A method and system for quantifying the strength of an oil and gas pipeline
By optimizing the physical decoder and encoder based on physical information neural networks, the problem of insufficient data in stress detection of oil and gas pipelines was solved, achieving accurate quantification of stress and improving the accuracy and reliability of detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SINOMACH SENSING TECH CO LTD
- Filing Date
- 2026-02-27
- Publication Date
- 2026-05-12
AI Technical Summary
In stress detection of oil and gas pipelines, existing technologies struggle to establish a quantitative relationship between ACSM signals and pipeline stress when data volume is insufficient, leading to inaccurate stress assessments.
A physical decoder based on physical information neural networks is used. By integrating the physical laws of electromagnetic fields to train the model and combining it with encoder optimization, accurate quantification of pipeline stress from ACSM measurement signals can be achieved.
Under small sample conditions, the stress of oil and gas pipelines was accurately quantified, reducing the reliance on large amounts of stress data and improving the accuracy and reliability of stress detection.
Smart Images
Figure CN121740295B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of stress detection equipment technology, and in particular to a method and system for quantifying stress in oil and gas transmission pipelines. Background Technology
[0002] The safe operation and maintenance of oil and gas pipelines requires real-time monitoring of stress status, which can be performed using online non-destructive testing technology. Electromagnetic sensors, such as Alternating Current Stress Measurement (ACSM) equipment, detect pipeline stress in a non-contact manner. ACSM sensors are based on the principle of electromagnetic detection, using an excitation source to magnetize the pipe wall, and a detection coil to measure changes in induced electromotive force. The inverse magnetostrictive effect causes stress to affect the magnetic permeability.
[0003] In some approaches, ACSM can be used to measure the change in pipe wall permeability, revealing stress concentrations or defect areas through signal distortion. However, during feature extraction and mapping of the ACSM signal, it is difficult to comprehensively analyze the complex physical field, and the extracted features may not cover all effective signal characteristics. Other approaches may directly employ machine learning methods such as deep learning; however, these methods face the problem of insufficient sample size, making it difficult to construct large-scale data-driven models.
[0004] In summary, for the above scheme, it is difficult to establish a quantitative relationship between ACSM signals and pipeline stress when the amount of data is insufficient. Summary of the Invention
[0005] This application provides a method and system for quantifying stress in oil and gas transmission pipelines to address the problem that it is difficult to establish a quantitative relationship between ACSM signals and pipeline stress when there is insufficient data.
[0006] In a first aspect, this application provides a method for stress quantification in oil and gas transmission pipelines, comprising:
[0007] Acquire measurement signals from ACSM sensors for oil and gas pipelines;
[0008] The measured signal is input into the encoder to obtain an estimated value of the pipe wall permeability;
[0009] The estimated value is input into a pre-trained physical decoder to obtain the predicted ACSM signal; wherein the physical decoder is a model based on a physical information neural network, and the training process of the physical decoder incorporates the physical laws of electromagnetic fields.
[0010] The predicted ACSM signal is compared with the measured signal to optimize the encoder, resulting in an optimized encoder.
[0011] The measurement signal is input into the optimized encoder to obtain the magnetic permeability of the pipe wall of the oil and gas transportation pipeline;
[0012] The stress value of the oil and gas transmission pipeline is determined based on the relationship between the magnetic permeability of the pipe wall and the magnetic permeability stress of the pipeline material.
[0013] By co-optimizing the encoder and the pre-trained physical decoder, accurate quantification of pipeline stress from ACSM measurement signals is achieved, solving the problem of difficulty in quantitatively assessing stress with small samples.
[0014] In some feasible embodiments, before inputting the estimated value into a pre-trained physical decoder to obtain the predicted ACSM signal, the following steps are included:
[0015] Acquire training data, which includes multiple sets of physical field parameters and the corresponding real ACSM signals;
[0016] Neural network for acquiring initial physical information;
[0017] The physical field parameters are input into the initial physical information neural network to output the magnetic vector potential and electric scalar potential of the spatial points;
[0018] Based on the magnetic vector potential, a predicted ACSM signal is output;
[0019] Based on the magnetic vector potential, the electric scalar potential, and the predicted ACSM signal, a training loss function is constructed.
[0020] The initial physical information neural network is trained by minimizing the training loss function to obtain the pre-trained physical decoder.
[0021] By integrating the physical laws of electromagnetic fields to train a physical decoder, it can accurately predict ACSM signals based on physical field parameters.
[0022] In some feasible embodiments, the training loss function includes a control equation loss term and an energy minimization loss term;
[0023] Construct the training loss function, including:
[0024] Calculate the residuals of the magnetic vector potential and the electric scalar potential under the electromagnetic field control equations;
[0025] Calculate the magnetic field energy density based on the magnetic vector potential and the electric scalar potential;
[0026] Within the computational domain, sampling and statistics are performed on the residuals to obtain residual results; and within the computational domain, sampling and statistics are performed on the magnetic field energy density to obtain density results;
[0027] Based on the residual results, the loss term of the governing equation is constructed, and based on the density results, the energy minimization loss term is constructed.
[0028] By controlling the loss terms of the control equations and minimizing the energy loss terms, the network output can be constrained to ensure that the predicted electromagnetic field quantities conform to physical laws, thereby improving the prediction accuracy of the physical decoder.
[0029] In some feasible embodiments, the training loss function further includes a boundary condition loss term;
[0030] Construct the training loss function, including:
[0031] At the boundary of the computational domain, the boundary values of the magnetic vector potential and the electric scalar potential are obtained;
[0032] The difference between the boundary value and the preset boundary condition value is calculated, and the boundary condition loss term is constructed based on the difference.
[0033] By constraining the field values on the domain boundary using boundary condition loss terms, the convergence and physical rationality of the electromagnetic field solution are ensured, thereby enhancing the numerical stability of the physical decoder.
