Artificial intelligence-based multi-physics monitoring stress field calculation method

Abstract

The application provides a multi-physical field monitoring stress field calculation method based on artificial intelligence, comprising: conducting indoor rock physical hydraulic fracturing experiments and multi-physical field monitoring; constructing a training data set; constructing a stress calculation module, the stress calculation module uses a three-channel substructure, one channel is a conventional multi-layer convolution structure for extracting multi-scale information of time, and two channels are UNet type substructures for extracting multi-scale information of time and space; training a detection network using the training data set; performing the same preprocessing step on all actual monitoring data, inputting the preprocessed acoustic emission data into the stress calculation module, and obtaining a stress field value; and updating the data set and the stress calculation module. The multi-physical field monitoring stress field calculation method based on artificial intelligence uses an adaptive AI network updating mechanism, dynamically monitors noise updating, and improves the adaptability of the stress calculation module.

Description

Technical Field

[0001] This invention relates to the field of "machine learning" technology, specifically "computation based on a specific computational model." In addition to machine learning, the multiphysics monitoring of this invention involves "obtaining an image of the object's interior by emitting ultrasonic waves or sound waves through it." Specifically, this invention relates to a method for calculating stress fields in multiphysics monitoring based on artificial intelligence. Background Technology

[0002] Hydraulic fracturing technology involves injecting high-pressure fluid into shale reservoirs to create complex artificial fractures, increasing reservoir connectivity and improving single-well production. Monitoring and evaluating different stages of hydraulic fracturing reservoir stimulation are prerequisites for efficient development and safe production. However, in hydraulic fracturing field monitoring, the complex conditions at the well site, the limited range of the observation system, and the inability to obtain accurate background data of the actual work area (velocity structure, rock mechanical parameters, etc.) make it difficult to verify the accuracy of fracture inversion results during hydraulic fracturing through field monitoring.

[0003] The laboratory's rock physics hydraulic fracturing experiments can provide stable and controllable stress loading conditions and signal measurement conditions, enabling the accurate acquisition of various property parameters of rock samples. This allows for detailed analysis and research on the changes in rocks during hydraulic fracturing, further providing guidance and assistance for actual hydraulic fracturing.

[0004] Indoor rock physics hydraulic fracturing experiments can perform multi-physics monitoring (ultrasound, acoustic emission, CT). In laboratory rock physics experiments, CT, ultrasound, and acoustic emission monitoring can be performed simultaneously. Computed Tomography (CT) is an imaging technique that acquires information about the object being tested in a non-destructive manner; CT imaging uses excited X-ray beams to perform tomographic scanning of rock samples, obtaining information about the internal structure of the rock sample. Active-source ultrasound uses piezoelectric ceramic sensors (PZTs) to excite seismic waves, which propagate within the rock sample. By inverting the received seismic waves, information about the internal medium, such as velocity, can be obtained. Passive-source acoustic emission uses piezoelectric ceramic sensors to continuously and passively acquire acoustic emission signals generated by internal rock fracturing. Analysis of these acoustic emission signals can obtain source parameters related to internal rock fracture. In these monitoring methods, piezoelectric ceramic sensors are installed on the rock surface and arranged in a specific array of detectors. These PZT sensors use a rapid automatic switching system to switch between receiving and transmitting functions, allowing for both passive acquisition of acoustic emission signals and active excitation of ultrasonic signals.

[0005] Existing stress field inversion methods mainly draw on seismic inversion methods, which are generally divided into two-step methods and one-step direct methods. When estimating the relative stress patterns in local areas, the following two basic assumptions are usually made: (1) the tectonic stress in the source area is uniform (homogeneous); (2) the fault slides along the direction of the decomposed shear stress, i.e., the Wallace-Bott assumption.

[0006] Two-step method: Determine the stress field and mode from a set of well-constrained multiple earthquake focal mechanisms. The specific implementation is as follows: 1. Determine the focal mechanism solution for each earthquake using focal mechanism inversion methods. Commonly used methods include: constraints based on initial motion polarity (first arrival P-wave), constraints based on amplitude-related information (using the absolute amplitude of P-waves or S-waves, the amplitude ratio of S-waves to P-waves), constraints based on waveform information (partial (or complete) waveform information of P-waves or S-waves in the earthquake event), and constraints based on a combination of multiple information (selecting two or more from initial motion polarity, absolute amplitude of P-waves (or S-waves), S / P amplitude ratio, and full waveform information as constraints to construct the objective function); 2. Using the focal mechanism solution determined in step 1, estimate the stress field using stress inversion methods. Common methods include the Michael method (Michael, 1984), the Angelier method (Angelier, 2002), and the Gephart and Forsyth method (Gephart & Forsyth, 1984). The disadvantages of this method are: 1. When the focal mechanism cannot be accurately determined, or when the determined focal mechanism has a large degree of ambiguity (such as a small number of observation stations or a narrow distribution of observation stations), the two-step stress field calculation method has a high degree of uncertainty; 2. Each step of the two-step inversion method requires fitting error. For example, the mechanism inversion error between different methods may be 5-10°. This cumulative inversion error leads to a large accumulation of uncertainty in the final result.

