A data collection system for fracture pressure in coalbed methane reservoirs

Abstract

This invention discloses a coalbed methane reservoir fracture pressure data collection system, comprising: a data acquisition layer: acquiring pressure, temperature, strain, and acoustic signals through a fiber Bragg grating array and piezoelectric sensors, and uploading them via LoRa network in the form of structured data packets after local filtering; a data transmission layer: edge nodes clean and compress data in real time, attach metadata tags, and prioritize the transmission of high-priority alarm signals through 5G channels and QoS routing; a data processing layer: fusing real-time data and geological information based on an ST-CNN dual-stream architecture, initializing the model through transfer learning and dynamically updating parameters, and adaptively regularizing generalization. It extracts spatiotemporal features and geological topological features to generate prediction maps, trends, and risk indicators, triggering early warning outputs visualized data, decision suggestions, and digital twin instructions to the application layer; and an application layer: integrating an intelligent decision-making platform and a digital twin module, supporting dynamic threshold setting, strategy optimization, and reverse adjustment of model parameters.

Description

Technical Field

[0001] This invention relates to the field of data collection technology, and in particular to a data collection system based on fracture pressure in coalbed methane reservoirs. Background Technology

[0002] Coalbed methane (CBM) reservoirs are unconventional natural gas resources primarily composed of methane, found within coal seams. They exist mainly in adsorbed form with some free form, exhibiting geological characteristics such as low permeability, strong heterogeneity, and a dual-pore structure. my country's CBM geological resources amount to approximately 37 trillion cubic meters. Their formation is closely related to coalification, with gas production beginning in the lignite stage and reaching maturity in the bituminous coal stage. Their calorific value is comparable to conventional natural gas, and combustion produces almost no pollution, making them an important supplement to clean energy. CBM development has a triple strategic value: ensuring mine safety, optimizing the energy structure, and reducing greenhouse gas emissions.

[0003] However, in the development of deep high-rank coal reservoirs, the problem of poor dynamic adaptability of fracture pressure data collection is particularly prominent. In the fracturing operation of a deep coalbed methane well in the Ordos Basin, the traditional pressure monitoring system experienced signal saturation distortion due to the sudden increase in reservoir pressure. The traditional static model could not dynamically calibrate the pressure acquisition gain, resulting in a fracture pressure gradient calculation error of up to 25%, which caused the optimization of fracturing parameters to lag and caused serious losses. Therefore, a fracture pressure data collection system based on coalbed methane reservoirs is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a coalbed methane reservoir fracture pressure data collection system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A coalbed methane reservoir fracture pressure data collection system includes:

[0007] Data acquisition layer: Multimodal dynamic sensing is performed through fiber Bragg grating (FBG) array and intelligent piezoelectric sensor to collect pressure, temperature, strain and sound wave signals in real time. After local filtering, the data is uploaded to the data transmission layer as structured data packets.

[0008] Data transmission layer: Edge computing nodes perform real-time cleaning, compression, and anomaly detection on the data, attach metadata tags, and prioritize the transmission of high-priority alarm signals to the data processing layer through 5G high-speed channels and QoS routing optimization. QoS stands for Quality of Service, which is one of the core basic concepts in the field of network communication. Its core goal is to differentiate and manage network resources based on the priority, latency, bandwidth, and other requirements of data streams.

[0009] Data processing layer: Based on the ST-CNN dual-stream architecture model, it integrates real-time monitoring data and geological prior information; it initializes the model through transfer learning, dynamically updates parameters by combining online incremental learning, and adaptively regularizes to balance generalization ability; it extracts spatiotemporal features and geological topological features in real time, and after dynamic fusion through an attention mechanism, it generates crack propagation prediction maps, pressure evolution trends and risk area identifiers, triggers adaptive early warning, and outputs three types of information to the application layer: visualized data, decision suggestions, and digital twin instructions, supporting real-time collaborative optimization. ST-CNN is an abbreviation for Spatio-Temporal Convolutional Neural Network.

[0010] Application layer: Information is integrated and displayed through an adaptive intelligent decision-making platform, supporting dynamic threshold setting and real-time strategy optimization. It also integrates a digital twin module to drive the real-time synchronization of the three-dimensional geological model, simulates the fracture propagation effect under different fracturing parameters, and generates dynamic decision reports. At the same time, it adjusts the model parameters of the processing layer or the QoS priority of the transmission layer in reverse.

[0011] The above technical solution further includes:

[0012] Furthermore, the fiber Bragg grating (FBG) array and the smart piezoelectric sensor perform multimodal dynamic sensing, including the following steps:

[0013] Fiber Bragg grating arrays sense strain and temperature by varying Bragg wavelengths, and their built-in microprocessors continuously monitor ambient temperature and pressure fluctuations. When temperature changes exceed ±5℃ or pressure surges exceed preset thresholds, an adaptive calibration procedure is triggered. For example, wavelength drift is corrected by comparing with a reference grating signal, and the effects of strain and temperature are quantified using the elastic-optical coefficient, thermal expansion coefficient, and thermo-optical coefficient. Intelligent piezoelectric sensors simultaneously detect acoustic signals (such as microseismic events) and adapt to changes in geological conditions by adjusting the sampling frequency and range.

[0014] The piezoelectric sensor outputs through a dual-channel capacitance-resistance circuit to decouple temperature / pressure signals. After local filtering, the data is uploaded to the wellhead edge node via the LoRa network in the form of a timestamped structured data packet.

[0015] The sensor's built-in microprocessor adjusts operating parameters in real time. For example, when the coefficient of variation of coal seam permeability increases, the fiber Bragg grating array automatically increases the sampling frequency to capture micro-strain; the piezoelectric sensor reduces gain in the high-pressure region to avoid signal saturation and increases sensitivity in the low-pressure region to detect minute vibrations.

[0016] Furthermore, the step of prioritizing the transmission of high-priority alarm signals to the data processing layer via 5G high-speed channels and QoS routing optimization includes the following steps:

[0017] Edge computing nodes parse the structured data packets uploaded by the acquisition layer in real time and identify high-priority alarm signals. When the fiber Bragg grating array detects a sudden increase in local pressure in the coal seam that exceeds the threshold, the system automatically marks the signal as "high priority" and adds metadata tags such as timestamp, geological location, and signal strength.

[0018] Edge nodes adjust transmission paths by invoking QoS dynamic routing algorithms (such as weighted fair queuing + priority queues) based on real-time network status and signal priority. High-priority signals are allocated to dedicated high-speed channels, while ordinary data is transmitted through low-power networks. When 5G network bandwidth is sufficient, high-priority signals are transmitted directly through 5G high-speed channels. If the 5G load is too high, the system automatically switches to a backup path and performs local preprocessing (such as data compression and feature extraction) through edge computing nodes to ensure that critical information is not lost.

[0019] After receiving the transmitted data, the data processing layer verifies the data integrity. If data loss or corruption is detected, the system automatically triggers a retransmission mechanism and optimizes the transmission path through QoS routing. Simultaneously, edge nodes report the transmission status (such as success / failure) to the acquisition layer and dynamically adjust the sensor sampling strategy (such as increasing the sampling frequency in high-risk areas). This is done through a three-step mechanism of threshold triggering, step size adjustment, and convergence verification. When the trigger frequency of crack propagation warnings in a certain area exceeds a preset threshold, the system increases the sampling frequency of the fiber Bragg grating array in that area to the target value by a fixed step size and prioritizes the transmission of high-risk data through QoS routing. After adjustment, the system continuously monitors the warning trigger rate. If the trigger frequency remains stable within ±10% of the target threshold or the false alarm rate is less than 3% within multiple consecutive verification cycles, it is determined to be in a convergence state, the frequency adjustment is stopped, and the current sampling parameters are fixed. For example, if crack propagation warnings are frequently triggered in a certain area, the system automatically increases the sampling frequency of the fiber Bragg grating array in that area to 1kHz and prioritizes the transmission of relevant data.

