Gas tunnel ventilation method and system based on geological simulation and deep learning

Abstract

The invention discloses a gas tunnel ventilation method and system based on geological simulation and deep learning, relates to the technical field of construction safety monitoring, and solves the technical problem that the construction safety of a gas tunnel cannot be fully guaranteed through gas concentration prediction before construction and gas concentration real-time monitoring during construction in the prior art. The method comprises the steps that a three-dimensional geologic model is constructed based on information of a geologic layer needing gas tunnel construction at present, then construction simulation is conducted, the change condition of the geologic layer in the construction process is simulated, the change condition of the geologic layer is input into a gas prediction model after training verification, and gas concentration data in a tunnel in the construction process is predicted. The ventilation intensity is adjusted in real time in the construction process according to the predicted gas concentration data; according to the method, on the basis of geological strata simulation, the relevance between geological strata changes and corresponding gas concentration changes is excavated, and the gas concentration changes in the construction process can be predicted on the basis of the geological conditions of the current construction tunnel.

Description

Technical Field

[0001] The present invention relates to the technical field of construction safety monitoring, and particularly to a ventilation method and system for gas tunnels based on geological simulation and deep learning. Background Art

[0002] With the rapid development of high-speed railway and highway construction, the construction of high-speed railway and highway continuously extends to the western remote areas. However, there are many areas with coal resources in the western mountainous areas, resulting in gas disasters in underground engineering construction. As a severe geological disaster, once a gas disaster occurs, it will cause great losses to construction workers, construction parties or construction progress. Therefore, strengthening the prevention of gas disasters is very important in the construction safety technology of gas tunnels.

[0003] Existing gas disaster prevention technologies include:

[0004] Strengthen the design of tunnel ventilation. Generally, by combining forced ventilation and exhaust ventilation, a stable intake airway and return airway are formed to avoid local gas accumulation. It also includes introducing intelligent variable-frequency fans to automatically adjust the wind speed and air volume according to the gas concentration to ensure ventilation efficiency. However, installing too many ventilation devices may cause waste of resources. At the same time, additional support measures are required for the installation of ventilation devices under complex geological conditions, which is not convenient for use.

[0005] Real-time monitor and warn the gas concentration in the tunnel. Generally, a combination of fixed sensors (such as methane and carbon monoxide sensors) and portable detectors is used to continuously monitor the gas concentration in the tunnel for 24 hours. When the monitored concentration exceeds the threshold, an audible and visual alarm is triggered, and the ventilation system is linked to control the wind speed. However, relying solely on sensor monitoring, due to the easy interference of sensors by environmental factors such as dust and humidity, faults may occur, resulting in inaccurate monitoring. And relying solely on monitoring cannot give timely warnings for large changes such as gas outbursts within a short period of time. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a ventilation method and system for gas tunnels based on geological simulation and deep learning, which solves the technical problem that the prior art cannot fully guarantee the construction safety of gas tunnels through pre-construction gas concentration prediction and real-time monitoring of gas concentration during construction.

[0007] A ventilation system for gas tunnels based on geological simulation and deep learning includes:

[0008] Construct a 3D geological model based on the geological layer information required for current gas tunnel construction, and then conduct construction simulation to simulate the changes in the geological layer during the construction process. Input the changes in the geological layer into the trained and verified gas prediction model to predict the gas concentration data in the tunnel during the construction process, and adjust the ventilation intensity in real time according to the predicted gas concentration data.

[0009] Furthermore, it also includes obtaining the gas concentration change trend during the construction process from the predicted gas concentration data, clarifying the corresponding time sequence nodes when the gas concentration shows an obvious upward trend and there are mutation situations, and then reversely determining the corresponding geological layer position based on this time sequence node. Set the range with a radius of r centered on this geological layer position as the dangerous area, and prepare corresponding preventive measures in advance for the dangerous area situation.

[0010] Furthermore, the geological layer information of the historical gas tunnel construction includes: overall rock layer information, coal seam burial depth, coal seam fault situation, coal seam surrounding rock, coal seam thickness, and nearby hydrogeological conditions.

[0011] Furthermore, the construction of the 3D geological model based on the geological layer information includes: learning the deep geometric morphology and spatial distribution characteristics of various types of rock layers in the geological layer through a convolutional neural network, and at the same time performing deep feature fusion on geological exploration data and multi-source data related to geospatial information. Then, use the continuous adversarial training of the generative adversarial network to generate a 3D geological model that highlights the characteristics of the coal seam. Finally, train and fit an optimal simulation model to achieve the correspondence between the geological layer information and the 3D geological model, and use it to directly generate a 3D geological model according to the geological layer information.