[0007] One-step method: A physical relationship between the stress field and the observed waveform is established through reasonable geomechanical assumptions. For a given set of initial stress fields, we first predict the dip angles of a set of fault strikes that satisfy the stress fields. Then, we simulate the theoretical waveforms and compare them with the observed waveforms using an objective function. Thus, a single inversion loop is constructed to invert the four stress parameters. If the stress field can generate a theoretical waveform that best fits the observed data, the inversion loop terminates, and the current stress field is estimated as the true stress field; otherwise, the stress field is regenerated and the loop continues until the exit condition is met. Rivera & Cisternas (1990), Horiuchi et al. (1995), and Abers & Gephart (2001) proposed methods for finding the optimal stress field to interpret P-wave initiation observed from multiple events; in these methods, given a sufficiently large number of events, the stress field can be found by obtaining only a small number of P-wave polarities from a single event. Iwata (2018) expanded on their work by incorporating the spatial heterogeneity of stress patterns in cases of uncertain focal mechanisms, where spatial patterns are directly inferred from P-wave initiations. Kuang & Zhang (2020) further proposed a method for direct estimation of tectonic stress fields using waveform matching, which improves the accuracy of estimation by using more informative seismic waveforms.

[0008] However, the acoustic emission signal-to-noise ratio acquired by indoor rock physics experiments is low, and the PZT detectors are sparsely deployed; the inversion process requires matching of four parameters and multiple iterations, all of which make the inversion process inefficient and result in high cumulative calculation errors. Summary of the Invention

[0009] The purpose of this invention is to at least partially overcome the shortcomings of the prior art and provide a multi-physics field monitoring stress field calculation method based on artificial intelligence.

[0010] The present invention also aims to provide a multi-physics field monitoring stress field calculation method based on artificial intelligence, which avoids the disadvantages of low efficiency and high calculation error in the stress field inversion process of the prior art.

[0011] The present invention also aims to provide a multi-physics field monitoring stress field calculation method based on artificial intelligence, which can improve the adaptability of the stress calculation module.

[0012] To achieve the above-mentioned objectives or one of them, the technical solution of the present invention is as follows:

[0013] A multi-physics field monitoring stress field calculation method based on artificial intelligence, the method comprising:

[0014] Indoor rock physical hydraulic fracturing experiments were conducted, and multi-physics field monitoring was performed.

[0015] Constructing the training dataset: The experiment was conducted on different rock samples to obtain data, which constituted the total dataset; in each experiment, stress field parameters were set to obtain two-dimensional acoustic emission waveform data, with each individual acoustic emission waveform data being a two-dimensional structure. , This represents the number of channels in the piezoelectric ceramic sensor. The number of sampling points for a single channel is set to 1000; the stress field parameters are a four-dimensional structure; the dataset consists of simulated synthetic data; the data is augmented to improve its generalizability; all data undergo the same preprocessing steps.

[0016] A stress calculation module is constructed, employing a three-channel substructure. One channel is a conventional multi-layer convolutional structure for extracting multi-scale temporal information, and two channels are UNet-type substructures for extracting multi-scale temporal and spatial information. These three substructures are connected through fully connected layers to fuse the extracted features. The final connected layer outputs the values ​​corresponding to four stress field parameters. The maximum value of each dimension is extracted as the four predicted stress field parameters. The input to the stress calculation module is multi-channel data of microseismic events, and the output is the predicted stress field parameters. The network uses a cross-entropy loss function to fit the error between the predicted and actual stress field parameter values, and backpropagation guides network parameter learning and representation learning.

[0017] Train the detection network using the training dataset;

[0018] All data obtained from actual monitoring are preprocessed using the same steps. The preprocessed acoustic emission data is then input into the stress calculation module to obtain stress field values.

[0019] Update the dataset and stress calculation module.

[0020] According to a preferred embodiment of the present invention, the "conducting indoor rock physical hydraulic fracturing experiments and performing multi-physics field monitoring" includes:

[0021] Rock samples of different sizes were collected according to the research objectives. The rock samples were covered with a prefabricated rubber sleeve. The rubber sleeve had multiple probe holes, but no piezoelectric ceramic sensors were attached to the holes.

[0022] A single CT data was collected from a rock sample without a piezoelectric ceramic sensor attached using an indoor hydraulic fracturing experimental setup. This CT data did not contain metal artifacts from the piezoelectric ceramic sensor and served as the first-stage CT data.

[0023] The piezoelectric ceramic sensor is placed in the probe hole and then bonded to the surface of the rock sample.

[0024] The indoor hydraulic fracturing test apparatus was used to collect CT data on a rock sample with a piezoelectric ceramic sensor attached. The spatial position of the piezoelectric ceramic sensor was then determined using the CT data. The CT data included metal artifacts of the piezoelectric ceramic sensor and was used as the second-stage CT data.

[0025] Different loading strategies were employed to pressurize the rock samples. The pressurization phase included an isotropic loading phase, an axial pressure increase phase, a water injection phase to increase pore pressure, and a pressure unloading phase. During the water injection phase to increase pore pressure, several CT data points were acquired at set intervals. These CT data points, containing piezoelectric ceramic sensor metal artifacts, served as the third-stage CT data. During the pressure unloading phase, one CT data point was acquired, which also contained piezoelectric ceramic sensor metal artifacts, serving as the fourth-stage CT data. Finally, the piezoelectric ceramic sensor was removed from the rock sample, and another CT data point was acquired. This new CT data point, free of piezoelectric ceramic sensor metal artifacts, served as the fifth-stage CT data.

[0026] All acquired CT data were processed.

[0027] According to a preferred embodiment of the present invention, the rock sample has the following specifications: a diameter of 50 mm and a length of 125 mm, and is cylindrical.

