An artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring system and monitoring method

Abstract

The application discloses a kind of based on artificial intelligence's optical fiber coiled tubing downhole fluid monitoring system and monitoring method, including composite modem, coiled tubing body, straight light unit, helical light unit, artificial intelligence decision unit, cloud data center, artificial seismic source module and ground detector array.The application deeply fuses artificial intelligence technology such as deep learning, knowledge graph, graph neural network and distributed optical fiber sensing technology, realizes the intelligent identification of multiphase fluid, the high-precision inversion of production fluid profile, the intelligent modeling of three-dimensional stress field and the automatic tracking and evolution prediction of fluid boundary.Physical information neural network is used to embed physical conservation law into training process, graph neural network is used to construct stress field graph structure, and time-space graph convolution network is used to track time-lapse fluid boundary.A knowledge graph associated with downhole fluid, formation and engineering parameters is constructed to realize closed-loop optimization from monitoring data to intelligent decision-making.

Description

Technical Field

[0001] This invention relates to the fields of well geophysics, intelligent well completion, distributed fiber optic sensing, and artificial intelligence, and particularly to an artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring system and method, which is applicable to long-term real-time dynamic monitoring of oil and gas production wells, intelligent identification of multiphase fluids, intelligent inversion of production profiles, three-dimensional stress field modeling, and well-to-ground time-lapse joint three-dimensional exploration. Background Technology

[0002] Coiled tubing (CT) is an uninterrupted tubing system manufactured using welding or seamless connection techniques. It can reach lengths of several kilometers and has a diameter of several tens of millimeters, exhibiting excellent flexibility. Coiled tubing equipment features pressurized operation and continuous tripping, is compact, has a fast turnaround time, and is low-cost. It is widely used in oilfield drilling, well completion, fracturing, acidizing, fluid drainage, oil testing, oil and gas production, and well workover. Currently, deploying permanent armored fiber optic cables outside metal casing faces challenges such as high construction difficulty, high cost, difficulty in fiber optic cable location detection, and immature downhole radiation avoidance technology. Furthermore, armored fiber optic cables deployed inside casing can only perform temporary data acquisition because they will affect or interfere with other downhole operations, making long-term monitoring within the casing impossible. While armored fiber optic cables can be installed within coiled tubing for microseismic monitoring and production profile measurement in horizontal well sections, this is only a temporary measure. It is impossible to permanently deploy coiled tubing with embedded armored fiber optic cables in horizontal wells without affecting downhole oil and gas production. Furthermore, existing technology has the following key drawbacks:

[0003] (1) Low level of intelligence in data processing: Traditional fiber optic monitoring systems rely on simple threshold judgments or empirical formulas for data interpretation, which cannot make full use of the inherent correlation between multi-source heterogeneous data, resulting in low fluid identification accuracy and large error in liquid production profile inversion.

[0004] (2) Lack of real-time intelligent decision-making capability: The existing system cannot realize real-time intelligent identification and prediction of fluid state, and cannot automatically optimize production system based on monitoring data, requiring a lot of manual intervention.

[0005] (3) Insufficient multi-parameter fusion analysis capability: Data from multiple sources such as DAS, DTS, DSS, and DPS are processed independently, lacking an effective multi-parameter joint inversion and intelligent fusion mechanism, making it difficult to accurately track the dynamic changes of the fluid interface.

[0006] (4) Difficulty in interpreting time-shifted monitoring data: Traditional time-shifted seismic data processing methods rely on manual interpretation, which is inefficient, subjective, and cannot achieve automatic tracking and evolution prediction of fluid boundaries.

[0007] With the rapid development of artificial intelligence technology, deep learning, knowledge graphs, graph neural networks, and other technologies have made breakthroughs in areas such as image recognition, time series prediction, and complex system modeling. Deeply integrating artificial intelligence technology with distributed fiber optic sensing technology to build an intelligent downhole fluid monitoring system is an important way to solve the aforementioned technical challenges. Summary of the Invention

[0008] The main objective of this invention is to address the technical problems in existing fiber optic coiled tubing monitoring systems, such as low intelligence in data processing, lack of real-time intelligent decision-making capabilities, insufficient multi-parameter fusion analysis capabilities, and difficulties in interpreting time-lapse monitoring data. This invention provides an artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring system and method. By deeply integrating artificial intelligence technologies such as deep learning, knowledge graphs, and graph neural networks into the distributed fiber optic sensor data processing flow, it achieves intelligent identification of multiphase fluids, high-precision inversion of production profiles, intelligent modeling of three-dimensional stress fields, and automatic tracking and evolution prediction of fluid boundaries, providing intelligent means for the scientific management of oil and gas reservoirs and improving oil recovery.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] An artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring system includes: a composite modulator / demodulator, a coiled tubing body, a straight optical unit, a spiral optical unit, an artificial intelligence decision-making unit, a cloud data center, an artificial seismic source module, and a ground detector array;

[0011] The composite modulator / demodulator is located near the wellhead and contains a DAS signal input module, a DTS signal input module, a DSS signal input module, and a DPS signal input module.

[0012] The straight optical unit and the spiral optical unit are installed on the outer wall of the coiled tubing body and are used to monitor and measure fluid noise data, strain data, temperature data, pressure data, seismic data, microseismic data and time-shifted well seismic data inside and outside the coiled tubing body in real time.

[0013] The artificial intelligence decision-making unit is connected to the composite modem for data communication and is used to intelligently process, extract features, recognize patterns, and output decisions on the collected multi-source heterogeneous data; the cloud data center is connected to the artificial intelligence decision-making unit for the storage, management, big data analysis, and knowledge graph construction of massive monitoring data.

[0014] The artificial seismic source module is set on the ground and is used to excite time-shifted three-component seismic data in wells;

[0015] The ground geophone array is set on the ground to collect ground seismic data, which, together with the well seismic data, constitutes a well-to-ground integrated three-dimensional exploration.

[0016] Furthermore, the straight optical unit includes a straight single-mode fiber, a straight multimode fiber, a straight single-mode strained fiber, and a straight continuous grating fiber or a straight microstructure fiber;

[0017] The signal output end of the straight single-mode optical fiber is connected to the DAS signal input module for data communication, and is used to collect distributed acoustic wave sensing signals.

[0018] The signal output end of the straight multimode optical fiber is connected to the DTS signal input module for data communication, and is used to collect distributed temperature sensing signals.

[0019] The signal output end of the straight single-mode strain fiber is connected to the DSS signal input module for data communication, and is used to acquire distributed stress sensing signals.

[0020] The signal output end of the straight continuous grating fiber or straight microstructure fiber is connected to the DPS signal input module for data communication and is used to collect distributed pressure sensing signals.

[0021] The straight single-mode fiber, straight multimode fiber, straight single-mode strain fiber, and straight continuous grating fiber or straight microstructure fiber are all high-temperature resistant and hydrogen-damage resistant fibers.

[0022] Furthermore, the spiral optical unit includes a spiral single-mode fiber, a spiral multimode fiber, a spiral single-mode strained fiber, and a spiral continuous grating fiber or a spiral microstructure fiber;

[0023] The signal output end of the spiral single-mode optical fiber is connected to the DAS signal input module for data communication, and is used to acquire three-component distributed acoustic wave sensing signals.

[0024] The signal output end of the spiral multimode optical fiber is connected to the DTS signal input module for data communication, and is used to acquire three-component distributed temperature sensing signals.

[0025] The signal output end of the spiral single-mode strain fiber is connected to the DSS signal input module for data communication, and is used to acquire three-component distributed stress sensing signals.