[0012] Furthermore, the training and verification of the gas prediction model include: collecting the geological layer information of the historical gas tunnel construction and the gas concentration data recorded during the construction process, constructing a 3D geological model based on the geological layer information, conducting construction simulation on the basis of the 3D geological model according to the construction plan, obtaining the changes in the geological layer during the simulated construction process through the construction simulation, processing the changes in the geological layer and the gas concentration data to obtain a data set, and using the data set to train and verify the gas prediction model.

[0013] Furthermore, obtaining the changes in the geological layer during the simulated construction process through the construction simulation includes obtaining the simulation results by using PLAXIS 3D to simulate the geometric morphology and distribution changes of each rock layer during the construction process on the basis of the 3D geological model, and further extracting the geometric morphology changes of the coal seam and adjacent rock layers from the simulation results and performing time sequence processing to obtain the time sequence data of the coal seam changes during the construction progress.

[0014] Further, the data set obtained by processing based on the geological layer change situation and gas concentration data includes: determining several key nodes of geological layer change in the historical construction progress according to the coal seam change time series data, then preprocessing the gas concentration data and the coal seam change time series data respectively to remove redundant, repeated and invalid data, and then aligning the preprocessed gas concentration data and the coal seam change time series data with several key nodes as the core to obtain the data set.

[0015] Further, the gas prediction model includes an input layer, a multi-modal time series decomposition encoder, a dynamic graph attention fusion layer, an adaptive collaborative decoder, a multi-scale residual connection, and an output layer. The input layer receives the coal seam change time series data. The multi-modal time series decomposition encoder separates the coal seam change time series data into a trend term, a periodic term, and a residual term and extracts corresponding features. The dynamic graph attention fusion layer constructs a dynamic adjacency matrix based on the extracted features to fuse spatio-temporal features. Finally, the adaptive collaborative decoder generates the predicted gas concentration time series data through dual-path decoding collaboration.

[0016] Further, the multi-modal time series decomposition encoder includes a trend sub-encoder, a periodic sub-encoder, and a residual sub-encoder. The trend sub-encoder uses a TCN network with dilated convolution to capture long-term trends. The periodic sub-encoder adopts a bidirectional quasi-recurrent neural network to capture multi-scale periodic features. The residual sub-encoder combines one-dimensional convolution and a gating mechanism to extract local mutation features.

[0017] Further, the gas prediction model also includes a multi-scale residual connection module. The multi-scale residual connection module constructs three-level skip connections between sub-encoders, directly connects the trend terms, makes weighted connections for the periodic terms, and makes gated connections for the residual terms.

[0018] The gas tunnel ventilation system based on geological simulation and deep learning includes a simulation module, a model training module, a gas prediction module, a ventilation management module, and a dangerous area determination module;

[0019] The simulation module is used to construct a three-dimensional geological model based on geological layer information, and then perform construction simulation on the basis of the three-dimensional geological model according to the construction plan, and obtain the change situation of the geological layer during the simulated construction process through the construction simulation.

[0020] The model training module is used to process the geological layer change situation and gas concentration data to obtain a data set, and use the data set to train and verify the gas prediction model.

[0021] The gas prediction module includes a trained and verified gas prediction model. The gas prediction model is used to predict the gas concentration data in the tunnel during the construction process according to the change situation of the geological layer.

[0022] The ventilation management module adjusts the ventilation intensity in real time during the construction process according to the predicted gas concentration data.

[0023] The dangerous area determination module is used to determine the dangerous areas during the construction process based on the predicted gas concentration data, and guide the construction personnel to prepare corresponding preventive measures in advance according to the situation of the dangerous areas.

[0024] The beneficial effects of the present invention include:

[0025] Based on the geological layer simulation, the present invention explores the correlation between the geological layer changes and the corresponding gas concentration changes. It can predict the gas concentration changes during the construction process only based on the geological conditions of the current construction tunnel, enabling the construction personnel to give early warnings about the locations where dangers may occur during the construction process, prepare corresponding measures in advance, and avoid the dangerous disasters that may be caused by gas outbursts.

