each candidate individual Substitute into the dynamic feedback multi-objective fitness function The fitness score is obtained and compared with the current individual. If the fitness is improved, the update is accepted; otherwise, the original individual is retained to form a new generation of population.

[0045] S55. Repeat the differential mutation, pollen throwing, mask update and fitness screening processes from S52 to S54 until the maximum number of iterations or the fitness convergence condition is met, and output the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector in the current iteration.

[0046] Optionally, the differential vector mutation operation randomly selects the graph attention network hyperparameter vectors of three different individuals from the current population, calculates the vector difference between two population individuals, and then weightedly superimposes it with the vector of another population individual to obtain a new candidate graph attention network hyperparameter vector. The differential mutation factor used in the calculation is used to control the amplification factor of the difference.

[0047] Optionally, the pollen throwing operation obtains a local perturbation sample by adding the vector difference between the optimal individual and the current individual on the basis of the graph attention network hyperparameter vector of the current individual and multiplying it by a local guide step length proportional coefficient, where the local guide step length proportional coefficient is used to control the size of the jump amplitude.

[0048] Optionally, the S6 includes the following steps:

[0049] S61. Using the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector Reconstruct the graph attention network model structure;

[0050] S62. Reconstruct the node attribute matrix X * Combined with the weighted adjacency matrix A to form the final input graph structure G * =(V,A,X * ), with G * The graph attention network model configured with the optimal graph attention network hyperparameter vector is used as input for training. The training objective is to minimize the cross entropy loss of the collapsibility level labels of all nodes. The training termination criterion is that the macro-average F1 value of the collapsibility level category on the validation set reaches a stable level and does not improve.

[0051] S63. Use the trained graph attention network model to perform inference on the final input graph structure and output the predicted probability distribution y of the collapsibility level of each node i =(p i,1 ,p i,2 ,...,p i,M ), where p i,m represents the predicted probability that the i-th node belongs to the collapsibility level category m;

[0052] S64. Based on the predicted probability distribution, select the level corresponding to the maximum probability as the final predicted label

[0053] S65. Introduce a judgment standard based on engineering indicators, combined with the natural moisture content indicator ω i , natural dry density index ρ d,i , porosity index e i and groundwater level depth index h i An automatic collapsibility grade identification rule is constructed to obtain the final collapsibility grade.

[0054] Optionally, the automatic determination rule of the collapsibility level is:

[0055] If it satisfies: natural moisture index ω i <12, natural dry density index ρ d,i >1.65g / cm 3 , porosity index e i <0.65, groundwater depth index h i >8m, the final predicted label Defined as slightly collapsible grade;

[0056] If the natural moisture content index is 12≤ω i <16, natural dry density index 1.50≤ρ d,i ≤1.65g / cm 3 , porosity index 0.65≤e i <0.80, groundwater level depth index 5≤h i ≤8m, the final predicted label Defined as medium collapsibility grade;

[0057] If the natural moisture content index is 16≤ω i <20, natural dry density index 1.35≤ρ d,i <1.50g / cm 3 , porosity index 0.80≤e i <1.00, groundwater level depth index 2≤h i <5m, the final predicted label Defined as strong collapsibility grade;

[0058] If it satisfies: natural moisture index ω i ≥20, natural dry density index ρ d,i <1.35g / cm 3 , porosity index e i ≥1.00, groundwater depth index h i <2m, the final predicted label Defined as an extremely collapsible grade.

[0059] The beneficial effects of the present invention are:

[0060] (1) The present invention adopts the differential flower pollination algorithm to jointly optimize the structural hyperparameters of the graph attention network and the node feature selection mask vector, introduces the differential vector mutation operation in the global search and the pollen throwing mechanism in the local search, so that the search process has stronger jumpiness and convergence in the exploration space. At the same time, combined with the multi-objective fitness function of adaptive risk perception, it dynamically evaluates the impact of misjudgment of different wetting levels on engineering safety, thereby guiding the algorithm to prioritize improving the recognition accuracy of high-risk levels.

[0061] (2) The present invention constructs a graph structure representation of weighted edge weights by integrating spatial Euclidean distance, hydrogeological indicator similarity and stratification consistency index, and generates a sparse graph adjacency matrix to input into the graph neural network, so that the model can effectively characterize the spatial coupling relationship between loess sampling units, hydrological environment consistency and stratum development trend, and solves the problem of context semantic loss caused by the isolated modeling of node relationships in the existing technology.