[0026] The signal output end of the spiral continuous grating fiber or spiral microstructure fiber is connected to the DPS signal input module for data communication, and is used to acquire three-component distributed pressure sensing signals.

[0027] Furthermore, the number of the spiral optical units is three groups, and the spiral winding angle of each group of the spiral optical units is different, wherein the spiral winding angles of the three groups of the spiral optical units are 30 degrees, 45 degrees and 60 degrees respectively.

[0028] When monitoring and measuring three-component stress signals, three-component microseismic events, and three-component well seismic signals in real time, the artificial intelligence decision-making unit uses an adaptive algorithm to select the fiber with the optimal real-time monitoring and measurement effect among the three helical angles for recording and calculation.

[0029] Furthermore, the artificial intelligence decision-making unit includes:

[0030] The intelligent preprocessing module is used to perform noise reduction, filtering, normalization, and temperature drift correction on the original DAS, DTS, DSS, and DPS signals.

[0031] The deep learning feature extraction module uses a convolutional neural network (CNN), long short-term memory network (LSTM) and Transformer architecture to extract multi-scale time-frequency features from the preprocessed data;

[0032] The multiphase fluid intelligent identification module, based on a deep learning classification network, identifies in real time the distribution and characteristics of oil, gas, water, oil-water mixture, gas-water mixture, oil clumps, water clumps and bubbles flowing in a continuous tubing, and estimates the gas content and water content.

[0033] The intelligent inversion module for production profiles, based on a physical constraint neural network and combined with a multi-parameter joint inversion method, calculates the production or gas profile, water intake or gas intake profile and their changes for each oil and gas production reservoir or oil and gas production well section downhole, and identifies channeling and invalid circulation.

[0034] The three-dimensional stress field intelligent modeling module, based on graph neural network (GNN), integrates three-component stress data to construct three-dimensional stress field distribution maps at various depths and directions underground, enabling intelligent early warning and prediction of pipe column damage in high-stress concentrated pipe sections.

[0035] The time-shifting fluid boundary intelligent tracking module, based on the spatiotemporal sequence prediction model, processes time-shifting seismic data collected at different times, automatically tracks the three-dimensional spatial distribution range and evolution of the oil-water two-phase and oil-gas-water three-phase fluid boundaries around the well, and predicts the distribution of remaining oil and the dynamic changes in recovery rate.

[0036] The knowledge graph reasoning module, based on the constructed knowledge graph relating downhole fluids, formations, and engineering parameters, integrates expert rules for intelligent reasoning, providing decision support for production system optimization.

[0037] Furthermore, the intelligent preprocessing module employs an adaptive wavelet threshold denoising algorithm to denoise the DAS signal, a Kalman filter algorithm to correct the temperature drift of the DTS signal, an empirical mode decomposition (EMD) algorithm to extract the trend term from the DSS signal, and a Savitzky-Golay filter algorithm to smooth the DPS signal.

[0038] Furthermore, the deep learning feature extraction module includes:

[0039] A one-dimensional convolutional neural network (1D-CNN) submodule is used to extract local temporal features of DAS signals;

[0040] A two-dimensional convolutional neural network (2D-CNN) submodule is used to extract frequency domain features from the time-frequency plot of DAS signals (such as STFT and CWT transform results);

[0041] The Long Short-Term Memory (LSTM) submodule is used to extract long-term sequence dependency features from DTS, DSS, and DPS signals.

[0042] The Transformer attention submodule is used to capture global correlation features between multi-source data;

[0043] The multimodal feature fusion layer is used to adaptively weight and fuse the features extracted from the above sub-modules to generate a unified high-dimensional feature representation.

[0044] Furthermore, the multiphase fluid intelligent identification module adopts an improved ResNet-50 architecture as the basic classification network, with the fused multimodal features as input and the fluid type probability distribution as output.

[0045] The improvements include: introducing a channel attention mechanism (SE Block) to enhance the weights of key feature channels; employing label smoothing regularization to alleviate the imbalance problem in training data; and using the Focal Loss loss function to focus on learning hard-to-classify samples.

[0046] The gas content and moisture content estimation uses a regression network based on an attention mechanism, with the classification result as a prior constraint, to output continuous gas content and moisture content values.

[0047] Furthermore, the intelligent inversion module for the produced fluid profile adopts a physical information neural network (PINN) architecture, embedding the mass conservation equation, momentum conservation equation, and energy conservation equation as physical constraints into the neural network loss function;

[0048] The inputs to the intelligent inversion module for the produced fluid profile include: DAS fluid noise characteristics, DTS temperature gradient characteristics, DPS pressure gradient characteristics, and multiphase fluid identification results;

[0049] The output of the intelligent inversion module for the product profile includes: volumetric flow rate at each depth, flow contribution rate of each phase, water / gas absorption profile, and probability distribution of channeling location.

[0050] Furthermore, the three-dimensional stress field intelligent modeling module adopts a graph neural network (GNN) architecture, using each monitoring point on the continuous tubing trajectory as a graph node, and the spatial distance and geological connectivity between nodes as edges to construct a downhole stress field graph structure;

[0051] The three-dimensional stress field intelligent modeling module includes: a graph attention network (GAT) layer for learning attention weights between nodes; a graph convolutional network (GCN) layer for aggregating features of neighboring nodes; and a temporal graph network layer for capturing the temporal evolution of the stress field.

[0052] The output of the three-dimensional stress field intelligent modeling module includes: the three-component stress values ​​of each monitoring point, the three-dimensional visualization results of the stress field, the early warning level of high stress concentration areas, and the prediction of the remaining life of the tubing.

[0053] Furthermore, the time-shifting fluid boundary intelligent tracking module adopts a spatiotemporal graph convolutional network (ST-GCN) architecture, taking seismic attribute volumes from different periods as input and fluid boundary positions as output;

[0054] The processing flow of the time-shifting fluid boundary intelligent tracking module includes:

[0055] (1) Perform surface consistency amplitude compensation processing, surface consistency deconvolution processing, and wavefield separation processing with relative amplitude preservation on time-shifted variable offset fiber vertical seismic profile data and time-shifted three-dimensional fiber vertical seismic profile data collected at different times.

[0056] (2) The processed seismic data were subjected to amplitude-preserving pre-stack depth migration to obtain the migrated seismic data;

[0057] (3) Extract fluid-sensitive attribute parameters from the migrated seismic data, including P-wave and S-wave velocities, P-wave and S-wave velocity ratio, Poisson's ratio, P-wave impedance, S-wave impedance, P-wave and S-wave impedance ratio, Young's modulus, shear modulus and pore fluid density.

[0058] (4) Obtain highlight volume data by combining S-transform;

[0059] (5) Input the above fluid-sensitive attribute parameters and highlight volume data into the spatiotemporal graph convolutional network to automatically identify and track the three-dimensional spatial distribution of the fluid boundary;

[0060] (6) Train a sequence prediction model based on historical time-shifted data to predict the future evolution trend of the fluid boundary.

[0061] Furthermore, the knowledge graph reasoning module includes:

[0062] The physical layer includes downhole fluid entities (oil, gas, water, and mixtures), formation entities (reservoirs, caprocks, and faults), engineering parameter entities (flow rate, pressure, temperature, and stress), and equipment entities (coiled tubing, packers, and perforated sections).

[0063] The relationship layer defines the semantic relationships between entities, including "located in", "belongs to", "affects", "causes", "optimizes", etc.

[0064] The rule layer embeds domain expert knowledge, including rules such as "high temperature and high pressure areas are prone to pipe damage", "sudden increase in water content indicates water channeling", and "stress concentration areas require optimization of perforation parameters".

[0065] The inference engine, based on graph embedding and logical rule-based joint reasoning, enables intelligent mapping from monitoring data to decision recommendations.