[0026] The present invention constructs a three-dimensional model of the geological layer and then obtains the geological layer change data mainly based on coal seams through construction simulation. The geological layer change data is mapped with the gas concentration time series data to train the gas prediction model. Thus, during the actual tunnel construction process, gas tunnel construction simulation can be carried out, and then the construction simulation data is input into the gas prediction model to obtain the gas concentration change situation during the construction process. The dangerous situations such as gas outbursts are predicted through the prediction data, and combined with means such as real-time monitoring by sensors during the construction process, a double construction safety guarantee before and during the construction is formed, further improving the safety of gas tunnel construction.

[0027] In addition, the ventilation intensity in the tunnel is adjusted in real time according to the predicted gas concentration changes, which not only ensures the safety of tunnel construction but also makes the ventilation equipment more in line with the construction needs, avoiding the impact of excessive installation of ventilation equipment on tunnel construction. Description of the Drawings

[0028] Figure 1 It is a flowchart of the gas tunnel ventilation method based on geological simulation and deep learning involved in the embodiments of the present application. Detailed Embodiments

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Therefore, the detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application to be protected, but only represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0030] Embodiment 1

[0031] The following will combine with the attached Figure 1 to make a detailed description of the specific embodiments of the present invention;

[0032] A gas tunnel ventilation method based on geological simulation and deep learning, including:

[0033] Construct a three-dimensional geological model based on the geological layer information of the gas tunnel construction currently required, and then conduct construction simulation to simulate the changes in the geological layer during the construction process. Input the changes in the geological layer into the trained and verified gas prediction model to predict the gas concentration data in the tunnel during the construction process, and adjust the ventilation intensity in real time according to the predicted gas concentration data.

[0034] In another embodiment, it also includes obtaining the gas concentration change trend during the construction process from the predicted gas concentration data, clarifying the time sequence nodes corresponding to the obvious upward trend and mutation of the gas concentration, and then reversely determining the corresponding geological layer position based on the time sequence nodes. The range within a radius r with this geological layer position as the core is set as the dangerous area, and corresponding preventive measures are prepared in advance for the situation of the dangerous area.

[0035] In another embodiment, the geological layer information of the historical gas tunnel construction includes: overall rock layer information, coal seam burial depth, coal seam fault situation, coal seam surrounding rock, coal seam thickness, and nearby hydrogeological conditions.

[0036] The influence of coal seam faults on gas includes many aspects. Near the fault, a series of tectonic fissures are generated in the coal seam and its roof and floor rock layers. The density, scale, and type of the fault all affect the gas emission volume. Normal faults belong to the open type with poor sealing performance, and the fault plane becomes the channel for gas migration. Reverse faults mostly belong to the compressive and compressional-shear types with good sealing performance. It is difficult for gas to migrate through the fault plane. At the same time, the area near the fault plane of the reverse fault becomes a tectonic stress concentration zone, increasing the coal seam pressure and the gas adsorption amount. For faults with a large scale, if the fault throw is large and the fault dislocation and fracture zone is wide, these are all conducive to gas dissipation.

[0037] The influence of coal seam burial depth on gas is specifically as follows: The deeper the burial depth of the coal seam, the longer the distance for the gas in the coal seam to migrate to the surface and the more difficult it is to dissipate; at the same time, the increase in depth also reduces the gas permeability of the coal seam under the action of in-situ stress, which is conducive to the preservation of gas. Due to the increase in coal seam gas pressure, the gas adsorption amount of coal increases, and the coal seam gas content also increases. In the methane zone, when the depth is not large, the gas content of the coal seam increases linearly with the depth.

[0038] The surrounding rock of the coal seam mainly refers to the strata within a certain thickness range including the immediate roof, main roof and immediate floor of the coal seam. The influence of the surrounding rock of the coal seam on gas is as follows: the greater the gas permeability of the coal seam and its surrounding rock, the easier the gas is to escape, and the smaller the gas content of the coal seam; on the contrary, the gas is easy to be preserved, and the gas content of the coal seam is high. Practice has proved that: the thicker the rock strata with low gas permeability in the roof and floor of the coal seam (such as mudstone and fine clastic rock with dense filling) and the larger the proportion it accounts for in the coal-bearing strata, the higher the gas content of the coal seam. On the contrary, when the surrounding rock is composed of thick medium-coarse sandstone, conglomerate or limestone with developed fissures and karst caves, the gas content of the coal seam is small.

