A multi-oil and gas pipeline stress sensor data fusion method and system

Abstract

The application relates to the field of pipeline detection, and provides a multi-oil-gas conveying pipeline stress sensor data fusion method and system.The method comprises the following steps: acquiring first detection data collected by an alternating current stress measurement sensor and second detection data collected by a residual magnetism method stress sensor; based on a first feature extraction model, performing first feature vector extraction on the first detection data to obtain a first feature vector; based on a second feature extraction model, performing feature vector extraction on the second detection data to obtain a second feature vector; fusing the first feature vector and the second feature vector to obtain fusion features, and mapping the fusion features into a pipeline stress value through a stress mapping model.The method is based on the implicit function fitting capability of a deep learning model and feature vectors, establishes a mapping relationship between sensor signals and actual pipeline stresses, realizes data analysis and fusion of the sensor, and solves the problems of insufficient mapping relationship quantitative precision and poor reliability.

Description

Technical Field

[0001] This application relates to the field of pipeline inspection technology, and in particular to a method and system for fusing stress sensor data from multiple oil and gas pipelines. Background Technology

[0002] The safe operation and maintenance of oil and gas pipelines relies on online non-destructive testing (NDT) technology. For stress detection, either an alternating current stress measurement (ACSM) sensor or a remnant magnetic field (RMF) sensor can be used. ACSM sensors utilize alternating magnetic field excitation to detect changes in induced voltage or impedance; they are sensitive to stress but susceptible to lift-off and attitude interference, resulting in significant data noise. RMF sensors detect leakage magnetic fields caused by stress concentration under the influence of the Earth's magnetic field; the signal is stable, but the quantitative relationship is unclear, and they are insensitive to low stress levels.

[0003] Alternatively, a dual-modal detection method combining active and passive sensors can be used to combine the two types of sensors. Hardware design and stress quantization algorithms can be set up, and the algorithm can extract feature values ​​such as vector modulus and gradient features from the sensor signals. It attempts to fit the mapping relationship between the sensor signals and the actual stress in the pipeline, so as to take advantage of the complementary advantages of the two types of sensors.

[0004] However, the dual-modal detection method that combines active and passive modes completes the detection process through multiple feature values, but the feature values ​​it relies on fail to fully cover the characteristics of pipeline stress-related signals, resulting in insufficient quantification accuracy and poor reliability of the established mapping relationship. Summary of the Invention

[0005] This application provides a method and system for fusing stress sensor data from multiple oil and gas pipelines to solve the problems of insufficient quantification accuracy and poor reliability of mapping relationships.

[0006] In a first aspect, this application provides a method for fusing stress sensor data from multiple oil and gas pipelines, including:

[0007] Acquire first detection data and second detection data, wherein the first detection data is data collected by an AC stress measurement sensor and the second detection data is data collected by a remanent magnetization stress sensor;

[0008] Based on the first feature extraction model, a first feature vector is extracted from the first detection data to obtain a first feature vector. The first feature extraction model is used to extract stress features based on geometric changes.

[0009] Based on the second feature extraction model, feature vector extraction is performed on the second detection data to obtain a second feature vector. The second feature extraction model is used to extract spatial features and temporal features.

[0010] The first feature vector and the second feature vector are fused to obtain fused features, and the fused features are mapped to pipeline stress values ​​through a stress mapping model.

[0011] In some feasible embodiments, based on a first feature extraction model, a first feature vector extraction is performed on the first detection data to obtain a first feature vector, prior to which the following steps are included:

[0012] A denoising model is constructed, which includes a first input channel and a second input channel, and the denoising model is obtained by training with a loss function.

[0013] A sinusoidal signal of the same length as the first detection data is generated, and the frequency of the sinusoidal signal is the same as the excitation frequency of the AC stress measurement sensor.

[0014] The first detection data is input to the first input channel, and the sine signal is input to the second input channel;

[0015] The attention block is used to fuse the features of the first detection data and the sinusoidal signal to output the denoised first detection data. The attention block is set at the front end of the denoising model.

[0016] In some feasible embodiments, based on a first feature extraction model, a first feature vector extraction is performed on the first detection data to obtain a first feature vector, including:

[0017] A graph neural network is constructed, wherein the graph neural network includes nodes and edges, the nodes correspond to the data acquisition points of the AC stress measurement sensor, the features of the nodes include spatial coordinates and pseudo-displacement, the edges are formed by constructing connections between the nodes through the spatial proximity relationship or time series proximity relationship of the data acquisition points, and the graph neural network is obtained after training with a physical constraint loss function;

[0018] The denoised first detection data is input into the graph neural network to calculate the stress feature vector of each node through the graph neural network;

[0019] The first feature vector is formed based on the stress feature vectors of all nodes.

[0020] In some feasible embodiments, the physical constraints include a first constraint and a second constraint; the first constraint is used to constrain the balance relationship of stress eigenvectors between adjacent nodes, and the second constraint is used to constrain the constitutive relationship between the stress eigenvectors and the strain calculated based on the pseudo-displacement;

[0021] Constructing a graph neural network includes:

[0022] The graph neural network is trained by minimizing the physical constraint loss function to output a stress feature vector, which is a vector that includes data features and conforms to the laws of pipeline mechanics.

[0023] In some feasible implementations, a denoising model is constructed, including:

[0024] The difference between the denoised output signal and the real signal in the first norm in the time domain is calculated and used as a time domain difference constraint.

[0025] Perform a short-time Fourier transform on the denoised output signal to obtain a first spectrum, and perform a short-time Fourier transform on the real signal to obtain a second spectrum;

[0026] Based on the excitation frequency of the AC stress measurement sensor, the center excitation frequency is determined to be a narrow frequency range of the excitation frequency.

[0027] Calculate the spectral component difference between the first spectrum and the second spectrum within the narrowband frequency range and apply a first weight; and calculate the spectral component difference outside the narrowband frequency range and apply a second weight, using the norm of the weighted spectral difference as a frequency domain distribution constraint.

[0028] The loss function is obtained by adding the time-domain difference constraint and the frequency-domain distribution constraint through a balance coefficient.

[0029] The loss function is minimized to generate a denoising model, wherein the signal waveform output by the denoising model is within a preset difference range from the sinusoidal shape.

[0030] In some feasible embodiments, the second feature extraction model includes a convolutional network model and a recurrent neural network; the second feature vector includes a spatial feature vector and a temporal feature vector.

[0031] Based on the second feature extraction model, feature vector extraction is performed on the second detection data to obtain a second feature vector, including:

[0032] The three-axis components of the second detection data are convolved by a convolutional neural network to extract spatial features and obtain a spatial feature vector.

[0033] The second detection data is processed by a recurrent neural network with a gated switch and a rotating cross-attention mechanism to extract features in the time dimension and obtain a time feature vector.