[0066] Furthermore, the cloud data center includes:

[0067] The distributed data storage layer adopts the Hadoop HDFS and Apache Kafka architecture to store massive amounts of time-series monitoring data.

[0068] The real-time computing layer uses the Apache Flink stream processing engine to realize real-time analysis and early warning of monitoring data;

[0069] The offline analysis layer uses the Spark MLlib machine learning library for large-scale model training and knowledge graph updates.

[0070] The visualization layer provides an interactive visualization interface for downhole three-dimensional stress field, fluid distribution, and production profile.

[0071] Furthermore, both the straight single-mode strain fiber and the spiral single-mode strain fiber are made by coating the outer wall of the straight single-mode fiber and the spiral single-mode fiber with a high-temperature resistant and high-strength flexible composite material, making the fiber diameter 1mm to 2mm, and then placing them in a stainless steel tube of the corresponding size, so as to sense the stress and strain outside the stainless steel tube.

[0072] Furthermore, the outer wall of the continuous tubing body is provided with grooves, and the straight light units are all placed in the middle of the weld seam of the continuous tubing body or symmetrically placed in the middle of the weld seam and the groove of the continuous tubing body before the continuous tubing body is welded into a pipe.

[0073] The straight optical unit is embedded and welded to the continuous tubing body during the continuous tubing body welding process.

[0074] After the straight light unit and the spiral light unit are placed on the outer wall of the continuous tubing body, they are filled and fixed with a high-temperature resistant flexible composite material, and then covered with a thin stainless steel sleeve.

[0075] An artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring method, the monitoring method being based on the aforementioned monitoring system, includes the following steps:

[0076] Step 1, System Initialization and Data Acquisition: Deploy the intelligent coiled tubing, which has built-in straight and spiral optical units, to the bottom of the vertical, deviated, or horizontal well; connect each optical fiber in the straight and spiral optical units to the corresponding signal input terminal of the composite modem; start the composite modem to acquire DAS, DTS, DSS, and DPS signals inside and outside the coiled tubing in real time; transmit the acquired raw data to the artificial intelligence decision unit for intelligent processing.

[0077] Step 2, Intelligent Preprocessing and Feature Extraction: The intelligent preprocessing module is used to perform adaptive denoising, filtering, normalization and temperature drift correction on the original DAS, DTS, DSS and DPS signals; the deep learning feature extraction module is used to extract multi-scale time-frequency features from the preprocessed data, including: (1) using 1D-CNN to extract local time-domain features of DAS signals; (2) using 2D-CNN to extract frequency domain features of DAS signal time-frequency maps; (3) using LSTM to extract long-term sequence dependency features of DTS, DSS and DPS signals; (4) using Transformer attention mechanism to capture global correlation features between multi-source data; (5) using multi-modal feature fusion layer to generate a unified high-dimensional feature representation.

[0078] Step 3, Multiphase Fluid Intelligent Identification: The fused high-dimensional features are input into the multiphase fluid intelligent identification module; based on the improved ResNet-50 classification network, the distribution status and characteristics of oil, gas, water, oil-water mixture, gas-water mixture, oil clumps, water clumps and bubbles flowing in the coiled tubing are identified in real time; based on the attention mechanism regression network, the gas content and water content of each depth segment are estimated; the fluid type probability distribution and phase content estimation results are output.

[0079] Step 4, Intelligent Inversion of Production Profile: Input the DAS fluid noise characteristics, DTS temperature gradient characteristics, DPS pressure gradient characteristics, and the fluid identification results obtained in Step 3 into the intelligent inversion module of production profile; based on the Physical Information Neural Network (PINN), under the physical constraints of mass conservation, momentum conservation, and energy conservation, calculate the production or gas production profile, water intake or gas intake profile of each oil and gas production reservoir or each oil and gas production well section downhole; identify the location of channeling and ineffective circulation areas; output the volumetric flow rate, the flow contribution rate of each phase, and the channeling probability distribution of each depth segment.

[0080] Step 5, Intelligent Modeling and Early Warning of 3D Stress Field: Using a composite modulator and demodulator, the DSS signal transmitted via a spiral single-mode strain fiber is modulated and demodulated to obtain three-component stress data at different depths and orientations outside the coiled tubing. Using monitoring points on the coiled tubing trajectory as graph nodes, a downhole stress field map structure is constructed. Based on a graph neural network (GNN), the three-component stress data is fused to construct a 3D stress field distribution map at various depths and orientations underground. Real-time long-term monitoring is conducted on well sections with high three-component stress to intelligently warn of the risk of local coiled tubing damage. Using a composite modulator and demodulator, the DAS signal transmitted via a spiral single-mode fiber is modulated and demodulated to obtain the distribution of three-component microseismic data in the 3D space around the well trajectory and the rupture mechanism of the seismic source point. Based on an artificial intelligence decision-making unit, the fracturing effect is intelligently evaluated, and the spacing between fracturing sections in horizontal wells and the spacing between multi-branch horizontal wells are automatically optimized.

[0081] Step Six: Intelligent Tracking and Recovery Prediction of Time-Shifted Fluid Boundaries: The artificial seismic source module located on the ground is periodically activated to collect time-shifted variable-offset fiber vertical seismic profile data and time-shifted three-dimensional fiber vertical seismic profile data via straight and spiral single-mode optical fibers. Surface-consistent amplitude compensation processing, surface-consistent deconvolution processing, and wavefield separation processing to maintain relative amplitude are applied to the time-shifted seismic data collected at different times. Amplitude-preserving pre-stack depth migration processing is used on the processed seismic data to extract fluid-sensitive attribute parameters. Highlight volume data is obtained by combining S-transform. The fluid-sensitive attribute parameters and highlight volume data are input into a spatiotemporal graph convolutional network (ST-GCN) to automatically identify and track the three-dimensional spatial distribution range of the oil-water two-phase and oil-gas-water three-phase fluid boundaries around the well. A sequence prediction model is trained based on historical time-shifted data to predict the future evolution trend of the fluid boundaries. The recovery status of underground oil and gas resources and the distribution status and characteristics of remaining oil and gas resources around the well are calculated and monitored. The development plan and production system are intelligently optimized and adjusted by combining a knowledge graph reasoning module.

[0082] Step 7, Intelligent Decision-Making and Closed-Loop Optimization: Based on the production profile data obtained in Step 4 and the fluid boundary evolution prediction results obtained in Step 6, intelligent decision-making is performed using the knowledge graph reasoning module; production system optimization suggestions are generated, including adjustments to production allocation schemes, optimization of water injection parameters, and improvement of fracturing schemes; the optimization suggestions are fed back to the surface control system to achieve intelligent closed-loop control of downhole operations; the optimization effect is continuously monitored, and the decision-making strategy is continuously iterated and optimized through reinforcement learning algorithms.

[0083] Compared with the prior art, the present invention has the following beneficial effects:

[0084] 1. This invention deeply integrates artificial intelligence technology with distributed fiber optic sensing technology, constructing a complete technology chain from data acquisition, intelligent preprocessing, feature extraction, pattern recognition to intelligent decision-making, realizing the intelligent upgrade of downhole fluid monitoring, and significantly improving data processing efficiency and interpretation accuracy.

[0085] 2. This invention employs advanced artificial intelligence technologies such as deep learning, knowledge graphs, and graph neural networks to achieve intelligent identification of multiphase fluids, high-precision inversion of the produced fluid profile, intelligent modeling of three-dimensional stress fields, and automatic tracking of fluid boundaries, thus solving the technical bottlenecks of traditional methods that rely on human experience, are inefficient, and have poor accuracy.

