Automatic pile foundation layout method and system for complex geological foundation

Abstract

The application provides a pile foundation automatic arrangement method and system for complex geological foundation, relates to the technical field of pile foundation automatic arrangement, and comprises the following steps: inputting a grid tensor into an improved ResNet-18 CNN network which has been trained, and outputting a geological determination parameter corresponding to each grid tensor; splicing the geological determination parameter of each grid according to the grid position to obtain a geological determination parameter map; inputting the geological determination parameter map and a current pile position matrix into a trained DQN network for reinforcement learning to obtain an optimal pile position coordinate set of all grids and corresponding pile length, pile number and total cost; traversing each optimal pile position to determine the geological determination parameter; and inputting the geological determination parameter map and the current pile position matrix into the DQN network for reinforcement learning, so that an optimal pile position coordinate set, pile length and pile number which consider both karst cave risk avoidance and total cost minimization can be autonomously learned in a huge pile position layout search space.

Description

Technical Field

[0001] This invention relates to the field of automatic pile foundation layout technology, specifically to an automatic pile foundation layout method and system for complex geological foundations. Background Technology

[0002] Traditional pile foundation engineering design primarily relies on geological exploration reports, empirical formulas, and standard manuals for manual pile placement decisions. This approach struggles to fully integrate multi-source geological information for global pile optimization when facing complex geological conditions such as karst cave zones or alternating soft and hard strata. It typically only estimates pile lengths locally based on individual borehole data, lacking a systematic utilization of the spatial distribution characteristics of the geological conditions across the entire exploration area. Furthermore, manual pile placement schemes often employ uniform or empirically-based densification methods, failing to differentiate layouts based on bearing capacity requirements and geological risks. This leads to problems such as an excessive number of piles, material waste, or insufficient local safety redundancy. In addition, the interpretation of geological parameters and pile location adjustments in traditional design processes depend on engineer experience, resulting in long design cycles and significant fluctuations in scheme quality, making it difficult to achieve cost optimization while ensuring project safety. While some existing BIM-assisted design methods can achieve 3D geological visualization and automatic pile length calculation, they are essentially still parametric modeling tools, lacking the ability to autonomously learn and globally optimize pile placement strategies, and thus unable to dynamically generate optimal pile placement schemes based on geological conditions. In the existing technology, document CN119720338A proposes a method for automatic layout of steel pipe pile foundations based on historical data statistical analysis and K-means clustering. Its core process involves collecting site data to form a database, obtaining layout reference targets through regional division and influencing factor identification, classifying historical construction sites using the K-means clustering algorithm, and combining box plots and the 3σ principle for outlier removal and statistical analysis. Finally, the predicted pile foundation layout target is determined by taking the average of the historical reference layout and the layout reference target. However, this method cannot output the specific coordinates of each pile, only providing the overall layout target parameters. Although geological factors are considered, the geological data is not converted into a spatially continuous geological judgment parameter map, failing to reflect the details of geological condition changes on a two-dimensional plane. It lacks a mechanism for identifying and fine-tuning high-risk locations such as karst caves and insufficient bearing capacity, and cannot perform pile-by-pile safety verification and position correction for local adverse geological conditions. Therefore, an automatic pile foundation layout method and system for complex geological foundations is urgently needed.

[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic pile foundation arrangement method and system for complex geological foundations, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The automatic pile foundation layout method for complex geological foundations includes the following steps: S1: Divide the exploration area into grids, set up boreholes in each grid, collect geological exploration parameters from each borehole, preprocess the geological exploration parameters to generate a raster tensor corresponding to each grid, and mark the pile positions in each grid to construct an initial pile position matrix; S2: Input the grid tensor of each grid into the trained improved ResNet-18CNN network, and the network outputs the geological determination parameters of each grid; stitch the geological determination parameters of each grid according to their spatial location to generate a geological determination parameter map covering the entire exploration area; input the geological determination parameter map and the initial pile position matrix into the trained DQN network for reinforcement learning; through the multi-round decision of the DQN network, output the optimal pile position coordinate set, as well as the pile length, total number of piles and total cost corresponding to each optimal pile position; S3: Traverse each optimal pile location in the set of optimal pile location coordinates and make a judgment based on the geological judgment parameters corresponding to its location: if the preset safety and bearing capacity conditions are met, the pile location is retained; if not, a new pile location is selected again using the polar coordinate generation method within the preset range of the pile location, until the new pile location meets the preset conditions, thereby obtaining the fine-tuned pile location coordinates, the corresponding pile length and the total cost. S4: Verify all pile positions after fine-tuning in S3 to determine if they meet the preset engineering constraints. If they do, output the final optimal pile position layout scheme. If they do not meet the constraints, retain the optimal pile position coordinates, corresponding pile lengths, total number of piles, and total cost output by the DQN network in S2, and output a warning.

[0006] Furthermore, the raster tensor for each grid cell is output, and the initial stake matrix is ​​constructed, specifically including: Based on the preset grid spacing, the exploration area is divided into grids, and boreholes are set at the center point of each grid to collect geological exploration parameters at each borehole. The geological exploration parameters include soil cohesion, internal friction angle, standard penetration test blow count, groundwater level depth, and a marker for the presence of karst caves. The presence of karst caves is marked as 1, and the absence of karst caves is marked as 0. The geological exploration parameters are preprocessed by normalization. The parameters of each grid after normalization are combined in a preset order to form the grid tensor corresponding to that grid. Simultaneously, the pile position status of each grid is marked: if a pile position has been placed at the center point of the grid, the pile position mark value of the grid is set to 1, otherwise it is set to 0; the pile position mark values ​​of all grids are arranged and spliced ​​according to the spatial row and column positions of each grid in the exploration area to construct an initial pile position matrix with the same spatial size as the grid tensor.

[0007] Furthermore, the ResNet-18 CNN network is improved to process the input raster tensor and output the geological determination parameters corresponding to each grid, specifically including: The improved ResNet-18 CNN network, based on the ResNet-18 backbone network, replaces the original fully connected layer at the back end with a 1×1 convolutional layer. This 1×1 convolutional layer compresses the feature map into three channels, which are used to output the predicted bearing capacity, the predicted probability of karst caves, and the predicted probability distribution of soil layer categories, respectively, serving as the geological determination parameters corresponding to the grid. The soil layer categories include clay, sand, and rock. The improved ResNet-18 CNN network can be trained in the following ways: The preprocessed raster tensors of historical exploration areas are collected as network input samples, and the bearing capacity, karst cave probability, and soil layer type of the corresponding grids, as determined by experts, are used as real labels. The network output is constrained using a multi-task loss function, which is obtained by weighted summation of the mean square error loss of bearing capacity prediction, the binary cross-entropy loss of karst cave probability prediction, and the cross-entropy loss of soil layer category prediction. The network parameters are updated using the backpropagation algorithm until the total loss value converges, thus completing the network training.