[0034] In some feasible embodiments, the training loss function further includes an excitation source constraint loss term; the physical field parameters include source current density;
[0035] Construct the training loss function, including:
[0036] Based on the magnetic vector potential and the electric scalar potential, the predicted source current density is calculated;
[0037] In the excitation source region, the predicted source current density is subjected to surface integration to obtain the total excitation current;
[0038] The difference between the total excitation current and the actual excitation current is calculated, and the excitation source constraint loss term is constructed based on the difference.
[0039] By constraining the loss term through the excitation source, we ensure that the predicted total excitation current equals the true value, guaranteeing the accuracy of the source term setting and improving the prediction reliability of the physical decoder in the excitation region.
[0040] In some feasible embodiments, the training loss function further includes an interface continuity loss term;
[0041] Construct the training loss function, including:
[0042] At the interface of different material regions, the predicted magnetic vector potentials on both sides of the interface are obtained respectively.
[0043] The difference in the normal component of the predicted magnetic vector potential on both sides of the interface is calculated, and the interface continuity loss term is constructed based on the difference.
[0044] By controlling the continuity of the electromagnetic field at the interface of different materials through the interface continuity loss term, the distortion of field quantities is avoided, and the simulation accuracy of the physical decoder in complex material environments is improved.
[0045] In some feasible embodiments, the training loss function further includes a data fitting loss term;
[0046] Construct the training loss function, including:
[0047] The difference between the predicted ACSM signal and the actual ACSM signal is calculated, and the data fitting loss term is constructed based on the difference.
[0048] By constraining the difference between the predicted signal and the real signal through data fitting loss term, the output of the physical decoder is ensured to match the real data, thereby enhancing the empirical accuracy of the model.
[0049] In some feasible embodiments, comparing the predicted ACSM signal with the measured signal to optimize the encoder, resulting in an optimized encoder, includes:
[0050] Multiple samples are taken from the probability distribution represented by the estimated value to obtain the permeability sample value;
[0051] Multiple permeability sample values are input into the pre-trained physical decoder to obtain multiple predicted ACSM signal sequences;
[0052] Obtain historical measurement signal sequences;
[0053] Calculate the likelihood between each of the predicted ACSM signal sequences and the historical measurement signal sequences;
[0054] Based on the likelihood, a data fitting loss term is constructed;
[0055] Calculate the divergence between the probability distribution and the preset prior probability distribution, and construct a distribution constraint loss term based on the divergence;
[0056] The parameters of the encoder are optimized by combining the data fitting loss term and the distribution constraint loss term to obtain an optimized encoder.
[0057] The encoder is optimized by sampling from the probability distribution and calculating the likelihood and divergence, enabling it to robustly inverse the probability distribution of the tube wall permeability from the ACSM signal.
[0058] In some feasible embodiments, the stress value of the oil and gas transmission pipeline is determined based on the relationship between the magnetic permeability of the pipe wall and the magnetic permeability stress of the pipeline material, including:
[0059] Based on the physical effect of contramagnetic stricture, a linear relationship between the change in the permeability of the tube wall and the stress value was confirmed.
[0060] Obtain the proportionality coefficient in the linear relationship;
[0061] Based on the proportionality coefficient, the magnetic permeability of the pipe wall is substituted into the linear relationship to obtain the stress value of the oil and gas transmission pipeline.
[0062] Based on the linear relationship confirmed by the inverse magnetostriction effect and the calibration proportionality coefficient, the permeability value is directly mapped to the stress value, realizing the rapid and accurate quantification of the final stress.
[0063] Secondly, this application also provides a stress quantification system for oil and gas transmission pipelines, used to perform the stress quantification method for oil and gas transmission pipelines described in the first aspect, comprising:
[0064] A data acquisition unit is used to acquire measurement signals from ACSM sensors on oil and gas pipelines; receive the measurement signals to obtain an estimate of the pipe wall permeability; and input the estimate to a pre-trained physical decoder to obtain a predicted ACSM signal; wherein the physical decoder is a model based on a physical information neural network, and the training process of the physical decoder incorporates the physical laws of electromagnetic fields.
[0065] An encoder optimization unit is used to compare the predicted ACSM signal with the measured signal to optimize the encoder and obtain an optimized encoder.
[0066] An encoder is used to receive the measurement signal to obtain the magnetic permeability of the pipe wall of the oil and gas transportation pipeline; and to determine the stress value of the oil and gas transportation pipeline based on the magnetic permeability of the pipe wall and the magnetic permeability-stress relationship of the pipeline material.
[0067] As can be seen from the above technical solutions, this application provides a method and system for stress quantification of oil and gas transportation pipelines. The method includes: acquiring measurement signals of the oil and gas transportation pipeline from an ACSM sensor; inputting the measurement signals into an encoder to obtain an estimated value of the pipe wall permeability; and inputting the estimated value into a pre-trained physical decoder to obtain a predicted ACSM signal. The physical decoder is a model based on a physical information neural network, and its training process incorporates electromagnetic field physical laws. The predicted ACSM signal is compared with the measurement signal to optimize the encoder, resulting in an optimized encoder. The measurement signal is input into the optimized encoder to obtain the pipe wall permeability of the oil and gas transportation pipeline. Then, based on the pipe wall permeability and the permeability-stress relationship of the pipeline material, the stress value of the oil and gas transportation pipeline is determined. Through the collaborative optimization of the encoder and the pre-trained physical decoder, accurate quantification of pipeline stress from ACSM measurement signals is achieved, solving the problem of difficulty in quantitatively assessing stress with small samples. Attached Figure Description
[0068] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0069] Figure 1 A schematic flowchart illustrating the stress quantification method for oil and gas transmission pipelines provided in this application embodiment;
[0070] Figure 2 This is a schematic diagram of the loss function construction process provided in the embodiments of this application;
[0071] Figure 3 This is a schematic diagram of the stress quantification system for oil and gas transmission pipelines provided in an embodiment of this application. Detailed Implementation
[0072] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following examples do not represent all embodiments consistent with this application.
