A method and system for identifying and locating defects in large-diameter pipeline spiral welds
Patent Information
- Application Number
- CN202611151552.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-08-28
AI Technical Summary
[0005]本发明所要解决的技术问题是现有管道焊缝缺陷识别与定位的精准度低
本发明基于构建的柱坐标系,通过激光测振仪检测装置采用双模式协同激励与感知采集模式,即激励点位固定多激光点位切换和激光点位固定多激励点位切换的采集模式,获得采集数据,提高对缺陷的辨识度,实现微缺陷特征的全面捕捉;从采集数据中提取几何、时域、时频域共12项核心特征,并对经标准化处理后的特征向量,通过哈达玛积实现特征与空间位置信息的绑定编码,得到有位置标识的输入向量,解决了传统方法中特征与位置信息分离处理导致的定位精度不足的瓶颈;液态神经网络的液态层通过动态神经元状态演化捕捉特征的空间-时序关联,再通过激活函数、特征聚合和Softmax函数,获得缺陷分类与缺陷点三维坐标,实现焊缝缺陷分类及定位的映射;通过引入QRFS算法对LNNs的高维参量(权重、偏置、神经元时间常数等参数)进行全局搜索与局部精调,并结合多任务损失函数与AdamW优化器,确保缺陷分类与定位的精度,解决现有模型优化过程易陷入局部最优的问题;从而实现了不同缺陷的识别,定位误差≤1mm,解决了现有技术对大孔径管道螺旋焊缝缺陷识别与定位的精准度低的问题。
Smart Images

Figure CN122651592A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nondestructive testing technology, and specifically discloses a method and system for defect identification and location of spiral welds in large-diameter pipes. Background Technology
[0002] In the energy sector, including long-distance pipelines for oil, chemicals, and natural gas, large-diameter pipelines are crucial structures for transporting high-temperature, high-pressure, and corrosive media. The quality of the spiral welds in these pipelines directly impacts pipeline safety; therefore, accurate identification and location of defects in the spiral welds are essential for ensuring safe pipeline operation. During the manufacturing of large-diameter pipelines, several issues can arise: insufficient heat input, excessively high welding speed leading to incomplete fusion between the weld metal and the base metal; damp flux, insufficient flux coverage, and excessively long arcs resulting in micropores within the weld; inadequate interpass cleaning, uneven flux particle size, and excessively high welding speeds preventing slag from bubbling out, leading to slag inclusions; and high hardening tendency of high-strength steel, insufficient preheating / interpass temperature, and high restraint stress causing hidden microcracks in the weld or heat-affected zone.
[0003] Currently, commonly used non-destructive testing methods include ultrasonic testing, X-ray testing, and eddy current testing. However, ultrasonic sensors are difficult to fit snugly against the curved surface of a pipe; X-ray testing equipment is expensive, poses radiation hazards, is insensitive to cracks parallel to the X-ray direction, and is unsuitable for online inspection; eddy current testing can only detect surface / near-surface defects and is insensitive to deep defects; furthermore, spiral welds have a periodic helical distribution, meaning defects can appear at any location within the weld, resulting in low efficiency for point-by-point scanning.
[0004] Pipeline weld defect detection often faces complex environments such as high temperature, humidity, oil contamination, and electromagnetic interference. When extracting defect features, traditional machine learning or static neural network models are often relied upon, which are difficult to effectively handle non-stationary vibration signals, cannot fully explore the spatiotemporal correlation features of defects, and ignore the intrinsic relationship between spatial location information and defect features, resulting in insufficient accuracy in defect identification and localization. At the same time, traditional neural networks are prone to getting stuck in local optima and slow convergence when optimizing key high-dimensional parameters such as weights, biases, and time constants, which limits the performance of the model in classifying and locating defects. Summary of the Invention
[0005] The technical problem to be solved by this invention is the low accuracy of existing pipeline weld defect identification and positioning.
[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A method for defect identification and location of spiral welds in large-diameter pipes includes: Based on the constructed cylindrical coordinate system, the vibration signal, excitation point cylindrical coordinates, and laser point cylindrical coordinates are obtained by using a dual-mode collaborative excitation and sensing acquisition mode through a laser vibration meter detection device. Based on the vibration signal, excitation point column coordinates, and laser point column coordinates, core features are extracted and feature vectors are generated. Based on the feature vector, spatial encoding is performed through the Hadamard product to generate an input vector with location identifiers; Based on the input vector with location identifiers, a well-trained liquid neural network is used to identify and locate defects, resulting in defect classification and three-dimensional coordinates of defect points.
[0007] Furthermore, the dual-mode collaborative excitation and sensing acquisition mode includes: The location of the excitation point is fixed, and the detection position of the laser point is switched. The detection position of the laser point is fixed, while the position of the excitation point is changed.
[0008] Furthermore, based on the vibration signal, the coordinates of the excitation point column, and the coordinates of the laser point column, core features are extracted to generate a feature vector, including: The core features include geometric features, temporal features, and time-frequency features. Geometric features include the spatial distance between the excitation point and the laser point, the circumferential angle difference, the axial distance difference, and the circumferential offset of the laser point relative to the center line of the spiral weld, as well as the regional factor. Temporal features include statistical peak value, mean, variance, and kurtosis. Time-frequency features include power spectral density peak value, dominant frequency, wavelet packet energy entropy, and time-frequency mean. Geometric features are obtained based on the coordinates of the excitation point column and the laser point column; Based on the vibration signal, the time-domain characteristics are obtained; Based on the vibration signal, time-frequency domain characteristics are obtained through time-frequency analysis; After standardizing the core features, feature vectors are generated.