[0020] Furthermore, the ST-CNN-based spatiotemporal convolutional neural network dual-stream architecture model, which integrates real-time monitoring data and prior geological information, includes the following steps:

[0021] The real-time data stream acquired by the data acquisition layer is input into the 3D convolutional branch of ST-CNN, and the spatiotemporal features of the dynamic signal are extracted through the spatiotemporal convolution kernel. For example, the propagation speed of microseismic events is quantified by the optical flow estimation module, capturing the millimeter-level spatial trajectory and millisecond-level temporal changes of crack propagation, and generating a three-dimensional feature map containing "pressure fluctuation - strain distribution - sound wave propagation".

[0022] Fixed geological information, such as well logging curves (e.g., resistivity, porosity), 3D geological models (e.g., permeability distribution maps), and historical fracturing cases, is format-converted and coordinate-calibrated before being input into a graph convolutional network (GCN) to generate geological prior flow. The geological prior flow extracts the topological features of the coal seam geological network through the GCN. The adjacency matrix is ​​used to quantify the geological correlations of different regions (e.g., the spatial distribution of high-permeability and low-permeability zones). The geological attributes (e.g., permeability, geostress) of adjacent nodes are aggregated through graph convolutional layers to generate a topological feature map containing "geological structure-historical patterns".

[0023] ST-CNN dual-stream architecture parameter settings:

[0024] 3D convolution branches:

[0025] Kernel size: 3×3×3 (temporal×spatial×channel), stride 1, padding "same";

[0026] Number of layers: 4 (including 2 residual blocks), activation function is Leaky ReLU (negative slope 0.01);

[0027] Pooling layer: Max pooling (2×2×2), used for dimensionality reduction to extract key spatiotemporal features;

[0028] GCN branch:

[0029] Graph convolutional layers: 3 layers, each layer outputs 64→128→256 dimensions, and the adjacency matrix is ​​dynamically constructed based on the coefficient of variation of coal seam permeability;

[0030] Activation functions: Sigmoid is used for topological feature extraction, and ReLU is used for feature fusion layers;

[0031] Attention mechanism:

[0032] Multi-head self-attention (8 heads), each head with 32 dimensions, calculates feature correlation through Q / K / V matrix;

[0033] Weighted fusion formula: ,in, For the output of the i-th attention head, For the attention head dimension, In this context, k explicitly represents K (bond), Q· To query the dot product of the key vector matrix (Q) and the transpose of the key vector matrix (K), where V is the value vector matrix, , , , , , X is the parameter matrix learned during model training. X is a comprehensive feature matrix composed of N spatiotemporal features and M geological topological features. The similarity matrix between features is calculated (such as the correlation strength between spatiotemporal features and geological topology).

[0034] For example: geological parameters:

[0035] Coal seam thickness: 10.2m, permeability variation coefficient: 0.53, geostress difference: 10MPa;

[0036] Porosity: 8.5%, gas content: 22 m³ / t, fault density: 0.3 faults / km²;

[0037] Model input:

[0038] Real-time data: Strain (±500με), temperature (±2℃), and acoustic wave signal (50-500Hz) from a fiber Bragg grating array.

[0039] Geological prerequisites: 3D geological model (resolution 1m×1m×0.5m), historical fracturing case database (including data from 12 wells);

[0040] Implementation results:

[0041] Fracture propagation prediction: Successfully predicted the fracture propagation path along high-permeability channels, with a 92% match rate with actual microseismic events;

[0042] Risk area identification: Three areas with pressure gradients exceeding the threshold (accounting for 18% of the total area) were identified, triggering a yellow alert;

[0043] Engineering optimization: Based on the prediction results, the fracturing parameters were adjusted (displacement rate was reduced from 8 m³ / min to 6 m³ / min), the fracture length control error was <5%, and the gas production was increased by 17%.

[0044] Furthermore, the process of initializing the model through transfer learning, dynamically updating parameters through online incremental learning, and adaptively regularizing to balance generalization ability includes the following steps:

[0045] Historical fracturing case data (such as well logging curves, fracture propagation records, and pressure evolution trends) are standardized, and a geological feature label library (such as permeability classification and geostress distribution) is constructed. This data is then fed into ST-CNN, enabling the model to have "basic geological feature recognition" capabilities from the initial deployment stage. This allows it to quickly distinguish fracture propagation patterns in high-permeability and low-permeability zones, shortening the cold start time. The geological prior flow branch in the dual-flow architecture is adopted, and topological features (such as geological structure correlations) in historical data are extracted through graph convolutional networks. The model weights are then adjusted using fine-tuning strategies in transfer learning, enabling the model to initially grasp the "correlation between coal seam geological patterns and fracture propagation."

[0046] The data processing layer continuously receives real-time multimodal signals (pressure / temperature / strain / sound waves) from the data acquisition layer; when a new fracture propagation pattern is detected (such as a "nonlinear propagation trajectory" that has not appeared in historical data), the ST-CNN updates its parameters; for example, when a new fracture bifurcation pattern is found in a high-pressure area, the model will adjust the convolution kernel weights of the corresponding region while retaining the learned geological features (such as permeability distribution) to avoid catastrophic forgetting.

[0047] New experience is continuously accumulated through a closed loop of "model update - prediction verification - strategy optimization". When the model predicts that the fracture propagation rate exceeds the threshold, the system automatically triggers an alarm and adjusts the pumping parameters. At the same time, the actual fracturing effect is fed back to the model to further optimize the prediction accuracy.

[0048] The system monitors the noise level of the data in real time through a dynamic noise detection module at the edge nodes (such as noise level assessment based on wavelet transform). When the noise is high, the model automatically increases the L2 regularization coefficient to suppress the risk of overfitting. When the data quality is stable, the regularization coefficient is reduced to improve the model's ability to capture subtle features.

[0049] Through adaptive regularization, the model can automatically balance complexity and generalization ability in scenarios of "high noise-low quality data" and "low noise-high quality data". In areas with large coefficient of variation of coal seam permeability, the model maintains accurate prediction of fracture propagation path through dynamic regularization, while avoiding misjudgment caused by data noise.

[0050] Furthermore, the real-time extraction of spatiotemporal features and geological topological features, after dynamic fusion via an attention mechanism, generates a fracture propagation prediction map, pressure evolution trend, and risk area identification, triggering an adaptive early warning, including the following steps:

[0051] A self-attention mechanism is employed to calculate the local similarity between spatiotemporal features (such as fracture tip displacement field) and geological topological features (such as coal seam porosity distribution) using cosine similarity or Euclidean distance, generating an initial weight matrix. This initial matrix is ​​then combined with real-time monitoring data (such as pressure surge signals) to trigger weight redistribution. For example, when an abnormal acoustic event is detected, the weight of the corresponding geological topological features is increased, enhancing the accuracy of risk area identification. The fusion formula is as follows:

[0052] ;

[0053] in, Let i be the dynamic weight of the i-th spatiotemporal feature. The dynamic weight of the j-th geological topological feature is calculated in real time by the attention mechanism based on the importance of the feature; Spatiotemporal characteristics, Let N be the i-th spatiotemporal feature (e.g., crack propagation rate), and N be the number of spatiotemporal features. Geological and topological features, Let be the j-th geological topological feature (e.g., fault distribution density), M be the number of geological topological features, and σ be the activation function (e.g., Sigmoid or ReLU) to ensure that the fusion result is non-linearly separable.

