Shield pile cutting stratum type identification method based on data and knowledge dual drive

Abstract

The invention relates to the technical field of tunnel and underground engineering, in particular to a shield pile cutting stratum type identification method based on data and knowledge dual driving. The method comprises the following steps: S1, installing a monitoring system on a shield tunneling machine; s2, processing original vibration and tunneling data obtained by the monitoring system on the shield tunneling machine; s3, knowledge driving of pile cutting stratum types is completed by dividing and classifying tasks; s4, performing a multi-modal classification task of pile cutting stratum types by adopting an MT model; and S5, taking the weighted accuracy as a model performance evaluation index. The shield tunneling parameters and the vibration signals are influenced by stratum types, meanwhile, an MT model combines the efficient classification capability of MLP and the strong feature capture capability of Transform, sequence data does not need to be generated in a classification task, the complexity and calculation overhead of the model are reduced, and an accurate classification result can be output.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnels and underground engineering, and in particular to a method for identifying shield pile cutting stratum types based on dual-drive of data and knowledge. Background Art

[0002] The rapid development of urban underground space has led to an increasing number of shield pile cutting projects, making pile cutting a cutting-edge technology. However, uncertainties in geological conditions and the location and number of pile foundations affect the control of shield tunneling parameters, building disturbance, and construction efficiency. Traditional methods rely on geological data and ground-penetrating radar to identify pile foundation location and number, but suffer from poor real-time performance and low accuracy. Existing research has largely focused on the dynamic process of shield pile cutting, lacking a real-time, accurate identification technology. Although deep learning has been widely applied to shield tunneling and can effectively identify stratum types with the advancement of computer technology, it remains difficult to meet the needs of actual pile cutting projects due to the closed nature of the pressure chamber, the lag in soil identification, and the heavy reliance on data accuracy. Furthermore, shield tunneling parameters and vibration signals are affected by stratum type, and while this can be used to predict stratum type, there is currently a lack of clear research on the vibration characteristics and feature inputs of shield pile cutting. Therefore, a real-time, high-precision intelligent identification method for stratum type and pile number for shield pile cutting is urgently needed. Summary of the Invention

[0003] The present invention provides a data- and knowledge-driven shield pile cutting stratum type identification method, which aims to solve the problems existing in the existing shield machine pile cutting dynamic parameter research.

[0004] The present invention provides a data- and knowledge-driven method for identifying stratum types for shield pile cutting, comprising the following steps: S1. Install the monitoring system on the shield machine; S2. Processing the raw vibration and tunneling data obtained by the monitoring system on the shield machine; S3. Complete knowledge-driven pile cutting stratum type by dividing and classifying tasks; S4. A hybrid model based on MLP and Transformer, the MT model, is used for multimodal classification of pile-cutting stratum types. S5. The weighted accuracy P w as an indicator for evaluating model performance.

[0005] As a further improvement of the present invention, the monitoring system includes three magnetoelectric vibration acceleration sensors, a signal acquisition instrument and a dynamic signal analysis system. Three test points are set on the shield machine, the second test point is located on the main bearing of the shield machine, the first test point and the third test point are located on both sides of the main bearing of the shield machine and are mirror-symmetrical. The three magnetoelectric vibration acceleration sensors are respectively installed on the three test points and collect the three-axis acceleration vibration signal of the shield machine. The magnetoelectric vibration acceleration sensors transmit the data to the signal acquisition instrument via wireless transmission, and the signal acquisition instrument uploads the data to the dynamic signal analysis system in real time via the wireless network.

[0006] As a further improvement of the present invention, the process of processing the original vibration and excavation data in step S2 specifically includes data cleaning, data combination, and data normalization.

[0007] As a further improvement of the present invention, the data cleaning is to clean up the invalid information contained: given the original excavation data R , traverse R To remove rows with zeros, we use the following formula: (1) in, Represents the data after data cleaning; F Indicates the action performed.

