Pipeline near-weld zone defect guided wave characteristic signal extraction method

Abstract

This invention relates to the field of non-destructive testing technology for pipelines, and discloses a method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline. The method involves acquiring raw guided wave signals through a sensor array deployed near the weld zone; mathematically modeling and actively separating the inherent weld structure echoes in the signals; locating potential defect responses in the purified signals using an iterative optimization approach; constructing a dynamic defect feature learning framework, using the separated weld model as a suppression reference, and autonomously learning and updating typical representation patterns of defect signals from continuous data streams; applying this framework to perform online scanning of real-time signals, identifying signal segments matching the learned patterns as candidate defect features; and integrating spatiotemporal information to achieve three-dimensional fusion localization and contour delineation of the defect source. This technology can suppress weld structure interference, extract weak defect signals, and possess adaptive identification and continuous optimization capabilities for defect features.

Description

Technical Field

[0001] This invention relates to the field of pipeline nondestructive testing technology, specifically a method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline. Background Technology

[0002] In the field of pipeline nondestructive testing, defect detection in the near-weld region has long been a technical challenge. Conventional guided wave testing techniques mainly rely on comparative analysis with baseline signals in a defect-free state, or use fixed thresholds and preset patterns to identify defect echoes. However, the pipeline weld itself is a geometric and material discontinuity, generating strong and stable structural echoes. When early, minute defects appear in the near-weld region, their weak signals are easily overwhelmed by the strong inherent weld echoes, resulting in an excessively low signal-to-noise ratio. Existing methods struggle to effectively extract the invariant weld structural response from complex mixed signals without requiring ideal baseline samples, thus severely limiting the sensitivity and accuracy of detecting minute defects.

[0003] Another limitation of existing technologies lies in their static defect identification patterns. Actual pipeline operating conditions are complex and varied, and the morphology and development process of defects are uncertain. Traditional signal processing and identification methods are typically based on fixed parameters or criteria, lacking self-adjustment and evolutionary capabilities. This makes it difficult for the system to adapt to individual differences in different pipeline welds, and it cannot effectively learn and track changes in the representation patterns of defect signals that may occur during long-term operation. A new signal processing method is needed that can actively suppress inherent structural interference and has the ability to adaptively learn defect features from continuous data streams. Summary of the Invention

[0004] The purpose of this invention is to provide a method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline, the method comprising: Acquire the raw guided wave echo signal collected by a sensor array arranged near the weld zone of the pipeline; Initial signal purification processing is performed to model and separate the inherent weld structure echo components in the original guided wave echo signal, resulting in the separated weld structure echo model and the initial purified signal data. The signal data of the initial purification is compared and analyzed in parallel with the echo model of the weld structure. Through iterative optimization, potential defect response signals are located and marked in the signal data of the initial purification, and a set of signal data marked with potential defect response signals is generated. Based on the signal data set labeled with potential defect response signals, a dynamic defect feature learning framework is constructed. The dynamic defect feature learning framework uses the weld structure echo model as a suppression reference to continuously learn and update the typical representation pattern of defect signals from the newly input data stream. The dynamic defect feature learning framework is used to perform online scanning of continuously acquired guided wave signals, identify signal segments that match the typical characterization pattern, and extract these signal segments as candidate defect feature signals. By integrating all the candidate defect feature signals and combining their temporal synchronization and spatial distribution information, the defect source is located and outlined in the three-dimensional space near the weld zone of the pipeline.

[0006] Preferably, the initial signal purification process involves modeling and separating the inherent weld structure echo components in the original guided wave echo signal to obtain the separated weld structure echo model and the initially purified signal data, specifically including: Signal samples from defect-free historical sections are selected from the original guided wave echo signals as the training basis; Multi-cycle alignment and averaging are performed on the signal samples of the defect-free historical section to generate a standard reference waveform representing the geometric characteristics of an ideal weld. The energy ratio and phase delay characteristics of the standard reference waveform under different guided wave modes are analyzed, and a mathematical model describing the scattering of the weld structure is established. The mathematical model is the weld structure echo model. The original guided wave echo signal is input into the weld structure echo model. Through adaptive filtering technology, the weld structure echo component predicted by the model is subtracted from the original signal, and the residual signal is output. The residual signal is the signal data of the initial purification.