[0034] In some feasible embodiments, the time feature vector is obtained, including:

[0035] Calculate the principal stress direction and principal stress intensity of the triaxial components at each time step;

[0036] When the change in the principal stress direction exceeds a set angle, or the change in the principal stress intensity exceeds a set percentage, the rotating cross-attention mechanism is activated to generate the time feature vector of the current axis based on the feature vectors of the historical states of the other two axes.

[0037] When the change in the principal stress direction does not exceed a set angle, or the change in the principal stress intensity does not exceed a set percentage, the self-attention mechanism is activated to generate a time feature vector of the current axis based on the feature vector of the current axis's own historical state.

[0038] In some feasible embodiments, the first feature vector and the second feature vector are fused to obtain fused features, including:

[0039] Feature fusion is performed through a hierarchical coding fusion network, which includes a self-attention encoder, a first feature compression network, a second feature compression network, and a cross-fusion block.

[0040] The first feature vector is input into the self-attention encoder to perform compression encoding, so as to output the first encoded vector;

[0041] The spatial feature vector is input into the first feature compression network to perform compression encoding, so as to output a spatial encoded vector;

[0042] The time feature vector is input into the second feature compression network to perform compression encoding, and a time encoded vector is output.

[0043] The spatial encoding vector and the temporal encoding vector are input into the cross-fusion block to obtain the fused features.

[0044] In some feasible embodiments, the fused features are mapped to pipe stress values ​​using a stress mapping model, including:

[0045] The fused features are input into the stress mapping model, which includes a cross-attention mechanism and a fully connected layer.

[0046] The fused features and the first encoding vector are input into the cross-attention mechanism to generate context-enhanced feature representations;

[0047] The context-enhanced feature representation is input into the fully connected layer, where linear transformation and dimensionality reduction are performed to output the pipe stress value.

[0048] Secondly, this application provides a multi-oil and gas pipeline stress sensor data fusion system for performing the multi-oil and gas pipeline stress sensor data fusion method described in the first aspect, comprising:

[0049] The acquisition module is used to acquire first detection data and second detection data. The first detection data is data acquired by an AC stress measurement sensor, and the second detection data is data acquired by a remanent magnetization stress sensor.

[0050] The feature extraction module is configured to perform a first feature vector extraction on the first detection data based on a first feature extraction model to obtain a first feature vector, wherein the first feature extraction model is used to extract stress features based on geometric changes; and to perform feature vector extraction on the second detection data based on a second feature extraction model to obtain a second feature vector, wherein the second feature extraction model is used to extract spatial features and temporal features.

[0051] The mapping module is used to fuse the first feature vector and the second feature vector to obtain fused features, and to map the fused features to pipeline stress values ​​through a stress mapping model.

[0052] As can be seen from the above technical solutions, this application provides a method and system for fusing stress sensor data from multiple oil and gas pipelines. The method includes: acquiring first detection data and second detection data, wherein the first detection data is data collected by an AC stress measurement sensor and the second detection data is data collected by a remanent magnetization stress sensor; then, based on a first feature extraction model, performing a first feature vector extraction on the first detection data to obtain a first feature vector, wherein the first feature extraction model is used to extract stress features based on geometric changes; and based on a second feature extraction model, performing feature vector extraction on the second detection data to obtain a second feature vector, wherein the second feature extraction model is used to extract spatial and temporal features; fusing the first feature vector and the second feature vector to obtain fused features, and mapping the fused features to pipeline stress values ​​through a stress mapping model. This method, based on the implicit function fitting capability of deep learning models and based on feature vectors, establishes a mapping relationship between sensor signals and actual pipeline stress, realizing sensor data analysis and fusion, and solving the problems of insufficient quantization accuracy and poor reliability of the mapping relationship. Attached Figure Description

[0053] 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.

[0054] Figure 1 A schematic flowchart illustrating the multi-oil and gas pipeline stress sensor data fusion method provided in this application embodiment;

[0055] Figure 2 This is a schematic diagram of spatial feature vector generation provided in an embodiment of this application;

[0056] Figure 3 This is a schematic diagram illustrating the processing of axial and radial stress data provided in the embodiments of this application;

[0057] Figure 4 This is a schematic diagram illustrating the generation of time feature vectors in an embodiment of this application.

[0058] Figure 5 This is a schematic diagram illustrating the generation of pipeline stress values ​​provided in an embodiment of this application. Detailed Implementation

[0059] 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.

[0060] In the field of pipeline stress detection technology, the active-passive collaborative dual-modal detection method combines AC stress measurement sensors and remanent magnetization stress sensors to construct a hardware system and design corresponding stress quantification algorithms. Specifically, it extracts a variety of preset feature values ​​from the signals of the two types of sensors, including manually designed indicators such as vector modulus and gradient features, and attempts to establish a mathematical mapping relationship between sensor signals and actual pipeline stress through the combination of these feature values, so as to take advantage of the complementary advantages between the stress sensitivity of AC stress measurement sensors and the signal stability of remanent magnetization stress sensors.

[0061] However, the above approach relies heavily on a limited set of pre-selected eigenvalues, which are based on researchers’ prior understanding of physical phenomena and have limitations in engineering practice.

[0062] In actual, complex pipeline operating conditions, stress-related signal characteristics exhibit diverse and mutually coupled forms. Pre-designed indicators such as vector modulus and gradient features are difficult to fully cover all signal patterns related to the actual stress state of the pipeline. This incompleteness leads to information gaps in the established mapping relationship model, making it difficult for the model to accurately capture subtle changes and complex patterns in the stress state.

[0063] Specifically, the set of feature values ​​extracted from sensor signals by quantization algorithms fails to fully characterize changes in pipeline geometry, such as the special stress distribution characteristics caused by abrupt changes in local shape, like elbows and welds. At the same time, the simple feature values ​​on which the rich spatiotemporal correlation information contained in the residual magnetism stress sensor data, including the spatial coordination relationship between triaxial data and the dynamic law of stress state evolution over time, lack an effective extraction and utilization mechanism. This artificial feature-based approach is limited by the designer's depth of understanding of complex physical phenomena and is difficult to adaptively discover those unrecognized but actually effective feature patterns from the raw data.

[0064] To address the aforementioned problems, some embodiments of this application provide a method for fusing stress sensor data from multiple oil and gas pipelines, such as... Figure 1 As shown, it includes the following steps S100-S400.

[0065] S100: Acquire the first detection data and the second detection data.

[0066] The first detection data is the data collected by the AC Stress Measurement Sensor (ACSM sensor). The ACSM sensor applies an alternating magnetic field of fixed frequency to the pipe wall through an excitation source, and uses a detection coil to sense the induced signal generated by the change in the magnetic permeability of the pipe wall due to stress. This is the first detection data, also known as ACSM.

[0067] The first detection data is a sequence that changes over time. Ideally, its waveform should be a sine wave. However, in actual signals, noise introduced by sensor lift-off effect, attitude change, and environmental electromagnetic interference will be mixed in.

