Pile sinking multi-dimensional index cooperative prediction method and system
The collaborative prediction method for multi-dimensional indicators of pile driving with construction mechanism consistency constraints solves the problem of fragmented modeling of global and sequential indicators in existing technologies, realizes high-precision and stability prediction of the pile driving construction process, adapts to complex geological conditions, and improves the reliability of construction decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- POWERCHINA HUADONG ENG CORP LTD
- Filing Date
- 2026-03-27
- Publication Date
- 2026-07-03
Smart Images
Figure CN122332893A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of building engineering and intelligent construction technology, and in particular to a method and system for collaborative prediction of multi-dimensional indicators of pile driving with construction mechanism consistency constraints. Background Technology
[0002] Pile driving is a crucial process in port engineering, bridge engineering, offshore wind power, and various infrastructure construction projects. Its construction quality and efficiency directly affect the pile foundation's bearing capacity, structural safety, and overall project cost. During pile driving, indicators such as pile stabilization readings, hammer readings, total hammer blows, and the number of blows per burst and energy per burst varying with pile depth are important parameters reflecting the construction status, ground response, and equipment effectiveness. Accurate prediction of these indicators is crucial for developing reasonable construction plans, reducing pile driving risks, and avoiding over- or under-driving piles.
[0003] Existing methods for predicting pile driving construction indicators mainly rely on empirical formulas, analogical analysis, or numerical simulation methods based on mechanical assumptions. On the one hand, these methods are highly dependent on the experience of engineers and are difficult to adapt to the complex and varied geological conditions in different regions; on the other hand, they often require idealization and simplification of soil constitutive relations, boundary conditions, and construction disturbance mechanisms, resulting in complex modeling, high computational costs, and difficulty in meeting the needs of rapid evaluation and batch application in actual engineering projects.
[0004] With the development of engineering informatization and sensing technology, a large amount of historical data has been accumulated during pile driving construction, including pile foundation geometric parameters, construction equipment parameters, and geological exploration data that varies with depth, such as CPTU data. Data-driven machine learning methods provide a new technical path for predicting key indicators in pile driving construction. However, pile driving construction is essentially a complex coupled engineering system involving multiple factors and the interaction between ground response and construction behavior, not a simple single-output regression problem. Existing data-driven methods still have limitations in their application in pile driving construction scenarios. Summary of the Invention
[0005] The purpose of this invention is to address the problem that existing data-driven models struggle to balance multi-source heterogeneous data fusion, stage-specific modeling, and consistency between global and sequential indicators. This invention proposes a collaborative prediction method and system for multi-dimensional indicators of pile driving with construction mechanism consistency constraints.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a method for synergistic prediction of multi-dimensional indicators for pile driving is provided, including the following steps: Feature extraction is performed based on the pile foundation parameters, equipment parameters, and geological CPTU data of the target pile driving project to obtain the corresponding basic feature vector of the target pile driving project; the basic feature vector includes static features corresponding to the pile foundation parameters and equipment parameters, as well as time-series features corresponding to the geological CPTU; Regression prediction is performed based on the aforementioned basic feature vector to obtain the corresponding global index prediction results. The global prediction results include single real number prediction results for the corresponding initial bearing capacity, stable pile body reading, pressure hammer pile body reading, and total number of hammer blows. Based on geological CPTU data, stage boundary points are identified, stage labels corresponding to each depth location are generated, and the corresponding stage label sequence is obtained. Based on the basic feature vector, global index prediction results and stage label sequence, the preset sequence prediction model outputs preliminary sequence prediction results indicating the number of hammer blows per burst and the energy of each burst at each depth position. Based on the preliminary sequence prediction results and the global index prediction results, consistency constraint correction is completed to obtain the final sequence prediction results.
[0007] This application can not only output global indicator prediction results and final sequence prediction results simultaneously, but also force the sequence cumulative results to maintain physical consistency with the global prediction results through subsequent consistency constraint correction steps. This fundamentally solves the problem of contradictory results caused by the separate modeling of the two types of indicators in the existing technology, and significantly improves the engineering credibility of the prediction results and the reliability of construction decisions.
[0008] As one possible implementation method, the static characteristics include one or more of the following: pile diameter, pile length, pile weight, pile bottom wall thickness, pile foundation penetration depth, hammer energy of the construction equipment, and hammer weight; The time-series characteristics include one or more of the mean, variance, skewness, kurtosis, and Fourier transform coefficients corresponding to the tip drag and side drag.
[0009] As one possible implementation, the features corresponding to the basic feature vector are the core features determined in advance based on feature selection.
