A method for establishing a refined model for the reuse of old piles

Abstract

This invention relates to the field of old pile reuse modeling technology, and discloses a method for establishing a refined old pile reuse model. The method acquires a set of detection data and a set of geological exploration data for the target old pile. The detection data set includes a text report of the old pile structure, image data of pile damage, and time-series data of pile bearing capacity. Semantic parsing is performed on the old pile structure text report to generate a damage semantic feature vector, and multi-scale degradation feature extraction is performed on the pile bearing capacity time-series data to generate a bearing capacity degradation feature vector. The two vectors are fused modally to generate a set of pile state features. A pre-configured geological parameter mapping model is called to encode the stratigraphic attributes of the geological exploration data set to generate a stratigraphic attribute feature vector. This vector is spatially correlated and fused with the set of pile state features to generate a fused geological pile feature matrix, thereby improving the accuracy and reliability of old pile reuse modeling.

Description

Technical Field

[0001] This invention relates to the field of old pile reuse modeling technology, specifically to a method for establishing a refined old pile reuse model. Background Technology

[0002] In current infrastructure upgrades and expansions, a large number of existing pile foundations face challenges in their disposal. With increasing emphasis on environmental protection and resource conservation, the reuse of old piles has gradually become a key focus in the industry. Compared to the traditional method of directly removing old piles and driving new ones, reuse can significantly reduce construction waste, lower project costs, and shorten the construction cycle. However, the compatibility of the pile's condition with the local geological environment is a crucial factor determining the feasibility and safety of reuse. Therefore, accurately constructing a reuse model for old piles has become a key issue hindering the large-scale promotion of this technology.

[0003] Traditional methods for assessing and modeling old piles have significant limitations. The detection and analysis of old pile conditions often rely on single types of data. For example, judging pile damage solely based on old pile structure reports, or assessing pile bearing capacity using only limited bearing capacity testing data, fails to fully integrate the value of multi-source testing data. While old pile structure reports contain information such as the location and type of pile defects, they are presented in natural language, making them difficult to use directly for model building. Pile damage image data can intuitively reflect the details of surface or internal damage, but it is often overlooked or used only as a qualitative reference, unable to be converted into quantitative features. Time-series data on pile bearing capacity records changes in the pile's bearing performance at different times; traditional methods often employ simple statistical analysis, failing to capture the multi-scale characteristics of bearing capacity degradation over time, leading to biases in the assessment of the pile's actual bearing capacity.

[0004] Traditional modeling methods often sever the connection between the condition of old piles and the geological environment. The bearing capacity and stability of old piles depend not only on their own structural condition but also on the geological properties of the surrounding area. The geomechanical properties, water content, and density of different strata directly affect the stress transfer patterns and long-term stability of old piles during reuse. In current technologies, the processing of geological exploration data is mostly limited to basic description, failing to effectively integrate the characteristics of geological strata with the characteristics of pile condition. This results in old pile reuse models that cannot accurately reflect the working state of the piles in the actual geological environment, leading to problems such as unreasonable reuse scheme design and increased engineering safety hazards.

[0005] Traditional modeling processes lack standardized feature extraction and fusion procedures. Different engineering teams often choose data processing methods based on experience, making it difficult to guarantee the consistency and reliability of the models. Even when some methods attempt to integrate multi-source data, they often use simple splicing or weighted superposition, failing to consider the modal differences and intrinsic relationships between different types of data. The fused features cannot accurately represent the compatibility between the overall state of the piles and the geological environment, ultimately affecting the scientific nature of decisions regarding the reuse of old piles and limiting the application effectiveness of old pile reuse technology in engineering practice. Summary of the Invention

[0006] The purpose of this invention is to provide a method for establishing a refined model of old pile reuse, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for establishing a refined model for the reuse of old piles, the method comprising:

[0008] Acquire the detection data set and geological exploration data set of the target old pile, wherein the detection data set includes the old pile structure text report, pile damage image data and pile foundation bearing capacity time series data;

[0009] Semantic parsing is performed on the old pile structure text report to generate a damage semantic feature vector. Multi-scale degradation feature extraction is performed on the pile foundation bearing capacity time series data to generate a bearing capacity degradation feature vector. The damage semantic feature vector and the bearing capacity degradation feature vector are then subjected to modal fusion processing to generate a pile state feature set.

[0010] The geological exploration data set is processed by calling a pre-configured geological parameter mapping model to encode the stratigraphic attributes, generating stratigraphic attribute feature vectors. The stratigraphic attribute feature vectors are then spatially correlated and fused with the pile body state feature set to generate a fused geological pile body feature matrix.

[0011] Preferably, the step of performing semantic parsing processing on the old pile structure text report to generate a damage semantic feature vector includes:

[0012] The damage description fragment segmentation process is performed on the old pile structure text report to obtain multiple damage description text units;

[0013] The pre-trained structural defect semantic model is invoked to perform context encoding on each of the damage description text units to generate an initial damage feature vector.

[0014] Match the set of defect entities associated with the current damage description text unit from the preset pile foundation defect knowledge base, input the set of defect entities into the structural defect semantic model for entity feature encoding processing, and generate entity defect feature vectors;

[0015] The entity defect feature vector and the initial damage feature vector are subjected to attention-weighted fusion processing to generate an enhanced damage feature vector.

