Anti-collision three-dimensional layout method for complex reinforcement mesh of fan pile bearing foundation

Abstract

The present application relates to the technical field of computer-aided architectural design, and in particular to a complex steel mesh anti-collision three-dimensional layout method for fan pile bearing foundation, which solves the technical problem that the existing technology causes cross-scene layout parameters to fluctuate dramatically due to over-fitting of model environment data, resulting in the risk of structural collision. The method comprises: obtaining historical engineering data; analyzing the correlation between each attribute parameter and sample three-dimensional layout data, determining the influence weight of each attribute parameter, and determining the value fluctuation degree according to the influence weight; determining the fluctuation sensitivity of each network layer according to the value fluctuation degree of each attribute parameter, the output error of the reference model and the output error of the layer-deleted model; setting the corresponding regularization strength according to the fluctuation sensitivity of each network layer for training to obtain the trained target model; and inputting the target engineering data into the target model to output the steel mesh three-dimensional layout data.

Description

Technical Field

[0001] This invention relates to the field of computer-aided architectural design technology, specifically to a method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations. Background Technology

[0002] Onshore wind turbine pile foundations are the core load-bearing structures of wind turbine generators. The quality of the spatial arrangement of the complex steel mesh inside directly determines the structural load-bearing safety, wind vibration fatigue resistance, and overall service life of the foundation. As the scale of wind power installations continues to expand into regions with complex geology and extreme climates, the requirements for maintaining stable load-bearing capacity and design reliability of foundation structures under different environmental conditions are becoming increasingly stringent. This presents higher technical challenges to the intelligent arrangement methods of steel mesh.

[0003] In current engineering practice, intelligent reinforcement methods based on neural networks have been introduced for the reinforcement mesh layout of wind turbine pile foundations, such as fully connected networks and convolutional neural networks. These methods, by mining and learning the potential mapping relationships in a large amount of historical engineering data, enable the automatic generation and prediction of key parameters for reinforcement layout (such as three-dimensional spatial coordinate data such as reinforcement spacing, diameter, and number of layers). This aims to significantly improve design efficiency and reduce the subjective experience bias of manual reinforcement layout, thus promoting the development of intelligent onshore wind turbine foundation design.

[0004] However, such neural network models are prone to deeply coupling with project-specific geological conditions, wind pressure loads, and seismic resistance levels during training. This results in highly discrete 3D layout data output by the trained reinforcement model when facing different wind fields or geological migration scenarios, making it impossible to maintain stable structural design standards. Even if the overall network fitting error is small, non-deterministic environmental parameters may still dominate the direction of network weight updates, severely interfering with the model's extraction of core structural features. This means that even minor environmental fluctuations can trigger significant changes in the output layout parameters. Summary of the Invention

[0005] To address the technical problem of existing technologies where overfitting of models to environmental data leads to drastic fluctuations in deployment parameters across different scenarios, posing a risk of structural collisions, this invention aims to provide a three-dimensional anti-collision deployment method for complex steel mesh reinforcement in wind turbine pile foundations. The specific technical solution adopted is as follows: Firstly, a method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations is provided, comprising: acquiring historical engineering data of the wind turbine pile foundation; the historical engineering data includes historical state data and sample three-dimensional layout data of the corresponding steel mesh layout scheme that meets the structural safety anti-collision requirements; analyzing the correlation between each attribute parameter in the historical state data and the sample three-dimensional layout data, determining the influence weight of each attribute parameter on the sample three-dimensional layout data, and determining the degree of fluctuation of each attribute parameter value according to the influence weight; dividing the historical engineering data into training data and validation data, and using the training data to test the preset... A baseline model is obtained through network training. For each network layer in the baseline model, a layer-deleting model is constructed by deleting the corresponding network layer. Validation data is input into both the baseline model and the layer-deleting model. Based on the fluctuation of each attribute parameter value, the output error of the baseline model, and the output error of the layer-deleting model, the fluctuation sensitivity of each network layer is determined. The corresponding regularization intensity is set according to the fluctuation sensitivity of each network layer for training until multiple network layers are trained, resulting in the trained target model. The target engineering data of the wind turbine pile foundation to be tested is input into the target model, and the three-dimensional layout data of the steel mesh of the wind turbine pile foundation to be tested is output.

[0006] Based on the above technical solution, in the complex steel mesh anti-collision three-dimensional layout method for wind turbine pile foundation provided by the present invention, historical engineering data containing steel mesh layout schemes that meet the structural safety anti-collision requirements are obtained as training targets. Combined with quantitative analysis of the fluctuation degree of attribute parameter values ​​and differential discrimination and regularization constraints on the fluctuation sensitivity of each layer of the neural network, the trained target model can effectively reduce the interference of external environmental fluctuations on the output of three-dimensional layout data. When facing cross-scenario applications such as different geological conditions and load conditions, the steel mesh three-dimensional layout data output by the model can still maintain a stable layout scheme that meets the structural safety anti-collision requirements.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the method for analyzing the correlation between each attribute parameter in the historical state data and the sample 3D layout data, determining the influence weight of each attribute parameter on the sample 3D layout data, and determining the degree of fluctuation of each attribute parameter's value based on the influence weight specifically includes: constructing a causal relationship graph based on each attribute parameter in the historical state data and the sample 3D layout data; determining the causal relationship paths between the attribute parameters and the sample 3D layout data, and among the attribute parameters themselves; determining the influence weight of each attribute parameter on the sample 3D layout data based on the path characteristics of each attribute parameter to the sample 3D layout data in the causal relationship path; and performing a reverse mapping on the influence weight of each attribute parameter to obtain the degree of fluctuation of each attribute parameter's value.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for determining the influence weight of each attribute parameter on the sample 3D layout data based on the path characteristics of each attribute parameter to the sample 3D layout data in the causal association path specifically includes: classifying the attribute parameters into direct association types (directly associated with the sample 3D layout data) and indirect association types (associated with the sample 3D layout data through other attribute parameters) based on the path characteristics of each attribute parameter to the sample 3D layout data; determining the influence weight corresponding to the attribute parameter of the direct association type based on the path length of the attribute parameter of the direct association type reaching the sample 3D layout data in the causal association path; determining the path influence weight of each indirect path based on the number of intermediate attribute parameters and the influence weight corresponding to the intermediate attribute parameters in at least one indirect path reaching the sample 3D layout data through the attribute parameter of the indirect association type in the causal association path; and determining the influence weight corresponding to the attribute parameter of the indirect association type based on the average of the path influence weights of at least one indirect path.

[0009] In conjunction with the first aspect above, in one possible implementation, the method for determining the path length from the attribute parameters of the directly associated type to the sample 3D layout data specifically includes: obtaining the normalized input parameter value of the attribute parameters of the directly associated type, and the normalized output parameter value of the output parameter corresponding to the attribute parameters of the directly associated type in the sample 3D layout data; mapping the normalized input parameter value and the normalized output parameter value to the same feature space respectively, to obtain the first feature point and the second feature point in the same feature space; and calculating the Euclidean distance between the first feature point and the second feature point as the path length.

[0010] In conjunction with the first aspect above, in one possible implementation, the method for constructing a causal relationship graph based on each attribute parameter in historical state data and sample 3D layout data specifically includes: using the attribute parameters and the output parameters contained in the sample 3D layout data as nodes in the causal relationship graph; testing the statistical correlation between nodes, and determining the directed relationship between statistically correlated node pairs based on the test results and preset causal direction determination rules; and determining the connecting edges between nodes in the causal relationship graph and the direction of the connecting edges based on the directed relationship.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the method of inputting validation data into the baseline model and the de-layered model respectively, and determining the fluctuation sensitivity of each network layer based on the fluctuation degree of each attribute parameter value, the output error of the baseline model, and the output error of the de-layered model, specifically includes: combining validation data in pairs to form sample pairs; weighting and summing the differences in the values ​​of each attribute parameter in the sample pairs according to the fluctuation degree of each attribute parameter value, to obtain the weighted difference value between the two validation samples in each sample pair; determining the output error of each validation sample under the baseline model based on the difference between the sample 3D layout data corresponding to the two validation samples in each sample pair and the prediction output of the baseline model; determining the output error of each validation sample under the de-layered model based on the difference between the sample 3D layout data corresponding to the two validation samples in each sample pair and the prediction output of the de-layered model; and determining the fluctuation sensitivity of the corresponding network layer based on the weighted difference value of multiple sample pairs, the output error of the baseline model corresponding to each validation sample, and the output error of the de-layered model.