[0054] Based on the spatiotemporal-geological feature map after dual-stream fusion, ST-CNN tracks the displacement field changes at the fracture tip through 3D convolutional kernels. Combined with the geological topological constraints of graph convolutional networks, it generates a fracture propagation prediction map in real time. In areas with a large coefficient of variation in coal seam permeability, the model prioritizes predicting the path of fracture extension along high-permeability channels and quantifies the propagation speed through an optical flow estimation module. A threshold segmentation algorithm is used to identify areas in the prediction map where the pressure gradient exceeds the safety threshold and marks them as high-risk areas.

[0055] Simultaneously, by combining real-time monitoring data with abnormal acoustic signals (such as a sudden increase in the density of microseismic events), an adaptive early warning mechanism is triggered. An initial threshold is determined based on historical data and a noise baseline. For example, in coalbed methane development, an initial safety threshold is set through correlation analysis between geological factors (such as the permeability coefficient of variation) and real-time monitoring data (such as pressure gradients). Referring to the case of ultrasonic metering devices, the threshold value is adjusted according to the step size of flow velocity changes or the magnitude of pressure mutations. For example, when an abnormal acoustic event is detected, the weight of the corresponding area's geological topological features is increased, and the threshold value is correspondingly widened to capture high-risk signals. In the presence of long-term interference (such as continuous pressure fluctuations), the signal amplifier gain is adjusted in conjunction with the threshold value to ensure a stable signal-to-noise ratio. For example, when the signal-to-noise ratio increases by a preset value, the threshold value is reduced to optimize detection sensitivity. A false alarm rate upper limit is set by training the model with historical data. When the adjusted parameters cause the false alarm rate to fall below the threshold, the adjustment stops. The probability density is estimated using kernel density. When the probability of a dynamic pixel exceeds a critical value, it is identified as a foreground point, triggering an early warning. When the area of ​​a risk zone exceeds a preset proportion, the system automatically upgrades the alarm level and prioritizes transmission to the application layer.

[0056] The real-time collected pressure data is mapped onto the surface of a 3D geological model using a contour interpolation algorithm to generate a pressure distribution heat map. High-pressure areas are marked with a red gradient, while low-pressure areas are marked with a blue gradient, providing an intuitive visualization of pressure changes.

[0057] Based on the crack propagation prediction map, the stress release process at the crack tip is simulated using the particle flow algorithm to generate a simulation animation; the animation frame rate is synchronized with the real-time data, enabling on-site engineers to observe the crack propagation trend in real time.

[0058] Combining prediction results with prior geological information, the system generates intelligent decision-making suggestions through a rule engine. When the predicted fracture propagation rate exceeds the threshold, the system automatically recommends "reducing the pumping flow rate by 20%" or "adjusting the fracturing fluid viscosity to 50 mPa·s", and verifies the effectiveness of the suggestions through a historical case library.

[0059] The digital twin module receives the fracture propagation prediction map and risk area markers output by the model. It dynamically adjusts the geometry of the 3D geological model through a mesh deformation algorithm. When the fracture propagation rate predicted by the model exceeds a preset threshold or the area change rate of the risk area exceeds 10%, the mesh deformation algorithm is triggered to adjust the 3D geological model. The adjustment adopts an adaptive step size strategy. Each iteration calculates the deformation amount based on the current error gradient (e.g., an initial step size of 0.1m, which decreases to 0.01m with the number of iterations) until the matching degree between the model geometry and real-time monitoring data (e.g., the distribution of microseismic events and changes in pressure gradient) reaches more than 95% or the error fluctuation of multiple consecutive iterations is less than 2%, which is considered a convergence state. For example, when a fracture is detected to be shifting in a certain direction, the model automatically updates the permeability distribution and geostress field of the corresponding area, so that the digital twin and the physical reservoir state are synchronized in real time.

[0060] The updated digital twin model is integrated and displayed through a GIS platform, supporting real-time collaboration among multiple departments such as geology, engineering, and safety. Geological experts can use the model to verify whether the fracture propagation path meets expectations, and engineering teams can adjust fracturing parameters based on model recommendations.

[0061] Furthermore, the integrated digital twin module drives the three-dimensional geological model to synchronize in real time, simulates the fracture propagation effect under different fracturing parameters, and generates a dynamic decision report, including the following steps:

[0062] The fracture propagation prediction map, pressure evolution trend, and risk area identification output by the data processing layer are injected into the digital twin module in real time; the coordinate data of the fracture propagation prediction map are directly mapped to the grid nodes of the three-dimensional geological model, and the pressure evolution trend curve is synchronized to the model attribute panel to perform real-time data alignment between the physical and digital twins.

[0063] Based on techniques such as mesh deformation algorithms (e.g., free deformation FFD) and attribute interpolation, the geometry and physical properties of the 3D geological model are dynamically updated with real-time data; when a fracture is detected to be shifting in a certain direction, the model automatically adjusts the permeability distribution of the corresponding area.

[0064] Multiple sets of fracturing parameters (such as pump flow rate range of 2-10 m³ / min, fracturing fluid viscosity gradient of 5-50 mPa·s, and proppant concentration gradient of 0.1-1%) are configured in the digital twin platform, and three-dimensional simulation scenarios of different fracturing schemes are generated by parameter scanning; for example, two sets of comparative schemes, "high flow rate + low viscosity" and "low flow rate + high viscosity", are set to simulate the difference in fracture propagation path;

[0065] Simulate the mechanical process of fracture propagation under different parameters. For example, quantify the stress effect of fracturing fluid flow on the fracture wall through a fluid-structure interaction model, predict the final fracture morphology (such as length, width, and number of branches) by combining geological topological constraints, and output quantitative indicators such as pressure evolution curves and fracture propagation rate.

[0066] By combining the pressure gradient, strain distribution, and acoustic anomaly signals in the simulation results, high-risk areas are identified. For example, when the simulation shows that the pressure gradient in a certain area exceeds the safety threshold, the system automatically marks the area and triggers a level-three warning, while generating targeted suggestions such as "reduce pump injection rate" and "adjust fracturing fluid ratio".

[0067] By leveraging the parallel simulation capabilities of the digital twin platform, multiple fracturing parameter schemes can be run simultaneously to compare fracture propagation effects (such as propagation length and branch density), pressure evolution trends, and risk zone areas. For example, the TOPSIS multi-attribute decision algorithm is used to comprehensively evaluate the safety, economy, and efficiency of each scheme, recommend the optimal scheme, and generate a dynamic decision report that includes scheme comparisons, risk assessments, and operational recommendations. The decision report includes 3D visualization results (such as fracture propagation animations and pressure distribution heatmaps), quantitative indicator comparison tables (such as propagation speed and risk zone area), a list of operational recommendations (such as adjusting pumping parameters and optimizing fracturing fluid ratios), and a risk assessment summary.