[0086] 3. This invention uses a Physical Information Neural Network (PINN) to embed physical conservation laws into the neural network training process, ensuring that the inversion results of the produced fluid profile meet physical constraints and improving the reliability and interpretability of the inversion results.

[0087] 4. This invention uses graph neural networks (GNN) to model the downhole stress field, making full use of the spatial correlation and geological connectivity between monitoring points to achieve high-precision three-dimensional reconstruction and intelligent early warning of the stress field.

[0088] 5. This invention uses a spatiotemporal graph convolutional network (ST-GCN) to automatically track and predict the evolution of time-shifting fluid boundaries, achieving a leap from "manual interpretation" to "intelligent prediction" and significantly improving the application value of time-shifting monitoring data.

[0089] 6. This invention constructs a knowledge graph linking downhole fluids, formations, and engineering parameters, integrates expert rules for intelligent reasoning, and realizes intelligent mapping from monitoring data to decision recommendations, providing intelligent decision support for the scientific management of oil and gas reservoirs.

[0090] 7. This invention enables centralized storage, management and analysis of massive monitoring data through a cloud data center, supports collaborative monitoring and intelligent comparative analysis of multiple wells and regions, and provides a data foundation for intelligent management at the oilfield level.

[0091] 8. This invention sets up three sets of helical optical units with different helix angles and combines them with an artificial intelligence adaptive algorithm to select the optimal helix angle for recording and calculation, so that it can achieve the best monitoring and measurement effect when facing three-component stress signals, three-component microseismic events and three-component well seismic signals.

[0092] 9. This invention realizes the function of well-to-ground time-shift joint three-dimensional exploration by setting up artificial seismic source modules and detector arrays on the ground and combining them with artificial intelligence time-shift data analysis methods. It also realizes long-term real-time intelligent monitoring and prediction of the dynamic changes of underground oil-water two-phase and oil-gas-water three-phase fluid interfaces.

[0093] 10. This invention enables long-term, real-time, dynamic, and intelligent monitoring of oil and gas production wells without interrupting production for downhole operations. It has the advantages of no production loss, no operating costs, no personnel risks, and no environmental pollution risks, providing an indispensable intelligent means for the scientific management of oil and gas reservoirs and improving recovery rates. Attached Figure Description

[0094] Figure 1 This is a schematic diagram of the overall architecture of the artificial intelligence-based fiber optic continuous tubing downhole fluid monitoring system of the present invention;

[0095] Figure 2 This is a flowchart of the artificial intelligence data processing and fluid recognition process of the present invention;

[0096] Figure 3 This is a schematic diagram of the integrated structure of the optical fiber continuous tubing optical unit of the present invention;

[0097] Figure 4 This is a schematic diagram of the well-to-ground time-lapse combined three-dimensional exploration and AI fluid boundary tracking of the present invention;

[0098] Figure 5 This is a detailed architecture diagram of the AI ​​intelligent decision-making unit of the present invention;

[0099] Figure 6 This is a schematic diagram of the downhole three-dimensional stress field modeling based on graph neural networks according to the present invention;

[0100] Figure 7 This is a GNN modeling diagram of the downhole three-dimensional stress field of the present invention;

[0101] Figure 8 This is a flowchart of the time-shifting fluid boundary intelligent tracking and recovery prediction process of the present invention;

[0102] Figure 9 This is a knowledge graph architecture diagram of the association between downhole fluids, formations, and engineering parameters in this invention;

[0103] Figure 10 This is a flowchart of the intelligent closed-loop optimization process for production systems based on reinforcement learning, as presented in this invention. Detailed Implementation

[0104] like Figures 1-3 As shown, an artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring system includes a composite modulator / demodulator 1, a coiled tubing body 2, a straight optical unit 3, a spiral optical unit 4, an artificial intelligence decision-making unit 10, a cloud data center 11, an artificial seismic source module 8, and a ground detector array 9.

[0105] The composite modulator / demodulator 1 is located near the wellhead and internally includes a DAS signal input module, a DTS signal input module, a DSS signal input module, and a DPS signal input module. The DAS signal is a distributed acoustic wave sensing signal, the DTS signal is a distributed temperature sensing signal, the DSS signal is a distributed stress sensing signal, and the DPS signal is a distributed pressure sensing signal.

[0106] like Figure 3 The straight optical unit 3 includes a straight single-mode fiber 301, a straight multimode fiber 302, a straight single-mode strain fiber 303, and a straight continuous grating fiber 304. The signal output end of the straight single-mode fiber 301 is connected to the DAS signal input module for data communication; the signal output end of the straight multimode fiber 302 is connected to the DTS signal input module for data communication; the signal output end of the straight single-mode strain fiber 303 is connected to the DSS signal input module for data communication; and the signal output end of the straight continuous grating fiber 304 is connected to the DPS signal input module for data communication. All straight single-mode fibers 301, 302, 303, and 304 are high-temperature resistant and hydrogen-loss resistant fibers. The spiral optical unit 4 includes a spiral single-mode fiber 401, a spiral multimode fiber 402, a spiral single-mode strain fiber 403, and a spiral continuous grating fiber 404. The signal output end of the spiral single-mode fiber 401 is connected to the DAS signal input module for data communication; the signal output end of the spiral multimode fiber 402 is connected to the DTS signal input module for data communication; the signal output end of the spiral single-mode strain fiber 403 is connected to the DSS signal input module for data communication; and the signal output end of the spiral continuous grating fiber 404 is connected to the DPS signal input module for data communication.

[0107] The number of spiral light units 4 is three groups, and the spiral winding angle of each group of spiral light units 4 is different. The spiral winding angles of the three groups of spiral light units 4 are 30 degrees, 45 degrees and 60 degrees respectively.

[0108] like Figure 2 The artificial intelligence decision-making unit 10 includes an intelligent preprocessing module 101, a deep learning feature extraction module 102, a multiphase fluid intelligent identification module 103, a product fluid profile intelligent inversion module 104, a three-dimensional stress field intelligent modeling module 105, a time-shifting fluid boundary intelligent tracking module 106, and a knowledge graph reasoning module 107.

[0109] like Figure 5As shown, the various modules of the artificial intelligence decision-making unit 10 form a hierarchical data processing and decision-making process: the original multi-source heterogeneous data is cleaned and corrected by the intelligent preprocessing module 101, and then input into the deep learning feature extraction module 102 to extract multi-scale time-frequency features; the fused high-dimensional features are input in parallel into the multiphase fluid intelligent identification module 103, the liquid production profile intelligent inversion module 104, the three-dimensional stress field intelligent modeling module 105, and the time-shifted fluid boundary intelligent tracking module 106 for specialized intelligent analysis; the analysis results of each module are converged into the knowledge graph reasoning module 107, and joint reasoning is performed in combination with domain expert rules to finally generate production system optimization decision suggestions and feed them back to the ground control system.

[0110] like Figure 6 and Figure 7 As shown, the three-dimensional stress field intelligent modeling module 105 uses each monitoring point on the continuous tubing trajectory as a graph node, and the spatial distance and geological connectivity between nodes as edges to construct a downhole stress field graph structure G=(V,E,A); the graph attention network (GAT) layer automatically identifies the neighboring nodes that have the greatest impact on the stress state of the current node by learning the attention weights αij between nodes; the graph convolutional network (GCN) layer aggregates the features of neighboring nodes through a message passing mechanism and updates the stress representation of each monitoring point; the temporal graph network layer introduces gated recurrent units (GRU) to capture the dynamic law of stress field evolution over time; finally, it outputs the three-component stress values ​​[σx,σy,σz] of each monitoring point, a three-dimensional stress field visualization distribution map, a high stress concentration area warning level, and a tubing string remaining life prediction based on Miner's cumulative damage theory.