[0007] Preferably, the step of performing parallel comparative analysis between the initial cleaned signal data and the weld structure echo model, and locating and marking potential defect response signals in the initial cleaned signal data through an iterative optimization method, and generating a signal data set marked with potential defect response signals, specifically includes: The signal data from the initial purification is divided into continuous signal analysis windows; Within each signal analysis window, the time-frequency energy distribution of the initial purified signal data is calculated and compared with the theoretical attenuation background of the weld structure echo model within the same time window; Design an optimization function with the significance of signal distortion as the objective, and find the signal region where the optimization function value reaches a local maximum by iteratively adjusting the detection threshold and window function parameters; Signal regions that meet the optimization criteria are identified as potential defect response signals, and markers are added to the corresponding time positions and sensor channels of the initial purified signal data to form a set of signal data marked with potential defect response signals.

[0008] Preferably, based on the signal data set labeled with potential defect response signals, a dynamic defect feature learning framework is constructed. This dynamic framework utilizes the weld structure echo model as a suppression reference to continuously learn and update the typical representation patterns of defect signals from newly input data streams. Specifically, this includes: From the signal data set labeled with potential defect response signals, a segment of the labeled signal is extracted as the initial training set; Extract multi-dimensional features for each signal segment, including but not limited to waveform kurtosis, spectral centroid shift, and cross-correlation delay between multiple sensors; Using the initial training set and its multi-dimensional features, an initial defect signal classifier is trained, which is used to distinguish defect signals from residual noise. During the online detection phase, the initial defect signal classifier interacts with the newly input signal that has been suppressed by the weld structure echo model. New defect signal samples with high classification confidence are automatically collected and used to periodically update the feature weights and discrimination boundaries of the defect signal classifier, thereby realizing the dynamic evolution of the typical representation mode.

[0009] Preferably, the step of using the dynamic defect feature learning framework to perform online scanning of continuously acquired guided wave signals, identifying signal segments that match the typical characterization pattern, and extracting these signal segments as candidate defect feature signals specifically includes: The guided wave signal stream, which is acquired in real time and purified by the initial signal, is input into the updated defect signal classifier. The defect signal classifier performs feature calculation and pattern matching on a point-by-point or window-by-window basis on the input signal stream and outputs the matching probability value of each analysis unit. A dynamic probability threshold is set. When the matching probability value of a certain analysis unit exceeds the dynamic probability threshold, the signal of the analysis unit and its preceding and following associated intervals is extracted. The extracted signal is refined at the boundary to remove the parts at both ends of the extracted signal that are mixed with background noise, while retaining the core response waveform to form a clean candidate defect feature signal.

[0010] Preferably, the integration of all the candidate defect feature signals, combined with their temporal synchronization and spatial distribution information, to perform defect source fusion localization and contour delineation in the three-dimensional space near the weld zone of the pipeline specifically includes: Collect temporally correlated candidate defect feature signals from different sensors in the sensor array; Calculate the precise time difference between each pair of associated candidate defect feature signals, and construct a set of hyperbolic equations about the defect source location based on the known sensor spatial coordinates and waveguide group velocity. The spatial grid search and consistency verification algorithm is used to solve the hyperbolic equation system to determine the two-dimensional circumferential and axial positions of the defect source on the pipe wall. By analyzing the amplitude ratio and waveform envelope shape of candidate defect feature signals from sensors from different directions, the orientation characteristics and possible size of the defect are inferred, and all information is combined to generate a probability distribution cloud map of the defect in three-dimensional space.

[0011] Preferably, the step of selecting signal samples from defect-free historical sections of the original guided wave echo signal as the training basis specifically includes: Retrieve and archive historical inspection data of the pipeline in the area near the weld when it was initially defect-free; Select multiple consecutive detection cycle data from the historical detection data archive that have a signal-to-noise ratio higher than a preset standard and no abnormal alarm records; The selected data from multiple consecutive detection cycles are overlaid in time to generate a signal sample set for the defect-free historical segment.

