Method and System for Identifying Geological Types of Shield Machine Construction Face

Abstract

The present invention provides a method and system for identifying the geological type of the working face during shield tunneling construction, including: Step S1: Collect the equipment status parameter data during construction and perform preprocessing; Step S2: Standardize the original data and extract features; Step S3: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain an identification result; Step S4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and perform training; Step S5: Evaluate the identification effect according to the test results of the test set. The present invention selects multiple tunneling parameters of the shield machine through correlation analysis, which is beneficial to more comprehensively reflect the geological information of the working face during shield machine construction; by combining multiple data to form a two-dimensional image data and inputting it into an improved multi-head self-attention block, the geological information of the current working face and the correlation information of adjacent working faces can be fully mined.

Description

Technical Field

[0001] The present invention relates to the technical field of parameter evaluation, and more specifically, to a method and system for identifying the geological type of the face of a shield machine during construction. Even more specifically, it relates to a method and system for identifying the geological type of the face of a shield machine during construction based on an improved multi-head self-attention mechanism and a two-dimensional convolutional neural network. Background Art

[0002] Compared with other excavation methods, the shield method has been increasingly applied in the excavation of subway tunnels, highway tunnels, and railway tunnels due to its significant advantages such as high efficiency and environmental protection. A shield machine is mainly used for tunneling in soft soil tunnels. Its structure is relatively enclosed, and it is difficult for operators to directly observe the surrounding geological conditions. This enclosed working environment increases the difficulty of the normal and stable operation of the shield machine. Since the construction process of a shield machine is greatly affected by geological conditions, many problems will arise if tunneling blindly without knowing the geological conditions ahead. On the one hand, it may cause collapses and ground settlements due to overly soft tunneling geology or excessive groundwater content, resulting in project delays and even causing casualties to a large number of residents and crew members. On the other hand, if the operating parameters of the shield machine cannot adapt to the current tunneling formation, construction accidents such as cutterhead blockage, shield body jamming, and water seepage inside the shield may occur. Therefore, the real-time and accurate identification of rock and soil types is an important prerequisite for selecting reasonable tunneling parameters and ensuring construction safety, and it helps to improve construction quality and excavation efficiency.

[0003] Patent document CN113657515A (application number: 202110957595X) discloses an invention of a classification method for the surrounding rock grade of a tunnel based on the identification of sensitive parameters of a rock machine and an improved FMC model, including: obtaining real-time TBM dynamic tunneling parameters, thrust F, propulsion speed v, cutterhead torque T, and cutterhead rotation speed n; cleaning data to obtain steady-state tunneling data; constructing a rock machine parameter database; identifying the importance of sensitive parameters of the rock machine; selecting sensitive parameters of the rock machine as training samples to train the FMC model; selecting sensitive parameters of the rock machine as identification samples and inputting them into the FMC model for surrounding rock identification; and outputting the surrounding rock grade identification result.

[0004] Patent document CN109635461A (application number: 2018115472944) discloses a method for automatically identifying surrounding rock grades by using parameters while drilling, including: preprocessing the data set of parameters while drilling collected; analyzing the preprocessed data set of parameters while drilling, applying different data dimensionality reduction methods to determine the relationships and contribution rates among variables in the parameters while drilling, applying the ordered weighted averaging operator method to perform weighted average calculations on multiple contribution rates calculated for each variable of the parameters while drilling, sorting and selecting the best according to the magnitudes of the combined contribution rates calculated, determining the main characteristic variables of the parameters while drilling, and on this basis, classifying the sample set of the main characteristic parameters; applying the established neural network and expert knowledge system to train the main characteristic parameters of different classifications, obtaining stable weight coefficients and thresholds, and verifying the surrounding rock identification of the established neural network mathematical model for surrounding rock identification by using the tested sample data.

[0005] The above patent only selects a small number of tunneling parameters and does not consider the geological information correlation between adjacent working faces, and the excavation of the information contained in the data is not sufficient.

[0006] Patent document CN108182440B (application number: CN201810019670.6) discloses a method for obtaining surrounding rock categories based on slag piece image recognition, and the steps are as follows: S1, obtaining slag piece images; S2, obtaining the processed slag piece images; S3, calculating the associated sensitive feature set of surrounding rock grades of the processed slag piece images; S4, using the AP clustering method to divide the data samples into k clusters; S5, fusing the Gaussian kernel function and the polynomial kernel function by weighting, and performing LSSVM regression on each cluster in step S4 to obtain k sub-models; S6, fusing the obtained k sub-models by weighting to obtain the surrounding rock category value. However, the invention does not use a two-dimensional convolutional neural network to identify the extracted two-dimensional features, and the accuracy improvement is limited. Summary of the Invention

[0007] Aiming at the defects in the prior art, the purpose of the present invention is to provide a method and system for identifying the geological type of the working face of a shield machine.

[0008] According to a method for identifying the geological type of the working face of a shield machine provided by the present invention, it includes:

[0009] Step S1: Collecting the equipment status parameter data during construction and performing preprocessing;

[0010] Step S2: Standardizing the original data and performing feature extraction;

[0011] Step S3: Extracting two-dimensional features of the same dimension and sending them to a two-dimensional convolutional neural network to obtain an identification result;

[0012] Step S4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and train it;

[0013] Step S5: Evaluate the recognition effect according to the test results of the test set.