[0016] The enhanced damage feature vectors of each damage description text unit are processed by sequence position encoding and concatenation to generate the damage semantic feature vector.

[0017] Preferably, the step of performing multi-scale degradation feature extraction processing on the time-series data of the pile foundation bearing capacity to generate a bearing capacity degradation feature vector includes:

[0018] Identify abnormal fluctuation intervals in the time series data of the pile foundation bearing capacity, and perform sliding window sampling processing within the abnormal fluctuation intervals to obtain multiple local bearing capacity segments;

[0019] The degradation mode extraction process is performed on each of the local bearing capacity segments by a pre-trained temporal convolutional network to generate a local degradation feature vector;

[0020] Calculate the pattern similarity score between each local degradation feature vector and the preset global degradation template;

[0021] Based on the pattern similarity score, each local degradation feature vector is dynamically weighted and aggregated to generate the bearing capacity degradation feature vector.

[0022] Preferably, the step of performing modal fusion processing on the damage semantic feature vector and the bearing capacity degradation feature vector to generate a pile state feature set includes:

[0023] The damage semantic feature vector is aligned to the time dimension and mapped to the same sampling timestamp sequence as the bearing capacity degradation feature vector;

[0024] Construct a damage-degradation cross-attention mechanism and calculate the correlation weight matrix between the semantic features of each timestamp in the damage semantic feature vector and the degradation features of the corresponding timestamp in the bearing capacity degradation feature vector;

[0025] Based on the correlation weight matrix, the damage semantic feature vector and the bearing capacity degradation feature vector are subjected to bidirectional feature interaction processing to generate interactive damage feature vector and interactive degradation feature vector.

[0026] The interactive damage feature vector and the interactive degradation feature vector are subjected to gated feature fusion processing to generate the pile state feature set.

[0027] Preferably, the step of calling a pre-configured geological parameter mapping model to perform stratigraphic attribute encoding processing on the geological exploration data set and generating stratigraphic attribute feature vectors includes:

[0028] The geological exploration data set is subjected to soil layer division processing to obtain the physical parameter data set of each soil layer;

[0029] The physical parameter data set is subjected to feature filtering processing through the sparse coding layer of the geological parameter mapping model to generate key soil layer feature vectors.

[0030] The high-order combination layer of the geological parameter mapping model is invoked to perform spatial correlation and combination processing on the key soil layer feature vectors to generate a layered geological feature matrix.

[0031] The layered geological feature matrix is ​​input into the aggregated output layer of the geological parameter mapping model for feature compression processing to generate the stratigraphic attribute feature vector.

[0032] Preferably, the step of spatially correlating and fusing the stratigraphic attribute feature vector with the pile state feature set to generate a fused geological pile feature matrix includes:

[0033] Spatial interpolation is performed on the stratum attribute feature vector based on the pile location coordinates to generate the pile location stratum attribute feature vector.

[0034] A pile-stratum graph convolutional network is constructed, and the set of pile state features is used as node features and the pile location stratum attribute feature vector is used as edge features. The input of the graph convolutional network is then used for neighborhood feature propagation processing.

[0035] The node update feature set is output through the multi-layer convolution operation of the graph convolutional network, and the node update feature set is matrix-reorganized to generate the fused geological pile feature matrix.

[0036] Preferably, the method further includes:

[0037] The pre-trained reuse scheme generation model is invoked to perform reuse parameter parsing processing on the fused geological pile feature matrix, generating a reuse parameter configuration set;

[0038] The preset pile foundation structure template is parametrically modified according to the reuse parameter configuration set to generate an initial reuse pile model.

[0039] Preferably, the step of calling the pre-trained reuse scheme generation model to perform reuse parameter parsing processing on the fused geological pile feature matrix to generate a reuse parameter configuration set, including:

[0040] The fused geological pile feature matrix is ​​input into the feature filtering layer of the reuse scheme generation model for redundancy filtering to obtain a set of core feature vectors.

[0041] The core feature vector set is subjected to multi-task parameter prediction processing through the parameter regression layer of the model generated by the reuse scheme, so as to generate pile diameter correction parameters, pile length correction parameters and material replacement parameters.

[0042] The pile diameter correction parameters, pile length correction parameters, and material replacement parameters are standardized to generate the reuse parameter configuration set.

[0043] Preferably, the method further includes:

[0044] Obtain a set of historical load case data, input the set of historical load case data into the initial reuse pile model for load response simulation processing, and generate simulated stress distribution data;

[0045] The simulated stress distribution data is processed to identify weak areas, generating a set of coordinates for high-stress areas;

[0046] Based on the coordinate set of the high-stress region, the initial reuse pile model is subjected to structural strengthening and correction processing to generate an optimized reuse pile model.

[0047] Preferably, the step of performing structural reinforcement correction processing on the initial reused pile model based on the coordinate set of the high-stress region to generate an optimized reused pile model includes:

[0048] Geometric data of the target region is extracted from the initial reused pile model based on the set of coordinates of the high-stress region.