[0039] The hydrogeological conditions include factors such as the groundwater level, groundwater pressure and groundwater flow. The influence of hydrogeological conditions on gas is as follows: the level of the groundwater level will affect the accumulation and release of gas in the coal seam. A high water level increases the pressure of gas accumulation, making it more difficult to release, thus increasing the risk of gas outburst. On the contrary, a low water level is conducive to the accumulation and release of gas. Groundwater pressure also has an impact on gas outburst. A higher groundwater pressure will increase the difficulty of gas migration and increase the risk of gas outburst. And groundwater flow will change the distribution and migration path of gas in the coal seam, which may cause gas to accumulate in some areas and increase the risk of gas outburst.

[0040] In another embodiment, the construction of the three-dimensional geological model of the geological layer based on the geological layer information includes: learning the deep geometric shape and spatial distribution characteristics of various types of rock strata in the geological layer through a convolutional neural network, and at the same time performing deep feature fusion on geological exploration data and multi-source data related to geospatial information, and then using the continuous adversarial training of the generative adversarial network to generate a three-dimensional geological model that conforms to the prominent coal seam characteristics. Finally, a best simulation model is trained and fitted to realize the correspondence between the geological layer information and the three-dimensional geological model, for directly generating a three-dimensional geological model according to the geological layer information. Specifically,

[0041] Geological exploration data includes geological mapping data, drilling data and geophysical exploration data;

[0042] Geological mapping data includes the topography and geomorphology, stratigraphic lithology, geological structure, hydrogeological conditions of the tunnel site area, as well as the distribution of strata, coal seam location, thickness, occurrence, and lithology of the roof and floor in the area where the tunnel is located.

[0043] Drilling data includes the coal seam location, color, thickness, structure, fracture development, whether there is a mined-out area, mined-out area location, scale, groundwater conditions, etc. exposed by the main boreholes of the gas tunnel in the coal-bearing strata.

[0044] For the boreholes of the gas tunnel in non-coal-bearing strata, it mainly involves the bubble phenomenon in the drilling fluid, the gas escape location and gas smell during the drilling process.

[0045] Geophysical exploration data includes parts with physical property differences in the tunnel site area, such as fault structures, fold cores, goafs, and accumulated water in old kilns, which are prone to gas and water accumulation.

[0046] Geospatial information includes position coordinates and three-dimensional spatial forms; the position coordinates include coordinate systems (latitude and longitude coordinates or plane projection coordinates of the start and end points of the tunnel and key nodes), elevations (altitude elevations of the tunnel entrance, exit, tunnel roof, and tunnel floor), and burial depths (thicknesses of overburden layers for each section of the tunnel). The three-dimensional spatial forms include linear parameters (plane alignment, longitudinal slope, cross-sectional shape), and geometric models (three-dimensional spatial models based on BIM or GIS, including details such as tunnel axes, section dimensions, and support structures).

[0047] In another embodiment, the training and verification of the gas prediction model include: collecting geological layer information of historical gas tunnel construction and gas concentration data recorded during the construction process, constructing a three-dimensional geological model based on the geological layer information, performing construction simulation on the basis of the three-dimensional geological model according to the construction plan, obtaining the changes in the geological layer during the simulated construction process through the construction simulation, processing the changes in the geological layer and the gas concentration data to obtain a data set, and using the data set to train and verify the gas prediction model.

[0048] The preprocessing of the data set to ensure source-target sequence alignment includes:

[0049] Data cleaning and standardization, removing noise data and unifying the text format to ensure semantic consistency of input and output.

[0050] Word segmentation and encoding, selecting a word segmentation method according to the task, splitting the text into word or sub-word units, constructing vocabulary tables for the source language and the target language respectively, mapping words to digital indices, and at the same time restricting the vocabulary size to reduce the computational burden. Low-frequency words are used <unk>Marker.

[0051] After sequence alignment and padding, after counting the length distribution of the corpus, set the maximum sequence length, and use <pad>Pad to a unified length, truncate long sequences to prevent out-of-memory during training; at the same time, add <sos>and <eos>Marks indicating the start and end of generation.

[0052] Batch generation and tensor conversion, organizing data by batches, shuffling the order to enhance generalization, and converting it into PyTorch / TensorFlow tensors.

[0053] In another embodiment, obtaining the changes in geological layers during the simulated construction process through construction simulation includes using PLAXIS 3D to simulate the changes in the geometric shapes and distributions of each rock layer during the construction process based on a three-dimensional geological model to obtain simulation results, further extracting the changes in the geometric shapes of the coal seam and adjacent rock layers from the simulation results and performing time series processing to obtain the time series data of the changes in the coal seam during the construction progress.