[0014] Preferably, in the said step S1:

[0015] Collect the equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain the equipment status parameter sequence;

[0016] Select the hydraulic oil tank temperature, the temperature of the return oil area of the hydraulic oil tank, the main drive cooling water flow rate, the main drive cooling return water temperature, the gear oil tank temperature, the main bearing oil flow rate, the pinion oil flow rate, the oil flow rates of the front and rear raceways of the pinion, the cutter head torque, the cutter head speed setting, the average propulsion speed, the penetration degree, the total propulsion force, the propulsion speed setting, the earth pressure, the average earth pressure, the screw conveyor speed setting value, the screw conveyor speed measured value, the screw conveyor pressure measured value, the screw conveyor supplementary oil pressure measured value, the screw conveyor earth pressure measured value, the screw conveyor torque, the mortar injection port pressure.

[0017] Preferably, in the said step S2:

[0018] Use a preset data frame to extract two-dimensional data and send it to an improved multi-head self-attention block for feature extraction;

[0019] Calculate the multi-head self-attention value using the following formula:

[0020]

[0021] f(Q, K) = Q T ·K

[0022] A i = softmax(f(Q i , K i ))

[0023] head i = A i ·V i

[0024] Output = Concat(head1, head2, head3)·W O

[0025] Among them, Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, i is the head corresponding to the multi-head self-attention, is to generate the corresponding Q for the i-th head iThe transformation matrix of values, Generate the corresponding K for the i-th head respectively i The transformation matrix of values, is to generate the corresponding V for the i-th head respectively i The transformation matrix of values, X is the original input of a preset dimension, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after softmax calculation for dot product attention, head i is the attention value finally obtained for the i-th head, Output is the feature result extracted by the multi-head attention finally output, W O is the linear transformation matrix for summarizing the features extracted by different heads;

[0026] The design of the multi-head self-attention block adopts:

[0027] Step A1: Randomly discard the input after passing through the multi-head self-attention layer;

[0028] Step A2: Perform a residual connection between the output after random discard and the input;

[0029] Step A3: Perform Batch Normalization on the residual connection result to obtain result 1;

[0030] Step A4: Perform a linear transformation on result 1 and then randomly discard;

[0031] Step A5: Perform a residual connection between the output after random discard and result 1;

[0032] Step A6: Perform Batch Normalization on the residual connection result to obtain the final output result.

[0033] Preferably, in the step S3:

[0034] The two-dimensional convolutional neural network includes a preset number of convolutional layers and a fully connected layer, adopts the method of stacking convolutional layers of multiple preset sizes, and outputs the recognition result through a fully connected neural network.

[0035] Preferably, in the step S4:

[0036] Use the Keras package under the TensorFlow framework to construct and train the neural network model for identifying the geological type of the shield machine construction face; use the trained neural network model for identifying the geological type of the shield machine construction face to identify the geological type of the subsequent construction working face;

[0037] In the step S5:

[0038] Calculate the accuracy rate and F1 score respectively according to the test results of the subsequent working face data set of the construction, and evaluate the recognition effect of the geological type.

[0039] A shield machine construction face geological type recognition system provided by the present invention includes:

[0040] Module M1: Collect the equipment status parameter data during construction and perform preprocessing;

[0041] Module M2: Standardize the original data and perform feature extraction;

[0042] Module M3: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result;

[0043] Module M4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and perform training;

[0044] Module M5: Evaluate the recognition effect according to the test results of the test set.

[0045] Preferably, in the module M1:

[0046] Collect the equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain the equipment status parameter sequence;

[0047] Select the hydraulic oil tank temperature, the hydraulic oil tank return oil area temperature, the main drive cooling water flow rate, the main drive cooling return water temperature, the gear oil tank temperature, the main bearing oil flow rate, the pinion oil flow rate, the front and rear raceway oil flow rates of the pinion, the cutter head torque, the cutter head speed setting, the average propulsion speed, the penetration degree, the total propulsion force, the propulsion speed setting, the earth pressure, the average earth pressure, the screw conveyor speed setting value, the screw conveyor speed measurement value, the screw conveyor pressure measurement value, the screw conveyor supplementary oil pressure measurement value, the screw conveyor earth pressure measurement value, the screw conveyor torque, and the mortar injection port pressure.

[0048] Preferably, in the module M2:

[0049] Use a preset data frame to extract two-dimensional data and send it to an improved multi-head self-attention block for feature extraction;

[0050] Calculate the multi-head self-attention value using the following formula:

[0051]

[0052] f(Q, K) = Q T ·K

[0053] A i = softmax(f(Q i , K i ))

[0054] head i = A i ·V i

[0055] Output = Concat(head1, head2, head3)·W O

[0056] Among them, Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, and i is the head corresponding to the multi-head self-attention, is the transformation matrix for generating the corresponding Q i value of the i-th head, is the transformation matrix for generating the corresponding K i value of the i-th head respectively, is the transformation matrix for generating the corresponding V i value of the i-th head respectively. X is the original input of a preset dimension, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after the softmax calculation of the dot product attention, head i is the attention value finally obtained by the i-th head, Output is the feature result extracted by the final output multi-head attention, W O is the linear transformation matrix for summarizing the features extracted by different heads;

[0057] The design of the multi-head self-attention block adopts:

[0058] Step A1: Randomly discard the input after passing through the multi-head self-attention layer;

[0059] Step A2: Perform a residual connection between the output after random discard and the input;

[0060] Step A3: Perform Batch Normalization on the residual connection result to obtain result 1;

[0061] Step A4: Perform a linear transformation on result 1 and then randomly discard;

[0062] Step A5: Perform a residual connection between the output after random discard and result 1;

[0063] Step A6: Perform Batch Normalization on the residual connection result to obtain the final output result.