[0009] Furthermore, the formula for calculating spatial coding is as follows: , in, Given a position-labeled input vector, For element-wise products, For spatial location normalized vectors, Geometric eigenvectors For time-domain feature vectors, This is the time-frequency domain feature vector.
[0010] Furthermore, based on the location-labeled input vector, a pre-trained liquid neural network is used for defect identification and localization, resulting in defect classification and three-dimensional coordinates of the defect points, including: Based on the location-labeled input vector, spatial-temporal correlations are captured through neuron state evolution; Based on spatial-temporal correlation, global aggregate features are extracted; Based on global aggregation features, defect classification and three-dimensional coordinates of defect points are obtained.
[0011] Furthermore, based on spatial-temporal correlations, global aggregated features are extracted, including: Based on spatial-temporal correlation, spatial-temporal coupled features are generated through activation functions; All spatial-temporal coupled features are combined to generate a liquid layer feature matrix; Based on the liquid layer feature matrix, the features are aggregated according to the time step sequence to extract global aggregated features.
[0012] Furthermore, based on the global aggregated features, the defect classification and the three-dimensional coordinates of the defect points are obtained, including: Based on the global aggregated features, the probability distribution of crack levels is obtained through the Softmax function; Defect classification is obtained based on the probability distribution of crack levels; Based on global aggregation features, the coordinates of defect points are output through a fully connected layer.
[0013] Furthermore, the optimized training of the liquid neural network includes: The liquid neural network is globally searched and locally fine-tuned using a quasi-random fractal optimization algorithm, and then trained using a multi-task loss function and the AdamW optimizer to obtain an optimized liquid neural network.
[0014] Furthermore, the multi-task loss function is: , , in, For multi-task loss function, The classification loss is cross-entropy. To regress the loss, This is a genuine defect label. This is the loss balance coefficient. This represents the probability distribution of crack grades. To smooth the L1 loss function, The coordinates of the defect points are output by the liquid neural network. These are the actual coordinates of the defect points.
[0015] This invention also relates to a defect identification and location system for spiral welds in large-diameter pipes, used in the aforementioned defect identification and location method for spiral welds in large-diameter pipes, comprising: The data acquisition module is used to acquire vibration signals, excitation point cylindrical coordinates, and laser point cylindrical coordinates based on the constructed cylindrical coordinate system and through the laser vibration meter detection device using a dual-mode collaborative excitation and sensing acquisition mode. The feature vector generation module is used to extract core features and generate feature vectors based on vibration signals, excitation point column coordinates, and laser point column coordinates. The spatial encoding module is used to perform spatial encoding based on feature vectors through the Hadamard product, generating input vectors with location identifiers; The defect identification and localization module is used to identify and locate defects based on the input vector with location identifiers, through an optimized and trained liquid neural network, to obtain defect classification and three-dimensional coordinates of defect points.
[0016] The beneficial effects of this invention are: This invention, based on a constructed cylindrical coordinate system, employs a dual-mode collaborative excitation and sensing acquisition mode using a laser vibrometer detection device. This mode combines fixed excitation points with multiple laser point switching and fixed laser point with multiple excitation point switching to acquire data, improving defect identification and achieving comprehensive capture of micro-defect features. Twelve core features (geometric, temporal, and time-frequency domains) are extracted from the acquired data. The standardized feature vectors are then bound to spatial location information via Hadamard product, resulting in a location-identified input vector. This overcomes the bottleneck of insufficient positioning accuracy caused by the separation of feature and location information in traditional methods. The liquid layer of the liquid neural network utilizes dynamic neural... Meta-state evolution captures the spatial-temporal correlation of features, and then through activation functions, feature aggregation, and the Softmax function, it obtains the defect classification and three-dimensional coordinates of defect points, realizing the mapping of weld defect classification and localization. By introducing the QRFS algorithm to perform global search and local fine-tuning of the high-dimensional parameters (weights, biases, neuron time constants, etc.) of LNNs, and combining multi-task loss functions and the AdamW optimizer, the accuracy of defect classification and localization is ensured, solving the problem that existing models are prone to getting trapped in local optima during the optimization process. Thus, it achieves the identification of different defects with a localization error of ≤1mm, solving the problem of low accuracy in the identification and localization of spiral weld defects in large-diameter pipelines in existing technologies. Attached Figure Description
[0017] Figure 1 This is a flowchart of a method for defect identification and location of spiral weld seams in large-diameter pipes according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the column coordinate system in an embodiment of the present invention; Figure 3 This is a block diagram of the laser vibration meter detection device in an embodiment of the present invention; Figure 4This is a block diagram of a defect identification and location system for spiral welds in large-diameter pipes according to an embodiment of the present invention. Figure 5 This is a schematic diagram of setting the excitation point and laser point in Embodiment 1 of the present invention.
[0018] The attached diagram lists the components represented by each number as follows: 1. Spiral weld; 2. Cross-section. Detailed Implementation
[0019] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0020] A method for defect identification and location in spiral welds of large-diameter pipes, such as... Figure 1 As shown, it includes the following steps: S1. Based on the constructed cylindrical coordinate system, the vibration signal, excitation point cylindrical coordinates, and laser point cylindrical coordinates are obtained by using a dual-mode collaborative excitation and sensing acquisition mode through a laser vibration meter detection device. like Figure 2 As shown, a cylindrical coordinate system is established for the pipeline under test, with the pipeline axis as the z-axis and the circumferential direction of the pipeline cross-section as the z-axis. Direction, i.e., circumferential angle is The radial direction is r; multiple excitation points and laser points are set. Based on the cylindrical coordinate system, excitation points, and laser points, a dual-mode collaborative excitation and sensing acquisition mode is adopted through a laser vibration meter detection device to collect vibration signals, excitation point cylindrical coordinates, and laser point cylindrical coordinates.