[0010] As one possible implementation method, the single real numerical variables are pre-grouped based on physical correlation to obtain several variable groups. The single real numerical variables include initial bearing capacity, pile body reading of stable pile, pile body reading of pressure hammer, and total number of hammer blows. Feature filtering is performed on the sequence dimension variables to obtain the corresponding first core indicator set, where the sequence dimension variables are the prediction targets of the sequence prediction model; For single real-valued variables, feature filtering is performed based on variable groups to obtain a second set of core indicators that corresponds one-to-one with the corresponding variable groups. The basic feature vector includes a first input vector corresponding to the first core indicator set, and a second input vector corresponding one-to-one with each of the second core indicator sets; Regression prediction is performed on the corresponding variable groups based on each second input vector to obtain the corresponding global index prediction results; Based on the first input vector, the global indicator prediction results, and the stage label sequence, the preset sequence prediction model outputs the corresponding preliminary sequence prediction results.
[0011] As one possible implementation method, the specific steps for completing consistency constraint correction and obtaining the final sequence prediction result based on the preliminary sequence prediction results and global index prediction results are as follows: Based on the number of hammer blows per set at each depth location in the preliminary sequence prediction results, the cumulative predicted number of hammer blows is calculated; The consistency ratio coefficient is calculated by comparing the cumulative predicted hammer blows with the total hammer blows predicted by the global index. Based on the consistency ratio coefficient, the number of hammer blows per burst in the preliminary sequence prediction results is proportionally corrected to obtain the target number of hammer blows per burst. The corresponding target hammer energy per burst is determined based on the hammer energy of each burst in the preliminary sequence prediction results; The final sequence prediction results are generated based on the number of hammer blows per burst at each depth location and the hammer blow energy per burst at the target.
[0012] As one possible implementation, the sequence prediction model includes a shared feature extraction layer, a first prediction branch, and a second prediction branch; Wherein: the input of the first prediction branch is the common hidden layer features extracted by the shared feature extraction layer, and the output is the first original output sequence used to indicate the number of hammer blows per array at each depth position; The input to the second prediction branch is the common hidden layer features extracted by the shared feature extraction layer, and the output is a second raw output sequence used to indicate the energy of each hammer blow at each depth location.
[0013] As one possible implementation, the normalized basic feature vector / first input vector is concatenated with the global index prediction result to generate the input features of the sequence prediction model; The shared feature extraction layer maps the input features of the sequence prediction model to the common hidden space, extracts the common hidden layer features through a shared linear layer and ReLU activation function, and selectively activates the common hidden layer features based on the corresponding stage mask vector or stage control weight vector for the stage label corresponding to each depth position of the stage label sequence.
[0014] This application uses the global index prediction results as prior information to guide the sequence prediction process, establishing a causal link from macroscopic goals to microscopic processes. At the same time, through the stage masking mechanism, the model can automatically adjust the neuron activation mode according to different construction stages, realizing fine-grained stage-differentiated modeling and significantly improving the ability to characterize the pile driving process under complex geological conditions.
[0015] As one possible implementation method, a geological resistance curve is determined based on geological CPTU data, wherein the geological resistance curve is a cone tip resistance, lateral resistance, or a combination of both curves; The geological resistance curve is smoothed, and the first-order difference of the smoothed geological resistance curve with the depth of soil penetration is calculated to obtain the corresponding gradient change characteristics. The corresponding gradient change rate is obtained based on the gradient change characteristics of adjacent soil penetration depths; Gradient abrupt change points are identified based on the obtained gradient change rate, and stage boundary points are obtained by filtering based on preset segmentation rules. Based on the obtained stage boundary points, corresponding stage labels are generated for the depth positions corresponding to each soil penetration depth. in: Points where the gradient rate of change exceeds a preset rate of change threshold are considered gradient abrupt change points; Segmentation rules are used to indicate the rules for filtering and merging the resulting gradient mutation points to determine the stage boundary points.
[0016] The stage division method proposed in this application is driven entirely by objective geological data, avoiding the subjectivity of manual experience-based division. By combining gradient change rate threshold and segmentation rules, it can automatically adapt to the stage division requirements under different geological conditions, and the generated stage label sequence provides a reliable physical basis for subsequent stage perception and prediction.
[0017] Secondly, a multi-dimensional index collaborative prediction system for pile driving is provided, including: The feature extraction module is used to extract features based on the pile foundation parameters, equipment parameters, and geological CPTU data of the target pile driving project to obtain the corresponding basic feature vector of the target pile driving project. The basic feature vector includes static features corresponding to the pile foundation parameters and equipment parameters, as well as temporal features corresponding to the geological CPTU. The single real number prediction module is used to perform regression prediction based on the basic feature vector to obtain the corresponding global index prediction results. The global prediction results include the single real number prediction results for the corresponding initial bearing capacity, stable pile body reading, pressure hammer pile body reading, and total number of hammer blows. The stage identification module is used to identify stage boundary points based on geological CPTU data, generate stage labels corresponding to each depth location, and obtain the corresponding stage label sequence. The sequence prediction module is used to output the corresponding preliminary sequence prediction results based on the basic feature vector, global index prediction results and stage label sequence, using a preset sequence prediction model. The constraint calibration module is used to perform consistency constraint correction based on the preliminary sequence prediction results and the global index prediction results, and obtain the final sequence prediction results.
