Simulation and monitoring method for intelligent early warning of large deformation of coal mine TBM tunneling roadway

Abstract

The invention relates to the technical field of deformation early warning of a coal mine TBM tunneling roadway, in particular to a simulation and monitoring method for intelligent early warning of large deformation of the coal mine TBM tunneling roadway. According to the technical scheme, the method comprises the steps of data acquisition and preprocessing, feature mining and fusion, intelligent early warning model construction and real-time monitoring and early warning, and the data acquisition and preprocessing specifically comprises the steps that multiple types of sensors are installed on a coal mine TBM tunneling roadway site to acquire multiple data, three-dimensional reconstruction is conducted on image data, and timestamps and space coordinates of the data are unified; a multi-dimensional spatio-temporal data set is formed by using wavelet transform denoising, and feature mining and fusion comprises the steps of extracting geologic structure image data features by using 3D-CNN and capturing spatio-temporal change features of other data by using LSTM. According to the method, high-quality data is provided, meanwhile, the model is more suitable for complex working conditions and more accurate, the generalization ability of the model is enhanced, real-time monitoring and early warning are timely, effective and visual, decision making is facilitated, and safety accidents and economic losses can be effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of deformation early warning for coal mine TBM tunneling headings, and particularly to a simulation and monitoring method for intelligent early warning of large deformation in coal mine TBM tunneling headings. Background Art

[0002] In the early warning of large deformation in coal mine TBM tunneling headings, one prominent problem at present is the insufficient utilization of a large amount of data generated during the tunneling process. Coal mine TBM tunneling generates a vast amount of data, including geological condition data (such as rock hardness, joint fracture conditions, in-situ stress information, etc.), tunneling parameter data (such as propulsion speed, cutterhead rotation speed, torque, etc.), and roadway deformation monitoring data (such as displacement, convergence amount, etc.). However, existing early warning methods often simply process some of the data and do not fully explore the complex relationships between these data.

[0003] Traditional methods set some simple thresholds to judge whether a large deformation may occur in the roadway. For example, an early warning is issued when the displacement exceeds a certain fixed value. However, this method ignores the comprehensive influence of geological conditions and tunneling parameters on deformation. In fact, under different geological conditions, the same tunneling parameters may lead to different deformation results, and the relationship between these data is often non-linear. Therefore, the present application proposes a simulation and monitoring method for intelligent early warning of large deformation in coal mine TBM tunneling headings. Summary of the Invention

[0004] The object of the present invention is to address the problem in the background art that the traditional early warning method for large deformation in coal mine TBM tunneling headings only sets some simple thresholds to judge whether a large deformation may occur in the roadway, ignoring the comprehensive influence of geological conditions and tunneling parameters on deformation, and to propose a simulation and monitoring method for intelligent early warning of large deformation in coal mine TBM tunneling headings.

[0005] The technical solution of the present invention: A simulation and monitoring method for intelligent early warning of large deformation in coal mine TBM tunneling headings, comprising the following steps:

[0006] Data acquisition and preprocessing: Install multiple types of sensors at the site of coal mine TBM tunneling headings to collect various data, perform three-dimensional reconstruction on the image data and unify the time stamps and spatial coordinates of the data, and use wavelet transform for denoising to form a multi-dimensional spatio-temporal data set;

[0007] Feature mining and fusion: Use 3D-CNN to extract the features of geological structure image data, use LSTM to capture the spatio-temporal change features of other data, and fuse the two through a feature fusion network based on the attention mechanism to form a fused feature vector;

[0008] Intelligent early warning model construction: An early warning model based on GAN is constructed with the fused feature vector as the input. During training, the Wasserstein distance is used as the loss function, and a large amount of historical data is used to train until the model converges;

[0009] Real-time monitoring and early warning: The real-time data is input into the trained model, and early warning signals of different levels are output according to the preset threshold, and the deformed area and trend are displayed.