[0064] Preferably, in the module M3:

[0065] The two-dimensional convolutional neural network includes a preset number of convolutional layers and a fully connected layer, and adopts a method of stacking multiple convolutional layers of preset sizes to output the recognition result through the fully connected neural network.

[0066] Preferably, in the module M4:

[0067] The Keras package under the TensorFlow framework is used to build and train a neural network model for identifying the geological type of the tunnel face of the shield machine. The trained neural network model for identifying the geological type of the tunnel face of the shield machine is used to identify the geological type of the subsequent working face of the construction.

[0068] In the module M5:

[0069] According to the test results of the subsequent construction working face data set, the accuracy and f1 index were calculated to evaluate the recognition effect of geological types.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] 1. The present invention fully analyzes the changes in shield machine tunneling parameters mapped to different surrounding rock grades by querying the national standard for surrounding rock grade classification, and combines the Pearson linear correlation method to analyze the linear correlation between tunneling parameters and geological conditions, and selects 30 shield machine tunneling parameters, which is conducive to more comprehensive reflection of the geological information of the tunnel face during shield machine construction;

[0072] 2. The present invention designs a new attention value calculation method based on the particularity of geological type recognition tasks, which avoids the problem that the gradient of the SoftMax function disappears due to the excessive increase of the dot product attention value;

[0073] 3. The present invention combines 10 30-dimensional data to form a two-dimensional image data and inputs the improved multi-head self-attention block to extract features. The structure uses the multi-head self-attention mechanism to extract the association information between each data in the two-dimensional input and the other nine data, and preserves the information of the original data itself through residual connection, and creatively adds the BN layer to process the output result. This improved multi-head self-attention block can fully and effectively mine the geological information of the current working face and the association information of the adjacent working faces;

[0074] 4. The present invention utilizes the strong ability of convolutional neural networks to recognize two-dimensional image data, and adopts a two-dimensional convolutional neural network to identify the extracted two-dimensional features, which can greatly improve the accuracy of identifying the surrounding rock type of the working face; it helps the crew to discover geological changes ahead and make timely adjustments to the excavation status; it helps to improve construction quality and excavation efficiency, and enhance the automation and intelligence level of the shield machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Other features, objects, and advantages of the present invention will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0076] Figure 1 is a flowchart of the implementation of the shield tunneling face geological type recognition method based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network proposed by the present invention;

[0077] Figure 2 is a structural diagram of the improved multi-head self-attention block proposed by the present invention;

[0078] Figure 3 is a structural diagram of the two-dimensional convolutional neural network model proposed by the present invention;

[0079] Figure 4 is a confusion matrix of the recognition results of the shield tunneling face geological type recognition method based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network proposed by the present invention in the test set. Detailed Embodiment

[0080] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0081] Example 1:

[0082] The present invention provides a shield tunneling face geological type recognition method and system based on an improved multi-head self-attention mechanism and a two-dimensional convolutional neural network, including: collecting 30 equipment status parameter data during shield tunneling; deleting data in non-tunneling states of the shield tunneling machine; standardizing the cleaned data and taking 10 30-dimensional data as a two-dimensional data to input into an improved multi-head self-attention block for feature extraction; on this basis, building a two-dimensional convolutional neural network to process the extracted two-dimensional features and training; evaluating the recognition effect of the surrounding rock grade of the model on a labeled test set. This model uses an improved self-attention mechanism, can fully extract the correlation information of adjacent working faces, and achieve a super-high accuracy recognition of the surrounding rock grade of the shield tunneling face. It helps the crew to process the changes in the surrounding rock grade in front in a timely manner, realizes efficient and safe construction, and improves the automation and intelligence level of the shield tunneling machine.

[0083] According to a shield tunneling face geological type recognition method provided by the present invention, as Figures 1 - 4 shown, including:

[0084] Step S1: Collect the equipment status parameter data during construction and perform preprocessing;

[0085] Specifically, in the said Step S1:

[0086] Collect the equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain the equipment status parameter sequence;

[0087] Select the hydraulic oil tank temperature, the temperature of the hydraulic oil tank return area, the main drive cooling water flow rate, the main drive cooling return water temperature, the gear oil tank temperature, the main bearing oil flow rate, the pinion oil flow rate, the front and rear raceway oil flow rates of the pinion, the cutter head torque, the cutter head speed setting, the average propulsion speed, the penetration degree, the total propulsion force, the propulsion speed setting, the earth pressure, the average earth pressure, the screw conveyor speed setting value, the screw conveyor speed measured value, the screw conveyor pressure measured value, the screw conveyor supplementary oil pressure measured value, the screw conveyor earth pressure measured value, the screw conveyor torque, and the mortar injection port pressure.