[0008] Furthermore, reinforcement learning is performed through the DQN network to output the optimal set of pile location coordinates and the corresponding pile length, total number of piles, and total cost, specifically including: The DQN network is initialized, consisting of a target network and a main network. The DQN network is then trained to generate the optimal set of pile location coordinates, along with the corresponding pile lengths, total number of piles, and total cost. At the start of each training round, the current geological assessment parameter map is obtained, and the constructed initial pile location matrix is ​​copied as the current pile location matrix. The two are then concatenated as the initial state. In each decision step, a greedy strategy is used to select the action: first, a random number between 0 and 1 is generated. If the random number is less than the current exploration rate, a pile location coordinate and pile length are randomly selected from the action space; otherwise, the current state is input into the main network, and the action corresponding to the maximum value among all Q values ​​output by the main network is selected. After each decision is made, the current exploration rate is gradually reduced according to a preset decay coefficient until it reaches a preset minimum value; When a selected action is executed, it is first determined whether the action is a termination action: if it is a termination action, the current round of training ends and the next round begins; if it is not a termination action, the corresponding position in the current pile position matrix is ​​marked as an existing pile position according to the coordinates and pile length specified in the action, and the total number of piles is increased by one. The determination of whether it is a termination action specifically involves checking whether the current round meets the preset termination conditions, including whether the current total number of piles has reached the preset upper limit, or whether the decrease ratio of the total cost of each round compared to the total cost of the previous round is lower than the preset ratio threshold; if any of the above termination conditions are met, the system will forcibly mark the current step as a termination action. Then calculate the immediate reward for performing the selected action: Specifically, calculate the total cost before and after performing the action, which includes material cost, construction cost and risk cost. Subtract the total cost before performing the action from the total cost after performing the action, and take the opposite number as the immediate reward. That is, a positive reward is obtained when the total cost decreases due to the new pile placement, and a negative reward is obtained otherwise. After the selected action is completed, the next state is obtained based on the updated current pile position matrix, and it is determined whether the current wheel meets the termination condition. If the termination condition is met, the termination flag is set to true; otherwise, it is set to false. The five elements—current state, selected action, immediate reward, next state, and termination flag—are stored as an experience sample in the experience replay buffer. When the number of experience samples stored in the buffer is greater than the preset small batch size, the network parameters are updated after each environmental interaction step. The specific update process is as follows: a small batch of experience samples is randomly and uniformly sampled from the buffer, and the target Q value is calculated using the target network; if the termination flag of the sample is true, the target Q value is directly equal to the immediate reward in the sample; if the termination flag is false, the next state is input into the target network to obtain the Q values ​​of all actions in the state, and the maximum value is selected, multiplied by the discount factor, and then added to the immediate reward as the target Q value. The Q-value predicted by the main network for the current state and the selected action is compared with the target Q-value. The difference is calculated and the squared average is taken to obtain the loss value. The gradient of the loss function with respect to the parameters of the main network is calculated by the backpropagation algorithm. The parameters of the main network are updated by the Adam optimizer so that the Q-value predicted by the main network gradually approaches the target Q-value. After each fixed number of decision steps, the current parameters of the main network are completely copied to the target network. When the total cost fluctuation of consecutive preset rounds is less than the preset percentage threshold and the cumulative reward of consecutive preset rounds no longer increases, the DQN network is determined to be trained and the optimal set of pile coordinates, as well as the corresponding pile length, total number of piles, and total cost are output.

[0009] Furthermore, the optimal pile location is determined and fine-tuned based on geological assessment parameters, including: Iterate through each optimal pile location in the set of optimal pile location coordinates and make a comprehensive judgment on the geological judgment parameters corresponding to its location. Specifically, if the probability of karst cave at the pile location is greater than the preset probability, or the bearing capacity is lower than the design requirement threshold, or the soil layer category belongs to the preset category that is not suitable for direct pile placement, then the pile location is determined not to meet the preset conditions, the pile location is deleted, and a new pile location is selected again within the preset radius of the current pile location using the polar coordinate generation method, until the new pile location meets all the preset conditions at the same time. If no pile location that meets all the conditions can be found within the preset radius, then the candidate point with the lowest probability of karst cave and the highest bearing capacity is selected. The rate of reduction in the probability of karst cave compared with the original pile location is calculated, and experts give a comprehensive score to determine whether to adopt the candidate point.

[0010] Furthermore, the optimal pile location is reselected using the polar coordinate generation method, including: Mark the pile locations that are currently determined not to meet the preset conditions as high-risk pile locations; calculate the maximum search radius step number based on the grid side length and the preset maximum allowable search radius, and determine the number of angle partitions under each radius step number; Based on the maximum search radius step count and the number of angle partitions, a candidate point set is generated: in, The x-coordinate represents the current high-risk pile location; The vertical coordinate represents the current high-risk pile location; This indicates the number of search radius steps, expressed in units of grid side length; Indicates the angular partition index; This indicates rounding to the nearest integer. This represents the number of angular partitions with radius k. Represents the set of candidate points; Indicates the number of steps for the search radius. Angular partitioning is The increment of the x-coordinate of the candidate points at that time; Indicates the number of steps for the search radius. Angular partitioning is The increment of the ordinate of the candidate point at that time; Indicates the maximum search radius in steps; This indicates the preset maximum allowed search radius; Indicates the grid side length; Indicates rounding up; For each candidate point in the candidate point set, the following conditions are comprehensively determined: the probability of karst cave is less than the preset probability threshold, the bearing capacity is greater than the minimum design requirement, and the soil layer type does not belong to the preset unsuitable pile placement category; the candidate point that meets all the conditions at the same time is taken as the new pile location and the search is terminated; if multiple candidate points meet the conditions at the same time, the one with the lowest probability of karst cave and the highest bearing capacity is selected from them.

[0011] Furthermore, a comprehensive verification of all the adjusted pile positions is conducted to determine whether they meet the preset engineering constraints, including: A comprehensive verification of all pile positions after S3 fine-tuning was performed to determine whether the following constraints were simultaneously met: The increase in total cost after adjustment compared to total cost before adjustment shall not exceed the preset allowable threshold, and the minimum distance between all piles shall not be less than the preset lower limit of pile distance. If all the above constraints are met, the final optimal pile location scheme will be directly output. If the conditions are not met simultaneously, the optimal pile location coordinates, corresponding pile lengths, total number of piles, and total cost directly output by the DQN network described in S2 are retained, and a warning message is output.

[0012] The present invention also provides an automatic pile foundation layout system for complex geological foundations, the automatic pile foundation layout system for complex geological foundations being used to execute the above-described automatic pile foundation layout method for complex geological foundations, comprising: Data acquisition module: used to divide the exploration area into grids, set up boreholes for each grid, collect geological exploration parameters from each borehole, preprocess the geological exploration parameters to generate a raster tensor corresponding to each grid, and mark the pile positions of each grid to construct an initial pile position matrix; The reinforcement learning module is used to input the grid tensor of each grid into the trained improved ResNet-18CNN network, which outputs the geological determination parameters of each grid. The geological determination parameters of each grid are stitched together according to their spatial location to generate a geological determination parameter map covering the entire exploration area. The geological determination parameter map and the initial pile position matrix are input into the trained DQN network for reinforcement learning. Through multiple rounds of decision-making by the DQN network, the optimal pile position coordinate set, as well as the pile length, total number of piles, and total cost corresponding to each optimal pile position are output. The optimal pile location generation module is used to traverse each optimal pile location in the set of optimal pile location coordinates and make a judgment based on the geological judgment parameters corresponding to its location: if the preset safety and bearing capacity conditions are met, the pile location is retained; if not, a new pile location is selected again using the polar coordinate generation method within the preset range of the pile location, until the new pile location meets the preset conditions, thereby obtaining the fine-tuned pile location coordinates, the corresponding pile length, and the total cost. The verification module is used to verify all pile positions after fine-tuning in S3 and determine whether they meet the preset engineering constraints. If they do, the module outputs the final optimal pile position layout scheme. If they do not meet the constraints, the module retains the optimal pile position coordinates, corresponding pile lengths, total number of piles, and total cost output by the DQN network in S2, and outputs a warning.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This application improves the ResNet-18 CNN network for end-to-end feature extraction from rasterized geological data, automatically identifying the bearing capacity, karst probability, and soil layer type at each grid location. It transforms discrete borehole data into a geological assessment parameter map covering the entire area, providing a systematic and structured geological information input for subsequent intelligent decision-making, overcoming the problem of insufficient geological data utilization in traditional methods. The geological assessment parameter map and pile location matrix are jointly input into a DQN network for reinforcement learning, with total cost reduction as the reward orientation. This enables the agent to autonomously explore the optimal pile placement strategy globally, outputting the cost-optimal combination of pile location coordinates, pile length, and number of piles, effectively reducing material costs. Regarding construction costs, and addressing the limitation that DQN-output pile locations might be situated in high-risk geological areas, a polar coordinate generation method was employed to conduct a local spiral search within a preset range. Safe pile locations were reselected based on a comprehensive assessment of factors such as karst cave probability, bearing capacity, and soil layer type. This approach achieved local safety correction while preserving overall cost advantages, embodying the technical principle of balancing global optimization with local safety. An overall engineering constraint verification step was implemented to double-verify the pile spacing and cost increase of the fine-tuned scheme. If the conditions were not met, the scheme reverted to the original DQN scheme and a warning was issued. This ensured the safety and reliability of the scheme and provided designers with traceable decision-making support. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 A graph showing the relationship between the maximum search radius step count and the grid side length; Figure 3 This is a schematic diagram of the overall system of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0016] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0017] Example: Please see Figures 1-2 The present invention provides a technical solution: The automatic pile foundation layout method for complex geological foundations includes the following steps: S1: Divide the exploration area into grids, set up boreholes in each grid, collect geological exploration parameters from each borehole, preprocess the geological exploration parameters to generate a raster tensor corresponding to each grid, and mark the pile positions in each grid to construct an initial pile position matrix; In this embodiment, the output of the raster tensor for each grid and the construction of the initial stake matrix specifically include: Based on the preset grid spacing, the exploration area is divided into grids, and boreholes are set at the center point of each grid to collect geological exploration parameters at each borehole. The geological exploration parameters include soil cohesion, internal friction angle, standard penetration test blow count, groundwater level depth, and a marker for the presence of karst caves. The presence of karst caves is marked as 1, and the absence of karst caves is marked as 0. The geological exploration parameters are preprocessed by normalization. The parameters of each grid after normalization are combined in a preset order to form the grid tensor corresponding to that grid. Simultaneously, the pile position status of each grid is marked: if a pile position has been placed at the center point of the grid, the pile position mark value of the grid is set to 1, otherwise it is set to 0; the pile position mark values ​​of all grids are arranged and spliced ​​according to the spatial row and column positions of each grid in the exploration area to construct an initial pile position matrix with the same spatial size as the grid tensor.