[0010] Optionally, the multi-type sensors include but are not limited to geological radar sensors, stress sensors, displacement sensors, and tunneling parameter sensors, which collect geological structure image data, in-situ stress data, roadway displacement data, TBM propulsion speed, and cutterhead torque data in real time.

[0011] Optionally, the specific process of the 3D-CNN extracting the features of the geological structure image data is as follows: The three-dimensional reconstructed geological structure image data is input into the three-dimensional convolutional neural network (3D-CNN), and the texture, shape, and spatial distribution features of the geological structure are automatically extracted through multiple convolutional layers and pooling layers of the 3D-CNN; the size of the convolutional kernel of the 3D-CNN is 3×3×3, the stride is 2, and the padding method is same. The formula of its loss function is:

[0012]

[0013] where N is the number of samples, C is the number of categories, X, Y, and Z are the dimensions of the three-dimensional image, y i,c,x,y,z is the true label, is the predicted value.

[0014] Optionally, when using LSTM to capture the spatio-temporal change features of other data, for in-situ stress data, roadway displacement data, and tunneling parameter data, a long short-term memory network is constructed, and the gating mechanism of LSTM is used to capture the trend features and long-term dependence relationships of the data changing over time. The number of hidden layer units of LSTM is 128, and the calculation formula of the forget gate is respectively: f t =σ(W f ·[h t-1 ,x t +b f )

[0015] The calculation formula of the input gate is respectively: i t =σ(W i ·[h t-1 ,x t +b i )

[0016] The calculation formula of the output gate is respectively: o t =σ(W o ·[h t-1 ,x t +bo )

[0017] The cell state update formula is as follows: h t = o t *tanh(C t ),

[0018] where f t : is the output of the forget gate; σ is the sigmoid function; W f is the weight matrix; h t-1 is the hidden state at the previous moment; x t is the current input; b f is the bias term; i t is the output of the input gate; Candidate cell state; W i , W o are both weight matrices; b i , b o are both bias terms; C t is the cell state update; o t is the output of the output gate; h t is the current hidden state.

[0019] Optionally, in the feature fusion network based on the attention mechanism, the attention weight calculation formula is:

[0020]

[0021] The calculation formula for the fused feature vector is:

[0022]

[0023] where a i is the attention weight of the i-th feature; e i is the energy value of the i-th feature calculated through a feed-forward neural network; n is the number of features; F is the fused feature vector; F i is the i-th feature vector.

[0024] Optionally, a warning model based on GAN is constructed with the fused feature vector as the input. Specifically, a warning model based on the generative adversarial network (GAN) is constructed with the fused feature vector as the input, where the generator is used to generate the prediction results of the roadway deformation, and the discriminator is used to judge the difference between the prediction results and the actual monitoring data.

[0025] Optionally, during the training process of the adversarial network, the Wasserstein distance is introduced as the loss function, and the formula for the Wasserstein distance loss function is:

[0026]

[0027] Among them, E is the expectation, and P data is the true data distribution; D(x) is the output of the discriminator for the true data; P z is the noise distribution input to the generator; D(G(z)) is the output of the discriminator for the data generated by the generator;

[0028] Optionally, input the real-time data into the trained model, and output early warning signals at different levels according to the preset threshold, and display the deformation area and trend. Specifically, it includes:

[0029] Input the multi-dimensional spatio-temporal data collected and preprocessed in real time into the trained intelligent early warning model, and the model outputs the prediction probability of large roadway deformation and the predicted value of the deformation degree;

[0030] According to the preset different-level deformation thresholds and probability thresholds, when the prediction probability and the predicted value of the deformation degree exceed the corresponding thresholds, issue early warning signals at different levels, and at the same time visually display the area where large deformation may occur and the deformation trend in the monitoring system;

[0031] The different-level deformation thresholds are determined by dividing the roadway deformation degree through the clustering analysis algorithm, and the probability thresholds are set accordingly. During the training process, the early stopping method is used to prevent overfitting, and the value range of the early stopping method parameters is 5-10.