[0068] The second set of detection data consists of data collected by a residual magnetism (RMF) stress sensor. The RMF sensor is used to detect the weak leakage magnetic field generated in a pipeline due to stress concentration, causing the internal magnetic domains to rearrange. The RMF sensor can simultaneously measure changes in magnetic flux or magnetic field strength in the axial, circumferential, and radial directions of the pipeline. Therefore, the second set of detection data includes triaxial components acquired synchronously over time. While the triaxial components are correlated with the pipeline's stress state, the quantitative mapping relationship is unclear, and the response is particularly insensitive in low-stress areas.

[0069] S200: Based on the first feature extraction model, perform first feature vector extraction on the first detection data to obtain the first feature vector.

[0070] The first feature extraction model is used to extract stress features based on geometric changes. For example, the geometry of a pipeline changes at bends, welds, and defective areas, resulting in characteristic changes in the local stress distribution. The first feature extraction model identifies and extracts these stress patterns corresponding to specific geometric structures from sensor signals, thereby forming a first feature vector.

[0071] The first feature extraction model is used to extract features from the first detection data. Specifically, it is designed for the ACSM sensor, which is affected by various factors such as the intensity and direction of the tube wall stress, the nonlinear change in permeability caused by the anisotropy of the material, coil lifting, and high noise caused by disturbances. Therefore, in this embodiment, the first feature extraction model includes two models: a denoising model and a graph neural network.

[0072] In this process, noise and outliers are removed from the original signal, i.e., the first detection data, using a denoising model, and then stress feature signals are extracted using a neural network to obtain the first feature vector.

[0073] First, we introduce the denoising model. In some embodiments, a denoising model is constructed before performing a first feature vector extraction on the first detection data based on a first feature extraction model to obtain the first feature vector. The denoising model is used to eliminate noise in the signal. In this embodiment, it is used to eliminate noise in the first detection data. After training, it can recover the effective underlying true signal from the noisy first detection data.

[0074] In one implementation, the denoising model can be a one-dimensional convolutional neural network, which extracts signal features through an encoder and then reconstructs a clean signal through a decoder. In another implementation, the denoising model can be a filter bank based on the transform domain, which suppresses noise components by transforming the signal to the frequency domain and then inversely transforming it back to the time domain.

[0075] Since the ACSM sensor's detection principle involves passing a fixed-frequency alternating current through the excitation coil, the detection coil should generate an induced signal with the same frequency, which is a sine wave. Therefore, in this embodiment, the denoising model is a U-Net (U-shaped Network) model, whose structure includes symmetrical encoder and decoder paths, and uses skip connections to concatenate the feature maps of each layer of the encoder with the feature maps of the corresponding layers of the decoder. The encoder consists of multiple cascaded convolutional layers and pooling layers, used for progressive downsampling and extracting multi-scale features; the decoder consists of multiple cascaded transposed convolutional layers or upsampling layers, used for progressive upsampling and reconstructing the signal.

[0076] The U-Net model includes a first input channel and a second input channel. The first input channel receives the signal acquired by the ACSM sensor, i.e., the first detection data. The second input channel receives a sine signal of the same length as the signal acquired by the ACSM sensor. The second input channel is set in parallel with the first input channel to ensure that the reference signal can be processed by the model synchronously with the original signal.

[0077] A sinusoidal signal is a periodic waveform signal with a fixed frequency. The waveform of a sinusoidal signal follows the mathematical laws of the sine function, and its frequency is completely consistent with the frequency of the alternating current flowing through the excitation coil inside the AC stress measurement sensor. See the following formula:

[0078] ;

[0079] in, The amplitude of the sine wave. The frequency of the sine wave is the fixed frequency of the alternating current flowing through the excitation coil in the ACSM sensor. For time, This represents the initial phase of the sine wave. The specific values ​​of amplitude and initial phase are obtained through finite element simulation of the electromagnetic response of the pipeline under typical stress conditions, or through sensor calibration experiments on actual calibrated pipeline sections.

[0080] The U-Net model is obtained through training with a loss function. This loss function guides the training process of the denoising model. By calculating the difference between the denoised output signal and the desired true signal, the loss function provides direction and measurement for optimizing the model parameters. During training, the denoising capability of the U-Net model is gradually improved by continuously adjusting the internal parameters of the model to minimize the value of the loss function.

[0081] Training the model requires a training dataset containing a large number of paired samples. Each sample includes a noisy ACSM signal generated in simulation or actually collected as input, and its corresponding real clean signal as a training label. The real clean signal can be obtained by simulating noiseless ideal operating conditions using high-precision electromagnetic simulation software, or by calibrating pipes under known stress states in a laboratory environment with strict shielding against external interference. To enhance the model's generalization ability, noisy samples can be constructed by artificially synthesizing noise of different intensities and types (such as Gaussian white noise and impulse noise) on top of the simulated or measured clean signal.

[0082] In some embodiments, prior to building the denoising model, the following is included:

[0083] The difference between the denoised output signal and the real signal in the first norm in the time domain is calculated and used as a time domain difference constraint.

[0084] Perform a short-time Fourier transform on the denoised output signal to obtain the first spectrum, and perform a short-time Fourier transform on the real signal to obtain the second spectrum;

[0085] Based on the excitation frequency of the AC stress measurement sensor, the center excitation frequency is determined to be a narrow frequency range of the excitation frequency.

[0086] Calculate the spectral component differences between the first spectrum and the second spectrum within the narrowband frequency range and apply a first weight; and calculate the spectral component differences outside the narrowband frequency range and apply a second weight, using the norm of the weighted spectral differences as a frequency domain distribution constraint.

[0087] The time-domain difference constraint and the frequency-domain distribution constraint are added together by a balancing coefficient to obtain the loss function;

[0088] Minimize the loss function to generate a denoised model. The signal waveform output by the denoised model has a difference range from the sinusoidal shape.

[0089] The process of constructing the loss function first requires calculating the time-domain difference constraint. In each training iteration, the denoised output signal is compared with the corresponding real signal point by point in the time domain. The denoised output signal is the noisy input signal of the denoised model to be trained, that is, the intermediate result generated after processing the first detection data. During the model training phase, the denoised output signal is the estimated value of the real clean signal calculated by the model based on the current parameters. There is a difference between it and the final expected real signal. This difference is used to guide the update of the model parameters.

[0090] Then, the sum of the absolute values ​​of the differences between the denoised output signal and the corresponding real signal is calculated, which is the first norm difference. This serves as a time-domain difference constraint term, representing the degree of similarity between the denoised signal and the ideal signal in the overall waveform profile. Furthermore, it is insensitive to sharp noise points that may exist in the signal, which is beneficial to the stability of training.

[0091] Next, the frequency domain distribution constraints are calculated. Specifically, short-time Fourier transforms are applied to both the denoised output signal and the true signal to obtain the first and second spectra. These spectra characterize the evolution of the signal frequency components over time. Then, based on the known excitation frequency of the AC stress measurement sensor, a narrow frequency band is determined. This narrow frequency band represents the region where the effective signal energy should be concentrated.