[0018] This invention, by adopting the above technical solutions, has significant technical effects: This invention introduces the global indicator prediction results into the sequence indicator prediction process, and establishes a construction mechanism connection between the global indicator and the sequence indicator at the prediction link level through consistency constraint correction. This avoids the problem of result separation caused by isolated modeling of the two and improves the physical rationality of the prediction results. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart illustrating the collaborative prediction method for multi-dimensional indicators of pile driving with construction mechanism consistency constraints according to the present invention. Detailed Implementation
[0021] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure.
[0022] Existing data-driven solutions have the following shortcomings in pile driving construction scenarios: Input data typically contains both static parameters with fixed dimensions and sequence parameters with significantly different lengths. It comes from diverse sources and has a heterogeneous structure, making it difficult to model uniformly. The prediction targets include both global indicators such as the total number of hammer blows and sequential indicators such as the number of hammer blows per burst and the energy of each burst, which vary with depth. Existing methods often model these two separately, failing to reflect the inherent construction mechanism relationship between the two types of indicators. The pile driving process has obvious stage differences. The construction response mechanism is different under different strata and different pile driving stages. However, existing models usually do not explicitly identify the construction stage, nor do they adjust the prediction mechanism according to the stage differences. For variable-length sequence outputs, existing methods often rely on complex recurrent networks or process them by means of forced truncation or uniform padding, which can easily introduce invalid information and affect prediction stability. Existing methods often lack constraints on the consistency between sequence prediction results and global prediction results, which may lead to discrepancies between the cumulative number of hammer blows per burst and the total number of hammer blows, thereby reducing engineering usability and result reliability.
[0023] Therefore, there is an urgent need for a pile driving construction prediction method that can integrate multi-source heterogeneous input data, explicitly consider the stage differences in the pile driving process, and perform collaborative prediction and consistency constraints on global and sequential indicators, so as to overcome the shortcomings of existing technologies in terms of accuracy, adaptability, mechanism consistency and engineering application.
[0024] To address the aforementioned issues, this application provides a collaborative prediction method for multi-dimensional indicators of pile driving with consistency constraints on construction mechanisms. This method no longer separates the global indicators and process sequence indicators of pile driving construction from each other. Instead, it first predicts the global key objectives, then uses the global prediction results as one of the inputs to the sequence model, and combines the stage identification results and consistency constraint mechanism to achieve collaborative prediction of multi-dimensional indicators. This method can be applied to engineering scenarios such as construction scheme design, construction parameter optimization, construction process control, and risk warning.
[0025] like Figure 1 As shown, the specific implementation process includes the following steps: S100. Based on the pile foundation parameters, equipment parameters and geological CPTU (cone penetration test) data of the target pile driving project, feature extraction is performed to obtain the corresponding basic feature vector of the target pile driving project; The basic feature vector includes static features corresponding to pile foundation parameters and equipment parameters, as well as time-series features corresponding to the geological CPTU; The static characteristics include one or more of the following: pile diameter, pile length, pile weight, pile bottom wall thickness, pile foundation penetration depth, hammer energy of the construction equipment, and hammer weight. The time-series characteristics include one or more of the following: mean, variance, skewness, kurtosis, and Fourier transform coefficients corresponding to the tip drag and side drag. Those skilled in the art can set the basic feature vectors themselves according to actual needs.
[0026] S200. Based on the aforementioned basic feature vector, perform regression prediction to obtain the corresponding global indicator prediction results; The global prediction results include single real number prediction results for the corresponding initial bearing capacity, stable pile body reading, pressure hammer pile body reading, and total number of hammer blows; In this embodiment, a pre-trained table-based model is selected as the core model for regression prediction; For example, the TabPFNv2 (Tabular Prior-Data Fitted Network v2) pre-trained table-based model can be used as the regression prediction model. This model is pre-trained on massive table datasets based on the Transformer architecture and has extremely strong few-shot learning and generalization capabilities, making it particularly suitable for the regression prediction requirements of fixed-dimensional features in this embodiment.
[0027] Those skilled in the art can set the number of models and training methods of the regression prediction model according to actual needs. They can train them separately using the "one indicator, one model" method or train them uniformly using the multi-output regression method. This specification does not impose detailed limitations on this.