[0062] (3) The present invention designs a dynamic feedback multi-objective fitness function, which integrates the adaptive weighted F1 risk index, the cross-profile generalization retention factor and the network deployment complexity index. Based on the performance feedback of the model in the validation set and the cross-interval test set at different stages, the evolution direction is dynamically adjusted to ensure that the model has stronger cross-regional generalization ability and deployability while maintaining high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0064] Figure 1 This is a flow chart of a method for automatically distinguishing collapsible levels based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0065] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0066] refer to Figure 1 , a method for automatically distinguishing collapsible levels based on deep learning, comprising the following steps:

[0067] S1. Collect multi-source loess geological engineering indicator data, construct a multi-source loess geological engineering indicator dataset, perform preprocessing on the multi-source loess geological engineering indicator dataset, and generate a preprocessed loess geological engineering indicator dataset;

[0068] S2. Construct a loess sampling unit graph structure based on the preprocessed loess geological engineering index dataset, generate weighted edge weights, and obtain a loess sampling unit graph structure dataset with a weighted adjacency matrix;

[0069] S3. Constructing a node attribute vector using each indicator in the preprocessed loess geological engineering indicator dataset, and embedding the node attribute vector into a loess sampling unit graph structure dataset with a weighted adjacency matrix to form a loess sampling unit weighted graph structure dataset with node attributes;

[0070] S4. Initialize the graph attention network model based on the weighted graph structure dataset of loess sampling units with node attributes and construct a multi-objective comprehensive fitness function;

[0071] S5. Randomly generate individuals in the differential flower pollination algorithm population, perform global search and local search on the initial population of the differential flower pollination algorithm, and iterate until the preset termination condition is met to obtain the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector;

[0072] S6. The graph attention network model is reconstructed using the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector to obtain the optimized graph attention network model. The weighted graph structure dataset of loess sampling units with node attributes is inferred to output the prediction results of the collapsibility level of each sampling unit.

[0073] In this embodiment, S1 includes the following steps:

[0074] S11. Collect multi-source loess geological engineering indicator raw data and construct a multi-source loess geological engineering indicator raw data set, where each loess sampling unit corresponds to an engineering indicator vector. The engineering indicator vector contains engineering indicators in multiple dimensions, where each dimension represents a natural moisture content indicator, a natural dry density indicator, a porosity ratio indicator, and a groundwater level depth indicator.

[0075] S12. Perform missing value filling processing on the original dataset of multi-source loess geological engineering indicators, using the nearest neighbor interpolation method to fill in the missing items, and form a multi-source loess geological engineering indicator dataset after missing value filling;

[0076] For a dimension of the engineering indicator vector with missing items, find an engineering indicator vector with valid values ​​in other spatially adjacent sampling units, calculate the average value of the adjacent engineering indicator vectors in the current dimension, and use the obtained average value to fill the value of the current missing dimension;

[0077] S13. Performing outlier removal processing on the multi-source loess geological engineering indicator dataset after missing value filling to form a multi-source loess geological engineering indicator dataset after outlier removal;

[0078] During the processing, for each engineering indicator dimension, the mean and standard deviation of all samples in the current dimension are calculated, and the anomaly detection interval of the current dimension is defined based on the mean and standard deviation. The anomaly detection interval is constructed by the upper and lower limits, which are the current dimension mean minus λ times the standard deviation and the mean plus λ times the standard deviation. If the current indicator dimension value of a loess sampling unit is not within the current anomaly detection interval, it is marked as an outlier and the current value is cleared, and refilled according to the nearest neighbor interpolation method in S12;

[0079] S14. Perform numerical normalization on the multi-source loess geological engineering indicator dataset after outliers are removed, and use the linear interval scaling method to standardize each engineering indicator dimension to form the normalized loess geological engineering indicator dataset D norm ;

[0080] For a certain indicator dimension value of each loess sampling unit, the minimum value in the current dimension is used as the lower limit and the maximum value as the upper limit to calculate the proportion position of the current value in the current interval. The current proportion position is the normalized value of the current indicator dimension. After normalization, all engineering indicators are compressed to a closed interval between zero and one.

[0081] In this embodiment, S2 includes the following steps:

[0082] S21. Each loess sampling unit in the normalized loess geological engineering index dataset is treated as a graph node to construct a node set V, where each node in the node set corresponds to a loess sampling unit, and the node attributes consist of the normalized engineering index vector of the loess sampling unit;

[0083] S22. By comparing the spatial position coordinates of the i-th node and the j-th node in the three-dimensional space Calculate the spatial Euclidean distance between any two loess sampling units Corresponding to the longitude, latitude and depth components respectively, the spatial Euclidean distance is used to reflect the spatial proximity of two nodes;

[0084] S23. Calculate the similarity of hydrogeological indicators between any two loess sampling units by calculating the absolute difference between the i-th node and the j-th node in the groundwater depth indicator dimension and taking the complementary value. The value range of the hydrogeological indicator similarity is a closed interval between zero and one. The greater the hydrogeological indicator similarity, the higher the similarity between the two nodes in hydrogeological characteristics.

[0085] S24. Calculate the stratification consistency index between any two loess sampling units by calculating the normalized value difference between the i-th node and the j-th node in the stratification age indicator dimension and multiplying the normalized value difference by the exponential function value after the adjustment coefficient. The stratification consistency index ranges from 0 to 1, a closed interval. The stratification consistency score is used to measure the consistency between two nodes in the stratification age indicator dimension. The adjustment coefficient is used to control the sensitivity of the score change. The stratification consistency score ranges from 0 to 1, an open interval.