[0068] Furthermore, the reverse adjustment of the model parameters of the processing layer or the QoS priority of the transport layer includes the following steps:

[0069] The application layer monitors the effectiveness of decision implementation, such as comparing the deviation between the actual crack propagation rate and the predicted value after the implementation of the "reduce pump flow rate" suggestion, whether the pressure evolution trend meets expectations, and whether the risk area has shrunk. If the effect does not meet expectations (such as the crack propagation rate still exceeding the threshold), a reverse adjustment mechanism is triggered.

[0070] The system evaluates the effectiveness of decision-making by combining prior geological information with real-time monitoring data, and dynamically adjusts the L2 regularization coefficient based on the data noise level. If real-time monitoring detects an increase in data noise in a certain area, the system automatically increases the regularization coefficient to suppress the risk of overfitting. If the data quality is stable, the coefficient is reduced to 0.001 to enhance the model's ability to capture subtle features. The learned geological features are retained, while new data features (such as nonlinear fracture propagation trajectories) are incorporated. The adjusted model parameters are fed back to the data processing layer through edge nodes, and the ST-CNN dual-stream architecture is rerun for prediction verification. If the prediction accuracy improves, the adjustment is confirmed to be effective. If the effect is unsatisfactory, a second adjustment is triggered or the model is switched to a backup model.

[0071] Edge nodes monitor the 5G / LoRa network status in real time (such as bandwidth utilization, latency, and packet loss rate). If the 5G network load is too high, causing the transmission delay of high-priority alarm signals to exceed 10ms, the system automatically switches some alarm signals to the LoRa backup path and adjusts the QoS routing policy. Bandwidth is dynamically allocated according to signal priority and network status. When the 5G network is idle, high-priority signals occupy 80% of the bandwidth. When congested, ordinary data bandwidth is automatically compressed to 30% to prioritize the transmission of alarm signals.

[0072] After the transmission layer is adjusted, the application layer monitors whether the transmission delay of the alarm signal is reduced and whether the data integrity meets the standards. If the effect meets the standards, the adjustment is confirmed to be effective. If it does not meet the standards, a second optimization is triggered or the system is switched to an alternative transmission path, forming a closed-loop feedback of "monitoring-adjustment-verification". The adjustments of the processing layer and the transmission layer are coordinated by the application layer. If the prediction accuracy is improved after the model parameters are adjusted, but the transmission delay is still high, the QoS priority of the transmission layer is optimized first. Conversely, if the data quality is improved after the transmission is optimized, the model parameters are further fine-tuned, forming a three-layer collaborative optimization of processing-transmission-application.

[0073] The present invention has the following beneficial effects:

[0074] In this invention, a fiber Bragg grating array and a smart piezoelectric sensor are used to achieve multimodal dynamic sensing. After local filtering, the data is uploaded via LoRa as structured data packets. Edge computing nodes perform real-time cleaning, compression, and anomaly detection, and high-priority signals are prioritized for transmission using 5G high-speed channels and QoS routing optimization. Based on the ST-CNN dual-stream architecture, real-time data and geological prior information are fused. The model is initialized through transfer learning and its parameters are dynamically updated through online incremental learning. With adaptive regularization to balance generalization capabilities, spatiotemporal features and geological topological features are extracted in real time and dynamically fused through an attention mechanism to generate prediction maps, trends, and risk indicators to trigger early warnings and output three types of information to the application layer. An adaptive intelligent decision-making platform supports dynamic threshold setting and strategy optimization, and an integrated digital twin module drives the real-time synchronization of the 3D geological model. The model parameters or QoS priorities are adjusted in reverse to form a full-chain dynamic adaptation mechanism, effectively solving the problem of poor dynamic adaptability in fracture pressure data collection. Attached Figure Description

[0075] Figure 1 This is a system block diagram of a coalbed methane reservoir fracture pressure data collection system proposed in this invention;

[0076] Figure 2 This is a demonstration diagram of a coalbed methane reservoir fracture pressure data collection system proposed in this invention. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] Please see Figures 1-2 As shown, the present invention is a data collection system for fracture pressure in coalbed methane reservoirs, comprising:

[0079] Data acquisition layer: Multimodal dynamic sensing is performed through fiber Bragg grating (FBG) array and intelligent piezoelectric sensor to collect pressure, temperature, strain and sound wave signals in real time. After local filtering, the data is uploaded to the data transmission layer as structured data packets.

[0080] Data transmission layer: Edge computing nodes perform real-time cleaning, compression, and anomaly detection on the data, attach metadata tags, and prioritize the transmission of high-priority alarm signals to the data processing layer through 5G high-speed channels and QoS routing optimization. QoS stands for Quality of Service, which is one of the core basic concepts in the field of network communication. Its core goal is to differentiate and manage network resources based on the priority, latency, bandwidth, and other requirements of data streams.

[0081] Data processing layer: Based on the ST-CNN dual-stream architecture model, it integrates real-time monitoring data and geological prior information; it initializes the model through transfer learning, dynamically updates parameters by combining online incremental learning, and adaptively regularizes to balance generalization ability; it extracts spatiotemporal features and geological topological features in real time, and after dynamic fusion through an attention mechanism, it generates crack propagation prediction maps, pressure evolution trends and risk area identifiers, triggers adaptive early warning, and outputs three types of information to the application layer: visualized data, decision suggestions, and digital twin instructions, supporting real-time collaborative optimization. ST-CNN is an abbreviation for Spatio-Temporal Convolutional Neural Network.

[0082] Application layer: Information is integrated and displayed through an adaptive intelligent decision-making platform, supporting dynamic threshold setting and real-time strategy optimization. It also integrates a digital twin module to drive the real-time synchronization of the three-dimensional geological model, simulates the fracture propagation effect under different fracturing parameters, and generates dynamic decision reports. At the same time, it adjusts the model parameters of the processing layer or the QoS priority of the transmission layer in reverse.

[0083] In one embodiment, the fiber Bragg grating (FBG) array and the smart piezoelectric sensor perform multimodal dynamic sensing, including the following steps:

[0084] Fiber Bragg grating arrays sense strain and temperature by varying Bragg wavelengths, and their built-in microprocessors continuously monitor ambient temperature and pressure fluctuations. When temperature changes exceed ±5℃ or pressure surges exceed preset thresholds, an adaptive calibration procedure is triggered. For example, wavelength drift is corrected by comparing with a reference grating signal, and the effects of strain and temperature are quantified using the elastic-optical coefficient, thermal expansion coefficient, and thermo-optical coefficient. Intelligent piezoelectric sensors simultaneously detect acoustic signals (such as microseismic events) and adapt to changes in geological conditions by adjusting the sampling frequency and range.

[0085] The piezoelectric sensor outputs through a dual-channel capacitance-resistance circuit to decouple temperature / pressure signals. After local filtering, the data is uploaded to the wellhead edge node via the LoRa network in the form of a timestamped structured data packet.

[0086] The sensor's built-in microprocessor adjusts operating parameters in real time. For example, when the coefficient of variation of coal seam permeability increases, the fiber Bragg grating array automatically increases the sampling frequency to capture micro-strain; the piezoelectric sensor reduces gain in the high-pressure region to avoid signal saturation and increases sensitivity in the low-pressure region to detect minute vibrations.

[0087] In one embodiment, the step of prioritizing the transmission of high-priority alarm signals to the data processing layer via 5G high-speed channels and QoS routing optimization includes the following steps:

[0088] Edge computing nodes parse the structured data packets uploaded by the acquisition layer in real time and identify high-priority alarm signals. When the fiber Bragg grating array detects a sudden increase in local pressure in the coal seam that exceeds the threshold, the system automatically marks the signal as "high priority" and adds metadata tags such as timestamp, geological location, and signal strength.