[0073] The stress quantification method for oil and gas transmission pipelines disclosed in this application is applicable to the safety operation and maintenance monitoring scenarios of oil and gas transmission pipelines. Pipelines are buried underground or exposed to the field for a long time. Due to factors such as soil pressure, temperature changes, and internal transmission pressure, stress concentration occurs in the pipe wall. The stress concentration area is a potential risk point of the pipeline, and traditional methods are difficult to achieve quantitative monitoring of stress levels.
[0074] In some technologies, signal distortion caused by changes in the magnetic permeability of the pipe wall is detected by sensors. This method can identify the approximate area of stress concentration or defects. However, the quantitative relationship between signal distortion features and stress is complex, and traditional feature extraction methods are difficult to establish an accurate mapping. On the other hand, deep learning models can be applied directly, but a large number of stress data samples are required for training. For pipeline stress samples, obtaining them is difficult and costly, making it difficult to establish a data-driven model.
[0075] In this application, the ACSM sensor is installed on the internal detector. The ACSM sensor continuously collects the induced electromotive force signal of the pipe wall. The induced electromotive force signal is transmitted to the encoder. The encoder analyzes the signal and outputs the probability distribution of the pipe wall permeability. Then, it samples multiple permeability values from the probability distribution and inputs them into the pre-trained physical decoder. The physical decoder calculates the corresponding predicted sensor signal based on the physical laws of electromagnetic fields.
[0076] Then, the predicted signal is compared with the actual measured signal, and the encoder parameters are optimized based on the difference. The optimization process can ensure that the encoder outputs a more accurate magnetic permeability value. After obtaining the magnetic permeability, the material calibration database is called. The database stores the magnetic permeability stress-correspondence curve of a specific pipe material. The magnetic permeability value is converted into a specific stress value through a linear relationship.
[0077] This method combines physical laws with data-driven approaches. The physical decoder is constructed through a physical information neural network, and its training is based on electromagnetic field equations. This approach reduces the reliance on a large number of stress samples. The encoder focuses on signal inversion, ensuring the rationality of the inversion results through physical constraints, ultimately achieving a quantitative output of stress levels.
[0078] Furthermore, by employing a phased processing strategy, first constructing a reliable forward model using physical laws, and then optimizing the data-driven inversion model, the requirement for a large sample size can be reduced while maintaining accuracy. Specifically, some embodiments of this application provide a method for stress quantification in oil and gas transportation pipelines, such as... Figure 1 As shown, it includes the following steps S100-S600.
[0079] S100: Acquire the measurement signal of the ACSM sensor for the oil and gas pipeline.
[0080] The ACSM sensor is a device based on the principle of alternating current stress measurement. The ACSM sensor magnetizes the pipe wall by using an excitation coil and uses a detection coil to measure the induced electromotive force generated by the change in the magnetic permeability of the pipe wall.
[0081] This application requires first obtaining the measurement signal obtained by the ACSM sensor to detect the oil and gas transmission pipeline. The measurement signal is the time series data of the induced electromotive force, which is a voltage sequence that changes over time and characterizes the physical effect of the pipeline stress state.
[0082] S200: Input the measurement signal into the encoder to obtain an estimated value of the pipe wall permeability.
[0083] An encoder is a data processing model that receives measurement signals acquired by an ACSM sensor and outputs an estimated value representing the permeability of the tube wall through an internal algorithm. The estimated value can be a specific numerical value or a probability distribution to characterize the uncertainty of the permeability.
[0084] One way to implement an encoder is as a recurrent neural network based on a long short-term memory network, used to process temporal signals; another way is as a convolutional neural network, used to extract spatial features from signals.
[0085] After acquiring the measurement signal, it is input to the encoder. The encoder extracts the features related to the permeability of the pipe wall from the measurement signal, which is the time series data. The encoder processes this time series data and outputs an estimate of the current permeability of the pipe wall to convert the sensor signal into an intermediate physical quantity.
[0086] S300: Input the estimated value into the pre-trained physical decoder to obtain the predicted ACSM signal.
[0087] The physical decoder is a model based on a physical information neural network. Its training process incorporates the physical laws of electromagnetic fields. Its structure may include multiple fully connected layers and activation functions, such as an input layer, hidden layers, and an output layer. The number of nodes in the input layer matches the dimension of the physical field parameters. Each hidden layer contains 128 neurons and uses the ReLU activation function. The number of nodes in the output layer matches the dimension of the electromagnetic field quantities. The input to the physical decoder includes various physical field parameters, including magnetic permeability, and the output is a predicted value of the spatial electromagnetic field quantities. By embedding physical laws as constraints into the training process, the physical decoder can learn and simulate the working principle of the ACSM sensor.
[0088] A pre-trained physical decoder is a physical decoder model that has already completed the pre-training process and whose network parameters are fixed.
[0089] Its training data comes from computer simulations or high-fidelity experiments. By minimizing the loss function that combines physical constraints with data fitting, the model is able to accurately predict electromagnetic fields and sensor signals based on input physical parameters.
[0090] Specifically, the estimated value of the tube wall permeability is input into a pre-trained physical decoder. The physical decoder encodes the physical laws governing the operation of the ACSM sensor and can calculate the signal that the sensor should theoretically produce based on the input physical parameters, i.e., predict the ACSM signal. The reason for using a pre-trained physical decoder is to introduce the constraints of physical laws.