[0086] S25. Based on the spatial Euclidean distance, hydrogeological index similarity and stratification consistency index, the weighted edge weight w between any i-th node and j-th node is calculated by fusion. i,j :

[0087]

[0088] Among them, β1, β2, and β3 are the weight coefficients of spatial Euclidean distance, hydrogeological index similarity, and stratification consistency index on edge weight, respectively;

[0089] S26. Construct an edge set between each node in the graph structure G = (V, E, W). The edge set consists of all node pairs whose weighted edge weights are greater than the edge weight threshold. The edge weight threshold is used to eliminate low-similarity node pairs to form a sparse graph structure, and the weighted adjacency matrix A is generated by filling the weighted edge weights. Each element of the weighted adjacency matrix indicates whether there is an edge connection between two nodes and its corresponding edge weight value. Output the loess sampling unit graph structure dataset G = (V, A) with the weighted adjacency matrix.

[0090] In this embodiment, S3 includes the following steps:

[0091] S31. Construct a node attribute vector for each loess sampling unit in the normalized loess geological engineering indicator dataset. The node attribute vector is used to represent the attribute information of the current loess sampling unit in multiple geological indicator dimensions. Each node attribute vector is sequentially arranged to form a node attribute matrix X. Each row in the node attribute matrix represents a loess sampling unit node, and each column represents an engineering indicator dimension. The number of rows in the node attribute matrix is ​​equal to the total number of loess sampling units, and the number of columns is equal to the number of engineering indicator dimensions.

[0092] S32. Combine the node attribute matrix and the weighted adjacency matrix to construct a weighted graph structure dataset G = (V, A, X) with node attributes;

[0093] The weighted graph structure dataset includes a node set, where each node corresponds to a loess sampling unit; a weighted adjacency matrix, where each element indicates whether there is an edge connection between two loess sampling unit nodes and its corresponding edge weight. The edge weight reflects the comprehensive correlation between the two sampling units in terms of spatial proximity, similarity of hydrogeological indicators, and consistency of stratification; and a node attribute matrix, where each row represents the geological indicator attribute information of the corresponding node, including the normalized values ​​of the natural water content indicator, natural dry density indicator, porosity indicator, and groundwater level depth indicator.

[0094] S33. Perform consistency dimension verification processing on the weighted graph structure dataset G = (V, A, X) with node attributes, so that the dimension K of the node attribute matrix X is consistent with the feature dimension of the graph attention network model input, and at the same time ensure that the number of nodes N matches the dimension of the weighted adjacency matrix A, and output the weighted graph structure dataset G = (V, A, X) of the loess sampling unit with node attributes.

[0095] In this embodiment, S4 includes the following steps:

[0096] S41. Initialize the graph attention network model based on the loess sampling unit weighted graph structure dataset with node attributes, and set the graph attention network hyperparameter vector, including the number of layers L of the graph attention network model and the number of attention heads per layer H. l , hidden unit dimension d l , nonlinear activation function σ l , Dropout ratio p l , initial learning rate η0, construct the initial graph attention network model structure set Θ0;

[0097] S42. Divide the dataset G into training set G train , validation set G val and the cross-interval test set G cross , cross-interval test set G cross Used to measure the generalization ability of the graph attention network model between different loess profiles; train the initial graph attention network model defined by the initial graph attention network model structure set, and record the key indicators of each stage, including the F1 value of the collapsible grade category m on the validation set val,m , macro-average F1 value on the cross-section test set F1 cross , the number of iterations and total number of model parameters required for training to converge on the validation set;

[0098] S43. Set a dynamic risk sensitivity weight for each collapsibility category The dynamic risk sensitivity weight is used to quantify the impact of the current collapsibility grade category on project safety in the event of a misjudgment. The dynamic risk sensitivity weight is updated in each generation based on the ratio of the number of misjudgments to the number of accurate judgments of the collapsibility grade category in the previous generation.

[0099]

[0100] in, is the number of misjudgments of collapsibility category m in round t, is the number of accurate identifications of collapsibility grade m, ∈ is a minimum constant to prevent division by zero, and λ is the weighting coefficient for misjudgment;

[0101] The S43 formula essentially establishes a dynamic feedback relationship between the risk weight of the collapsibility grade category m$ and the misjudgment rate of the grade in the current generation of the model. λ is used as a regulating factor to introduce the ratio of real-time misjudgments to accurate judgments.

[0102] Dynamic risk sensitivity weights are no longer traditional constants statically assigned by engineers based on experience, nor are they manually assigned in the post-processing stage. Instead, they are deeply integrated into the model optimization process and dynamically driven by the actual difficulty of engineering category identification and the risk of misjudgment. Each evolution reflects the model's actual adaptability to different levels of categories. This implementation breaks through the static evaluation limitations of conventional F1 macro / micro weighted indicators, allowing the model to focus more on error-prone and high-risk categories, improving its practical value oriented towards engineering safety and significantly enhancing the model's adaptability and robustness to engineering scenarios.

[0103] S44. Multiply the macro-average F1 value of each collapsible grade category on the validation set by the corresponding dynamic risk sensitivity weight and then sum them up. Average them by the number of collapsible grade categories to obtain the adaptive macro-average F1 weighted index. The adaptive macro-average F1 weighted index is used to reflect the difference in the impact of different collapsibility levels on the performance of the graph attention network model after misjudgment. The greater the weight of the high-risk level category, the more the graph attention network model tends to prioritize improving its discrimination accuracy.