[0032] Optionally, according to the design specifications of the coal mine roadway and the historical large deformation accident data, divide the roadway deformation degree into three levels: mild, moderate, and severe through the clustering analysis algorithm, and the corresponding deformation thresholds are d 1 、d 2 、d 3 ,and d 1 <d 2 <d 3 ; the probability thresholds are set as the mild early warning probability threshold p 1 、the moderate early warning probability threshold p 2 、the severe early warning probability threshold p 3 ,and p 1 <p 2 <p 3 。

[0033] Optionally, the installation positions and quantities of multiple types of sensors are dynamically adjusted according to the geological conditions and driving length of the coal mine TBM driving roadway, and the optimal layout of the sensors is determined through finite element simulation analysis.

[0034] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0035] 1. Through the reasonable layout of multiple types of sensors at key positions and the high-precision data acquisition frequency in this application, it is possible to comprehensively obtain geological and tunneling parameter information during the tunneling process of coal mine TBM, providing a rich data basis for subsequent accurate early warning. 3D reconstruction and unified coordinate processing make the data have better spatial consistency and analyzability, facilitating the exploration of the internal relationships between data. The wavelet transform denoising method can specifically remove the noise generated in complex environments, improve the data quality, avoid the interference of noise on subsequent model training and prediction, enable the model to better learn the true features in the data, and improve the accuracy of early warning.

[0036] 2. 3D-CNN can automatically learn the complex features of geological structure images, overcoming the problem that it is difficult to manually extract effective geological features by traditional methods. It can capture the subtle texture, shape, and spatial distribution changes in the geological structure, and these features are of great significance for judging large roadway deformations.

[0037] 3. Through the processing ability of LSTM for time series data, it can well capture the trends and long-term dependence relationships of in-situ stress, displacement, and tunneling parameters, etc., that change over time. This helps to discover potential deformation trends during the tunneling process and issue early warnings in advance. The feature fusion network based on the attention mechanism can dynamically allocate weights according to the importance of different features in the prediction of large roadway deformations, making the model pay more attention to key features, avoiding the limitation of treating all features equally in traditional fusion methods, and further improving the adaptability and prediction accuracy of the model to complex working conditions.

[0038] 4. The use of the early warning model based on GAN and the Wasserstein distance loss function enables the model to generate prediction results closer to the real situation, and improves the stability and convergence speed of model training. Compared with traditional early warning models, it can better handle complex non-linear relationships, reduce overfitting phenomena, and thus more accurately predict the occurrence of large roadway deformations. Training with a large amount of historical data and labeled samples enables the model to learn the deformation laws under various geological conditions and tunneling working conditions, improves the generalization ability of the model, and can adapt to the actual situations of different coal mines.

[0039] 5. This invention issues early warnings according to different levels of deformation thresholds and probability thresholds, and visually displays the deformation area and trend in the monitoring system, enabling the staff to timely understand the safety status of the roadway and take corresponding measures. This hierarchical early warning and visualization display method helps to improve the pertinence and timeliness of decision-making, and avoid or reduce safety accidents and economic losses caused by large roadway deformations.

[0040] In summary, through the reasonable layout of multiple types of sensors, high-frequency acquisition, combined with 3D reconstruction, coordinate unification, and wavelet transform denoising, the present invention provides high-quality data for subsequent processing. By extracting the depth features of geological images through 3D-CNN, capturing the time series features through LSTM, and dynamically fusing the attention mechanism, the model is more adaptable to complex working conditions and more accurate. The intelligent early warning model is constructed through GAN and Wasserstein distance loss function, which is stable and accurate. By using a large amount of historical data to enhance the generalization ability, real-time monitoring and early warning are timely, effective, and visual, which is conducive to decision-making and can reduce safety accidents and economic losses. Description of the Drawings

[0041] Figure 1 It is a flowchart of a simulation and monitoring method for intelligent early warning of large deformation in a coal mine TBM driving roadway. Detailed Embodiments

[0042] The following uses specific specific examples to illustrate the embodiments of the present invention. Those skilled in the art can easily understand the other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0043] Embodiment 1

[0044] As Figure 1 shown, a simulation and monitoring method for intelligent early warning of large deformation in a coal mine TBM driving roadway proposed by the present invention includes four parts: data collection and preprocessing, feature mining and fusion, intelligent early warning model construction, and real-time monitoring and early warning. The following details its process.