[0088] Step S2: Standardize the original data and perform feature extraction;

[0089] Specifically, in the said Step S2:

[0090] Use a preset data frame to extract two-dimensional data and send it to an improved multi-head self-attention block for feature extraction;

[0091] Calculate the multi-head self-attention value using the following formula:

[0092]

[0093] f(Q, K) = Q T ·K

[0094] A i = softmax(f(Q i , K i ))

[0095] head i = A i ·V i

[0096] Output = Concat(head1, head2, head3)·W O

[0097] where Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, i is the head corresponding to the multi-head self-attention, is to generate the corresponding Q for the i-th headi The transformation matrix of values Generate corresponding K for the i-th head respectively i The transformation matrix of values is to generate corresponding V for the i-th head respectively i The transformation matrix of values, X is the original input of a preset dimension, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after softmax calculation for dot product attention, head i is the attention value finally obtained for the i-th head, Output is the feature result extracted by the multi-head attention of the final output, W O is the linear transformation matrix for summarizing the features extracted by different heads;

[0098] The design of the multi-head self-attention block adopts:

[0099] Step A1: The input is randomly discarded after passing through the multi-head self-attention layer;

[0100] Step A2: The output after random discard is connected with the input by residual connection;

[0101] Step A3: The result of residual connection is batch-normalized to obtain result 1;

[0102] Step A4: The result 1 is linearly transformed and then randomly discarded;

[0103] Step A5: The output after random discard is connected with result 1 by residual connection;

[0104] Step A6: The result of residual connection is batch-normalized to obtain the final output result.

[0105] Step S3: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result;

[0106] Specifically, in the step S3:

[0107] The two-dimensional convolutional neural network includes a preset number of convolutional layers and a fully connected layer, adopts the method of stacking convolutional layers of multiple preset sizes, and outputs the recognition result through a fully connected neural network.

[0108] Step S4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and train;

[0109] Specifically, in the step S4:

[0110] Build and train a neural network model for identifying the geological type of the tunnel face during shield machine construction using the Keras package under the TensorFlow framework; use the trained neural network model for identifying the geological type of the tunnel face during shield machine construction to identify the geological type of the subsequent working face during construction.

[0111] Step S5: Evaluate the recognition effect based on the test results of the test set.

[0112] In the said step S5:

[0113] Calculate the accuracy rate and F1 index respectively based on the test results of the subsequent working face dataset during construction, and evaluate the recognition effect of the geological type.

[0114] Example 2:

[0115] Embodiment 2 is a preferred example of Embodiment 1 to illustrate the present invention more specifically.

[0116] Those skilled in the art can understand a method for identifying the geological type of the tunnel face during shield machine construction provided by the present invention as a specific implementation manner of a system for identifying the geological type of the tunnel face during shield machine construction, that is, the system for identifying the geological type of the tunnel face during shield machine construction can be implemented by executing the step process of the method for identifying the geological type of the tunnel face during shield machine construction.

[0117] The present invention selects more tunneling parameters through correlation analysis, and inputs multiple data combinations into the improved multi-head self-attention block for feature extraction. It can not only fully excavate the geological information of the current working face contained in each sample data, but also extract the correlation information between adjacent working faces. Combining the characteristics of the convolutional neural network having strong recognition ability for two-dimensional image data, the extracted features are input into the two-dimensional convolutional neural network for recognition and classification, greatly improving the recognition accuracy of the surrounding rock grade of the tunnel face.

[0118] According to a system for identifying the geological type of the tunnel face during shield machine construction provided by the present invention, it includes:

[0119] Module M1: Collect the equipment status parameter data during construction and perform preprocessing.

[0120] Specifically, in the said module M1:

[0121] Collect the equipment status parameter data during shield machine construction and perform preprocessing to obtain an equipment status parameter sequence.

[0122] Select the hydraulic oil tank temperature, hydraulic oil tank return oil area temperature, main drive cooling water flow rate, main drive cooling return water temperature, gear oil tank temperature, main bearing oil flow rate, pinion oil flow rate, pinion front and rear raceway oil flow rate, cutter head torque, cutter head speed setting, average propulsion speed, penetration rate, total propulsion force, propulsion speed setting, earth pressure, average earth pressure, screw conveyor speed setting value, screw conveyor speed measured value, screw conveyor pressure measured value, screw conveyor supplementary oil pressure measured value, screw conveyor earth pressure measured value, screw conveyor torque, mortar injection port pressure.

[0123] Module M2: Standardize the original data and perform feature extraction;

[0124] Specifically, in the module M2:

[0125] Extract two-dimensional data using a preset data frame and send it to an improved multi-head self-attention block for feature extraction;

[0126] Calculate the multi-head self-attention value using the following formula:

[0127]

[0128] f(Q, K) = Q T ·K

[0129] A i = softmax(f(Q i , K i ))

[0130] head i = A i ·V i

[0131] Output = Concat(head1, head2, head3)·W O

[0132] Among them, Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, i is the head corresponding to the multi-head self-attention, is the transformation matrix for generating the corresponding Q i value for the i-th head, is the transformation matrix for generating the corresponding K i value for the i-th head respectively, is the one that generates the corresponding V iThe transformation matrix of values, X is the original input of a preset dimension, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after softmax calculation for dot product attention, headi is the attention value finally obtained by the i-th head, Output is the feature result extracted by the final output multi-head attention, W O is the linear transformation matrix for summarizing the features extracted by different heads;

[0133] The design of the multi-head self-attention block adopts:

[0134] Step A1: The input is randomly discarded after passing through the multi-head self-attention layer;

[0135] Step A2: The output after random discard is connected with the input through residual connection;

[0136] Step A3: The result of residual connection is subjected to Batch Normalization to obtain Result 1;

[0137] Step A4: The Result 1 is linearly transformed and then randomly discarded;

[0138] Step A5: The output after random discard is connected with Result 1 through residual connection;

[0139] Step A6: The result of residual connection is subjected to Batch Normalization to obtain the final output result.