[0092] Based on the obtained spectrum, the spectral component differences are calculated. Specifically, the differences between the first and second spectra at all frequency points are calculated, and then spatially varying weights are applied to these differences.

[0093] Within the previously defined narrowband frequency range, a larger first weight, for example, 1, is applied to the spectral component differences; outside the narrowband frequency range, a smaller second weight, for example, 0.2, is applied. The square of the L2 norm of all weighted spectral differences is calculated, and the result is used as a frequency domain distribution constraint term. This forces the energy of the denoised signal output by the model to be mainly distributed near the expected excitation frequency, thereby ensuring that the basic shape of the output signal approaches a sine wave.

[0094] After obtaining the time-domain difference constraint and the frequency-domain distribution constraint respectively, the combined loss function is shown in the following equation:

[0095] ;

[0096] in, The output signal is the signal output by the denoising model. For real signals, They represent The results after the short-time Fourier transform, namely the first spectrum and the second spectrum, These are the coefficients at different frequencies.

[0097] At excitation frequency Set up a narrow strip nearby ,area Inside The value is 1 for the first position and 0.2 for the rest. The coefficient used to balance the losses of the two parts, i.e., the balance coefficient.

[0098] The first part of the loss function is used to control the difference between the denoised result and the true result. The 1 norm is used to improve the robustness to outliers. The second part is used to control the components of the denoised result at different frequencies, ensuring that most of the signal is concentrated near the excitation frequency, thereby ensuring that the basic shape of the signal is sinusoidal.

[0099] During training, stochastic gradient descent or its variants are used as the optimizer, with initial learning rate, batch size, and training epochs set. In the forward propagation process, the noisy signal and a sinusoidal reference signal are input into the U-Net model to obtain the predicted signal y. pred Calculate the value of the loss function L. During backpropagation, calculate the gradient of the loss function with respect to the parameters of each model layer, and use the optimizer to update the weights and bias parameters in all convolutional and attention layers. Iterate through forward and backpropagation until the loss function converges or the predetermined number of training epochs is reached. After training is complete, save the model parameters.

[0100] By minimizing the loss function and using optimization algorithms such as gradient descent, all internal parameters of the denoising model are iteratively updated. The training objective is to find a set of model parameters that enable the model to produce an output signal for any noisy input. This output signal closely follows the real signal in the time domain and its energy is highly concentrated in a narrow band near the excitation frequency in the frequency domain. The model is trained independently, similar to a conventional U-Net, by progressively adding noise to simulated or measured signals and using the results before and after each step as training samples to train the model.

[0101] When the trained denoising model is applied to a real-world pipeline inspection scenario, the first detection data collected in real time and the corresponding sinusoidal reference signal are input into the model. The model output is the denoised first detection data, which will be used as the input to the subsequent first feature extraction model to extract stress features related to the pipeline geometry.

[0102] After training the U-Net, denoised first detection data can be obtained based on the first detection data. In some embodiments, a sinusoidal signal of the same length as the first detection data is generated, and the frequency of the sinusoidal signal is the same as the excitation frequency of the AC stress measurement sensor. The first detection data is input to the first input channel, and the sinusoidal signal is input to the second input channel. The features of the first detection data and the sinusoidal signal are fused through an attention block to output the denoised first detection data. The attention block is set at the front end of the denoising model.

[0103] The first detection data, which contains noise, is input into the first input channel of the denoising model. At the same time, a sinusoidal signal is input into the second input channel of the denoising model. The two signals enter the model synchronously and are processed by the attention block set at the front end of the model. Attention is used to realize feature interaction and fusion between different signal sources. By calculating the correlation weights between query vector, key vector and value vector, the model is allowed to dynamically extract useful feature information from the reference sinusoidal signal when processing the original sensor signal and fuse it into the feature representation of the main signal.

[0104] The attention block uses the signal from the first input channel as the dominant information and the ideal signal from the second input channel as the reference information. It calculates the correlation between the dominant signal and the reference signal at various positions through an internal mechanism, and selectively injects the waveform characteristics of the ideal sinusoidal signal into the feature representation of the dominant signal based on these correlation weights, thereby achieving deep fusion of the two signals at the feature level.

[0105] After being enhanced by attention blocks, the signal continues to pass through the backbone network of the denoising model, such as a convolutional network with an encoder-decoder structure, for deeper feature extraction and signal reconstruction. Finally, the denoising model outputs the processed result, namely the denoised first detection data. This output data will replace the first detection data and be sent to the graph neural network for feature extraction.

[0106] Graph neural networks provide stress signal feature vectors based on the basic geometry of the pipe wall, mainly extracting stress features formed by geometric changes in pipe components such as bends and welds, as well as defects.

[0107] To construct a basic mechanical graph based on the spatial structure, the graph neural network takes the first detection data after denoising as input and outputs the stress feature vector.

[0108] In some embodiments, the graph neural network is composed of multiple stacked graph convolutional layers or graph attention layers. The graph neural network includes nodes and edges, where nodes correspond to data acquisition points of the AC stress measurement sensor, and the features of the nodes include spatial coordinates. and pseudo-displacement Edges are formed by connecting nodes through spatial or temporal proximity relationships of data collection points.

[0109] Spatial coordinates are derived from the sensor's positioning system within the pipeline. The pseudo-displacement is a learnable parameter, initially set to zero. Next, edges need to be established between nodes. Edge construction can be based on two relationships: spatial proximity, where the Euclidean distance between nodes is calculated based on the 3D coordinates of all data acquisition points, and connections are established between node pairs whose distance is less than a set threshold; and temporal proximity, where connections are established between nodes corresponding to adjacent data acquisition points according to the time sequence of data acquisition. Through these edges, the graph neural network forms a topology capable of information transmission.

[0110] When training a graph neural network, it is trained using a physical constraint loss function. In some embodiments, the physical constraints include a first constraint and a second constraint. The first constraint is used to constrain the balance relationship of stress feature vectors between adjacent nodes, and the second constraint is used to constrain the constitutive relationship between the stress feature vectors and the strain calculated based on pseudo-displacement.

[0111] The loss function is a physical constraint term, and its specific form is as follows:

[0112] ;

[0113] in, Inversely proportional to The coefficient of the three-dimensional spatial distance between nodes. The feature value extracted for node i The feature value extracted for node j The stiffness coefficient of the pipe wall material. For the strain between nodes, we can approximate it using the relative displacement change rate between adjacent nodes, which is the ratio of the difference in pseudo-displacement between adjacent nodes to the distance between nodes, as shown in the following formula:

[0114] ;

[0115] The first part of the loss function is used to constrain the stress balance between adjacent nodes to prevent stress distortion. The greater the distance, the weaker the constraint becomes. The second part is based on Hooke's law to constrain the constitutive relation of the material, so that the law learned by the model conforms to the physical relationship.