[0008] As a further improvement of the present invention, the data combination is to match vibration data with tunneling data: taking the X direction of shield machine tunneling as an example, the specific steps for each sampling period are: a1. The sampling frequency of the magnetoelectric vibration acceleration sensor is once per minute. The vibration data is processed and the calculation formula is: (2) (3) (4) in, is the peak value of the vibration; is the effective value of the vibration; is the time domain series of vibrations, ; N is the number of vibration signal sampling points; is the frequency; is the amplitude; a2. The tunneling data acquisition system of the shield machine collects data three times per minute, and processes the data to obtain tunneling data. The processing calculation formula is: (5) in, To obtain excavation data within a period at intervals of one minute D The average value of is the driving parameter sequence within one minute; m is the number of sampling points in one minute, and the number of sampling points is 3; a3. Processed vibration data 、 、 and tunneling data A set of training data is obtained by combining them every minute.

[0009] As a further improvement of the present invention, the data normalization is to adjust the feature data to the same scale: given feature data , the characteristic data include excavation parameters and vibration parameters, calculate the mean and standard deviation , the formula is as follows: (6) in, is the normalized value, is the original value, .

[0010] As a further improvement of the present invention, in step S3, the classification task is divided in the following manner: when the shield machine is about to reach the grouting layer or is in the grouting layer, the predicted pile cutting stratum type is divided into RGP3, including rock stratum, grouting layer, and grouting layer-pile composite stratum; when the shield machine is in the grouting layer-pile composite stratum, the predicted pile cutting stratum type is divided into RGP4 and RGP5, RGP4 includes rock stratum, grouting layer, 1 to 2 piles and 3, 4 or 5 piles, and RGP5 includes rock stratum, reinforcement area, 1 pile, 2 piles and 3, 4 or 5 piles.

[0011] As a further improvement of the present invention, in step S4, when the MT model performs a multimodal classification task, in the pile-cutting stratum type prediction task, the Transformer encoder obtains each position number of the pile-cutting stratum, and each position number corresponds to a position vector. By combining the position vector and the input vector of the training data as input features, position information is introduced into each input feature, and the input features are extracted and the global context information is modeled through the multi-head self-attention mechanism, and the extracted input features are output to the MLP classifier to complete the task classification.

[0012] As a further improvement of the present invention, in the encoder of the Transformer, a unique code is generated for each frequency in the input feature sequence and added to the frequency feature: the input feature is linearly transformed by the embedding layer and added to the position code to obtain the input matrix X, and the matrix X is input into the multi-head attention layer and linearly projected h times to obtain the query matrix Q, key matrix K and value matrix V, where h is the number of attention heads of the multi-head attention.

[0013] As a further improvement of the present invention, the model performance evaluation indicators include accuracy, precision, recall rate, F1-score and weighted accuracy P w , the calculation equation and formula are as follows: (7) (8) (9) (10) (11) Among them, TP represents true positive, FP represents false positive, TN represents true negative, and FN represents false negative; is a category i The weight is determined by the importance of the construction and is adjusted according to the construction task. The sum of the weights is 1; is a category i The classification accuracy of .

[0014] The beneficial effects of the present invention are: utilizing the influence of stratum type on shield tunneling parameters and vibration signals, and the MT model combines the efficient classification capability of MLP with the strong feature capture capability of Transformer. It does not need to generate sequence data in the classification task. Compared with other models, it has excellent comprehensive performance and more flexible application scenarios, can reduce the complexity and computational overhead of the model, reduce training time, and can output accurate classification results, thereby realizing intelligent identification of stratum type and number of piles cut, and ensuring the safe excavation of shield pile cutting projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a shield pile cutting stratum type identification method based on dual-drive of data and knowledge of the present invention; Figure 2 This is the overall structure diagram of the TM model of the shield pile cutting stratum type identification method based on data and knowledge dual drive of the present invention. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0017] like Figure 1 As shown, a data and knowledge-driven shield pile cutting stratum type identification method of the present invention includes the following steps: S1. Install the monitoring system on the shield machine; S2. Processing the raw vibration and tunneling data obtained by the monitoring system on the shield machine; S3. Complete knowledge-driven pile cutting stratum type by dividing and classifying tasks; In order to avoid inaccurate data, input parameters and sensitivity analysis can be performed after dividing the classification tasks. The selection of input parameters mainly relies on experience and is selected based on data analysis. Sensitivity analysis systematically changes one or more input parameters and observes the extent of change in the output results, thereby identifying the key factors that have the greatest impact on the results and helping decision makers understand the stability, uncertainty and potential risks of the model.

