A method and system for automatically extracting fracture characteristics of a fracturing pump pressure curve

Abstract

The application discloses a kind of fracturing pump pressure curve breakage feature automatic extraction method and system, belong to the technical field of hydraulic fracturing physical simulation experiment.The method includes: synchronous acquisition pump pressure time series and acoustic emission event stream;With the time shift cross-correlation peak of acoustic emission event energy density and pump pressure drop component, automatically generate the pseudo-label of crack time, realize the automatic generation of label under weak supervision;Causal convolution denoising network of building coding fracturing fluid mass conservation equation and quasi-static crack propagation criterion, through the discontinuous mask of crack point neighborhood, while filtering noise, crack pressure sudden drop is completely retained;Multi-task data processing outputs crack time, breakage pressure, expansion section pressure gradient and pump stop feature, and uses acoustic emission positioning, CT scanning and profile sample data to check.The application can realize the objective, high-precision, real-time automatic extraction of breakage feature, greatly reduce the cost of artificial labeling, solve the technical problem that denoising and mutation fidelity cannot be compatible.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic fracturing physical simulation experiment technology, specifically involving an automatic extraction method and system for fracturing pump pressure curve rupture characteristics. It can be used for the automatic interpretation of rupture characteristic parameters such as initiation pressure, propagation pressure plateau, and instantaneous pump stop pressure in laboratory true triaxial hydraulic fracturing experiments, and can also be applied to the intelligent analysis of fracturing operation curves in the field. Background Technology

[0002] Hydraulic fracturing physical simulation experiments are an important means of studying fracture initiation mechanisms, propagation laws, and fracturing parameter design. A true triaxial hydraulic fracturing experimental system typically consists of a three-dimensional independent servo loading system, a constant-flow, constant-pressure plunger pump system, and a multi-channel monitoring system. Pressure sensors and acoustic emission (AE) sensors are simultaneously deployed on the specimen to record the injection pressure history and rock micro-fracture events. The pump-pressure curve is the core basis for experimental data analysis, containing key characteristic values ​​such as initiation pressure, propagation pressure plateau, instantaneous pump shutdown pressure, and closure pressure. These characteristic values ​​are the core inputs for subsequent inversion of parameters such as rock fracture toughness, fracturing fluid filtration coefficient, and in-situ stress state. The accuracy of the initiation pressure interpretation directly determines the reliability of fracture toughness inversion and numerical model calibration. The pressure gradient in the propagation stage reflects the coupling effect of fracture propagation resistance and filtration strength, while the pump shutdown pressure drop data is widely used for interpreting fracture parameters based on pressure reduction analysis.

[0003] The measured pump pressure curve contains various noise components. In a laboratory environment, pressure pulsations caused by the reciprocating motion of the plunger pump, transient disturbances caused by the action of the servo loading system, micro-vibrations generated by the friction between the indenter and the specimen end face, and electromagnetic interference from the data acquisition link are all superimposed on the effective signal. In particular, the sudden pressure drop at the moment of crack initiation lasts only tens to hundreds of milliseconds, which overlaps with the aforementioned noise in both frequency band and amplitude. Conventional low-pass and band-pass filters aim to smooth the signal, but while filtering out noise, they inevitably weaken or even erase the sudden crack initiation characteristics, leading to a systematic deviation in the interpretation of crack initiation pressure and crack initiation time.

[0004] Currently, laboratories both domestically and internationally generally use manual interpretation to determine rupture characteristics from pump-pressure curves. Researchers manually mark the rupture initiation time and pressure based on experience, using the peak point, slope inflection point, and the timing of the accompanying acoustic emission event on the pressure-time curve. This method has three inherent drawbacks: First, interpretation results vary from person to person; the rupture initiation time can vary by 1 to 2 seconds from multiple interpretations of the same curve, and the rupture initiation pressure can vary by 0.05 MPa to 0.2 MPa, directly affecting the objectivity of inter-experimental comparisons. Second, the interpretation process is time-consuming and labor-intensive, failing to meet the need for real-time acquisition of characteristic values ​​during experiments; researchers often only complete all data processing several hours after the experiment ends. Third, manual interpretation does not fully utilize the synchronously recorded acoustic emission information; the physical coupling relationship between the acoustic emission event timing and pressure response is only used as qualitative circumstantial evidence and not transformed into quantitative interpretation criteria.