[0054] In another embodiment, processing the data based on the changes in geological layers and gas concentration data to obtain a data set includes: determining several key nodes of the changes in geological layers during the historical construction progress according to the time series data of the changes in the coal seam, then preprocessing the gas concentration data and the time series data of the changes in the coal seam respectively to remove redundant, repeated, and invalid data, and then aligning the preprocessed gas concentration data and the time series data of the changes in the coal seam with several key nodes as the core to obtain a data set.

[0055] In another embodiment, the gas prediction model includes an input layer, a multi-modal time series decomposition encoder, a dynamic graph attention fusion layer, an adaptive collaborative decoder, multi-scale residual connections, and an output layer. The input layer receives the time series data of the changes in the coal seam. The multi-modal time series decomposition encoder separates the time series data of the changes in the coal seam into a trend term, a periodic term, and a residual term and extracts corresponding features. The dynamic graph attention fusion layer constructs a dynamic adjacency matrix based on the extracted features to fuse spatio-temporal features. Finally, the adaptive collaborative decoder generates the predicted time series data of gas concentration through dual-path decoding collaboration.

[0056] Specifically, the trend term is extracted by the moving average method. Let the time series data of the m-th modality of the input be X m =[x m,1 ,x m,2 …,x m,T , and the window size is k. Then the calculation of the trend term T m =[t m,1 ,t m,2 ,…t m,T is as follows:

[0057]

[0058] The extraction of the periodic term includes: first removing the trend term from the original data to obtain the data after the trend term Then group the detrended data by period, calculate the average value of each group, and then expand these average values to the entire time series to obtain the periodic term S m ; The period can be 1 day, 1 week or 1 month, adjusted according to the final prediction effect.

[0059] Residual term extraction involves subtracting trend terms and period terms from the original data.

[0060] Specifically, the trend term mainly represents the long-term change law, such as the increasing trend of gas leakage caused by the increase of 3°C in geothermal gradient for every 100-meter increase in coal mining depth; the change of gas permeability caused by the redistribution of coal body stress during the advancement of the coal mining face. The periodic term mainly represents repetitive fluctuations, such as the periodic changes in tunnel ventilation caused by the three-shift operation system every day, which causes the gas concentration to fluctuate day and night; seasonal rainfall causes changes in coal seam moisture content, which in turn affects the quarterly cycle of gas adsorption-desorption equilibrium. The residual term mainly represents random fluctuations, such as the instantaneous gas outburst caused by the sudden collision between the coal mining machine pick and the interlayer; the sudden change of air pressure disturbance caused by the sudden breakthrough of adjacent tunnels.

[0061] The adaptive collaborative decoder includes a dual-path decoding structure: a main decoding path: generating prediction results based on dynamic attention context; an auxiliary decoding path: generating future geological state discrimination signals through adversarial training; at the same time, a teacher forcing mechanism is introduced. The teacher forcing mechanism is a method for training recurrent neural networks, especially in sequence-to-sequence models. The core idea is that during training, the model's own predicted output is not used as the input for the next step, but the real and correct label is used as the input for the next step.

[0062] The dynamic graph attention fusion layer extracts information features of modalities from multiple dimensions based on the multimodal representation of the multi-head attention mechanism. The core of the attention layer is the feature extractor, which captures potential contextual features based on the entire sequence. Each embedding contains other embedded information, which can generate meaningful features and remove irrelevant noise.

[0063] Specifically, it includes calculating the similarity matrix S∈R between modalities through cosine similarity M×M , the specific cosine similarity calculation is:

[0064]

[0065] In the formula, h i and h j are the eigenvectors of the i-th and j-th modes respectively.

[0066] Then for each modality m, calculate its attention weight to other modalities:

[0067]

[0068] Where a∈R 2H is the learnable parameter vector of the attention layer, [h m ‖h n represents the concatenation of two vectors.

[0069] Finally, feature fusion is performed:

[0070] In another embodiment, the multi-modal time series decomposition encoder includes a trend sub-encoder, a periodic sub-encoder, and a residual sub-encoder to encode the extracted trend term, periodic term, and residual term respectively. The trend sub-encoder uses a TCN network with dilated convolutions to capture long-term trends. The periodic sub-encoder adopts a bidirectional quasi-recurrent neural network to capture multi-scale periodic features. The residual sub-encoder combines one-dimensional convolution and gating mechanism to extract local mutation features.