[0018] In the above process, the exploration area is first uniformly divided by a preset grid spacing and a borehole is set at the center point of each grid. The purpose is to discretize the continuous spatial area into a finite number of regular grid units, so that each grid corresponds to a set of clear spatial coordinates, thereby establishing a one-to-one mapping relationship between geological parameters and spatial location, and providing a standardized data structure for subsequent convolutional neural network to extract spatial features. Secondly, the collected geological exploration parameters, such as soil cohesion, internal friction angle, standard penetration test blow count, groundwater level depth, and cavern presence markers, are preprocessed by normalization. This is because the physical meaning, dimensions, and numerical ranges of these parameters vary. If unnormalized data is directly input into the neural network, features with larger numerical ranges will dominate the gradient update direction of the loss function, leading to unstable model training and slow convergence. Normalization scales all features to the same scale, ensuring that geological parameters of each dimension are treated equally during training, thereby improving the model's convergence speed and prediction accuracy. Combining the normalized parameters in a preset order to form the raster tensor corresponding to each grid binds the multidimensional geological data of the same grid into a whole in the feature dimension. This ensures that the tensor representation of each grid contains all the geological information of that location while maintaining the structural relationships between the parameters. Simultaneously, the pile position status of each grid is binarized and marked, and all marked values ​​are arranged and spliced ​​according to the spatial row and column positions of the grid in the exploration area to form an initial pile position matrix. This is to record the distribution of pile foundations in the entire exploration area in a simple binary encoding method. Since the initial pile position matrix and the grid tensor have the same spatial size, they can be directly spliced ​​together in subsequent steps and used as the state input of the DQN network. This allows the agent to perceive the spatial distribution characteristics of geological conditions and understand the density and distribution pattern of the currently deployed piles when making decisions at each step, so as to make a globally optimal pile deployment decision that takes into account both geological conditions and the current pile deployment status. Through the above processing, the originally scattered and heterogeneous geological report data and pile location record data are uniformly converted into regular numerical representations in tensor and matrix form, providing a standardized, complete and computable data foundation for improving the geological parameter inference of ResNet-18 CNN network and the reinforcement learning decision-making of DQN network.

[0019] S2: Input the grid tensor of each grid into the trained improved ResNet-18CNN network, and the network outputs the geological determination parameters of each grid; stitch the geological determination parameters of each grid according to their spatial location to generate a geological determination parameter map covering the entire exploration area; input the geological determination parameter map and the initial pile position matrix into the trained DQN network for reinforcement learning; through the multi-round decision of the DQN network, output the optimal pile position coordinate set, as well as the pile length, total number of piles and total cost corresponding to each optimal pile position; In this embodiment, the improved ResNet-18 CNN network is used to process the input raster tensor and output the geological determination parameters corresponding to each grid, specifically including: The improved ResNet-18 CNN network, based on the ResNet-18 backbone network, replaces the original fully connected layer at the back end with a 1×1 convolutional layer. This 1×1 convolutional layer compresses the feature map into three channels, which are used to output the predicted bearing capacity, the predicted probability of karst caves, and the predicted probability distribution of soil layer categories, respectively, serving as the geological determination parameters corresponding to the grid. The soil layer categories include clay, sand, and rock. The improved ResNet-18 CNN network can be trained in the following ways: The preprocessed raster tensors of historical exploration areas are collected as network input samples, and the bearing capacity, karst cave probability, and soil layer type of the corresponding grids, as determined by experts, are used as real labels. The network output is constrained using a multi-task loss function, which is obtained by weighted summation of the mean square error loss of bearing capacity prediction, the binary cross-entropy loss of karst cave probability prediction, and the cross-entropy loss of soil layer category prediction. The network parameters are updated using the backpropagation algorithm until the total loss value converges, thus completing the network training.

[0020] In the above process, the ResNet-18 CNN network structure is improved. First, 64 7*7 convolutional kernels are used to perform convolution operations on the input grid tensor with a stride of 2. Then, a 3*3 pooling layer with a stride of 2 is connected. After this basic feature extraction, a 1*1 convolutional layer is used to compress the feature map into 3 channels. Each channel outputs bearing capacity, cave probability, and soil layer type, including clay, sand, and rock, which constitute the geological determination parameters corresponding to each grid. First, the original fully connected layers at the back end of the ResNet-18 backbone network are replaced with 1×1 convolutional layers. This is because the original fully connected layers of ResNet-18 flatten the feature map into a one-dimensional vector and output a classification result with a fixed dimension. This process loses the spatial information of each grid position in the feature map, causing the network to only make a single overall prediction for the entire input map, and cannot output the corresponding geological determination parameters for each grid position in the input raster tensor. After replacing the fully connected layers with 1×1 convolutional layers, the convolutional layers independently perform linear transformations of the channel dimension for each spatial position on the feature map, making the height and width of the output feature map completely consistent with the spatial size of the input raster tensor. This enables point-by-point prediction of geological parameters for each grid position, providing an output structure that is strictly aligned with the input space for subsequent steps to stitch together the geological determination parameters of each grid into a geological determination parameter map covering the entire exploration area. The feature map is compressed into three channels using a 1×1 convolutional layer, which are used to output the predicted bearing capacity, the predicted cavern probability, and the predicted probability distribution of soil layer categories, respectively. Different output representations are designed for the different mathematical properties of the three geological parameters: bearing capacity is a continuous value, which is suitable for direct regression output by a single neuron; the cavern probability is a binary confidence level between 0 and 1, which is suitable for outputting the probability value using the Sigmoid activation function; the soil layer categories involve three mutually exclusive categories: clay, sand, and rock, which are suitable for outputting the probability distribution of the three categories using the Softmax activation function. The three outputs do not interfere with each other and share the underlying ResNet-18 feature extraction network, realizing parameter sharing and joint optimization in a multi-task learning framework. In terms of training methods, preprocessed raster tensors from historical exploration areas are collected as input samples for the network, and the bearing capacity, karst cave probability, and soil layer type of the corresponding grids calibrated by experts are used as real labels. This is to utilize expert knowledge to provide training signals for supervised learning, enabling the network to learn the mapping relationship between raster tensors and geological judgment parameters. A multi-task loss function is used to constrain the network output. Bearing capacity uses mean squared error loss, cave probability uses binary cross-entropy loss, and soil layer category uses cross-entropy loss. The total loss is obtained by weighted summation of these three loss functions. This is because the three prediction tasks have different natures: bearing capacity regression focuses on the numerical deviation between the predicted and true values, making mean squared error loss, which measures the error in continuous value prediction, suitable; cave probability prediction is essentially a binary classification problem to determine the existence of caves, making binary cross-entropy loss, which measures the difference between binary confidence and the true label, suitable; and soil layer category prediction is a 3-class classification problem, making cross-entropy loss, which measures the difference between the predicted probability distribution and the one-hot encoding of the true category, suitable. Weighted summation of these three loss functions allows the network to simultaneously consider all three tasks during training, leveraging the regularization effect of multi-task learning to improve the model's generalization ability. Finally, all trainable parameters in the network are updated using backpropagation and stochastic gradient descent until the total loss value converges, enabling the network to achieve satisfactory prediction accuracy on three tasks simultaneously: bearing capacity regression, karst cave probability binary classification, and soil layer category triclass classification. Through the aforementioned network structure improvements and multi-task training strategies, the improved ResNet-18 CNN network can output accurate and comprehensive geological determination parameters for the raster tensor of each grid, providing refined geological information input covering three dimensions—bearing capacity, karst cave risk, and soil layer type—for the subsequent pile placement decision of the DQN network.