[0005] With the development of artificial intelligence technology in the oil and gas industry, several machine learning-based automatic processing methods for fracturing data have been proposed. Existing machine learning-based construction data event recognition methods rely on large-scale manual annotation to train models, resulting in extremely high annotation costs. Furthermore, the recognition process does not incorporate physical equation constraints, making it difficult to guarantee generalization ability on high-frequency laboratory data in noisy environments. In addition, they only output single event markers and cannot output multi-dimensional fracture features in one step.

[0006] In the application of physical information neural networks, existing technologies mostly use physical constraints for forward modeling tasks such as pressure prediction and formation fracture pressure profile prediction, rather than for fidelity denoising tasks. They have not resolved the core contradiction that filtering and denoising and abrupt change preservation are mutually exclusive, and the data sources are all field construction data or well logging data, without involving pump pressure-acoustic emission synchronization data from laboratory true triaxial experiments.

[0007] In terms of fidelity filtering, existing fracturing pump shutdown data filtering methods based on convolutional neural networks protect the peak value of pump shutdown pressure through statistical loss functions. However, their protection of abrupt change points relies on manually designed statistical losses and does not encode the fracturing fluid mass conservation and fracture propagation equations. This is a "statistical level peak protection" rather than an "equation level constraint." When the noise morphology changes or multiple candidate abrupt change points appear, it is impossible to determine which abrupt change point is the physical fracturing initiation point. Furthermore, it only targets the single abrupt change peak value of the water hammer signal in the field pump shutdown and is not suitable for the multi-stage and multi-feature structure of the pump pressure curve in the laboratory, which includes pressure rise, fracturing initiation, propagation, and pump shutdown. At the same time, it does not involve the multi-task output of fracture features and the automatic generation of training labels.

[0008] In terms of intelligent processing of acoustic emission, existing technologies mostly focus on waveform screening and rupture source localization for the acoustic emission data stream itself, without establishing a quantitative correlation between the pump pressure curve and the acoustic emission data, without using the pressure response as a physical constraint on the acoustic emission event, and the training of acoustic emission signal recognition models generally relies on manually labeled synthetic datasets, resulting in high data acquisition costs and limited generalization ability.

[0009] In summary, existing technologies for extracting fracture features from fracturing pump pressure curves have three core technical problems: First, the acquisition of training labels relies on manual annotation, lacking a mechanism for automatically generating labels using the physical coupling of pump pressure and acoustic emission; second, the preservation of abrupt changes during the denoising process relies on statistical loss rather than physical equation constraints, making it impossible to determine the authenticity of abrupt change points at the physical level; and third, feature extraction targets a single event marker or a single feature, lacking a closed-loop verification method between a multi-task integrated output structure and ground truth. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of existing technologies by providing an automatic extraction method and system for fracturing pump pressure curves, which solves the technical problems of existing technologies such as reliance on manual annotation, inability to simultaneously achieve noise reduction and abrupt change fidelity, fragmented feature extraction process, and lack of quantitative verification.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows: On one hand, the present invention provides an automatic extraction method for fracturing pump pressure curve fracture characteristics, comprising the following steps: S1. Synchronous signal acquisition: During the true triaxial hydraulic fracturing experiment, the pump pressure time series of the specimen is acquired synchronously. The acoustic emission event stream; the sampling rate of the pump pressure time series is not less than 50Hz, and the acoustic emission event stream includes the occurrence time of each rupture event. With energy The time base error between the pump pressure signal and the acoustic emission signal is less than 1ms.

[0012] S2. Automatic generation of weakly supervised pseudo-labels for crack initiation time: The discrete acoustic emission event stream is converted into a continuous acoustic emission energy density function. The pressure drop component of the pump pressure time series is calculated. Time-shift cross-correlation is performed on the acoustic emission energy density function and the pressure drop component. Pseudo-labels for crack initiation time are obtained by filtering based on the peak position of the cross-correlation function and an adaptive threshold, and corresponding pseudo-labels for crack initiation pressure are obtained.

[0013] S3. Training the Physically Constrained Fidelity Denoising Network: A causal temporal convolutional network is constructed as the denoising network. The network input is the original pump pressure time series segment, and the output is the denoised pump pressure sequence and the fracturing state sequence. The denoising network is trained with a total loss function, which includes data fitting loss, physical constraint loss, and smoothing loss. A discontinuity mask is set in a preset neighborhood of the pseudo-label at the fracturing moment. The discontinuity mask is used to shield the data fitting loss, physical constraint loss, and smoothing loss within this neighborhood, allowing the network to output pressure abrupt changes at the fracturing point. The physical constraint loss includes the fracturing fluid mass conservation residual term and the quasi-static fracture propagation residual term.