[0028] S300: Based on geological CPTU data, identify stage boundary points, generate stage labels corresponding to each depth location, and obtain a stage label sequence; In this embodiment, the pile driving process can be divided into several stages, such as the easy-penetration stage, the transition stage, and the difficult-penetration stage. Those skilled in the art can adjust the number of stages and the corresponding segmentation rules according to the actual geological conditions of the project, such as manually setting the stage boundary points corresponding to each stage. In this embodiment, based on the gradient variation characteristics of geological parameters with depth, a corresponding stage label sequence is generated for each depth location to achieve staged division of the pile driving construction process. Specifically: Based on geological CPTU data, geological resistance curves (cone resistance c2, lateral resistance c3, or a combination of both) are determined, and the geological resistance curves are smoothed to eliminate the interference of data noise on stage identification. The first-order difference of the smoothed geological resistance curve with soil penetration depth is calculated to obtain the corresponding gradient variation characteristics. The corresponding gradient change rate is obtained based on the gradient change characteristics of adjacent soil penetration depths; Gradient abrupt change points are identified based on the obtained gradient change rate, and stage boundary points are obtained by filtering based on preset segmentation rules. Based on the obtained stage boundary points, corresponding construction stage labels are generated for each depth of soil penetration, forming a stage label sequence Pi. The specific expression is as follows: p i =[p i1 ,p i2 ,...,p iDepthi ]; Where, p ij This indicates the construction stage label corresponding to the j-th depth position of the pile driving project.
[0029] In this embodiment, points where the gradient rate of change exceeds a preset rate of change threshold are taken as gradient abrupt change points; Segmentation rules are used to indicate the rules for screening and merging the obtained gradient abrupt change points to determine the stage boundary points. After knowing the number of stages and geological conditions, those skilled in the art can set the corresponding segmentation rules to screen out the stage boundary points from the obtained gradient abrupt change points, so that the stage boundary points correspond to the divided construction stages, thereby identifying the construction stage corresponding to the location by recognizing the relationship between the soil penetration depth and the stage boundary points.
[0030] In addition, this step provides a basis for the stage division of the subsequent sequence prediction model. The sequence prediction model preferably adopts the L1 loss function based on the mask region to reduce the impact of outliers on the model training process.
[0031] S400: Based on the basic feature vector, global indicator prediction results and stage label sequence, the preset sequence prediction model outputs the corresponding preliminary sequence prediction results. The preliminary sequence prediction results are the prediction results of the number of hammer blows per burst and the energy of each burst corresponding to each soil penetration depth (depth position); The sequence prediction model includes a shared feature extraction layer, a first prediction branch, and a second prediction branch; in: The input to the first prediction branch is the common hidden layer features extracted by the shared feature extraction layer, and the output is the first raw output sequence used to indicate the number of hammer blows per array at each depth position; The input to the second prediction branch is the common hidden layer features extracted by the shared feature extraction layer, and the output is a second raw output sequence used to indicate the energy of each hammer blow at each depth location; The specific steps are as follows: S410. Construct input features for the sequence prediction model; The normalized basic feature vector X i Compared with the global indicator prediction results Concatenate the data to generate input features for a sequence prediction model. The specific expression is as follows: ; S420. Based on the shared feature extraction layer, stage-selective feature activation is completed; The shared feature extraction layer inputs features into the sequence prediction model. Mapping to the common hidden space, we extract the common hidden layer features H by using a shared linear layer and the ReLU activation function. The specific expression is as follows: ; Among them, Linear share To share linear transformation operations; D is the hidden layer dimension. .
[0032] In this embodiment, the stage label sequence P is targeted. i The stage label p corresponding to each depth position j ij Based on the corresponding stage mask vector or stage control weight vector M pij Selective activation is performed on the common hidden layer features, as shown in the following expression: ; in, This represents element-wise multiplication, which activates different groups of neurons according to different construction stages, thereby achieving stage difference modeling.
[0033] The stage mask vector or stage control weight vector is automatically obtained during the training phase.
[0034] S430. The first prediction branch generates a corresponding first original output sequence based on the obtained common hidden layer features, and the second prediction branch generates a corresponding second original output sequence based on the obtained common hidden layer features. In this embodiment, the activation functions for both the first and second prediction branches are LeakyReLU, and the corresponding calculation formula is as follows: ; ; Among them, O f2 For the first original output sequence, O f3 This is the second original output sequence. , where max-depth is the maximum length of the pile depth record corresponding to all samples during the training process. In order to achieve uniform adaptation to samples with different pile depths, it is used as the output dimension.
[0035] The model is pre-trained using relevant pile driving project samples. Separate independent linear layers are constructed for the number of hammer blows and the energy of each hammer blow. The common hidden layer features H are mapped to a sequence of fixed length (max-depth). Those skilled in the art, after understanding the model architecture and the corresponding inputs and outputs, can train the corresponding model on their own based on existing model training techniques. Therefore, it will not be described in detail.
[0036] S440. Mask the obtained first and second original output sequences to generate preliminary sequence prediction results. Due to the different actual pile driving depths (Depth) of different samples i To address discrepancies and remove invalid data from the fixed-length output, a depth mask matrix matching the actual pile driving depth is constructed: data within the effective depth range is set to 1, while data exceeding the depth is set to 0. i The part is set to 0.