[0012] In conjunction with the first aspect above, in one possible implementation, the method for determining the volatility sensitivity of a corresponding network layer based on the weighted difference values ​​of multiple sample pairs, the output error of the benchmark model corresponding to each validation sample, and the output error of the layer-removing model specifically includes: comparing the output errors of the two validation samples in each sample pair under the benchmark model to determine the first degree of closeness between the prediction errors of the two validation samples under the benchmark model, and combining the weighted difference values ​​to determine the first single-sample pair sensitivity of each sample pair under the benchmark model; comparing the output errors of the two validation samples in each sample pair under the layer-removing model to determine the second degree of closeness between the prediction errors of the two validation samples under the layer-removing model, and combining the weighted difference values ​​to determine the second single-sample pair sensitivity of each sample pair under the layer-removing model; determining the single-sample pair volatility sensitivity of each sample pair based on the difference between the first and second single-sample pair sensitivities of each sample pair; and determining the volatility sensitivity of the network layer based on the dispersion of the single-sample pair volatility sensitivities of multiple sample pairs.

[0013] In conjunction with the first aspect above, in one possible implementation, the regularization strength is achieved by randomly discarding neurons in the network layer with a corresponding regularization drop-off probability during training. The method of setting the corresponding regularization strength according to the fluctuation sensitivity of each network layer for training until multiple network layers are trained to obtain the trained target model specifically includes: mapping the fluctuation sensitivity of each network layer to the corresponding regularization drop-off probability; according to the arrangement order of the network layers from the output end to the input end in the preset neural network, sequentially constraining the parameter update process of the network layer to be constrained according to the regularization drop-off probability corresponding to each network layer to be constrained; fixing the weight parameters of the constrained network layers after training is completed until multiple network layers are constrained to obtain the trained target model.

[0014] In conjunction with the first aspect above, in one possible implementation, the method of constraining the parameter update process of the constrained network layer according to the regularized dropout probability corresponding to the constrained network layer specifically includes: based on the trained network model, performing layer deletion analysis on the constrained network layer to determine the current fluctuation sensitivity corresponding to the constrained network layer, and mapping the current fluctuation sensitivity to the current regularized dropout probability; constraining the parameter update process of the constrained network layer according to the current regularized dropout probability to obtain the updated network model, and using the updated network model as the network model for the next training operation.

[0015] In conjunction with the first aspect above, in one possible implementation, the method for obtaining historical engineering data of wind turbine pile foundations specifically includes: obtaining historical engineering records of multiple completed wind turbine pile foundations, extracting corresponding site environmental parameters, external load parameters, structural design parameters, and three-dimensional layout data of steel mesh that has passed construction acceptance; using the site environmental parameters, external load parameters, and structural design parameters as attribute parameters in the historical state data, and using the three-dimensional layout data of steel mesh that has passed construction acceptance as sample three-dimensional layout data.

[0016] In a second aspect, a three-dimensional anti-collision layout device for complex steel mesh for wind turbine pile foundations is provided, including a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it can perform the actions described in the first aspect or any possible implementation of the first aspect.

[0017] The present invention has the following beneficial effects: By acquiring historical engineering data containing steel mesh layout schemes that meet structural safety and collision prevention requirements as training targets, and combining quantitative analysis of the fluctuation of attribute parameter values ​​with differentiated discrimination and regularization constraints on the fluctuation sensitivity of each layer of the neural network, the trained target model can effectively reduce the interference of external environmental fluctuations on the output of three-dimensional layout data. When facing cross-scenario applications with different geological conditions and load conditions, the three-dimensional layout data of steel mesh output by the model can still maintain a stable layout scheme that meets the structural safety and collision prevention requirements. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations, as provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of the hardware structure of a three-dimensional anti-collision layout device for complex steel mesh for wind turbine pile foundations, provided as an embodiment of the present invention. Detailed Implementation

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a three-dimensional anti-collision layout method for complex steel mesh for wind turbine pile foundations provided by the present invention.

[0020] Please see Figure 1 This document illustrates a flowchart of a three-dimensional layout method for preventing collisions of complex steel mesh in wind turbine pile foundations, provided by an embodiment of the present invention. The method includes: S1. Obtain historical engineering data of wind turbine pile foundations.

[0021] Historical engineering data serves as the foundational data source for training the neural network. This includes historical state data and sample 3D layout data of the corresponding rebar mesh layout schemes that meet structural safety and collision prevention requirements. Historical state data characterizes the design conditions and external constraints of wind turbine pile foundations in various historical engineering projects, including site environmental parameters, external load parameters, and the foundation's own structural design parameters. This data will serve as the input features to the neural network model. Sample 3D layout data characterizes the actual rebar mesh spatial layout schemes that have passed construction acceptance, corresponding to the aforementioned historical state data. This data will serve as the target output for training the neural network model. The sample 3D layout data is taken from rebar mesh layout schemes that meet structural safety and collision prevention requirements. This means that the rebar mesh corresponding to this 3D layout data has been confirmed by structural safety testing after actual construction to meet load-bearing and collision prevention requirements, thus ensuring the safety and compliance of the data used as training labels, providing prior assurance for the model to learn correct layout patterns.

[0022] In some implementation methods, historical engineering records of multiple completed wind turbine pile foundations are obtained. Specifically, the original data of the wind turbine pile foundations recorded during the design and construction phases are extracted from the design archives, construction records, and completion acceptance documents of onshore wind turbine foundation projects. Each historical engineering record corresponds to a completed wind turbine pile foundation project, and the corresponding site environmental parameters, external load parameters, structural design parameters, and three-dimensional layout data of the steel reinforcement mesh that has passed construction acceptance are extracted.

[0023] Site environmental parameters are those significantly influenced by external natural conditions of the engineering site. These include, but are not limited to, site geological information and seismic design parameters. Site environmental parameters exhibit significant variability across different engineering projects, and changes in these parameters typically do not directly lead to adjustments in the seismic layout plan. Site geological information specifically includes parameters characterizing the site's geotechnical engineering conditions, such as foundation bearing capacity, soil corrosivity level, and groundwater level depth. Seismic design parameters specifically include parameters characterizing the seismic action conditions of the project area, such as seismic fortification intensity and basic design acceleration.

[0024] External load parameters are the external mechanical forces transmitted to the foundation during wind turbine operation, including but not limited to wind load parameters. These parameters vary with the climate conditions of the project area. Specifically, wind load parameters include basic wind speed, turbulence intensity, wind pressure coefficient, and other parameters characterizing the wind environment conditions of the project area.

[0025] Structural design parameters are the inherent property parameters of the foundation determined according to structural design codes, including but not limited to concrete strength grade parameters and steel reinforcement specification parameters. These parameters are usually uniformly specified by design standards and their values ​​are relatively stable across different projects. The concrete strength grade parameter characterizes the strength grade of the concrete material in the foundation structure; the steel reinforcement specification parameter characterizes the material properties of the steel reinforcement, such as its diameter, strength grade, and protective layer thickness.

[0026] The three-dimensional layout data of the steel mesh is a complete set of parameters characterizing the spatial structure of the steel cage, including the diameter of the steel bars, the spacing of the steel mesh in each layer, the number of steel bar layers, and the relative position coordinates of various types of steel bars in three-dimensional space. This type of data directly determines the stress performance of the steel mesh inside the foundation and the spatial avoidance relationship with other structural components.