[0012] Preferably, the step of designing an optimization function with signal distortion significance as the objective, and finding the signal region where the optimization function value reaches a local maximum by iteratively adjusting the detection threshold and window function parameters, specifically includes: The significance of signal distortion is defined as a function of the ratio of time-frequency energy to background model energy within the signal analysis window, and a penalty term for instantaneous phase changes in the signal is added. Initialize a broad detection threshold and a default rectangular window function; During the iteration process, the detection threshold is gradually narrowed and the shape and length of the window function are changed, and the signal distortion significance of each signal analysis window is recalculated; The system monitors changes in the signal distortion saliency value. When the signal distortion saliency value of a certain signal region continues to increase in several consecutive iterations and eventually stabilizes at a high level, the iteration is stopped and the signal region with a signal distortion saliency value higher than a preset threshold is locked.

[0013] Preferably, the step of performing boundary trimming on the truncated signal, removing the portions at both ends of the truncated signal that are mixed with background noise, and retaining the core response waveform, specifically includes: Calculate the short-time average energy of the signal from the center of the extracted signal outwards to both ends; Determine the location point where the short-time average energy drops to a certain proportion of the peak energy of the intercepted signal; Using these two locations as the new boundaries, discard the data points outside the boundaries to obtain the signal segments after boundary refinement.

[0014] Preferably, the step of employing a spatial grid search and consistency check algorithm to solve the hyperbolic equation system and determine the two-dimensional circumferential and axial positions of the defect source on the pipe wall specifically includes: The pipe surface is unfolded into a two-dimensional grid with axial and circumferential coordinates; For each potential defect source point in the grid, calculate its theoretical time of arrival to each sensor; The theoretical arrival time is compared with the measured arrival time of the candidate defect characteristic signal, and the sum of squares of the time residuals of all sensors is calculated. Traverse the entire two-dimensional grid to find the grid point that minimizes the sum of squares of the time residuals, and use the grid point as the initial solution for the location of the defect source. Cross-validation is performed using time difference information from multiple sensors near the initial solution to eliminate inconsistent sensor data. The remaining data is then used to recalculate and obtain the final optimized defect source location coordinates.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By actively modeling the inherent weld structure echo components mathematically and separating them from mixed signals, precise suppression of background interference is achieved. This technique avoids the absolute dependence of traditional methods on defect-free baseline signals, directly extracting a stable weld response model from the measured signal. An iterative optimization approach is used for parallel comparison and localization in the initial cleaned signal, gradually converging and identifying signal components with slight deviations from the model. This process improves the signal-to-noise ratio, revealing early, minor defect responses that were previously masked by strong weld echoes. The separation and iteration mechanism enhances the detection system's sensitivity to the initial state of defects, reducing false positives and false negatives caused by background interference. This method can adapt to signal differences caused by weld geometries, improving the stability and applicability of the detection process.

[0016] A dynamic defect feature learning framework, using a weld echo model as a suppression reference, was constructed, enabling the system to continuously evolve. This framework can handle continuously input online data streams and autonomously learn and update typical representation patterns of defect signals. Through the learning mechanism, the defect feature patterns identified by the system are no longer fixed but are continuously optimized and adjusted with the input of new data. This improves the ability to track and identify defect morphological expansion or slow feature changes. Online scanning and real-time matching reduce the subjectivity and lag of manually setting fixed thresholds, achieving automated identification of defect signals. The adaptive learning process allows the entire detection system to better match the actual state changes of pipelines during long-term operation, improving the reliability of long-term monitoring and the timeliness of early warning. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the pipe near-weld zone defect guided wave characteristic signal extraction method described in this invention. Figure 2 A flowchart for the initial signal purification process; Figure 3 A flowchart for parallel comparative analysis and potential defect labeling; Figure 4 A biaxial bar graph of multi-sensor signal characteristics for guided wave detection of defects near the weld zone of a pipeline. Figure 5 A bar chart showing the amplitude distribution of multi-sensor signals for guided wave detection of defects near the weld zone of a pipeline. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a method for extracting guided wave feature signals of defects in the near-weld zone of a pipeline. The method includes: acquiring raw guided wave echo signals collected by a sensor array arranged in the near-weld zone of the pipeline; performing initial signal purification processing to model and separate the inherent weld structure echo components in the raw guided wave echo signals, obtaining a separated weld structure echo model and initially purified signal data; performing parallel comparative analysis between the initially purified signal data and the weld structure echo model, and locating and marking potential defect response signals in the initially purified signal data through iterative optimization, generating a signal data set marked with potential defect response signals; constructing a dynamic defect feature learning framework based on the signal data set marked with potential defect response signals, which uses the weld structure echo model as a suppression reference to continuously learn and update typical representation patterns of defect signals from newly input data streams; and using the dynamic defect feature learning framework to perform online scanning of continuously acquired guided wave signals, identifying signal segments that match typical representation patterns, and extracting these signal segments as candidate defect feature signals. By integrating all candidate defect feature signals and combining their temporal synchronization and spatial distribution information, the defect source is located and outlined in the three-dimensional space near the weld zone of the pipeline.