[0037] The first and second original output sequences are masked based on the depth mask matrix to remove invalid depth prediction data. The number of hammer blows per array is then determined based on the masked result. And the energy of each hammer blow Generate corresponding preliminary sequence prediction results The specific expression is as follows: .
[0038] Note: During the training of the sequence prediction model, only valid sample points within the masked area are used to calculate the loss function to avoid interference from invalid padding values on the model training.
[0039] S500: Based on the preliminary sequence prediction results and the global index prediction results, complete the consistency constraint correction and obtain the final sequence prediction results; In this embodiment, the total number of hammer blows in the global index prediction result is used as a benchmark to perform mechanism consistency correction on the preliminary sequence prediction result, ensuring that the cumulative number of hammer blows in the sequence is consistent with the total number of hammer blows globally, and thus obtaining the final sequence prediction result. The specific steps are as follows: S510. Based on the number of hammer blows per burst at each depth position in the preliminary sequence prediction results, calculate the cumulative predicted hammer blow count. The specific calculation method is as follows: ; Among them, Depth i The actual pile driving depth record length for the i-th pile driving project (target pile driving project); This represents the number of hammer blows per set for the i-th pile driving project (target pile driving project) at the j-th depth position.
[0040] S520. Compare the cumulative predicted hammer blows with the total hammer blows predicted by the global index to calculate the consistency ratio coefficient. The specific calculation method is as follows: ; in, The total number of hammer blows is the prediction result of the global indicator. ε is the cumulative predicted number of hammer blows; ε is a very small constant to prevent the denominator from being zero, which can be set by those skilled in the art.
[0041] S530. Perform proportional correction on the number of hammer blows per burst in the preliminary sequence prediction results to obtain the target number of hammer blows per burst. In this embodiment, the number of hammer blows per burst in the preliminary sequence prediction results is constrained and adjusted based on the consistency ratio coefficient. The specific calculation method is as follows: ; in, This is the proportionality coefficient. This refers to the number of hammer blows per set for the i-th pile driving project (target pile driving project) at the j-th depth position in the preliminary sequence prediction results. The number of hammer blows per round for the corresponding target.
[0042] After correction, the mechanism of cumulative hammer blows per burst is matched with the total number of hammer blows worldwide, thereby improving the physical rationality and engineering interpretability of the output results.
[0043] S530. Determine the corresponding target hammer energy per hammer strike based on the hammer energy of each hammer strike in the preliminary sequence prediction results; The hammer energy of each hammer strike in the preliminary sequence prediction results can be directly used as the target hammer energy of each hammer strike. Alternatively, the hammer energy of each hammer strike in the preliminary sequence prediction results can be smoothed or the stage boundary continuity can be corrected according to the actual engineering needs to obtain the corresponding target hammer energy of each hammer strike.
[0044] S540. Based on the number of hammer blows per burst at each depth location and the target's hammer blow energy per burst, generate the corresponding final sequence prediction results; Right now: ; Among them, Depth i The actual pile driving depth recording length for the i-th pile driving project (target pile driving project).
[0045] In this embodiment, consistency constraint correction can be performed during the model inference stage or as a post-training processing step.
[0046] S600, Result Output; In this embodiment, the single real-value global indicators of the pile driving project (such as initial driving bearing capacity, total number of hammer blows, etc.) are output based on the global indicator prediction results, and the sequence indicator curves (such as number of hammer blows per burst, energy of hammer blows per burst) that change with depth are output based on the final sequence prediction results, providing data support for pile driving construction scheme design, construction parameter optimization, construction process control and risk warning.
[0047] As one possible implementation method, the features corresponding to the basic feature vector are the core features determined in advance based on feature selection from historical pile driving projects; The specific steps are as follows: S710. Extract the input and output data corresponding to historical pile driving projects; The specific definitions of input and output data for each historical pile driving project are as follows: The input data includes fixed-dimensional static variables. and sequence variables , specifically: Fixed-dimensional static variables This includes 7 static features, specifically [A1, A2, A3, A4, A5, B1, B2]. Where: A1 is the pile diameter, A2 is the pile length, A3 is the pile weight, A4 is the pile bottom wall thickness, and A5 is the pile depth; B1 is the hammer energy of the construction equipment, and B2 is the hammer weight. These are the aforementioned static features; this part of the data has a fixed dimension and no sequential characteristics.
[0048] Sequence variables That is, the geological CPTU (pore pressure static cone penetration test) parameter matrix; Among them, L i The length of the CPTU detection record for the i-th sample is typically between 1,000 and 10,000 sets. Each row of the matrix corresponds to a set of detection data, with three columns in total. Each column is: c1 (penetration depth, used as a sequence index), c2 (cone tip resistance), and c3 (lateral resistance).
[0049] The output data includes single real-valued variables. and sequence dimension variables , specifically: Single real-valued variable Specifically, [D2, E1, E2, E3]; Wherein, D2 is the initial bearing capacity, E1 is the reading of the stable pile body, E2 is the reading of the pressure hammer pile body, and E3 is the total number of hammer blows, which is the prediction result of the global index that will be regressed and predicted later. Here, it is the sample measured value.