[0089] Edge nodes adjust transmission paths by invoking QoS dynamic routing algorithms (such as weighted fair queuing + priority queues) based on real-time network status and signal priority. High-priority signals are allocated to dedicated high-speed channels, while ordinary data is transmitted through low-power networks. When 5G network bandwidth is sufficient, high-priority signals are transmitted directly through 5G high-speed channels. If the 5G load is too high, the system automatically switches to a backup path and performs local preprocessing (such as data compression and feature extraction) through edge computing nodes to ensure that critical information is not lost.

[0090] After receiving the transmitted data, the data processing layer verifies the data integrity. If data loss or corruption is detected, the system automatically triggers a retransmission mechanism and optimizes the transmission path through QoS routing. Simultaneously, edge nodes report the transmission status (such as success / failure) to the acquisition layer and dynamically adjust the sensor sampling strategy (such as increasing the sampling frequency in high-risk areas). This is done through a three-step mechanism of threshold triggering, step size adjustment, and convergence verification. When the trigger frequency of crack propagation warnings in a certain area exceeds a preset threshold, the system increases the sampling frequency of the fiber Bragg grating array in that area to the target value by a fixed step size and prioritizes the transmission of high-risk data through QoS routing. After adjustment, the system continuously monitors the warning trigger rate. If the trigger frequency remains stable within ±10% of the target threshold or the false alarm rate is less than 3% within multiple consecutive verification cycles, it is determined to be in a convergence state, the frequency adjustment is stopped, and the current sampling parameters are fixed. For example, if crack propagation warnings are frequently triggered in a certain area, the system automatically increases the sampling frequency of the fiber Bragg grating array in that area to 1kHz and prioritizes the transmission of relevant data.

[0091] In one embodiment, the ST-CNN-based spatiotemporal convolutional neural network dual-stream architecture model, which integrates real-time monitoring data and prior geological information, includes the following steps:

[0092] The real-time data stream acquired by the data acquisition layer is input into the 3D convolutional branch of ST-CNN, and the spatiotemporal features of the dynamic signal are extracted through the spatiotemporal convolution kernel. For example, the propagation speed of microseismic events is quantified by the optical flow estimation module, capturing the millimeter-level spatial trajectory and millisecond-level temporal changes of crack propagation, and generating a three-dimensional feature map containing "pressure fluctuation - strain distribution - sound wave propagation".

[0093] Fixed geological information, such as well logging curves (e.g., resistivity, porosity), 3D geological models (e.g., permeability distribution maps), and historical fracturing cases, is format-converted and coordinate-calibrated before being input into a graph convolutional network (GCN) to generate geological prior flow. The geological prior flow extracts the topological features of the coal seam geological network through the GCN. The adjacency matrix is ​​used to quantify the geological correlations of different regions (e.g., the spatial distribution of high-permeability and low-permeability zones). The geological attributes (e.g., permeability, geostress) of adjacent nodes are aggregated through graph convolutional layers to generate a topological feature map containing "geological structure-historical patterns".

[0094] ST-CNN dual-stream architecture parameter settings:

[0095] 3D convolution branches:

[0096] Kernel size: 3×3×3 (temporal×spatial×channel), stride 1, padding "same";

[0097] Number of layers: 4 (including 2 residual blocks), activation function is Leaky ReLU (negative slope 0.01);

[0098] Pooling layer: Max pooling (2×2×2), used for dimensionality reduction to extract key spatiotemporal features;

[0099] GCN branch:

[0100] Graph convolutional layers: 3 layers, each layer outputs 64→128→256 dimensions, and the adjacency matrix is ​​dynamically constructed based on the coefficient of variation of coal seam permeability;

[0101] Activation functions: Sigmoid is used for topological feature extraction, and ReLU is used for feature fusion layers;

[0102] Attention mechanism:

[0103] Multi-head self-attention (8 heads), each head with 32 dimensions, calculates feature correlation through Q / K / V matrix;

[0104] Weighted fusion formula: ,in, For the output of the i-th attention head, For the attention head dimension, In this context, k explicitly represents K (bond), Q· To query the dot product of the key vector matrix (Q) and the transpose of the key vector matrix (K), where V is the value vector matrix, , , , , , X is the parameter matrix learned during model training. X is a comprehensive feature matrix composed of N spatiotemporal features and M geological topological features. The similarity matrix between features is calculated (such as the correlation strength between spatiotemporal features and geological topology).

[0105] For example: geological parameters:

[0106] Coal seam thickness: 10.2m, permeability variation coefficient: 0.53, geostress difference: 10MPa;

[0107] Porosity: 8.5%, gas content: 22 m³ / t, fault density: 0.3 faults / km²;

[0108] Model input:

[0109] Real-time data: Strain (±500με), temperature (±2℃), and acoustic wave signal (50-500Hz) from a fiber Bragg grating array.

[0110] Geological prerequisites: 3D geological model (resolution 1m×1m×0.5m), historical fracturing case database (including data from 12 wells);

[0111] Implementation results:

[0112] Fracture propagation prediction: Successfully predicted the fracture propagation path along high-permeability channels, with a 92% match rate with actual microseismic events;

[0113] Risk area identification: Three areas with pressure gradients exceeding the threshold (accounting for 18% of the total area) were identified, triggering a yellow alert;

[0114] Engineering optimization: Based on the prediction results, the fracturing parameters were adjusted (displacement rate was reduced from 8 m³ / min to 6 m³ / min), the fracture length control error was <5%, and the gas production was increased by 17%.

[0115] In one embodiment, the process of initializing the model through transfer learning, dynamically updating parameters through online incremental learning, and adaptively regularizing to balance generalization ability includes the following steps:

[0116] Historical fracturing case data (such as well logging curves, fracture propagation records, and pressure evolution trends) are standardized, and a geological feature label library (such as permeability classification and geostress distribution) is constructed. This data is then fed into ST-CNN, enabling the model to have "basic geological feature recognition" capabilities from the initial deployment stage. This allows it to quickly distinguish fracture propagation patterns in high-permeability and low-permeability zones, shortening the cold start time. The geological prior flow branch in the dual-flow architecture is adopted, and topological features (such as geological structure correlations) in historical data are extracted through graph convolutional networks. The model weights are then adjusted using fine-tuning strategies in transfer learning, enabling the model to initially grasp the "correlation between coal seam geological patterns and fracture propagation."

[0117] The data processing layer continuously receives real-time multimodal signals (pressure / temperature / strain / sound waves) from the data acquisition layer; when a new fracture propagation pattern is detected (such as a "nonlinear propagation trajectory" that has not appeared in historical data), the ST-CNN updates its parameters; for example, when a new fracture bifurcation pattern is found in a high-pressure area, the model will adjust the convolution kernel weights of the corresponding region while retaining the learned geological features (such as permeability distribution) to avoid catastrophic forgetting.

[0118] New experience is continuously accumulated through a closed loop of "model update - prediction verification - strategy optimization". When the model predicts that the fracture propagation rate exceeds the threshold, the system automatically triggers an alarm and adjusts the pumping parameters. At the same time, the actual fracturing effect is fed back to the model to further optimize the prediction accuracy.