[0020] In one embodiment of the present invention, see [reference] Figure 2The initial signal purification process is performed to model and separate the inherent weld structure echo components in the original guided wave echo signal. The specific operations are as follows: Signal samples from defect-free historical sections are selected from the original guided wave echo signal as the training basis. This involves retrieving historical inspection data archives of the pipeline in the near-weld region under initial defect-free conditions. Multiple consecutive inspection cycles with a signal-to-noise ratio higher than a preset standard and no abnormal alarm records are selected from the historical inspection data archives. These selected consecutive inspection cycles are then time-synchronized and superimposed to generate a set of signal samples from defect-free historical sections. Multi-cycle alignment and averaging are performed on the signal samples from defect-free historical sections to generate a standard reference waveform representing the ideal weld geometry. The energy ratio and phase delay characteristics of the standard reference waveform under different guided wave modes are analyzed to establish a mathematical model describing weld structure scattering; this mathematical model is the weld structure echo model. Specifically, the standard reference waveform is first subjected to dispersion compensation processing to eliminate waveform broadening and distortion caused by guided wave dispersion effects, thereby obtaining clear mode separation results. Then, a time-frequency analysis-based mode decomposition algorithm is used to decompose the processed waveform into several main guided wave mode components, such as the symmetric mode S0 and the antisymmetric mode A0. For each separated mode component, its instantaneous amplitude envelope is extracted using Hilbert transform, and the proportion coefficient of that mode component in the total energy of the entire standard reference waveform is calculated. Simultaneously, the start time delay of each modal component relative to the excitation pulse signal is measured using a cross-correlation function. Based on the above analysis results, a mathematical model for scattering of multimodal superposition weld structures is constructed. The frequency domain expression of this model is:

[0021] in, The frequency domain transfer function representing the scattering model of the weld structure. For frequency variables, The total number of main guided wave modes involved in the modeling. For modal indexing, For the first Energy proportionality coefficients for each mode For the first The normalized amplitude spectrum of each mode is obtained by performing a Fast Fourier Transform on the waveform of the corresponding mode component and then normalizing it. The imaginary unit, For the first The time delay of each mode. This mathematical model serves as the echo model for the weld structure, used to predict the frequency domain characteristics of the inherent echo signal generated by the weld geometry under given excitation conditions.

[0022] The original guided wave echo signal is input into the weld structure echo model. Through adaptive filtering technology, the weld structure echo component predicted by the model is subtracted from the original signal, and the residual signal is output. This residual signal is the signal data of the initial purification.