[0027] Site environmental parameters, external load parameters, and structural design parameters are used as attribute parameters in the historical state data, while the 3D layout data of the steel mesh that has passed construction acceptance is used as sample 3D layout data. After aggregating the data from multiple historical project records, an engineering dataset consisting of multiple sets of samples is formed. Each set of samples contains an attribute parameter vector and a corresponding sample 3D layout data vector.

[0028] S2. Analyze the correlation between each attribute parameter in the historical state data and the sample 3D layout data, determine the influence weight of each attribute parameter on the sample 3D layout data, and determine the degree of fluctuation of each attribute parameter value based on the influence weight.

[0029] Assuming that the 3D layout data in the training samples are all taken from layout schemes that meet structural safety and collision avoidance requirements, the 3D layout output should remain stable under the constraints of the specifications. Therefore, if the value of a certain attribute parameter has a large dispersion among different engineering samples but the 3D layout output does not change accordingly, it indicates that the fluctuation of this attribute parameter does not affect the layout decision and is a parameter dominated by environmental factors. Conversely, if a small change in a certain attribute parameter corresponds to a significant difference in the 3D layout output, it indicates that this attribute parameter has a decisive role in the layout decision and is a key control parameter in structural design. The influence weight is used to quantify the contribution of each attribute parameter to this decisive role. The larger the value, the higher the contribution of the attribute parameter to the 3D layout output decision, and the attribute parameter is a key structural parameter that is less affected by environmental factors. The smaller the value, the lower the contribution of the attribute parameter to the 3D layout output decision, and the fluctuation of the attribute parameter mainly comes from changes in the external environment of the engineering site. The degree of value fluctuation is used to characterize the degree of numerical dispersion of each attribute parameter among different engineering samples. It is obtained by inverse mapping of the influence weight and is negatively correlated with the influence weight. That is, the higher the influence weight, the lower the degree of value fluctuation, and the lower the influence weight, the higher the degree of value fluctuation.

[0030] In some implementations, a causal relationship graph is constructed based on each attribute parameter in the historical state data and the sample 3D layout data.

[0031] Specifically, the attribute parameters and the output parameters contained in the 3D layout data of the samples are used as nodes in the causal relationship graph. The output parameters are the dimensional components contained in the 3D layout data of the samples, such as rebar spacing parameters, rebar diameter parameters, rebar layer quantity parameters, and spatial coordinate parameters of each layer of rebar mesh. Each output parameter corresponds to an independent node in the causal relationship graph. The statistical correlation between nodes is tested to determine whether there is a statistically significant association between any two nodes. The specific algorithm for the statistical correlation test can be determined according to the data characteristics. For example, Pearson correlation coefficient test is used for continuous numerical parameters, Spearman rank correlation coefficient test is used for discrete or ordinal parameters, and chi-square independence test is used for categorical variables. When performing the statistical correlation test, for each pair of nodes to be tested, the correlation metric and the corresponding significance probability value are calculated. When the significance probability value is less than the significance level threshold (set to 0.05 according to the industry practice of controlling the Type I error probability to no more than 5% in statistical hypothesis testing, to balance the sensitivity and specificity of causal relationship identification under limited sample conditions), it is determined that there is a statistical correlation between the node pair.

[0032] Based on the test results, for node pairs determined to have statistical correlation, a pre-defined causal direction determination rule is used to determine the directed association between these node pairs. This causal direction determination rule includes at least one of the following: a time-series-based rule, where if the parameter represented by the first node is determined before the parameter represented by the second node in the engineering timeline, the direction of the directed association is determined to be from the first node to the second node; a domain-knowledge-based rule, where the causal direction between parameters is pre-defined based on the parameter dependencies in the wind turbine pile foundation design specifications; or a conditional independence test rule, where, while controlling for other nodes, the conditional independence between two nodes is tested to eliminate spurious correlations and determine the causal direction. For example, for the "foundation bearing capacity" node and the "reinforcement spacing" node, since the foundation bearing capacity is determined by geological survey and takes precedence over the design of reinforcement spacing in time sequence, the direction is determined from "foundation bearing capacity" to "reinforcement spacing" according to the time sequence-based determination rule; for the "concrete strength grade" node and the "reinforcement diameter" node, the direction is determined from "concrete strength grade" to "reinforcement diameter" according to the design requirements for the matching of concrete and reinforcement strength in the structural design code.

[0033] The causal relationship graph is constructed by determining the connecting edges and their directions between nodes based on directed associations. Specifically, a directed edge is established between two nodes with a directed association, pointing from the cause node to the result node, thus completing the construction of the causal relationship graph. The purpose of constructing a causal relationship graph is to explicitly express the causal relationships between attribute parameters and between attribute parameters and output parameters in the form of a graph structure.

[0034] This study aims to identify causal paths between attribute parameters and sample 3D layout data, as well as among the attribute parameters themselves, within a causal relationship graph. A causal relationship path is a directed path in the graph that begins at a starting node, proceeds along directed edges, and ultimately reaches an ending node. Each path contains at least one directed edge. In the causal relationship graph, the starting point of a path is the node corresponding to a specific attribute parameter, and the ending point is either the node corresponding to an output parameter or another attribute parameter. The method for determining all causal relationship paths is as follows: traverse the causal relationship graph, starting from each attribute parameter node, and perform a depth-first search or breadth-first search along directed edges until a node with no outgoing edges is reached. Each complete path traversed during the search is considered a causal relationship path, and the sequence of nodes contained in each path and the directions of the connecting edges between the nodes are recorded.

[0035] Based on the path characteristics from each attribute parameter in the causal relationship path to the sample 3D layout data, the influence weight of each attribute parameter on the sample 3D layout data is determined.

[0036] Specifically, the path features include at least: whether there is a direct directed edge connection between the attribute parameter node and the output parameter node, and the path length between the attribute parameter node and the output parameter node.

[0037] Based on the path characteristics from each attribute parameter to the sample 3D layout data, attribute parameters are divided into direct association type and indirect association type. A direct association type means that the attribute parameter has at least one path in the causal relationship graph ending at the output parameter, and there is a direct directed edge connection between the attribute parameter node and the output parameter node in this path; that is, there are no other intermediate nodes between the attribute parameter node and the output parameter node in the path sequence. An indirect association type means that the attribute parameter has only one path in the causal relationship graph ending at another attribute parameter, or although there is a path ending at the output parameter, there is at least one intermediate attribute parameter node between the attribute parameter node and the output parameter node in this path.

[0038] Based on the path length of the attribute parameters of the directly related type to reach the sample 3D layout data in the causal relationship path, the influence weight of the attribute parameters of the directly related type is determined.

[0039] The process of determining the path length from the directly associated attribute parameters to the sample 3D layout data includes: obtaining the normalized input parameter values ​​of the directly associated attribute parameters, and the normalized output parameter values ​​of the output parameters corresponding to the directly associated attribute parameters in the sample 3D layout data. The purpose of normalization is to unify parameter values ​​of different dimensions to the same numerical scale to avoid distortion in Euclidean distance calculations due to dimensional differences. Normalization employs a preset normalization method, such as obtaining the maximum and minimum values ​​of the attribute parameter in all historical engineering samples, and using max-min linear normalization to obtain the normalized input parameter value of the attribute parameter in the current sample (the portion exceeding the maximum and minimum values ​​is truncated to avoid the influence of extreme values). The same normalization method is used for the output parameter, obtaining its maximum and minimum values ​​in all historical engineering samples, and calculating the normalized output parameter value in the current sample. The normalized input parameter values ​​and normalized output parameter values ​​are then mapped to the same feature space to obtain the first and second feature points in the same feature space. The same feature space refers to mapping the normalized values ​​of the attribute parameters and the normalized values ​​of the output parameters to a vector space of the same dimension through the same embedding function or encoder, so as to facilitate the calculation of the Euclidean distance between them. Since the attribute parameters and output parameters may differ in the dimension of the original data, independent mapping matrices matching their respective dimensions are used for projection. Specifically, the normalized input parameter values ​​are multiplied and projected using the input feature mapping matrix matching the dimension of the normalized input parameter values ​​to obtain the first feature point. At the same time, the normalized output parameter values ​​are multiplied and projected using the output feature mapping matrix matching the dimension of the normalized output parameter values ​​to obtain the second feature point. The input feature mapping matrix and the output feature mapping matrix are independent mapping matrices. The number of columns of each depends on the dimension of the input parameters and the dimension of the output parameters, respectively, and the number of rows of each is equal to the preset feature space dimension (for example, this dimension can be set to 64 to 256 dimensions according to the complexity of the input and output parameters, and the specific value is determined before model training through cross-validation), so that the first and second feature points obtained after projection have the same vector length. Calculate the Euclidean distance between the first feature point and the second feature point, and use this Euclidean distance as the path length to the sample 3D layout data through the attribute parameter of this directly associated type.