[0021] The DQN network is used for reinforcement learning to output the optimal set of pile location coordinates and the corresponding pile length, total number of piles, and total cost. Specifically, this includes: The DQN network is initialized, consisting of a target network and a main network. The DQN network is then trained to generate the optimal set of pile location coordinates, along with the corresponding pile lengths, total number of piles, and total cost. At the start of each training round, the current geological assessment parameter map is obtained, and the constructed initial pile location matrix is ​​copied as the current pile location matrix. The two are then concatenated as the initial state. In each decision step, a greedy strategy is used to select the action: first, a random number between 0 and 1 is generated. If the random number is less than the current exploration rate, a pile location coordinate and pile length are randomly selected from the action space; otherwise, the current state is input into the main network, and the action corresponding to the maximum value among all Q values ​​output by the main network is selected. After each decision is made, the current exploration rate is gradually reduced according to a preset decay coefficient until it reaches a preset minimum value; When a selected action is executed, it is first determined whether the action is a termination action: if it is a termination action, the current round of training ends and the next round begins; if it is not a termination action, the corresponding position in the current pile position matrix is ​​marked as an existing pile position according to the coordinates and pile length specified in the action, and the total number of piles is increased by one. The determination of whether it is a termination action specifically involves checking whether the current round meets the preset termination conditions, including whether the current total number of piles has reached the preset upper limit, or whether the decrease ratio of the total cost of each round compared to the total cost of the previous round is lower than the preset ratio threshold; if any of the above termination conditions are met, the system will forcibly mark the current step as a termination action. Then calculate the immediate reward for performing the selected action: Specifically, calculate the total cost before and after performing the action, which includes material cost, construction cost and risk cost. Subtract the total cost before performing the action from the total cost after performing the action, and take the opposite number as the immediate reward. That is, a positive reward is obtained when the total cost decreases due to the new pile placement, and a negative reward is obtained otherwise. After the selected action is completed, the next state is obtained based on the updated current pile position matrix, and it is determined whether the current wheel meets the termination condition. If the termination condition is met, the termination flag is set to true; otherwise, it is set to false. The five elements—current state, selected action, immediate reward, next state, and termination flag—are stored as an experience sample in the experience replay buffer. When the number of experience samples stored in the buffer is greater than the preset small batch size, the network parameters are updated after each environmental interaction step. The specific update process is as follows: a small batch of experience samples is randomly and uniformly sampled from the buffer, and the target Q value is calculated using the target network; if the termination flag of the sample is true, the target Q value is directly equal to the immediate reward in the sample; if the termination flag is false, the next state is input into the target network to obtain the Q values ​​of all actions in the state, and the maximum value is selected, multiplied by the discount factor, and then added to the immediate reward as the target Q value. The Q-value predicted by the main network for the current state and the selected action is compared with the target Q-value. The difference is calculated and the squared average is taken to obtain the loss value. The gradient of the loss function with respect to the parameters of the main network is calculated by the backpropagation algorithm. The parameters of the main network are updated by the Adam optimizer so that the Q-value predicted by the main network gradually approaches the target Q-value. After each fixed number of decision steps, the current parameters of the main network are completely copied to the target network. When the total cost fluctuation of consecutive preset rounds is less than the preset percentage threshold and the cumulative reward of consecutive preset rounds no longer increases, the DQN network is determined to be trained and the optimal set of pile coordinates, as well as the corresponding pile length, total number of piles, and total cost are output.

[0022] In the above process, the preset decay coefficient refers to the proportional factor by which the current exploration rate is multiplied and reduced after each round of training or each step of decision. It is determined as follows: before the start of training, it is set according to the total number of training rounds and the expected exploration rate decay rate. The value is usually between 0.99 and 0.9999. The specific value is calculated by the following formula: Preset decay coefficient = (Preset minimum exploration rate / Initial exploration rate)^(1 / Total number of training rounds). This setting ensures that the exploration rate decays smoothly from the initial value to the preset minimum value within the preset total number of training rounds, and the decay process is exponential rather than nonlinear. The preset minimum value refers to the lowest limit value that the exploration rate can reach during the decay process. It is determined by setting it according to the preset expectation of the exploration needs in the later stage of training. It is usually between 0.01 and 0.05. That is, when the training is completed, the agent still makes random action selection with a probability of 1% to 5% rather than greedily selecting the action with the largest Q value. This is because even in the later stage of training, retaining a very low probability of random exploration still helps to prevent the DQN network from getting stuck in a local optimum due to prematurely stopping exploration. At the same time, the value should not be too low to prevent the complete loss of the opportunity to explore potential better stub schemes, nor should it be too high to avoid the policy fluctuations being too large in the later stage of training and failing to converge stably. The "Preset Upper Limit" refers to the maximum number of piles allowed in a training round. It is determined by estimating the theoretically required minimum and maximum number of piles based on the total area of ​​the exploration area and the load requirements of the superstructure, and then setting the preset upper limit as the maximum number of piles. The "Relative Change Rate Threshold" is a threshold used to determine whether the training has converged. It is set before training begins based on the expected convergence accuracy of the DQN network, and is usually between 0.1% and 1.0%. When the decrease in total cost compared to the previous round is consistently less than this percentage in multiple training rounds, it indicates that the network can no longer further optimize the pile placement scheme based on the current strategy, thus triggering a termination flag and ending the training. The reason for adopting the above-mentioned DQN network structure and training process is to model the pile foundation layout problem as a sequential decision problem, so that the agent can autonomously learn the optimal pile layout strategy that takes into account both the avoidance of karst cave risk and the minimization of total cost in the huge pile location layout search space.

[0023] Specifically, the DQN network is first initialized and the target network and main network are set up. This is because in the DQN algorithm, the main network is used to predict the action Q value in the current state in real time and guide the action selection, while the target network is used to provide a relatively stable benchmark for calculating the target Q value. By separating the two and synchronizing the parameters periodically, the temporal correlation between training data can be effectively broken, the overestimation problem in the bootstrap process can be suppressed, and the network can be prevented from oscillating or even diverging during the update process, thereby ensuring the stability and convergence of training.

[0024] The use of a greedy strategy to select actions and the gradual decay of the exploration rate are intended to achieve a balance between exploration and exploitation in reinforcement learning. In the early stages of training, a higher exploration rate allows the agent to try a wide range of different peg placement combinations, avoiding getting stuck in local optima. As training progresses, the exploration rate gradually decreases, and the agent gradually shifts to using the Q-value prediction of the main network to select the optimal action under the current cognition, thus achieving a smooth transition from exploration-oriented to exploitation-oriented learning.