[0049] The preset reinforcement rule base is invoked to perform reinforcement strategy matching processing on the geometric data of the target area, and a set of area reinforcement parameters is generated;

[0050] The set of regional reinforcement parameters is written into the corresponding geometric attribute fields of the initial reuse pile model to generate the optimized reuse pile model.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] This method achieves deep integration and efficient utilization of multi-source detection data in the modeling process of old pile reuse, effectively compensating for the shortcomings of traditional modeling methods that rely on single data, leading to evaluation bias. By performing semantic parsing on the text report of old pile structure, the damage information originally existing in natural language form can be transformed into standardized damage semantic feature vectors, enabling the accurate extraction of implicit damage details in the text and avoiding subjective errors that may occur when manually interpreting text in traditional methods. At the same time, multi-scale degradation feature extraction is carried out on the time series data of pile foundation bearing capacity, which can capture the changing patterns of bearing capacity from different time dimensions and fully present the degradation process of pile bearing performance. Compared with traditional simple statistical analysis, it can better reflect the actual bearing capacity state of the pile and provide more realistic pile performance data for subsequent modeling.

[0053] In the feature fusion stage, this method employs modal fusion technology to combine damage semantic feature vectors with bearing capacity degradation feature vectors. This fully considers the modal differences and intrinsic correlations between the two feature vectors. The resulting pile state feature set comprehensively characterizes the structural damage and bearing capacity status of the pile, avoiding the feature fragmentation problem caused by simple splicing in traditional data integration methods. This fusion approach makes the description of the pile state more comprehensive and accurate, laying a high-quality foundation of pile state data for subsequent modeling. It also helps to more accurately determine whether old piles are suitable for reuse and which reuse scenarios are appropriate.

[0054] This method also focuses on the correlation between the geological environment and the condition of existing piles. By calling a pre-configured geological parameter mapping model to encode the stratigraphic attributes of the geological exploration data set, the complex geological exploration data is transformed into standardized stratigraphic attribute feature vectors, achieving effective quantification of geological data. Based on this, spatial correlation fusion processing combines the stratigraphic attribute feature vectors with the pile condition feature set to generate a fused geological pile feature matrix. This fusion method fully reflects the influence of stratigraphic attributes on the working state of the piles, making the constructed old pile reuse model more closely aligned with the actual engineering geological environment. Compared to the traditional modeling approach that separates geology from pile condition, the model constructed by this method can more accurately predict the stress and long-term stability of old piles under actual geological conditions, reducing the problem of unreasonable reuse schemes due to neglecting geological factors and lowering engineering safety risks.

[0055] This method achieves standardized processing across the entire process, from data acquisition and feature extraction to feature fusion, avoiding the inconsistencies and unstable results caused by reliance on experience in traditional modeling. Whether it's semantic parsing, multi-scale degradation feature extraction, or stratigraphic attribute encoding and spatial correlation fusion, all follow pre-defined standardized processes and model parameters, ensuring consistency and reliability in old pile reuse modeling across different engineering scenarios. This standardization facilitates the widespread application of this method within the industry, promoting the transformation of old pile reuse technology from experience-driven to data-driven, facilitating technological upgrading and standardized development in the entire field of old pile reuse, further leveraging the advantages of old pile reuse in resource conservation, environmental protection and emission reduction, and cost control, and providing more scientific and efficient technical support for infrastructure renewal and expansion projects. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working principle of the refined old pile reuse modeling method described in this invention.

[0057] Figure 2 A flowchart of the semantic feature vector of damage;

[0058] Figure 3 A flowchart of the load-bearing capacity degradation feature vector;

[0059] Figure 4 This is a flowchart of the feature vectors of formation attributes. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1 This invention provides a method for establishing a refined model for the reuse of old piles, the method comprising:

[0062] The process involves acquiring a dataset of detection data and a dataset of geological exploration data for the target old piles. The detection data includes a text report of the old pile structure, image data of pile damage, and time-series data of pile bearing capacity. Semantic parsing is performed on the old pile structure text report to generate a damage semantic feature vector. Multi-scale degradation feature extraction is performed on the pile bearing capacity time-series data to generate a bearing capacity degradation feature vector. The damage semantic feature vector and the bearing capacity degradation feature vector are then fused modally to generate a set of pile state features. A pre-configured geological parameter mapping model is used to encode the stratigraphic attributes of the geological exploration data, generating a stratigraphic attribute feature vector. Finally, the stratigraphic attribute feature vector and the pile state feature set are spatially correlated and fused to generate a fused geological pile feature matrix.

[0063] Example 1: See Figure 2 This involves the specific process of performing semantic parsing on text reports of old pile structures to generate damage semantic feature vectors. The process begins with segmenting the acquired text reports of old pile structures into damage description fragments. These reports typically contain textual descriptions from inspectors of conditions such as surface cracks, concrete spalling, and steel reinforcement corrosion. These descriptions may be scattered across different chapters or paragraphs of the report. During processing, initial segmentation is performed based on natural paragraph divisions and descriptive punctuation (such as periods and semicolons) to form continuous semantic fragments. Subsequently, further refined segmentation is performed according to common damage description patterns in civil engineering. For example, a text describing "a longitudinal crack exists in the middle of the pile, approximately 1.2m long and 0.5-1mm wide; localized concrete spalling is visible 3m from the top of the pile, covering an area of ​​approximately 0.08m², with exposed and rusted steel reinforcement" will be segmented into two independent damage description text units, corresponding to the descriptions of the cracks and concrete spalling, respectively.