[0028] Rock samples include two types: ordinary rock samples and disturbed rock samples;

[0029] The common rock sample types include sandstone and shale; they are further divided into five categories based on whether they contain bedding and the direction of bedding: bedding directions of 0±20°, 45±20°, 90±20°, 135±20°, and homogeneous media without bedding; and two categories based on axial compressive strength: 50±10MPa and 90±10MPa. Five rock samples were selected from each category, for a total of 100 rock samples.

[0030] The interfering rock samples include two types of rocks: sandstone and shale. They are homogeneous media without bedding and contain sediments. Five rock samples of each type are selected, for a total of 10 rock samples.

[0031] According to a preferred embodiment of the present invention, the generation of simulated synthetic data includes:

[0032] Based on rock sample information from both ordinary and disturbed rock samples, different stress field parameters are set according to the positions of piezoelectric ceramic sensors at different locations. The range is , , , Angular resolution is , The resolution is 0.01. ;

[0033] Synthesize the acoustic emission waveform data corresponding to each piezoelectric ceramic sensor;

[0034] Based on the resolution of the stress field parameters mentioned above, the acoustic emission waveform data corresponding to each piezoelectric ceramic sensor is repeatedly synthesized to synthesize multiple acoustic emission waveform data.

[0035] By selecting different types of rock samples and repeating the above process, multiple acoustic emission waveform data can be synthesized.

[0036] According to a preferred embodiment of the present invention, in the step of “training the detection network using the training dataset”, the dataset is divided into a training set and a test set in a ratio of 8:2; the stress calculation module adopts the stochastic gradient descent optimization method, sets a dynamic learning rate, sets the initial value to 0.0001, reduces it by half every 50 times, sets the batch size to 40, and sets the number of iterations to 200; the training of the stress calculation module is performed on the GPU image processing unit.

[0037] According to a preferred embodiment of the present invention, after "inputting the preprocessed acoustic emission data into the stress calculation module to obtain the stress field value", the method includes:

[0038] Analyze the stress field values ​​of two consecutive (nth and (n-1th) stress values Differences;

[0039] At given time intervals, the preprocessed CT data is processed to obtain the current CT image. The location of the piezoelectric ceramic sensor was obtained from CT images. The positional differences of the piezoelectric ceramic sensor in two adjacent images (the nth and the (n-1th)th images) were analyzed. :

[0040]

[0041] At given time intervals, the preprocessed ultrasonic data is processed to obtain the current ultrasonic velocity field value. The difference in velocity field values ​​between two consecutive measurements (nth and n-1th) of the position of the piezoelectric ceramic sensor was analyzed. :

[0042]

[0043] Noise data for each piezoelectric ceramic sensor was collected at given time intervals during the experiment. Background noise was selected, and its energy was calculated using a noise energy calculation method. The ratio of this background noise energy to the maximum noise energy in the data augmentation was then calculated. .

[0044] According to a preferred embodiment of the present invention, the training dataset is updated when any of the following conditions are triggered:

[0045] Condition 1: Stress field values ​​of two consecutive stress events (nth and (n-1th) When the difference is greater than 20;

[0046] Condition 2: The position difference of the piezoelectric ceramic sensor in two adjacent images (the nth and the (n-1th)th images) When it is greater than 10;

[0047] Condition 3: When the position of the piezoelectric ceramic sensor differs from the velocity field values ​​in two consecutive (nth and (n-1th) times) When it is greater than 10%; or

[0048] Condition 4: When When it is greater than 0.9.

[0049] According to a preferred embodiment of the present invention, the stress calculation module is triggered to update when the following conditions are met:

[0050] Within a predetermined time period, condition 1 is met consecutively more than twice, and the stress field values ​​of two adjacent stress field values ​​(the nth and (n-1th) times) are... The difference is greater than 40;

[0051] When condition 2 is satisfied; or

[0052] When condition 3 is met, and the positional difference in condition 2 is satisfied... The value is greater than 5.

[0053] According to a preferred embodiment of the present invention, the indoor hydraulic fracturing experimental apparatus includes a pressure vessel, a loading system, an acoustic emission counting and waveform acquisition system, and a CT monitoring system.

[0054] According to a preferred embodiment of the present invention, during the pressurization of rock samples using different loading strategies, active source ultrasonic data are collected at set times. During active source ultrasonic monitoring, some piezoelectric ceramic sensors are used as transmitting probes to excite ultrasonic signals, and the remaining piezoelectric ceramic sensors are used as receiving probes to receive ultrasonic signals.

[0055] During the pressurization of rock samples using different loading strategies, the piezoelectric ceramic sensor is used as both a transmitting and receiving probe during active source ultrasonic acquisition, and at other times it serves as a receiving probe to receive acoustic emission signals generated by changes in the rock sample.

[0056] The stress field calculation method based on artificial intelligence for multiphysics monitoring in this invention utilizes an adaptive AI network update mechanism to dynamically monitor noise updates and improve the adaptability of the stress calculation module. In addition, the stress field calculation method based on artificial intelligence for multiphysics monitoring in this invention is based on a real-time ultrasonic constraint update velocity model, and simultaneously updates the stress calculation module in real time according to the relationship between stress field direction and velocity. Attached Figure Description

[0057] Figure 1 The arrangement of a rock sample and a piezoelectric ceramic sensor according to an embodiment of the present invention is shown. The left side shows the rock sample, and the right side shows the arrangement of PZT on the surface of the rock sample.