[0091] In this application scenario, physical field parameters are used as inputs, including spatial coordinates, excitation current parameters, and material electromagnetic properties. The outputs are the predicted values of the magnetic vector potential and electric scalar potential at the corresponding spatial coordinate points. These are then converted into predicted ACSM signals through integration, thereby establishing a physical mapping relationship from parameters to signals.
[0092] In some embodiments, before inputting the estimated value into a pre-trained physical decoder to obtain a predicted ACSM signal, the process includes: acquiring training data and an initial physical information neural network; inputting the physical field parameters into the initial physical information neural network to output the magnetic vector potential and electric scalar potential of a spatial point; outputting a predicted ACSM signal based on the magnetic vector potential; constructing a training loss function based on the magnetic vector potential, electric scalar potential, and predicted ACSM signal; and training the initial physical information neural network by minimizing the training loss function to obtain the pre-trained physical decoder.
[0093] For the physical decoder, the goal of the training process is to obtain a model that can accurately simulate the physical principles of the ACSM sensor. First, training data needs to be acquired, which includes multiple sets of physical field parameters and their corresponding real ACSM signals. The neural network weights are initialized; the Adam optimizer is used, with an initial learning rate of 0.001 and a learning rate decay strategy applied; during training, the composite training loss function, consisting of the control equation loss, boundary condition loss, excitation source constraint loss, and data fitting loss, is minimized. The number of training epochs is set to 1000 until the training loss function converges.
[0094] Physical field parameters characterize the physical environment in which the sensor operates; the actual ACSM signal is the standard output that the sensor should produce under these conditions. Specifically, physical field parameters may include the frequency of the excitation source. Amplitude of the excitation source The geometric dimensions and spatial coordinates of the detection coil. Material properties, such as magnetic permeability Electrical conductivity Ambient temperature Pipe wall curvature etc., where the geometric dimensions of the detection coil include the radius. Number of turns Distance between the two coils , lift value The actual ACSM signal can be calculated using high-precision electromagnetic field simulation software, or it can be obtained through actual measurement on a high-stress calibration device.
[0095] The initial physical information neural network is the basic framework of the physical decoder. Its structure determines the model's learning ability. Before training begins, the number of input layer nodes of the initial physical information neural network is matched with the dimension of the physical field parameters, and the number of output layer nodes is matched with the dimension of the electromagnetic field quantity. The structure of the initial physical information neural network can be a multilayer perceptron containing multiple hidden layers or a residual network.
[0096] Next, the high-dimensional vector composed of all physical parameters of the physical field is generated. The data is input into a physical information neural network, which processes the input data and outputs the magnetic vector potential at the corresponding spatial point. With electric scalar potential The network calculates the predicted values internally to establish a mapping relationship from input parameters to field outputs.
[0097] It is important to note that, due to the anisotropy of the pipe wall, in the pipe wall region, the magnetic permeability and electrical conductivity are not scalars, but vectors, decomposed into three components along the pipe axis, circumferential direction, and radial direction, as shown in the following equation:
[0098] ;
[0099] ;
[0100] in, For the axial permeability component, For the circumferential permeability component, This is the radial permeability component; For the axial conductivity component, For the circumferential conductivity component, This represents the radial conductivity component.
[0101] Among them, magnetic vector potential and electric scalar potential are the basic physical quantities describing electromagnetic fields. Magnetic vector potential is a vector function that characterizes the distribution of magnetic field, and electric scalar potential is a scalar function that characterizes the distribution of electric field. In the electromagnetic field calculation of ACSM sensor, the electric field strength and magnetic induction intensity can be indirectly obtained by solving magnetic vector potential and electric scalar potential.
[0102] Based on the magnetic vector potential, the predicted ACSM signal can be obtained through integration calculation, simulating the actual physical process inside the ACSM sensor, that is, generating induced electromotive force through magnetic field changes, converting the potential function into a measurable electrical signal, so that the model output can be directly compared with the real signal in the training data.
[0103] At other locations, the permeability and conductivity remain scalars. Therefore, the induced electromotive force in the detection coil is calculated by performing a line integral over the magnetic vector potential of the detection coil. The induced electromotive force is the predicted ACSM signal, as shown in the following formula:
[0104] ;
[0105] in, The imaginary unit, Angular frequency, For closed path integrals, It is the differential line element of the integration path.
[0106] Based on the predicted ACSM signal, a training loss function is constructed. The training loss function includes not only the data fitting error between the predicted signal and the real signal, but also the physical constraints based on Maxwell's equations to ensure that the model's prediction results are physically reasonable. Finally, the parameters of the initial physical information neural network are adjusted by minimizing the training loss function through an optimization algorithm.
[0107] It is important to note that this iterative process continues until the training loss function converges to a predetermined threshold. At this point, the network has learned the physical laws and data features and becomes a pre-trained physical decoder.
[0108] By embedding the physical laws of electromagnetic fields into the training loss function as constraints, the physical encoder not only learns the statistical features in the training data but also gains a deeper understanding of the underlying physical mechanisms. This training method reduces the reliance on large amounts of experimental data and can make full use of simulation data to build sensor physical models. The resulting pre-trained physical decoder has good extrapolation ability and physical consistency. Even for combinations of physical parameters not covered by the training data, it can give reasonable predictions that conform to physical laws, providing a benchmark for subsequent encoder optimization. This is the foundation for achieving high-precision measurements under small sample conditions.
[0109] like Figure 2 As shown, the key to the physical encoder lies in the design of the training loss function. In order for the decoder to learn physical laws, Maxwell's equations are incorporated into the training loss function of the decoder. In some embodiments, the training loss function includes a control equation loss term, an energy minimization loss term, a boundary condition loss term, an excitation source constraint loss term, an interface continuity loss term, and a data fitting loss term. These loss terms are explained below.