[0021] like Figure 3 As shown, the laser vibration meter detection device includes a laser vibration meter, a modal force hammer, a data acquisition unit, a control unit, and a signal data platform. The control unit is connected to the signal data platform and the data acquisition unit, and the data acquisition unit is connected to the laser vibration meter and the modal force hammer.
[0022] After fixing the test specimen (i.e., the pipe to be tested) with the defect to be tested, adjust the appropriate distance between the test specimen and the laser vibrometer. Set the laser intensity, sensitivity, laser focal length, and other parameters on the operation interface of the signal data platform. Use the laser vibrometer to emit a laser signal (i.e., an excitation signal) and use a modal force hammer to strike the test specimen. Then, use the laser vibrometer to receive the interference light (i.e., laser) reflected back from the test specimen due to vibration to obtain the vibration signal. The data acquisition unit converts the vibration signal from an optical signal into an electrical signal. After the control unit performs preliminary processing on the electrical signal, the vibration signal is displayed on the signal data platform. The vibration signal includes an excitation signal and a response signal. The excitation signal is the data of the striking force and the duration of the modal force hammer on the test specimen. The response signal is the data of the vibration displacement and the vibration time.
[0023] The dual-mode collaborative excitation and sensing acquisition mode includes: Acquisition mode 1, where the position of the excitation point is fixed, the detection position of the laser point is switched, and multiple sets of data are acquired; Acquisition mode 2, where the detection position of the laser point is fixed, the position of the excitation point is changed, and multiple sets of data are acquired.
[0024] This invention utilizes a dual-mode collaborative excitation and sensing acquisition mode to fully cover the pipeline vibration propagation path and defect-sensitive areas, thereby acquiring more comprehensive defect features and avoiding the problems of missing defect features and low identification rate of minor defects in single acquisition methods.
[0025] S2. Extract core features and generate feature vectors based on vibration signals, excitation point column coordinates, and laser point column coordinates; The core features include geometric features, temporal features, and time-frequency features. Geometric features include the spatial distance d between the excitation point and the laser point, and the circumferential angle difference. axial distance difference And the circumferential offset of the laser spot relative to the centerline of the spiral weld, area factor Time-domain characteristics include statistical peaks mean ,variance , cliff Time-frequency domain characteristics include power spectral density peak value. , main frequency Wavelet packet energy entropy Time-frequency domain mean It has a total of 12 core features.
[0026] Geometric features are used to reflect the changes in vibration propagation path caused by defects, time-domain features are used to reflect the influence of defects on vibration amplitude and dispersion, and time-frequency domain features are used to reflect the influence of defects on natural frequency and energy distribution. Defects at different locations will cause regular differences in geometric features, time-domain features, and time-frequency domain features under the laser point-excitation point combination.
[0027] S21. Obtain the geometric features based on the coordinates of the excitation point column and the laser point column; The excitation point cylindrical coordinates are The coordinates of the laser point column are ,in, (The outer radius of the pipe under test, the excitation signal and the laser are all applied to the outer surface of the pipe under test).
[0028] Since the wall thickness of the pipe under test is much smaller than its diameter, the radial coordinate r is approximately a constant R (the outer radius of the pipe under test). Therefore, defect location mainly focuses on the circumferential angle. The axial position z and radial depth r are output as the third dimension.
[0029] The formula for calculating the spatial distance between the excitation point and the laser point is: , The formula for calculating the circumferential angle difference between the excitation point and the laser point is: , The formula for calculating the difference in axial distance between the excitation point and the laser point is as follows: , The formula for calculating the circumferential offset of the laser point relative to the centerline of the spiral weld is as follows: , in, The circumferential angle of the excitation point, The circumferential angle of the laser point. The circumferential angle of the spiral weld centerline; The chord length corresponding to the arc length is approximated; for simplification, the arc length can be used. approximate.
[0030] S22. Based on the vibration signal, obtain the time-domain characteristics; Based on vibration signals Statistical peak mean ,variance , cliff Four parameters. When defects are present, they reduce structural stiffness, causing changes in peak values. The larger the defect, the more significant the peak value attenuation. Defects also cause a shift in the signal mean value, leading to increased signal dispersion, increased vibration uncertainty, increased variance, significantly increased kurtosis, and enhanced impact response.
[0031] The formulas for calculating the statistical peak, mean, variance, and kurtosis are as follows: , , , , Where T is the oscillation period.
[0032] S23. Based on the vibration signal, obtain the time-frequency domain characteristics through time-frequency analysis; Since vibration signals are non-stationary signals, time-frequency characteristics are obtained by extracting parameters that reflect the frequency distribution characteristics of the signal through time-frequency analysis; defects will change the natural frequency of the pipe under test, and time-frequency characteristics can reflect the frequency changes.
[0033] The formulas for calculating the peak power spectral density, dominant frequency, wavelet packet energy entropy, and time-frequency domain mean are as follows: , , , , , in, For wavelet packet decomposition, the first k Each one carries a signal; To normalize energy, defects increase the disorder of energy distribution, and the entropy value increases. For Fast Fourier Transform; F is the number of frequency points, reflecting the average amplitude of the time-frequency matrix, and defects will cause the mean to decrease; STFT is Short-Time Fourier Transform; T is the vibration period; M is the number of sub-bands; For discrete frequency indexes.