[0064] After obtaining multiple damage description text units, a pre-trained structural defect semantic model is invoked to perform context encoding on each unit. This model is a language model pre-trained based on a large number of civil engineering structural inspection reports and technical specification texts, capable of understanding professional terminology and semantic relationships in engineering contexts. The model maps the word sequence in each text unit into a high-dimensional vector representation, captures the contextual dependencies between words, and generates an initial damage feature vector. This vector represents the semantic information of the text description itself.

[0065] To further enhance the engineering semantic accuracy of the features, a set of defect entities associated with the current damage description text unit is matched from a pre-defined pile foundation defect knowledge base. This knowledge base constructs common defect types, causes, morphological characteristics, and interrelationships in pile foundation engineering. For example, for a text unit describing "exposed rebar with rust," the knowledge base may match entity concepts such as "rebar corrosion," "protective layer damage," and "electrochemical corrosion." These matched defect entity sets are then input into a structural defect semantic model for entity feature encoding. The model treats each entity as an independent semantic unit, combining its definition and associated attributes in the knowledge base to generate an entity defect feature vector. This vector supplements domain knowledge that may be implicit or not explicitly expressed in the text description.

[0066] The entity defect feature vector and the initial damage feature vector are fused using an attention-weighted process. This process calculates the attention weight of entity features for different words in the text description, dynamically emphasizing text parts highly relevant to entity knowledge. For example, in the description "partial concrete spalling, exposed rebar," the semantic importance of "exposed rebar" is amplified after receiving the entity knowledge of "damaged protective layer." Through operations such as weighted summation or concatenation, an enhanced damage feature vector is generated, which simultaneously integrates the original text semantics and external domain knowledge.

[0067] After performing the above processing on each damage description text unit, a set of enhanced damage feature vectors is obtained. These vectors need to be integrated into a single vector that comprehensively represents the damage status of the entire report. Sequence position encoding and concatenation processing is then performed. This process considers the order in which different damage descriptions appear in the report (which may implicitly suggest the logical order or importance of damage discovery), adding positional encoding information to each enhanced damage feature vector. Then, all positionally encoded vectors are sequentially concatenated into a high-dimensional long vector. Sometimes, dimensionality reduction or normalization operations are performed based on the vector dimension, ultimately generating a fixed-dimensional damage semantic feature vector. This serves as a numerical and structured representation of the damage information in the old pile structure text report for subsequent modal fusion processing. The entire semantic parsing process transforms the unstructured text report into a structured feature vector, achieving deep information extraction of the pile damage status.

[0068] Example 2: See Figure 3This involves performing multi-scale degradation feature extraction processing on time-series data of pile foundation bearing capacity to generate a bearing capacity degradation feature vector, and then performing modal fusion processing on this vector and a damage semantic feature vector to generate a set of pile state features. This process begins with the analysis of time-series data of pile foundation bearing capacity. This time-series data typically originates from pile foundation load test results recorded by a long-term monitoring system at different time points. Data points may include parameters such as timestamps, load values, and settlement. Identifying abnormal fluctuation intervals in the data sequence is a crucial step. These intervals manifest as sudden changes in the load-settlement curve, drastic numerical oscillations, or anomalies deviating from the long-term trend. For example, within a certain time period, if the settlement increases abnormally without a significant increase in load, or if the load value experiences an unexpected sharp drop, these areas are marked as abnormal fluctuation intervals.

[0069] Within each identified anomalous fluctuation interval, a sliding window sampling process is performed. The sliding window moves within the interval at a fixed time step, capturing a subsequence of time as a local bearing capacity segment at each movement. The window size and step size need to consider the data sampling frequency and fluctuation characteristics. For example, a 24-hour window with a 12-hour step can be used to capture daily variation patterns; or a smaller 1-hour window with a 30-minute step can be used to capture finer fluctuation details. Through this multi-scale sampling, multiple local segments reflecting bearing capacity changes at different time granularities are obtained.

[0070] A pre-trained temporal convolutional network is used to extract degradation patterns from each local bearing capacity segment. This network, trained on a large amount of historical pile foundation monitoring data, can identify various typical bearing capacity degradation patterns, such as slow decay, step-like decline, and abrupt changes. The network performs multi-layer convolutional operations on each input segment to extract its local temporal features and pattern information, outputting a fixed-dimensional local degradation feature vector. This vector encodes the specific degradation characteristics exhibited by that segment.

[0071] The similarity score between each local degradation feature vector and a predefined global degradation template is calculated. The global degradation template is a feature representation of a series of typical degradation patterns predefined based on engineering experience, such as uniform settlement templates, differential settlement templates, and brittle failure precursor templates. The similarity score uses distance metrics or correlation calculations in vector space to quantify the degree of matching between the current local pattern and each standard pattern. This score reflects the typicality of the current degradation fragment and its correlation with known patterns.