[0014] S4. Online Inference and Multi-Task Feature Extraction: The raw pump pressure time series collected in real time is input into the trained denoising network, and the denoised pump pressure sequence and the fracture initiation state sequence are output point by point; based on the denoised pump pressure sequence and the fracture initiation state sequence, a multi-dimensional fracture feature set including the fracture initiation time, fracture pressure, pressure gradient of the propagation segment, and instantaneous pump stop pressure is output simultaneously.

[0015] Further, step S2 specifically includes: S21 Acoustic emission energy density estimation: A kernel function is used to smooth the acoustic emission event flow to obtain a continuous acoustic emission energy density function. The calculation formula is:

[0016] in For the first The moment of the rupture event, For the first The energy of a rupture event The kernel function is Gaussian, and the preferred kernel is a Gaussian kernel. nuclear width The time interval is set to 0.1s to 0.5s to smooth out the discreteness of event counting. S22. Pump pressure drop component extraction: Calculate the first-order difference of the pump pressure time series and obtain the pressure change rate after smoothing. Extraction pressure drop component This component exhibits a distinct single-peak characteristic at the moment of crack initiation; S23. Time-shift cross-correlation calculation: Calculate the cross-correlation function between the acoustic emission energy density function and the pressure drop component. ,in, The start time of pumping Time shift is the time when the pumping ends or the data is truncated. The range of values ​​is , Take 2s; the cross-correlation peak corresponds to the moment when the acoustic emission event and the pressure drop are most strongly coupled; S24. Pseudo-label selection: The time corresponding to the peak value of the cross-correlation function is taken as the candidate value for the fracture initiation time. An adaptive threshold is constructed based on the statistical characteristics of the cross-correlation function in the early stage of the boosting phase. When the peak value of the cross-correlation function exceeds the adaptive threshold and the candidate value for the fracture initiation time is located within the boosting phase, it is marked as a pseudo-label for the fracture initiation time. The corresponding pseudo-label of the crack initiation pressure is obtained. The adaptive threshold ,in and These represent the mean and standard deviation of the cross-correlation function during the first 20% of the infusion period. Take 3 to 5.

[0017] Furthermore, step S2 also includes a pseudo-tag quality verification step: using the ratio of the peak density of acoustic emission events within a preset window before and after the pseudo-tag at the fracturing initiation time to the average density of the entire segment as a confidence index, samples with confidence levels lower than a preset threshold are removed and transferred to manual review; at the same time, the temporal relationship between the peak pressure time and the pseudo-tag at the fracturing initiation time is verified, and only samples with peak pressure times earlier than or equal to the pseudo-tag at the fracturing initiation time are retained to ensure consistency with the physical laws of fracturing.

[0018] Furthermore, in the physical constraint loss: The fracturing fluid mass conservation residual term for:

[0019] in The pump flow rate is obtained from the pump system setpoint or actual measurement by the flow meter. This is the equivalent compressibility coefficient of the pipeline and fracturing fluid; This is a noise-reduced pump pressure sequence; The crack volume is defined as follows: the crack volume change rate is 0 before crack initiation, and is determined by the quasi-static crack propagation model after crack initiation. The quasi-static crack propagation residual term for:

[0020] in For the minimum principal stress, This represents the power-law mapping relationship between crack volume and net pressure corresponding to the plane strain crack model. The physical constraint loss The sum of the mean squares of the two residuals:

[0021] in The dimensional balance coefficient, This represents the number of sampling points.

[0022] Furthermore, the discontinuous mask The neighborhood of the pseudo-label at the moment of crack initiation The value is 0 during the inner period and 1 during other periods. Take a time of 0.2s to 0.5s; The data fitting loss for:

[0023] in, These are either pseudo-label data or labeled data that has been manually verified.

[0024] The smoothing loss for:

[0025] Total loss function ,in , These are the weighting coefficients.

[0026] Furthermore, the causal temporal convolutional network comprises multiple residual blocks, each residual block consisting of a causal dilated convolution, weight normalization, and a corrected linear unit, with the dilation factor adjusted according to... The duration of the boost phase is increased incrementally to ensure that the network's receptive field fully covers the boost phase.