[0058] Figure 2 The modular architecture of a multi-physics field monitoring stress field calculation method based on artificial intelligence according to an embodiment of the present invention is shown, which mainly includes three parts: a stress calculation module, a CT and ultrasound-assisted calculation module, and an update module. Detailed Implementation

[0059] Exemplary embodiments of the present invention are described in detail below with reference to the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements. Furthermore, in the following detailed description, numerous specific details are set forth for ease of explanation to provide a thorough understanding of the embodiments disclosed herein. However, it will be apparent that one or more embodiments may be practiced without these specific details. In other instances, well-known structures and apparatuses are illustrated to simplify the drawings.

[0060] As mentioned earlier, in indoor rock physics hydraulic fracturing experiments, simultaneous acquisition of data from CT, ultrasound, and acoustic emission enables multi-physics field monitoring. However, in indoor rock physics experiments, the number of PZT detectors used for acoustic emission acquisition is limited and their deployment is sparse; the loading process is complex, resulting in low signal-to-noise ratios in the acquired acoustic emission data; and conventional stress field inversion processes require multiple iterations to match the four parameters. These limitations in data acquisition and inversion methods lead to inefficient and inaccurate stress analysis processes.

[0061] This application proposes a direct calculation method for stress field monitoring based on artificial intelligence using multiphysics. The method employs a dual-channel model, utilizing a waveform submodule with 1D waveform information and a spectral submodule with 2D time-spectrum information. The invention uses different types of rock samples (rock type, directional bedding, axial compressive strength) and sets stress field parameters to construct a training dataset. A real-time update module based on multiphysics is designed, using three methods—rock damage assessment based on CT images, velocity assessment based on ultrasonic data, and stress field difference assessment—to adaptively update data and module parameters.

[0062] Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or computers-controlled machines to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learn-by-doing.

[0063] The following describes the specific process of the stress field calculation method for multiphysics monitoring based on artificial intelligence, according to an embodiment of the present invention.

[0064] 1. Basic situation of the rock physics experiment involved in this invention

[0065] 1.1 Preparation of experimental samples:

[0066] Experimental samples are generally derived from rocks from specific regions. The size of the rock samples can vary depending on the actual research purpose. The standard size used for conventional indoor hydraulic fracturing experiments is a cylinder with a diameter of 50 mm and a length of 125 mm.

[0067] Two types of rock samples are set up: ordinary rock samples and disturbed rock samples.

[0068] Common rock sample types include: rock type (sandstone, shale, etc., 2 types), bedding with different orientations (bedding directions are approximately 0±20°, 45±20°, 90±20°, 135±20°, etc., and homogeneous media without obvious bedding, etc., 5 types), and axial compressive strength (50±10MPa, 90±10MPa, etc., 2 types); 5 rock samples are selected from each type, for a total of 100 rock samples.

[0069] Interfering rock sample types include: rock types (sandstone, shale, etc.), homogeneous media without obvious bedding, and rock samples containing obvious sediments, such as heavy mineral impurities and clastic grains. Five rock samples are selected from each type, for a total of 10 rock samples.

[0070] The rock sample is covered with a prefabricated rubber sleeve containing 24 PZT probe holes (the number is not limited to 24; it depends on the sleeve's design). The PZT probes are placed in the holes, and the bottom of the PZT probes is glued to the rock sample surface. The positions of the PZT probe holes on the rubber sleeve are shown below. Figure 1As shown, four columns of probes are arranged along the rock sample at 0°, 90°, 180° and 270° respectively, with each column having a different probe height, to ensure that the four columns of probes do not appear at the same height at the same time.

[0071] 1.2 Experimental System Composition:

[0072] The experimental system is an indoor hydraulic fracturing experimental device, mainly consisting of a pressure vessel, a loading system, and an acoustic emission counting and waveform acquisition system.

[0073] (1) The pressure vessel is a high-pressure resistant metal container, and the loaded rock sample is placed inside the container.

[0074] (2) The loading system generally adopts a triaxial compressive stress loading system. This system can provide axial pressure, injection pressure and confining pressure to the inside of the pressure vessel to simulate the in-situ formation conditions of the rock sample.

[0075] (3) The acoustic emission counting and waveform acquisition system consists of a PZT, a preamplifier and a high-speed acquisition device.

[0076] (4) The CT monitoring system is composed of CT units, with a field of view greater than 125mm, and the scanning mode is plain scan, usually 256-slice CT.

[0077] 1.3 Experimental Procedure:

[0078] The experimental system can employ different loading strategies depending on the actual research objective. The experimental procedure of this application is as follows:

[0079] (1) Place the rock sample inside the rubber sleeve.

[0080] (2) Paste PZT (or PZT probe) into the PZT probe hole of the outer rubber sleeve of the rock sample.

[0081] (3) Place the rock sample with PZT (or PZT probe) attached inside the pressure vessel and collect data once. Calibrate the spatial position of all probes. (Because the PZT probe aperture is larger than the probe diameter, there will be spatial deviations when installing the PZT probe in each experiment. Therefore, it is necessary to accurately extract the position using CT images after installing the PZT probe.)

[0082] (4) Isotropic loading stage (containment pressure).

[0083] (5) Add axial compression stage (axial compression).

[0084] (6) Water injection to increase pore pressure (injection pressure) until the rock fracturing stage and unloading stage. Data is collected at set times during this process.

[0085] (7) Pressure unloading stage (injection pressure, confining pressure and axial pressure unloaded to 0), collect data once (after the rock sample is fractured, there are cracks inside); take the rock sample out of the container.

