A real-time stratum intelligent identification method based on sparse construction data of a rotary drilling rig

Abstract

The present application relates to the field of pile foundation construction, and specifically discloses a real-time stratum intelligent identification method based on sparse construction data of a rotary drilling rig, comprising the following steps: S1, collecting sparse construction data of a constructed pile hole and obtaining a stratum distribution label; S2, using a data preprocessing screening method to obtain construction data of a net drilling interval of the rotary pile; S3, using a data dynamic enhancement method to perform sliding window reconstruction on the construction data of the net drilling interval, constructing a stratum information matrix, and forming an enhanced data set; S4, inputting the enhanced data set into a real-time stratum intelligent identification model for training, stopping the training and storing model parameters after reaching a predetermined soil layer identification accuracy; and S5, collecting real-time construction data during the construction process, inputting the processed data into the trained model, and outputting a real-time identification result. The present application solves the problem of low sampling data interpretation difficulty in traditional construction relying on manual experience, provides a basis for construction parameter adjustment, and improves the intelligent level, quality and efficiency of pile foundation construction.

Description

Technical Field

[0001] This invention relates to the field of pile foundation construction, specifically to a real-time intelligent stratum identification method based on sparse construction data from rotary drilling rigs. Background Technology

[0002] As a crucial component of engineering construction, the quality and efficiency of pile foundation construction are decisive factors for project safety and operation and maintenance. Traditional rotary drilling piles are widely used in various engineering projects due to their convenient construction, high degree of mechanization, and high bearing capacity. Over the years, a mature and stable construction method has been developed.

[0003] However, certain limitations still exist in the actual construction process of rotary drilling piles. Firstly, the current hole-forming process of rotary drilling piles largely depends on the operator's skill level. As a large-scale equipment with highly integrated electromechanical and hydraulic systems, the precision and efficiency of its operation heavily rely on the operator's skill. This leads to inconsistent hole-forming quality and efficiency, and a lack of effective process control methods. Secondly, the adjustment of construction parameters during the hole-forming process heavily relies on the preliminary geological survey report. However, underground space exhibits high spatial variability, and relying solely on limited and sparse geological survey data makes it difficult to effectively adjust and optimize construction parameters. To achieve effective dynamic optimization and adjustment of construction parameters during the construction process, identifying the stratum type based on construction parameters is a necessary prerequisite for realizing intelligent pile foundation construction.

[0004] Traditional methods for identifying formations based on drilling parameters mainly fall into two categories. The first is theoretical methods, such as the specific work principle and net drilling speed determination. However, these methods have limited application scope and accuracy, making them unsuitable for everyday use. The second method uses deep learning to extract features from drilling parameters for formation type identification. However, this method heavily relies on high-precision, high-frequency data acquisition and multiple features. In actual field applications, the environment is complex, and the maintenance and acquisition of high-precision sensors are difficult, resulting in high acquisition difficulty, large data volume, and high computational complexity. Furthermore, in practical engineering applications, data from manufacturers and equipment is often confidential, providing only low-sampling-frequency data, leading to significant information loss and difficulty in parameter interpretation. Therefore, existing methods are insufficient for real-time formation identification applications under complex daily working conditions.

[0005] Therefore, there is an urgent need for a real-time intelligent geological identification method for rotary drilling rig construction data that can identify geological types using real-time measured construction data, achieving the goal of identification while construction is underway, so as to support the optimization of construction parameters and construction decisions. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention provides a real-time intelligent geological formation identification method based on sparse construction data from rotary drilling rigs. This method effectively solves the problem of heavy reliance on manual experience and lack of process quality control methods during rotary pile construction, significantly improving data richness and reducing the difficulty of data interpretation.

[0007] To achieve the above objectives, this invention proposes a real-time intelligent formation identification method based on sparse construction data from rotary drilling rigs, comprising: S1. Collect sparse construction data of the constructed pile holes, and simultaneously obtain the stratum distribution labels of the constructed pile holes; the construction data includes construction depth, drill bit rotation speed, torque, and vertical thrust. S2. Using data preprocessing and screening methods, with construction depth as the criterion, construction data of the net drilling range of rotary piles are obtained; S3. The construction data of the net drilling interval is reconstructed by sliding window using the data dynamic augmentation method to build a stratigraphic information matrix containing the stratigraphic sequence variation law and form an augmented dataset. S4. Input the enhanced dataset into the real-time intelligent soil identification model for training. After the predetermined soil layer identification accuracy is achieved, stop training and store the model parameters. S5. Collect real-time construction data during the construction process, and after automatic processing by the data preprocessing and filtering method described in step S2, input the data into the trained stratum identification model and output the real-time stratum identification results.