[0110] For the loss term of the control equation The magnetic vector potential and electric scalar potential are calculated, and the residuals under the electromagnetic field control equations are obtained. Within the computational domain, the residuals are sampled and statistically analyzed to obtain the residual results. Based on the residual results, the loss terms of the control equations are constructed, where the residual results are the loss terms of the control equations.
[0111] Specifically, according to Ampere's law, in a time-harmonic field, the electromagnetic field throughout the entire computational domain (including the tube wall, coil, and air) must satisfy the following equation:
[0112] ;
[0113] in, The gradient operator describes the rate of change of a scalar field. The magnetic field strength, For conduction current density, Angular frequency, , It is the electric displacement vector. The displacement current density is generated by the changing electric field and the change of the electric displacement vector over time.
[0114] Combining Maxwell's equations, the above formula can be further expanded into the following equation:
[0115] ;
[0116] in, The source current density is the current density generated by the excitation coil. These are the magnetic permeability and electrical conductivity at the point in space corresponding to the coordinates of the input point. Let be the dielectric constant of the material. Based on this, the loss term of the governing equation is designed as follows:
[0117] ;
[0118] in, The number of sampling points in space. This is the predicted value of the magnetic vector potential.
[0119] The manifestation of this loss term differs in different regions. Inside the excitation coil, there are source current and displacement current; therefore, The coil material is generally a non-ferromagnetic material. Approximately equal to the permeability of free space, in insulating regions such as air, only displacement current exists. , It is approximately equal to the permeability of free space.
[0120] Notably, in the tube wall region, conduction current is absolutely dominant, and the material can be ferromagnetic. At this point, since both magnetic permeability and electrical conductivity are vectors, the loss term becomes the following:
[0121] ;
[0122] Based on this, different training loss functions for the control equations are constructed for different regions and their averages are calculated.
[0123] For energy minimization loss term Based on the magnetic vector potential and the electric scalar potential, the magnetic field energy density is calculated. Within the computational domain, the magnetic field energy density is sampled and statistically analyzed to obtain the density result. Based on the density result, the energy minimization loss term is constructed. The density result is the energy minimization loss term.
[0124] This loss is based on the principle of energy minimization. Under the constraints of the existing governing equations, the model is guided to consider the anisotropy of the material. Due to the anisotropy of the material, the permeability varies in different directions, and the energy density also varies. The energy density is given by the following equation:
[0125] ;
[0126] in, The magnetic field strength, is the magnetic flux density.
[0127] Expanding Maxwell's equations, we get the following:
[0128] ;
[0129] in, The number of spatial sampling points set within the computational domain. This is the predicted value of the magnetic flux density.
[0130] in, ,and Similarly, in the air and coil region, the permeability is a scalar, while in the tube wall region, the permeability is a tensor, requiring the following transformation:
[0131] .
[0132] For boundary condition loss terms At the boundary of the computational domain, the boundary values of the magnetic vector potential and the electric scalar potential are obtained; the difference between the boundary value and the preset boundary condition value is calculated, and the boundary condition loss term is constructed based on the difference.
[0133] Specifically, the boundary loss term, based on the Dirichlet boundary condition, restricts the magnetic vector potential at the region edge to 0 to ensure the convergence of the physical model. Its specific form is as follows:
[0134] ;
[0135] in, Let be the normal unit vector of the region boundary. The number of points at the region boundary. This represents the difference between the boundary value and the preset boundary condition value.
[0136] For the incentive source constraint loss term Based on the magnetic vector potential and the electric scalar potential, the predicted source current density is calculated; in the excitation source region, the predicted source current density is subjected to surface integration to obtain the total excitation current; the difference between the total excitation current and the actual excitation current is calculated, and the excitation source constraint loss term is constructed based on the difference, wherein the difference is the excitation source constraint loss term.
[0137] Specifically, this constraint is used to ensure that within the excitation coil region, the predicted total current should be equal to the applied excitation current, in the following form:
[0138] ;
[0139] in, The predicted value of the source current density is calculated using the same formula as the method for calculating the loss function term in the governing equation. The actual value of the source current is given, and the known quantity is given. This represents the total excitation current.
[0140] For the interface continuity loss term At the interface of different material regions, the predicted magnetic vector potentials on both sides of the interface are obtained; the difference between the predicted magnetic vector potentials on both sides of the interface in the normal component is calculated, and the interface continuity loss term is constructed based on the difference, wherein the difference is the interface continuity loss term.
[0141] This portion of the loss is used to control the continuity of the electromagnetic field at the interface (air and pipe wall), avoiding distortion of the electromagnetic field at the material interface. Specifically, it takes the following form:
[0142] ;
[0143] in, The number of points on the interface. Let be the normal unit vector at the interface. The predicted magnetic vector potential on the air side, The predicted magnetic vector potential on the steel side, air permeability, denoted as ρ, where ρ is the material's magnetic permeability.
[0144] in, as well as These are known terms.
[0145] For the data fitting loss term The difference between the predicted ACSM signal and the real ACSM signal is calculated, and the data fitting loss term is constructed based on the difference, wherein the difference is the data fitting loss term.
[0146] Specifically, this part of the loss is based on high-fidelity data or measured data, and further constraints are applied at the corresponding parameter points, in the form of the following formula:
[0147] ;
[0148] in, The number of data points As given by the formula mentioned above, To predict ACSM signals, This is a true ACSM signal.
[0149] The weighted sum of the above six loss terms is used as the loss function during model training. Finally, the model directly predicts the field variables under the input parameter combinations, as shown in the following equation:
[0150] ;
[0151] Where real is the real part and imag is the imaginary part.
[0152] It is a magnetic vector potential, a complex vector, so it has six components. It is an electric scalar potential, a complex scalar, so it has two components.
[0153] S400: Compare the predicted ACSM signal with the measured signal to optimize the encoder and obtain an optimized encoder.