[0104] S45. Calculate the cross-profile generalization ability retention factor R gen The cross-profile generalization ability retention factor is used to measure the performance retention of the graph attention network model when migrating between profiles. It is calculated by dividing the macro-average F1 value on the cross-profile test set by the macro-average F1 value on the validation set. It indicates the generalization stability of the graph attention network model under new profile conditions relative to the known dataset. The closer the cross-profile generalization ability retention factor is to 1, the stronger the cross-scene discrimination ability of the graph attention network model is.

[0105] S46. Calculate the network deployment complexity index C deployThe network deployment complexity index is used to quantify the resource overhead of the graph attention network model during deployment. It is obtained by multiplying the logarithm of the total number of parameters of the graph attention network model by the ratio of the training iteration rounds. The training iteration rounds are normalized with the standard reference round number. The higher the deployment complexity index, the more difficult it is to execute the graph attention network model on resource-constrained devices.

[0106]

[0107] Among them, T ref is the standard reference training round number;

[0108] The total amount of model parameters (C params , i.e., resource consumption of deploying the model) and training rounds (T train , that is, convergence efficiency). The construction method does not simply take the number of parameters or the number of training rounds as the optimization target, but combines the two to suppress the risk of resource surge in large-scale models through a logarithmic function. At the same time, the convergence efficiency is reflected by normalizing the number of training rounds, focusing on the difficulty of engineering implementation in scenarios with limited software and hardware resources during actual deployment.

[0109] In existing methods, model optimization is mostly limited to minimizing accuracy, loss, and simple parameter count, and deployment resource consumption and model training efficiency are rarely integrated into quantitative optimization objectives. This implementation seamlessly embeds engineering deployability indicators into the main loop of the differential flower pollination algorithm, allowing the automated tuning results to be directly connected to practical applications.

[0110] S47. A dynamic feedback multi-objective fitness function is constructed by weighting the adaptive macro-average F1 weighted index, the cross-profile generalization capability preservation factor, and the network deployment complexity index according to the weighted coefficients.

[0111]

[0112] Among them, ω1, ω2, and ω3 are weighting coefficients, and the weights are set according to actual engineering requirements.

[0113] In this embodiment, S5 includes the following steps:

[0114] S51. Randomly generate the initial population of the differential flower pollination algorithm in the search space. Let the population size be P. Then each individual in the population Expressed as a joint vector form in, Represents the node feature selection mask vector of the i-th individual, corresponding to the K engineering indicator dimensions in the node attribute matrix X. The value of each mask dimension is 1, indicating that the engineering indicator is retained, and 0, indicating that it is blocked;

[0115] S52. Perform global search and local search on the initial population of the differential flower pollination algorithm. The global search and local search are respectively a differential vector mutation operation and a pollen tossing operation. The differential vector mutation operation is used to perform breadth exploration in the graph attention network hyperparameter vector space, and the pollen tossing operation is used to perform local search near the optimal solution.

[0116] S53. Selecting mask vector for node features The binary probability resampling mechanism is used, combined with the mask granularity control factor δ to perform the following update:

[0117]

[0118] Among them, rand() represents a randomly generated uniform variable. The current dimension validity score is obtained by aggregating and normalizing the attention weights of each dimension in the graph attention network model, which is used to guide the high-weight feature dimensions to have a higher probability of being retained.

[0119] The formula is an adaptive update mechanism for node feature selection mask vectors. The core idea is to dynamically adjust whether each engineering indicator participates in subsequent model input based on the current feature importance distribution of the model during the evolutionary search process.

[0120] S55 uses the attention weight within the model as the feature dimension validity score and couples it with a probabilistic random mechanism to achieve dynamic adaptive evolution of the feature selection subspace. That is, whether a feature is retained not only depends on the random perturbation of the external search, but also combines the real-time feedback of the internal discriminant contribution of the deep model, effectively avoiding the problems of discarding excellent features or repeatedly selecting redundant features caused by traditional global / random mutations.

[0121] The formula continuously re-evaluates the importance of features in each generation of evolution based on the attention actually learned by the model. Compared with traditional genetic algorithms or particle swarms that use simple probabilistic mask mutation, the present invention achieves deep coupling of structural optimization and learning optimization, and improves the robustness and generalization ability of discrimination under heterogeneous and multi-redundant engineering data.

[0122] S54. For each candidate individual Substitute into the dynamic feedback multi-objective fitness function The fitness score is obtained and compared with the current individual. If the fitness is improved, the update is accepted; otherwise, the original individual is retained to form a new generation of population.

[0123] S55. Repeat the differential mutation, pollen throwing, mask update and fitness screening processes from S52 to S54 until the maximum number of iterations or the fitness convergence condition is met, and output the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector in the current iteration.

[0124] In this implementation, the differential vector mutation operation randomly selects the graph attention network hyperparameter vectors of three different individuals from the current population, calculates the vector difference between two population individuals, and then performs weighted superposition with the vector of the other population individual to obtain a new candidate graph attention network hyperparameter vector. The differential mutation factor used in the calculation is used to control the amplification factor of the difference. The purpose of the differential vector mutation operation is to guide the search direction to jump towards the high-quality solution area and optimize the global exploration capability.