[0008] Preferably, in S1, the construction data is collected by means of the sensors built into the rotary drilling rig itself, or by adding corresponding sensors, or by means of the data collection function of the equipment manufacturer's online gimbal.

[0009] Preferably, in S1, the stratigraphic distribution labels are obtained by determining them through the geological report of the previous pile-by-pile exploration, through on-site geological drilling, or by observing the rotary drilling spoil in real time to determine the actual stratigraphic distribution labels at different depths of the construction pile holes.

[0010] Preferably, in S2, the calculation formula for the data preprocessing and filtering method is: ; ; In the formula, This is the original construction data sequence. The length of the overall construction sequence. For construction depth, Construction function for net drilling data. This represents net drilling data.

[0011] Preferably, in S3, the specific implementation of the data dynamic augmentation method is as follows: The length is The dynamic sliding window slides along the net drilling data in the depth direction, and the data within each window is recorded as an effective layer information matrix. Its corresponding tag is The real geological labels corresponding to the data entries are slid one step at a time until the entire data sequence is traversed, thus completing the construction of the augmented dataset.

[0012] Preferably, in step S4, the real-time intelligent formation identification model includes a data augmentation layer, a long short-term memory model layer, a sliding window attention mechanism layer, and a result output layer connected in sequence. The data augmentation layer is an automated implementation of the dynamic data augmentation method described in step S3, used to construct a parameter matrix of formation sequence information. The long short-term memory model layer is composed of a long short-term memory neural network, used to extract data features over a long time span of construction parameters. The sliding window attention mechanism layer is a structure combining the sliding window method and the attention mechanism, used for local feature extraction and weight calculation. The result output layer uses a fully connected neural network to summarize the data feature weight rules and output the predicted value of the formation type to be classified.

[0013] Preferably, the training parameters for the Long Short-Term Memory (LSTM) model layer are set as follows: the input dimension is... a, b, and c represent batch size, step size, and dimension, respectively. The batch size, step size, and dimension are adjusted according to the actual application scenario, and the number of hidden neurons in the module is set according to actual needs. After the Long Short-Term Memory model layer captures the features of the input augmented data, it outputs a feature vector. The output dimension is determined by adapting to the input dimension and the number of hidden neurons.

[0014] Preferably, the processing procedure of the sliding window attention mechanism layer is as follows: S31. Data padding is performed on the feature vector output by the Long Short-Term Memory model layer. The size of the time sliding window is adjusted according to the actual application scenario. The data dimension after padding is determined by adapting the input feature vector dimension and the sliding window size. S32. Calculate the attention weight for each position. The dimension of the attention weight is determined by adapting the dimension of the data after filling. S33. Calculate the context vector based on attention weights. The dimension of the context vector is consistent with the dimension of the feature vector output by the long short-term memory model layer.

[0015] Preferably, in S4, the specific process of model training also includes: The augmented dataset was divided into a training set and a validation set in a 7:3 ratio. The input dimension of the output layer is consistent with the context vector dimension, and the output dimension is the number of classification categories. The output of the last time step of the sequence is taken for classification to obtain the classification result, and the test set classification accuracy, precision, recall and F1 score are output.

[0016] Preferably, in S5, after outputting the real-time stratum identification results, the newly collected construction data and the corresponding stratum identification results are incorporated into the stratum identification model, and the model is optimized and updated in real time.

[0017] Therefore, this invention proposes a real-time intelligent formation identification method based on sparse construction data from rotary drilling rigs, the beneficial effects of which are as follows: (1) Improve the information richness of low sampling frequency data, and couple the dynamic changes of construction parameters with the actual geological conditions to improve data richness and reduce the difficulty of data interpretation.