[0072] Based on pattern similarity scores, each local degradation feature vector is dynamically weighted and aggregated. Feature vectors with higher scores indicate that the degradation pattern they represent is more typical or important, and are given higher weights during the aggregation process; vectors with lower scores have lower weights. Through weighted averaging or weighted concatenation, all local feature vectors are merged into a comprehensive bearing capacity degradation feature vector. This vector comprehensively represents the pile foundation bearing capacity degradation status throughout the entire time series data from a multi-scale perspective.

[0073] The generated bearing capacity degradation feature vector and the damage semantic feature vector obtained from the text report are subjected to modal fusion processing. First, the damage semantic feature vector is aligned in the time dimension. Since the text report may correspond to detection records at multiple time points, its semantic features need to be mapped to the same sampling timestamp sequence as the bearing capacity data. For time points without corresponding text records, interpolation or nearest neighbor methods can be used to estimate their semantic feature values, forming a continuous semantic feature representation on the time series.

[0074] A damage-degradation cross-attention mechanism is constructed to calculate the correlation between features of two modalities. This mechanism calculates the correlation weight matrix between the semantic features of each timestamp in the damage semantic feature vector and the degradation features of the corresponding timestamp in the bearing capacity degradation feature vector. Specifically, for each time point, the attention score of the semantic feature vector and the degradation feature vector is calculated. This score reflects the correlation between the damage status described in the text and the measured change in bearing capacity at that time point. The attention weights are obtained through softmax normalization, forming the correlation weight matrix.

[0075] Based on the correlation weight matrix, a bidirectional feature interaction processing is performed on the damage semantic feature vector and the bearing capacity degradation feature vector. In the direction from semantics to degradation, the weight matrix is ​​used to weight the semantic features, emphasizing textual description information that is highly correlated with bearing capacity changes, generating an interactive damage feature vector. In the direction from degradation to semantics, the weight matrix is ​​used to weight the degradation features, highlighting bearing capacity change features corresponding to textual damage records, generating an interactive degradation feature vector. This bidirectional processing ensures that information from both modalities is fully interacted.

[0076] Interactive damage feature vectors and interactive degradation feature vectors are fused using a gated feature fusion process. The gating mechanism dynamically adjusts the fusion ratio based on the signal-to-noise ratio and importance of the two features. For time points with clear textual descriptions and high consistency with bearing capacity changes, semantic features are given higher weights; for time points with high-quality monitoring data but insufficient textual descriptions, degradation features are emphasized. Through adjustment of the gating function, a final pile state feature set is generated. This feature set integrates qualitative damage information from textual descriptions and quantitative change information from monitoring data, forming a multimodal unified representation of the pile state.

[0077] This approach enables a comprehensive assessment of pile foundation conditions, from identifying abnormal fluctuations in bearing capacity data to multi-scale feature extraction and deep interaction with textual semantic features. Ultimately, it generates a feature set that comprehensively reflects the pile's condition, providing an information foundation for subsequent geological correlation and reuse analysis. This processing method fully considers the characteristics and interrelationships of different data modalities, achieving effective information integration and utilization.

[0078] Example 3: See Figure 4 This involves calling a pre-configured geological parameter mapping model to encode the stratigraphic attributes of the geological exploration data set, generating stratigraphic attribute feature vectors, and then spatially fusing these vectors with the pile body state feature set to generate a fused geological pile body feature matrix. This process begins with the processing of the geological exploration data set, which typically includes borehole sampling records, geotechnical test results, geophysical exploration data, etc., containing physical and mechanical parameters of each soil layer such as depth, thickness, lithology, water content, void ratio, compression modulus, internal friction angle, and cohesion.

[0079] The geological exploration data set is processed by soil stratification. Based on borehole columnar sections and geotechnical engineering stratification standards, the soil within the exploration depth range is divided into several engineering geological layers, each considered a relatively homogeneous unit. For example, it may include fill layers, clay layers, silty clay layers, sand layers, and gravel layers. Each soil layer corresponds to a set of physical parameter data, including typical physical and mechanical properties of the soil in that layer. Feature filtering of the physical parameter data set is performed through a sparse coding layer of the geological parameter mapping model. This layer employs an autoencoder structure, which, through sparsity constraints, identifies key parameters that significantly affect pile foundation behavior from numerous soil layer parameters. For example, for friction piles, parameters related to side friction resistance may be given higher importance; for end-bearing piles, the strength parameters of the soil layer at the pile tip may be prioritized. The generated key soil layer feature vector focuses on the most relevant geotechnical parameters.

[0080] A high-order composite layer of the geological parameter mapping model is invoked to perform spatial correlation and combination processing on the feature vectors of key soil layers. This layer considers not only the characteristics of individual soil layers but also the mutual influence and combination effects between upper and lower soil layers. For example, the combination of a weak underlying layer and a hard overlying crust layer behaves differently from that of a homogeneous soil layer. This layer captures these interlayer interactions through feature interaction operations:

[0081]

[0082] in: This represents the interaction characteristics between the i-th and j-th soil layers. For activation function, These are learnable weight parameters. and For the characteristic transformation function, and These represent the k-th components of the key soil layer feature vectors of the i-th and j-th layers, respectively. This represents the feature interaction operator, where N is the dimension of the feature vector. Through this processing, the generated layered geological feature matrix not only contains independent information about each soil layer, but also combined information about the interactions between layers.