[0125] In this embodiment, the pollen tossing operation is performed by adding the vector difference between the optimal individual and the current individual on the basis of the graph attention network hyperparameter vector of the current individual, and multiplying it by the local guide step ratio coefficient to obtain a local perturbation sample. The local guide step ratio coefficient is used to control the size of the jump amplitude. The purpose of the pollen tossing operation is to perform a fine-tuning search near the current optimal individual in the population, so as to improve the local convergence performance of the differential flower pollination algorithm in the later stage of the search.

[0126] In this embodiment, S6 includes the following steps:

[0127] S61. Using the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector Reconstruct the graph attention network model structure;

[0128] S62. Reconstruct the node attribute matrix X * Combined with the weighted adjacency matrix A to form the final input graph structure G * =(V,A,X * ), with G * As input, the graph attention network model configured with the optimal graph attention network hyperparameter vector is trained. The training objective is to minimize the collapsible level label cross entropy loss of all nodes. The training termination criterion is that the macro-average F1 value of the validation set reaches a stable state and does not improve.

[0129] S63. Use the trained graph attention network model to the final input graph structure G * =(V,A,X * ) performs inference and outputs the predicted probability distribution y of the collapsibility level of each node i =(p i,1 ,p i,2 ,...,p i,M ), where p i,m represents the predicted probability that the i-th node belongs to the collapsibility level category m;

[0130] S64. Based on the predicted probability distribution, select the level corresponding to the maximum probability as the final predicted label:

[0131]

[0132] S65. Introduce a judgment standard based on engineering indicators, combined with the natural moisture content indicator ω i , natural dry density index ρ d,i , porosity index e i and groundwater level depth index h i , construct the automatic judgment rules of collapsibility grade and obtain the final collapsibility grade.

[0133] In this embodiment, the automatic determination rule of collapsibility level is as follows:

[0134] If it satisfies: natural moisture index ω i <12, natural dry density index ρ d,i >1.65g / cm 3 , porosity index e i <0.65, groundwater depth index h i >8m, the final predicted label Defined as slightly collapsible grade;

[0135] If the natural moisture content index is 12≤ω i <16, natural dry density index 1.50≤ρ d,i ≤1.65g / cm 3 , porosity index 0.65≤e i <0.80, groundwater level depth index 5≤h i ≤8m, the final predicted label Defined as medium collapsibility grade;

[0136] If the natural moisture content index is 16≤ω i <20, natural dry density index 1.35≤ρ d,i <1.50g / cm 3 , porosity index 0.80≤e i <1.00, groundwater level depth index 2≤h i <5m, the final predicted label Defined as strong collapsibility grade;

[0137] If it satisfies: natural moisture index ω i ≥20, natural dry density index ρ d,i <1.35g / cm 3 , porosity index e i ≥1.00, groundwater depth index h i <2m, the final predicted label Defined as the extreme collapsibility grade;

[0138] If the predicted labels of the graph attention network model are inconsistent with the engineering rule division results, the level corresponding to the higher collapse risk is preferentially selected as the final output.

[0139] Example 1: Located in the construction section of Line 2 of a certain urban rail transit project, the new subway station and section tunnel involve a typical loess-covered area, with a geological risk of large-scale distribution of collapsible loess layers. The project is located in the southeastern part of Area A, with a site length of approximately 2.2 kilometers from east to west and a width of approximately 0.7 kilometers from north to south, with a total land area of ​​over 1.3 square kilometers. The terrain is mainly gentle slopes, and the thickness of the surface loess ranges from 8 to 28 meters.

[0140] Traditional practices generally rely on manual stratified sampling, indoor physical and mechanical tests, and manual identification of collapsibility grades. It is difficult to obtain high-density classification maps in a short period of time, and different experts often have subjective differences in indicator weighting and classification thresholds. It is difficult to meet the needs of rapid and automatic identification of large-scale engineering sections. In the early geological survey stage of this project, due to the uneven distribution of data points and large differences in the physical properties of soil layers at different depths, the traditional small-sample machine learning methods based on empirical formulas, support vector machines, decision trees, etc., are used, but the spatial consistency and cross-interval generalization ability of the identification results are unsatisfactory.

[0141] The engineering geological survey team used this method to compare with the traditional SVM discrimination method and the manual expert experience method to verify the degree of automation, discrimination accuracy, engineering applicability and actual effect of this method.

[0142] The construction site geological engineers conducted standard drilling, static penetration, on-site shear and water immersion compression tests, laying out 68 boreholes within the site. Each borehole obtained multiple sampling units at a depth of 2 meters, resulting in a total of 352 loess sampling unit data samples. Each sample contains the following engineering geological indicators:

[0143] Natural moisture content ω(%): 8.1~23.6, natural dry density ρ d (g / cm 3 ): 1.28~1.73, porosity index e: 0.61~1.12, groundwater level depth h(m): 0.9~11.4;

[0144] All the original data were formatted, anomalies eliminated and normalized, and finally a structured multi-source loess geological engineering indicator dataset was formed.

[0145] The data from all 352 sampling units and their corresponding spatial coordinates were used to construct a node set and a weighted adjacency matrix. Node attributes were the aforementioned nine-dimensional indices, and edge weights were integrated with spatial Euclidean distance, hydrogeological depth difference, and stratigraphic age consistency weights. Low-similarity edges were automatically removed to form a sparse graph structure. Graph visualization was used to clearly visualize node distribution and spatial topological relationships.