[0116] The graph neural network is trained by minimizing the physical constraint loss function to output a stress feature vector, which is a vector that includes data features and conforms to the laws of pipe mechanics.

[0117] Training a graph neural network requires input data and labels. The input data includes the coordinates of each node, initial pseudo-displacement values, and initial features mapped to that node from the denoised ACSM signal. Node coordinates are provided by the positioning system of the pipe detector. Initial pseudo-displacement values ​​can be set to a zero vector or a small random quantity. Label data consists of the actual stress values ​​in the pipe, which can be obtained through finite element simulation to calculate the stress at the corresponding node location, or through actual measurements using strain gauges in the destructive testing calibration section.

[0118] During training, node features and adjacency matrices are input into the graph neural network. The network forward propagates and outputs the predicted stress feature vector for each node. The physical constraint loss function L2 is calculated. Here, the stiffness coefficient is the known elastic modulus of the pipe material, the distance is calculated based on the node coordinates, and the strain is calculated based on the pseudo-displacement of the node in the current iteration.

[0119] When training a graph neural network, the first and second constraints are used together to form the physical constraint loss function. The training process uses an optimization algorithm, such as stochastic gradient descent, to continuously adjust the weight parameters of the graph neural network to minimize this composite loss function. Specifically, the optimizer is set to Adam, and the learning rate, batch size, and total number of training epochs are defined. In each iteration, the gradient of the L2 loss with respect to the weight parameters of each layer of the graph neural network and the pseudo-displacements of the nodes is calculated, and these parameters are updated using the backpropagation algorithm.

[0120] The process of minimization is essentially about finding a set of model parameters that allow the stress feature vector output by the model to not only extract features well from the input data, but also to satisfy the two major physical laws of stress balance and material constitutive properties to the greatest extent possible.

[0121] By embedding stress balance relationships and material constitutive relations into the model training objectives, the problem of purely data-driven deep learning models potentially outputting physically unreasonable results can be solved.

[0122] After constructing the graph structure, the denoised first detection data is combined with node features and input into the graph neural network for computation. The network uses multi-layer graph convolution or graph attention operations, with each layer aggregating feature information from neighboring nodes to update the feature representation of the central node. After multiple iterations, each node fuses information from its local neighborhood to generate a high-dimensional numerical vector, i.e., the stress feature vector of that node. All the stress feature vectors of all these nodes constitute the first feature vector extracted from the first detection data.

[0123] The trained graph neural network can be applied to real-world data. The node coordinates obtained from actual pipeline inspection and the denoised ACSM signal are input into the network. The stress feature vector output by the network serves as the first feature vector, used for subsequent feature fusion and stress mapping.

[0124] S300: Based on the second feature extraction model, feature vector extraction is performed on the second detection data to obtain the second feature vector. The second feature extraction model is used to extract spatial and temporal features.

[0125] The second feature extraction model is used to extract features from the second detection data simultaneously in both spatial and temporal dimensions, and combine them into a second feature vector. That is, the second feature vector includes both spatial and temporal feature vectors. Spatially, the second feature extraction model processes the interrelationships between triaxial stress data at the same time point; temporally, it analyzes the dynamic patterns of each axial data point over time. By integrating spatial and temporal information, the second feature vector can more comprehensively characterize the state and evolution trend of pipeline stress.

[0126] The second feature extraction model is used to extract features from the second detection data. Specifically, it is designed for the RMF sensor. Since the RMF sensor records time-series triaxial pipe stress data, which includes both spatial and temporal features, models are set up at both the temporal and spatial scales to extract stress features. In some embodiments, the second feature extraction model includes a convolutional network model and a recurrent neural network model.

[0127] Convolutional neural network (CNN) models are used to process data with a grid-like topology. They extract spatial features by using a sliding window operation on the input data through convolutional kernels and employing local connectivity and weight sharing mechanisms. In one implementation, the model can be a one-dimensional CNN where the convolutional kernels move only in the time or channel dimension, suitable for processing sequential signals. In another implementation, the model can be a two-dimensional CNN where the convolutional kernels move in both dimensions, suitable for processing image-format data.

[0128] In this embodiment, the convolutional network model is a network model with M 1×3 convolutional kernels. The convolutional neural network performs convolution on the three-axis components of the second detection data to extract spatial dimension features and obtain a spatial feature vector. The three-axis components are stress-related signals in three directions synchronously acquired by the remanent magnetization stress sensor. These three directions are defined as the axial, circumferential, and radial directions of the pipe, which together constitute a three-dimensional observation of the stress state of the pipe surface.

[0129] Specifically, the stress data collected at each sampling time along the three axes are regarded as a multi-channel data point. These data points are organized into a two-dimensional matrix in chronological order, where the row direction represents the time series and the column direction represents the three axial channels. The convolutional network model uses multiple convolutional kernels with a width of 1 and a length of 3 to perform sliding convolution operations along the time dimension.

[0130] Each convolution kernel corresponds to three axial weights. By performing convolution operations with the three-axis data, local spatial features under different axial combinations are extracted. After the convolution result is transformed nonlinearly by the activation function, a set of feature maps are formed. These feature maps are further processed and compressed into a fixed-length numerical vector, namely the spatial feature vector, which encodes the stress distribution pattern in three-dimensional space.

[0131] like Figure 2 As shown, for example, M, such as 16 1×3 convolution kernels, are used to convolve the triaxial stress data at each sampling time, i.e. the axial stress data in the figure. After processing by the ReLU (Rectified Linear Unit) function, M columns of feature vectors are obtained, which are then spliced ​​along the channel to obtain the feature vectors calculated in space by RMF.

[0132] Recurrent Neural Networks (RNNs) have a recurrent connection structure used to process sequential data, enabling them to retain historical information and thus capture temporal dependencies in the data. In one implementation, the RNN can be a simple RNN where hidden layer states are passed through fully connected connections. In another implementation, the RNN can be a Long Short-Term Memory (LSTM) network, which uses gating mechanisms to control the forgetting and updating of information, mitigating the vanishing gradient problem.

[0133] The second detection data is processed by a recurrent neural network (RNN) with a gated, rotating cross-attention mechanism to extract features in the time dimension and obtain a time feature vector.

[0134] Specifically, the RNN simultaneously receives stress data sequences along three axes. The core of the network lies in its ability to dynamically adjust its internal information processing strategy based on changes in the stress field. When the pipeline stress field is relatively stable, the network primarily relies on historical information for each axis for prediction. When a significant distortion in the stress field is detected, the network activates an inter-axis information exchange mechanism, allowing the processing of one axis to reference the historical states of the other two axes. This adaptive mechanism is implemented through gating switches and alternating cross-attention modules, ultimately outputting a fixed-length numerical vector that integrates the dynamic temporal information of the three axes, i.e., a time feature vector.