[0111] The time-shifting fluid boundary intelligent tracking module 106 adopts a spatiotemporal graph convolutional network (ST-GCN) architecture to automatically identify and track the three-dimensional spatial distribution of the fluid boundary and predict its evolution trend.

[0112] like Figure 8 As shown, the processing flow of the time-shifted fluid boundary intelligent tracking module 106 includes: processing time-shifted variable offset fiber vertical seismic profile (VSP) data and time-shifted three-dimensional optical fiber data acquired at different times.

[0113] The process involves convolution and wavefield separation with relative amplitude preservation. Amplitude-preserving pre-stack depth migration is applied to the processed seismic data to obtain high-resolution migrated seismic data volumes. Fluid-sensitive attribute parameters, such as P-wave and S-wave velocity, P-wave / S-wave velocity ratio, Poisson's ratio, P-wave impedance, S-wave impedance, P-wave / S-wave impedance ratio, Young's modulus, shear modulus, and pore fluid density, are extracted from the migrated data and combined with S-transform to obtain highlight volume data. Multi-period fluid-sensitive attribute parameters and highlight volume data are input into a spatiotemporal graph convolutional network (ST-GCN). Using seismic attribute volumes from different periods as input and fluid boundary locations as output, the network automatically identifies and tracks the three-dimensional spatial distribution range of oil-water two-phase and oil-gas-water three-phase fluid boundaries around the well. A sequence prediction model is trained based on historical time-shifted data to predict the future evolution trend of fluid boundaries, thereby assessing the recovery status of underground oil and gas resources and the distribution of remaining oil.

[0114] The knowledge graph reasoning module 107 includes an entity layer, a relation layer, a rule layer, and a reasoning engine. Based on graph embedding and logical rule joint reasoning, it realizes intelligent mapping from monitoring data to decision suggestions.

[0115] like Figure 9 As shown, the knowledge graph reasoning module 107 constructs a downhole fluid-formation-engineering parameter association knowledge graph, which includes: an entity layer containing downhole fluid entities (oil, gas, water, and mixtures), formation entities (reservoirs, caprocks, and faults), engineering parameter entities (flow rate, pressure, temperature, and stress), and equipment entities (coiled tubing, packers, and perforation sections); a relationship layer defining semantic relationships between entities, including "located in," "belongs to," "influences," "leads to," and "optimizes"; a rule layer embedding domain expert knowledge, including rules such as "high temperature and high pressure areas are prone to tubing damage," "sudden increase in water cut indicates water channeling," and "stress concentration areas require optimization of perforation parameters"; the reasoning engine, based on graph embedding and logical rule joint reasoning, maps multiphase fluid identification results, production profile inversion data, three-dimensional stress field distribution, and time-shifted fluid boundary tracking results into production system optimization decision suggestions, realizing a knowledge-driven mapping from monitoring data to intelligent decision-making.

[0116] The intelligent preprocessing module 101 uses an adaptive wavelet threshold denoising algorithm to denoise the DAS signal, a Kalman filter algorithm to correct the temperature drift of the DTS signal, an empirical mode decomposition (EMD) algorithm to extract the trend term of the DSS signal, and a Savitzky-Golay filter algorithm to smooth the DPS signal.

[0117] The deep learning feature extraction module 102 includes a one-dimensional convolutional neural network (1D-CNN) submodule, a two-dimensional convolutional neural network (2D-CNN) submodule, a long short-term memory network (LSTM) submodule, a Transformer attention submodule, and a multimodal feature fusion layer.

[0118] The multiphase fluid intelligent identification module 103 adopts an improved ResNet-50 architecture as the basic classification network, introduces a channel attention mechanism (SE Block), and uses label smoothing regularization and Focal Loss loss function. Gas content and water content estimation employ a regression network based on the attention mechanism.

[0119] The intelligent inversion module 104 for producing fluid profile adopts a physical information neural network (PINN) architecture, embedding the mass conservation equation, momentum conservation equation, and energy conservation equation as physical constraints into the neural network loss function.

[0120] The three-dimensional stress field intelligent modeling module 105 adopts a graph neural network (GNN) architecture, including a graph attention network (GAT) layer, a graph convolutional network (GCN) layer, and a temporal graph network layer.

[0121] The time-shifting fluid boundary intelligent tracking module 106 adopts a spatiotemporal graph convolutional network (ST-GCN) architecture to automatically identify and track the three-dimensional spatial distribution of the fluid boundary and predict its evolution trend.

[0122] The knowledge graph reasoning module 107 includes an entity layer, a relation layer, a rule layer, and a reasoning engine. Based on graph embedding and logical rule joint reasoning, it realizes intelligent mapping from monitoring data to decision suggestions.

[0123] The cloud data center 11 includes a distributed data storage layer, a real-time computing layer, an offline analysis layer, and a visualization layer.

[0124] Both the straight single-mode strain fiber 303 and the spiral single-mode strain fiber 403 are made by coating the outer wall of the straight single-mode fiber 301 and the spiral single-mode fiber 401 with a high-temperature resistant and high-strength flexible composite material, making the fiber diameter 1mm to 2mm, and then placing them in a stainless steel tube of the corresponding size. They are used to sense the stress and strain outside the stainless steel tube.

[0125] The outer wall of the coiled tubing body 2 has grooves 5. Before the coiled tubing body 2 is welded into a pipe, all the straight optical units 3 are placed in the middle of the weld seam of the coiled tubing body 2, or symmetrically placed in the middle of the weld seam and the groove 5 of the coiled tubing body 2. The straight optical units 3 are embedded and welded to the coiled tubing body 2 during welding. After the straight optical units 3 and spiral optical units 4 are placed on the outer wall of the coiled tubing body 2, they are filled and fixed using a high-temperature resistant flexible composite material 6, and then covered with a thin stainless steel sleeve 7.

[0126] like Figure 1 and Figure 4 As shown, the artificial seismic source module 8 is installed on the ground. The composite modulator / demodulator 1 uses a straight single-mode fiber 301 and a spiral single-mode fiber 401 to acquire the time-shifted three-component well seismic data excited by the artificial seismic source module 8. The ground geophone array 9 is installed on the ground to acquire ground seismic data. Simultaneously, it acquires well seismic data through the straight single-mode fiber 301 and the spiral single-mode fiber 401, thus forming a combined well-ground three-dimensional exploration system.

[0127] The time-shifted fluid boundary intelligent tracking module 106 performs surface consistency amplitude compensation processing, surface consistency deconvolution processing, and wavefield separation processing that maintains relative amplitude on time-shifted variable offset fiber vertical seismic profile data and time-shifted three-dimensional fiber vertical seismic profile data collected at different times. It then applies amplitude-preserving pre-stack depth migration to the processed seismic data to extract fluid-sensitive attribute parameters, including P-wave and S-wave velocities, P-wave / S-wave velocity ratio, Poisson's ratio, P-wave impedance, S-wave impedance, P-wave / S-wave impedance ratio, Young's modulus, shear modulus, and pore fluid density. Combined with S-transform, it obtains high-brightness volume data. This data is then input into a spatiotemporal graph convolutional network to automatically identify and track fluid boundaries.

[0128] An artificial intelligence-based method for monitoring downhole fluids in fiber optic coiled tubing includes the following steps:

[0129] Step 1, System Initialization and Data Acquisition: Deploy the intelligent coiled tubing, which has built-in straight optical unit 3 and spiral optical unit 4, to the bottom of the vertical, deviated, or horizontal well; connect each optical fiber to the corresponding signal input terminal of the composite modulator / demodulator 1; start the composite modulator / demodulator 1 to acquire DAS, DTS, DSS, and DPS signals in real time; transmit the raw data to the artificial intelligence decision unit 10.