[0146] During the model initialization stage, the differential flower pollination algorithm is used to automatically search for the key structural parameters of the graph attention network (3-5 layer structure, 4-8 attention heads per layer, 64-256 hidden units, Dropout 0.2-0.5), feature selection mask, and set the dynamic risk weight, adaptive F1 index, and cross-profile generalization ability index optimization targets. Traditional models such as SVM use manual feature screening and grid method to optimize hyperparameters.

[0147] All samples were randomly divided into a training set (60%, 211 cases), a validation set (20%, 71 cases), and a cross-section test set (20%, 70 cases), ensuring a uniform distribution of different stratum thicknesses and groundwater level variations across the intervals. All models were compared based on the same partitioning.

[0148] During the model training phase, the method of the present invention guides evolutionary parameter adjustment through a differential flower pollination algorithm and a multi-objective fitness function, automatically obtaining the optimal structure and feature subset after approximately 100 generations. Convergence is achieved after an average of 37 training rounds, far superior to the SVM method (an average of 112 rounds). Hyperparameter and feature selection are automatically completed with minimal human intervention.

[0149] In the inference stage, the graph attention network model directly outputs the probability distribution of the collapsibility level of all samples and calibrates it in combination with the engineering discrimination rules. i <12 and ρ d,i >1.65g / cm 3 When the low collapsibility conditions of the engineering specifications are met, it is automatically determined to be a slightly collapsible grade; otherwise, it is automatically classified according to the probability distribution and rule priority.

[0150] Ultimately, the collapsibility grade identification labels of all nodes are automatically summarized to form a multi-level spatial distribution map and grading report, which provides a basis for intensified surveys and foundation pit design in high-risk areas of subway projects. For areas with extremely strong and severe collapsibility, engineers can quickly locate the relevant drill holes and plot numbers to achieve differentiated foundation treatment.

[0151] Table 1 Comparison of model discrimination accuracy and generalization ability

[0152]

[0153] Table 2 Typical sample display (part)

[0154]

[0155] As shown in Tables 1 and 2, the proposed method achieved a 95.7% agreement between collapsibility classification and later-stage foundation settlement measurements across all 70 cross-section test samples, exceeding the 82.8% achieved by the SVM method and the 85.7% achieved by expert judgment. The algorithm automatically achieves a collaborative optimization of features, structure, risk, and generalization. Even in engineering scenarios characterized by extremely uneven sample distribution, strong spatial heterogeneity, and inconsistent human experience, it maintains a high degree of automation, objectivity, and adaptability, significantly reducing the burden of manual intervention and empirical judgment.

[0156] The traditional SVM missed 7 cases and misjudged 3 cases in the extremely strong and strong collapsible samples. The manual interpretation of experts had 6 disagreements due to subjective weighting differences. However, the method of the present invention accurately identified all high-risk samples.

[0157] The present invention adopts the differential flower pollination algorithm to jointly optimize the structural hyperparameters of the graph attention network and the node feature selection mask vector. It introduces the differential vector mutation operation in the global search and the pollen throwing mechanism in the local search, so that the search process has stronger jumpiness and convergence in the exploration space. At the same time, combined with the multi-objective fitness function of adaptive risk perception, it dynamically evaluates the impact of misjudgment of different wetting levels on engineering safety, thereby guiding the algorithm to prioritize improving the recognition accuracy of high-risk levels.

[0158] The present invention constructs a graph structure representation of weighted edge weights by integrating spatial Euclidean distance, hydrogeological indicator similarity and stratification consistency index, and generates a sparse graph adjacency matrix to input into the graph neural network, so that the model can effectively characterize the spatial coupling relationship, hydrological environment consistency and stratum development trend between loess sampling units, and solves the problem of context semantic loss caused by the isolated modeling of node relationships in the existing technology.

[0159] The present invention designs a dynamic feedback multi-objective fitness function, which integrates the adaptive weighted F1 risk index, the cross-profile generalization retention factor and the network deployment complexity index. Based on the performance feedback of the models at different stages in the validation set and the cross-interval test set, the evolution direction is dynamically adjusted to ensure that the model has stronger cross-regional generalization ability and deployability while maintaining high accuracy.

[0160] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for automatically distinguishing collapsible levels based on deep learning, characterized in that: The steps include: S1. Collect multi-source loess geological engineering indicator data, construct a multi-source loess geological engineering indicator dataset, perform preprocessing on the multi-source loess geological engineering indicator dataset, and generate a preprocessed loess geological engineering indicator dataset; S2. Construct a loess sampling unit graph structure based on the preprocessed loess geological engineering index dataset, generate weighted edge weights, and obtain a loess sampling unit graph structure dataset with a weighted adjacency matrix; S3. Constructing a node attribute vector using each indicator in the preprocessed loess geological engineering indicator dataset, and embedding the node attribute vector into a loess sampling unit graph structure dataset with a weighted adjacency matrix to form a loess sampling unit weighted graph structure dataset with node attributes; S4. Initialize the graph attention network model based on the weighted graph structure dataset of loess sampling units with node attributes and construct a multi-objective comprehensive fitness function; S5. Randomly generate individuals in the differential flower pollination algorithm population, perform global search and local search on the initial population of the differential flower pollination algorithm, and iterate until the preset termination condition is met to obtain the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector; S6. The graph attention network model is reconstructed using the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector to obtain the optimized graph attention network model. The weighted graph structure dataset of loess sampling units with node attributes is inferred to output the prediction results of the collapsibility level of each sampling unit.