[0045] 1. Data Collection and Preprocessing

[0046] Sensor Installation and Data Collection:

[0047] Install geological radar sensors at different key positions in the roadway during coal mine TBM tunneling. Install them on the top and both side walls of the roadway every 5 - 10 meters along the tunneling direction to obtain geological structure image data with a resolution accurate to the centimeter level. At the same time, install stress sensors in areas where ground stress may concentrate, install displacement sensors on the roadway wall and top, and install tunneling parameter sensors on the TBM equipment. The propulsion speed sensor is installed near the propulsion cylinder, and the cutter head torque sensor is installed at the cutter head drive system to collect data at a frequency of once per second. The reasonable layout of multiple types of sensors at key positions and the high-precision data acquisition frequency can comprehensively obtain geological and tunneling parameter information during the coal mine TBM tunneling process, providing a rich data basis for subsequent accurate early warning. Three-dimensional reconstruction and unified coordinate processing make the data have better spatial consistency and analyzability, facilitating the excavation of the internal relationships between data.

[0048] These sensors transmit data to the data acquisition terminal through wired or wireless means, including Zigbee or LoRa low-power wireless communication technologies, considering the complex environment underground in coal mines.

[0049] Data preprocessing:

[0050] For the image data collected by the geological radar, use a three-dimensional reconstruction algorithm to synthesize multiple two-dimensional images into a three-dimensional image. The three-dimensional reconstruction algorithm includes methods based on the principle of stereo vision or laser scan matching. Then align all sensor data according to the timestamp, with the GPS time or a unified high-precision clock as the reference, and at the same time unify the spatial coordinates to the same coordinate system, with the starting point of the roadway as the origin, the tunneling direction as the X-axis, the vertical direction as the Y-axis, and the horizontal direction as the Z-axis.

[0051] Use wavelet transform to denoise the data set. Select an appropriate wavelet basis function, the Daubechies wavelet, and the decomposition level. According to the data noise characteristics, 3 - 5 levels can be selected. Decompose the noisy signal into sub-bands of different frequencies, remove the high-frequency noise part through threshold processing, and then perform reconstruction to obtain the denoised data. This can effectively remove the noise data generated due to the vibration of tunneling equipment, electromagnetic interference, and complex geological environment, improving the data quality. The wavelet transform denoising method can specifically remove the noise generated in a complex environment, improve the data quality, avoid the interference of noise on subsequent model training and prediction, enable the model to better learn the true features in the data, and improve the accuracy of early warning.

[0052] 2. Feature mining and fusion implementation methods

[0053] 3D-CNN feature extraction:

[0054] Build a 3D-CNN model with a convolutional kernel size of 3×3×3, a stride of 2, and a padding method of same. Input the three-dimensional geological structure image data into the model. After passing through multiple convolutional layers and pooling layers, set 3 - 5 convolutional layers and 2 - 3 pooling layers to automatically extract the texture, shape, and spatial distribution features of the geological structure. In the convolutional layer, each convolutional kernel slides on the three-dimensional image to extract local features, and the pooling layer downsamples the features to reduce the data dimension while retaining key features.

[0055] Automatically extract the texture, shape, and spatial distribution features of the geological structure through multiple convolutional layers and pooling layers of 3D-CNN; the convolutional kernel size of 3D-CNN is 3×3×3, the stride is 2, and the padding method is same. Its loss function formula is:

[0056]

[0057] where N is the number of samples, C is the number of classes, X, Y, and Z are the dimensions of the three-dimensional image, and y i,c,x,y,z is the true label, is the predicted value.