[0130] Step 2, Intelligent Preprocessing and Feature Extraction: The intelligent preprocessing module 101 is used to perform adaptive denoising, filtering, normalization and temperature drift correction on the original signal; the deep learning feature extraction module 102 is used to extract multi-scale time-frequency features and generate a unified high-dimensional feature representation.

[0131] like Figure 2 As shown, in step two, the intelligent preprocessing module 101 performs adaptive denoising, filtering, normalization, and temperature drift correction on the original DAS, DTS, DSS, and DPS signals: an adaptive wavelet threshold denoising algorithm is used to denoise the DAS signal, a Kalman filter algorithm is used to correct the temperature drift on the DTS signal, an Empirical Mode Decomposition (EMD) algorithm is used to extract the trend term from the DSS signal, and a Savitzky-Golay filter algorithm is used to smooth the DPS signal; the deep learning feature extraction module 102 extracts multi-scale time-frequency features from the preprocessed data: using a one-dimensional convolutional neural network... The local temporal features of the DAS signal are extracted using a 1D-CNN network, and the frequency domain features of the DAS signal time-frequency map (STFT and CWT transform results) are extracted using a 2D-CNN network. Long Short-Term Memory (LSTM) network is used to extract the long-term sequence dependency features of DTS, DSS, and DPS signals. The Transformer attention mechanism is used to capture the global correlation features between multi-source data. Finally, the features extracted by the above sub-modules are adaptively weighted and fused through a multimodal feature fusion layer to generate a unified high-dimensional feature representation, providing high-quality input data for subsequent intelligent recognition and inversion.

[0132] Step 3, Intelligent identification of multiphase fluids: The fused features are input into the intelligent identification module 103 of multiphase fluids. Based on the improved ResNet-50 network, the fluid types such as oil, gas, water and mixed liquid are identified in real time. The gas content and water content are estimated based on the attention regression network.

[0133] like Figure 2 and Figure 5 As shown, in step three, the fused high-dimensional features generated in step two are input into the multiphase fluid intelligent recognition module 103; based on the improved ResNet-50 classification network, a channel attention mechanism (SE Block) is introduced to enhance the channel weights of key features, label smoothing regularization is used to alleviate the imbalance problem of training data, and the Focal Loss loss function is used to focus on the learning of difficult-to-classify samples, thereby identifying the distribution status and characteristics of oil, gas, water, oil-water mixture, gas-water mixture, oil clumps, water clumps and bubbles flowing in continuous tubing in real time; based on the attention mechanism, a regression network is used, with the classification results as prior constraints, to output continuous gas and water content estimates for each depth segment; finally, the fluid type probability distribution and phase content estimation results are output, providing key input for the production profile inversion.

[0134] Step 4, Intelligent Inversion of Production Profile: Input the multi-source features and fluid identification results into the Intelligent Inversion Module 104 for Production Profile. Based on PINN, calculate the production / gas profile and water / gas absorption profile of each reservoir under physical constraints, and identify the location of cross-channeling.

[0135] like Figure 5 As shown, in step four, the DAS fluid noise characteristics, DTS temperature gradient characteristics, DPS pressure gradient characteristics, and the fluid identification results obtained in step three are input into the intelligent inversion module 104 for production profiles. Based on the Physical Information Neural Network (PINN) architecture, under the physical constraints of the mass conservation equation, momentum conservation equation, and energy conservation equation, the production or gas production profile, water intake or gas intake profile of each oil and gas production reservoir or each oil and gas production well section in the well are calculated. By analyzing the abnormal flow distribution at each depth, the location of channeling and ineffective circulation areas are identified. The volumetric flow rate, the flow contribution rate of each phase, and the channeling probability distribution at each depth are output, providing a basis for adjusting the production system.

[0136] Step 5, Intelligent Modeling and Early Warning of 3D Stress Field: Modulate and demodulate the DSS signal of the spiral single-mode strain fiber 403 to obtain three-component stress data; construct a 3D stress field distribution map based on GNN; provide intelligent early warning for high-stress well sections; modulate and demodulate the DAS signal of the spiral single-mode fiber 401 to obtain three-component microseismic data; AI intelligently evaluates the fracturing effect and automatically optimizes the spacing between fracturing sections and well spacing.

[0137] like Figure 6 and Figure 7 As shown, in step five, the DSS signal transmitted by the spiral single-mode strain fiber 403 is modulated and demodulated using the composite modulator-demodulator 1 to obtain the three-component stress data [σx, σy, σz] at different depths and orientations on the outer side of the coiled tubing body; using each monitoring point on the coiled tubing trajectory as a graph node and the spatial distance and geological connectivity between nodes as edges, a downhole stress field graph structure G=(V,E,A) is constructed; the graph attention network (GAT) layer based on the graph neural network (GNN) learns the attention weights between nodes, the graph convolutional network (GCN) layer aggregates the features of neighboring nodes, and the temporal graph network layer introduces GRU capture. The system captures the temporal evolution of the stress field and integrates three-component stress data to construct a three-dimensional stress field distribution map at various depths and directions underground. It performs real-time long-term monitoring of well sections where the three-component stress exceeds a set threshold, providing intelligent early warning of the risk of damage to the local continuous tubing. Simultaneously, it uses a composite modulator-demodulator 1 to modulate and demodulate the DAS signal transmitted via a spiral single-mode fiber optic cable 401, obtaining the distribution of three-component microseismic data in the three-dimensional space around the well trajectory and the rupture mechanism of the source point. Based on an artificial intelligence decision-making unit 10, it intelligently evaluates the fracturing effect and automatically optimizes the segment spacing of horizontal well fracturing sections and the well spacing of multi-branch horizontal wells.

[0138] Step 6, Time-lapse Fluid Boundary Intelligent Tracking and Recovery Prediction: Periodically excite the artificial seismic source module 8 to collect time-lapse VSP data; perform surface consistency processing and amplitude-preserving pre-stack depth migration; extract fluid-sensitive attribute parameters; input ST-GCN to automatically track fluid boundaries; predict evolution trends; assess recovery status and remaining oil distribution; and combine knowledge graphs to intelligently optimize development schemes.

[0139] like Figure 8 As shown, in step six, the artificial seismic source module 8 is periodically excited to acquire time-shifted variable offset fiber vertical seismic profile (VSP) data and time-shifted three-dimensional fiber VSP data through straight single-mode fiber 301 and spiral single-mode fiber 401; the time-shifted fluid boundary intelligent tracking module 106 is used to perform surface consistency amplitude compensation, surface consistency deconvolution, and wavefield separation processing to maintain relative amplitude on the acquired multi-stage seismic data, and high-resolution migrated seismic data volume is obtained by amplitude-preserving pre-stack depth migration; P-wave and S-wave velocities, Poisson's ratio, wave impedance, and Young's elasticity are extracted from the migrated data. Fluid-sensitive property parameters such as modulus, shear elastic modulus, and pore fluid density are combined with S-transform to obtain bright volume data. Multi-period fluid-sensitive property parameters and bright volume data are input into a spatiotemporal graph convolutional network (ST-GCN) to automatically identify and track the three-dimensional spatial distribution range of the oil-water two-phase and oil-gas-water three-phase fluid boundaries around the well. Based on historical time-shifted data, a sequence prediction model is trained to predict the future evolution trend of fluid boundaries, thereby assessing the recovery status of underground oil and gas resources and the distribution of remaining oil. The development plan and production system are intelligently optimized and adjusted in conjunction with the knowledge graph reasoning module 107.