[0034] S24. After standardizing the core features, generate feature vectors.
[0035] The formula for standardization is: , , , in, For the k-th standardized core feature, For the k-th core feature before standardization, Let the mean of the k-th core feature in the training set be . Let the standard deviation be the k-th core feature of the training set. For the k-th core feature value of the i-th sample in the training set, The total number of samples in the training set. .
[0036] The 12 standardized core features are fused to generate a feature vector; the feature vector includes a geometric feature vector, a time-domain feature vector, and a time-frequency domain feature vector. The geometric eigenvectors are The time-domain feature vector is The time-frequency domain eigenvector is .
[0037] S3. Based on the feature vector, spatial encoding is performed through the Hadamard product to generate an input vector with positional identifiers; To enable the liquid neural network to perceive the spatial location corresponding to the features, the geometric feature vector, time-domain feature vector, time-frequency domain feature vector, and excitation point cylindrical coordinates and laser point cylindrical coordinates are spatially encoded using the Hadamard product. This ensures that each type of feature contains spatial information of the acquisition point, providing feature basis for subsequent defect localization.
[0038] The formula for calculating spatial coding is: , , in, is the input vector with location identifier; t is the time step, corresponding to the sampling time of the vibration signal; The element-wise product, also known as the Hadamard product, binds features to locations. For spatial location normalized vectors, ; The maximum distance between the laser point and the excitation point. The axial length of the pipe to be measured is given.
[0039] This invention couples and encodes feature vectors with spatial location information, enabling the input vector of a liquid neural network to contain coupled information of defect-sensitive features and acquisition point coordinates. This allows the liquid neural network to simultaneously perceive feature information and spatial correlation, thus solving the positioning accuracy bottleneck caused by the separation of feature and location information processing in traditional methods.
[0040] S4. Based on the input vector with location identifiers, defect identification and localization are performed through the optimized and trained liquid neural network to obtain defect classification and three-dimensional coordinates of defect points.
[0041] Traditional defect identification and localization typically employs Recurrent Neural Networks (RNNs) or Long Short-Term Memory Networks (LSTMs). The weights and structures of RNNs and LSTMs are static. RNNs focus only on temporal correlations and suffer from severe gradient vanishing problems. LSTMs memorize temporal information through cell states, requiring three gating units to mitigate the gradient vanishing problem, resulting in complex structures and increased computational load. Furthermore, neither RNNs nor LSTMs directly integrate spatial location information. The memory mechanisms of RNNs and LSTMs are purely mathematical modeling, which is not directly related to vibration changes caused by weld defects, and they have low sensitivity to micro-defects.
[0042] The liquid neural network (LNNs) used in this application can dynamically adjust the neuron weights and neuron time constants with time steps, which can adaptively capture the non-stationary characteristics of vibration signals and capture the coupling relationship between temporal changes and spatial positions. The neuron state evolution of LNNs can simulate the propagation and attenuation process of vibration signals, which fits the temporal dependence of vibration displacement changes.
[0043] A liquid neural network consists of an input layer, a liquid layer, and an output layer. The liquid layer contains multiple neurons, and the output layer is a fully connected layer.
[0044] The parameters of the liquid neural network are optimized using a quasi-random fractal optimization algorithm (QRFS algorithm). This involves global search and local fine-tuning of the high-dimensional parameter vector of the liquid neural network using QRFS. The liquid neural network is then trained using a multi-task loss function and the AdamW optimizer, resulting in an optimized and trained liquid neural network. Specifically, this includes: (1) The multi-task loss function is: , , in, This is the multi-task loss function, i.e., the total loss function; Cross-entropy is used for classification loss to optimize the accuracy of different defect recognition. The regression loss is used to optimize the 3D localization accuracy of the defect; Labels for genuine defects; This is the loss balance coefficient; This represents the probability distribution of crack grades. For smoothing L1 loss function; The coordinates of the defect point output by the liquid neural network; These are the actual coordinates of the defect points.
[0045] (2) The training of the liquid neural network includes: dividing the collected labeled data into training set, validation set, and test set according to a preset ratio; and using the AdamW optimizer. In the middle, set the learning rate and weight decay The initial value is then determined, and the parameters are subsequently optimized by the QRFS algorithm to determine the convergence condition for the number of iteration rounds.
[0046] (3) Optimization of the liquid neural network includes: The entire optimization process is implemented on the MATLAB platform. The QRFS algorithm is linked with LNN training to calculate the LNN loss corresponding to the parameters in real time and iteratively update the parameter population. During the optimization process, the optimal parameters of each round need to be saved to facilitate subsequent LNN reproduction and performance analysis.
[0047] 1) Define the set of parameters to be optimized and their value range. Define a high-dimensional parameter vector (i.e., the set of parameters to be optimized), the learning rate of AdamW. Weight decay The set of parameters to be optimized in the QRFS algorithm is as follows: , While searching for the core parameters of LNNs, the QRFS algorithm simultaneously searches for the optimal hyperparameters (learning rate and weight decay, etc.) of AdamW, ensuring that AdamW's fine-tuning capability precisely matches the dynamic characteristics of LNNs and the global optimal solution of the QRFS algorithm. The optimized formula for AdamW is as follows: , in, This is the corrected high-dimensional parameter vector. The initial values of the parameters provided for the current round of the ORFS algorithm. The high-dimensional parameter vector before correction. To prevent the numerical stability constant from being divided by zero, The gradient vector of the total loss. The learning rate for AdamW optimized for the QRFS algorithm. Weight decay for AdamW optimized for the QRFS algorithm.