[0040] The influence weight of the attribute parameter of the directly related type is determined based on the path length. The smaller the path length, the closer the value of the attribute parameter is to the value of the corresponding output parameter in the feature space, meaning the direct influence of the attribute parameter on the output parameter is more significant, and therefore the influence weight is larger. Conversely, the larger the path length, the smaller the influence weight. For example, the mapping relationship between path length and influence weight is an inverse proportional function or a negative exponential function, ensuring that the influence weight has an upper limit of 1 when the path length is 0, and the influence weight tends to 0 when the path length tends to infinity.

[0041] For attribute parameters belonging to the indirect association type, they may be associated with the output parameter through multiple indirect paths in the causal relationship graph. Each indirect path contains at least one intermediate attribute parameter node. For example, if attribute parameter A is connected to the output parameter Y through intermediate attribute parameter B (path A→B→Y), then the intermediate attribute parameter is B, and the number of intermediate attribute parameters is 1. If attribute parameter A is connected to the output parameter Y through intermediate attribute parameters B and C (path A→B→C→Y), then the intermediate attribute parameters are B and C, and the number of intermediate attribute parameters is 2. Based on the number of intermediate attribute parameters and the corresponding influence weights of the intermediate attribute parameters in the at least one indirect path from the attribute parameter of the indirect association type to the sample 3D layout data, the path influence weight of each indirect path is determined. The more intermediate attribute parameters there are, the longer the causal transmission chain, and the more intermediate nodes the influence of the attribute parameter is attenuated during transmission, thus the smaller the path influence weight. The larger the influence weight of each intermediate attribute parameter, the higher the contribution of each intermediate node in the transmission chain to the output decision, thus the larger the path influence weight. For example, the path influence weight can be obtained by multiplying the influence weights of each intermediate attribute parameter and then dividing by the number of intermediate attribute parameters, so as to simultaneously reflect the length decay of the transmission chain and the influence transmission of intermediate nodes. After obtaining the path influence weights of all indirect paths, the mean of the path influence weights of all indirect paths is calculated as the influence weight corresponding to the attribute parameter of the indirect association type.

[0042] It should be noted that in a causal relationship graph, an attribute parameter may belong to two types simultaneously. That is, the attribute parameter may have a direct path to an output parameter, and it may also have a path indirectly connected to the output parameter through other attribute parameters. In this case, the attribute parameter has both direct and indirect influence weights. The direct and indirect influence weights are then combined to obtain the comprehensive influence weight of the attribute parameter on the sample's 3D layout data. An exemplary method of combining these weights is to add the direct and indirect influence weights and take the average, or to perform a weighted sum based on preset weight coefficients. The direct influence weight is calculated based on the direct directed edge connection between the attribute parameter node and the output parameter node, reflecting the direct numerical correspondence between the two parameters without any intermediate links. This provides more direct and certain causal evidence with less information loss and uncertainty from intermediate links. The indirect influence weight is calculated based on the transmission path through one or more intermediate attribute parameters, reflecting the transmission mechanism by which the attribute parameter indirectly affects the output parameter by influencing intermediate attribute parameters. This provides relatively indirect causal evidence, and as the transmission path lengthens, the uncertainty of intermediate links accumulates, information attenuation intensifies, and the confidence of the causal inference decreases accordingly. Based on the above differences, the weight coefficient of the component that directly affects the weight is higher than that of the component that indirectly affects the weight. For example, the weight coefficient of the component that directly affects the weight can be set to 0.7, while the weight coefficient of the component that indirectly affects the weight can be set to 0.3.

[0043] For example, in a causal relationship graph, there exists a directed path: Concrete strength grade (attribute parameter) → Rebar diameter (output parameter). This is because, in structural design codes, the concrete strength grade directly affects the anchorage length and minimum reinforcement ratio requirements of the reinforcing bars, thus directly determining the diameter of the reinforcing bars used. In this path, there is a direct directed edge connection between the concrete strength grade and the reinforcing bar diameter, with no other attribute parameter nodes in between. Therefore, the concrete strength grade belongs to the direct correlation type, directly affecting the weight calculated based on the Euclidean distance between the normalized parameter value of the concrete strength grade and the normalized output parameter value of the reinforcing bar diameter. Simultaneously, another directed path exists in the causal relationship graph: Concrete strength grade (attribute parameter) → Effective cross-sectional height of foundation cap (intermediate attribute parameter) → Spatial coordinates of reinforcing mesh layers (output parameter). As the concrete strength grade increases, the shear capacity of the foundation cap increases accordingly. Under the same external load conditions, the effective cross-sectional height of the foundation cap can be appropriately reduced. The change in the effective cross-sectional height of the foundation cap directly determines the vertical layering arrangement of the internal steel reinforcement mesh (e.g., two or three layers), thus altering the coordinate positions of each layer of the reinforcement mesh in three-dimensional space. In this path, there is no direct directed edge connection between the concrete strength grade and the spatial coordinates of the steel reinforcement mesh layers; instead, the association is achieved through the effective cross-sectional height of the foundation cap, with only one intermediate attribute parameter. Therefore, the concrete strength grade also belongs to the indirect association type, and the indirect influence weight is determined by the number of intermediate attribute parameters in this indirect path and the influence weight corresponding to each intermediate attribute parameter.

[0044] The influence weights of each attribute parameter are inversely mapped to obtain the degree of fluctuation in each attribute parameter's value. Inverse mapping refers to reversing the numerical direction of the influence weights, so that originally larger influence weights are mapped to smaller fluctuation values, thus allowing the degree of fluctuation to directly characterize the extent to which the attribute parameter is affected by environmental factors. For example, the difference between 1 and the influence weights can be taken as the degree of fluctuation. A higher degree of fluctuation indicates a greater range of value variation for the attribute parameter across different engineering projects, and this variation does not have a decisive impact on the 3D layout output; therefore, this attribute parameter is a parameter governed by environmental factors. Conversely, a lower degree of fluctuation indicates that the attribute parameter's value tends to be stable across different engineering projects, has a decisive impact on the 3D layout output, and is a key control parameter in structural design.

[0045] S3. Divide historical engineering data into training data and validation data. Use the training data to train the preset neural network to obtain the benchmark model. For each network layer in the benchmark model, construct a layer-deleting model to delete the corresponding network layer.

[0046] Training data is used to update the weight parameters of the neural network, enabling the model to learn the mapping relationship from historical state data to sample 3D layout data, thereby obtaining a baseline model capable of reasonably predicting the 3D layout scheme of the rebar mesh. Validation data does not participate in weight updates; its purpose is to serve as an independent evaluation dataset for analyzing the sensitivity of each network layer to data fluctuation characteristics. The baseline model refers to the complete neural network model obtained after normal training using the training data based on the pre-defined neural network structure, serving as the foundation and comparison benchmark for all subsequent layer-removed models. A layer-removed model is a neural network model obtained by removing a specific network layer from the baseline model. Compared to the baseline model, its structure lacks the deleted network layer and connection parameters. For each network layer to be analyzed in the baseline model, a corresponding layer-removed model is constructed.