[0025] The reason for setting termination actions and termination conditions is that each round of pile placement decisions must terminate within a reasonable number of piles. When the current total number of piles reaches the preset upper limit or the total cost reduction ratio is lower than the preset threshold for multiple consecutive rounds, it indicates that continuing to add piles in the current round can no longer bring significant cost optimization. Forcibly triggering the termination action can avoid invalid decisions. At the same time, the termination flag is stored in the buffer as part of the experience sample, which simplifies the calculation of the target Q value in the termination state to be directly equal to the immediate reward. This is in line with the mathematical definition that there is no future reward in the termination state in the Markov decision process. The negative of the cost reduction is used as the immediate reward because the DQN network aims to maximize the cumulative reward. This design allows the agent to receive a positive reward when the total cost decreases and a negative reward when it increases, thereby driving the agent to actively discover piling strategies that can effectively reduce material costs, construction costs and risk costs. The five elements—current state, selected action, immediate reward, next state, and termination flag—are stored as experience samples in the experience replay buffer, and network parameters are updated by randomly and uniformly sampling from it. This is to break the temporal correlation between continuous decision samples, make the training data closer to the independent and identically distributed hypothesis, and thus improve training efficiency and model generalization ability.

[0026] The target Q value is calculated using the Bellman equation. If the termination flag is false, the next state is input into the target network to obtain the maximum Q value of all actions in that state, multiplied by the discount factor, and then the immediate reward is added. This design ensures that the value of the current decision not only includes the immediate reward obtained at the moment, but also proactively considers the long-term impact of the current placement decision on the future placement space, enabling the agent to make a globally optimal decision that takes into account both the current cost and the future placement feasibility. The DQN network employs a mean squared error loss function, backpropagation algorithm, and Adam optimizer to update the main network parameters, enabling the Q-value predicted by the main network to gradually approach the target Q-value, thus achieving temporal difference learning. After each fixed number of steps, the main network parameters are completely copied to the target network, providing a stable reference benchmark for the calculation of the target Q-value and further suppressing training oscillations. Training is considered complete when the total cost fluctuation over consecutive preset rounds is less than a preset percentage threshold and the cumulative reward no longer increases, ensuring that the network has converged to a stable optimal policy. Through the above design and training process of the DQN network, this scheme can efficiently search in a combinatorial optimization space containing a large number of grid positions and output a globally optimal slinging scheme that is autonomously optimized through reinforcement learning.

[0027] Among them, the full name of Q value is action value function value. Its core function is to quantitatively evaluate the long-term comprehensive benefits of the action of arranging a specific length of pile foundation at a specific coordinate under given geological conditions and current pile layout.

[0028] S3: Traverse each optimal pile location in the set of optimal pile location coordinates and make a judgment based on the geological judgment parameters corresponding to its location: if the preset safety and bearing capacity conditions are met, the pile location is retained; if not, a new pile location is selected again using the polar coordinate generation method within the preset range of the pile location, until the new pile location meets the preset conditions, thereby obtaining the fine-tuned pile location coordinates, the corresponding pile length and the total cost. In this embodiment, the optimal pile location is determined and fine-tuned based on geological assessment parameters, including: Iterate through each optimal pile location in the set of optimal pile location coordinates and make a comprehensive judgment on the geological judgment parameters corresponding to its location. Specifically, if the probability of karst cave at the pile location is greater than the preset probability, or the bearing capacity is lower than the design requirement threshold, or the soil layer category belongs to the preset category that is not suitable for direct pile placement, then the pile location is determined not to meet the preset conditions, the pile location is deleted, and a new pile location is selected again within the preset radius of the current pile location using the polar coordinate generation method, until the new pile location meets all the preset conditions at the same time. If no pile location that meets all the conditions can be found within the preset radius, then the candidate point with the lowest probability of karst cave and the highest bearing capacity is selected. The rate of reduction in the probability of karst cave compared with the original pile location is calculated, and experts give a comprehensive score to determine whether to adopt the candidate point.

[0029] In the above process, the optimal pile location output by the DQN network is subjected to pile-by-pile geological assessment and fine-tuning because the DQN network performs global optimization with the core optimization objective of minimizing total cost. Its decision-making process is mainly based on the cost-benefit trade-off between bearing capacity, karst cave probability, and soil layer type distribution reflected in the geological assessment parameter map. In some cases, pile locations in areas with high karst cave probability or low bearing capacity may be selected in order to reduce overall cost. For example, although there is a certain geological risk, the cost of a single pile is low, which is attractive from the perspective of global cost. However, such pile locations may not meet the hard design requirements of safety and bearing capacity in engineering practice. Therefore, a safety verification step needs to be added on the global cost optimization result of the DQN network. Each optimal pile location is traversed, and the probability of karst caves, bearing capacity, and soil layer type corresponding to its location are judged item by item. When any parameter does not meet the preset conditions, the pile location is deleted and the polar coordinate generation method is triggered to reselect the location. This is because an excessively high probability of karst caves means that the pile foundation construction may encounter cavities that may cause hole collapse or uncontrolled grouting. Bearing capacity is lower than the design requirement threshold, which means that the pile foundation cannot provide sufficient support for the superstructure. Soil layer type belongs to the preset unsuitable for direct pile placement, which means that the soil layer, such as highly compressible soft soil or strongly weathered and broken rock layer, does not have the basic conditions to serve as a bearing layer at the pile end. If any of the three conditions is not met, it is enough to constitute a safety hazard in the project, and the location must be corrected. The reason for using polar coordinate generation within the preset radius of the current pile location instead of simply selecting points randomly around it is that polar coordinate generation, through the preset maximum search radius, the search step size in units of grid side length, and the number of angle partitions that increases with the radius, can form a uniform candidate point grid that spirals outward around the original pile location. This ensures that the search process will not miss nearby high-quality points due to excessive step size, nor will it over-compute in invalid areas due to excessive step size. At the same time, the number of angle partitions increases synchronously with the radius to ensure that the search density remains uniform across all radius layers, thereby efficiently traversing all candidate locations around the original pile location within a limited computational cost. If a pile location that simultaneously meets all preset conditions cannot be found within the preset radius, then the candidate point with the lowest probability of karst cave and the highest bearing capacity is selected, and its rate of reduction in karst cave probability compared with the original pile location is calculated. Experts will then give a comprehensive score to determine whether to adopt the candidate point. This is because if a pile location that simultaneously meets all the hard conditions cannot be found within the local search radius, it indicates that there are certain geological defects in the area surrounding the original pile location. If it is forced to meet all the conditions at the same time, the search may fail and the pile location may have to be abandoned, which will affect the overall pile number distribution and bearing capacity balance. Therefore, by selecting the relatively optimal candidate point and calculating the rate of reduction in karst cave probability, a quantitative indicator is used. The preset radius refers to the local boundary value used in the polar coordinate generation method to search for alternative pile locations. It is determined by comprehensively considering the influence range of the pile group effect in the pile foundation project and the positioning accuracy of the construction equipment, typically taking a value of 1 to 3 times the grid side length. The preset probability refers to the threshold value for determining whether a pile location needs to be deleted or adjusted due to excessive karst cave risk. It is usually determined by geotechnical engineers based on the intensity of karst cave development at the site and construction experience, generally ranging from 0.3 to 0.7. The design requirement threshold refers to the threshold value used to determine whether the pile location's bearing capacity meets the requirements of the superstructure. The minimum limit value of structural load requirements is determined by first calculating the standard and ultimate values ​​of the vertical bearing capacity that a single pile needs to bear based on the dead load, live load, and seismic load of the superstructure. Then, based on the calculation method of the vertical bearing capacity of a single pile in the "Code for Design of Building Pile Foundations", and combined with the physical and mechanical parameters of each soil layer on site (such as soil cohesion, internal friction angle, standard penetration blow count, etc.), the minimum bearing capacity value required to meet the design load requirements is calculated. At the same time, a certain safety reserve factor (usually 2.0) is considered, and the final lower limit value of the bearing capacity is used as the design requirement threshold.