[0048] 2) Set the key parameters and value range of the QRFS algorithm. Key parameters of the QRFS algorithm include population size N and maximum number of iterations. Local step size initial value Quasi-random sequence basis and sensitivity coefficient .
[0049] 3) Initialize LNN parameters using Halton sequences. Dynamic weights Initialization: A customized Halton sequence is used to initialize the population to avoid a disconnect between parameters and physical processes. Based on feature importance, differential initial values are set for geometric features (j∈[1,4]), temporal features (j∈[5,8]), and time-frequency features (j∈[9,12]) to reflect the sensitivity of features to defects. Initial amplitude coefficients for feature contributions are set to 0.8, 0.6, and 0.4, respectively, for dynamic weights. Initialization; the dynamic weight initialization formula is: , in, The dynamic weights from the initial input layer to the liquid layer; The d-th quasi-random number generated for the Halton sequence. ∈[0,1].
[0050] Neuron time constant Initialization: Based on the main frequency of the pipeline under test Regarding the time constant Initialization, ensure Matching the vibration period T; the neuron time constant initialization formula is: , in, Let i be the time constant of the i-th neuron in the 0th generation (initial); The dominant frequency is T; the oscillation period is T. ; As an attenuation factor, the time constant of the neuron is initialized to match the characteristics of the vibration signal, thus avoiding disconnection from the physical process.
[0051] Loss balance coefficient Initialization: Calculate the initial classification loss on the training set using MATLAB. Regression loss Dynamically balance the weights of dual tasks To ensure that the initial loss ratio is appropriate and to avoid training dominated by a single task.
[0052] Hyperparameter initialization: Set the initial learning rate for AdamW and initial weight decay Initial sensitivity coefficient , , .
[0053] 4) Introduce a fractal expansion factor to dynamically adjust the search intensity for global fractal search. Adapting the dynamic characteristics of LNNs through fractal global search.
[0054] During the global search phase, the QRFS algorithm introduces a fractal expansion factor F(t) to dynamically adjust the search intensity, ensuring uniform exploration of the high-dimensional parameter space. The fractal expansion factor F(t) is determined based on the dynamic weights of the LNNs. Adaptive adjustment is performed to make the search intensity follow the changes in feature contribution: , Where t is the number of iterations in the current global search; This represents the maximum number of iterations for the global search. F(t) is the sum of the absolute values of all elements in the 12×128 dimensional dynamic weight matrix, used to reflect the overall contribution level of the features. The higher the contribution, the larger F(t) is, and the more aggressive the search.
[0055] During global parameter updates, based on the Halton quasi-random perturbation term ∈[0,1], the driving parameters are explored globally; the global parameter update formula is: , in, Let d be the d-th dimension parameter of the n-th candidate solution after the t-th generation global search; Let d be the d-th dimension parameter of the global optimal solution in generation t-1; Let d be the d-th dimension parameter of the n-th candidate solution in the (t-1)-th generation; The quasi-random perturbation term generated for the Halton sequence ensures uniformity in the global search. ; is the sensitivity coefficient of the d-th dimension parameter, used to fully search for key parameters.
[0056] During population selection, the total model loss corresponding to each candidate parameter is calculated. The population diversity is maintained by retaining the first preset percentage of individuals with the least loss, and the rest are supplemented by newly generated Halton sequence individuals.
[0057] 5) Use MATLAB to calculate characteristic time series for local fine-tuning. Dynamic weights Fine-tuning: Differentiated local optimization is performed on the core parameters of LNNs. The standard deviation std(Xj(t)) of the feature time series is calculated using MATLAB to precisely optimize the feature contribution weights. The dynamic weight fine-tuning formula is as follows: , in, The dynamic weights from the input layer to the liquid layer after local fine-tuning in generation t; The dynamic weights from the input layer to the liquid layer after the t-th generation global search; The step size is used for local fine-tuning, and its size varies with the number of iterations and the time constant of the previous round; Let be the standard deviation of the time series of the j-th feature. The time series data for the j-th feature; Quasi-random perturbation term generated for Halton sequences.
[0058] Neuron time constant Fine-tuning: Numerical solution of neuron state-time derivatives using MATLAB This allows the update of the neuron's time constant to match the temporal changes of the vibration signal; the fine-tuning formula for the neuron's time constant is: , in, Let be the time constant of the i-th neuron after the t-th generation of local fine-tuning; The time derivative of the state of the i-th neuron; To output ±1, ensure that the fine-tuning direction of the neuron's time constant closely matches the trend of neuron state changes; Quasi-random perturbation term generated for Halton sequences.
[0059] Fine-tuning the neuron's time constant ensures that the direction of fine-tuning aligns with the neuron's state evolution trend.
[0060] Loss balance coefficient Fine-tuning: Combining the rate of change of dual-task losses , The balance between classification and localization is dynamically adjusted; the formula for fine-tuning the loss balance coefficient is: , in, The loss balance coefficient after local fine-tuning in generation t; The loss balance coefficient is the result of the global search in the t-th generation. The classification loss is the rate of change of cross-entropy; The rate of change of regression loss; Quasi-random perturbation term generated for Halton sequences.
[0061] Hyperparameter fine-tuning: The initial learning rate and weight decay of AdamW are updated simultaneously to achieve a good fit between the optimizer parameters and the LNN parameters; the update formula is: , , in, Let be the initial learning rate of AdamW after local fine-tuning in generation t; Let be the initial learning rate of AdamW after the t-th generation global search; The weight decay of AdamW after local fine-tuning in generation t; The weight decay of AdamW after the t-th generation global search; Quasi-random perturbation term generated for Halton sequences.