2. The method for automatically determining the collapsibility level based on deep learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect multi-source loess geological engineering indicator raw data and construct a multi-source loess geological engineering indicator raw data set, where each loess sampling unit corresponds to an engineering indicator vector. The engineering indicator vector contains engineering indicators in multiple dimensions, where each dimension represents a natural moisture content indicator, a natural dry density indicator, a porosity ratio indicator, and a groundwater level depth indicator. S12. Perform missing value filling processing on the original dataset of multi-source loess geological engineering indicators, using the nearest neighbor interpolation method to fill in the missing items, and form a multi-source loess geological engineering indicator dataset after missing value filling; S13. Performing outlier removal processing on the multi-source loess geological engineering indicator dataset after missing value filling to form a multi-source loess geological engineering indicator dataset after outlier removal; S14. Perform numerical normalization on the multi-source loess geological engineering indicator dataset after outliers are removed, and use the linear interval scaling method to standardize each engineering indicator dimension to form the normalized loess geological engineering indicator dataset D norm .

3. The method for automatically distinguishing collapsible levels based on deep learning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Each loess sampling unit in the normalized loess geological engineering index dataset is treated as a graph node to construct a node set V, where each node in the node set corresponds to a loess sampling unit, and the node attributes consist of the normalized engineering index vector of the loess sampling unit; S22. By comparing the spatial position coordinates of the i-th node and the j-th node in the three-dimensional space Calculate the spatial Euclidean distance between any two loess sampling units Corresponding to the longitude, latitude and depth components respectively; S23. Calculate the similarity of hydrogeological indicators between any two loess sampling units by calculating the absolute difference between the i-th node and the j-th node in the groundwater depth indicator dimension and taking the complementary value. S24. Calculate the stratification consistency index between any two loess sampling units by calculating the normalized value difference between the i-th node and the j-th node in the stratification age indicator dimension and multiplying the normalized value difference by the exponential function value after the adjustment coefficient. S25. Based on the spatial Euclidean distance, hydrogeological index similarity and stratification consistency index, the weighted edge weight w between any i-th node and j-th node is calculated by edge weight fusion. i,j ; S26. Construct an edge set between each node in the graph structure G = (V, E, W). The edge set consists of all node pairs whose weighted edge weights are greater than the edge weight threshold. The edge weight threshold is used to eliminate low-similarity node pairs to form a sparse graph structure, and the weighted adjacency matrix A is generated by filling the weighted edge weights. Each element of the weighted adjacency matrix indicates whether there is an edge connection between two nodes and its corresponding edge weight value. Output the loess sampling unit graph structure dataset G = (V, A) with the weighted adjacency matrix.

4. The method for automatically determining the collapsibility level based on deep learning according to claim 3 is characterized in that: The S3 includes the following steps: S31. Construct a node attribute vector for each loess sampling unit in the normalized loess geological engineering indicator dataset, and arrange each node attribute vector in sequence to form a node attribute matrix X, where each row in the node attribute matrix represents a loess sampling unit node, and each column represents an engineering indicator dimension; S32. Combine the node attribute matrix and the weighted adjacency matrix to construct a weighted graph structure dataset G = (V, A, X) with node attributes; S33. Perform consistency dimension verification processing on the weighted graph structure dataset G=(V, A, X) with node attributes, and output the formed loess sampling unit weighted graph structure dataset G=(V, A, X) with node attributes.

5. The method for automatically determining the collapsibility level based on deep learning according to claim 4 is characterized in that: The S4 comprises the following steps: S41. Initialize the graph attention network model based on the loess sampling unit weighted graph structure dataset with node attributes, and set the graph attention network hyperparameter vector, including the number of layers L of the graph attention network model and the number of attention heads per layer H. l , hidden unit dimension d l , nonlinear activation function σ l , Dropout ratio p l , initial learning rate η0, construct the initial graph attention network model structure set Θ0; S42. Divide the dataset G into training set G train , validation set G val and the cross-interval test set G cross , cross-interval test set G cross Used to measure the generalization ability of the graph attention network model between different loess profiles; train the initial graph attention network model defined by the initial graph attention network model structure set, and record the key indicators of each stage, including the macro-average F1 value of the collapsibility grade category on the validation set val,m , macro-average F1 value on the cross-section test set F1 cross , the number of iterations and total number of model parameters required for training to converge on the validation set; S43. Set a dynamic risk sensitivity weight for each collapsibility category The dynamic risk sensitivity weight is updated in each generation of evolution according to the ratio of the number of misjudgments to the number of accurate judgments of the collapsibility grade category in the previous generation; S44. Multiply the macro-average F1 value of each collapsible grade category on the validation set by the corresponding dynamic risk sensitivity weight and then sum them up. Average them by the number of collapsible grade categories to obtain the adaptive macro-average F1 weighted index. S45. Calculate the cross-profile generalization ability retention factor R gen ,The cross-section generalization ability preservation factor is used to measure the performance preservation of the graph attention network model when migrating between sections; S46. Calculate the network deployment complexity index C deploy ,The network deployment complexity index is used to quantify the resource overhead of the graph attention network model during deployment; S47. A dynamic feedback multi-objective fitness function is constructed by weighting the adaptive macro-average F1 weighted index, the cross-profile generalization capability preservation factor, and the network deployment complexity index according to the weighted coefficients.