[0119] The system monitors the noise level of the data in real time through a dynamic noise detection module at the edge nodes (such as noise level assessment based on wavelet transform). When the noise is high, the model automatically increases the L2 regularization coefficient to suppress the risk of overfitting. When the data quality is stable, the regularization coefficient is reduced to improve the model's ability to capture subtle features.

[0120] Through adaptive regularization, the model can automatically balance complexity and generalization ability in scenarios of "high noise-low quality data" and "low noise-high quality data". In areas with large coefficient of variation of coal seam permeability, the model maintains accurate prediction of fracture propagation path through dynamic regularization, while avoiding misjudgment caused by data noise.

[0121] In one embodiment, the real-time extraction of spatiotemporal features and geological topological features, dynamically fused through an attention mechanism, generates a fracture propagation prediction map, pressure evolution trend, and risk area identifiers, triggering an adaptive early warning, including the following steps:

[0122] A self-attention mechanism is employed to calculate the local similarity between spatiotemporal features (such as fracture tip displacement field) and geological topological features (such as coal seam porosity distribution) using cosine similarity or Euclidean distance, generating an initial weight matrix. This initial matrix is ​​then combined with real-time monitoring data (such as pressure surge signals) to trigger weight redistribution. For example, when an abnormal acoustic event is detected, the weight of the corresponding geological topological features is increased, enhancing the accuracy of risk area identification. The fusion formula is as follows:

[0123] ;

[0124] in, Let i be the dynamic weight of the i-th spatiotemporal feature. The dynamic weight of the j-th geological topological feature is calculated in real time by the attention mechanism based on the importance of the feature; Spatiotemporal characteristics, Let N be the i-th spatiotemporal feature (e.g., crack propagation rate), and N be the number of spatiotemporal features. Geological and topological features, Let be the j-th geological topological feature (e.g., fault distribution density), M be the number of geological topological features, and σ be the activation function (e.g., Sigmoid or ReLU) to ensure that the fusion result is non-linearly separable.

[0125] Based on the spatiotemporal-geological feature map after dual-stream fusion, ST-CNN tracks the displacement field changes at the fracture tip through 3D convolutional kernels. Combined with the geological topological constraints of graph convolutional networks, it generates a fracture propagation prediction map in real time. In areas with a large coefficient of variation in coal seam permeability, the model prioritizes predicting the path of fracture extension along high-permeability channels and quantifies the propagation speed through an optical flow estimation module. A threshold segmentation algorithm is used to identify areas in the prediction map where the pressure gradient exceeds the safety threshold and marks them as high-risk areas.

[0126] Simultaneously, by combining real-time monitoring data with abnormal acoustic signals (such as a sudden increase in the density of microseismic events), an adaptive early warning mechanism is triggered. An initial threshold is determined based on historical data and a noise baseline. For example, in coalbed methane development, an initial safety threshold is set through correlation analysis between geological factors (such as the permeability coefficient of variation) and real-time monitoring data (such as pressure gradients). Referring to the case of ultrasonic metering devices, the threshold value is adjusted according to the step size of flow velocity changes or the magnitude of pressure mutations. For example, when an abnormal acoustic event is detected, the weight of the corresponding area's geological topological features is increased, and the threshold value is correspondingly widened to capture high-risk signals. In the presence of long-term interference (such as continuous pressure fluctuations), the signal amplifier gain is adjusted in conjunction with the threshold value to ensure a stable signal-to-noise ratio. For example, when the signal-to-noise ratio increases by a preset value, the threshold value is reduced to optimize detection sensitivity. A false alarm rate upper limit is set by training the model with historical data. When the adjusted parameters cause the false alarm rate to fall below the threshold, the adjustment stops. The probability density is estimated using kernel density. When the probability of a dynamic pixel exceeds a critical value, it is identified as a foreground point, triggering an early warning. When the area of ​​a risk zone exceeds a preset proportion, the system automatically upgrades the alarm level and prioritizes transmission to the application layer.

[0127] The real-time collected pressure data is mapped onto the surface of a 3D geological model using a contour interpolation algorithm to generate a pressure distribution heat map. High-pressure areas are marked with a red gradient, while low-pressure areas are marked with a blue gradient, providing an intuitive visualization of pressure changes.

[0128] Based on the crack propagation prediction map, the stress release process at the crack tip is simulated using the particle flow algorithm to generate a simulation animation; the animation frame rate is synchronized with the real-time data, enabling on-site engineers to observe the crack propagation trend in real time.

[0129] Combining prediction results with prior geological information, the system generates intelligent decision-making suggestions through a rule engine. When the predicted fracture propagation rate exceeds the threshold, the system automatically recommends "reducing the pumping flow rate by 20%" or "adjusting the fracturing fluid viscosity to 50 mPa·s", and verifies the effectiveness of the suggestions through a historical case library.

[0130] The digital twin module receives the fracture propagation prediction map and risk area markers output by the model. It dynamically adjusts the geometry of the 3D geological model through a mesh deformation algorithm. When the fracture propagation rate predicted by the model exceeds a preset threshold or the area change rate of the risk area exceeds 10%, the mesh deformation algorithm is triggered to adjust the 3D geological model. The adjustment adopts an adaptive step size strategy. Each iteration calculates the deformation amount based on the current error gradient (e.g., an initial step size of 0.1m, which decreases to 0.01m with the number of iterations) until the matching degree between the model geometry and real-time monitoring data (e.g., the distribution of microseismic events and changes in pressure gradient) reaches more than 95% or the error fluctuation of multiple consecutive iterations is less than 2%, which is considered a convergence state. For example, when a fracture is detected to be shifting in a certain direction, the model automatically updates the permeability distribution and geostress field of the corresponding area, so that the digital twin and the physical reservoir state are synchronized in real time.

[0131] The updated digital twin model is integrated and displayed through a GIS platform, supporting real-time collaboration among multiple departments such as geology, engineering, and safety. Geological experts can use the model to verify whether the fracture propagation path meets expectations, and engineering teams can adjust fracturing parameters based on model recommendations.

[0132] In one embodiment, the integrated digital twin module drives the three-dimensional geological model to synchronize in real time, simulate the fracture propagation effect under different fracturing parameters, and generate a dynamic decision report, including the following steps:

[0133] The fracture propagation prediction map, pressure evolution trend, and risk area identification output by the data processing layer are injected into the digital twin module in real time; the coordinate data of the fracture propagation prediction map are directly mapped to the grid nodes of the three-dimensional geological model, and the pressure evolution trend curve is synchronized to the model attribute panel to perform real-time data alignment between the physical and digital twins.

[0134] Based on techniques such as mesh deformation algorithms (e.g., free deformation FFD) and attribute interpolation, the geometry and physical properties of the 3D geological model are dynamically updated with real-time data; when a fracture is detected to be shifting in a certain direction, the model automatically adjusts the permeability distribution of the corresponding area.

[0135] Multiple sets of fracturing parameters (such as pump flow rate range of 2-10 m³ / min, fracturing fluid viscosity gradient of 5-50 mPa·s, and proppant concentration gradient of 0.1-1%) are configured in the digital twin platform, and three-dimensional simulation scenarios of different fracturing schemes are generated by parameter scanning; for example, two sets of comparative schemes, "high flow rate + low viscosity" and "low flow rate + high viscosity", are set to simulate the difference in fracture propagation path;

[0136] Simulate the mechanical process of fracture propagation under different parameters. For example, quantify the stress effect of fracturing fluid flow on the fracture wall through a fluid-structure interaction model, predict the final fracture morphology (such as length, width, and number of branches) by combining geological topological constraints, and output quantitative indicators such as pressure evolution curves and fracture propagation rate.