[0135] On the time scale, for the stress data of each axis, an RNN is used to process the time series signal. However, since the stress data of the three axes are not independent, in this embodiment, the RNN is an RNN model with a gated switch and a rotating cross-attention mechanism, including an input layer, a hidden layer and an output layer.

[0136] In some embodiments, the second detection data is processed by a recurrent neural network to extract features in the time dimension, resulting in a time feature vector, including:

[0137] Calculate the principal stress direction and principal stress intensity of the triaxial components at each time step;

[0138] When the change in the principal stress direction exceeds a set angle, or the change in the principal stress intensity exceeds a set percentage, the rotating cross-attention mechanism is activated to generate the time feature vector of the current axis based on the feature vectors of the historical states of the other two axes.

[0139] When the change in the principal stress direction does not exceed the set angle, or the change in the principal stress intensity does not exceed the set percentage, the self-attention mechanism is activated to generate the time feature vector of the current axis based on the feature vector of the current axis's own historical state.

[0140] Specifically, such as Figure 3 As shown, three RNNs with single hidden layers are used to process triaxial stress data respectively. The RNNs are fully connected neural networks with feedback loops. Unlike isolated RNNs, the hidden layer output of each RNN, i.e. the "memory" of the current state, is connected to the hidden layer input of the other two RNNs. The historical state of the single-axis stress data is assigned to the processing of the other two axes. In order to balance the historical states between different axes, a rotating cross-attention mechanism with gated switches is introduced.

[0141] At each moment, a decision signal, namely the state of the gate switch, needs to be calculated. This can be based on an assessment of the stability of the current stress field. The assessment method is to calculate the principal stress direction and principal stress intensity of the three-axis components at each moment.

[0142] The principal stress direction is an angular parameter describing the location of the maximum normal stress in a stress state at a point. It can be obtained by principal component analysis of the triaxial remanent magnetization stress data at each moment. Specifically, the spatial azimuth angle of the eigenvector corresponding to the first principal component is used as the measure of the principal stress direction.

[0143] Principal stress intensity is a numerical parameter describing the magnitude of the maximum normal stress in a stress state at a point. It can be calculated by principal component analysis of the triaxial remanent magnetization stress data at each moment. Specifically, the eigenvalue corresponding to the first principal component is used as the measure of principal stress intensity.

[0144] The direction of the first principal component obtained from the analysis determines the principal stress direction angle, and the eigenvalue of the first principal component determines the principal stress intensity. By comparing whether the change in the principal stress direction angle at the current moment exceeds the set angle and whether the rate of change of the principal stress intensity exceeds the set percentage, a comprehensive judgment is made on whether the stress field has undergone significant distortion.

[0145] Principal stresses are calculated using Principal Component Analysis (PCA). PCA is performed on the triaxial RMF values ​​at each time step. The azimuth of the first eigenvector of the PCA analysis result is taken as the principal stress direction, and the first eigenvalue is taken as the principal stress intensity. On the time scale, if there is a significant change in the principal stress direction angle, for example, exceeding 10°, or a significant change in the principal stress intensity, for example, exceeding 10%, then the stress field is considered to have a relatively significant distortion, and the alternating cross-attention mechanism is used. Otherwise, the stress field is considered relatively stable, and the self-attention mechanism is used instead.

[0146] Taking an RNN that processes axial data as an example, this mechanism calculates the cross-attention between the hidden state of the axial unit and the hidden states of the circumferential and radial units at the previous time step. Specifically, the hidden state of the axial unit is used as the query vector, and attention is calculated with the hidden states of the circumferential and radial units as the key and value vectors, respectively, to generate two feature vectors that fuse cross-axis information. These two feature vectors are then concatenated and used as additional input for the axial unit at the next time step.

[0147] For cross-attention, when the gating switch is open (stress field distortion), a rotating cross-attention mechanism is used, and settings are configured. The matrix, with axial stress data, is output from the hidden layer of the previous cycle. For example, for the radial stress number in the hidden layer output of the previous cycle ,Will As an attention mechanism Vector, calculation of the subject Impact The eigenvectors are shown in the following formula:

[0148] ;

[0149] Similarly, the circumferential stress data is also calculated. Impact Feature vector The two feature vectors are then concatenated along the channels to form a feature matrix. This serves as the input for the next iteration of the axial RNN.

[0150] When the gate switch is closed, the stress field is relatively stable, and a self-attention mechanism is used for processing. Feature vector See the following formula:

[0151] ;

[0152] The input for the next loop is now a matrix. .

[0153] When the gating switch is closed, meaning that the changes in the principal stress direction and intensity are both within the threshold range, the stress field is considered relatively stable, and the self-attention mechanism is activated. Taking the axial element as an example, this mechanism uses the hidden state of the axial element itself as the query vector, key vector, and value vector simultaneously for self-attention calculation, which is equivalent to reweighting and enhancing its own historical features. The generated feature vectors are then concatenated and input into the next time step.

[0154] like Figure 4 As shown, the final output sequence signal of the entire RNN is the feature vector of the axial data on the time scale. Similarly, for circumferential stress data and radial stress data, the feature vector of that axis under the influence of the other two axes is calculated alternately. After the three-axis calculation is completed, the three feature vectors are concatenated along the channels to serve as the feature vector of the RMF sensor on the time scale. This vector dynamically integrates the independent features of each axis during the stable period with the collaborative features across axes during the distortion period.

[0155] S400: The first feature vector and the second feature vector are fused to obtain the fused feature, and the fused feature is mapped to the pipeline stress value through the stress mapping model.

[0156] After the feature vectors of the ACSM and RMF sensors are calculated, feature fusion is performed based on the sensor characteristics. Since the first and second feature vectors originate from sensors based on different physical principles, they characterize the stress state of the pipeline from different perspectives. Feature fusion is not a simple data splicing, but rather uses a specific algorithmic structure, such as a neural network incorporating an attention mechanism, to generate a more informative and robust fused feature that combines the advantages of both sensors.

[0157] In this embodiment, feature fusion adopts a hierarchical coding fusion network. The hierarchical coding fusion network adopts a hierarchical processing strategy, first encodes and compresses various types of input features independently, and then performs cross-fusion at a higher level. This structure is conducive to preserving the uniqueness of various features, while exploring the deep correlation between them.

[0158] like Figure 5 As shown, the hierarchical coding fusion network includes a self-attention encoder, a first feature compression network, a second feature compression network, and a cross-fusion block.

[0159] Among them, the self-attention encoder is based on the self-attention mechanism, which can calculate the correlation between each element in the input sequence and all elements in the sequence (including itself), and generate a new sequence representation by weighted summation, thereby capturing the long-term dependencies and contextual information within the sequence.

[0160] The first feature compression network is used to reduce and refine the features. It takes a high-dimensional feature vector as input and compresses it into a feature representation with higher information density and lower dimensionality through a specific network structure, while preserving as much key information as possible from the original features. The second feature compression network works in parallel with the first feature compression network and is responsible for compressing and encoding different types of input features, ensuring that different types of features can obtain appropriate optimized representations before fusion.