[0086] In the above (4)-(6) process, active source ultrasound data are collected according to the set time. When performing active source ultrasound monitoring, some PZT probes are used as transmitting probes (excitation probes) to excite ultrasound signals, and the remaining probes are used as receiving probes to receive ultrasound signals.

[0087] During the above (4)-(6) process, due to the loading of external forces, the process will be accompanied by acoustic emission events. Except when the PZT probe is used as the transmitting probe during active source ultrasonic acquisition, it is used as the receiving probe for the rest of the time to receive the acoustic emission signals generated by the changes in the rock sample.

[0088] 1.4 Main data analysis contents:

[0089] The CT data processing workflow includes: data acquisition, CT imaging, PZT metal artifact suppression, and joint analysis with acoustic emission results.

[0090] The acoustic emission data processing flow includes: valid event picking (picking data segments with obvious phases), first arrival picking (extracting the first arrival time of the waveform from the valid event data), source location (obtaining the source location of the event based on the picked first arrival time and the spatial position of the probe), source mechanism analysis (inverting the source mechanism of the event based on the waveform of the event to obtain the rupture characteristics of the event), magnitude calculation (inverting the magnitude of the event based on the waveform of the event), and stress field analysis (based on the waveform of the event or the source mechanism of the event), etc.

[0091] The ultrasonic data processing workflow includes: ultrasonic event identification (identifying signals received by other channels based on ultrasonic excitation time), ultrasonic first arrival pickup (obtaining the first arrival time of the received signal), and velocity inversion analysis (inverting the velocity model of the rock sample based on the first arrival time, waveform, and other information of the signal).

[0092] 2. Construct the training dataset:

[0093] 2.1 Overview of the dataset.

[0094] Using the ordinary rock samples and interfering rock samples prepared in 1.1, repeat the process. These simulations constitute the total dataset.

[0095] In each simulation, the stress field parameters are set. Two-dimensional acoustic emission waveform data were obtained. A single acoustic emission waveform data has a two-dimensional structure. , This is the number of channels for the PZT probe (24 in this example). This represents the number of sampling points per channel (can be set to 1000). The stress field parameters are a four-dimensional structure, with each dimension having a length of [missing value]. , which are the Gaussian distribution data of the center of the stress field parameter value for each dimension (the value of each dimension follows a Gaussian distribution with the center value of the parameter as the mean).

[0096] 2.2 Dataset generation.

[0097] The dataset consists of simulated synthetic data.

[0098] (1) Generation of simulated synthetic data (synthetic rock sample data):

[0099] Based on the rock sample information (velocity) of ordinary and interfering rock samples in section 1.1; given the PZT probe positions at different locations. ; Set different stress field parameter ranges as , , , Angular resolution is , The resolution is 0.01. The vector.

[0100] The waveform corresponding to each PZT probe was synthesized using the methods described in section 2.3. (Acoustic emission waveform data, acoustic emission data).

[0101] Based on the above stress field parameter resolution, uniformly select the above stress field parameters, repeat the above synthesis process, and synthesize multiple acoustic emission waveform data.

[0102] Select different types of rock samples from 1.1, repeat the above process, and synthesize multiple acoustic emission waveform data.

[0103] (2) Random noise:

[0104] Based on the ordinary rock samples described in section 1.1, complete the rock physics experiments described in section 1.3, and collect noise data for each PZT probe during the experiments. In different experimental phases (phases 3-6 in 1.3), background noise was selected that contained no consistent signals (e.g., acoustic emission signals, consistent mechanical vibrations, etc.). Random noise data Two-dimensional structure .

[0105] (3) Consistent noise:

[0106] Based on the ordinary rock samples described in section 1.1, complete the rock physics experiments described in section 1.3, and collect noise data for each PZT probe during the experiments. ; at different experimental stages (stages (3)-(6) in 1.3). Consistent noise was selected, including mechanical vibration of the pressure pump (transmitted to the pressure vessel through the high-pressure pipeline), mechanical vibration of the CT scanner, etc. Consistent noise data Two-dimensional structure .

[0107] 2.3 Data Synthesis Related Algorithms

[0108] 2.3.1 Green's function generation method based on 3D velocity model:

[0109] Green's function calculation for 3D velocity model based on SPECFEM3D software.

[0110] (1) Obtain a 3D velocity model;

[0111] (2) Set the position of the detector;

[0112] (3) Set the simulation parameters of the SPECFEM3D software, using Ricker wavelet;

[0113] (4) Run the simulation using SPECFEM3D software to obtain waveforms of all detectors;

[0114] (5) Based on the known Ricker wavelet, deconvolve the waveform of the detector to extract the Green function.

[0115] 2.3.2 Physical Relationship and Calculation Method between Stress Field and Recorded Waveform:

[0116] Assuming a dual-couple source model, the theoretical waveform can be expressed as:

[0117]

[0118] in Indicates the focal term. The Green's function can be obtained from step 2.3.1. `str` represents the direction, `dip` represents the tilt angle, and `rake` represents the slip direction.

[0119] The relative stress mode includes the directions of three principal axes ( ,in , and These are the maximum, intermediate, and minimum principal stresses, respectively, and the relative magnitudes of the three principal stresses (stress ratios). The directions of the three axes are represented by three Euler angles. , and (Definition of three Euler angles:) The direction, negative tilt angle, The tilt angle. The relative magnitudes between the three principal stresses (defined as parameters) can be determined. The goal is to estimate the three Euler angles and the parameters. Assuming the fault slip direction is along the shear traction direction, then:

[0120]

[0121] This represents the fault slippage. For shear traction force, The normal vector of the fault plane. For the deviatoric stress tensor:

[0122]

[0123] in, These are the fault direction parameters (i.e., strike and dip angle). For stress tensor, It is a sliding vector. Let be the event sequence number. Therefore, the deviatoric stress tensor can be expressed as a linear equation and inverted based on the known focal mechanism (i.e., strike (str), dip (dip), and slip direction (rake). This method is a conventional approach for stress inversion based on the inversion focal plane mechanism.