[0030] The optimal pile location is reselected using the polar coordinate generation method, including: Mark the pile locations that are currently determined not to meet the preset conditions as high-risk pile locations; calculate the maximum search radius step number based on the grid side length and the preset maximum allowable search radius, and determine the number of angle partitions under each radius step number; Based on the maximum search radius step count and the number of angle partitions, a candidate point set is generated: in, The x-coordinate represents the current high-risk pile location; The vertical coordinate represents the current high-risk pile location; This indicates the number of search radius steps, expressed in units of grid side length; Indicates the angular partition index; This indicates rounding to the nearest integer. This represents the number of angular partitions with radius k. Represents the set of candidate points; Indicates the number of steps for the search radius. Angular partitioning is The increment of the x-coordinate of the candidate points at that time; Indicates the number of steps for the search radius. Angular partitioning is The increment of the ordinate of the candidate point at that time; Indicates the maximum search radius in steps; This indicates the preset maximum allowed search radius; Indicates the grid side length; Indicates rounding up; For each candidate point in the candidate point set, the following conditions are comprehensively determined: the probability of karst cave is less than the preset probability threshold, the bearing capacity is greater than the minimum design requirement, and the soil layer type does not belong to the preset unsuitable pile placement category; the candidate point that meets all the conditions at the same time is taken as the new pile location and the search is terminated; if multiple candidate points meet the conditions at the same time, the one with the lowest probability of karst cave and the highest bearing capacity is selected from them.

[0031] In the above process, the pile locations that are currently determined not to meet the preset conditions are first marked as high-risk pile locations. This is to clarify that the pile locations have safety hazards in terms of karst cave probability, bearing capacity, or soil layer type, and cannot be directly used, thereby triggering the subsequent local reselection process. At the same time, the coordinates of the pile location are marked as the reference origin of the polar coordinate search, so that all the candidate points generated subsequently are distributed around the origin, ensuring that the alternative pile locations still serve the same superstructure load transfer area. The maximum search radius steps are calculated based on the grid edge length and the preset maximum allowable search radius. It uses the edge length of a single grid cell as the basic search step size, discretizes the continuous maximum allowable search radius into integer multiples of the grid step size, and uses rounding up to ensure that the actual search range completely covers the preset maximum allowable search radius without omitting candidate positions near the boundary; at the same time, according to Determine the number of angle partitions for each radius step, such that the number of angle partitions increases with the radius step. Linear growth means that the larger the radius and the longer the circumference, the more angle partitions there are, thus ensuring that the arc length interval between adjacent candidate points on each layer is roughly uniform, avoiding the omission of candidate points due to too few angle partitions in the near area or the generation of redundant calculations due to too many angle partitions in the far area. Candidate point set and The calculation formula for the radius step in polar coordinates. and angular partition index The coordinates are converted to integer increments in grid units, and rounded to ensure that the candidate point coordinates are aligned with the grid center. This allows each candidate point to directly correspond to a specific grid on the existing geological assessment parameter map. This enables rapid and comprehensive assessment of the probability of karst caves, bearing capacity, and soil layer type at that location, without the need for interpolation or resampling, significantly improving search efficiency. (Complete formula) Candidate points are generated layer by layer from the inside out, forming a spiral expansion search order. That is, the search is first conducted on the circle closest to the original high-risk pile position. If no candidate point meets the conditions, the search is expanded to the next circle. This search strategy from near to far prioritizes the load transfer path of the original pile position and only expands the search range when necessary, thus minimizing the impact on the overall pile layout scheme. For each candidate point in the candidate point set, a comprehensive assessment is conducted to determine whether three conditions are simultaneously met: the probability of karst cavern is less than a preset probability threshold, the bearing capacity is greater than the minimum design requirement, and the soil layer type does not belong to a preset unsuitable pile placement category. This is because these three conditions are independent and necessary hard requirements for pile site safety and bearing capacity. A high probability of karst cavern implies a risk of borehole collapse or over-pouring of concrete; insufficient bearing capacity means the inability to support the superstructure; and an unsuitable soil layer type indicates poor engineering properties of the bearing stratum. All three conditions must be met simultaneously to ensure the engineering safety of the pile site; none can be omitted. When multiple candidate points simultaneously meet all conditions, the one with the lowest probability of karst cavern and the highest bearing capacity is selected. This is because, while meeting basic safety conditions, a lower probability of karst cavern results in lower construction risk, and higher bearing capacity provides a larger safety margin. This selection strategy optimizes multiple feasible candidate points, further enhancing the safety margin of the pile site. If a candidate point that meets the conditions is found in a certain circle, the search is terminated immediately. This is because the search order of the polar coordinate generation method is arranged from near to far. The first candidate point that meets the conditions must be the closest to the original high-risk pile position among all feasible candidate points. There is no need to continue searching for more distant circles, thus greatly shortening the calculation time while ensuring the optimal result. Preset maximum allowable search radius This refers to the maximum straight-line distance limit that allows new candidate pile locations to deviate from the original high-risk pile locations when performing polar coordinate search. The value is usually taken in the range of 1 to 5 times the grid side length, so that the search range is sufficient to cover areas where geological conditions may improve, but not so large as to affect the overall stress balance of the pile foundation and the collaborative working mechanism of the pile group.

[0032] In the above embodiments, 20 sets of data on the grid side length and the corresponding maximum search radius steps are given to reflect the change of the maximum search radius steps as the grid side length changes, as shown in Table 1: Table 1: Relationship between grid side length and corresponding maximum search radius step count The data in Table 1 above are based on the preset maximum allowable search radius. The value was calculated using a fixed unit of meters, as can be seen from the data in the table. With grid side length The relationship is inversely proportional: the smaller the grid side length, the larger the maximum search radius step size, meaning more search layers, denser candidate points, and higher search accuracy, but also a corresponding increase in computational cost; conversely, the larger the grid side length, the smaller the maximum search radius step size, fewer search layers, sparser candidate points, improved search efficiency, but potentially missing some candidate positions. Rice time, This indicates that a search needs to be conducted 20 times around the original high-risk pile location, corresponding to an actual search radius of 10 meters, completely covering the preset maximum allowable search radius; when Rice time, This means that only one search cycle is needed, corresponding to an actual search radius of 10 meters, which also completely covers the preset maximum allowable search radius; the rounding up operation ensures that the actual search radius is not less than the preset maximum allowable search radius under various grid side lengths. It will not miss candidate points near the boundary due to discretization.

[0033] S4: Verify all pile positions after fine-tuning in S3 to determine if they meet the preset engineering constraints. If they do, output the final optimal pile position layout scheme. If they do not meet the constraints, retain the optimal pile position coordinates, corresponding pile lengths, total number of piles, and total cost output by the DQN network in S2, and output a warning.

[0034] A comprehensive verification of all adjusted pile positions was conducted to determine whether they met the preset engineering constraints, including: A comprehensive verification of all pile positions after S3 fine-tuning was performed to determine whether the following constraints were simultaneously met: The increase in total cost after adjustment compared to total cost before adjustment shall not exceed the preset allowable threshold, and the minimum distance between all piles shall not be less than the preset lower limit of pile distance. If all the above constraints are met, the final optimal pile location scheme will be directly output. If the conditions are not met simultaneously, the optimal pile location coordinates, corresponding pile lengths, total number of piles, and total cost directly output by the DQN network described in S2 are retained, and a warning message is output.