[0050] Sequence dimension variables , which is the training target label of the sequence prediction model, and here it is the measured value; Among them, Depth i The actual pile driving depth recording length for the i-th sample; each row of the matrix corresponds to the pile driving data sampled at a specific interval, with two columns, each column being: f2 (number of hammer blows per set) and f3 (energy of hammer blows per set). In this embodiment, the sampling frequency at a specific interval is 1m. Those skilled in the art can also use other fixed interval sampling according to the actual engineering situation.
[0051] S720, Data Preprocessing, specifically: In this embodiment, a differentiated approach is used to complete missing value completion based on the different data characteristics of static variables and sequence variables. Simultaneously, the geological sequence data is truncated according to the target pile driving depth to remove invalid deep geological information, reduce noise interference, and lower the complexity of subsequent modeling, resulting in purified effective data. The specific steps are as follows: Handling missing values; For fixed-dimensional static variables For missing numerical variables, the mean is used to fill in the missing values; for sequence variables... For missing sequence dimension variables, interpolation using the mean of preceding and following terms was used to fill in the missing variables, in order to maintain the continuity of the geological sequence data.
[0052] Sequence truncation; Given that geological CPTU parameters beyond the target pile driving depth have no substantial guiding significance for the current task, the geological CPTU sequence variables are adjusted based on the target pile driving depth. The data is truncated (in this embodiment, it is truncated to 100 record points after the target depth), and only the data within the effective depth range related to the current pile driving task is retained. Invalid deep geological data is removed, thereby reducing noise interference and reducing modeling complexity.
[0053] S730. Based on the preprocessed feature data, perform feature selection to determine the corresponding core features; In this embodiment, variable-length sequence data and fixed-dimensional static data are fused into a feature vector of a unified dimension through feature extraction, concatenation, and selection, thus solving the problem of inconsistent modeling of multi-source heterogeneous data. The specific steps are as follows: S731, Sequence Feature Extraction; With sequence variables Using the soil penetration depth c1 as an index, an automated feature extraction algorithm is used to mine the temporal features of the cone tip resistance c2 and the side resistance c3. High-dimensional statistical features, including mean, variance, skewness, kurtosis, and Fourier transform coefficients, are extracted from the truncated CPTU sequence to generate a sequence feature matrix. , specifically: ; Among them, F raw This represents the initial number of sequence features extracted.
[0054] Preferably, this embodiment uses the Python open-source library tsfresh.
[0055] S732, Feature splicing; 7-dimensional fixed-dimensional static variables With sequence feature matrix By splicing the data, we obtain the initial features. This achieves the integration of engineering static parameters with geological sequence characteristics, specifically: ; Among them, F raw This represents the initial number of sequence features extracted.
[0056] S733, Feature Selection; To address the redundancy and multicollinearity issues in the initial features, a model-based feature selection method is employed. This method calculates the predictive correlation between features and the output variable, selecting the F core features that contribute the most to the prediction of the target variable, thus obtaining the corresponding basic feature vector X. i , specifically: ; In this embodiment, the model-based feature selection method used is sklearn.feature_selection.
[0057] The basic feature vector X i It includes both static engineering parameter information and geological sequence information, which will provide a unified input basis for subsequent global indicator prediction, stage identification and sequence indicator prediction.
[0058] Furthermore, during the model training phase, to prevent overfitting, the input features X of the training set are... i Applying Gaussian noise for data augmentation enriches the sample distribution and improves the robustness of the model.
[0059] In practical applications, the corresponding core features are directly extracted to construct the corresponding basic feature vectors.
[0060] As one possible implementation, single real-valued variables are pre-grouped, and optimization is performed based on the grouping results during the feature selection stage; This embodiment takes into account the physical correlation between indicators, groups single real-value variables to obtain several variable groups, and constructs a core indicator set corresponding to each variable group by combining the physical correlation of indicators in the feature selection stage. That is, the feature selection stage: Select the corresponding first core indicator set for the sequence dimension variable, and train the corresponding sequence prediction model; For a single real-valued variable, a second core indicator set that corresponds one-to-one with the corresponding variable group is selected, and the regression prediction model for the corresponding variable group is trained using the second core indicator set. For example, group optimization can be performed according to [D2], [E1, E2], [E3]; where D2 is the initial bearing capacity, E1 is the reading of the stable pile body, E2 is the reading of the pressure hammer pile body, and E3 is the total number of hammer blows.