[0124] Assuming three Euler angles, Given the orientation of the fault plane, and based on the second assumption, the expected slip direction of the fault can be theoretically estimated. The stress tensor is a second-order tensor. In the principal coordinate system, the stress tensor... Represented as:

[0125]

[0126] Assuming three Euler angles are given , , Representing the directions of the three principal stresses, we can first transform the stress tensor to a geographic coordinate system:

[0127]

[0128] in This represents the stress tensor in a geographic coordinate system. In the geographic coordinate system, this stress tensor can be projected onto a fault plane in any direction.

[0129]

[0130] in, This represents the stress tensor in the coordinate system of the fault plane. and For rotation, the direction of slip under a given stress field can be predicted by analyzing the stress vector on the fault plane (by projecting the stress tensor onto the fault plane).

[0131] Since the slip angle can be directly predicted using known fault planes and stress tensors, the physical relationship equation between the recorded waveform and stress field variables can be rewritten as:

[0132]

[0133] Based on the above formula, we can directly determine the stress mode by evaluating the fit between the theoretical waveform and the recorded waveform.

[0134] 2.3.3 Noise Energy Calculation Method:

[0135]

[0136] 2.4 Data augmentation.

[0137] During data generation, data augmentation is performed to improve the generalizability of the dataset. Data augmentation methods include:

[0138] (1) Add the background noise actually monitored in 2.2(2) to the synthetic data in 2.2(1). When adding, use different scale coefficients. The scale coefficient range is a random number from -2 to 2. The addition method is to multiply the noise by the scale coefficient and add it to the synthetic data. Use the noise energy calculation method in 2.3.3 to calculate the added noise energy value.

[0139] (2) Add the consistent noise actually monitored in 2.2 (3) to the synthetic data in 2.2 (1). When adding, use different scaling factors. The scaling factor range is a random number from -1 to 1. The addition method is to multiply the noise by the scaling factor and add it to the synthetic data.

[0140] 2.5 Data preprocessing.

[0141] All data undergo the same preprocessing steps.

[0142] 3. Constructing a stress field calculation method:

[0143] 3.1 Module architecture.

[0144] Multiphysics monitoring stress field calculation methods, such as Figure 2 As shown, it includes experimental setup, multiphysics data acquisition module (data preprocessing module), stress calculation module, CT data calculation module, ultrasound data calculation module, and parameter update module.

[0145] The experimental setup and multiphysics data acquisition module mainly includes the experimental section in section 1.3, which corresponds to this section.

[0146] The stress calculation module contains the main structure of the stress calculation method, mainly comprising section 3.2.

[0147] The parameter update module mainly includes fine-tuning and updating the stress calculation module based on the updated training dataset.

[0148] 3.2 Stress Calculation Module

[0149] The stress calculation module uses a three-channel substructure. One channel is a conventional multi-layer convolutional structure used to extract multi-scale temporal information (while preserving spatial invariance); the other two channels are UNet-type substructures used to extract multi-scale temporal and spatial information.

[0150] The three-channel substructures mentioned above are connected through a fully connected layer to fuse the extracted features; finally, the connection layer outputs the values ​​corresponding to the four stress parameters; finally, the maximum value of each dimension is extracted as the four stress parameters to be predicted.

[0151] The stress calculation module takes multi-channel data of microseismic events as input and outputs predicted stress field parameters (Gaussian distribution). The network uses a cross-entropy loss function (Softmax function) to fit the error between the predicted stress field parameter values ​​and the actual stress field parameters, and backpropagation guides the network parameter learning and representation learning.

[0152] 4. Train the detection network using the training dataset.

[0153] 4.1 Divide the dataset into a training set and a test set in a ratio of 8:2.

[0154] 4.2 This stress calculation module uses the stochastic gradient descent optimization method; a dynamic learning rate is set, with an initial value of 0.0001, which is halved every 50 iterations. The batch size is set to 40; the number of iterations is 200.

[0155] 4.3 The training of the stress calculation module is carried out on the GPU image processing unit.

[0156] 5. Analysis and processing of actual monitoring data:

[0157] 5.1. Actual data preprocessing.

[0158] All data obtained from actual monitoring undergo the same preprocessing steps: removing the mean and normalizing the data.

[0159] 5.2 Actual data analysis.

[0160] The preprocessed acoustic emission data is input into the stress calculation module to obtain the stress field values. .

[0161] (1) Analyze the stress field values ​​of two consecutive stresses (the nth stress and the (n-1th stress) Differences .

[0162]

[0163] (2) At given time intervals, the preprocessed CT data is processed to obtain the current CT image. To obtain the location of the PZT probe in CT images Analyze the positional differences of the probe in two adjacent images (the nth and the (n-1th)th images). .

[0164]

[0165] (3) At given time intervals, the preprocessed ultrasonic data is processed to obtain the current ultrasonic velocity field value. The difference in velocity field values ​​between two consecutive measurements (the nth and the (n-1th)th measurements) at the probe position was analyzed. .