[0035] In the above process, when verifying all pile positions after S3 fine-tuning, it is necessary to simultaneously verify two conditions: the increase in total cost after adjustment compared to the total cost before adjustment does not exceed the preset allowable threshold, and the minimum distance between all pile positions is not less than the preset lower limit of pile spacing. This is because although the polar coordinate local fine-tuning of individual high-risk pile positions in S3 ensures that each pile position meets the safety and bearing capacity conditions at the pile-by-pile level, the cumulative effect of multiple fine-tuning operations may lead to two global problems: First, each fine-tuning operation moves the pile position from the high-risk area to the surrounding safe area, which may increase the pile length or construction difficulty, thus gradually increasing the total cost. If the cost increase exceeds the preset allowable threshold, it means that the fine-tuned scheme is no longer economically acceptable. Second, the independent fine-tuning of multiple pile positions may cause adjacent pile positions with reasonable spacing to move closer to each other after local adjustment, resulting in the pile spacing being less than the lower limit of pile spacing required by the specification, thereby causing engineering problems such as stress superposition, bearing capacity reduction, and mutual interference between adjacent piles in the pile group effect. Therefore, it is necessary to verify these two conditions simultaneously at the overall level. If both constraints are met, the final optimal pile location scheme is directly output. This is because the fine-tuned scheme at this point successfully corrects all local safety hazards while maintaining the global cost optimization results of the DQN network. Furthermore, the overall economic efficiency and construction feasibility have been verified by engineering specifications, and the scheme can be directly used to guide construction. If the two constraints are not met simultaneously, the optimal pile coordinates, corresponding pile length, total number of piles, and total cost directly output by the DQN network in S2 are retained, and a warning message is output. This is because when fine-tuning leads to cost exceeding the limit or insufficient pile spacing, it indicates that although the local correction scheme of S3 improves the safety of individual pile positions, it has an unacceptable negative impact on the overall scheme. At this time, there is an irreconcilable conflict between safety and economy or construction feasibility. The safest decision in the project is to revert to the original scheme optimized by reinforcement learning and prompt the designer to conduct manual review. That is, the original scheme of the DQN network is guaranteed to be globally cost-optimal. Although there are geological risks at individual pile positions, they can be compensated by special measures in construction, such as follow-up casing protection and grouting reinforcement. However, if the fine-tuning scheme violates the cost constraint or pile spacing constraint, it is a systemic defect that is difficult to compensate for one by one. Therefore, adopting the revert strategy and issuing a warning can both ensure the bottom line of project safety and provide the designer with a basis for decision-making, reflecting the engineering wisdom of combining automated methods with human experience. The preset allowable threshold is usually set as a fixed percentage value, typically ranging from 3% to 10% of the total pile foundation cost, giving the S3 fine-tuning operation sufficient room to correct safety defects.

[0036] Please see Figure 3The present invention also provides an automatic pile foundation layout system for complex geological foundations, wherein the inspection is used to execute the above-mentioned automatic pile foundation layout method for complex geological foundations, including: Data acquisition module: used to divide the exploration area into grids, set up boreholes for each grid, collect geological exploration parameters from each borehole, preprocess the geological exploration parameters to generate a raster tensor corresponding to each grid, and mark the pile positions of each grid to construct an initial pile position matrix; The reinforcement learning module is used to input the grid tensor of each grid into the trained improved ResNet-18CNN network, which outputs the geological determination parameters of each grid. The geological determination parameters of each grid are stitched together according to their spatial location to generate a geological determination parameter map covering the entire exploration area. The geological determination parameter map and the initial pile position matrix are input into the trained DQN network for reinforcement learning. Through multiple rounds of decision-making by the DQN network, the optimal pile position coordinate set, as well as the pile length, total number of piles, and total cost corresponding to each optimal pile position are output. The optimal pile location generation module is used to traverse each optimal pile location in the set of optimal pile location coordinates and make a judgment based on the geological judgment parameters corresponding to its location: if the preset safety and bearing capacity conditions are met, the pile location is retained; if not, a new pile location is selected again using the polar coordinate generation method within the preset range of the pile location, until the new pile location meets the preset conditions, thereby obtaining the fine-tuned pile location coordinates, the corresponding pile length, and the total cost. The verification module is used to verify all pile positions after fine-tuning in S3 and determine whether they meet the preset engineering constraints. If they do, the module outputs the final optimal pile position layout scheme. If they do not meet the constraints, the module retains the optimal pile position coordinates, corresponding pile lengths, total number of piles, and total cost output by the DQN network in S2, and outputs a warning.

[0037] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0038] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An automatic pile foundation layout method for complex geological foundations, characterized in that, The specific steps include: S1: Divide the exploration area into grids, set up boreholes in each grid, collect geological exploration parameters from each borehole, preprocess the geological exploration parameters to generate a raster tensor corresponding to each grid, and mark the pile positions in each grid to construct an initial pile position matrix; S2: Input the grid tensor of each grid into the trained improved ResNet-18CNN network, and the network outputs the geological determination parameters of each grid; stitch the geological determination parameters of each grid according to their spatial location to generate a geological determination parameter map covering the entire exploration area; input the geological determination parameter map and the initial pile position matrix into the trained DQN network for reinforcement learning; through the multi-round decision of the DQN network, output the optimal pile position coordinate set, as well as the pile length, total number of piles and total cost corresponding to each optimal pile position; S3: Traverse each optimal pile location in the set of optimal pile location coordinates and make a judgment based on the geological judgment parameters corresponding to its location: if the preset safety and bearing capacity conditions are met, the pile location is retained; if not, a new pile location is selected again using the polar coordinate generation method within the preset range of the pile location, until the new pile location meets the preset conditions, thereby obtaining the fine-tuned pile location coordinates, the corresponding pile length and the total cost. S4: Verify all pile positions after fine-tuning in S3 to determine if they meet the preset engineering constraints. If they do, output the final optimal pile position layout scheme. If they do not meet the constraints, retain the optimal pile position coordinates, corresponding pile lengths, total number of piles, and total cost output by the DQN network in S2, and output a warning.

2. The automatic pile foundation layout method for complex geological foundations according to claim 1, characterized in that, Output the raster tensor for each grid cell and construct the initial stake matrix, specifically including: Based on the preset grid spacing, the exploration area is divided into grids, and boreholes are set at the center point of each grid to collect geological exploration parameters at each borehole. The geological exploration parameters include soil cohesion, internal friction angle, standard penetration test blow count, groundwater level depth, and a marker for the presence of karst caves. The presence of karst caves is marked as 1, and the absence of karst caves is marked as 0. The geological exploration parameters are preprocessed by normalization. The parameters of each grid after normalization are combined in a preset order to form the grid tensor corresponding to that grid. Simultaneously, the pile position status of each grid is marked: if a pile position has been placed at the center point of the grid, the pile position mark value of the grid is set to 1, otherwise it is set to 0; the pile position mark values ​​of all grids are arranged and spliced ​​according to the spatial row and column positions of each grid in the exploration area to construct an initial pile position matrix with the same spatial size as the grid tensor.

3. The automatic pile foundation layout method for complex geological foundations according to claim 1, characterized in that, The improved ResNet-18 CNN network is used to process the input raster tensor and output the geological determination parameters corresponding to each grid, specifically including: The improved ResNet-18 CNN network, based on the ResNet-18 backbone network, replaces the original fully connected layer at the back end with a 1×1 convolutional layer. This 1×1 convolutional layer compresses the feature map into three channels, which are used to output the predicted bearing capacity, the predicted probability of karst caves, and the predicted probability distribution of soil layer categories, respectively, serving as the geological determination parameters corresponding to the grid. The soil layer categories include clay, sand, and rock. The improved ResNet-18 CNN network can be trained in the following ways: The preprocessed raster tensors of historical exploration areas are collected as network input samples, and the bearing capacity, karst cave probability, and soil layer type of the corresponding grids, as determined by experts, are used as real labels. The network output is constrained using a multi-task loss function, which is obtained by weighted summation of the mean square error loss of bearing capacity prediction, the binary cross-entropy loss of karst cave probability prediction, and the cross-entropy loss of soil layer category prediction. The network parameters are updated using the backpropagation algorithm until the total loss value converges, thus completing the network training.