[0161] Cross-fusion blocks enable information interaction between different feature sources. Through attention mechanisms or other cross operations, feature vectors from different sources can query and respond to each other, thereby generating a unified feature representation that integrates information from multiple sources.

[0162] In some embodiments, feature fusion is performed between the first feature vector and the second feature vector to obtain fused features, including:

[0163] The first feature vector is input into the self-attention encoder to perform compression encoding, and the first encoded vector is output.

[0164] The spatial feature vector is input into the first feature compression network to perform compression encoding, and outputs a spatial encoded vector;

[0165] The time feature vector is input into the second feature compression network to perform compression encoding, and outputs a time-encoded vector;

[0166] Spatial and temporal encoding vectors are input into the cross-fusion block to obtain fused features.

[0167] In the first layer of encoding, various input features are fed into the corresponding encoders for processing. The sequence feature vectors extracted by ACSM are processed by an independent self-attention encoder. The encoded feature vector, i.e. the first encoded vector, is output by integrating contextual information. This vector not only retains the original feature information, but also strengthens the long-term dependencies within the sequence.

[0168] Meanwhile, for the spatial and temporal feature vectors extracted from the RMF, since spatial and temporal features are correlated both within and between groups, a graph neural network (GNN) is used for feature fusion and compression within groups, outputting a total feature vector that represents both the temporal and spatial features of the RMF, i.e., the fused vector. Between groups, a cross-attention block is used to fuse the temporal and spatial features of the RMF, resulting in the encoded total RMF feature vector.

[0169] In some embodiments, mapping fused features to pipe stress values ​​using a stress mapping model includes:

[0170] The fused features are input into the stress mapping model;

[0171] The fused features and the first encoding vector are input into a cross-attention mechanism to generate context-enhanced feature representations;

[0172] The context-enhanced feature representation is input into a fully connected layer, where linear transformations and dimensionality reductions are performed to output pipe stress values.

[0173] The stress mapping model is a computational model that converts fused features into pipeline stress values. The model receives fused features as input, performs a nonlinear mapping from a high-dimensional feature space to a one-dimensional stress value through its internal computational structure, and outputs a quantitative stress value.

[0174] See you again Figure 5 In some embodiments, the stress mapping model includes a cross-attention mechanism and a fully connected layer. The cross-attention mechanism can handle feature sequences from two different sources. This mechanism establishes the association weights between the elements of the two sequences by having the features of one sequence query the features of the other sequence, thereby generating an enhanced feature representation that integrates dual-source information.

[0175] A fully connected layer is a type of neural network layer in which each neuron is connected to all neurons in the layer above. Fully connected layers can integrate information from all input features and, through linear transformations combined with activation functions, achieve feature recombination and dimensional transformation.

[0176] In the cross-attention stage, the fused feature and the first encoding vector are simultaneously input into the cross-attention mechanism. This mechanism uses the fused feature as the query vector and the first encoding vector as the key and value vectors. By calculating the similarity between each element in the fused feature and all elements in the first encoding vector, a set of attention weights is obtained. These weights are then used to perform a weighted summation of the first encoding vector, generating a weighted context vector. This context vector is then superimposed on the original fused feature to form a context-enhanced feature representation. This process allows the fused feature to retain its own information while selectively absorbing relevant supplementary information from the first encoding vector.

[0177] After obtaining the context-enhanced feature representation, it is input into a fully connected layer. The fully connected layer consists of multiple linear layers connected in sequence. Each linear layer is usually followed by a non-linear activation function. These layers gradually perform linear transformations and non-linear processing on the input features, continuously combining and refining the effective information in the features. The number of neurons in the last fully connected layer is set to 1. The output is scaled appropriately to obtain the final pipe stress value.

[0178] For example, suppose the fused feature is a 96-dimensional vector and the first encoding vector is a 128-dimensional vector. In the cross-attention mechanism, the 96-dimensional fused feature is used as the query, and attention is calculated on the 128-dimensional first encoding vector to output a 96-dimensional context enhancement feature. This enhancement feature is then input into a three-layer fully connected network. The first layer maps the 96-dimensional input to 64 dimensions, the second layer maps the 64-dimensional input to 32 dimensions, and the third layer maps the 32-dimensional input to 1 dimension. Finally, a specific numerical value is output, such as 150.3 MPa, which is the predicted pipeline stress value.

[0179] This application combines the physical laws of sensors with deep learning methods, overcoming the limitations of existing methods in feature extraction. It uses deep learning to model the mapping relationship between sensor signals and pipeline stress values, thereby improving accuracy and reliability.

[0180] Based on the above-mentioned method for fusing stress sensor data from multiple oil and gas pipelines, this application provides a system for fusing stress sensor data from multiple oil and gas pipelines, comprising:

[0181] The acquisition module is used to acquire first detection data and second detection data. The first detection data is data acquired by the AC stress measurement sensor, and the second detection data is data acquired by the remanent magnetization stress sensor.

[0182] The feature extraction module is used to perform a first feature vector extraction on the first detection data based on a first feature extraction model to obtain a first feature vector, wherein the first feature extraction model is used to extract stress features based on geometric changes; and to perform feature vector extraction on the second detection data based on a second feature extraction model to obtain a second feature vector, wherein the second feature extraction model is used to extract spatial features and temporal features.

[0183] The mapping module is used to fuse the first feature vector and the second feature vector to obtain the fused feature, and then map the fused feature to the pipeline stress value through the stress mapping model.

[0184] 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 fusing stress sensor data from multiple oil and gas pipelines, characterized in that, include: Acquire first detection data and second detection data, wherein the first detection data is data collected by an AC stress measurement sensor and the second detection data is data collected by a remanent magnetization stress sensor; Based on the first feature extraction model, a first feature vector is extracted from the first detection data to obtain a first feature vector. The first feature extraction model is used to extract stress features based on geometric changes. Based on the first feature extraction model, a first feature vector is extracted from the first detection data to obtain a first feature vector. Prior to this, the process includes: A denoising model is constructed, comprising a first input channel and a second input channel, and obtained by training with a loss function. The construction of the denoising model includes: calculating the first norm difference between the denoised output signal and the real signal in the time domain, as a time-domain difference constraint; performing a short-time Fourier transform on the denoised output signal to obtain a first spectrum, and performing a short-time Fourier transform on the real signal to obtain a second spectrum; determining a narrowband frequency range with the center excitation frequency based on the excitation frequency of the AC stress measurement sensor; calculating the spectral component differences between the first spectrum and the second spectrum within the narrowband frequency range and applying a first weight; calculating the spectral component differences outside the narrowband frequency range and applying a second weight, using the norm of the weighted spectral differences as a frequency domain distribution constraint; adding the time-domain difference constraint and the frequency domain distribution constraint through a balance coefficient to obtain the loss function; minimizing the loss function to generate a denoising model, wherein the signal waveform output by the denoising model has a sinusoidal shape within a preset difference range. A sinusoidal signal of the same length as the first detection data is generated, and the frequency of the sinusoidal signal is the same as the excitation frequency of the AC stress measurement sensor. The first detection data is input to the first input channel, and the sine signal is input to the second input channel; The first detection data and the features of the sinusoidal signal are fused through an attention block to output the denoised first detection data. The attention block is set at the front end of the denoising model. Based on the second feature extraction model, feature vector extraction is performed on the second detection data to obtain a second feature vector. The second feature extraction model is used to extract spatial features and temporal features. The first feature vector and the second feature vector are fused to obtain fused features, and the fused features are mapped to pipeline stress values ​​through a stress mapping model.