[0023] In practice, initial signal purification is performed to model and separate the inherent weld structure echo components in the original guided wave echo signal. One example scenario involves in-service guided wave detection of the circumferential weld of an oil and gas pipeline. A sensor array deployed near the weld consists of eight piezoelectric sensors arranged in an equally spaced ring array on one side of the weld. The original guided wave echo signal acquired by the sensor array includes the excitation main wave packet, strong scattered echoes from the weld structure, possible defect scattered signals, and environmental noise. Signal samples from defect-free historical sections are selected from the original guided wave echo signal as the training basis. Historical detection data archives of the pipeline in the initial defect-free state near the weld are retrieved. These archives contain ten sets of complete guided wave echo data recorded during baseline detection before pipeline commissioning. Multiple consecutive detection cycles with a signal-to-noise ratio higher than a preset standard and no abnormal alarm records are selected from the historical detection data archive. The preset standard is an overall signal-to-noise ratio greater than 20 decibels. All ten selected data sets meet this condition. Multiple consecutive detection cycle data selected are time-synchronized and superimposed to generate a signal sample set of defect-free historical sections. Time synchronization and superposition is achieved by aligning the direct incident wavefront in each cycle signal using a cross-correlation algorithm. Then, the arithmetic mean of the ten aligned signals is calculated, and the averaged signal waveform serves as the representative waveform of the signal sample set of defect-free historical sections.

[0024] Multi-cycle alignment and averaging are performed on signal samples from defect-free historical sections to generate a standard reference waveform representing the ideal weld geometry. This multi-cycle alignment and averaging is completed during the generation of the signal sample set for defect-free historical sections, and the resulting average waveform is the standard reference waveform. The energy ratio and phase delay characteristics of the standard reference waveform under different guided wave modes are analyzed to establish a mathematical model describing weld structure scattering. In practice, the standard reference waveform undergoes dispersion compensation and mode decomposition to separate the response components of the bending and tensile modes. The amplitude ratio of each mode component is quantified, and its time delay relative to the excitation signal is measured. The mathematical model, i.e., the weld structure echo model, is constructed using a parameterized transfer function. The transfer function simulates the filtering effect of the weld as a scatterer on different incident guided wave modes. Its core is a weighted multi-delay superposition structure used to synthesize the predicted weld structure echo components. The mathematical expression of the transfer function is:

[0025] in: This represents the frequency domain response model of the weld structure. Represents frequency variables. This represents the total number of guided wave modes involved in the modeling. It is a frequency-dependent complex coefficient, characterizing the first... Amplitude and phase modulation of each mode, Represents the imaginary unit. Representing the Each modal component has a fixed time delay relative to the excitation signal. Parameters are optimized and determined by fitting measured data of a standard reference waveform. and .

[0026] The original guided wave echo signal is input into the weld structure echo model. Adaptive filtering technology is used to subtract the weld structure echo component predicted by the model from the original signal. In some embodiments, a least mean square adaptive filter is used for processing. The filter is based on the weld structure echo model. The unit impulse response in the time domain is used as the reference input. The filtering process is performed in the time domain, and the algorithm continuously adjusts the filter weight coefficients to minimize the mean square error between the filter's output signal and the original guided wave echo signal. At this point, the filter output is the best estimate of the weld structure echo component in the original signal. This estimated component is subtracted point by point from the original guided wave echo signal, and the residual signal is output. The residual signal is the signal data of the initial purification. In the data comparison of the example scenario, the amplitude of the original signal before processing fluctuates significantly within the weld echo time window, while the amplitude fluctuation of the signal data after processing and initial purification is significantly reduced within the same time window, the background tends to be stable, and the potential defect scattering peaks are more clearly revealed.

[0027] In one embodiment of the present invention, see [reference] Figure 3 The initial cleaned signal data and the weld structure echo model are compared and analyzed in parallel. Through iterative optimization, potential defect response signals are located and marked in the initial cleaned signal data, generating a set of signal data marked with potential defect response signals. The initial cleaned signal data is divided into continuous signal analysis windows. Within each signal analysis window, the time-frequency energy distribution of the initial cleaned signal data is calculated and compared with the theoretical attenuation background of the weld structure echo model within the same time window. An optimization function targeting signal distortion significance is designed. By iteratively adjusting the detection threshold and window function parameters, the function seeks to find the signal region where the optimization function value reaches a local maximum. The process involves defining signal distortion significance as a function of the ratio of time-frequency energy to background model energy within the signal analysis window, incorporating a penalty term for instantaneous phase abrupt changes in the signal. A broad detection threshold and a default rectangular window function are initialized. During iteration, the detection threshold is gradually narrowed, and the shape and length of the window function are changed. The signal distortion significance of each signal analysis window is recalculated, and changes in the signal distortion significance value are monitored. When the signal distortion significance value of a certain signal region continues to increase over several iterations and eventually stabilizes at a high level, the iteration stops, and the signal region with a signal distortion significance value higher than the preset threshold is locked. Signal regions that meet the optimization criteria are identified as potential defect response signals, and markers are added to the corresponding time positions and sensor channels of the initially cleaned signal data, forming a signal data set marked with potential defect response signals.