[0140] Module M3: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result;

[0141] Specifically, in the module M3:

[0142] The two-dimensional convolutional neural network includes a preset number of convolutional layers and a fully connected layer, adopts the method of stacking multiple convolutional layers of preset sizes, and outputs the recognition result through a fully connected neural network.

[0143] Module M4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and perform training;

[0144] Specifically, in the module M4:

[0145] Use the Keras package under the TensorFlow framework to build a neural network model for shield tunneling face geological type recognition and perform training; use the trained neural network model for shield tunneling face geological type recognition to recognize the geological type of the subsequent working face of the construction;

[0146] Module M5: Evaluate the recognition effect according to the test results of the test set.

[0147] In the said Module M5:

[0148] Calculate the accuracy rate and F1 index respectively according to the test results of the subsequent construction working face data set, and evaluate the recognition effect of the geological type.

[0149] Example 3:

[0150] Embodiment 3 is a preferred example of Embodiment 1 to illustrate the present invention more specifically.

[0151] Aiming at the problems that the geological conditions of the shield machine construction face cannot be observed at present and the recognition accuracy of the current geological type recognition method is not high, the present invention provides a shield machine construction face geological type recognition method and system based on an improved multi-head self-attention mechanism and a two-dimensional convolutional neural network.

[0152] A shield machine construction face geological type recognition method based on an improved multi-head self-attention mechanism and a two-dimensional convolutional neural network provided by the present invention includes:

[0153] Step S1: Collect 30 equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain an equipment status parameter sequence;

[0154] Step S1 includes selecting the hydraulic oil tank temperature (°C), the hydraulic oil tank return oil area temperature (°C), the main drive cooling water flow rate (L / min), the main drive cooling return water temperature (°C), the 1# gear oil tank temperature (°C), the main bearing oil flow rate (L / min), the pinion oil flow rate (L / min), the pinion front and rear raceway oil flow rate (L / min), the cutter head torque (kN·m), the cutter head speed setting (%)), the average propulsion speed (mm / min), the penetration degree (mm / r), the total propulsion force (kN), the propulsion speed setting (%), the earth pressure 1# (bar), the earth pressure 2# (bar), the earth pressure 3# (bar), the earth pressure 4# (bar), the earth pressure 5# (bar), the earth pressure 6# (bar), the average earth pressure (bar), the screw conveyor speed setting value (%), the screw conveyor speed measured value (rpm), the screw conveyor pressure measured value (bar), the screw conveyor supplementary oil pressure measured value (bar), the screw conveyor earth pressure measured value after (bar), the screw conveyor torque (kN·m), the mortar injection port 3 pressure (bar), the mortar injection port 4 pressure (bar), the mortar injection port 5 pressure (bar).

[0155] Step S2: Standardize the original data by using the mean-standard deviation method;

[0156] Step S3: Extract two-dimensional data using a 10*30 data frame and transmit it to the improved multi-head self-attention block for feature extraction;

[0157] In step S3, the following is adopted:

[0158]

[0159] f(Q, K) = Q T ·K

[0160] A i = softmax(f(Q i , K i ))

[0161] head i = A i ·V i

[0162] Output = Concat(head1, head2, head3)·W O

[0163] where Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, i is the head corresponding to the multi-head self-attention, is the transformation matrix for generating the corresponding Q i value for the i-th head, is the transformation matrix for generating the corresponding K i value for the i-th head respectively, is the transformation matrix for generating the corresponding V i value for the i-th head respectively, X is the original input of 10*30 dimensions, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after softmax calculation of the dot product attention, head i is the attention value finally obtained for the i-th head, Output is the feature result extracted by the final output multi-head attention, W O is the linear transformation matrix for summarizing the features extracted by different heads.

[0164] The design of the multi-head self-attention block adopts:

[0165] Step S3.1: Randomly discard the input after passing through the multi-head self-attention layer;

[0166] Step S3.2: Perform a residual connection between the output after random discard and the input;

[0167] Step S3.3: Perform Batch Normalization on the residual connection result to obtain Result 1;

[0168] Step S3.4: Perform a linear transformation on Result 1 and then perform random dropout;

[0169] Step S3.5: Perform a residual connection between the output after random dropout and Result 1;

[0170] Step S3.6: Perform Batch Normalization on the residual connection result to obtain the final output result;

[0171] Step S4: Extract two-dimensional features of the same dimension and feed them into a two-dimensional convolutional neural network to obtain the recognition result;

[0172] The described two-dimensional convolutional neural network includes 11 convolutional layers and one fully connected layer, and adopts the method of stacking multiple convolutional layers with a size of 3*3. The quantities are 3-3-2-2-1 respectively, and the recognition result is output through the fully connected neural network.

[0173] Step S5: Use the Keras package under the TensorFlow framework to construct a neural network model for the geological type recognition of the shield machine construction face and perform training;

[0174] Step S6: Through the trained neural network model for the geological type recognition of the shield machine construction face, recognize the geological type of the subsequent construction working face;

[0175] Step S7: Calculate the accuracy rate and f1 index respectively according to the test results of the subsequent construction working face dataset to evaluate the recognition effect of the geological type.