[0027] Further, in step S4: the initiation time is the moment when the initiation state sequence jumps from 0 to 1; the rupture pressure is the pressure value corresponding to the denoised pump pressure sequence at the initiation time; the propagation segment pressure gradient is the slope of the linear regression of the denoised pump pressure sequence between the initiation time and the pump stop time; the pump stop characteristics include the instantaneous pump stop pressure and the pressure drop within a preset time after pump stop.

[0028] Furthermore, the method also includes step S5 ground truth closed-loop verification: comparing the output fracture feature set with acoustic emission (AE) positioning results, CT scan crack morphology, and specimen profile data to generate a verification report, and adjusting the parameters of the denoising network according to the verification results to form an optimized closed loop.

[0029] On the other hand, this invention provides an automatic extraction system for fracturing pump pressure curve fracture features based on acoustic emission weakly supervised pseudo-labels and physical information constraints, including a hardware acquisition unit and a data processing server: The hardware acquisition unit includes a true triaxial loading system, a pumping system, a pressure sensor, an acoustic emission sensor array, and a data acquisition module. The true triaxial loading system is used to apply triaxial independent stress loads to the specimen to simulate the stress state of the target reservoir. The pumping system is a constant flow and constant pressure plunger pump used to inject fracturing fluid into the specimen wellbore at a set flow rate. The pressure sensor is installed between the outlet of the pumping pipeline and the inlet of the specimen wellbore to collect pump pressure time series with a sampling rate of not less than 50Hz. The acoustic emission sensor array contains no less than 6 acoustic emission sensors arranged on the surface of the specimen to collect acoustic emission event flow during loading and fracturing. The data acquisition module is used to synchronously acquire pressure signals and acoustic emission signals, ensuring that the time base error of the two types of signals is less than 1ms.

[0030] The data processing server includes a pseudo-label generation module, a physical constraint denoising network module, and a feature output module. The pseudo-label generation module is used to execute the above-mentioned pseudo-label generation steps to automatically generate pseudo-labels for fracture initiation time and fracture initiation pressure. The physical constraint denoising network module is used to execute the above-mentioned network training and online denoising inference steps to output denoised pump pressure sequences and fracture initiation state sequences. The feature output module is used to execute the above-mentioned multi-task feature extraction steps to output multi-dimensional fracture feature sets.

[0031] Furthermore, the data processing server also includes a ground truth verification module, which is used to compare and verify the fracture feature set with acoustic emission (AE) positioning data, CT scan data, and specimen profile data, and to provide feedback for adjusting network parameters.

[0032] Furthermore, the data processing server also includes a storage module for storing the original collected data, pseudo-labels, denoising results, and fragmentation feature sets.

[0033] Compared with the prior art, the present invention has the following advantages: 1. This invention utilizes the time-shift cross-correlation between the energy density of acoustic emission events and the pump pressure drop component to automatically generate pseudo-labels for the crack initiation time under weak supervision. The cross-correlation peak corresponds to the moment when the physical coupling between the two is strongest. With the help of adaptive threshold and confidence verification, the pseudo-labels have a high consistency rate with the manually labeled ones. Compared with the existing schemes that rely on manual labeling, the training and labeling costs are significantly reduced, and only a small number of low-confidence samples need to be manually verified.

[0034] 2. This invention encodes the fracturing fluid mass conservation equation and the quasi-static fracture propagation criterion into a network in the form of loss terms. By removing the constraints through the discontinuity mask of the fracture initiation point neighborhood, the network can completely retain the pressure drop at the moment of fracture initiation while filtering out pump pulsation, servo disturbance and electromagnetic noise. Compared with the existing statistical peak preservation method, the physical equation constraint can determine the authenticity of the abrupt change point from a physical level, significantly enhances the robustness to changes in noise morphology, achieves a balance between noise removal and abrupt change preservation, and improves the accuracy of fracture initiation feature interpretation.

[0035] 3. This invention can output multi-dimensional parameters such as crack initiation time, fracture pressure, pressure gradient of propagation section, and pump shutdown characteristics in one step, directly connecting fracture toughness and filtration coefficient inversion, shortening the post-experiment processing time from hours to minutes; at the same time, through the ground truth verification module, the output is quantitatively verified using acoustic emission positioning results, CT scan and specimen profile data, so that the feature extraction results have repeatable and verifiable objectivity.