2. The method for fusing stress sensor data from multiple oil and gas pipelines according to claim 1, characterized in that, Based on the first feature extraction model, a first feature vector is extracted from the first detection data to obtain a first feature vector, including: A graph neural network is constructed, wherein the graph neural network includes nodes and edges, the nodes correspond to the data acquisition points of the AC stress measurement sensor, the features of the nodes include spatial coordinates and pseudo-displacement, the edges are formed by constructing connections between the nodes through the spatial proximity relationship or time series proximity relationship of the data acquisition points, and the graph neural network is obtained after training with a physical constraint loss function; The denoised first detection data is input into the graph neural network to calculate the stress feature vector of each node through the graph neural network; The first feature vector is formed based on the stress feature vectors of all nodes.

3. The method for fusing stress sensor data from multiple oil and gas pipelines according to claim 2, characterized in that, The physical constraints include a first constraint and a second constraint; the first constraint is used to constrain the balance relationship of stress eigenvectors between adjacent nodes, and the second constraint is used to constrain the constitutive relationship between the stress eigenvectors and the strain calculated based on the pseudo-displacement. Constructing a graph neural network includes: The graph neural network is trained by minimizing the physical constraint loss function to output a stress feature vector, which is a vector that includes data features and conforms to the laws of pipeline mechanics.

4. The method for fusing stress sensor data from multiple oil and gas pipelines according to claim 1, characterized in that, The second feature extraction model includes a convolutional network model and a recurrent neural network; the second feature vector includes a spatial feature vector and a temporal feature vector. Based on the second feature extraction model, feature vector extraction is performed on the second detection data to obtain a second feature vector, including: The three-axis components of the second detection data are convolved by a convolutional neural network to extract spatial features and obtain a spatial feature vector. The second detection data is processed by a recurrent neural network with a gated switch and a rotating cross-attention mechanism to extract features in the time dimension and obtain a time feature vector.

5. The method for fusing stress sensor data from multiple oil and gas pipelines according to claim 4, characterized in that, The time feature vector is obtained, including: Calculate the principal stress direction and principal stress intensity of the triaxial components at each time step; When the change in the principal stress direction exceeds a set angle, or the change in the principal stress intensity exceeds a set percentage, the rotating cross-attention mechanism is activated to generate the time feature vector of the current axis based on the feature vectors of the historical states of the other two axes. When the change in the principal stress direction does not exceed a set angle, or the change in the principal stress intensity does not exceed a set percentage, the self-attention mechanism is activated to generate a time feature vector of the current axis based on the feature vector of the current axis's own historical state.

6. The method for fusing stress sensor data from multiple oil and gas pipelines according to claim 5, characterized in that, The first feature vector and the second feature vector are fused to obtain fused features, including: Feature fusion is performed through a hierarchical coding fusion network, which includes a self-attention encoder, a first feature compression network, a second feature compression network, and a cross-fusion block. The first feature vector is input into the self-attention encoder to perform compression encoding, so as to output the first encoded vector; The spatial feature vector is input into the first feature compression network to perform compression encoding, so as to output a spatial encoded vector; The time feature vector is input into the second feature compression network to perform compression encoding, and a time encoded vector is output. The spatial encoding vector and the temporal encoding vector are input into the cross-fusion block to obtain the fused features.

7. The method for fusing stress sensor data from multiple oil and gas pipelines according to claim 6, characterized in that, The fused features are mapped to pipe stress values ​​using a stress mapping model, including: The fused features are input into the stress mapping model, which includes a cross-attention mechanism and a fully connected layer. The fused features and the first encoding vector are input into the cross-attention mechanism to generate context-enhanced feature representations; The context-enhanced feature representation is input into the fully connected layer, where linear transformation and dimensionality reduction are performed to output the pipe stress value.

8. A data fusion system for stress sensors in multiple oil and gas pipelines, characterized in that, The method for performing stress sensor data fusion of multiple oil and gas pipelines according to any one of claims 1-7 includes: The acquisition module is used to acquire first detection data and second detection data. The first detection data is data acquired by an AC stress measurement sensor, and the second detection data is data acquired by a remanent magnetization stress sensor. The feature extraction module is used to perform a first feature vector extraction on the first detection data based on a first feature extraction model to obtain a first feature vector. The first feature extraction model is used to extract stress features based on geometric changes. Based on the first feature extraction model, a first feature vector is extracted from the first detection data to obtain a first feature vector. Prior to this, the process includes: A denoising model is constructed, comprising a first input channel and a second input channel, and obtained by training with a loss function. The construction of the denoising model includes: calculating the first norm difference between the denoised output signal and the real signal in the time domain, as a time-domain difference constraint; performing a short-time Fourier transform on the denoised output signal to obtain a first spectrum, and performing a short-time Fourier transform on the real signal to obtain a second spectrum; determining a narrowband frequency range with the center excitation frequency based on the excitation frequency of the AC stress measurement sensor; calculating the spectral component differences between the first spectrum and the second spectrum within the narrowband frequency range and applying a first weight; calculating the spectral component differences outside the narrowband frequency range and applying a second weight, using the norm of the weighted spectral differences as a frequency domain distribution constraint; adding the time-domain difference constraint and the frequency domain distribution constraint through a balance coefficient to obtain the loss function; minimizing the loss function to generate a denoising model, wherein the signal waveform output by the denoising model has a sinusoidal shape within a preset difference range. A sinusoidal signal of the same length as the first detection data is generated, and the frequency of the sinusoidal signal is the same as the excitation frequency of the AC stress measurement sensor. The first detection data is input to the first input channel, and the sine signal is input to the second input channel; The first detection data and the features of the sinusoidal signal are fused through an attention block to output the denoised first detection data. The attention block is set at the front end of the denoising model. The feature extraction module is also used to perform feature vector extraction on the second detection data based on the second feature extraction model to obtain a second feature vector. The second feature extraction model is used to extract spatial features and temporal features. The mapping module is used to fuse the first feature vector and the second feature vector to obtain fused features, and to map the fused features to pipeline stress values ​​through a stress mapping model.

Images