[0083] The layered geological feature matrix is ​​input into the aggregated output layer of the geological parameter mapping model for feature compression. This layer employs attention pooling or hierarchical aggregation methods to compress the feature matrix containing information from multiple soil layers into a fixed-dimensional stratigraphic attribute feature vector. The aggregation process considers the relative importance of different soil layers on pile foundation behavior; for example, shallow soil layers may have a higher weighting on friction piles than deep soil layers. The stratigraphic attribute feature vector is then spatially correlated and fused with the pile state feature set. First, spatial interpolation is performed on the stratigraphic attribute feature vector based on the pile location coordinates. Since geological exploration points are usually limited, the stratigraphic attributes at the pile location need to be estimated using methods such as Kriging interpolation or inverse distance weighting, based on the relative position of the pile location and the exploration points, to generate the pile location stratigraphic attribute feature vector.

[0084] A pile-stratum map convolutional network is constructed for feature fusion. The set of pile state features is used as node features, with each node corresponding to a pile unit or monitoring point. The pile location stratum attribute feature vector is used as edge features, representing the interaction between the pile and the stratum. The graph structure is constructed based on the spatial location of the pile and the distribution of strata, and connections are established between adjacent pile nodes and between pile nodes and stratum units.

[0085] Node and edge features are input into a graph convolutional network for neighborhood feature propagation. Each graph convolutional layer performs the following operations: nodes aggregate feature information from their neighboring nodes and update their features by combining the stratigraphic features of the connecting edges. Through multiple convolutional operations, the features of each node continuously fuse the state information of its neighboring nodes and related stratigraphic attribute information, achieving deep interaction between the pile state and geological conditions. The graph convolutional network outputs a set of updated node features. Each node's feature 'now' not only contains its own state information but also the influence of the surrounding pile state and stratigraphic attributes. These updated node features are then matrix-reorganized and rearranged into a two-dimensional feature matrix according to the spatial location or function of the piles, generating the final fused geological pile feature matrix. This matrix comprehensively reflects the multidimensional information of the pile's own state, the influence of neighboring piles, and the stratigraphic geological conditions.

[0086] Example 4: This involves calling a pre-trained reuse scheme generation model to perform reuse parameter parsing on the fused geological pile feature matrix, generating a reuse parameter configuration set, and then parametrically modifying the preset pile foundation structure template based on this set to generate an initial reuse pile model. This process begins by inputting the fused geological pile feature matrix into the feature filtering layer of the reuse scheme generation model. This layer employs a feature selection mechanism based on importance scoring, filtering out redundant or low-contribution feature dimensions by analyzing the correlation between features and pile foundation reuse performance indicators. For example, for friction pile reuse, features with low correlation to side friction resistance may be downweighted or eliminated; for end-bearing piles, more features related to the strength of the soil at the pile tip may be retained. After filtering, a core feature vector set is obtained, which retains the key information most influential on the pile foundation reuse design.

[0087] The model generates a reuse scheme and uses a parametric regression layer to perform multi-task parameter prediction on the core feature vector set. This layer employs a multi-task learning architecture, simultaneously predicting multiple reuse design parameters. The network output includes pile diameter correction parameters, pile length correction parameters, and material replacement parameters. The pile diameter correction parameters indicate the required increase or decrease in the original pile diameter; the pile length correction parameters reflect the adjustment amount required to lengthen or shorten the pile; and the material replacement parameters provide suggestions for strengthening or replacing the pile material, such as increasing the concrete strength grade or increasing the steel reinforcement ratio.

[0088] The predicted pile diameter correction parameters, pile length correction parameters, and material replacement parameters are standardized. This process discretizes continuous numerical parameters into commonly used engineering standard values, such as rounding the pile diameter correction value to the closest standard pile diameter series and adjusting the concrete strength grade to the grade specified in the code. Simultaneously, the material replacement scheme is coded and standardized to form a parameter expression form that conforms to engineering design standards. Finally, a structured set of reuse parameter configurations is generated.

[0089] The system performs parametric modification on the preset pile foundation structure template based on the reuse parameter configuration set. The pile foundation structure template is a parametric model built based on typical pile foundation types, containing modifiable parameter fields such as pile diameter, pile length, and material properties. The system reads the parameter values ​​from the parameter configuration set, automatically replaces the corresponding parameters in the template, and generates an initial reused pile model adapted to specific old pile conditions. The entire parametric modification process ensures that the model meets engineering specifications and is compatible with geological conditions and the current state of the piles.

[0090] Table 1: Examples of Reuse Parameter Configuration

[0091]

[0092] This implementation process utilizes a professional reuse scheme generation model to transform the feature matrix obtained from multi-source data fusion into specific and feasible engineering design parameters. The feature selection layer ensures that the analysis process focuses on the most critical influencing factors, while the multi-task prediction capability of the parameter regression layer guarantees the coordination and consistency of various reuse parameters. Parameter standardization ensures that the output results meet the requirements of actual engineering applications, while the parameter modification mechanism enables rapid conversion from parameters to specific design schemes. The generated initial reuse pile model considers both the existing state of the old piles and geological conditions, providing a reasonable foundation for subsequent optimization design. The entire process embodies an intelligent design method that combines data-driven design with engineering experience, providing scientific basis and technical support for the reuse of old piles.