[0047] In some implementations, historical engineering data comprises multiple sets of samples, each set including a historical state data vector and a corresponding 3D layout data vector. Data partitioning uses a preset ratio, randomly dividing all samples into training and validation data according to this ratio. The rule for setting the partition ratio is based on the sample size requirements for subsequent neural network training and the statistical significance requirements of the validation data. For example, the partition ratio is 8:2. This ratio is based on the rule of thumb in machine learning (i.e., the training set should account for no less than 70% to ensure sufficient model learning, and the validation set should account for no less than 10% to ensure the reliability of statistical evaluation). Random sampling is used to divide 80% of all samples into training data and 20% into validation data, ensuring sufficient training samples while making the validation data statistically representative.

[0048] The preset neural network is a feedforward neural network structure with multiple network layers. Each network layer contains multiple neurons, and neurons in adjacent network layers are fully or partially connected through weight parameters. For example, the preset neural network includes an input layer, three hidden layers, and an output layer. The number of neurons in the input layer equals the number of attribute parameters in the historical state data (e.g., 15). Each hidden layer has the same number of neurons (e.g., each hidden layer contains 64 neurons), and the number of neurons in the output layer equals the number of output parameters contained in the sample 3D layout data (e.g., 8). Each hidden layer uses a non-linear activation function to introduce non-linear transformation capability; for example, the ReLU activation function (corrected linear unit, i.e., outputting 0 for negative inputs and the input itself for positive inputs) is used. The output layer uses a linear activation function to output continuous values.

[0049] Before training begins, the weight and bias parameters of each layer in the pre-defined neural network are initialized using a weight initialization algorithm, exemplarily Xavier uniform distribution initialization (a weight initialization method that determines the random initialization range based on the number of input and output neurons). The specific process of training the pre-defined neural network using training data is as follows: historical state data from the training data is input into the pre-defined neural network as input features, and the predicted 3D layout data is obtained after forward propagation. The difference between the predicted 3D layout data and the corresponding sample 3D layout data in the training data is calculated as the training loss. This training loss is calculated using a loss function, exemplarily the mean squared error loss function, which is the average of the squares of the differences between the predicted and true values. The gradient of the weight parameters of each network layer is calculated based on the training loss. An optimization algorithm is used to update the weight and bias parameters of each network layer along the gradient descent direction to gradually reduce the training loss. An example optimization algorithm is the Adam optimization algorithm (an adaptive moment estimation stochastic gradient descent optimization method).

[0050] The training process repeats the forward propagation, loss calculation, backpropagation, and parameter update steps described above until the training stopping condition is met. The training stopping condition includes one or a combination of the following: the training loss decreases below a preset loss threshold (exemplarily, the loss threshold is set to 0.001); the prediction error on the validation set no longer decreases within a preset number of consecutive training epochs (exemplarily 50 epochs) (i.e., an early stopping strategy); or the number of training epochs reaches a preset maximum number of training epochs (exemplarily 1000 epochs). After training is complete, the weight parameters and bias parameters of each network layer obtained from the final training are fixed to form a baseline model.

[0051] Furthermore, for each network layer in the baseline model, a layer-deletion model is constructed to remove the corresponding network layer. The baseline model contains multiple network layers, labeled as layer 1, layer 2, ... in order from input to output, where layer 1 is the input layer, the last layer is the output layer, and the intermediate layers are hidden layers. For each network layer to be analyzed in the baseline model (i.e., each hidden layer and output layer; the input layer, as a data input interface, does not participate in layer-deletion analysis, and the output layer, being at the end of the network, does not have a subsequent receiving end for cross-layer reconnection and therefore does not participate in layer-deletion or cross-layer reconnection analysis), a corresponding layer-deletion model is constructed. The specific method for constructing the layer-deletion model is as follows: the complete structure and all parameters of the baseline model are copied, the specified network layer is deleted from the copied model structure, and the input and output connections of the deleted network layer are re-established directly to ensure that the data flow can be directly transmitted from the layer above the deleted network layer to the layer below the deleted network layer, thereby maintaining the integrity of the model's data flow. Since the baseline model has the same number of neurons in each hidden layer, the number of channels in the output tensor feature of the deleted network layer is consistent with the input dimension of its receiving network. Therefore, direct concatenation and transfer of data between upper and lower layers can be achieved without introducing adaptation parameters, effectively avoiding program execution interruptions caused by dimension mismatch. For example, when deleting the kth intermediate hidden layer of the baseline model, the output of the (k-1)th layer is directly used as the input of the (k+1)th layer without any nonlinear transformation, thus eliminating the contribution of this network layer to the model processing while maintaining the continuity of the data flow. When constructing the layer-deleting model, the original parameters of the deleted network layer are discarded, and the re-established connections do not introduce new trainable parameters. Following this method, a corresponding layer-deleting model is constructed for each network layer to be analyzed.

[0052] S4. Input the validation data into the baseline model and the de-layered model respectively. Determine the fluctuation sensitivity of each network layer based on the fluctuation of each attribute parameter value, the output error of the baseline model and the output error of the de-layered model.

[0053] By inputting the same validation sample into both the baseline model and the de-layered model, and comparing the prediction error differences between the output 3D layout data and the actual sample 3D layout data, the functional characteristics of the deleted network layer in the model processing can be determined. The degree of value fluctuation characterizes the numerical dispersion of each attribute parameter across different engineering samples. When the two validation samples show significant differences in the values ​​of attribute parameters with large value fluctuations, it indicates that the environmental conditions of the two samples are significantly different. Output error refers to the degree of difference between the model's predicted output and the sample 3D layout data, reflecting the model's prediction accuracy for the current validation sample. After inputting the validation sample into both the baseline model and the de-layered model, the similarity of the prediction errors of the same model for the two validation samples is compared. If the prediction errors are similar, it indicates that the model can maintain consistent prediction accuracy on two samples with different environmental conditions, demonstrating that the model is not affected by environmental differences between samples.

[0054] When the baseline model's prediction errors are similar on two samples with significant environmental differences, while the deleted layer model shows a greater difference in prediction errors on the same two samples, it indicates that the deleted network layer in the baseline model originally served to eliminate the influence of environmental differences. That is, the network layer is not sensitive to environmental fluctuations and can actively suppress the interference caused by environmental differences on the output, and its fluctuation sensitivity is low. Conversely, when the baseline model's prediction errors differ significantly on two samples with significant environmental differences, while the deleted layer model shows similar prediction errors on these two samples, it indicates that the deleted network layer in the baseline model originally served to amplify the influence of environmental differences. That is, the network layer is highly sensitive to environmental fluctuations and may cause the model to output different deployment results under different environmental conditions, and its fluctuation sensitivity is high.

[0055] In some implementations, validation data is paired to form sample pairs. The differences in the values ​​of each attribute parameter in each sample pair are weighted and summed based on the degree of fluctuation of each attribute parameter, resulting in a weighted difference value between the two validation samples in each sample pair. Specifically, the validation data contains multiple validation samples. Each validation sample contains a historical state data vector (composed of the values ​​of each attribute parameter) and a corresponding three-dimensional sample layout data vector. The validation samples in the validation data are paired, with each pair forming a sample pair. For any sample pair, the values ​​of the two validation samples on the same attribute parameter are obtained, and the absolute value of their difference is calculated as the difference in the attribute parameter value. Then, using the degree of fluctuation of the attribute parameter value as a weighting coefficient, the differences in the attribute parameter values ​​are weighted and summed to obtain a weighted difference value, which reflects the portion of the overall difference between the two validation samples in the sample pair across all attribute parameters contributed by the highly volatile attribute parameter. If two validation samples have significant differences in attribute parameters with high volatility, the weighted difference value will be large, indicating that the differences between the two samples mainly come from environmental volatility attribute parameters rather than structural key attribute parameters; if two validation samples have small differences in attribute parameters with high volatility, the weighted difference value will be small, indicating that the environmental conditions of the two samples are relatively similar.