[0036] 4. This invention adopts a causal convolutional network structure, which can perform online inference point by point. The fracturing state can be judged in real time during the experiment and linked with the automatic pump stop logic to avoid the risk of overpressure in the experiment. At the same time, it can be transferred to the real-time fracture identification and early warning of the fracturing construction curve in the field. Attached Figure Description

[0037] Figure 1 This is a diagram showing the overall architecture of the system described in this invention.

[0038] Figure 2 This is an overall flowchart of the method described in this invention.

[0039] Figure 3 This is a schematic diagram illustrating the principle of pseudo-tag generation according to the present invention; wherein, Figure 3 In the graph, A represents the change of pump pressure over time t. Figure 3 In the graph, B represents the change of acoustic emission energy density with time t. Figure 3 C in the figure represents the change of time-shift cross-correlation with time t.

[0040] Figure 4 This is a schematic diagram showing the arrangement of the acoustic emission sensor array on both sides of the specimen in an embodiment of the present invention. Detailed Implementation

[0041] The present invention will be further described below.

[0042] This embodiment targets a true triaxial hydraulic fracturing physical simulation experiment scenario, achieving automatic extraction of fracturing characteristics from the fracturing pump pressure curve. The specific implementation process is as follows: like Figure 1 As shown, this system includes a hardware acquisition unit and a data processing server.

[0043] The hardware acquisition unit includes a true triaxial loading system, a pumping system, a pressure sensor, an acoustic emission sensor array, and a data acquisition module.

[0044] In this embodiment, the specimen used was a 300mm×300mm×300mm Longmaxi Formation shale with a central borehole diameter of 20mm and a depth of 200mm. The simulated wellbore was bonded with epoxy resin. A true triaxial loading system applied three independent geostresses to the specimen: a minimum principal stress of 12MPa, an intermediate principal stress of 16MPa, and a maximum principal stress of 20MPa, simulating the stress state of the target reservoir. The pumping system used an ISCO constant flow and constant pressure plunger pump with a flow rate range of 0–50mL / min and a pressure range of 0–50MPa. In the experiment, slickwater was injected at a constant flow rate of 10mL / min. The pressure sensor used was a strain gauge pressure sensor with a range of 60MPa and an accuracy of 0.01MPa, installed between the pumping pipeline outlet and the specimen wellbore inlet. The sampling rate was set to 100Hz to record the pump pressure time series. The acoustic emission sensor array consists of eight acoustic emission sensors with a resonant frequency of 150kHz, arranged with four sensors on each side of the specimen. Figure 4 As shown, the acquisition card has a sampling rate of 1MHz and uses threshold triggering mode to record acoustic emission (AE) event streams. Each event includes the time of occurrence. With energy The data acquisition module uses a PXI chassis to achieve synchronous acquisition of pressure and acoustic emission signals, with a time base error of less than 1ms, ensuring the time alignment accuracy of the two types of signals.

[0045] The data processing server includes a pseudo-label generation module, a physical constraint denoising network, a feature output module, a storage module, and a ground truth verification module.

[0046] The pseudo-tag generation module performs the following steps, the principle of which is as follows: Figure 3 As shown: Acoustic emission energy density estimation: A Gaussian kernel is used to smooth the discrete acoustic emission event stream to obtain a continuous acoustic emission energy density function.

[0047] The Gaussian kernel function In this embodiment, the kernel width We set the time to 0.2s to eliminate the influence of event discreteness.

[0048] Pump pressure drop component extraction: analysis of pump pressure time series The pressure change rate is obtained by calculating the first difference and smoothing it with a 0.5s moving average. Extraction pressure drop component Only the signal component of pressure drop is retained.

[0049] Time-shift cross-correlation calculation: Calculation of acoustic emission energy density function With pressure drop component Cross-correlation function between them:

[0050] in The start time of pumping For the data truncation time, time shift The value range is [-2s, 2s]. Physically, there is a temporal coupling between the acoustic emission event and the pressure drop, and the cross-correlation peak corresponds to the moment when the coupling between the two is strongest.