[0093] Example 5: This involves acquiring a historical load case dataset, inputting this dataset into an initial reused pile model for load response simulation, and generating simulated stress distribution data. The historical load case data originates from long-term monitoring records of similar engineering projects in the area, containing actual stress conditions under various load combinations. This data typically includes information such as load magnitude, location, and duration under different load conditions, including static load, live load, wind load, and seismic action. The load response simulation employs the finite element method, discretizing the initial reused pile model into a unit mesh, applying various load conditions from the historical load data, and calculating the stress response of the pile under each condition. Numerical calculations yield the stress magnitude and distribution at various parts of the pile, forming detailed simulated stress distribution data.

[0094] Weak area identification processing is performed on simulated stress distribution data. This processing identifies potential high-stress areas by analyzing the statistical characteristics and spatial distribution patterns of the stress data. A combined approach of stress threshold analysis and gradient change detection is employed. First, a stress threshold based on material strength is set to mark areas exceeding the allowable stress level. Second, the spatial gradient of stress changes is analyzed to identify locations with high stress concentration. By combining these two methods, a set of coordinates for high-stress areas is generated, which accurately identifies structurally weak points within the pile that require special attention.

[0095] The initial reused pile model was structurally strengthened and corrected based on a coordinate set of high-stress areas. Geometric data of the target areas, including their dimensions, shapes, and spatial locations, were extracted from the initial model using the coordinate set. This geometric data reflects the specific morphological characteristics of the high-stress areas, providing a geometric basis for subsequent reinforcement strategy selection.

[0096] The system invokes a pre-defined reinforcement rule base to perform reinforcement strategy matching on the geometric data of the target area. This rule base is an expert knowledge system built upon extensive engineering practice and theoretical research, containing reinforcement solutions for various typical high-stress conditions. Based on the geometric characteristics of the high-stress area (such as area size, shape complexity, and spatial location) and stress characteristics (such as stress level and stress type), the rule base matches the most suitable reinforcement measures. For example, for areas with localized stress concentration, a solution that increases the local cross-sectional area might be matched; for areas with insufficient ductility, a solution that increases the reinforcement ratio might be matched; and for areas with weak shear resistance, a solution that increases the configuration of stirrups might be matched. Through this intelligent matching, a set of regional reinforcement parameters is generated for each high-stress area, containing detailed parameters such as specific reinforcement dimensions, materials, and construction details.

[0097] The set of regional reinforcement parameters is written into the corresponding geometric attribute fields of the initial reused pile model. This process is achieved through parametric modeling technology. The system automatically identifies the locations of areas requiring reinforcement and modifies the geometric and material properties of the model according to the reinforcement parameters. For example, for areas requiring larger cross-sections, the system automatically expands the cross-sectional dimensions of that area; for areas requiring enhanced reinforcement, the system modifies the rebar configuration parameters. The entire modification process maintains the overall coordination and geometric continuity of the model, ensuring a reasonable connection between the reinforcement measures and the original structure. Through this precise parametric modification, an optimized reused pile model is finally generated. This model not only retains the overall design concept of the initial scheme but also specifically strengthens the identified weak points, improving the overall structural safety and reliability.

[0098] This implementation process optimized and improved the initial reuse pile model through systematic load simulation, stress analysis, and targeted reinforcement. From inputting historical load data to calculating stress distribution, from identifying weak areas to matching reinforcement strategies, the process ultimately completed the parametric correction of the model, resulting in a safer and more reliable optimized design scheme. The entire process embodies the design optimization concept based on actual stress performance, providing more scientific and reliable technical support for the reuse of old piles.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for establishing a refined model for the reuse of old piles, characterized in that, The method includes: Acquire the detection data set and geological exploration data set of the target old pile, wherein the detection data set includes the old pile structure text report, pile body damage image data and pile foundation bearing capacity time series data; Semantic parsing is performed on the old pile structure text report to generate a damage semantic feature vector. Multi-scale degradation feature extraction is performed on the pile foundation bearing capacity time series data to generate a bearing capacity degradation feature vector. The damage semantic feature vector and the bearing capacity degradation feature vector are then subjected to modal fusion processing to generate a pile state feature set. The geological exploration data set is processed by calling a pre-configured geological parameter mapping model to encode the stratigraphic attributes, generating stratigraphic attribute feature vectors. The stratigraphic attribute feature vectors are then spatially correlated and fused with the pile state feature set to generate a fused geological pile feature matrix. The step of spatially correlating and fusing the stratigraphic attribute feature vector with the pile state feature set to generate a fused geological pile feature matrix includes: Spatial interpolation is performed on the stratum attribute feature vector based on the pile location coordinates to generate the pile location stratum attribute feature vector. A pile-stratum graph convolutional network is constructed, and the set of pile state features is used as node features and the pile location stratum attribute feature vector is used as edge features. The input of the graph convolutional network is then used for neighborhood feature propagation processing. The node update feature set is output through the multi-layer convolution operation of the graph convolutional network, and the node update feature set is matrix-reorganized to generate the fused geological pile feature matrix. The method further includes: The pre-trained reuse scheme generation model is invoked to perform reuse parameter parsing processing on the fused geological pile feature matrix, generating a reuse parameter configuration set; The preset pile foundation structure template is parametrically modified according to the reuse parameter configuration set to generate an initial reuse pile model.