[0071] In another embodiment, the gas prediction model further includes a multi-scale residual connection module. The multi-scale residual connection module constructs three-level skip connections between sub-encoders, directly connects the trend term, performs weighted connection on the periodic term, and performs gated connection on the residual term.

[0072] In another embodiment, a gas tunnel ventilation system based on geological simulation and deep learning is involved, including a simulation module, a model training module, a gas prediction module, a ventilation management module, and a dangerous area determination module;

[0073] The simulation module is used to construct a three-dimensional geological model based on geological layer information, and then perform construction simulation on the basis of the three-dimensional geological model according to the construction plan, and obtain the change situation of the geological layer during the simulated construction process through the construction simulation.

[0074] The model training module is used to process the change situation of the geological layer and gas concentration data to obtain a data set, and use the data set to train and verify the gas prediction model.

[0075] The gas prediction module includes a trained and verified gas prediction model, and the gas prediction model is used to predict the gas concentration data in the tunnel during the construction process according to the change situation of the geological layer.

[0076] The ventilation management module adjusts the ventilation intensity in real time during the construction process according to the predicted gas concentration data.

[0077] The dangerous area determination module is used to determine the dangerous area during the construction process based on the predicted gas concentration data, and guide the construction personnel to prepare corresponding preventive measures in advance according to the situation of the dangerous area.

[0078] The above-described embodiments merely represent the specific implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the protection scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the technical solution of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application.< / eos> < / sos> at the beginning and end of the target sequence <sos>and <eos>Marks indicating the start and end of generation.

[0052] Batch generation and tensor conversion, organizing data by batches, shuffling the order to enhance generalization, and converting it into PyTorch / TensorFlow tensors.

[0053] In another embodiment, obtaining the changes in geological layers during the simulated construction process through construction simulation includes using PLAXIS 3D to simulate the changes in the geometric shapes and distributions of each rock layer during the construction process based on a three-dimensional geological model to obtain simulation results, further extracting the changes in the geometric shapes of the coal seam and adjacent rock layers from the simulation results and performing time series processing to obtain the time series data of the changes in the coal seam during the construction progress.

[0054] In another embodiment, processing the data based on the changes in geological layers and gas concentration data to obtain a data set includes: determining several key nodes of the changes in geological layers during the historical construction progress according to the time series data of the changes in the coal seam, then preprocessing the gas concentration data and the time series data of the changes in the coal seam respectively to remove redundant, repeated, and invalid data, and then aligning the preprocessed gas concentration data and the time series data of the changes in the coal seam with several key nodes as the core to obtain a data set.

[0055] In another embodiment, the gas prediction model includes an input layer, a multi-modal time series decomposition encoder, a dynamic graph attention fusion layer, an adaptive collaborative decoder, multi-scale residual connections, and an output layer. The input layer receives the time series data of the changes in the coal seam. The multi-modal time series decomposition encoder separates the time series data of the changes in the coal seam into a trend term, a periodic term, and a residual term and extracts corresponding features. The dynamic graph attention fusion layer constructs a dynamic adjacency matrix based on the extracted features to fuse spatio-temporal features. Finally, the adaptive collaborative decoder generates the predicted time series data of gas concentration through dual-path decoding collaboration.

[0056] Specifically, the trend term is extracted by the moving average method. Let the time series data of the m-th modality of the input be X m =[x m,1 ,x m,2 …,x m,T , and the window size is k. Then the calculation of the trend term T m =[t m,1 ,t m,2 ,…t m,T is as follows:

[0057]

[0058] The extraction of the periodic term includes: first removing the trend term from the original data to obtain the data after the trend term Then group the detrended data by period, calculate the average value of each group, and then expand these average values to the entire time series to obtain the periodic term S m ; The period can be 1 day, 1 week or 1 month, adjusted according to the final prediction effect.

[0059] Residual term extraction involves subtracting trend terms and period terms from the original data.