[0140] Step 7, Intelligent Decision-Making and Closed-Loop Optimization: Based on the predicted results of the produced fluid profile and fluid boundary, the knowledge graph reasoning module 107 generates production system optimization suggestions; these suggestions are fed back to the ground control system to achieve intelligent closed-loop control; and the decision-making strategy is continuously iterated and optimized through reinforcement learning. The electronic components and artificial intelligence algorithm modules used in this invention are all general standard parts or components / algorithms known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0141] like Figure 10As shown, in the intelligent decision-making and closed-loop optimization process in step seven, based on the production / gas production profile data of each reservoir output by the intelligent production profile inversion module 104 and the fluid boundary evolution prediction results output by the time-shifting fluid boundary intelligent tracking module 106, the knowledge graph reasoning module 107 first performs production system optimization decision-making under multi-objective constraints, generating specific suggestions including production allocation scheme adjustment, water injection parameter optimization, and fracturing scheme improvement; the optimization suggestions are fed back to the surface control system, and intelligent closed-loop control of downhole operations is achieved by adjusting the wellhead valve opening, water injection pressure, and fracturing pump injection parameters; after the closed-loop control is executed, the dynamic response data of production is continuously monitored, and the decision-making effect is evaluated and the strategy network parameters are updated using reinforcement learning algorithms (such as Proximal Policy Optimization (PPO) or Deep Deterministic Policy Gradient (DDPG)). The decision-making strategy is continuously iterated and optimized to achieve a complete optimization loop from monitoring data to intelligent decision-making and then to closed-loop execution.

Claims

1. An artificial intelligence-based fiber optic coiled tubing downhole fluid monitoring system, comprising a composite modulator / demodulator (1), a coiled tubing body (2), a straight optical unit (3), and a spiral optical unit (4), characterized in that: It also includes an artificial intelligence decision-making unit (10), a cloud data center (11), an artificial seismic source module (8), and a ground detector array (9). The composite modulator (1) is located near the wellhead and contains a DAS signal input module, a DTS signal input module, a DSS signal input module and a DPS signal input module. The straight optical unit (3) and the spiral optical unit (4) are located on the outer wall of the coiled tubing body (2) and are used to monitor and measure fluid noise data, strain data, temperature data, pressure data, seismic data in the formation around the wellbore, microseismic data and seismic data in the moving well in real time. The artificial intelligence decision-making unit (10) is connected to the composite modem (1) for data communication and is used to intelligently process, extract features, recognize patterns and output decisions on the collected multi-source heterogeneous data; The cloud data center (11) is network-connected to the artificial intelligence decision-making unit (10) for the storage, management, big data analysis and knowledge graph construction of massive monitoring data; The artificial seismic source module (8) is set on the ground and is used to excite time-shifted three-component seismic data in wells; The ground geophone array (9) is set on the ground to collect ground seismic data and form a well-ground combined three-dimensional exploration with the well seismic data.

2. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 1, characterized in that: The straight optical unit (3) includes a straight single-mode fiber (301), a straight multimode fiber (302), a straight single-mode strained fiber (303), and a straight continuous grating fiber or a straight microstructure fiber (304). The signal output end of the straight single-mode fiber (301) is connected to the DAS signal input module for data communication; the signal output end of the straight multimode fiber (302) is connected to the DTS signal input module for data communication; the signal output end of the straight single-mode strain fiber (303) is connected to the DSS signal input module for data communication; and the signal output end of the straight continuous grating fiber or straight microstructure fiber (304) is connected to the DPS signal input module for data communication. The straight single-mode fiber (301), straight multimode fiber (302), straight single-mode strain fiber (303), and straight continuous grating fiber or straight microstructure fiber (304) are all high-temperature resistant and hydrogen-damage resistant fibers.

3. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 1, characterized in that: The spiral optical unit (4) includes a spiral single-mode fiber (401), a spiral multimode fiber (402), a spiral single-mode strain fiber (403), and a spiral continuous grating fiber or a spiral microstructure fiber (404). The signal output end of the spiral single-mode fiber (401) is connected to the DAS signal input module for data communication. The signal output end of the spiral multimode fiber (402) is connected to the DTS signal input module for data communication. The signal output end of the spiral single-mode strain fiber (403) is connected to the DSS signal input module for data communication. The signal output end of the spiral continuous grating fiber or the spiral microstructure fiber (404) is connected to the DPS signal input module for data communication.

4. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 3, characterized in that: The number of the spiral light units (4) is three groups, and the spiral winding angle of each group of the spiral light units (4) is different. The spiral winding angles of the three groups of the spiral light units (4) are 30 degrees, 45 degrees and 60 degrees respectively. When monitoring and measuring three-component stress signals, three-component microseismic events and three-component well seismic signals in real time, the artificial intelligence decision unit (10) selects the fiber with the best real-time monitoring and measurement effect among the three helical angles through an adaptive algorithm for recording and calculation.

5. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 1, characterized in that: The artificial intelligence decision-making unit (10) includes: The intelligent preprocessing module (101) is used to perform noise reduction, filtering, normalization and temperature drift correction on the original DAS, DTS, DSS and DPS signals; The deep learning feature extraction module (102) uses a convolutional neural network, a long short-term memory network and a Transformer architecture to extract multi-scale time-frequency features from the preprocessed data; The multiphase fluid intelligent identification module (103) is based on a deep learning classification network to identify the distribution status and characteristics of oil, gas, water, oil-water mixture, gas-water mixture, oil clumps, water clumps and bubbles flowing in the continuous tubing in real time, and to estimate the gas content and water content. The intelligent inversion module for production profile (104) is based on a physical constraint neural network and combined with a multi-parameter joint inversion method to calculate the production profile or gas production profile, water absorption or gas absorption profile and their changes for each oil and gas production reservoir or oil and gas production well section downhole, and to identify channeling and invalid circulation. The three-dimensional stress field intelligent modeling module (105) is based on graph neural network, integrates three-component stress data, constructs three-dimensional stress field distribution maps at various depths and directions underground, and realizes intelligent early warning and pipe column damage prediction for high stress concentration pipe sections. The time-shifting fluid boundary intelligent tracking module (106) processes time-shifting seismic data collected at different times based on a spatiotemporal sequence prediction model, automatically tracks the three-dimensional spatial distribution range and evolution of the oil-water two-phase and oil-gas-water three-phase fluid boundaries around the well, and predicts the distribution of remaining oil and the dynamic changes in recovery rate. The knowledge graph reasoning module (107) is based on the constructed knowledge graph of the association between downhole fluid, formation and engineering parameters. It integrates expert rules to perform intelligent reasoning and provides decision support for the optimization of production system.

6. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 5, characterized in that: The intelligent preprocessing module (101) uses an adaptive wavelet threshold denoising algorithm to denoise the DAS signal, a Kalman filter algorithm to correct the temperature drift of the DTS signal, an empirical mode decomposition algorithm to extract the trend term of the DSS signal, and a Savitzky-Golay filter algorithm to smooth the DPS signal. The deep learning feature extraction module (102) includes: A one-dimensional convolutional neural network submodule is used to extract local temporal features of DAS signals; A two-dimensional convolutional neural network submodule is used to extract frequency domain features from the time-frequency plot of DAS signals; a long short-term memory network submodule is used to extract long-term sequence dependency features from DTS, DSS, and DPS signals. The Transformer attention submodule is used to capture global correlation features between multi-source data; The multimodal feature fusion layer is used to adaptively weight and fuse the features extracted from the above sub-modules to generate a unified high-dimensional feature representation; The multiphase fluid intelligent recognition module (103) adopts an improved ResNet-50 architecture as the basic classification network, introduces a channel attention mechanism to enhance the weight of key feature channels, uses label smoothing regularization to alleviate the problem of imbalanced training data, and uses the Focal Loss loss function to focus on the learning of difficult-to-classify samples. The gas content and moisture content estimation uses an attention-based regression network, with the classification result as a prior constraint, to output continuous gas content and moisture content values. The intelligent inversion module (104) for producing fluid profile adopts a physical information neural network architecture, embedding the mass conservation equation, momentum conservation equation and energy conservation equation as physical constraints into the neural network loss function; The inputs to the intelligent inversion module (104) for the produced fluid profile include: DAS fluid noise characteristics, DTS temperature gradient characteristics, DPS pressure gradient characteristics, and multiphase fluid identification results; The output of the intelligent inversion module (104) for the product profile includes: volumetric flow rate at each depth, flow contribution rate of each phase, water / gas absorption profile, and probability distribution of channeling location. The three-dimensional stress field intelligent modeling module (105) adopts a graph neural network architecture, using each monitoring point on the continuous tubing trajectory as a graph node, and the spatial distance and geological connectivity between nodes as edges to construct a downhole stress field graph structure; the three-dimensional stress field intelligent modeling module (105) includes: a graph attention network layer, used to learn the attention weights between nodes; a graph convolutional network layer, used to aggregate the features of neighboring nodes; and a temporal graph network layer, used to capture the temporal evolution law of the stress field; The output of the three-dimensional stress field intelligent modeling module (105) includes: the three-component stress values ​​of each monitoring point, the three-dimensional visualization results of the stress field, the early warning level of the high stress concentration area, and the prediction of the remaining life of the tubing. The time-shifting fluid boundary intelligent tracking module (106) adopts a spatiotemporal graph convolutional network architecture, taking earthquake attribute volumes of different periods as input and fluid boundary positions as output; The processing flow of the time-shifted fluid boundary intelligent tracking module (106) includes: performing surface consistency amplitude compensation processing, surface consistency deconvolution processing, and wavefield separation processing with relative amplitude preservation on time-shifted variable offset fiber vertical seismic profile data and time-shifted three-dimensional fiber vertical seismic profile data collected at different times; performing amplitude-preserving pre-stack depth migration processing on the processed seismic data to obtain migrated seismic data; extracting fluid-sensitive attribute parameters from the migrated seismic data, including P-wave and S-wave velocities, P-wave and S-wave velocity ratio, Poisson's ratio, P-wave impedance, S-wave impedance, P-wave and S-wave impedance ratio, Young's modulus, shear modulus, and fluid density in pores; obtaining high-brightness volume data by combining S-transform; inputting the above fluid-sensitive attribute parameters and high-brightness volume data into a spatiotemporal graph convolutional network to automatically identify and track the three-dimensional spatial distribution of the fluid boundary; training a sequence prediction model based on historical time-shifted data to predict the future evolution trend of the fluid boundary; The knowledge graph reasoning module (107) includes: an entity layer, which includes downhole fluid entities, formation entities, engineering parameter entities and equipment entities; a relation layer, which defines the semantic relationships between entities; a rule layer, which embeds domain expert knowledge; and a reasoning engine, which realizes intelligent mapping from monitoring data to decision suggestions based on graph embedding and logical rule joint reasoning.

7. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 1, characterized in that: The cloud data center (11) includes: a distributed data storage layer, which adopts the Hadoop HDFS and Apache Kafka architecture to store massive amounts of time-series monitoring data; a real-time computing layer, which adopts the Apache Flink stream processing engine to realize real-time analysis and early warning of monitoring data; an offline analysis layer, which adopts the Spark MLlib machine learning library to perform large-scale model training and knowledge graph updates; and a visualization layer, which provides an interactive visualization interface for downhole three-dimensional stress field, fluid distribution, and production profile.

8. The fiber optic coiled tubing downhole fluid monitoring system based on artificial intelligence according to claim 1, characterized in that: The outer wall of the continuous tubing body (2) is provided with a groove (5). The straight light unit (3) is placed in the middle of the weld of the continuous tubing body (2) or symmetrically placed in the middle of the weld and the groove (5) of the continuous tubing body (2) before the continuous tubing body (2) is welded into a pipe. The straight light unit (3) is embedded and welded to the continuous oil pipe body (2) during the welding and forming process; After the straight light unit (3) and the spiral light unit (4) are placed on the outer wall of the continuous tubing body (2), they are filled and fixed with high temperature resistant flexible composite material (6), and then covered with a thin layer of stainless steel sleeve (7) after fixing.

9. A fiber optic coiled tubing downhole fluid monitoring method based on artificial intelligence, based on the monitoring system described in any one of claims 1-8, characterized in that, Includes the following steps: Step 1, System Initialization and Data Acquisition: Deploy the intelligent coiled tubing, which has a built-in straight optical unit (3) and a spiral optical unit (4), to the bottom of the vertical, deviated, or horizontal well; connect each optical fiber to the corresponding signal input terminal of the composite modulator (1); start the composite modulator (1) and acquire downhole DAS, DTS, DSS, and DPS signals in real time; transmit the raw data to the artificial intelligence decision unit (10). Step 2, Intelligent preprocessing and feature extraction: The intelligent preprocessing module (101) of the artificial intelligence decision unit (10) is used to perform adaptive denoising, filtering, normalization and temperature drift correction on the original signal; the deep learning feature extraction module (102) of the artificial intelligence decision unit (10) is used to extract multi-scale time-frequency features from the preprocessed data to generate a unified high-dimensional feature representation; Step 3, Multiphase fluid intelligent identification: Input the fused features into the multiphase fluid intelligent identification module (103) of the artificial intelligence decision unit (10), and identify the distribution status and characteristics of oil, gas, water, oil-water mixture, gas-water mixture, oil clumps, water clumps and bubbles flowing in the continuous tubing in real time based on the deep learning classification network, and estimate the gas content and water content; Step 4, Intelligent inversion of production profile: Input the multi-source features and fluid identification results into the intelligent inversion module (104) of the artificial intelligence decision unit (10), and calculate the production or gas profile, water intake or gas intake profile of each oil and gas production reservoir or each oil and gas production well section under physical constraints based on the physical information neural network, and identify the location of the channeling. Step 5, Intelligent Modeling and Early Warning of Three-Dimensional Stress Field: Obtain three-component stress data from the DSS signal of the helical single-mode strain fiber (403) of the modulation and demodulation helical optical unit (4); Based on graph neural network fusion of three-component stress data, a three-dimensional stress field distribution map of various depths and directions underground is constructed. Intelligent early warning for high-stress well sections; Three-component microseismic data were obtained by modulating and demodulating the DAS signal of a spiral single-mode fiber (401). Based on the artificial intelligence decision-making unit (10), the fracturing effect is intelligently evaluated, and the spacing between fracturing sections and well spacing is automatically optimized. Step 6, Intelligent tracking and recovery prediction of time-shifted fluid boundaries: Periodically excite the artificial seismic source module (8) to collect time-shifted seismic data; perform surface consistency processing and amplitude-preserving pre-stack depth migration; extract fluid-sensitive attribute parameters; input the spatiotemporal graph convolutional network to automatically track the three-dimensional spatial distribution of fluid boundaries; predict evolution trends; assess recovery status and remaining oil distribution; combine knowledge graph to intelligently optimize development schemes; Step 7, Intelligent Decision-Making and Closed-Loop Optimization: Based on the prediction results of the production fluid profile and fluid boundary, the knowledge graph reasoning module (107) of the artificial intelligence decision-making unit (10) is used to generate production system optimization suggestions; Feedback is sent to the ground control system to achieve intelligent closed-loop control; the decision-making strategy is continuously optimized through reinforcement learning.

Images