[0056] Based on the differences between the 3D layout data of the two validation samples in each sample pair and the predicted output of the benchmark model, the output error of each validation sample under the benchmark model is determined. Specifically, for each validation sample in the validation data, the historical state data vector is input into the benchmark model, and the predicted 3D layout data of the validation sample under the benchmark model is obtained after forward propagation calculation. The corresponding 3D layout data of the validation sample (i.e., the 3D layout data of the steel mesh that has passed construction acceptance, as the true target value of the prediction) is obtained, and the difference between the predicted 3D layout data and the sample 3D layout data is calculated. The difference is calculated using a preset error metric function. For example, Euclidean distance or mean square error is used as the error value. The error value reflects the accuracy of the benchmark model's prediction for the validation sample. The smaller the error value, the closer the predicted output of the benchmark model is to the actual layout scheme.

[0057] Based on the differences between the 3D layout data of the two validation samples in each sample pair and the predicted output of the layer-removed model, the output error of each validation sample under the layer-removed model is determined. Specifically, for each validation sample in the validation data, the historical state data vector is input into the layer-removed model, and the predicted 3D layout data of the validation sample under the layer-removed model is obtained after forward propagation calculation. The 3D layout data of the corresponding validation sample is obtained, and the difference between the predicted 3D layout data and the sample 3D layout data is calculated using the same preset error metric function as the baseline model. The difference metric is consistent with the metric method used to calculate the output error under the baseline model. For each layer-removed model corresponding to the network layer to be analyzed, each validation sample in the validation data obtains a corresponding output error value, which reflects the prediction accuracy of the model for the validation sample after removing the network layer.

[0058] The fluctuation sensitivity of the corresponding network layer is determined based on the weighted difference values ​​of multiple sample pairs, the output error of the baseline model corresponding to each validation sample, and the output error of the deleting model.

[0059] Specifically, by comparing the output errors of the two validation samples in each sample pair under the baseline model, the first degree of similarity between the prediction errors of the two validation samples under the baseline model is determined. This can be achieved by taking the reciprocal of the absolute value of the difference between the output errors of the two validation samples under the baseline model (to avoid the denominator being zero, a very small positive value, such as 10, can be added to the denominator when calculating the reciprocal). -6 The first degree of closeness is represented by a weighted difference value. The smaller the absolute value of the difference, the closer the prediction errors of the two validation samples under the baseline model are, meaning the baseline model's prediction accuracy for the two samples is consistent. The sensitivity of each sample to the first single-sample pair under the baseline model is determined by combining the weighted difference value with the first degree of closeness. Specifically, the weighted difference value is multiplied by the first degree of closeness to obtain the sensitivity of that sample to the first single-sample pair under the baseline model. The larger the weighted difference value (i.e., the greater the environmental difference between the two samples), the closer the prediction errors of the baseline model for the two samples are (i.e., the greater the first degree of closeness). Therefore, the baseline model can make consistent and accurate predictions for the two samples regardless of environmental differences, and thus the sensitivity of the first single-sample pair is greater.

[0060] Similarly, by comparing the output errors of the two validation samples in each sample pair under the layer-removing model, the second degree of similarity between the prediction errors of the two validation samples under the layer-removing model is determined. Specifically, this can be achieved by taking the reciprocal of the absolute value of the difference between the output errors of the two validation samples under the layer-removing model (to avoid the denominator being zero, a very small positive value, such as 10, can be added to the denominator when calculating the reciprocal). -6The difference represents the second degree of closeness. The smaller the absolute value of the difference, the closer the prediction errors of the two validation samples under the layer-removed model are, meaning that the prediction accuracy of the layer-removed model is consistent for the two samples. The second single-sample pair sensitivity of each sample under the layer-removed model is determined by combining the weighted difference value with the second degree of closeness. Specifically, the weighted difference value is multiplied by the second degree of closeness to obtain the second single-sample pair sensitivity of that sample under the layer-removed model. The physical meaning of the second single-sample pair sensitivity is similar to that of the first single-sample pair sensitivity, but it reflects the performance characteristics of the layer-removed model, that is, whether the model can still maintain consistent prediction accuracy on two samples with significantly different environments after removing the network layers.

[0061] The single-sample pair volatility sensitivity of each sample pair is determined based on the difference between the sensitivity of the first and second single-sample pairs for each sample pair. Specifically, the difference between the sensitivity of the second and first single-sample pairs is calculated as the single-sample pair volatility sensitivity of that sample pair under the network layer to be analyzed. If the sensitivity of the second single-sample pair is greater than that of the first single-sample pair after removing the network layer, it indicates that the model can make more consistent and accurate predictions for two samples with large environmental differences after removing the network layer. That is, the original function of the network layer was to amplify the impact of environmental differences on the output, and therefore the network layer has a high volatility sensitivity. If the sensitivity of the second single-sample pair is less than that of the first single-sample pair after removing the network layer, it indicates that the environmental adaptability of the model decreases after removing the network layer. That is, the original function of the network layer was to suppress the impact of environmental differences on the output, and therefore the network layer has a low volatility sensitivity.

[0062] The volatility sensitivity of a network layer is determined by the dispersion of its single-sample-pair volatility sensitivity across multiple sample pairs. Specifically, the dispersion can be represented by the standard deviation. A larger standard deviation of the volatility sensitivity across all sample pairs indicates a greater difference in volatility sensitivity across sample pairs with varying environmental differences, thus indicating higher volatility sensitivity for the network layer. Conversely, a smaller standard deviation indicates consistent performance across different sample pairs, thus indicating lower volatility sensitivity for the network layer.

[0063] Based on this, the standard deviation of a single sample's sensitivity to volatility can be directly used as the volatility sensitivity. Furthermore, when the weighted difference values ​​between validation sample pairs have significant dispersion, even if the network layer itself is not highly sensitive to volatility features, the standard deviation of a single sample's sensitivity to volatility may be amplified due to the dispersed distribution of environmental differences between sample pairs, thus overestimating the network layer's volatility sensitivity. Conversely, when the weighted difference values ​​between sample pairs are concentrated, even if the network layer itself is highly sensitive to volatility features, the standard deviation of a single sample's sensitivity to volatility may be compressed into a smaller range, thus underestimating the network layer's volatility sensitivity. Therefore, the standard deviation of the weighted difference values ​​can be introduced as a normalization factor. Dividing the standard deviation of a single sample's sensitivity to volatility by the standard deviation of the weighted difference values ​​yields the normalized volatility sensitivity dispersion. This ratio eliminates the direct influence of the dispersion of the sample pair's own environmental differences on the dispersion of a single sample's sensitivity to volatility, ensuring that the comparison of volatility sensitivity between different network layers is no longer affected by the distribution characteristics of the validation data itself. Meanwhile, the normalized ratio reflects the strength of the network layer's volatility sensitivity relative to the dispersion of samples in response to environmental differences. A larger ratio indicates that the network layer's volatility sensitivity fluctuates more drastically with changes in the sample's environment, meaning the network layer has a higher sensitivity to environmental fluctuations. The normalized ratio is then multiplied by the standard deviation of the single-sample volatility sensitivity, combining the absolute dispersion of the network layer's volatility sensitivity with the strength of its response to environmental differences. Multiplying these two values ​​yields the final volatility sensitivity of the network layer; a larger value indicates higher volatility sensitivity.

[0064] S5. Set the corresponding regularization intensity according to the fluctuation sensitivity of each network layer and train until multiple network layers are trained to obtain the trained target model.

[0065] By applying strong regularization constraints to network layers with high volatility sensitivity during training, the excessive extraction and transmission of environmental volatility features is suppressed, while weaker regularization constraints are applied to network layers with low volatility sensitivity to retain the ability to extract structurally stable features. This improves the overall generalization ability of the model to output stable 3D layout data across environmental scenarios.