4. The automatic pile foundation layout method for complex geological foundations according to claim 1, characterized in that, The DQN network is used for reinforcement learning to output the optimal set of pile location coordinates and the corresponding pile length, total number of piles, and total cost. Specifically, this includes: The DQN network is initialized, consisting of a target network and a main network. The DQN network is then trained to generate the optimal set of pile location coordinates, along with the corresponding pile lengths, total number of piles, and total cost. At the start of each training round, the current geological assessment parameter map is obtained, and the constructed initial pile location matrix is ​​copied as the current pile location matrix. The two are then concatenated as the initial state. In each decision step, a greedy strategy is used to select the action: first, a random number between 0 and 1 is generated. If the random number is less than the current exploration rate, a pile location coordinate and pile length are randomly selected from the action space; otherwise, the current state is input into the main network, and the action corresponding to the maximum value among all Q values ​​output by the main network is selected. After each decision is made, the current exploration rate is gradually reduced according to a preset decay coefficient until it reaches a preset minimum value; When a selected action is executed, it is first determined whether the action is a termination action: if it is a termination action, the current round of training ends and the next round begins; if it is not a termination action, the corresponding position in the current pile position matrix is ​​marked as an existing pile position according to the coordinates and pile length specified in the action, and the total number of piles is increased by one. The specific termination conditions for determining whether an action is terminated are as follows: check whether the current round meets the preset termination conditions, including whether the current total number of piles has reached the preset upper limit, or whether the decrease ratio of the total cost of each round compared to the total cost of the previous round is lower than the preset ratio threshold; if any of the above termination conditions are met, the system will forcibly mark the current step as a terminated action. Then calculate the immediate reward for performing the selected action: Specifically, calculate the total cost before and after performing the action, which includes material cost, construction cost and risk cost. Subtract the total cost before performing the action from the total cost after performing the action, and take the opposite number as the immediate reward. That is, a positive reward is obtained when the total cost decreases due to the new pile placement, and a negative reward is obtained otherwise. After the selected action is completed, the next state is obtained based on the updated current pile position matrix, and it is determined whether the current wheel meets the termination condition. If the termination condition is met, the termination flag is set to true; otherwise, it is set to false. The five elements—current state, selected action, immediate reward, next state, and termination flag—are stored as an experience sample in the experience replay buffer. When the number of experience samples stored in the buffer is greater than the preset small batch size, the network parameters are updated after each environmental interaction step. The specific update process is as follows: a small batch of experience samples is randomly and uniformly sampled from the buffer, and the target Q value is calculated using the target network; if the termination flag of the sample is true, the target Q value is directly equal to the immediate reward in the sample; if the termination flag is false, the next state is input into the target network to obtain the Q values ​​of all actions in the state, and the maximum value is selected, multiplied by the discount factor, and then added to the immediate reward as the target Q value. The Q-value predicted by the main network for the current state and the selected action is compared with the target Q-value. The difference is calculated and the squared average is taken to obtain the loss value. The gradient of the loss function with respect to the parameters of the main network is calculated by the backpropagation algorithm. The parameters of the main network are updated by the Adam optimizer so that the Q-value predicted by the main network gradually approaches the target Q-value. After each fixed number of decision steps, the current parameters of the main network are completely copied to the target network. When the total cost fluctuation of consecutive preset rounds is less than the preset percentage threshold and the cumulative reward of consecutive preset rounds no longer increases, the DQN network is determined to be trained and the optimal set of pile coordinates, as well as the corresponding pile length, total number of piles, and total cost are output.

5. The automatic pile foundation layout method for complex geological foundations according to claim 1, characterized in that, The optimal pile location is determined and fine-tuned based on geological assessment parameters, including: Iterate through each optimal pile location in the set of optimal pile location coordinates and make a comprehensive judgment on the geological judgment parameters corresponding to its location. Specifically, if the probability of karst cave at the pile location is greater than the preset probability, or the bearing capacity is lower than the design requirement threshold, or the soil layer category belongs to the preset category that is not suitable for direct pile placement, then the pile location is determined not to meet the preset conditions, the pile location is deleted, and a new pile location is selected again within the preset radius of the current pile location using the polar coordinate generation method, until the new pile location meets all the preset conditions at the same time. If no pile location that meets all the conditions can be found within the preset radius, then the candidate point with the lowest probability of karst cave and the highest bearing capacity is selected. The rate of reduction in the probability of karst cave compared with the original pile location is calculated, and experts give a comprehensive score to determine whether to adopt the candidate point.

6. The automatic pile foundation layout method for complex geological foundations according to claim 5, characterized in that, The optimal pile location is reselected using the polar coordinate generation method, including: Mark the pile locations that are currently determined not to meet the preset conditions as high-risk pile locations; calculate the maximum search radius step number based on the grid side length and the preset maximum allowable search radius, and determine the number of angle partitions under each radius step number; Based on the maximum search radius step count and the number of angle partitions, a candidate point set is generated: in, The x-coordinate represents the current high-risk pile location; The vertical coordinate represents the current high-risk pile location; This indicates the number of search radius steps, expressed in units of grid side length; Indicates the angular partition index; This indicates rounding to the nearest integer. This represents the number of angular partitions with radius k. Represents the set of candidate points; Indicates the number of steps for the search radius. Angular partitioning is The increment of the x-coordinate of the candidate points at that time; Indicates the number of steps for the search radius. Angular partitioning is The increment of the ordinate of the candidate point at that time; Indicates the maximum search radius in steps; This indicates the preset maximum allowed search radius; Indicates the grid side length; Indicates rounding up; For each candidate point in the candidate point set, the following conditions are comprehensively determined: the probability of karst cave is less than the preset probability threshold, the bearing capacity is greater than the minimum design requirement, and the soil layer type does not belong to the preset unsuitable pile placement category; the candidate point that meets all the conditions at the same time is taken as the new pile location and the search is terminated; if multiple candidate points meet the conditions at the same time, the one with the lowest probability of karst cave and the highest bearing capacity is selected from them.

7. The automatic pile foundation layout method for complex geological foundations according to claim 1, characterized in that, A comprehensive verification of all adjusted pile positions was conducted to determine whether they met the preset engineering constraints, including: A comprehensive verification of all pile positions after S3 fine-tuning was performed to determine whether the following constraints were simultaneously met: The increase in total cost after adjustment compared to total cost before adjustment shall not exceed the preset allowable threshold, and the minimum distance between all piles shall not be less than the preset lower limit of pile distance. If all the above constraints are met, the final optimal pile location scheme will be directly output. If the conditions are not met simultaneously, the optimal pile location coordinates, corresponding pile lengths, total number of piles, and total cost directly output by the DQN network described in S2 are retained, and a warning message is output.

8. An automatic pile foundation layout system for complex geological foundations, wherein the automatic pile foundation layout system for complex geological foundations is used to implement the automatic pile foundation layout method for complex geological foundations as described in any one of claims 1-7, characterized in that, include: Data acquisition module: used to divide the exploration area into grids, set up boreholes for each grid, collect geological exploration parameters from each borehole, preprocess the geological exploration parameters to generate a raster tensor corresponding to each grid, and mark the pile positions of each grid to construct an initial pile position matrix; Reinforcement learning module: used to input the grid tensor of each grid into the trained improved ResNet-18CNN network, and the network outputs the geological determination parameters of each grid. The geological determination parameters of each grid are spliced ​​together according to their spatial location to generate a geological determination parameter map covering the entire exploration area. The geological determination parameter map and the initial pile position matrix are input into the trained DQN network for reinforcement learning. Through the multi-round decision-making of the DQN network, the optimal pile position coordinate set, as well as the pile length, total number of piles and total cost corresponding to each optimal pile position are output. The optimal pile location generation module is used to traverse each optimal pile location in the set of optimal pile location coordinates and make a judgment based on the geological judgment parameters corresponding to its location: if the preset safety and bearing capacity conditions are met, the pile location is retained; if not, a new pile location is selected again using the polar coordinate generation method within the preset range of the pile location, until the new pile location meets the preset conditions, thereby obtaining the fine-tuned pile location coordinates, the corresponding pile length, and the total cost. The verification module is used to verify all pile positions after fine-tuning in S3 and determine whether they meet the preset engineering constraints. If they do, the module outputs the final optimal pile position layout scheme. If they do not meet the constraints, the module retains the optimal pile position coordinates, corresponding pile lengths, total number of piles, and total cost output by the DQN network in S2, and outputs a warning.

Images