[0051] Adaptive threshold selection: The time corresponding to the peak of the cross-correlation function is taken as the candidate value for crack initiation time. Construct an adaptive threshold ,in and These represent the mean and standard deviation of the cross-correlation function during the first 20% of the time period after the start of infusion, as described in this embodiment. Take 4. When and When the pressure rises, this moment is marked as a pseudo-label for the crack initiation moment, corresponding to the pressure value. This is a false label indicating the initiation pressure.

[0052] False tag quality verification: The ratio of the peak density of acoustic emission events within ±1s before and after the false tag at the crack initiation time to the average density of the entire segment is used as the confidence index. In this embodiment, the confidence threshold is set to 5. Samples below the threshold are automatically removed and transferred to the manual review list. At the same time, the temporal relationship between the peak pressure time and the false tag at the crack initiation time is verified. Only samples with peak pressure earlier than or equal to the crack initiation time are retained, which conforms to the physical law of "pressure rises to the peak and then suddenly drops after crack initiation".

[0053] In this embodiment, a total of 20 fracturing experiments were completed. After the above steps, a total of 19 valid pseudo-labels were obtained, and 1 low-confidence sample was transferred to manual verification. When the pseudo-labels were compared with the manual interpretation results of senior experimental personnel, the deviation of the fracturing time was less than 0.5s, and the consistency rate reached 100%.

[0054] The physical constraint denoising network adopts a causal temporal convolutional network (TCN) structure, with the following specific configuration: The network contains 6 residual blocks, each consisting of a causal dilated convolution, a weight normalization, and a corrected linear unit (ReLU). The dilation factors are 1, 2, 4, 8, 16, and 32, respectively. The convolution kernel size is 7, and the number of channels per layer is 64. The network's receptive field covers 6000 sampling points, corresponding to a duration of 60 seconds at a sampling rate of 100Hz, which can completely cover the entire boost phase.

[0055] The network input is a sliding window segment of the original pump pressure time series, and the output is a denoised pump pressure sequence. With crack initiation state sequence The crack initiation state sequence is 0 before crack initiation and 1 after crack initiation.

[0056] The total loss function for network training is:

[0057] In this embodiment Take 0.5, Take 0.1.

[0058] Data fitting loss :

[0059] in For a discontinuous mask, the neighborhood of the pseudo-label at the moment of splitting. The value is 0 for the inner period and 1 for the rest of the time period; in this embodiment The time is set to 0.3s. The purpose of the mask is to: mark the initiation point as the first-order discontinuity of the pressure function, and not apply data fitting penalty within this region, allowing the network to output pressure abruptly.

[0060] Physical constraint loss It includes two terms: mass conservation residual and quasi-static extended residual.

[0061] Residual term due to mass conservation:

[0062] in The pump flow rate is a constant value of 10 mL / min in this embodiment; The equivalent compressibility coefficient of the pipeline and fracturing fluid is calibrated to 0.4 mL / MPa based on the pressure-flow relationship in the booster section. The crack volume is given. Before crack initiation, the crack volume change rate is 0. After crack initiation, it is determined by the power-law relationship of the plane strain crack model. ,in Let the plane strain elastic modulus be taken as... MPa; Net pressure, The minimum principal stress is 12 MPa; These are the power-law coefficients, calibrated online from the extended segment data.

[0063] Quasi-static extended residual term:

[0064] The physical constraint loss is the sum of the mean squares of the two residuals:

[0065] In this embodiment Set to 1 to balance the dimensions of the two residuals.

[0066] The physical constraint loss is also subject to the discontinuity mask constraint: the mask is 0 in the neighborhood of the crack initiation point, and no pressure drop is penalized; outside the crack initiation point, the physical term forces the network output to satisfy the mass conservation and crack propagation law, achieving high-fidelity noise reduction.

[0067] Smoothing loss :

[0068] The smoothing loss is masked in the neighborhood of the crack initiation point, ensuring noise filtering without weakening the crack initiation abrupt change.

[0069] During model training, the experimental data from 20 rounds were divided into training, validation, and test sets in a 7:2:1 ratio; the Adam optimizer was used, with an initial learning rate of... The batch size is 64, the training rounds are 200, and the early stopping criterion is that the validation set loss does not decrease for 20 consecutive rounds.

[0070] After training is complete, enter online inference mode. The overall process is as follows: Figure 2 As shown.