[0018] (2) Effectively interpret the correspondence between construction data and stratum type characteristics, effectively identify stratum types in dynamic construction process, provide data and technical support for construction decision-making and construction parameter adjustment, and effectively solve the problem of heavy reliance on manual experience and lack of process quality control methods in rotary pile construction.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a data augmentation construction method for a real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs, according to the present invention. Figure 2 This is a diagram of the architecture of a real-time identification model for underground strata types based on sparse rotary drilling data, which is a real-time intelligent identification method for strata based on sparse rotary drilling data according to the present invention. Figure 3 This is a schematic diagram of the LSTM structure of a real-time intelligent formation identification method based on sparse rotary drilling rig construction data according to the present invention. Figure 4 This is a schematic diagram of the sliding window attention mechanism of a real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to the present invention; Figure 5 This is a flowchart of a real-time intelligent formation identification method based on sparse construction data from rotary drilling rigs, according to the present invention. Figure 6 This is an example diagram of the training output index of the identification model of the real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to the present invention; Figure 7 This is a visualization result of the formation identification of a real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to the present invention. Detailed Implementation

[0021] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figures 1-7 As shown, the present invention provides a real-time intelligent formation identification method based on sparse construction data from rotary drilling rigs, comprising: S1. Collect sparse construction data of the constructed pile holes, and simultaneously obtain the stratigraphic distribution labels of the constructed pile holes; the construction data includes construction depth, drill bit rotation speed, torque, and vertical thrust. Construction data is collected through sensors built into the rotary drilling rig itself, by adding corresponding sensors, or through the data collection function of the equipment manufacturer's online cloud platform.

[0024] The stratigraphic distribution labels are obtained by determining them through the geological report of the previous pile-by-pile exploration, through on-site geological drilling, or by observing the rotary drilling spoil in real time to determine the actual stratigraphic distribution labels at different depths of the construction pile holes.

[0025] S2. Using data preprocessing and screening methods, with construction depth as the criterion, construction data of the net drilling range of rotary piles are obtained; The calculation formula for the data preprocessing and filtering method is as follows: ; ; In the formula, This is the original construction data sequence. The length of the overall construction sequence. For construction depth, Construction function for net drilling data. This represents net drilling data.

[0026] S3. The construction data of the net drilling interval is reconstructed by sliding window using the data dynamic augmentation method to build a stratigraphic information matrix containing the stratigraphic sequence variation law and form an augmented dataset. The specific implementation of the data dynamic augmentation method is as follows: The length is The dynamic sliding window slides along the net drilling data in the depth direction, and the data within each window is recorded as an effective layer information matrix. Its corresponding tag is The real geological labels corresponding to the data entries are slid one step at a time until the entire data sequence is traversed, thus completing the construction of the augmented dataset.

[0027] S4. Input the enhanced dataset into the real-time intelligent soil identification model for training. After the predetermined soil layer identification accuracy is achieved, stop training and store the model parameters. The S4 real-time intelligent stratigraphic identification model consists of a data augmentation layer, a long short-term memory model layer, a sliding window attention mechanism layer, and a result output layer, which are connected in sequence.

[0028] The data augmentation layer is an automated implementation of the dynamic data augmentation method in step S3, and is used to construct the parameter matrix of the stratigraphic sequence information; The Long Short-Term Memory (LSTM) model layer is composed of a LSTM neural network and is used to extract data features of construction parameters over a long time span. The sliding window attention mechanism layer is a structure that combines the sliding window method with the attention mechanism, and is used for local feature extraction and weight calculation. The output layer uses a fully connected neural network to summarize the data feature weight patterns and output the predicted value of the stratum type to be classified.

[0029] The training parameters for the Long Short-Term Memory (LSTM) model layer are set as follows: the input dimension is... a, b, and c represent batch size, step size, and dimension, respectively. The batch size, step size, and dimension are adjusted according to the actual application scenario, and the number of hidden neurons in the module is set according to actual needs. After the Long Short-Term Memory model layer captures the features of the input augmented data, it outputs a feature vector. The output dimension is determined by adapting to the input dimension and the number of hidden neurons.

[0030] The processing procedure of the sliding window attention mechanism layer is as follows: S31. Data padding is performed on the feature vector output by the Long Short-Term Memory model layer. The size of the time sliding window is adjusted according to the actual application scenario. The data dimension after padding is determined by adapting the input feature vector dimension and the sliding window size. S32. Calculate the attention weight for each position. The dimension of the attention weight is determined by adapting the dimension of the data after filling. S33. Calculate the context vector based on attention weights. The dimension of the context vector is consistent with the dimension of the feature vector output by the long short-term memory model layer.

[0031] The specific process of model training also includes: The augmented dataset was divided into a training set and a validation set in a 7:3 ratio. The input dimension of the output layer is consistent with the context vector dimension, and the output dimension is the number of classification categories. The output of the last time step of the sequence is taken for classification to obtain the classification result, and the test set classification accuracy, precision, recall and F1 score are output.