[0137] By combining the pressure gradient, strain distribution, and acoustic anomaly signals in the simulation results, high-risk areas are identified. For example, when the simulation shows that the pressure gradient in a certain area exceeds the safety threshold, the system automatically marks the area and triggers a level-three warning, while generating targeted suggestions such as "reduce pump injection rate" and "adjust fracturing fluid ratio".

[0138] By leveraging the parallel simulation capabilities of the digital twin platform, multiple fracturing parameter schemes can be run simultaneously to compare fracture propagation effects (such as propagation length and branch density), pressure evolution trends, and risk zone areas. For example, the TOPSIS multi-attribute decision algorithm is used to comprehensively evaluate the safety, economy, and efficiency of each scheme, recommend the optimal scheme, and generate a dynamic decision report that includes scheme comparisons, risk assessments, and operational recommendations. The decision report includes 3D visualization results (such as fracture propagation animations and pressure distribution heatmaps), quantitative indicator comparison tables (such as propagation speed and risk zone area), a list of operational recommendations (such as adjusting pumping parameters and optimizing fracturing fluid ratios), and a risk assessment summary.

[0139] In one embodiment, the reverse adjustment of the model parameters of the processing layer or the QoS priority of the transport layer includes the following steps:

[0140] The application layer monitors the effectiveness of decision implementation, such as comparing the deviation between the actual crack propagation rate and the predicted value after the implementation of the "reduce pump flow rate" suggestion, whether the pressure evolution trend meets expectations, and whether the risk area has shrunk. If the effect does not meet expectations (such as the crack propagation rate still exceeding the threshold), a reverse adjustment mechanism is triggered.

[0141] The system evaluates the effectiveness of decision-making by combining prior geological information with real-time monitoring data, and dynamically adjusts the L2 regularization coefficient based on the data noise level. If real-time monitoring detects an increase in data noise in a certain area, the system automatically increases the regularization coefficient to suppress the risk of overfitting. If the data quality is stable, the coefficient is reduced to 0.001 to enhance the model's ability to capture subtle features. The learned geological features are retained, while new data features (such as nonlinear fracture propagation trajectories) are incorporated. The adjusted model parameters are fed back to the data processing layer through edge nodes, and the ST-CNN dual-stream architecture is rerun for prediction verification. If the prediction accuracy improves, the adjustment is confirmed to be effective. If the effect is unsatisfactory, a second adjustment is triggered or the model is switched to a backup model.

[0142] Edge nodes monitor the 5G / LoRa network status in real time (such as bandwidth utilization, latency, and packet loss rate). If the 5G network load is too high, causing the transmission delay of high-priority alarm signals to exceed 10ms, the system automatically switches some alarm signals to the LoRa backup path and adjusts the QoS routing policy. Bandwidth is dynamically allocated according to signal priority and network status. When the 5G network is idle, high-priority signals occupy 80% of the bandwidth. When congested, ordinary data bandwidth is automatically compressed to 30% to prioritize the transmission of alarm signals.

[0143] After the transmission layer is adjusted, the application layer monitors whether the transmission delay of the alarm signal is reduced and whether the data integrity meets the standards. If the effect meets the standards, the adjustment is confirmed to be effective. If it does not meet the standards, a second optimization is triggered or the system is switched to an alternative transmission path, forming a closed-loop feedback of "monitoring-adjustment-verification". The adjustments of the processing layer and the transmission layer are coordinated by the application layer. If the prediction accuracy is improved after the model parameters are adjusted, but the transmission delay is still high, the QoS priority of the transmission layer is optimized first. Conversely, if the data quality is improved after the transmission is optimized, the model parameters are further fine-tuned, forming a three-layer collaborative optimization of processing-transmission-application.

[0144] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for collecting fracture pressure data in coalbed methane reservoirs, characterized in that, include: Data acquisition layer: Multimodal dynamic sensing is performed through fiber Bragg grating array and intelligent piezoelectric sensor to collect pressure, temperature, strain and sound wave signals in real time. After local filtering, the data is uploaded to the data transmission layer in the form of structured data packets. Data transmission layer: Edge computing nodes perform real-time cleaning, compression, and anomaly detection on the data, attach metadata tags, and prioritize the transmission of high-priority alarm signals to the data processing layer via 5G high-speed channels and QoS routing optimization. Data Processing Layer: Based on the ST-CNN spatiotemporal convolutional neural network dual-stream architecture model, it integrates real-time monitoring data and geological prior information. Historical fracturing case data, after standardization, is injected into the geological prior stream branch of ST-CNN. Fine-tuning strategies in transfer learning are used to adjust model weights to initialize the model. Real-time multimodal signals are continuously received. When a new fracture propagation pattern is detected, parameters are updated through a closed-loop process of model update-prediction verification-strategy optimization. Dynamic noise detection modules at edge nodes monitor data noise; when noise exceeds a threshold, the L2 regularization coefficient is increased, and when data quality is stable, the regularization coefficient is decreased. The model is initialized through transfer learning, and parameters are dynamically updated using online incremental learning, adaptively balancing generalization capabilities through regularization. In real time, spatiotemporal features and geological topological features are extracted and dynamically fused through an attention mechanism to generate a crack propagation prediction map, pressure evolution trend and risk area identification, trigger adaptive early warning, and output three types of information to the application layer: visualized data, decision suggestions and digital twin instructions. Application Layer: Information is integrated and displayed through an adaptive intelligent decision-making platform, supporting dynamic threshold setting and real-time strategy optimization. It also integrates a digital twin module to drive real-time synchronization of the 3D geological model, simulating the fracture propagation effect under different fracturing parameters and generating dynamic decision reports. The application layer monitors the decision execution effect, dynamically adjusts the L2 regularization coefficient according to the data noise level, and sends the adjusted model parameters back to the data processing layer for prediction verification. Edge nodes monitor the network status in real time. If the 5G network load is detected to be higher than the preset threshold, the high-priority alarm signal is switched to the LoRa backup path, and bandwidth is dynamically allocated according to signal priority and network status. Adjustments to the processing layer and the transmission layer are uniformly coordinated through the application layer.

2. The coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The fiber Bragg grating array and the smart piezoelectric sensor perform multimodal dynamic sensing, including the following steps: The fiber Bragg grating array senses strain and temperature through changes in Bragg wavelength. The built-in microprocessor continuously monitors ambient temperature and pressure fluctuations. When the temperature change exceeds 5°C or the pressure change exceeds a preset threshold, adaptive calibration is triggered. The intelligent piezoelectric sensor synchronously detects the acoustic signal. By adjusting the sampling frequency and range, it adapts to changes in geological conditions. After local filtering, the collected data is uploaded to the wellhead edge node via the LoRa network in the form of a timestamped structured data packet.

3. The coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The method of prioritizing the transmission of high-priority alarm signals to the data processing layer via 5G high-speed channels and QoS routing optimization includes the following steps: Edge computing nodes parse structured data packets uploaded by the acquisition layer in real time and identify high-priority alarm signals. When the fiber Bragg grating array detects a sudden increase in local pressure in the coal seam exceeding a threshold, the system automatically marks the signal as high priority and adds metadata tags. Based on real-time network status and signal priority, edge nodes call QoS dynamic routing algorithms to adjust the transmission path. High-priority signals are allocated to dedicated high-speed channels, while ordinary data is transmitted through low-power networks. After receiving the transmitted data, if the data processing layer detects data loss or corruption, it triggers a retransmission mechanism and adjusts the transmission path through QoS routing optimization. At the same time, edge nodes feed back the transmission status to the acquisition layer and dynamically adjust the sensor sampling strategy.