[0071] During the experiment, the data acquisition module continuously pushes real-time pump pressure data, and the physical constraint denoising network outputs denoised pressure values ​​and crack initiation states point by point in a causal manner, with single-point processing time less than 1ms. The feature output module monitors the crack initiation state sequence in real time. When the crack initiation state changes from 0 to 1, it immediately determines that crack initiation has occurred, outputs the crack initiation time and rupture pressure, and can trigger an alert signal, which is linked to the automatic pump stop logic of the pumping system.

[0072] After the experiment, the feature output module outputs a multi-dimensional fracture feature set based on the complete denoised pump pressure sequence and fracture initiation state sequence: fracture initiation time. The moment when the fracture initiation state sequence jumps from 0 to 1; fracture pressure. : Denoising pump pressure sequence in Pressure value at any given time; pressure gradient in the extended segment From the moment of crack initiation to the moment of pump shutdown Linear regression was performed on the denoised pressure to obtain the slope value; Pump stop characteristics: instantaneous pump stop pressure The data includes the pressure drop within 30 seconds after pump shutdown. This characteristic data is stored in the storage module and can be directly input into subsequent fracture toughness inversion and filtration coefficient inversion modules.

[0073] The ground truth verification module compares and verifies the output fracture characteristics with the ground truth values: Acoustic emission (AE) localization verification: The event cloud of acoustic emission three-dimensional localization is aggregated in a planar manner along the rupture surface. The rupture initiation time corresponds to the start time of the event cloud outbreak and is compared with the rupture initiation time of the feature output.

[0074] CT scan and specimen section verification: The internal crack morphology of the specimen is obtained by CT scan, or the specimen is dissected after the experiment to observe the crack morphology and verify the correspondence between the crack propagation stage and the pressure propagation stage.

[0075] After verification is completed, a verification report is generated. If the deviation exceeds the preset range, feedback is provided to adjust the network parameters and loss weights, thus forming an optimization and adjustment process.

[0076] In this embodiment, the verification results of four rounds of experiments on the test set show that: the denoising curve is smooth in the pressure rise section and the sudden drop feature of the crack initiation point is complete; the deviation between the crack initiation pressure interpretation value and the manual interpretation value is less than 0.05 MPa, and the deviation of the crack initiation time is less than 0.3 s; the deviation between the pressure gradient in the propagation section and the manual linear regression result is less than 3%; the deviation of the pump stop pressure interpretation is less than 0.02 MPa; and all ground truth verifications are passed, verifying the accuracy and reliability of this method.

Claims

1. A method for automatically extracting fracturing characteristics from fracturing pump pressure curves, characterized in that, Includes the following steps: S1. During the true triaxial hydraulic fracturing experiment, the pump pressure time sequence and acoustic emission event flow of the specimen were collected simultaneously. S2. Convert the discrete acoustic emission event stream into a continuous acoustic emission energy density function, calculate the pressure drop component of the pump pressure time series, perform time-shift cross-correlation operation on the acoustic emission energy density function and the pressure drop component, obtain the pseudo-label of the crack initiation time based on the peak position of the cross-correlation function and the adaptive threshold, and obtain the corresponding pseudo-label of the crack initiation pressure. S3. Construct a causal temporal convolutional network as a denoising network. The network input is the original pumping time series segment, and the output is the denoised pumping sequence and the crack initiation state sequence. Train the denoising network with a total loss function, which includes data fitting loss, physical constraint loss and smoothing loss. Set a discontinuous mask within the preset neighborhood of the pseudo-label at the moment of crack initiation; S4. Input the real-time collected raw pump pressure time series into the trained denoising network. Based on the point-by-point output denoised pump pressure sequence and fracture initiation state sequence, finally obtain a multi-dimensional fracture feature set including fracture initiation time, fracture pressure, propagation section pressure gradient, and instantaneous pump stop pressure.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: S21. Acoustic emission energy density estimation: A kernel function is used to smooth the acoustic emission event flow to obtain a continuous acoustic emission energy density function. The calculation formula is: in For the first The moment of the rupture event, For the first The energy of a rupture event The Gaussian kernel function; S22. Pump pressure drop component extraction: Calculate the first difference of the pump pressure time series and smooth it to obtain the pressure change rate, then extract the pressure drop component. S23. Time-shift cross-correlation calculation: Calculate the cross-correlation function between the acoustic emission energy density function and the pressure drop component. ; S24. Pseudo-label selection: The time corresponding to the peak value of the cross-correlation function is taken as the candidate value for the fracture initiation time. An adaptive threshold is constructed based on the statistical characteristics of the cross-correlation function in the early stage of the boosting phase. When the peak value of the cross-correlation function exceeds the adaptive threshold and the candidate value for the fracture initiation time is located within the boosting phase, it is marked as a pseudo-label for the fracture initiation time. The corresponding pseudo-label of the crack initiation pressure is obtained. Adaptive threshold ,in and These represent the mean and standard deviation of the cross-correlation function during the first 20% of the time period after the start of the infusion.