[0028] In practical implementation, the initial purified signal data and the weld structure echo model are compared and analyzed in parallel. An example scenario continues the detection of circumferential welds in oil and gas pipelines from the previous embodiment. The initial purified signal data originates from sensor channel number three in the sensor array. This channel's signal still exhibits amplitude fluctuations within the time range of 150 to 250 microseconds. The initial purified signal data is divided into continuous signal analysis windows, using a 20-microsecond Hanning window with a 15-microsecond overlap between adjacent windows, thus covering the entire detection time range of interest. Within each signal analysis window, the time-frequency energy distribution of the initial purified signal data is calculated and compared with the theoretical attenuation background of the weld structure echo model within the same time window. The time-frequency energy distribution is calculated using a short-time Fourier transform, while the theoretical attenuation background of the weld structure echo model is generated by convolving the amplitude response curve of the weld structure echo model at the corresponding center frequency with an exponential decay function.

[0029] Design an optimization function targeting signal distortion significance. Iteratively adjust the detection threshold and window function parameters to find the signal region where the optimization function value reaches a local maximum. Define signal distortion significance as a function of the ratio of time-frequency energy to background model energy within the signal analysis window, and incorporate a penalty term for instantaneous phase abrupt changes in the signal. The formula for calculating signal distortion significance is:

[0030] in: Represents the significance of signal distortion. This represents the total number of time-frequency points within the signal analysis window. Representing the Energy value at each time frequency point The background model represents the first The theoretical energy value at each time frequency point It is a tiny positive number used to prevent division by zero. It is the weighting coefficient of the penalty term. This is a quantitative measure of instantaneous phase abrupt changes. A broad detection threshold and a default rectangular window function are initialized. The broad detection threshold is set to twice the baseline value of signal distortion significance, and the default rectangular window function length is set to 30 microseconds. During iteration, the detection threshold is gradually narrowed, and the shape and length of the window function are changed. The signal distortion significance of each signal analysis window is recalculated. The detection threshold is narrowed by a step size of 5% per iteration. The window function shape alternates between rectangular, Hanning, and Blackman windows, and the window function length is adjusted between 15 and 40 microseconds. Changes in the signal distortion significance value are monitored. When the signal distortion significance value of a certain signal region continues to increase over several iterations and eventually stabilizes at a high level, the iteration stops, and the signal region with a signal distortion significance value higher than the preset threshold is locked. In the example scenario, for sensor channel number 3, the iteration process converged after the fifth iteration, locking a signal region centered at 180 microseconds and with a width of 22 microseconds.

[0031] Signal regions meeting the optimization criteria are identified as potential defect response signals. Marks are added to the corresponding time positions and sensor channels of the initially purified signal data, forming a signal data set marked with potential defect response signals. In some embodiments, the marking information is stored in metadata form, which includes the start timestamp, end timestamp, center time, maximum signal distortion significance value, and the physical number of the associated sensor channel of the locked signal region. It can be understood that the signal data set marked with potential defect response signals not only includes the original initially purified signal data samples but also is associated with structured marked data describing the spatial and characteristic attributes of the potential defect response signals. At the data comparison level, in the initially purified signal data before iterative optimization, potential defect response signals are mixed in with residual fluctuations and are difficult to directly identify. After parallel comparative analysis and iterative optimization localization, the signal data set marked with potential defect response signals clearly indicates that sensor channel number three has an abnormal signal segment with high signal distortion significance around 180 microseconds.