[0058] LSTM feature extraction:

[0059] For data arranged in time series such as in-situ stress data, roadway displacement data, and tunneling parameter data, build an LSTM network with 128 hidden layer units. Calculate according to the set calculation formulas for the forget gate, input gate, and output gate. For the forget gate, linearly combine the previous hidden state h t-1 and the current input x t with the weight matrix W f , add the bias term b f , and finally obtain the output of the forget gate through the sigmoid function. In this way, the LSTM network can capture the trend features and long-term dependence relationships of these data over time, such as the change trend of in-situ stress with the tunneling process, the dynamic correlation between displacement and tunneling parameters, etc. Build a long short-term memory network, use the gating mechanism of LSTM to capture the trend features and long-term dependence relationships of data over time. The number of hidden layer units of LSTM is 128, and the calculation formulas for the forget gate are respectively: f t =σ(W f ·[h t-1 ,x t +b f )

[0060] The calculation formulas for the input gate are respectively: i t =σ(W i ·[h t-1 ,x t +b i )

[0061] The output gate calculation formulas are as follows: o t = σ(W o ·[h t-1 , x t + b o )

[0062] The cell state update formula is: h t = o t *tanh(C t )

[0063] Among them, f t : is the output of the forget gate; σ is the sigmoid function; W f is the weight matrix; h t-1 is the previous hidden state; x t is the current input; b f is the bias term; i t is the output of the input gate; Candidate cell state; W i , W o are both weight matrices; b i , b o are both bias terms; C t is the cell state update; o t is the output of the output gate; h t is the current hidden state. Deep feature extraction: 3D-CNN can automatically learn the complex features of geological structure images, overcoming the problem that it is difficult to manually extract effective geological features by traditional methods. It can capture the subtle texture, shape, and spatial distribution changes in geological structures, and these features are of great significance for judging large roadway deformations.

[0064] Time series feature capture: The processing ability of LSTM for time series data enables it to well capture the trends and long-term dependence relationships of in-situ stress, displacement, and tunneling parameters over time. This helps to discover potential deformation trends during tunneling and issue early warnings.

[0065] Feature fusion based on the attention mechanism:

[0066] In the feature fusion network based on the attention mechanism, for the geological features extracted by 3D-CNN and the spatio-temporal change features extracted by LSTM, the energy value of each feature is calculated through a feed-forward neural network. This feed-forward neural network can be a simple multi-layer perceptron (such as including an input layer, a hidden layer, and an output layer). Then, according to the attention weight calculation formula:

[0067]

[0068] Calculate the attention weight a for each feature i , where n is the total number of features. Finally, according to the formula for the fused feature vector

[0069]

[0070] where, a i is the attention weight of the i-th feature; e i is the energy value of the i-th feature calculated by a feed-forward neural network; n is the number of features; F is the fused feature vector; F i is the i-th feature vector.

[0071] All features are weighted and fused to obtain a fused feature vector. In this way, features can be dynamically fused according to the influence weights of different features on the large deformation of the roadway, enabling key features to play a more important role in subsequent models. The feature fusion network based on the attention mechanism can dynamically allocate weights according to the importance of different features in the prediction of roadway large deformation, making the model pay more attention to key features, avoiding the limitation of treating all features equally in traditional fusion methods, and further improving the adaptability and prediction accuracy of the model to complex working conditions.

[0072] 3. Implementation method of intelligent early warning model construction

[0073] Construction and training of GAN model:

[0074] Construct a GAN-based early warning model with the fused feature vector as the input. Both the generator and the discriminator adopt a multi-layer neural network structure. The generator can include transposed convolutional layers for generating data, and the discriminator includes convolutional layers for judging the authenticity of data. During the training process, the Wasserstein distance is used as the loss function. Sample real data from the real data distribution and input it into the discriminator. At the same time, sample noise from the noise distribution input by the generator, and the generator generates data according to the noise and also inputs it into the discriminator. By continuously adjusting the parameters of the generator and the discriminator, it is made difficult for the discriminator to distinguish between real data and generated data, and at the same time, the prediction results generated by the generator are made closer to the real situation. The formula for the Wasserstein distance loss function is:

[0075]

[0076] where, E is the expectation, P data is the real data distribution; D(x) is the output of the discriminator for real data; P z is the noise distribution input by the generator; D(G(z)) is the output of the discriminator for the data generated by the generator.