[0061] That is, for variable groups [D2] and [E3], the corresponding second core indicator sets are used to train the model one indicator at a time. For variable groups [E1, E2], the corresponding second core indicator sets are used to train the model in a unified manner using a multi-output regression approach, so as to obtain the corresponding regression prediction sets that correspond one-to-one with the variable groups. In practical applications, the basic feature vectors include a first input vector corresponding to the first core indicator set and a second input vector corresponding to each of the second core indicator sets. Subsequently, regression prediction is performed on the corresponding variable groups based on each second input vector to obtain the corresponding global indicator prediction results. Based on the first input vector, the global indicator prediction results and the stage label sequence, the preset sequence prediction model outputs the corresponding preliminary sequence prediction results. As one possible implementation, the corresponding inputs and outputs are normalized when training the sequence prediction model; That is, for the corresponding input feature vector X i and target sequence dimension variables MinMax normalization is performed to map the data to the [0,1] interval, which improves the convergence speed and stability of model training.
[0062] This invention incorporates global index prediction results into the sequence index prediction process and uses consistency constraint correction to maintain a coordinated relationship between the cumulative number of hammer blows per burst and the total number of hammer blows. It establishes a construction mechanism connection between global and sequence indices at the prediction link level, avoiding the problem of fragmented results caused by isolated modeling and improving the physical rationality of the prediction results. Furthermore, this invention identifies stages of the gradient characteristics of geological parameters changing with depth and uses stage-controlled hidden layer weights or stage mask vectors in the sequence prediction model to differentiate between different stages. This better characterizes the stage differences in pile driving construction and improves the prediction accuracy and stability under complex geological conditions. Finally, this invention effectively solves the technical challenge of multi-source heterogeneous data fusion by extracting statistical features from variable-length CPTU sequences and constructing input feature vectors in conjunction with static variables. Simultaneously, the combination of fixed-length output and mask processing mechanisms improves the model's adaptability to variable-length pile driving sequence prediction tasks and reduces interference from invalid information.
[0063] Secondly, a multi-dimensional index collaborative prediction system for pile driving is provided, including: The feature extraction module is used to extract features based on the pile foundation parameters, equipment parameters, and geological CPTU data of the target pile driving project to obtain the corresponding basic feature vector of the target pile driving project. The basic feature vector includes static features corresponding to the pile foundation parameters and equipment parameters, as well as temporal features corresponding to the geological CPTU. The single real number prediction module is used to perform regression prediction based on the basic feature vector to obtain the corresponding global index prediction results. The global prediction results include the single real number prediction results for the corresponding initial bearing capacity, stable pile body reading, pressure hammer pile body reading, and total number of hammer blows. The stage identification module is used to identify stage boundary points based on geological CPTU data, generate stage labels corresponding to each depth location, and obtain the corresponding stage label sequence. The sequence prediction module is used to output the corresponding preliminary sequence prediction results based on the basic feature vector, global index prediction results and stage label sequence, using a preset sequence prediction model. The constraint calibration module is used to perform consistency constraint correction based on the preliminary sequence prediction results and the global index prediction results, and obtain the final sequence prediction results.
[0064] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.
[0065] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0066] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0067] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0068] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0069] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0070] It should be noted that: The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.
[0071] Furthermore, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.
[0072] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.
[0073] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.
Claims
1. A method for synergistic prediction of multi-dimensional indicators for pile driving, characterized in that, Includes the following steps: Feature extraction is performed based on the pile foundation parameters, equipment parameters, and geological CPTU data of the target pile driving project to obtain the corresponding basic feature vector of the target pile driving project; the basic feature vector includes static features corresponding to the pile foundation parameters and equipment parameters, as well as time-series features corresponding to the geological CPTU; Regression prediction is performed based on the aforementioned basic feature vector to obtain the corresponding global index prediction results. The global prediction results include single real number prediction results for the corresponding initial bearing capacity, stable pile body reading, pressure hammer pile body reading, and total number of hammer blows. Based on geological CPTU data, stage boundary points are identified, stage labels corresponding to each depth location are generated, and the corresponding stage label sequence is obtained. Based on the basic feature vector, global index prediction results and stage label sequence, the preset sequence prediction model outputs preliminary sequence prediction results indicating the number of hammer blows per burst and the energy of each burst at each depth position. Based on the preliminary sequence prediction results and the global index prediction results, consistency constraint correction is completed to obtain the final sequence prediction results.
2. The method for collaborative prediction of multi-dimensional indicators of pile driving according to claim 1, characterized in that: The static characteristics include one or more of the following: pile diameter, pile length, pile weight, pile bottom wall thickness, pile foundation penetration depth, hammer energy of the construction equipment, and hammer weight. The time-series characteristics include one or more of the mean, variance, skewness, kurtosis, and Fourier transform coefficients corresponding to the tip drag and side drag.
3. The method for synergistic prediction of multi-dimensional indicators for pile driving according to claim 1, characterized in that, The features corresponding to the basic feature vectors are the core features determined in advance based on feature selection.