[0066] In some implementations, regularization strength is used to characterize the constraint on the network layer parameter update process. This can be achieved by randomly dropping neurons in the network layer during training with a corresponding regularized dropout probability. The volatility sensitivity of each network layer is mapped to a corresponding regularized dropout probability. The regularized dropout probability, as a probability value, must be within the interval [0, 1]. Therefore, a predefined mapping function is used to convert the volatility sensitivity into the corresponding dropout probability. The mapping function uses a monotonically increasing non-linear function to ensure a positive correlation between volatility sensitivity and dropout probability; that is, the higher the volatility sensitivity, the higher the dropout probability. For example, a sigmoid-type mapping function is used:

[0067] In the formula, Let be the regularized dropout probability of the k-th network layer. The fluctuation sensitivity of the k-th network layer, The scaling factor is a preset value (1.0 for example; when the fluctuation sensitivity distribution of all network layers is relatively concentrated, α can be increased to enhance the discriminative power; when the distribution is relatively dispersed, α can be decreased to avoid probability oversaturation). A preset offset coefficient (exemplarily taken as the mean of the volatility sensitivity of all network layers) is used to adjust the center point of the mapping function to the average level of the volatility sensitivity of all network layers, so that network layers with volatility sensitivity below the mean receive a drop probability below 0.5, and network layers with volatility sensitivity above the mean receive a drop probability above 0.5. Through the above mapping method, each network layer obtains a regularized drop probability corresponding to its volatility sensitivity.

[0068] Simultaneously, following the order of network layers from output to input in the pre-defined neural network, each network layer is sequentially designated as a layer to be constrained. The input layer serves as the data input interface, receiving raw input data and passing it to the first hidden layer. It does not involve weight connections or nonlinear transformations between neurons and therefore does not participate in regularization constraint processing. The parameter update process of the constrained network layer is constrained according to the regularization dropout probability corresponding to that layer. After training, the weight parameters of the constrained network layers are fixed until multiple network layers are constrained, resulting in the trained target model.

[0069] The parameter update process includes: based on the trained network model, performing layer deletion analysis on the network layers to be constrained, determining the current volatility sensitivity of the corresponding network layers, and mapping the current volatility sensitivity to the current regularization drop probability. Specifically, when processing the network layers to be constrained, the network layers closer to the output end have been trained and fixed, and the parameters of these network layers have converged and no longer change. The network model that has completed the training of some network layers at this time is called the current network model. Based on the current network model, a current layer-deleted model is constructed after deleting network layers from the current network model. Each validation sample in the validation data is input into the current network model and the current layer-deleted model respectively, and the output error of each validation sample under the current network model and the current layer-deleted model is calculated. Combining the volatility of the values ​​of each attribute parameter, the weighted difference value and output similarity of each sample pair are calculated to determine the current volatility sensitivity of the network layer under the current parameter state. Then, using the same mapping function as described above, the current volatility sensitivity is mapped to the current regularization drop probability. By using the above method, we ensure that the discard probability is recalculated based on the parameter state of the current network model, reflecting the actual fluctuation sensitivity characteristics of the network layer in the current training phase, rather than the historical value calculated based on the benchmark model.

[0070] The parameter update process of the network layer to be constrained is constrained according to the current regularization dropout probability, resulting in an updated network model, which is then used as the network model for the next training operation. Specifically, based on the current network model, neurons of the network layer to be constrained are randomly dropped according to the current regularization dropout probability for forward and backward propagation. The weight and bias parameters of the network layer are updated according to the optimization algorithm, while the parameters of other network layers (including fixed network layers near the output and unprocessed network layers near the input) are not updated or are temporarily not updated. After a preset number of training rounds, the parameters of the network layer to be constrained tend to converge. At this point, the parameters of the network layer are fixed, resulting in an updated network model. This updated network model is used as the network model for the next training operation. That is, when processing the next network layer to be constrained, the above steps of re-determining the current fluctuation sensitivity, mapping the current regularization dropout probability, and constraining parameter updates are continued based on the updated network model. Through the above iterative process, the differentiated regularization training of each network layer is completed layer by layer from the output layer to the input layer, finally obtaining a target model in which all network layers have been trained and the parameters are fixed. In the above implementation, based on the trained network model, a layer deletion analysis is performed on the network layer to be constrained to determine the current fluctuation sensitivity of the network layer to be constrained. The current fluctuation sensitivity is then mapped to the current regularization drop probability. The parameter update process of the network layer to be constrained is constrained according to the current regularization drop probability to obtain the updated network model. The updated network model is then used as the network model for the next training operation. This ensures that the regularization strength of each network layer is set based on the fluctuation sensitivity under the actual training state, avoiding the problem of mismatch between the regularization strength and the actual state of the network layer caused by using the static analysis results of the baseline model before training, and improving the effectiveness of differentiated regularization.

[0071] S6. Input the target engineering data of the pile foundation of the wind turbine to be tested into the target model, and output the three-dimensional layout data of the steel mesh of the pile foundation of the wind turbine to be tested.

[0072] The target engineering data refers to the design conditions and external constraint parameters of the wind turbine pile foundation under test in the current engineering scenario. The parameter types and formats are consistent with historical data, including site geological information, wind load parameters, seismic design parameters, concrete strength grade parameters, and steel reinforcement specifications. However, unlike historical engineering data which originates from measured values ​​of completed projects, the target engineering data originates from the design conditions and survey results of the current project under test. The target model refers to a neural network model obtained after layer-by-layer differential regularization training. The network structure is consistent with the baseline model, but the weight and bias parameters of each network layer have been updated during training to output stable 3D layout data independent of environmental fluctuations. After inputting the target engineering data into the target model, the forward propagation calculation of the target model outputs the 3D layout data of the steel reinforcement mesh for the wind turbine pile foundation under test. This output 3D layout data is the final steel reinforcement mesh spatial layout scheme provided by this solution that meets the structural safety and anti-collision requirements. It includes a complete set of parameters such as the diameter of the steel reinforcement, the spacing between each layer of steel reinforcement mesh, the number of steel reinforcement layers, and the relative position coordinates of various types of steel reinforcement in 3D space. Because the target model weakens the dependence of each network layer on environmental fluctuations through differential regularization training during the training process, the output results can remain stable under different environmental conditions and will not change drastically due to fluctuations in the values ​​of environmental-related attribute parameters (such as foundation bearing capacity, wind load, etc.) in the target engineering data. This ensures the consistency and reliability of the output layout scheme in cross-scenario applications and avoids structural safety risks such as steel bar collisions caused by excessive differences in layout parameters due to changes in environmental data.

[0073] Based on the above technical solution, by acquiring historical engineering data containing steel mesh layout schemes that meet structural safety and anti-collision requirements as training targets, and combining quantitative analysis of the fluctuation degree of attribute parameter values ​​with differentiated discrimination and regularization constraints on the fluctuation sensitivity of each layer of the neural network, the trained target model can effectively reduce the interference of external environmental fluctuations on the output of three-dimensional layout data. When facing cross-scenario applications with different geological conditions and load conditions, the three-dimensional layout data of steel mesh output by the model can still maintain a stable layout scheme that meets the structural safety and anti-collision requirements.

[0074] This invention also provides a schematic diagram of the hardware structure of a three-dimensional anti-collision layout device for complex steel mesh in wind turbine pile foundations, see [link / reference]. Figure 2 The complex steel mesh anti-collision three-dimensional layout device 200 for wind turbine pile foundation includes a processor 201, and optionally, a memory 202 connected to the processor 201.