[0062] 6) Verify multi-dimensional convergence determination using MATLAB calculations. Convergence criteria are used to ensure that parameters are fully optimized.
[0063] The convergence conditions of multiple dimensions must be met simultaneously: QRFS algorithm convergence, LNN task performance meeting the target, and neuron state stability, to avoid parameter overfitting or insufficient optimization.
[0064] in, Let be the fitness corresponding to the globally optimal parameters in generation t; For the first The fitness corresponding to the globally optimal parameters; Let be the classification loss with optimal parameters in generation t; Let be the regression loss of the optimal parameters in generation t; denoted as the standard deviation of all neuron states.
[0065] By constructing a collaborative framework of dynamic characteristics of LNNs and QRFS optimization mechanism, this paper proposes a fractal quasi-random dual-mechanism adaptation optimization for LNNs, which addresses the high-dimensional non-convexity characteristics of dynamic neuron states, time-varying weights, and coupling parameters. The high-dimensional space uniformity of Halton sequences solves the problem of insufficient global exploration of LNN parameters, and the local fine-tuning characteristics of fractal perturbation adapt to the dynamic coupling relationship of LNN parameters. This makes up for the shortcomings of traditional optimization algorithms in convergence imbalance during parameter optimization, and achieves a synergistic improvement in the ability of parameter optimization and model spatiotemporal correlation capture.
[0066] This invention employs the Quasi-Random Fractal Optimization (QRFS) algorithm to perform global optimal search and local precise tuning of the high-dimensional parameter vector of the liquid neural network. It also combines a multi-task loss function and the AdamW optimizer to train the model, ensuring the accuracy of defect identification and localization, and ensuring the performance of LNNs with a defect identification rate of ≥95% and a localization error of ≤1mm.
[0067] S41. Based on the location-identified input vector, capture spatial-temporal correlations through neuron state evolution; The position-labeled input vector is input into the input layer of the liquid neural network to obtain a sequence of temporal feature vectors. .
[0068] The liquid layer captures spatial-temporal correlations from temporal feature vector sequences through neuronal state evolution.
[0069] The neuron state evolution equation is , in, Let represent the state of the i-th neuron at time step t, where i = 1, 2, ..., K; K is the number of neurons in the liquid layer, which can be 128. This is the neuron's time constant, used to regulate the neuron's sensitivity to spatial information; The dynamic weights from the input layer to the liquid layer are used to fuse the contributions of the three types of features; The input vector has a location identifier; The dynamic weights of neurons inside the liquid layer, with a constrained range. ; As an activation function, it can alleviate the vanishing gradient problem under low-frequency features; The neurons in the liquid layer are biased, with a constraint range of [-0.5, 0.5]. This is a time delay used to simulate the timing dependence of vibration signals; Let be the state of the k-th neuron at the time delay.
[0070] The formula for calculating the dynamic weights from the input layer to the liquid layer is: , in, The dynamic weights from the input layer to the liquid layer. As the initial weights, This is the weight gain coefficient. This is a regional factor representing the circumferential offset of the laser point relative to the centerline of the spiral weld. The area factor representing the circumferential offset of the standardized laser point position relative to the centerline of the spiral weld. For the standardized kurtosis, This is the standardized wavelet packet energy entropy.
[0071] S42. Extract global aggregate features based on spatial-temporal correlation; S421. Based on spatial-temporal correlation, spatial-temporal coupling features are generated through activation functions; Based on spatial-temporal correlation, the liquid layer maps neuron states through activation functions to generate spatial-temporal coupled features.
[0072] The spatial-temporal coupling characteristics are: .
[0073] S422. Combine all spatial-temporal coupling features to generate a liquid layer feature matrix; There are a total of 128 spatial-temporal coupling features, which are the outputs of 128 neurons.
[0074] The feature matrix of the liquid layer is: .
[0075] S423. Based on the liquid layer feature matrix, aggregate according to the time step sequence to extract global aggregate features.
[0076] Based on the feature matrix of the liquid layer, the states of all neurons in the liquid layer are arranged in time step sequence. Perform aggregation to extract global aggregated features across time.
[0077] The aggregation method is as follows: , Where S represents the global aggregated feature; The peak value of the neuron state over time is used to reflect signal mutations caused by defects; The time-domain valley value is used to capture vibration decay characteristics; The time-domain mean is used to characterize the overall shift of the vibration signal; The standard deviation in the time domain reflects the dispersion of the signal; after aggregation... This means compressing the time series information into global features, and then flattening it out to get a 512-dimensional feature vector, thus completing the compression of time series information into static features.
[0078] This invention couples the dynamic neuron state evolution of LNNs with the physical processes of vibration signal propagation, attenuation, and reflection in defective welded structures. It simulates vibration time delay changes by dynamically adjusting the neuron time constant and captures signal distortion caused by defects through the nonlinear characteristics of the activation function. This breaks through the limitation of traditional mathematical modeling being disconnected from physical processes and achieves adaptive spatiotemporal correlation capture of non-stationary vibration signals.
[0079] S43. Based on the global aggregation features, obtain the defect classification and the three-dimensional coordinates of the defect points.
[0080] S431. The output layer obtains the probability distribution of crack levels based on the global aggregated features and the Softmax function; based on the probability distribution of crack levels, the defect classification is obtained. Crack grades range from 0 to 4. Grade 0 corresponds to no defects, Grade 1 corresponds to non-fusion defects, Grade 2 corresponds to porosity defects, Grade 3 corresponds to slag inclusion defects, and Grade 4 corresponds to crack defects.