4. The method for collaborative prediction of multi-dimensional indicators of pile driving according to claim 1, characterized in that: The individual real-value variables are pre-grouped based on physical correlation to obtain several variable groups. The individual real-value variables include initial bearing capacity, pile body reading of stable pile, pile body reading of pressure hammer, and total number of hammer blows. Feature filtering is performed on the sequence dimension variables to obtain the corresponding first core indicator set, where the sequence dimension variables are the prediction targets of the sequence prediction model; For single real-valued variables, feature filtering is performed based on variable groups to obtain a second set of core indicators that corresponds one-to-one with the corresponding variable groups. The basic feature vector includes a first input vector corresponding to the first core indicator set, and a second input vector corresponding one-to-one with each of the second core indicator sets; Regression prediction is performed on the corresponding variable groups based on each second input vector to obtain the corresponding global index prediction results; Based on the first input vector, the global indicator prediction results, and the stage label sequence, the preset sequence prediction model outputs the corresponding preliminary sequence prediction results.
5. A method for synergistic prediction of multi-dimensional indicators for pile driving according to any one of claims 1 to 4, characterized in that, Based on the preliminary sequence prediction results and the global index prediction results, the specific steps to complete the consistency constraint correction and obtain the final sequence prediction results are as follows: Based on the number of hammer blows per set at each depth location in the preliminary sequence prediction results, the cumulative predicted number of hammer blows is calculated; The consistency ratio coefficient is calculated by comparing the cumulative predicted hammer blows with the total hammer blows predicted by the global index. Based on the consistency ratio coefficient, the number of hammer blows per burst in the preliminary sequence prediction results is proportionally corrected to obtain the target number of hammer blows per burst. The corresponding target hammer energy per burst is determined based on the hammer energy of each burst in the preliminary sequence prediction results; The final sequence prediction results are generated based on the number of hammer blows per burst at each depth location and the hammer blow energy per burst at the target.
6. A method for synergistic prediction of multi-dimensional indicators for pile driving according to any one of claims 1 to 4, characterized in that: The sequence prediction model includes a shared feature extraction layer, a first prediction branch, and a second prediction branch; in: The input to the first prediction branch is the common hidden layer features extracted by the shared feature extraction layer, and the output is the first raw output sequence used to indicate the number of hammer blows per array at each depth position; The input to the second prediction branch is the common hidden layer features extracted by the shared feature extraction layer, and the output is a second raw output sequence used to indicate the energy of each hammer blow at each depth location.
7. The method for synergistic prediction of multi-dimensional indicators for pile driving according to claim 6, characterized in that: The normalized basic feature vector / first input vector is concatenated with the global index prediction result to generate the input features of the sequence prediction model; The shared feature extraction layer maps the input features of the sequence prediction model to a common hidden space, and extracts the features of the common hidden layer through a shared linear layer and the ReLU activation function; For each depth position of the stage label sequence, the common hidden layer features are selectively activated based on the corresponding stage mask vector or stage control weight vector.
8. A method for collaborative prediction of multi-dimensional indicators of pile driving according to any one of claims 1 to 4, characterized in that: Geological resistance curves are determined based on geological CPTU data, wherein the geological resistance curves are cone tip resistance, lateral resistance, or a combination of both. The geological resistance curve is smoothed, and the first-order difference of the smoothed geological resistance curve with the depth of soil penetration is calculated to obtain the corresponding gradient change characteristics. The corresponding gradient change rate is obtained based on the gradient change characteristics of adjacent soil penetration depths; Gradient abrupt change points are identified based on the obtained gradient change rate, and stage boundary points are obtained by filtering based on preset segmentation rules. Based on the obtained stage boundary points, corresponding stage labels are generated for the depth positions corresponding to each soil penetration depth. in: Points where the gradient rate of change exceeds a preset rate of change threshold are considered gradient abrupt change points; Segmentation rules are used to indicate the rules for filtering and merging the resulting gradient mutation points to determine the stage boundary points.
9. A multi-dimensional index collaborative prediction system for pile driving, characterized in that, include: The feature extraction module is used to extract features based on the pile foundation parameters, equipment parameters, and geological CPTU data of the target pile driving project to obtain the corresponding basic feature vector of the target pile driving project. The basic feature vector includes static features corresponding to the pile foundation parameters and equipment parameters, as well as temporal features corresponding to the geological CPTU. The single real number prediction module is used to perform regression prediction based on the basic feature vector to obtain the corresponding global index prediction results. The global prediction results include the single real number prediction results for the corresponding initial bearing capacity, stable pile body reading, pressure hammer pile body reading, and total number of hammer blows. The stage identification module is used to identify stage boundary points based on geological CPTU data, generate stage labels corresponding to each depth location, and obtain the corresponding stage label sequence. The sequence prediction module is used to output the corresponding preliminary sequence prediction results based on the basic feature vector, global index prediction results and stage label sequence, using a preset sequence prediction model. The constraint calibration module is used to perform consistency constraint correction based on the preliminary sequence prediction results and the global index prediction results, and obtain the final sequence prediction results.