[0154] Optimizing an encoder involves comparing the predicted signal with the measured signal and adjusting the parameters of the encoder model. The optimization process aims to reduce the difference between the estimated permeability of the encoder output and the actual measured signal after conversion by the physical decoder, thereby improving the encoder's inversion capability.
[0155] After obtaining the physical decoder based on the PINN (Physics-Informed Neural Network) model, the next stage is to build and train a Bayesian network-based encoder. The input to this encoder is the real ACSM signal, i.e., the induced electromotive force in the detection coil. The distribution of output permeability, which varies with time.
[0156] Because real ACSM signals may contain noise and other disturbances, a Bayesian LSTM (Long Short-Term Memory) network structure is used to ensure robustness of the output. Its Bayesian property is achieved through variational inference, whereby the weight parameters in the LSTM are treated as probability distributions (usually assumed to be Gaussian), and the mean and variance of these distributions are learned through training. For example, the LSTM structure includes an input layer, an LSTM layer with 128 hidden units, and an output layer. The output layer maps the final hidden state of the LSTM to parameters characterizing the posterior distribution of permeability (e.g., the mean and variance of a Gaussian distribution). The output is the permeability distribution, not a numerical value.
[0157] Specifically, in some embodiments, comparing the predicted ACSM signal with the measured signal to optimize the encoder and obtain an optimized encoder includes: performing multiple samplings from the probability distribution characterized by the estimated permeability of the tube wall to obtain permeability sample values; inputting the multiple permeability sample values into the pre-trained physical decoder to obtain multiple predicted ACSM signal sequences; acquiring historical measurement signal sequences; calculating the likelihood between each predicted ACSM signal sequence and the historical measurement signal sequence; constructing a data fitting loss term based on the likelihood; calculating the divergence between the probability distribution and a preset prior probability distribution to construct a distribution constraint loss term based on the divergence; and combining the data fitting loss term and the distribution constraint loss term to optimize the encoder parameters to obtain an optimized encoder.
[0158] Before training begins, a training dataset needs to be prepared. Specifically, this includes: using a physical decoder, generating a corresponding theoretical ACSM signal sequence as historical measurement signals based on a set of known and varying simulation parameters of the tube wall permeability; simultaneously, simulating real disturbances by adding noise of different intensities to the theoretical signals, constructing input-output sample pairs, where the input is the noisy ACSM signal sequence and the output is the corresponding real permeability value (for verification).
[0159] First, multiple samples are taken from the probability distribution represented by the estimated permeability of the tube wall. The encoder output is not a single definite value, but a probability distribution, which means that there is uncertainty in its estimation of permeability. The sampling operation is to explore this uncertainty by extracting multiple sample values to cover the possible range of permeability values. The number of samplings needs to be large enough to ensure that it can represent the characteristics of the entire distribution.
[0160] The obtained permeability sample values are then input one by one into the pre-trained physical decoder. The physical decoder is a model that has determined the physical laws. It can calculate the theoretical signal that the ACSM sensor should produce under the input permeability parameters. This operation is performed for each sample value to obtain multiple predicted ACSM signal sequences. These sequences constitute a set of predictions for the possible output of the sensor.
[0161] Specifically, for each moment ACSM signal prior to this moment As input, the permeability is calculated using a Bayesian LSTM. ,exist Posterior distribution at time 1 K parameter values are sampled from this distribution. These K sampled values are input into the optimization decoder. The remaining physical parameters in the optimization decoder input are all measurable and are known quantities. Make predictions to obtain K prediction sequences. .
[0162] Historical measurement signal sequences are known real data and serve as a benchmark for evaluating prediction accuracy. The likelihood between each predicted ACSM signal sequence and the historical measurement signal sequence is then calculated. The likelihood measure quantifies the degree of agreement between each predicted sequence and the real signal. The closer the predicted sequence is to the real sequence, the higher its likelihood.
[0163] The data fitting loss term is constructed based on the calculated likelihood. The core idea of the data fitting loss term is that the permeability sample values corresponding to high likelihood are more likely to be correct. Therefore, the loss function encourages the encoder to output a probability distribution that makes the predicted signal close to the actual measured signal. At the same time, the divergence between the probability distribution and the preset prior probability distribution is calculated. The prior distribution represents the initial perception of permeability before the measured signal is seen, and the divergence measures the difference between the current probability distribution and this prior distribution.
[0164] Next, the training loss function is constructed, and the parameters in the Bayesian LSTM are optimized using gradient descent. During this optimization process, the decoder weights are frozen, and only the encoder is trained. The training loss function consists of two parts: data fitting loss. That is, the data fitting loss term and the KL divergence loss. That is, the distribution constraint loss term.
[0165] Among them, for data fitting loss The loss function controls the deviation between the predicted sequence and the actual input sequence, using a negative log-likelihood function, as shown in the following formula:
[0166] ;
[0167] in, For loss control prediction sequence, Let be the probability.
[0168] For KL divergence loss This part of the loss is used to apply the posterior distribution learned by the model. The model is guided towards a pre-defined distribution to prevent it from learning to an unreasonable distribution; here, a normal distribution is chosen. The KL divergence of the two distributions is used as the second loss function, and the specific form of the loss is as follows:
[0169] ;
[0170] The loss function combines the data fitting loss term and the distribution constraint loss term, i.e., the weighted sum of the two loss functions. For example, using the Adam optimizer with a learning rate of 0.001 and a batch size of 32, a Bayesian LSTM encoder is trained for 300 epochs until the total loss converges. The two factors together guide the update of the encoder parameters. Parameters are iteratively adjusted using optimization algorithms such as gradient descent, ultimately resulting in an optimized encoder that more accurately maps the measured signal to the pipe wall permeability.