3. The method according to claim 2, characterized in that, Step S2 further includes a pseudo-label quality verification step: using the ratio of the peak density of acoustic emission events within a preset window before and after the pseudo-label at the crack initiation time to the average density of the entire segment as a confidence index, samples with confidence levels lower than a preset threshold are removed and transferred to manual review; at the same time, the temporal relationship between the peak pressure time and the pseudo-label at the crack initiation time is verified, and only samples with peak pressure times earlier than or equal to the pseudo-label at the crack initiation time are retained.

4. The method according to claim 1, characterized in that, The physical constraint loss includes the fracturing fluid mass conservation residual term and the quasi-static fracture propagation residual term: The fracturing fluid mass conservation residual term for: in For pump flow rate, The equivalent compressibility coefficient of the pipeline and fracturing fluid. For the noise-reducing pump pressure sequence, The crack volume is defined as follows: the crack volume change rate is 0 before crack initiation, and is determined by the quasi-static crack propagation model after crack initiation. The quasi-static crack propagation residual term for: in For the minimum principal stress, This represents the power-law mapping relationship between crack volume and net pressure corresponding to the plane strain crack model. The physical constraint loss The sum of the mean squares of the two residuals: in The dimensional balance coefficient, This represents the number of sampling points.

5. The method according to claim 1, characterized in that, The neighborhood of the pseudo-label at the moment of the discontinuity mask's initiation. The value is 0 during the inner period and 1 during other periods. The data fitting loss for: in, This is either pseudo-label data or labeled data that has been manually verified. It is a discontinuous mask; The smoothing loss for: Total loss function ,in , These are the weighting coefficients.

6. The method according to claim 1, characterized in that, The causal temporal convolutional network contains multiple residual blocks, each of which consists of a causal dilated convolution, a weight normalization, and a modified linear unit.

7. The method according to claim 5, characterized in that, In step S4: the initiation time is the moment when the initiation state sequence jumps from 0 to 1; the rupture pressure is the pressure value corresponding to the denoised pump pressure sequence at the initiation time; the propagation segment pressure gradient is the slope of the linear regression of the denoised pump pressure sequence between the initiation time and the pump stop time; the pump stop characteristics include the instantaneous pump stop pressure and the pressure drop within a preset time after pump stop.

8. The method according to claim 1, characterized in that, It also includes step S5 ground truth closed-loop verification: comparing the output fracture feature set with the acoustic emission localization results, CT scan crack morphology, and specimen profile data, generating a verification report, and adjusting the parameters of the denoising network according to the verification results.

9. An automatic extraction system for fracturing pump pressure curve fracture characteristics, characterized in that, The method for performing any one of claims 1 to 8 includes a hardware acquisition unit and a data processing server: The hardware acquisition unit includes a true triaxial loading system, a pumping system, a pressure sensor, an acoustic emission sensor array, and a data acquisition module. The true triaxial loading system is used to apply triaxial independent stress loads to the specimen. The pumping system is used to inject fracturing fluid into the specimen wellbore at a set flow rate. The pressure sensor is used to acquire pump pressure time series. The acoustic emission sensor array contains multiple acoustic emission sensors arranged on the specimen surface to acquire acoustic emission event flows. The data acquisition module is used to simultaneously acquire pressure signals and acoustic emission signals. The data processing server includes a pseudo-label generation module, a physical constraint denoising network module, and a feature output module. The pseudo-label generation module is used to automatically generate pseudo-labels for the fracture initiation time and fracture initiation pressure. The physical constraint denoising network module is used to output a denoised pump pressure sequence and a fracture initiation state sequence. The feature output module is used to output a multi-dimensional fracture feature set.

10. The system according to claim 9, characterized in that, The data processing server also includes a ground truth verification module, which is used to compare and verify the fracture feature set with acoustic emission positioning data, CT scan data, and specimen profile data, and to provide feedback for adjusting network parameters.

Images