[0081] The formula for calculating the probability distribution of crack grade is: , in, This represents the probability distribution of crack levels; c = 0, 1, ..., 4 represents the crack level, i.e., the defect label. For classification weights, ; For classification bias, ; The classification weights are optimized using the QRFS algorithm; The classification bias optimized using the QRFS algorithm; To aggregate global features Flattened, it becomes a 512-dimensional vector.
[0082] S432. Based on the global aggregation features, output the coordinates of the defect points through the fully connected layer.
[0083] The three-dimensional coordinates of the defect point are , , , , in, To map the output to the [0,1] interval, ensure that the coordinates are within the range of the pipe to be tested; The circumferential angle of the defect point, with an output range of [0, 2π); Here are the axial coordinates of the defect point, and the range is... ; Radial depth of the defect point, range S represents the global aggregation feature. The axial length of the pipe to be measured is... The wall thickness of the pipe to be measured. This is the circumferential angle weight matrix. For circumferential angular offset, This is the axial weight matrix. For axial offset The radial weight matrix is... For radial offset.
[0084] By coupling the dynamic neuron state evolution in liquid neural networks (LNNs) with the physical processes of vibration signal propagation, attenuation, and reflection in defective welded structures, and by simulating vibration time delay changes through dynamic adjustment of neuron time constants, and capturing signal distortion caused by defects through the nonlinear characteristics of activation functions, this method overcomes the limitations of traditional mathematical modeling being disconnected from physical processes, and achieves adaptive spatiotemporal correlation capture of non-stationary vibration signals.
[0085] This invention also relates to a defect identification and location system for spiral welds in large-diameter pipes, such as... Figure 4 As shown, it includes: The data acquisition module is used to acquire vibration signals, excitation point cylindrical coordinates, and laser point cylindrical coordinates based on the constructed cylindrical coordinate system and through the laser vibration meter detection device using a dual-mode collaborative excitation and sensing acquisition mode. The feature vector generation module is used to extract core features and generate feature vectors based on vibration signals, excitation point column coordinates, and laser point column coordinates. The spatial encoding module is used to perform spatial encoding based on feature vectors through the Hadamard product, generating input vectors with location identifiers; The defect identification and localization module is used to identify and locate defects based on the input vector with location identifiers, through an optimized and trained liquid neural network, to obtain defect classification and three-dimensional coordinates of defect points.
[0086] Example 1 For example, consider a large-diameter pipe with a defect to be tested. The pipe is made of X90M high-strength steel, with an outer radius R = 600 mm and a wall thickness of [missing information]. =20mm, axial length =800mm, the weld seam has defects such as incomplete fusion, micropores, inclusions and microcracks during the manufacturing process.
[0087] The laser vibration meter testing device uses the Julight laser vibration meter, and the modal force measuring hammer uses the CT-LC-01 model. The horizontal distance between the pipe and the laser vibration meter should be adjusted to approximately 90cm.
[0088] Establish a cylindrical coordinate system for the pipeline (with the center of one end of the pipeline as the origin, the axis as the z-axis, and the circumferential angle as the z-axis). Counting from the weld start line r; as... Figure 5 As shown, excitation points and laser points are set on the surface of the pipe covering the spiral weld 1. There is one section every 45° along the circumference and three sections 2 are taken in the axial direction, for a total of 6×3=18 points.
[0089] To ensure the integrity of the excitation signal, the signal data platform was set with a trigger delay of -10ms, a trigger threshold of 0.5%, and a trigger hysteresis of 0.3%. The sampling frequency of the laser vibrometer was 50kHz-100kHz, the laser focal length was adjusted to 90-100cm, and the striking force of the modal force hammer was approximately 50-100N. Multiple sets of data (i.e., vibration signals) were obtained using a dual-mode collaborative excitation and sensing acquisition mode. Five sets of data were collected at each point, for a total of 90 sets of data collected in both acquisition modes.
[0090] Based on the excitation point and laser point, and using MATLAB to process the acquired data, 12 core features were analyzed in the geometric, time-domain, and time-frequency domains. In the time-frequency domain feature extraction of wavelet packet energy entropy, the wavelet basis function was db4, the decomposition level was 3, and the number of subbands was M=8; the short-time Fourier transform (STFT) window length was 256 points, the overlap rate was 50%, and the number of frequency points was F=1024.
[0091] Training the liquid neural network: The collected data was divided into training, validation, and test sets in a 7:2:1 ratio, and the parameters were initialized. The number of neurons in the liquid layer of the LNN was K=128, and the initial learning rate of the AdamW optimizer was set. Weight decay Loss balance coefficient The convergence condition is that the change in the total loss of the verification set is less than 1e-6 for 10 consecutive rounds or 200 rounds.
[0092] Define the dimensions and value ranges of the parameters to be optimized, ensuring that the parameters match the pipeline characteristics and the structure of the liquid neural network. Set the value ranges of the parameters to be optimized, as shown in Table 1: Table 1. Range of values for the parameters to be optimized The key parameters and their value ranges for the QRFS algorithm are set as shown in Table 2: Table 2 Key parameters and value ranges of the QRFS algorithm Neuron time constant Initialization: Combined with the main frequency of the pipeline (Vibration period T = 1 ms), ensuring that the neuron's time constant matches the vibration period: ,set up ∈[0.05ms, 0.2ms], The attenuation factor (approximately 0.6065) helps prevent the initial value from deviating from the actual physical properties.
[0093] Loss balance coefficient Initialization: Calculate the initial classification loss on the training set using MATLAB. Regression loss Dynamically balance the weights of the two tasks: To ensure that the initial loss ratio is appropriate and to avoid training dominated by a single task.