[0176] The described improved multi-head self-attention mechanism adopts a new attention calculation method in combination with the characteristics of the geological type recognition task and is more in line with the characteristics of the geological type recognition task in the design of the multi-head self-attention block.

[0177] First, select 30 equipment status parameters with strong correlation from the data recorded in the shield machine on-site construction as the input of the intelligent recognition model. Then establish a shield machine construction face geological type recognition model based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network, use ten pieces of data to form a two-dimensional data as the input, and the corresponding surrounding rock grade as the output. And use the data recorded in the shield machine construction site to train the model. The trained model can realize the real-time perception of the geological type of the working face, which helps the crew to discover the geological changes ahead and make timely adjustments to the tunneling state, improve the construction quality and excavation efficiency, and enhance the automation and intelligence level of the shield machine.

[0178] A shield machine construction face geological type recognition system based on an improved multi-head self-attention mechanism and a two-dimensional convolutional neural network provided by the present invention includes:

[0179] Module M1: Collect 30 equipment status parameter data during shield machine construction and perform preprocessing to obtain an equipment status parameter sequence;

[0180] The module M1 selects the hydraulic oil tank temperature (°C), the hydraulic oil tank return oil area temperature (°C), the main drive cooling water flow rate (L / min), the main drive cooling return water temperature (°C), the 1# gear oil tank temperature (°C), the main bearing engine oil flow rate (L / min), the pinion engine oil flow rate (L / min), the front and rear raceway engine oil flow rate of the pinion (L / min), the cutter head torque (kNm), the cutter head speed setting (%)), the average propulsion speed (mm / min), the penetration (mm / r), the total propulsion force (kN), the propulsion speed setting (%), the earth pressure 1# (bar), the earth pressure 2# (bar), the earth pressure 3# (bar), the earth pressure 4# (bar), the earth pressure 5# (bar), the earth pressure 6# (bar), the average earth pressure (bar), the screw conveyor speed setting value (%), the screw conveyor speed measured value (rpm), the screw conveyor pressure measured value (bar), the screw conveyor supplementary oil pressure measured value (bar), the screw conveyor earth pressure measured value after (bar), the screw conveyor torque (kNm), the mortar injection port 3 pressure (bar), the mortar injection port 4 pressure (bar), the mortar injection port 5 pressure (bar).

[0181] Module M2: Standardize the original data using the mean-standard deviation method;

[0182] Module M3: Extract two-dimensional data using a 10*30 data frame and send it to an improved multi-head self-attention block for feature extraction;

[0183] In the module M3, the following is adopted:

[0184]

[0185] f(Q, K) = Q T ·K

[0186] A i = softmax(f(Q i , K i ))

[0187] head i = A i ·V i

[0188] Output = Concat(head1, head2, head3)·WO

[0189] where Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value corresponding to the i-th head, and i is the head corresponding to the multi-head self-attention, is the transformation matrix for generating the corresponding Q i value for the i-th head, is the transformation matrix for generating the corresponding K i value for the i-th head respectively, is the transformation matrix for generating the corresponding V i value for the i-th head respectively. X is the 10*30-dimensional original input, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after the softmax calculation of the dot product attention, head i is the attention value finally obtained for the i-th head, Output is the feature result extracted by the final output multi-head attention, W O is the linear transformation matrix for summarizing the features extracted by different heads.

[0190] The design of the multi-head self-attention block in the module M3 adopts:

[0191] Module M3.1: Random dropout is performed after the input passes through the multi-head self-attention layer;

[0192] Module M3.2: The output after random dropout is connected with the input through a residual connection;

[0193] Module M3.3: The result of the residual connection is batch-normalized to obtain result 1;

[0194] Module M3.4: Linear transformation is performed on result 1 and then random dropout is performed;

[0195] Module M3.5: The output after random dropout is connected with result 1 through a residual connection;

[0196] Module M3.6: The result of the residual connection is batch-normalized to obtain the final output result;

[0197] Module M4: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result;

[0198] The two-dimensional convolutional neural network in the module M4 includes 11 convolutional layers and a fully connected layer, and adopts the method of stacking multiple convolutional layers with a size of 3*3. The quantities are 3-3-2-2-1 respectively, and the recognition results are output through the fully connected neural network.

[0199] Module M5: Use the Keras package under the TensorFlow framework to construct and train a neural network model for identifying the geological type of the shield machine working face.

[0200] Module M6: Identify the geological type of the subsequent working face of the construction through the trained neural network model for identifying the geological type of the shield machine working face.

[0201] Module M7: Calculate the accuracy rate and f1 index respectively according to the test results of the subsequent working face dataset of the construction, and evaluate the recognition effect of the geological type.

[0202] The improved multi-head self-attention mechanism adopts a new attention calculation method in combination with the characteristics of the geological type recognition task and is more in line with the characteristics of the geological type recognition task in the design of the multi-head self-attention block.

[0203] First, select 30 equipment status parameters with strong correlation from the data recorded in the on-site construction of the shield machine as the input of the intelligent recognition model. Then establish a shield machine working face geological type recognition model based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network, use a two-dimensional data composed of ten data as the input, and the corresponding surrounding rock grade as the output. And use the data recorded in the shield machine construction site to train the model. The trained model can realize the real-time perception of the geological type of the working face, which helps the crew to discover the geological changes ahead and make timely adjustments to the tunneling state, improve the construction quality and excavation efficiency, and enhance the automation and intelligence level of the shield machine.