2. The method for establishing a refined model for the reuse of old piles according to claim 1, characterized in that, The semantic parsing process performed on the old pile structure text report to generate a damage semantic feature vector includes: The damage description fragment segmentation process is performed on the old pile structure text report to obtain multiple damage description text units; The pre-trained structural defect semantic model is invoked to perform context encoding on each of the damage description text units to generate an initial damage feature vector. Match the set of defect entities associated with the current damage description text unit from the preset pile foundation defect knowledge base, input the set of defect entities into the structural defect semantic model for entity feature encoding processing, and generate entity defect feature vectors; The entity defect feature vector and the initial damage feature vector are subjected to attention-weighted fusion processing to generate an enhanced damage feature vector. The enhanced damage feature vectors of each damage description text unit are processed by sequence position encoding and concatenation to generate the damage semantic feature vector.

3. The method for establishing a refined model for the reuse of old piles according to claim 1, characterized in that, The process of performing multi-scale degradation feature extraction on the time-series data of the pile foundation bearing capacity to generate a bearing capacity degradation feature vector includes: Identify abnormal fluctuation intervals in the time series data of the pile foundation bearing capacity, and perform sliding window sampling processing within the abnormal fluctuation intervals to obtain multiple local bearing capacity segments; The degradation mode extraction process is performed on each of the local bearing capacity segments by a pre-trained temporal convolutional network to generate a local degradation feature vector; Calculate the pattern similarity score between each local degradation feature vector and the preset global degradation template; Based on the pattern similarity score, each local degradation feature vector is dynamically weighted and aggregated to generate the bearing capacity degradation feature vector.

4. The method for establishing a refined model for the reuse of old piles according to claim 1, characterized in that, The step of performing modal fusion processing on the damage semantic feature vector and the bearing capacity degradation feature vector to generate a pile state feature set includes: The damage semantic feature vector is aligned to the time dimension and mapped to the same sampling timestamp sequence as the bearing capacity degradation feature vector; Construct a damage-degradation cross-attention mechanism and calculate the correlation weight matrix between the semantic features of each timestamp in the damage semantic feature vector and the degradation features of the corresponding timestamp in the bearing capacity degradation feature vector; Based on the correlation weight matrix, the damage semantic feature vector and the bearing capacity degradation feature vector are subjected to bidirectional feature interaction processing to generate interactive damage feature vector and interactive degradation feature vector. The interactive damage feature vector and the interactive degradation feature vector are subjected to gated feature fusion processing to generate the pile state feature set.

5. The method for establishing a refined model for the reuse of old piles according to claim 1, characterized in that, The step of calling a pre-configured geological parameter mapping model to perform stratigraphic attribute encoding processing on the geological exploration data set, generating stratigraphic attribute feature vectors, includes: The geological exploration data set is subjected to soil layer division processing to obtain the physical parameter data set of each soil layer; The physical parameter data set is subjected to feature filtering processing through the sparse coding layer of the geological parameter mapping model to generate key soil layer feature vectors. The high-order combination layer of the geological parameter mapping model is invoked to perform spatial correlation and combination processing on the key soil layer feature vectors to generate a layered geological feature matrix. The layered geological feature matrix is ​​input into the aggregated output layer of the geological parameter mapping model for feature compression processing to generate the stratigraphic attribute feature vector.

6. The method for establishing a refined model for the reuse of old piles according to claim 1, characterized in that, The pre-trained reuse scheme generation model is invoked to perform reuse parameter parsing processing on the fused geological pile feature matrix, generating a reuse parameter configuration set, including: The fused geological pile feature matrix is ​​input into the feature filtering layer of the reuse scheme generation model for redundancy filtering to obtain a set of core feature vectors. The core feature vector set is subjected to multi-task parameter prediction processing through the parameter regression layer of the model generated by the reuse scheme, generating pile diameter correction parameters, pile length correction parameters and material replacement parameters. The pile diameter correction parameters, pile length correction parameters, and material replacement parameters are standardized to generate the reuse parameter configuration set.

7. The method for establishing a refined model for the reuse of old piles according to claim 1, characterized in that, The method further includes: Obtain a set of historical load case data, input the set of historical load case data into the initial reuse pile model for load response simulation processing, and generate simulated stress distribution data; The simulated stress distribution data is processed to identify weak areas, generating a set of coordinates for high-stress areas; Based on the coordinate set of the high-stress region, the initial reuse pile model is subjected to structural strengthening and correction processing to generate an optimized reuse pile model.

8. The method for establishing a refined model for the reuse of old piles according to claim 7, characterized in that, The process of performing structural reinforcement and correction on the initial reused pile model based on the coordinate set of the high-stress region to generate an optimized reused pile model includes: Geometric data of the target region is extracted from the initial reused pile model based on the set of coordinates of the high-stress region. The preset reinforcement rule base is invoked to perform reinforcement strategy matching processing on the geometric data of the target area, and a set of area reinforcement parameters is generated; The set of regional reinforcement parameters is written into the corresponding geometric attribute fields of the initial reuse pile model to generate the optimized reuse pile model.

Images