[0060] Specifically, the trend term mainly represents the long-term change law, such as the increasing trend of gas leakage caused by the increase of 3°C in geothermal gradient for every 100-meter increase in coal mining depth; the change of gas permeability caused by the redistribution of coal body stress during the advancement of the coal mining face. The periodic term mainly represents repetitive fluctuations, such as the periodic changes in tunnel ventilation caused by the three-shift operation system every day, which causes the gas concentration to fluctuate day and night; seasonal rainfall causes changes in coal seam moisture content, which in turn affects the quarterly cycle of gas adsorption-desorption equilibrium. The residual term mainly represents random fluctuations, such as the instantaneous gas outburst caused by the sudden collision between the coal mining machine pick and the interlayer; the sudden change of air pressure disturbance caused by the sudden breakthrough of adjacent tunnels.

[0061] The adaptive collaborative decoder includes a dual-path decoding structure: a main decoding path: generating prediction results based on dynamic attention context; an auxiliary decoding path: generating future geological state discrimination signals through adversarial training; at the same time, a teacher forcing mechanism is introduced. The teacher forcing mechanism is a method for training recurrent neural networks, especially in sequence-to-sequence models. The core idea is that during training, the model's own predicted output is not used as the input for the next step, but the real and correct label is used as the input for the next step.

[0062] The dynamic graph attention fusion layer extracts information features of modalities from multiple dimensions based on the multimodal representation of the multi-head attention mechanism. The core of the attention layer is the feature extractor, which captures potential contextual features based on the entire sequence. Each embedding contains other embedded information, which can generate meaningful features and remove irrelevant noise.

[0063] Specifically, it includes calculating the similarity matrix S∈R between modalities through cosine similarity M×M , the specific cosine similarity calculation is:

[0064]

[0065] In the formula, h i and h j are the eigenvectors of the i-th and j-th modes respectively.

[0066] Then for each modality m, calculate its attention weight to other modalities:

[0067]

[0068] Where a∈R 2H is the learnable parameter vector of the attention layer, [h m ‖h n represents the concatenation of two vectors.

[0069] Finally, feature fusion is performed:

[0070] In another embodiment, the multi-modal time series decomposition encoder includes a trend sub-encoder, a periodic sub-encoder, and a residual sub-encoder to encode the extracted trend term, periodic term, and residual term respectively. The trend sub-encoder uses a TCN network with dilated convolutions to capture long-term trends. The periodic sub-encoder adopts a bidirectional quasi-recurrent neural network to capture multi-scale periodic features. The residual sub-encoder combines one-dimensional convolution and gating mechanism to extract local mutation features.

[0071] In another embodiment, the gas prediction model further includes a multi-scale residual connection module. The multi-scale residual connection module constructs three-level skip connections between sub-encoders, directly connects the trend term, performs weighted connection on the periodic term, and performs gated connection on the residual term.

[0072] In another embodiment, a gas tunnel ventilation system based on geological simulation and deep learning is involved, including a simulation module, a model training module, a gas prediction module, a ventilation management module, and a dangerous area determination module;

[0073] The simulation module is used to construct a three-dimensional geological model based on geological layer information, and then perform construction simulation on the basis of the three-dimensional geological model according to the construction plan, and obtain the change situation of the geological layer during the simulated construction process through the construction simulation.

[0074] The model training module is used to process the change situation of the geological layer and gas concentration data to obtain a data set, and use the data set to train and verify the gas prediction model.

[0075] The gas prediction module includes a trained and verified gas prediction model, and the gas prediction model is used to predict the gas concentration data in the tunnel during the construction process according to the change situation of the geological layer.

[0076] The ventilation management module adjusts the ventilation intensity in real time during the construction process according to the predicted gas concentration data.

[0077] The dangerous area determination module is used to determine the dangerous area during the construction process based on the predicted gas concentration data, and guide the construction personnel to prepare corresponding preventive measures in advance according to the situation of the dangerous area.

[0078] The above-described embodiments merely represent the specific implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the protection scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the technical solution of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application.< / eos> < / sos> < / pad> < / unk>

Claims

1. Gas tunnel ventilation system based on geological simulation and deep learning, characterized by: include: A three-dimensional geological model is constructed based on the geological layer information where gas tunnel construction is currently required, and then a construction simulation is carried out to simulate the changes in the geological layer during the construction process. The changes in the geological layer are input into the trained and verified gas prediction model to predict the gas concentration data in the tunnel during the construction process. The ventilation intensity is adjusted in real time during the construction process according to the predicted gas concentration data.

2. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 1 is characterized in that: It also includes obtaining the changing trend of gas concentration during the construction process from the predicted gas concentration data, clarifying the corresponding time nodes when the gas concentration shows an obvious upward trend and a sudden change, and then reversing the corresponding geological layer position based on the time node, setting the range of radius r with the geological layer position as the core as the danger zone, and preparing corresponding preventive measures in advance for the situation in the danger zone.

3. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 1, characterized in that: The method of constructing a three-dimensional geological layer model based on geological layer information includes: learning the deep geometric morphology and spatial distribution characteristics of various types of rock layers in the geological layer through a convolutional neural network, and at the same time performing deep feature fusion on geological survey data and multi-source data related to geographic spatial information, and then using continuous adversarial training of a generative adversarial network to generate a three-dimensional geological model that meets the characteristics of the highlighted coal seam. Finally, an optimal simulation model is fitted through training to achieve the correspondence between the geological layer information and the three-dimensional geological model, so as to directly generate a three-dimensional geological model based on the geological layer information.

4. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 1 is characterized in that: The training and verification of the gas prediction model includes: collecting geological layer information of historical gas tunnel construction and gas concentration data recorded during the construction process, constructing a three-dimensional geological model based on the geological layer information, performing construction simulation on the basis of the three-dimensional geological model based on the construction plan, obtaining changes in the geological layer during the simulated construction process through construction simulation, processing based on the changes in the geological layer and the gas concentration data to obtain a data set, and using the data set to train and verify the gas prediction model.

5. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 1, characterized in that: The obtaining of changes in geological strata during the construction process through construction simulation includes obtaining simulation results by simulating the changes in the geometry and distribution of each rock stratum during the construction process using PLAXIS 3D on the basis of a three-dimensional geological model, further extracting the changes in the geometry of the coal seam and adjacent rock strata from the simulation results and performing time series processing to obtain time series data of changes in the coal seam during construction progress.

6. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 1, characterized in that: The data set obtained by processing the geological layer changes and gas concentration data includes: determining several key nodes of geological layer changes in historical construction progress based on the coal seam change time series data, and then preprocessing the gas concentration data and coal seam change time series data respectively to remove redundant, repeated and invalid data, and then aligning the preprocessed gas concentration data and coal seam change time series data with several key nodes as the core to obtain a data set.

7. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 1, characterized in that: The gas prediction model includes an input layer, a multimodal time series decomposition encoder, a dynamic graph attention fusion layer, an adaptive collaborative decoder, a multi-scale residual connection and an output layer. The input layer receives the coal seam change time series data. The multimodal time series decomposition encoder separates the coal seam change time series data into trend items, period items and residual items and extracts corresponding features. The dynamic graph attention fusion layer constructs a dynamic adjacency matrix based on the extracted features to fuse the spatiotemporal features. Finally, the adaptive collaborative decoder collaboratively generates the predicted gas concentration time series data through dual-path decoding.

8. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 7, characterized in that: The multimodal time series decomposition encoder includes a trend sub-encoder, a period sub-encoder and a residual sub-encoder. The trend sub-encoder uses a TCN network with hole convolution to capture long-term trends, the period sub-encoder uses a bidirectional quasi-recurrent neural network to capture multi-scale periodic features, and the residual sub-encoder combines one-dimensional convolution with a gating mechanism to extract local mutation features.

9. The gas tunnel ventilation method based on geological simulation and deep learning according to claim 8, characterized in that: The gas prediction model also includes a multi-scale residual connection module, which constructs a three-level skip connection between sub-encoders, directly connects the trend terms, performs weighted connections on the period terms, and performs gated connections on the residual terms.

10. Gas tunnel ventilation system based on geological simulation and deep learning, characterized by: It includes simulation module, model training module, gas prediction module, ventilation management module and hazardous area determination module; The simulation module is used to construct a three-dimensional geological model based on geological layer information, and then perform construction simulation on the three-dimensional geological model based on the construction plan, and obtain the changes of the geological layer during the construction process through construction simulation. The model training module is used to process the data set based on the change of geological layers and gas concentration data, and use the data set to train and verify the gas prediction model; The gas prediction module includes a trained and verified gas prediction model, which is used to predict the gas concentration data in the tunnel during the construction process according to the changes in the geological layer; The ventilation management module adjusts the ventilation intensity in real time during the construction process according to the predicted gas concentration data; The hazardous area determination module is used to determine hazardous areas during the construction process based on the predicted gas concentration data, and to guide construction personnel to prepare corresponding preventive measures in advance according to the situation in the hazardous areas.

Images