[0171] In this ACSM stress measurement scenario, the encoder input is specifically set as a time window, for example, an induced electromotive force sequence with a length of 100 sampling points, where the input dimension for each time step is 1 (voltage value). The output is the posterior distribution of the permeability parameter at the current moment, specifically represented by two scalars: the mean μ of the distribution. mean and variance μ var This fully characterizes the uncertainty of permeability estimation. This setting clarifies the way the algorithm is combined with the sensor time-series signal processing scenario, as well as the intrinsic relationship between input and output data.
[0172] S500: Input the measurement signal into the optimized encoder to obtain the magnetic permeability of the pipe wall of the oil and gas transmission pipeline.
[0173] The original measurement signal is then input into the optimized encoder. Since the encoder has been optimized, its output tube wall permeability value will be more accurate and reliable.
[0174] S600: Based on the relationship between the magnetic permeability of the pipe wall and the magnetic permeability stress of the pipe material, determine the stress value of the oil and gas transmission pipeline.
[0175] The permeability-stress relationship characterizes the physical correlation between the permeability of a pipe material and the stress it is subjected to. This relationship is based on the inverse magnetostriction effect. For ferromagnetic materials, the change in permeability is linearly proportional to the applied stress value within a certain range.
[0176] Based on the pipe wall permeability provided by the optimized encoder and the permeability-stress relationship calibrated in advance through material experiments, the stress value of the oil and gas transportation pipeline is determined. Since the permeability-stress relationship is linear, the permeability value can be mapped to a specific stress value through simple linear calculation, thereby achieving stress quantification.
[0177] In some embodiments, determining the stress value of the oil and gas transmission pipeline based on the pipe wall permeability and the permeability-stress relationship of the pipeline material includes: confirming the linear relationship between the change in pipe wall permeability and the stress value according to the inverse magnetostrictive physical effect; obtaining the proportionality coefficient in the linear relationship; and substituting the pipe wall permeability into the linear relationship based on the proportionality coefficient to obtain the stress value of the oil and gas transmission pipeline.
[0178] Specifically, due to the contramagnetic-strictive physical effect of ferromagnetic materials, meaning that the change in magnetic permeability of ferromagnetic materials is proportional to the pipe stress, the proportionality coefficient can be calculated by mapping the pipe wall permeability to the pipe stress using the following formula:
[0179] ;
[0180] in, This represents the stress value of the pipe wall. The saturation magnetostriction coefficient is 1. The saturation magnetic flux density It represents the relative permeability under a constant external magnetic field.
[0181] For a certain material, the above parameters, except Since all external values are constant, the relationship between the permeability of the pipe wall and the applied stress can be considered as a linear function. By calibrating a very small amount of data on the pipe wall material, the corresponding curve of material permeability-stress can be obtained. In other words, the proportional coefficient in the linear relationship can be obtained through material calibration. Since the relationship is linear, the calculation process is simplified to a single multiplication operation. The permeability value obtained by real-time inversion is subtracted from the reference permeability under stress-free conditions to obtain the change in permeability. Multiplying this change by the reciprocal of the proportional coefficient, the stress value can be directly output.
[0182] As can be seen from the above technical solution, the oil and gas pipeline stress quantification method provided in this embodiment introduces a pre-trained physical decoder and incorporates physical laws as constraints into the encoder optimization process. This method utilizes the forward mapping capability of the physical decoder from physical parameters to sensor signals to guide the encoder to learn the inverse mapping relationship of physical parameters from sensor signals. This makes the encoder training no longer completely dependent on a large number of hard-to-obtain pipeline stress real value samples, but can use easily obtained simulation data or calibration data to train the physical decoder, and then optimize the encoder through the principle of physical consistency.
[0183] Therefore, this method addresses the challenge of establishing a quantitative relationship between ACSM signals and pipeline stress when the actual sample size of pipeline stress is insufficient. Furthermore, because the physical decoder incorporates electromagnetic field physics, it can comprehensively analyze complex physical fields, avoiding the problems of incomplete feature extraction and inability to cover all effective signal features found in traditional methods. Ultimately, this method achieves accurate quantitative assessment of oil and gas pipeline stress under small sample conditions.
[0184] Based on the above-mentioned method for stress quantification in oil and gas transmission pipelines, such as Figure 3 As shown, some embodiments of this application provide a stress quantification system for oil and gas transmission pipelines, including:
[0185] The data acquisition unit is used to acquire the measurement signal of the ACSM sensor on the oil and gas pipeline; and to receive the measurement signal to obtain an estimated value of the magnetic permeability of the pipe wall; and to input the estimated value into a pre-trained physical decoder to obtain a predicted ACSM signal; wherein the physical decoder is a model based on a physical information neural network, and the training process of the physical decoder incorporates the physical laws of electromagnetic fields.
[0186] An encoder optimization unit is used to compare the predicted ACSM signal with the measured signal to optimize the encoder and obtain an optimized encoder.
[0187] An encoder is used to receive the measurement signal to obtain the magnetic permeability of the pipe wall of the oil and gas transportation pipeline; and to determine the stress value of the oil and gas transportation pipeline based on the magnetic permeability of the pipe wall and the magnetic permeability-stress relationship of the pipeline material.
[0188] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.