[0018] S4. A hybrid model based on MLP and Transformer, the MT model, is used for multimodal classification of pile-cutting stratum types. In the MT model, the MLP is connected to the output of the Transformer, meaning that data is first encoded in the Transformer and then classified in the MLP. S5. The weighted accuracy P w as an indicator for evaluating model performance.

[0019] Specifically, the monitoring system includes three iVS101A magnetoelectric vibration accelerometers, a signal acquisition instrument and a dynamic signal analysis system. Three test points are set on the shield machine. The second test point is located on the main bearing of the shield machine. The first test point and the third test point are located on both sides of the main bearing of the shield machine and are mirror-symmetrical. The bases of the three magnetoelectric vibration accelerometers are adsorbed on the three test points on the shield through magnets and collect the three-axis acceleration vibration signals of the shield machine. The magnetoelectric vibration accelerometer transmits the data to the signal acquisition instrument via wireless transmission. The signal acquisition instrument uploads the data to the cloud database of the dynamic signal analysis system in real time through the wireless network in the shield machine.

[0020] like Figure 2 As shown, the process of processing the original vibration and excavation data in step S2 specifically includes data cleaning, data combination, and data normalization.

[0021] Data cleaning is to clean up the invalid information contained: given the original excavation data R , traverse R To remove rows with zeros, we use the following formula: (1) in, Represents the data after data cleaning; F Indicates the action performed.

[0022] Data combination is to match vibration data with tunneling data: Taking the X direction of shield machine tunneling as an example, the specific steps for each sampling period are as follows: a1. The sampling frequency of the magnetoelectric vibration acceleration sensor is once per minute. The vibration data is processed and the calculation formula is: (2) (3) (4) in, is the peak value of the vibration; is the effective value of the vibration; is the time domain series of vibrations, ; N is the number of sampling points of the vibration signal, that is, the number of samples in each sampling period; is the frequency; is the amplitude; the number of sampling points during the monitoring period is 1024.

[0023] a2. The tunneling data acquisition system of the shield machine collects data three times per minute and processes the data to obtain tunneling data. The processing calculation formula is: (5) in, To obtain excavation data within a period at intervals of one minute D The average value of is the driving parameter sequence within one minute; m is the number of sampling points in one minute, that is, the number of sampling points of the shield machine's own data acquisition system in one minute, and the number of sampling points is 3; a3. Processed vibration data 、 、 and tunneling data A set of training data is obtained by combining them every minute.

[0024] Data normalization is to adjust the feature data to the same scale: given feature data , the characteristic data include excavation parameters and vibration parameters, calculate the mean and standard deviation , the formula is as follows: (6) in, is the normalized value, is the original value, .

[0025] This invention, based on the shield machine's dynamic response monitoring system and the position encoding technology employed by the MT model, effectively identifies the positional sequence of input parameters and uses the top five frequencies with the largest corresponding amplitudes as one of the inputs, thereby improving recognition accuracy. This facilitates breakthroughs in both stratum type identification and pile cutting quantity prediction, providing a new technical paradigm for intelligent decision-making in shield construction under complex geological conditions. This invention is highly practical and can be applied not only to shield pile cutting projects but also as a reference for shield tunneling through underground structures such as diaphragm walls and anchor cables.