[0032] S5. Collect real-time construction data during the construction process, and after automatic processing by the data preprocessing and filtering method described in step S2, input the data into the trained stratum identification model and output the real-time stratum identification results.

[0033] After outputting the real-time stratum identification results, the newly collected construction data and the corresponding stratum identification results are incorporated into the stratum identification model, and the model is optimized and updated in real time.

[0034] This invention uses a rotary drilling pile construction site for a high-speed railway in Shandong Province as the verification object. This site has complex geological conditions, encompassing various strata including silty clay, silt, fine sand, and silty sand. Furthermore, only low-frequency construction data can be obtained on-site, which closely matches the complex working conditions commonly encountered in actual engineering projects, making it suitable for verifying the accuracy and real-time performance of the method used in this invention for stratum identification. The specific implementation process is as follows: Site preparation and data collection: 25 completed pile holes in the construction site were selected as the training data source. The depth, torque, rotation speed and vertical thrust data of each pile hole were collected by the sensors on the rotary drilling rig during the construction process. The data sampling interval was 20 seconds. Combined with the site's one-pile-one-exploration geological report and on-site spoil observation, the real stratum labels corresponding to different depths of each pile hole were determined, and the original dataset was established.

[0035] Data preprocessing: A preprocessing method based on construction depth was adopted. Construction data for the net drilling range of each pile hole were obtained by filtering using a formula, and data interference from non-drilling stages such as drill lifting and hole cleaning was eliminated. The formula is as follows: ; ; In the formula, This is the original construction data sequence. The length of the overall construction sequence. For construction depth, Construction function for net drilling data. This represents net drilling data.

[0036] Perform "dynamic data augmentation" on the raw data: Set the length of the dynamic sliding window. The data in each window is slid along the depth direction of the net drilling data with a step size of 1. The data in each window is used to construct a formation information matrix. The formation label corresponding to the 8th data in the middle of the window is used as the label of the matrix. The augmented dataset is constructed, and a total of 7200 valid samples are obtained.

[0037] Model Training: The augmented dataset is input into the proposed stratum identification method based on rotary drilling rig construction data. A parameter matrix with a step size of 15 is input into a defined LSTM module for spatiotemporal data feature extraction, then into a sliding window attention mechanism module for dynamic feature calculation, and finally summarized into a fully connected neural network module. Specifically: The training set and validation set were divided into a 7:3 ratio. The parameters of the real-time intelligent stratigraphic identification model were configured as follows: the number of hidden neurons in the LSTM layer was set to 64, the time window size of the sliding window attention mechanism was set to 7, and the output dimension of the fully connected layer was consistent with the number of stratigraphic types (4 types) in the site.

[0038] The above information is summarized and the classification results are output, such as... Figure 6 The image shows the output of training indicators for a certain venue.

[0039] The training set is input into the model, and the training is optimized using the cross-entropy loss function. The output training results are then evaluated to determine if they meet the preset accuracy requirement, which is set to 85% in this example. If not, training continues; if they do meet the requirement, the model is saved and set as the identification model for this construction site. The training process outputs accuracy, precision, recall, and F1 score as evaluation metrics.

[0040] Real-time stratigraphic identification verification: A newly started pile hole within the site was selected as the test object. Real-time construction data was collected synchronously during construction, and after automated preprocessing and dynamic enhancement, it was input into the trained model. The model outputs stratigraphic identification results in real time and displays them on the operation screen, such as... Figure 7 As shown.

[0041] Results verification and model optimization: After construction, the model identification results were compared with the actual geological survey data of the pile hole to calculate the accuracy of stratum identification. The construction data and real labels of the pile hole were added to the dataset, and the model was iteratively optimized to improve the identification accuracy of subsequent construction.