[0166]

[0167] (4) Collect noise data for each PZT probe during the experiment at given intervals; select background noise that does not contain any consistent signals (e.g., acoustic emission signals, consistent mechanical vibrations, etc.). Calculate the energy of the background noise segment using the noise energy calculation method in 2.3.3. Calculate the ratio of the background noise energy segment to the maximum noise energy value in the data augmentation in 2.4(1). .

[0168] 5.3 Dataset Update Mechanism

[0169] Update the training dataset when any of the following conditions are triggered.

[0170] (1) Differences in stress field values When it is greater than 20.

[0171] (2) When the PZT probe position difference in 5.2(2) When it is greater than 10. Based on the stress parameters obtained from the previous result. Set range , , It quickly calculates the synthesized waveform and updates it to the dataset.

[0172] (3) When the velocity field difference in 5.2(3) When it is greater than 10%, use the current speed value. Stress parameters compared to the previous result Set range , , It quickly calculates the synthesized waveform and updates it to the dataset.

[0173] (4) When 5.2(4) When the value is greater than 0.9, the data augmentation method 2.4(1) is used to limit the range of added noise to 0.9. arrive , arrive A random number is generated. 10% of the data in the dataset is randomly selected and updated.

[0174] 6. Parameter update module and parameter update:

[0175] 6.1 Module update mechanism.

[0176] The module triggers an update when the following conditions are met.

[0177] (1) When condition (1) in 5.3 is met more than twice in a short period of time (e.g., 5s), and the difference value is greater than 40. The main consideration is that when the stress parameters of the rock sample change rapidly in a short period of time, it means that the rock sample is in the stage of imminent fracture.

[0178] (2) When condition (2) in 5.3 is met. The main consideration is that when the PZT position of the rock sample changes significantly, it means that the rock sample has been damaged.

[0179] (3) When condition (3) in 5.3 is met, and the PZT probe position difference in (2) in 5.3 is... When the value is greater than 5, the main consideration is that when the rock sample velocity changes, and the PZT probe position differs, local fracturing may occur.

[0180] 6.2 Parameter update method.

[0181] Use the new data added after the start of this experiment, that is, the data updated in 5.3 (both data remain the same). (To maintain consistent quantity), the entire stress calculation module is trained using the current stress calculation module parameters as initial values ​​to maintain update efficiency.

[0182] 6.3 Update the parameters of the stress calculation module.

[0183] The stress field calculation method based on artificial intelligence for multiphysics monitoring in this invention utilizes an adaptive AI network update mechanism to dynamically monitor noise updates and improve the adaptability of the stress calculation module. In addition, the stress field calculation method based on artificial intelligence for multiphysics monitoring in this invention is based on a real-time ultrasonic constraint update velocity model, and simultaneously updates the stress calculation module in real time according to the relationship between stress field direction and velocity.

[0184] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that variations may be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for calculating stress field in multiphysics monitoring based on artificial intelligence, characterized in that, The method includes: Indoor rock physical hydraulic fracturing experiments were conducted, and multi-physics field monitoring was performed. Constructing the training dataset: The experiments were conducted on different rock samples to obtain data, forming the overall dataset; in each experiment, stress field parameters were set to obtain two-dimensional acoustic emission waveform data, with each individual acoustic emission waveform representing a two-dimensional structure (N). sta N t ), N sta N represents the number of channels in the piezoelectric ceramic sensor. t The number of sampling points for a single channel is set to 1000; the stress field parameters are a four-dimensional structure; the dataset consists of simulated synthetic data; the data is augmented to improve its generalizability; all data undergo the same preprocessing steps. A stress calculation module is constructed, employing a three-channel substructure. One channel is a conventional multi-layer convolutional structure for extracting multi-scale temporal information, and two channels are UNet-type substructures for extracting multi-scale temporal and spatial information. These three substructures are connected through fully connected layers to fuse the extracted features. The final connected layer outputs the values ​​corresponding to four stress field parameters. The maximum value of each dimension is extracted as the four predicted stress field parameters. The input to the stress calculation module is multi-channel data of microseismic events, and the output is the predicted stress field parameters. The network uses a cross-entropy loss function to fit the error between the predicted and actual stress field parameter values, and backpropagation guides network parameter learning and representation learning. Train the detection network using the training dataset; All data obtained from actual monitoring are preprocessed using the same steps. The preprocessed acoustic emission data is then input into the stress calculation module to obtain stress field values. Update the dataset and stress calculation module. After "inputting the preprocessed acoustic emission data into the stress calculation module to obtain the stress field value", the method includes: Analyze the differences between two consecutive stress field values; At given time intervals, the preprocessed CT data is processed to obtain the current CT image Img. n The location of the piezoelectric ceramic sensor, Sta, is obtained from CT images. n Analyze the positional difference E of the piezoelectric ceramic sensor in two adjacent images. Sta : At given time intervals, the preprocessed ultrasonic data is processed to obtain the current ultrasonic velocity field value Vel. n The position of the piezoelectric ceramic sensor was analyzed based on the difference E between two adjacent velocity field values. Vel : Noise data for each piezoelectric ceramic sensor was collected at given time intervals during the experiment; background noise was selected, and its energy was calculated using a noise energy calculation method. The ratio E between this background noise energy and the maximum noise energy in the data augmentation was then calculated. noise ; The stress calculation module will be updated when the following conditions are met: Within the predetermined time, condition 1 is met more than twice consecutively, and the difference between two adjacent stress field values ​​is greater than 40. When condition 2 is satisfied; or When condition 3 is satisfied, and the positional difference E in condition 2 is satisfied. Sta The value is greater than 5.