[0026] In step S3, the classification task is divided into the following ways: when the shield machine is about to reach the grouting layer or is in the grouting layer, the predicted pile cutting formation type is classified as RGP3, which includes rock layer, grouting layer, and grouting layer-pile composite formation; when the shield machine is in the grouting layer-pile composite formation, the predicted pile cutting formation type is classified as RGP4 and RGP5. RGP4 includes rock layer, grouting layer, 1-2 piles, and 3, 4, or 5 piles; RGP5 includes rock layer, reinforced area, 1 pile, 2 piles, and 3, 4, or 5 piles. RGP stands for rock layer, grouting layer, and pile.

[0027] In step S4, the MT model discards the decoder. When performing multimodal classification tasks, for example, in the pile-cutting stratum type prediction task, the Transformer's encoder extracts input features and models global context information through a multi-head self-attention mechanism, while the subsequent MLP classifier completes the classification. Specifically, the encoder obtains each position number of the pile-cutting stratum, each corresponding to a position vector. By combining the position vector with the input vector of a set of training data (including shield machine vibration and tunneling training data) as input features, position information is introduced into each input feature. The multi-head attention mechanism can then distinguish data at different locations.

[0028] The MT model uses positional encoding technology. The Transformer encoder generates a unique code for each frequency in the input sequence and adds it to the frequency features. Specifically, the input features undergo a linear transformation in the embedding layer and are then added to the positional encoding to produce the input matrix X. This matrix X is then fed into the multi-head attention layer and linearly projected h times to produce the query matrix Q, key matrix K, and value matrix V, where h is the number of attention heads in the multi-head attention layer.

[0029] Model performance evaluation indicators include accuracy, precision, recall, F1-score and weighted accuracy P w , the calculation equation and formula are as follows: (7) (8) (9) (10) (11) Among them, TP represents true positive, FP represents false positive, TN represents true negative, and FN represents false negative; is a category i The weight is determined by the importance of the construction and is adjusted according to the construction task. The sum of the weights is 1; is a category i The classification accuracy of .

[0030] By integrating multiple model performance evaluation indicators, the weighted accuracy rate, which is more suitable for engineering, is adopted as the optimal model performance evaluation indicator.

[0031] This paper installs a monitoring system on the shield machine and uses the influence of stratum type on shield tunneling parameters and vibration signals to develop a MT model to identify the stratum type of shield pile cutting. The developed MT model combines the efficient classification capabilities of the MLP with the strong feature capture capabilities of the Transformer. It does not require the generation of sequence data for classification tasks. Compared with other models, it has superior overall performance and more flexible application scenarios. It can reduce model complexity and computational overhead, shorten training time, and output accurate classification results, thus achieving intelligent identification of stratum type and number of piles to be cut, ensuring the safe excavation of shield pile cutting projects.

[0032] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A data and knowledge-driven shield pile cutting stratum type identification method, characterized in that: The following steps are involved: S1. Install the monitoring system on the shield machine; S2. Processing the raw vibration and tunneling data obtained by the monitoring system on the shield machine; S3. Complete knowledge-driven pile cutting stratum type by dividing and classifying tasks; S4. A hybrid model based on MLP and Transformer, the MT model, is used for multimodal classification of pile-cutting stratum types. S5. The weighted accuracy P w as an indicator for evaluating model performance.

2. The shield pile cutting stratum type identification method based on dual-driven data and knowledge according to claim 1 is characterized in that: The monitoring system includes three magnetoelectric vibration acceleration sensors, a signal acquisition instrument and a dynamic signal analysis system. Three test points are set on the shield machine. The second test point is located on the main bearing of the shield machine. The first test point and the third test point are located on both sides of the main bearing of the shield machine and are mirror-symmetrical. The three magnetoelectric vibration acceleration sensors are respectively installed on the three test points and collect the three-axis acceleration vibration signal of the shield machine. The magnetoelectric vibration acceleration sensors transmit the data to the signal acquisition instrument via wireless transmission. The signal acquisition instrument uploads the data to the dynamic signal analysis system in real time via the wireless network.