Claims
1. A method for stress quantification in oil and gas transmission pipelines, characterized in that, include: Acquire measurement signals from ACSM sensors for oil and gas pipelines; The measured signal is input into the encoder to obtain an estimated value of the pipe wall permeability; The estimated value is input into a pre-trained physical decoder to obtain the predicted ACSM signal; wherein the physical decoder is a model based on a physical information neural network, and the training process of the physical decoder incorporates the physical laws of electromagnetic fields. The predicted ACSM signal is compared with the measured signal to optimize the encoder, resulting in an optimized encoder. The measurement signal is input into the optimized encoder to obtain the magnetic permeability of the pipe wall of the oil and gas transportation pipeline; Based on the relationship between the magnetic permeability of the pipe wall and the magnetic permeability stress of the pipe material, the stress value of the oil and gas transmission pipeline is determined. Before inputting the estimated value into a pre-trained physical decoder to obtain the predicted ACSM signal, the process includes: Acquire training data, which includes multiple sets of physical field parameters and the corresponding real ACSM signals; Neural network for acquiring initial physical information; The physical field parameters are input into the initial physical information neural network to output the magnetic vector potential and electric scalar potential of the spatial points; Based on the magnetic vector potential, a predicted ACSM signal is output; Based on the magnetic vector potential, the electric scalar potential, and the predicted ACSM signal, a training loss function is constructed. The initial physical information neural network is trained by minimizing the training loss function to obtain the pre-trained physical decoder; The training loss function is obtained by weighted summation of six loss terms, which include the control equation loss term, energy minimization loss term, boundary condition loss term, excitation source constraint loss term, interface continuity loss term, and data fitting loss term.
2. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, The training loss function includes a control equation loss term and an energy minimization loss term; Construct the training loss function, including: Calculate the residuals of the magnetic vector potential and the electric scalar potential under the electromagnetic field control equations; Calculate the magnetic field energy density based on the magnetic vector potential and the electric scalar potential; Within the computational domain, sampling and statistics are performed on the residuals to obtain residual results; and within the computational domain, sampling and statistics are performed on the magnetic field energy density to obtain density results; Based on the residual results, the loss term of the governing equation is constructed, and based on the density results, the energy minimization loss term is constructed.
3. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, The training loss function also includes a boundary condition loss term; Construct the training loss function, including: At the boundary of the computational domain, the boundary values of the magnetic vector potential and the electric scalar potential are obtained; The difference between the boundary value and the preset boundary condition value is calculated, and the boundary condition loss term is constructed based on the difference.
4. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, The training loss function further includes an excitation source constraint loss term; the physical field parameters include source current density. Construct the training loss function, including: Based on the magnetic vector potential and the electric scalar potential, the predicted source current density is calculated; In the excitation source region, the predicted source current density is subjected to surface integration to obtain the total excitation current; The difference between the total excitation current and the actual excitation current is calculated, and the excitation source constraint loss term is constructed based on the difference.
5. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, The training loss function also includes an interface continuity loss term; Construct the training loss function, including: At the interface between different material regions, the predicted magnetic vector potentials on both sides of the interface are obtained. The difference in the normal component of the predicted magnetic vector potential on both sides of the interface is calculated, and the interface continuity loss term is constructed based on the difference.
6. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, The training loss function also includes a data fitting loss term; Construct the training loss function, including: The difference between the predicted ACSM signal and the actual ACSM signal is calculated, and the data fitting loss term is constructed based on the difference.
7. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, Comparing the predicted ACSM signal with the measured signal to optimize the encoder, resulting in an optimized encoder, includes: Multiple samples are taken from the probability distribution represented by the estimated value to obtain the permeability sample value; Multiple permeability sample values are input into the pre-trained physical decoder to obtain multiple predicted ACSM signal sequences; Obtain historical measurement signal sequences; Calculate the likelihood between each of the predicted ACSM signal sequences and the historical measurement signal sequences; Based on the likelihood, a data fitting loss term is constructed; Calculate the divergence between the probability distribution and the preset prior probability distribution, and construct a distribution constraint loss term based on the divergence; The encoder parameters are optimized by combining the data fitting loss term and the distribution constraint loss term to obtain an optimized encoder.
8. The method for stress quantification in oil and gas transmission pipelines according to claim 1, characterized in that, Based on the relationship between the magnetic permeability of the pipe wall and the magnetic permeability-stress relationship of the pipe material, the stress value of the oil and gas transmission pipeline is determined, including: Based on the physical effect of contramagnetic stricture, a linear relationship between the change in the permeability of the tube wall and the stress value was confirmed. Obtain the proportionality coefficient in the linear relationship; Based on the proportionality coefficient, the magnetic permeability of the pipe wall is substituted into the linear relationship to obtain the stress value of the oil and gas transmission pipeline.
9. A stress quantification system for oil and gas transmission pipelines, characterized in that, A method for quantifying stress in an oil and gas pipeline according to any one of claims 1-8 includes: A data acquisition unit is used to acquire measurement signals from ACSM sensors on oil and gas pipelines; receive the measurement signals to obtain an estimate of the pipe wall permeability; and input the estimate to a pre-trained physical decoder to obtain a predicted ACSM signal; wherein the physical decoder is a model based on a physical information neural network, and the training process of the physical decoder incorporates the physical laws of electromagnetic fields. An encoder optimization unit is used to compare the predicted ACSM signal with the measured signal to optimize the encoder and obtain an optimized encoder. An encoder is used to receive the measurement signal to obtain the magnetic permeability of the pipe wall of the oil and gas transportation pipeline; and to determine the stress value of the oil and gas transportation pipeline based on the magnetic permeability of the pipe wall and the magnetic permeability-stress relationship of the pipeline material. Before inputting the estimated value into a pre-trained physical decoder to obtain the predicted ACSM signal, the process includes: acquiring training data, which contains multiple sets of physical field parameters and corresponding real ACSM signals; acquiring an initial physical information neural network; inputting the physical field parameters into the initial physical information neural network to output the magnetic vector potential and electric scalar potential of spatial points; outputting the predicted ACSM signal based on the magnetic vector potential; constructing a training loss function based on the magnetic vector potential, electric scalar potential, and predicted ACSM signal; and training the initial physical information neural network by minimizing the training loss function to obtain the pre-trained physical decoder. The training loss function is obtained by weighted summation of six loss terms, which include the control equation loss term, energy minimization loss term, boundary condition loss term, excitation source constraint loss term, interface continuity loss term, and data fitting loss term.