4. The coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The ST-CNN-based spatiotemporal convolutional neural network dual-stream architecture model, which integrates real-time monitoring data and prior geological information, includes the following steps: The real-time data stream acquired by the data acquisition layer is input into the 3D convolutional branch of ST-CNN to extract the spatiotemporal features of dynamic signals. Fixed geological information such as well logging curves, 3D geological models, and historical fracturing cases are input into the graph convolutional network GCN to generate geological prior flow. The geological prior flow extracts the topological features of the coal seam geological network through GCN. The geological correlation of different regions is quantified by using the adjacency matrix. The geological attributes of adjacent nodes are aggregated through the graph convolutional layer to generate a topological feature map containing geological structure-historical patterns.

5. A coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The process of initializing the model through transfer learning, dynamically updating parameters through online incremental learning, and adaptively regularizing to balance generalization ability includes the following steps: Historical fracturing case data were standardized and a geological feature label library was constructed. The data was then injected into the ST-CNN spatiotemporal convolutional neural network dual-stream architecture model. The geological prior flow branch in the dual-stream architecture was adopted. Topological features in the historical data were extracted through graph convolutional networks, and the model weights were adjusted using fine-tuning strategies in transfer learning. The data processing layer continuously receives real-time multimodal signals from the data acquisition layer. When a new fracture propagation mode is detected, ST-CNN automatically updates its parameters. The update experience is continuously accumulated through a closed loop of model update-prediction verification-strategy optimization. When the model predicts that the fracture propagation rate exceeds the threshold, an alarm is triggered and the pumping parameters are adjusted. At the same time, the actual fracturing effect is fed back to the model to further optimize the prediction accuracy. The dynamic noise detection module of the edge nodes monitors the data noise level in real time. When the noise is high, the model automatically increases the L2 regularization coefficient. When the data quality is stable, the regularization coefficient is decreased.

6. A coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The real-time extraction of spatiotemporal features and geological topological features, dynamically fused through an attention mechanism, generates a fracture propagation prediction map, pressure evolution trend, and risk area identification, triggering an adaptive early warning. This process includes the following steps: Based on the spatiotemporal-geological feature map after dual-stream fusion, ST-CNN tracks the displacement field changes at the fracture tip through 3D convolutional kernels. Combined with the geological topological constraints of graph convolutional networks, it generates a fracture propagation prediction map in real time. In areas with large coefficients of variation of coal seam permeability, it prioritizes predicting the path of fractures extending along high-permeability channels and quantifies the propagation speed. It identifies areas in the prediction map where the pressure gradient exceeds the safety threshold through a threshold segmentation algorithm and marks them as high-risk areas. At the same time, combined with the abnormal acoustic signals in the real-time monitoring data, it triggers an adaptive early warning mechanism. When the area of ​​the risk area exceeds a preset proportion, the alarm level is upgraded and the data is transmitted to the application layer first. Real-time collected pressure data is projected onto the surface of a 3D geological model to generate a dynamically updated pressure distribution heat map. High-pressure areas are marked with a red gradient, while low-pressure areas are marked with a blue gradient. Based on the fracture propagation prediction map, the stress release process at the fracture tip is simulated to generate a dynamic simulation animation. By combining prediction results with prior geological information, intelligent decision-making suggestions are generated through a rule engine. The digital twin module receives the fracture propagation prediction map and risk area identification output by the model, adjusts the geometry of the three-dimensional geological model, and automatically updates the permeability distribution and geostress field of the corresponding area when a fracture is detected to be shifting in a certain direction. This ensures that the digital twin is synchronized with the physical reservoir state in real time. The updated digital twin model is integrated and displayed through the platform, supporting real-time collaboration among multiple departments, including geology, engineering, and safety.

7. A coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The integrated digital twin module drives the three-dimensional geological model to synchronize in real time, simulates the fracture propagation effect under different fracturing parameters, and generates a dynamic decision report, including the following steps: The fracture propagation prediction map, pressure evolution trend, and risk area identification output from the data processing layer are injected into the digital twin module in real time. The geometry and physical properties of the 3D geological model are dynamically updated with real-time data. Multiple sets of fracturing parameters are configured in the digital twin platform, and 3D simulation scenarios of different fracturing schemes are generated through parameter scanning. Based on the coupled algorithm of finite element analysis and computational fluid dynamics, the mechanical process of fracture propagation under different parameters is simulated. By quantifying the stress effect of fracturing fluid flow on the fracture wall, the final fracture morphology is predicted in combination with geological topological constraints, and the pressure evolution curve and quantitative indicators of fracture propagation rate are output. Combined with the pressure gradient, strain distribution, and acoustic anomaly signal in the simulation results, high-risk areas are dynamically identified. When the simulation shows that the pressure gradient of a certain area exceeds the safety threshold, this area is marked and a level 3 warning is triggered, and targeted suggestions are generated at the same time. Multiple fracturing parameter schemes are run simultaneously through a digital twin platform to compare fracture propagation effects, pressure evolution trends, and risk area areas. The safety, economy, and efficiency of each scheme are comprehensively evaluated, the optimal scheme is recommended, and a dynamic decision report is generated that includes scheme comparison, risk assessment, and operational recommendations. The decision report includes three-dimensional visualization results, a quantitative indicator comparison table, a list of operational recommendations, and a risk assessment summary.

8. A coalbed methane reservoir fracture pressure data collection system according to claim 1, characterized in that, The adjustment of the processing layer and the transport layer includes the following steps: The application layer monitors the effectiveness of decision execution. Combining prior geological information with real-time monitoring data, a weighted scoring system is used to evaluate the effectiveness of decisions. The L2 regularization coefficient is dynamically adjusted based on the data noise level. If the real-time monitoring shows an increase in data noise in a certain area, the regularization coefficient is increased. If the data quality is stable, the coefficient is reduced to 0.001, retaining the learned geological features while incorporating new data features. The adjusted model parameters are fed back to the data processing layer through edge nodes, and ST-CNN is rerun for prediction verification. If the prediction accuracy improves, the adjustment is confirmed to be effective. If the effect is not good, a second adjustment is triggered or the model is switched to a backup model. Edge nodes monitor the status of 5G or LoRa networks in real time. If the 5G network load is too high, causing the transmission delay of high-priority alarm signals to be greater than 10ms, some alarm signals will be switched to the LoRa backup path, and the QoS routing policy will be adjusted. Bandwidth will be dynamically allocated according to signal priority and network status. When the 5G network is idle, high-priority signals will occupy 80% of the bandwidth. When congested, ordinary data bandwidth will be automatically compressed to 30% to ensure the transmission of alarm signals. After the transmission layer is adjusted, the application layer monitors whether the transmission delay of the alarm signal is reduced and whether the data integrity meets the standards. If the effect meets the standards, the adjustment is confirmed to be effective. If it does not meet the standards, a second optimization is triggered or the system is switched to an alternative transmission path. The adjustments of the processing layer and the transmission layer are coordinated by the application layer. If the prediction accuracy is improved after the model parameters are adjusted, but the transmission delay is still high, the QoS priority of the transmission layer is optimized first. Conversely, if the data quality is improved after the transmission is optimized, the model parameters are further fine-tuned.

Images