[0077] The warning model is trained using a large amount of historical data and samples marked with roadway large deformation conditions. The dataset is divided into a training set and a validation set according to 8:2. During the training process, the mini-batch gradient descent algorithm is adopted, and the batch size is set to 32 to update the model parameters until the loss of the model on the validation set no longer decreases or reaches the preset number of training epochs between 100 and 200, obtaining a trained intelligent warning model. The use of the warning model based on GAN and the Wasserstein distance loss function enables the model to generate prediction results closer to the real situation and improves the stability and convergence speed of model training. Compared with the traditional warning model, it can better handle complex non-linear relationships, reduce the overfitting phenomenon, and thus more accurately predict the occurrence of roadway large deformation.

[0078] Make full use of historical data: Use a large amount of historical data and marked samples for training, enabling the model to learn the deformation laws under various geological conditions and tunneling working conditions, improving the generalization ability of the model, and being able to adapt to the actual situations of different coal mines.

[0079] 4. Implementation methods of real-time monitoring and warning

[0080] Real-time data input and prediction:

[0081] Input the multi-dimensional spatio-temporal data collected and preprocessed in real time into the trained intelligent warning model. The model outputs the prediction probability of roadway large deformation and the predicted value of the deformation degree according to the input fusion feature vector.

[0082] Warning and display:

[0083] Give warnings according to different preset deformation thresholds and probability thresholds of different levels. Through the clustering analysis algorithm (K-means clustering), analyze the design specifications of coal mine roadways and historical large deformation accident data, and divide the roadway deformation degree into three levels: mild (deformation threshold d 1 ), moderate (d 2 ), and severe (d 3 ), d 1 <d 2 <d 3 , and the probability thresholds are set as the mild warning probability threshold p 1 , the moderate warning probability threshold p 2 , the severe warning probability threshold p 3 , and p 1 <p 2 <p 3 . When the prediction probability and the predicted value of the deformation degree exceed the corresponding thresholds, different levels of warning signals are issued. When the prediction probability p 1 is greater than and the predicted value of the deformation degree is greater than d 1When it is [a certain value], a mild warning signal is issued; when the predicted probability is greater than p 2 and the predicted value of the deformation degree is greater than d 2 When it is [a certain value], a moderate warning signal is issued; when the predicted probability p 3 is greater than and the predicted value of the deformation degree is greater than d 3 When it is [a certain value], a severe warning signal is issued. At the same time, the area where large deformation may occur is visually displayed on the interface of the monitoring system. By marking the discolored area and the deformation trend in the 3D model of the roadway, the deformation development direction is shown through animation. Warning is carried out according to different levels of deformation thresholds and probability thresholds, and the deformation area and trend are visually displayed in the monitoring system, which can enable the staff to timely understand the safety status of the roadway and take corresponding measures. This hierarchical warning and visualization display method helps to improve the pertinence and timeliness of decision-making, and avoid or reduce safety accidents and economic losses caused by large deformation of the roadway.

[0084] Embodiment 2

[0085] Based on Embodiment 1, considering the influence of complex environmental factor changes in the coal mine underground on the sensor performance and data acquisition, an adaptive calibration and fault detection mechanism for the sensor is added.

[0086] For the adaptive calibration of the sensor, a set of standard reference data is collected every hour, and these standard reference data are obtained under known stable geological conditions and tunneling parameters. By comparing the real-time collected data with the standard reference data, a calibration model is established using support vector regression in machine learning to calibrate the deviation of the sensor in real time to ensure the accuracy of data acquisition.