[0094] Hyperparameter initialization: Set the initial learning rate for AdamW. Initial weight decay Initial sensitivity coefficient =2.0, =1.5, =1.2.
[0095] The fractal global search phase iterates 1-60 times; during population selection, the top 50% of individuals with the least loss are retained.
[0096] In the multi-dimensional convergence determination stage, MATLAB calculations were used for verification during the iteration process. Classification loss Regression loss Standard deviation of neuron states If the convergence condition is not met after 100 iterations, the parameter with the minimum current loss is output as the optimal solution. The optimal parameters obtained through QRFS optimization are then used. Substitute the LNNs into the test set and then fine-tune them using the AdamW optimizer to ensure that the LNNs have a defect recognition rate of ≥95% and a positioning error of ≤1mm on the test set.
[0097] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for defect identification and location of spiral welds in large-diameter pipes, characterized in that, include: Based on the constructed cylindrical coordinate system, the vibration signal, excitation point cylindrical coordinates, and laser point cylindrical coordinates are obtained by using a dual-mode collaborative excitation and sensing acquisition mode through a laser vibration meter detection device. Based on the vibration signal, excitation point column coordinates, and laser point column coordinates, core features are extracted and feature vectors are generated. Based on the feature vector, spatial encoding is performed through the Hadamard product to generate an input vector with location identifiers; Based on the input vector with location identifiers, a well-trained liquid neural network is used to identify and locate defects, resulting in defect classification and three-dimensional coordinates of defect points.
2. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 1, characterized in that, The dual-mode collaborative excitation and sensing acquisition mode includes: The location of the excitation point is fixed, and the detection position of the laser point is switched. The detection position of the laser point is fixed, while the position of the excitation point is changed.
3. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 1, characterized in that, The process of extracting core features and generating feature vectors based on vibration signals, excitation point column coordinates, and laser point column coordinates includes: The core features include geometric features, time-domain features, and time-frequency domain features. The geometric features include the spatial distance between the excitation point and the laser point, the circumferential angle difference, the axial distance difference, and the circumferential offset of the laser point relative to the center line of the spiral weld as a region factor. The time-domain features include statistical peak value, mean, variance, and kurtosis. The time-frequency domain features include power spectral density peak value, dominant frequency, wavelet packet energy entropy, and time-frequency domain mean. Geometric features are obtained based on the coordinates of the excitation point column and the laser point column; Based on the vibration signal, the time-domain characteristics are obtained; Based on the vibration signal, time-frequency domain characteristics are obtained through time-frequency analysis; After standardizing the core features, feature vectors are generated.
4. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 1, characterized in that, The formula for calculating spatial coding is: , in, Given a position-labeled input vector, For element-wise products, For spatial location normalized vectors, Geometric eigenvectors For time-domain feature vectors, This is the time-frequency domain feature vector.
5. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 1, characterized in that, The process of identifying and locating defects using a pre-trained liquid neural network based on a location-identified input vector, to obtain defect classification and three-dimensional coordinates of defect points, includes: Based on the location-labeled input vector, spatial-temporal correlations are captured through neuron state evolution; Based on spatial-temporal correlation, global aggregate features are extracted; Based on global aggregation features, defect classification and three-dimensional coordinates of defect points are obtained.
6. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 5, characterized in that, The extraction of global aggregated features based on spatial-temporal correlation includes: Based on spatial-temporal correlation, spatial-temporal coupled features are generated through activation functions; All spatial-temporal coupled features are combined to generate a liquid layer feature matrix; Based on the liquid layer feature matrix, the features are aggregated according to the time step sequence to extract global aggregated features.
7. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 5, characterized in that, The step of obtaining defect classification and three-dimensional coordinates of defect points based on global aggregated features includes: Based on the global aggregated features, the probability distribution of crack levels is obtained through the Softmax function; based on the probability distribution of crack levels, defect classification is obtained. Based on global aggregation features, the coordinates of defect points are output through a fully connected layer.
8. The method for defect identification and location of spiral welds in large-diameter pipes according to claim 1, characterized in that, The optimization training of liquid neural networks includes: The liquid neural network is globally searched and locally fine-tuned using a quasi-random fractal optimization algorithm, and then trained using a multi-task loss function and the AdamW optimizer to obtain an optimized liquid neural network.
9. A method for defect identification and location of spiral welds in large-diameter pipes according to claim 8, characterized in that, The multi-task loss function is: , , in, For multi-task loss function, The classification loss is cross-entropy. To regress the loss, This is a genuine defect label. This is the loss balance coefficient. This represents the probability distribution of crack grades. To smooth the L1 loss function, The coordinates of the defect points are output by the liquid neural network. These are the actual coordinates of the defect points.
10. A defect identification and location system for spiral welds in large-diameter pipes, characterized in that, A method for defect identification and location of spiral welds in large-diameter pipes according to any one of claims 1-9 includes: The data acquisition module is used to acquire vibration signals, excitation point cylindrical coordinates, and laser point cylindrical coordinates based on the constructed cylindrical coordinate system and through the laser vibration meter detection device using a dual-mode collaborative excitation and sensing acquisition mode. The feature vector generation module is used to extract core features and generate feature vectors based on vibration signals, excitation point column coordinates, and laser point column coordinates. The spatial encoding module is used to perform spatial encoding based on feature vectors through the Hadamard product, generating input vectors with location identifiers; The defect identification and localization module is used to identify and locate defects based on the input vector with location identifiers, through an optimized and trained liquid neural network, to obtain defect classification and three-dimensional coordinates of defect points.