6. The method for automatically determining the collapsibility level based on deep learning according to claim 5, characterized in that: The S5 comprises the following steps: S51. Randomly generate the initial population of the differential flower pollination algorithm in the search space. Let the population size be P. Then each individual in the population Expressed as a joint vector form in, Represents the node feature selection mask vector of the i-th individual; S52. Perform global search and local search on the initial population of the differential flower pollination algorithm. The global search and local search are respectively a differential vector mutation operation and a pollen tossing operation. The differential vector mutation operation is used to perform breadth exploration in the graph attention network hyperparameter vector space, and the pollen tossing operation is used to perform local search near the optimal solution. S53. Selecting mask vector for node features The binary probability resampling mechanism is used, combined with the mask granularity control factor δ to perform the following update: Among them, rand() represents a randomly generated uniform variable, and the current dimension validity score is obtained by aggregating and normalizing the attention weights of each dimension in the graph attention network model; S54. For each candidate individual Substitute into the dynamic feedback multi-objective fitness function The fitness score is obtained and compared with the current individual. If the fitness is improved, the update is accepted; otherwise, the original individual is retained to form a new generation of population. S55. Repeat the differential mutation, pollen throwing, mask update and fitness screening processes from S52 to S54 until the maximum number of iterations or the fitness convergence condition is met, and output the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector in the current iteration.

7. The method for automatically determining the collapsibility level based on deep learning according to claim 6, characterized in that: The differential vector mutation operation randomly selects the graph attention network hyperparameter vectors of three different individuals from the current population, calculates the vector difference between two population individuals, and then weightedly superimposes it with the vector of the other population individual to obtain a new candidate graph attention network hyperparameter vector. The differential mutation factor used in the calculation is used to control the amplification factor of the difference.

8. The method for automatically determining the collapsibility level based on deep learning according to claim 6, characterized in that: The pollen tossing operation adds the vector difference between the optimal individual and the current individual to the graph attention network hyperparameter vector of the current individual, and multiplies it by the local guide step scale coefficient to obtain a local perturbation sample. The local guide step scale coefficient is used to control the size of the jump amplitude.

9. The method for automatically determining the collapsibility level based on deep learning according to claim 6, characterized in that: The S6 comprises the following steps: S61. Using the optimal graph attention network hyperparameter vector and the optimal node feature selection mask vector Reconstruct the graph attention network model structure; S62. Reconstruct the node attribute matrix X * Combined with the weighted adjacency matrix A to form the final input graph structure G * =(V,A,X * ), with G * The graph attention network model configured with the optimal graph attention network hyperparameter vector is used as input for training. The training objective is to minimize the cross entropy loss of the collapsibility level labels of all nodes. The training termination criterion is that the macro-average F1 value of the collapsibility level category on the validation set reaches a stable level and does not improve. S63. Use the trained graph attention network model to perform inference on the final input graph structure and output the predicted probability distribution y of the collapsibility level of each node i =(p i,1 ,p i,2 ,...,p i,M ), where p i,m represents the predicted probability that the i-th node belongs to the collapsibility level category m; S64. Based on the predicted probability distribution, select the level corresponding to the maximum probability as the final predicted label S65. Introduce a judgment standard based on engineering indicators, combined with the natural moisture content indicator ω i , natural dry density index ρ d,i , porosity index e i and groundwater level depth index h i An automatic collapsibility grade identification rule is constructed to obtain the final collapsibility grade.

10. The method for automatically determining the collapsibility level based on deep learning according to claim 6, characterized in that: The automatic determination rule of collapsibility grade is as follows: If it satisfies: natural moisture index ω i <12, natural dry density index ρ d,i >1.65g / cm 3 , porosity index e i <0.65, groundwater depth index h i >8m, the final predicted label Defined as slightly collapsible grade; If the natural moisture content index is 12≤ω i <16, natural dry density index 1.50≤ρ d,i ≤1.65g / cm 3 , porosity index 0.65≤e i <0.80, groundwater level depth index 5≤h i ≤8m, the final predicted label Defined as medium collapsibility grade; If the natural moisture content index is 16≤ω i <20, natural dry density index 1.35≤ρ d,i <1.50g / cm 3 , porosity index 0.80≤e i <1.00, groundwater level depth index 2≤h i <5m, the final predicted label Defined as strong collapsibility grade; If it satisfies: natural moisture index ω i ≥20, natural dry density index ρ d,i <1.35g / cm 3 , porosity index e i ≥1.00, groundwater depth index h i <2m, the final predicted label Defined as an extremely collapsible grade.

Images