2. The method for calculating stress field based on artificial intelligence for multiphysics field monitoring according to claim 1, characterized in that, The phrase "conducting indoor rock physical hydraulic fracturing experiments and conducting multi-physics field monitoring" includes: Rock samples of different sizes were collected according to the research objectives. The rock samples were covered with a prefabricated rubber sleeve. The rubber sleeve had multiple probe holes, but no piezoelectric ceramic sensors were attached to the holes. A single CT data was collected from a rock sample without a piezoelectric ceramic sensor attached using an indoor hydraulic fracturing experimental setup. This CT data did not contain metal artifacts from the piezoelectric ceramic sensor and served as the first-stage CT data. The piezoelectric ceramic sensor is placed in the probe hole and then bonded to the surface of the rock sample. The indoor hydraulic fracturing test apparatus was used to collect CT data on a rock sample with a piezoelectric ceramic sensor attached. The spatial position of the piezoelectric ceramic sensor was then determined using the CT data. The CT data included metal artifacts of the piezoelectric ceramic sensor and was used as the second-stage CT data. Different loading strategies were employed to pressurize the rock samples. The pressurization phase included an isotropic loading phase, an axial pressure increase phase, a water injection phase to increase pore pressure, and a pressure unloading phase. During the water injection phase to increase pore pressure, several CT data points were acquired at set intervals. These CT data points, containing piezoelectric ceramic sensor metal artifacts, served as the third-stage CT data. During the pressure unloading phase, one CT data point was acquired, which also contained piezoelectric ceramic sensor metal artifacts, serving as the fourth-stage CT data. Finally, the piezoelectric ceramic sensor was removed from the rock sample, and another CT data point was acquired. This new CT data point, free of piezoelectric ceramic sensor metal artifacts, served as the fifth-stage CT data. All acquired CT data were processed.

3. The multiphysics-based stress field monitoring and calculation method according to claim 2, characterized in that: The rock sample has the following dimensions: 50 mm in diameter and 125 mm in length; it is cylindrical. Rock samples include two types: ordinary rock samples and disturbed rock samples; The common rock sample types include sandstone and shale; they are further divided into five categories based on whether they contain bedding and the direction of bedding: bedding directions of 0±20°, 45±20°, 90±20°, 135±20°, and homogeneous media without bedding; and two categories based on axial compressive strength: 50±10MPa and 90±10MPa. Five rock samples were selected from each category, for a total of 100 rock samples. The interfering rock samples include two types of rocks: sandstone and shale. They are homogeneous media without bedding and contain sediments. Five rock samples of each type are selected, for a total of 10 rock samples.

4. The multi-physics field monitoring stress field calculation method based on artificial intelligence according to claim 3, characterized in that, The generation of simulated synthetic data includes: Based on the rock sample information of ordinary rock samples and interfering rock samples, given the positions of piezoelectric ceramic sensors at different locations, different stress field parameters (a,b,c,Φ) are set in the range of 0°≤a≤360°, 0°≤b≤90°, 0°≤c≤90°, with an angular resolution of 5° for (a,b,c) and a resolution of 0.01 for Φ, which is a vector of 0≤Φ≤1. Synthesize the acoustic emission waveform data corresponding to each piezoelectric ceramic sensor; Based on the resolution of the stress field parameters mentioned above, the acoustic emission waveform data corresponding to each piezoelectric ceramic sensor is repeatedly synthesized to synthesize multiple acoustic emission waveform data. By selecting different types of rock samples and repeating the above process, multiple acoustic emission waveform data can be synthesized.

5. The multi-physics field monitoring stress field calculation method based on artificial intelligence according to claim 1, characterized in that: In the "Training the Detection Network Using the Training Dataset" step, the dataset is divided into a training set and a test set in a ratio of 8:

2. The stress calculation module adopts the stochastic gradient descent optimization method, sets a dynamic learning rate with an initial value of 0.0001, and reduces it by half every 50 iterations. The batch size is set to 40, and the number of iterations is 200. The training of the stress calculation module is performed on the GPU image processing unit.

6. The method for calculating stress field based on artificial intelligence for multiphysics field monitoring according to claim 1, characterized in that, The training dataset is updated when any of the following conditions are met: Condition 1: The difference between two consecutive stress field values ​​is greater than 20; Condition 2: When the position difference E between two adjacent images of the piezoelectric ceramic sensor Sta When it is greater than 10; Condition 3: When the position of the piezoelectric ceramic sensor is within the difference E between two adjacent velocity field values Vel When it is greater than 10%; or Condition 4: When E noise When it is greater than 0.

9.

7. The multiphysics-based stress field monitoring and calculation method according to claim 2, characterized in that: The indoor hydraulic fracturing experimental setup includes a pressure vessel, a loading system, an acoustic emission counting and waveform acquisition system, and a CT monitoring system.

8. The multiphysics-based stress field monitoring and calculation method according to claim 2, characterized in that: During the pressurization of rock samples using different loading strategies, active source ultrasonic data were collected at set times. During active source ultrasonic monitoring, some piezoelectric ceramic sensors acted as transmitting probes to excite ultrasonic signals, while the remaining piezoelectric ceramic sensors acted as receiving probes to receive ultrasonic signals. During the pressurization of rock samples using different loading strategies, the piezoelectric ceramic sensor is used as both a transmitting and receiving probe during active source ultrasonic acquisition, and at other times it serves as a receiving probe to receive acoustic emission signals generated by changes in the rock sample.

Images