[0204] Example 4:

[0205] Embodiment 4 is a preferred example of Embodiment 1 to more specifically illustrate the present invention.

[0206] Reference Figures 1 to 4 , the present invention provides a method for identifying the geological type of the shield machine working face based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network, including the following steps:

[0207] Step S1: Collect 30 equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain the equipment status parameter sequence.

[0208] Step S2: Standardize the original data by using the mean-standard deviation method.

[0209] Step S3: Extract two-dimensional data using a 10*30 data frame and send it to an improved multi-head self-attention block for feature extraction. The design structure of the multi-head self-attention block is as Figure 2 shown. The 10*30 two-dimensional input is randomly dropped with a dropout rate of 0.2 after passing through the multi-head self-attention layer. The output after random dropout is connected to the input with a residual connection; the result of the residual connection is batch-normalized to obtain Result 1. Result 1 is linearly transformed and then randomly dropped with a dropout rate of 0.2; the output after random dropout is connected to Result 1 with a residual connection. The result of the residual connection is batch-normalized to obtain the final output feature result;

[0210] Step S4: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result. The structure of the two-dimensional convolutional neural network is as Figure 3 shown. It includes five convolutional groups. The first convolutional group includes 3 convolutional layers with a size of 3*3, and the number of convolutional kernels in each layer is 32, plus a max-pooling layer and a BN layer. The second convolutional group includes 3 convolutional layers with a size of 3*3, and the number of convolutional kernels in each layer is 64, plus a max-pooling layer and a BN layer. The third convolutional group includes 2 convolutional layers with a size of 3*3, and the number of convolutional kernels in each layer is 128, plus a max-pooling layer and a BN layer. The fourth convolutional group includes 2 convolutional layers with a size of 3*3, and the number of convolutional kernels in each layer is 256, plus a max-pooling layer and a BN layer. The fifth convolutional group includes 1 convolutional layer with a size of 3*3, and the number of convolutional kernels in each layer is 512, plus a max-pooling layer and a BN layer. Finally, it is flattened and input into a fully connected layer

[0211] Step S5: Use the Keras package under the TensorFlow framework to build a neural network model for shield tunneling face geological type recognition and train it; use the sparse categorical cross-entropy loss function and the Adam optimizer as the loss function and optimizer respectively, and the batch size of model training is set to 30. The training set includes 128,004 shield tunneling machine record data rows, and the test set includes 35,529 shield tunneling machine record data rows, obtaining a shield tunneling face geological type recognition model after training;

[0212] Step S6: Use the trained neural network model for shield tunneling face geological type recognition to recognize the geological type of the subsequent construction working face;

[0213] Step S7: Calculate the accuracy rate according to the test results of the subsequent construction working face dataset to evaluate the recognition effect of the geological type.

[0214] From Figure 4It can be seen that the proposed shield tunneling face geological type recognition model based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network is very accurate in identifying geological types. The recognition accuracy on this dataset is 96.68%. It shows that the proposed method for identifying the geological type of the shield tunneling face based on the improved multi-head self-attention mechanism and two-dimensional convolutional neural network has a high recognition accuracy.

[0215] Those skilled in the art know that in addition to implementing the systems, devices, and their respective modules provided by the present invention in the form of pure computer-readable program code, the method steps can be logically programmed to enable the systems, devices, and their respective modules provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same program. Therefore, the systems, devices, and their respective modules provided by the present invention can be regarded as a kind of hardware component, and the modules included therein for implementing various programs can also be regarded as the structure within the hardware component; the modules for implementing various functions can also be regarded as either software programs for implementing the method or the structure within the hardware component.

[0216] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for identifying the geological type of the working face during shield tunneling construction, characterized in that Including: Step S1: Collect the equipment status parameter data during construction and perform preprocessing; Step S2: Standardize the original data and perform feature extraction; Step S3: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result; Step S4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and perform training; Step S5: Evaluate the recognition effect according to the test results of the test set; In the said Step S2: Use a preset data frame to extract two-dimensional data and send it to an improved multi-head self-attention block for feature extraction; Calculate the multi-head self-attention value using the following formula: f(Q, K) = Q T ·K A i = softmax(f(Q i , K i )) head i = A i ·V i Output = Concat(head1, head2, head3) · W O Among them, Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, and i is the head corresponding to the multi-head self-attention, is the transformation matrix for generating the corresponding Q i value for the i-th head, is the transformation matrix for generating the corresponding K i value for the i-th head respectively, is the transformation matrix for generating the corresponding V i value for the i-th head respectively. X is the original input of a preset dimension, Q is the independent variable of the dot-product attention value calculation function f(Q, K), K is the independent variable of the dot-product attention value calculation function f(Q, K), A i is the value after the softmax calculation of the dot-product attention, head i is the attention value finally obtained for the i-th head, Output is the feature result extracted by the final multi-head attention, and W O is the linear transformation matrix for summarizing the features extracted by different heads; The design of the said multi-head self-attention block adopts: Step A1: Randomly discard the input after passing through the multi-head self-attention layer; Step A2: Perform residual connection on the output after random discard and the input; Step A3: Perform Batch Normalization on the result of the residual connection to obtain Result 1; Step A4: Perform linear transformation on Result 1 and then randomly discard; Step A5: Perform residual connection on the output after random discard and Result 1; Step A6: Perform Batch Normalization on the result of the residual connection to obtain the final output result; In the said Step S3: The said two-dimensional convolutional neural network includes a preset number of convolutional layers and a fully connected layer, adopts the method of stacking convolutional layers of multiple preset sizes, and outputs the recognition result through a fully connected neural network.