[0087] At the same time, for possible faults of the sensor, a fault detection method based on signal feature analysis and machine learning classification is adopted. Analyze the data signal features in the normal working and common fault states of the sensor, such as the frequency, amplitude, phase, etc. of the signal. Use classification algorithms such as decision trees to train the fault detection model. When the data signal features of the sensor exceed the normal range, the sensor fault is determined in time and an alarm is issued to notify the staff to repair or replace. This mechanism can further ensure the reliability of data acquisition, thereby improving the stability and accuracy of the entire warning system.

[0088] In addition, in the feature mining and fusion link, a dynamic update mechanism for feature importance evaluation is introduced. As new data is continuously collected and the model is continuously trained, the distribution of the data may change. Re-evaluate the importance of the geological features extracted by 3D-CNN and the spatio-temporal change features extracted by LSTM every 100 meters of tunneling, and adjust the weight parameters in the feature fusion network based on the attention mechanism according to the new evaluation results, so that the model can better adapt to the dynamic changes of the complex environment in the coal mine underground and improve the timeliness and accuracy of warning.

[0089] Considering the differences in geological conditions and tunneling techniques between different coal mines, the transfer learning ability of the model is enhanced. When training the model with the data of a new coal mine, the model parameters that have been trained well in other coal mines are used as the initial values, and then a small amount of data from the new coal mine is used for fine-tuning. In this way, when the data is limited, the model can quickly adapt to the environment of the new coal mine, reducing the model training time and resource consumption, and improving the versatility and scalability of the model.

[0090] The above specific embodiments are merely several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A simulation and monitoring method for intelligent early warning of large deformation of TBM tunneling tunnels in coal mines, characterized in that: The following steps are involved: Data collection and preprocessing: installing multiple types of sensors on site in coal mine TBM tunneling tunnels to collect multiple data, reconstructing image data in three dimensions and unifying the timestamp and spatial coordinates of the data, and using wavelet transform to remove noise to form a multi-dimensional spatiotemporal data set; Feature mining and fusion: use 3D-CNN to extract the features of geological structure image data, use LSTM to capture the spatiotemporal variation features of other data, and fuse the two to form a fused feature vector through a feature fusion network based on the attention mechanism; Intelligent early warning model construction, using fused feature vectors as input to build a GAN-based early warning model, using Wasserstein distance as the loss function during training, and using a large amount of historical data to train until the model converges; Real-time monitoring and early warning, input real-time data into the trained model, output different levels of early warning signals according to preset thresholds and display deformation areas and trends.

2. The method for simulating and monitoring large deformation intelligent early warning of coal mine TBM tunneling tunnel according to claim 1 is characterized in that: Multiple types of sensors include but are not limited to geological radar sensors, stress sensors, displacement sensors, and tunneling parameter sensors, which collect geological structure image data, ground stress data, tunnel displacement data, TBM propulsion speed, and cutterhead torque data in real time.

3. The method for simulating and monitoring large deformation intelligent early warning of TBM tunneling tunnel in coal mine according to claim 1 is characterized in that: The specific method of extracting geological structure image data features by 3D-CNN is as follows: the geological structure image data after 3D reconstruction is input into the 3D convolutional neural network (3D-CNN), and the texture, shape and spatial distribution characteristics of the geological structure are automatically extracted through multiple convolutional layers and pooling layers of 3D-CNN; The convolution kernel size of 3D-CNN is 3×3×3, the step size is 2, the padding method is the same, and its loss function formula is: Among them, N is the number of samples, C is the number of categories, X, Y, Z are the dimensions of the three-dimensional image, and y i,c,x,y,z is the true label, is the predicted value.