Claims

1. A three-dimensional method for preventing collisions of complex steel mesh in wind turbine pile foundations, characterized in that, include: Acquire historical engineering data of wind turbine pile foundations; the historical engineering data includes historical status data and three-dimensional layout data of samples that meet the structural safety and anti-collision requirements in the corresponding steel mesh layout scheme; Analyze the correlation between each attribute parameter in the historical state data and the sample 3D layout data, determine the influence weight of each attribute parameter on the sample 3D layout data, and determine the degree of value fluctuation of each attribute parameter based on the influence weight. The historical engineering data is divided into training data and validation data. The training data is used to train a preset neural network to obtain a benchmark model. The benchmark model contains multiple network layers, which are, in order from the input side to the output side, the input layer, the hidden layer, and the output layer. The baseline model has the same number of neurons in each hidden layer; The network layers in the baseline model are subjected to layer deletion and cross-layer reconnection analysis, while the input and output layers are not involved in the layer deletion and cross-layer reconnection analysis. The analysis of layer deletion and cross-layer reconnection includes: for each network layer in the baseline model, constructing a layer deletion model that deletes the corresponding network layer; The verification data is input into the baseline model and the de-layered model respectively. Based on the fluctuation degree of each attribute parameter value, the output error of the baseline model and the output error of the de-layered model, the fluctuation sensitivity of each network layer is determined. The training process is repeated until multiple network layers are trained, and the target model is obtained by setting the corresponding regularization intensity according to the fluctuation sensitivity of each network layer. Input the target engineering data of the pile foundation of the wind turbine to be tested into the target model, and output the three-dimensional layout data of the steel mesh of the pile foundation of the wind turbine to be tested.

2. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 1, characterized in that, Analyze the correlation between each attribute parameter in the historical state data and the sample 3D layout data, determine the influence weight of each attribute parameter on the sample 3D layout data, and determine the degree of fluctuation of each attribute parameter value based on the influence weight, including: A causal relationship diagram is constructed based on each attribute parameter in the historical state data and the sample 3D layout data, and the causal relationship paths between the attribute parameters and the sample 3D layout data, as well as between the attribute parameters themselves, are determined in the causal relationship diagram. Based on the path characteristics from each attribute parameter in the causal relationship path to the sample 3D layout data, determine the influence weight of each attribute parameter on the sample 3D layout data; By inversely mapping the influence weight of each attribute parameter, the degree of fluctuation in the value of each attribute parameter can be obtained.

3. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 2, characterized in that, Based on the path characteristics from each attribute parameter in the causal relationship path to the sample 3D layout data, the influence weight of each attribute parameter on the sample 3D layout data is determined, including: Based on the path characteristics of each attribute parameter to the sample 3D layout data, the attribute parameters are divided into direct association type, which is directly associated with the sample 3D layout data, and indirect association type, which is associated with the sample 3D layout data through other attribute parameters. Based on the path length of the attribute parameters of the directly associated type to the sample 3D layout data in the causal association path, the influence weight corresponding to the attribute parameters of the directly associated type is determined. Based on the number of intermediate attribute parameters and the influence weights corresponding to the intermediate attribute parameters in at least one indirect path leading to the sample 3D layout data through the attribute parameters of the indirect association type in the causal association path, the path influence weight of each indirect path is determined, and the influence weight corresponding to the attribute parameter of the indirect association type is determined based on the mean of the path influence weights of the at least one indirect path.

4. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 3, characterized in that, Determining the path length from the attribute parameters of the directly associated type to the sample 3D layout data includes: Obtain the normalized input parameter values ​​of the attribute parameters of the directly associated type, and the normalized output parameter values ​​of the output parameters corresponding to the attribute parameters of the directly associated type in the sample 3D layout data; The normalized input parameter value and the normalized output parameter value are mapped to the same feature space to obtain the first feature point and the second feature point in the same feature space. The Euclidean distance between the first feature point and the second feature point is calculated and used as the path length.

5. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 2, characterized in that, A causal relationship diagram is constructed based on each attribute parameter in the historical state data and the sample 3D layout data, including: The attribute parameters and the output parameters contained in the sample 3D layout data are used as nodes in the causal relationship graph; The statistical correlation between nodes is tested, and the directed association between statistically correlated node pairs is determined based on the test results and the pre-set causal direction determination rules. The connection edges between nodes in the causal relationship graph and their directions are determined based on the directed association relationship.

6. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 1, characterized in that, The validation data is input into the baseline model and the de-layered model, respectively. Based on the fluctuation degree of each attribute parameter value, the output error of the baseline model, and the output error of the de-layered model, the fluctuation sensitivity of each network layer is determined, including: The verification data are combined in pairs to form sample pairs. The differences in the values ​​of each attribute parameter in the sample pair are weighted and summed according to the degree of fluctuation of the value corresponding to each attribute parameter, so as to obtain the weighted difference value between the two verification samples in each sample pair. Based on the difference between the sample 3D layout data corresponding to the two verification samples in each sample pair and the prediction output of the benchmark model, the output error of each verification sample under the benchmark model is determined. Based on the difference between the sample 3D layout data corresponding to the two validation samples in each sample pair and the prediction output of the layer-removing model, the output error of each validation sample under the layer-removing model is determined; The fluctuation sensitivity of the corresponding network layer is determined based on the weighted difference values ​​of multiple sample pairs, the output error of the baseline model corresponding to each validation sample, and the output error of the deleting model.

7. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 6, characterized in that, The fluctuation sensitivity of the corresponding network layer is determined based on the weighted difference values ​​of multiple sample pairs, the output error of the baseline model corresponding to each validation sample, and the output error of the de-layer model, including: By comparing the output errors of the two validation samples in each sample pair under the baseline model, the first degree of closeness between the prediction errors of the two validation samples under the baseline model is determined, and the first single-sample pair sensitivity of each sample pair under the baseline model is determined by combining the weighted difference value. By comparing the output errors of the two validation samples in each sample pair under the layer-removing model, the second degree of closeness between the prediction errors of the two validation samples under the layer-removing model is determined, and the second single-sample pair sensitivity of each sample pair under the layer-removing model is determined by combining the weighted difference value. The single-sample pair fluctuation sensitivity of each sample pair is determined based on the difference between the first single-sample pair sensitivity and the second single-sample pair sensitivity of each sample pair. The volatility sensitivity of a network layer is determined based on the dispersion of the volatility sensitivity of a single sample pair among multiple sample pairs.

8. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 1, characterized in that, The regularization strength is achieved by randomly discarding neurons in the network layer during training with a corresponding regularization drop probability; Training is performed by setting the corresponding regularization strength according to the fluctuation sensitivity of each network layer until multiple network layers have been trained, resulting in the trained target model, including: Map the fluctuation sensitivity of each network layer to the corresponding regularized dropout probability; According to the arrangement order of the network layers from the output end to the input end in the preset neural network, the parameter update process of the network layer to be constrained is constrained according to the regularized dropout probability corresponding to each network layer to be constrained. After training, the weight parameters of the constrained network layers are fixed until multiple network layers are constrained, and the trained target model is obtained.

9. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 8, characterized in that, The parameter update process of the network layer to be constrained according to the regularized dropout probability corresponding to the network layer to be constrained includes: Based on the trained network model, a layer deletion analysis is performed on the network layer to be constrained to determine the current fluctuation sensitivity of the network layer to be constrained, and the current fluctuation sensitivity is mapped to the current regularization drop probability. The parameter update process of the network layer to be constrained is constrained according to the current regularized dropout probability to obtain the updated network model, and the updated network model is used as the network model for the next training operation.

10. The method for three-dimensional anti-collision layout of complex steel mesh for wind turbine pile foundations according to claim 1, characterized in that, Obtain historical engineering data for wind turbine pile foundations, including: Obtain historical engineering records of multiple completed wind turbine pile foundations, and extract corresponding site environmental parameters, external load parameters, structural design parameters, and three-dimensional layout data of steel mesh that has passed construction acceptance. The site environment parameters, external load parameters, and structural design parameters are used as attribute parameters in the historical state data, and the three-dimensional layout data of the steel mesh that has passed construction acceptance is used as the sample three-dimensional layout data.

Images