2. The method for identifying the geological type of the heading face in shield tunneling construction according to claim 1, wherein In the said Step S1: Collect the equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain the equipment status parameter sequence; Select the hydraulic oil tank temperature, hydraulic oil tank return oil area temperature, main drive cooling water flow, main drive cooling return water temperature, gear oil tank temperature, main bearing oil flow, pinion oil flow, pinion front and rear raceway oil flow, cutter head torque, cutter head speed setting, average propulsion speed, penetration degree, total propulsion force, propulsion speed setting, earth pressure, average earth pressure, screw conveyor speed setting value, screw conveyor speed measurement value, screw conveyor pressure measurement value, screw conveyor supplementary oil pressure measurement value, screw conveyor earth pressure measurement value, screw conveyor torque, mortar injection port pressure.

3. The method for identifying the geological type of the working face during the construction of a shield machine according to claim 1, characterized in that: In the said Step S4: Use the Keras package under the TensorFlow framework to build a neural network model for identifying the geological type of the working face during the construction of the shield machine and perform training; use the trained neural network model for identifying the geological type of the working face during the construction of the shield machine to identify the geological type of the subsequent working face during construction; In the said Step S5: Calculate the accuracy rate and f1 index respectively according to the test results of the subsequent working face dataset during construction, and evaluate the recognition effect of the geological type.

4. A geological type identification system for the working face of a shield machine construction, characterized in that, Including: Module M1: Collect the equipment status parameter data during construction and perform preprocessing; Module M2: Standardize the original data and perform feature extraction; Module M3: Extract two-dimensional features of the same dimension and send them to a two-dimensional convolutional neural network to obtain the recognition result; Module M4: Build a two-dimensional convolutional neural network to process the extracted two-dimensional features and perform training; Module M5: Evaluate the recognition effect according to the test results of the test set; In the said module M2: Extract two-dimensional data using a preset data frame and deliver it to an improved multi-head self-attention block for feature extraction; Calculate the multi-head self-attention value using the following formula: f(Q, K) = Q T ·K A i = softmax(f(Q i , K i )) head i = A i ·V i Output = Concat(head1, head2, head3) · W O Among them, Q i is the query value corresponding to the i-th head, K i is the key value corresponding to the i-th head, V i is the value value corresponding to the i-th head, where i is the head corresponding to the multi-head self-attention, is the transformation matrix for generating the corresponding Q i value for the i-th head, is the transformation matrix for generating the corresponding K i value for the i-th head respectively, is the transformation matrix for generating the corresponding V i value for the i-th head respectively. X is the original input of the preset dimension, Q is the independent variable of the dot product attention value calculation function f(Q, K), K is the independent variable of the dot product attention value calculation function f(Q, K), A i is the value after the softmax calculation of the dot product attention, head i is the attention value finally obtained for the i-th head, Output is the feature result extracted by the final multi-head attention output, W O is the linear transformation matrix for summarizing the features extracted by different heads; The design of the said multi-head self-attention block adopts: Step A1: The input is randomly discarded after passing through the multi-head self-attention layer; Step A2: Perform a residual connection between the output after random discard and the input; Step A3: Perform Batch Normalization on the result of the residual connection to obtain result 1; Step A4: Perform a linear transformation on result 1 and then perform random discard; Step A5: Perform a residual connection between the output after random discard and result 1; Step A6: Perform Batch Normalization on the result of the residual connection to obtain the final output result; In the said module M3: The said two-dimensional convolutional neural network includes a preset number of convolutional layers and a fully connected layer, adopts a method of stacking multiple convolutional layers of preset sizes, and outputs the recognition result through a fully connected neural network.

5. The shield machine construction face geological type identification system according to claim 4, characterized in that In the said module M1: Collect the equipment status parameter data during the construction of the shield machine and perform preprocessing to obtain the equipment status parameter sequence; Select the hydraulic oil tank temperature, hydraulic oil tank return oil area temperature, main drive cooling water flow rate, main drive cooling return water temperature, gear oil tank temperature, main bearing oil flow rate, pinion oil flow rate, pinion front and rear raceway oil flow rate, cutter head torque, cutter head speed setting, average propulsion speed, penetration, total propulsion force, propulsion speed setting, earth pressure, average earth pressure, screw conveyor speed setting value, screw conveyor speed measurement value, screw conveyor pressure measurement value, screw conveyor supplementary oil pressure measurement value, screw conveyor earth pressure measurement value, screw conveyor torque, mortar injection port pressure.

6. The shield machine construction face geological type identification system according to claim 4, characterized in that: In the said module M4: Use the Keras package under the TensorFlow framework to construct a neural network model for identifying the geological type of the shield machine construction face and train it; use the trained neural network model for identifying the geological type of the shield machine construction face to identify the geological type of the subsequent construction working face; In the said module M5: Calculate the accuracy rate and f1 index respectively according to the test results of the subsequent construction working face dataset, and evaluate the recognition effect of the geological type.

Images