4. The method for simulating and monitoring large deformation intelligent early warning of TBM tunneling tunnel in coal mines according to claim 3 is characterized in that: In the process of using LSTM to capture the spatiotemporal variation characteristics of other data, a long short-term memory network is constructed for ground stress data, tunnel displacement data, and excavation parameter data. The gating mechanism of LSTM is used to capture the trend characteristics and long-term dependencies of data over time. The number of hidden layer units of LSTM is 128, and the calculation formulas of the forget gate are: f t =σ(W f ·[h t-1 ,x t ]+b f ) The input gate calculation formulas are: t =σ(W i ·[h t-1 ,x t ]+b i ) The calculation formulas for the output gate are: t =σ(W o ·[h t-1 ,x t ]+b o ) The cell state update formula is: h t =o t *tanh(C t ), Among them, f t : is the output of the forget gate; σ is the sigmoid function; W f is the weight matrix; h t-1 is the hidden state at the last moment; x t is the current input; b f is the bias term; i t is the input gate output; Candidate cell state; W i , W o are all weight matrices; b i 、b o All are bias terms; C t Update for cell status; o t Output gate output; h t The current hidden state.

5. The method for simulating and monitoring large deformation intelligent early warning of TBM tunneling tunnel in coal mines according to claim 4 is characterized in that: In the feature fusion network based on the attention mechanism, the attention weight calculation formula is: The calculation formula of the fused feature vector is: Among them, a i is the attention weight of the i-th feature; e i is the energy value of the i-th feature calculated by a feedforward neural network; n is the number of features; F is the fused feature vector; F i The i-th eigenvector.

6. The method for simulating and monitoring large deformation intelligent early warning of coal mine TBM tunneling tunnel according to claim 1 is characterized in that: Taking the fused feature vector as input, a GAN-based early warning model is constructed. Specifically, taking the fused feature vector as input, a generative adversarial network (GAN)-based early warning model is constructed, wherein the generator is used to generate the prediction results of the tunnel deformation, and the discriminator is used to judge the difference between the prediction results and the actual monitoring data.

7. The method for simulating and monitoring large deformation intelligent early warning of TBM tunneling tunnel in coal mines according to claim 6 is characterized in that: In the training process of the adversarial network, Wasserstein distance is introduced as the loss function. The formula of Wasserstein distance loss function is: Among them, E is the expectation, P data is the real data distribution; D(x) is the output of the discriminator for the real data; P z is the noise distribution of the generator input; D(G(z)) is the output of the discriminator for the data generated by the generator.

8. The method for simulating and monitoring large deformation intelligent early warning of coal mine TBM tunneling tunnel according to claim 1 is characterized in that: The real-time data is input into the trained model, and different levels of warning signals are output according to the preset thresholds, and the deformation areas and trends are displayed, including: The multi-dimensional spatiotemporal data collected in real time and preprocessed are input into the trained intelligent early warning model, and the model outputs the predicted probability and deformation degree of the tunnel; According to the pre-set deformation thresholds and probability thresholds of different levels, when the predicted probability and deformation degree predicted values ​​exceed the corresponding thresholds, different levels of early warning signals are issued, and the areas where large deformation may occur and the deformation trend are intuitively displayed in the monitoring system; The deformation thresholds of different levels are determined by dividing the deformation degree of the tunnel through the cluster analysis algorithm, and the probability threshold is set accordingly. The early stopping method is used in the training process to prevent overfitting, and the parameter value range of the early stopping method is 5-10.

9. The method for simulating and monitoring large deformation intelligent early warning of coal mine TBM tunneling tunnel according to claim 8 is characterized in that: According to the design specifications of coal mine roadways and historical data of large deformation accidents, the degree of roadway deformation is divided into three levels: mild, moderate, and severe, through a clustering analysis algorithm. The corresponding deformation thresholds are d1, d2, and d3, respectively, and d1 < d2 < d3; the probability thresholds are set as the mild warning probability threshold p1, the moderate warning probability threshold p2, and the severe warning probability threshold p3, and p1 < p2 < p3.

10. The method for simulating and monitoring large deformation intelligent early warning of coal mine TBM tunneling tunnel according to claim 1 is characterized in that: The installation positions and quantities of multiple types of sensors are dynamically adjusted according to the geological conditions and driving length of the coal mine TBM driving roadway, and the optimal layout of the sensors is determined through finite element simulation analysis.

Images