[0042] Therefore, this invention provides a real-time intelligent stratum identification method based on sparse construction data from rotary drilling rigs. This method effectively eliminates the over-reliance on manual experience and limited geological survey reports in traditional rotary pile construction, solves the problem of difficult interpretation of low-sampling-frequency construction data, and achieves real-time accurate identification of stratum types during construction. This method provides a reliable basis for the dynamic optimization and adjustment of construction parameters, strengthens process quality control in pile foundation construction, reduces construction risks caused by deviations in stratum identification, improves construction efficiency and overall project safety, promotes the development of intelligent and precise pile foundation construction, adapts to the actual application needs of complex engineering conditions, and reduces uncertainty and resource waste during construction.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A real-time intelligent formation identification method based on sparse construction data from rotary drilling rigs, characterized in that, include: S1. Collect sparse construction data of the constructed pile holes, and simultaneously obtain the stratum distribution labels of the constructed pile holes; the construction data includes construction depth, drill bit rotation speed, torque, and vertical thrust. S2. Using data preprocessing and screening methods, with construction depth as the criterion, construction data of the net drilling range of rotary piles are obtained; S3. The construction data of the net drilling interval is reconstructed by sliding window using the data dynamic augmentation method to build a stratigraphic information matrix containing the stratigraphic sequence variation law and form an augmented dataset. S4. Input the enhanced dataset into the real-time intelligent soil identification model for training. After the predetermined soil layer identification accuracy is achieved, stop training and store the model parameters. S5. Collect real-time construction data during the construction process, and after automatic processing by the data preprocessing and filtering method described in step S2, input the data into the trained stratum identification model and output the real-time stratum identification results.

2. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 1, characterized in that, In S1, construction data is collected through sensors built into the rotary drilling rig itself, by adding corresponding sensors, or through the data collection function of the equipment manufacturer's online PTZ.

3. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 1, characterized in that, In S1, the stratigraphic distribution labels are obtained by determining them through the geological report of the previous pile-by-pile exploration, through on-site geological drilling, or by observing the rotary drilling spoil in real time to determine the actual stratigraphic distribution labels at different depths of the construction pile holes.

4. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 1, characterized in that, In S2, the calculation formula for the data preprocessing and filtering method is: ； ； In the formula, This is the original construction data sequence. The length of the overall construction sequence. For construction depth, Construction function for net drilling data. This represents net drilling data.

5. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 1, characterized in that, In S3, the specific implementation of the data dynamic augmentation method is as follows: The length is The dynamic sliding window slides along the net drilling data in the depth direction, and the data within each window is recorded as an effective layer information matrix. Its corresponding tag is The real geological labels corresponding to the data entries are slid one step at a time until the entire data sequence is traversed, thus completing the construction of the augmented dataset.

6. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 1, characterized in that, In S4, the real-time intelligent formation identification model includes a data augmentation layer, a long short-term memory model layer, a sliding window attention mechanism layer, and a result output layer connected in sequence. The data augmentation layer is an automated implementation of the data dynamic augmentation method described in step S3, used to construct a parameter matrix of formation sequence information. The long short-term memory model layer is composed of a long short-term memory neural network, used to extract data features over a long time span of construction parameters. The sliding window attention mechanism layer is a structure combining the sliding window method and the attention mechanism, used for local feature extraction and weight calculation. The result output layer uses a fully connected neural network to summarize the data feature weight rules and output the predicted value of the formation type to be classified.

7. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 6, characterized in that, The training parameters for the Long Short-Term Memory (LSTM) model layer are set as follows: the input dimension is... a, b, and c represent batch size, step size, and dimension, respectively. The batch size, step size, and dimension are adjusted according to the actual application scenario, and the number of hidden neurons in the module is set according to actual needs. After the Long Short-Term Memory model layer captures the features of the input augmented data, it outputs a feature vector. The output dimension is determined by adapting to the input dimension and the number of hidden neurons.

8. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 6, characterized in that, The processing procedure of the sliding window attention mechanism layer is as follows: S31. Data padding is performed on the feature vector output by the Long Short-Term Memory model layer. The size of the time sliding window is adjusted according to the actual application scenario. The data dimension after padding is determined by adapting the input feature vector dimension and the sliding window size. S32. Calculate the attention weight for each position. The dimension of the attention weight is determined by adapting the dimension of the data after filling. S33. Calculate the context vector based on attention weights. The dimension of the context vector is consistent with the dimension of the feature vector output by the long short-term memory model layer.

9. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 6, characterized in that, In S4, the specific process of model training also includes: The augmented dataset was divided into a training set and a validation set in a 7:3 ratio. The input dimension of the output layer is consistent with the context vector dimension, and the output dimension is the number of classification categories. The output of the last time step of the sequence is taken for classification to obtain the classification result, and the test set classification accuracy, precision, recall and F1 score are output.

10. The real-time intelligent formation identification method based on sparse construction data of rotary drilling rigs according to claim 1, characterized in that, In S5, after outputting the real-time stratum identification results, the newly collected construction data and the corresponding stratum identification results are incorporated into the stratum identification model to optimize and update the model in real time.

Images