3. The shield pile cutting stratum type identification method based on dual-driven data and knowledge according to claim 1 is characterized in that: The process of processing the original vibration and excavation data in step S2 specifically includes data cleaning, data combination, and data normalization.

4. The shield pile cutting stratum type identification method based on dual-driven data and knowledge according to claim 3 is characterized in that: The data cleaning is to clean up the invalid information contained: given the original excavation data R , traverse R To remove rows with zeros, we use the following formula: (1) in, Represents the data after data cleaning; F Indicates the action performed.

5. The shield pile cutting stratum type identification method based on dual-driven data and knowledge according to claim 3 is characterized in that: The data combination is to match vibration data with tunneling data: taking the X direction of shield machine tunneling as an example, the specific steps for each sampling period are: a1. The sampling frequency of the magnetoelectric vibration acceleration sensor is once per minute. The vibration data is processed and the calculation formula is: (2) (3) (4) in, is the peak value of the vibration; is the effective value of the vibration; is the time domain series of vibrations, ; N is the number of vibration signal sampling points; is the frequency; is the amplitude; a2. The tunneling data acquisition system of the shield machine collects data three times per minute, and processes the data to obtain tunneling data. The processing calculation formula is: (5) in, To obtain excavation data within a period at intervals of one minute D The average value of is the driving parameter sequence within one minute; m is the number of sampling points in one minute, and the number of sampling points is 3; a3. Processed vibration data 、 、 and tunneling data A set of training data is obtained by combining them every minute.

6. The shield pile cutting stratum type identification method based on dual drive of data and knowledge according to claim 3 is characterized in that: The data normalization is to adjust the feature data to the same scale: given feature data , the characteristic data include excavation parameters and vibration parameters, calculate the mean and standard deviation , the formula is as follows: (6) in, is the normalized value, is the original value, .

7. The shield pile cutting stratum type identification method based on dual drive of data and knowledge according to claim 1 is characterized in that: In step S3, the classification task is divided in the following manner: when the shield machine is about to reach the grouting layer or is in the grouting layer, the predicted pile cutting stratum type is divided into RGP3, which includes the rock layer, the grouting layer, and the grouting layer-pile composite stratum; when the shield machine is in the grouting layer-pile composite stratum, the predicted pile cutting stratum type is divided into RGP4 and RGP5, RGP4 includes the rock layer, the grouting layer, 1 to 2 piles and 3, 4 or 5 piles, and RGP5 includes the rock layer, the reinforcement area, 1 pile, 2 piles and 3, 4 or 5 piles.

8. The shield pile cutting stratum type identification method based on dual drive of data and knowledge according to claim 1 is characterized in that: In step S4, when the MT model performs a multimodal classification task, in the pile-cutting stratum type prediction task, the Transformer encoder obtains each position number of the pile-cutting stratum, and each position number corresponds to a position vector. By combining the position vector and the input vector of the training data as input features, position information is introduced into each input feature. The input features are extracted and the global context information is modeled through the multi-head self-attention mechanism, and the extracted input features are output to the MLP classifier to complete the task classification.

9. The shield pile cutting stratum type identification method based on dual drive of data and knowledge according to claim 8 is characterized in that: In the Transformer encoder, a unique code is generated for each frequency in the input feature sequence and added to the frequency feature: the input feature is linearly transformed by the embedding layer and added to the position code to obtain the input matrix X. The matrix X is input to the multi-head attention layer and linearly projected h times to obtain the query matrix Q, key matrix K and value matrix V, where h is the number of attention heads of the multi-head attention.

10. The shield pile cutting stratum type identification method based on dual drive of data and knowledge according to claim 1 is characterized in that: The model performance evaluation indicators include accuracy, precision, recall, F1-score and weighted accuracy P w , the calculation equation and formula are as follows: (7) (8) (9) (10) (11) Among them, TP represents true positive, FP represents false positive, TN represents true negative, and FN represents false negative; is a category i The weight is determined by the importance of the construction and is adjusted according to the construction task. The sum of the weights is 1; is a category i The classification accuracy of .

Images