[0032] In one embodiment of the present invention, a dynamic defect feature learning framework is constructed based on a signal data set labeled with potential defect response signals. This framework utilizes a weld structure echo model as a suppression reference to continuously learn and update the typical representation patterns of defect signals from newly input data streams. Labeled signal segments are extracted from the signal data set labeled with potential defect response signals as an initial training set. Multi-dimensional features of each signal segment are extracted, including but not limited to waveform kurtosis, spectral centroid shift, and cross-correlation delay between multiple sensors. An initial defect signal classifier is trained using the initial training set and its multi-dimensional features to distinguish defect signals from residual noise. During the online detection phase, the initial defect signal classifier interacts with newly input signals suppressed by the weld structure echo model. New defect signal samples with high classification confidence are automatically collected and used to periodically update the feature weights and discrimination boundaries of the defect signal classifier, achieving dynamic evolution of typical representation patterns.

[0033] In practical implementation, a dynamic defect feature learning framework is constructed based on a signal data set labeled with potential defect response signals. An example scenario continues to focus on the detection of the oil and gas pipeline. The signal data set labeled with potential defect response signals already contains labeling information from eight sensor channels. Sensor channels numbered 3, 5, and 7 are all labeled with potential defect response signals within the time interval of 170 to 190 microseconds. From the signal data set labeled with potential defect response signals, labeled signal segments are extracted as the initial training set. Specifically, based on the start and end timestamps of each label record, complete signal waveform segments are extracted from the corresponding initially cleaned signal data. These segments, along with their respective sensor channel information, constitute the samples of the initial training set.

[0034] Multi-dimensional features are extracted from each signal segment, including waveform kurtosis, spectral centroid shift, and cross-correlation delay between multiple sensors. Waveform kurtosis measures the steepness of the signal amplitude distribution, spectral centroid shift describes the movement of signal energy around the center of its frequency distribution, and cross-correlation delay between multiple sensors quantifies the time lag of the same event between signals from different sensors. The formula for calculating waveform kurtosis is:

[0035] in: Represents waveform kurtosis. Represents the total number of sampling points in a signal segment. The first one representing the signal sequence The amplitude of each sampling point This represents the average amplitude of all sampled points in the signal sequence. The standard deviation of the amplitude across all sampling points of the signal sequence. Spectral centroid offset. The cross-correlation delay is obtained by calculating the weighted average frequency of the amplitude spectrum of the Fast Fourier Transform (FFT) of the signal segment after Hanning window weighting and subtracting it from the spectral centroid frequency of a defect-free background reference spectrum. The determination is made by calculating the cross-correlation function between the labeled signal segment and the signal in the corresponding time interval of a reference sensor channel, and then finding the time shift corresponding to the peak position of the cross-correlation function.

[0036] Using the initial training set and its multi-dimensional features, an initial defect signal classifier is trained. In some embodiments, the initial defect signal classifier employs a support vector machine algorithm, with the input feature vector being the waveform kurtosis calculated for each signal segment. Spectral barycenter shift and cross-correlation delay The resulting vectors are either two-dimensional or three-dimensional. The training process uses all samples in the signal dataset labeled with potential defect response signals, and automatically converts these samples into binary class labels. Labeled signal segments are assigned the label "defect signal," while an equal number of signal segments randomly selected from unlabeled time regions are assigned the label "residual noise." In essence, the initial defect signal classifier solves an optimization problem to find a hyperplane in the feature space that maximizes the separation between the two classes, thus gaining an initial ability to distinguish between defect signals and residual noise.

[0037] During the online detection phase, the initial defect signal classifier interacts with newly input signals, which have been suppressed by the weld structure echo model. New defect signal samples with high classification confidence are automatically collected and used to periodically update the feature weights and discrimination boundaries of the defect signal classifier, achieving dynamic evolution of typical representation patterns. Optionally, during online detection, the system processes newly acquired guided wave echo signals in real time. After suppression by the weld structure echo model, a new, initially purified signal data stream is generated. The defect signal classifier performs sliding window analysis on the data stream, calculating a feature vector for each analysis window and outputting a classification confidence score. When the classification confidence score of a certain analysis window exceeds a preset high threshold, the system automatically archives the signal segment of that window and its calculated feature vector into a dynamic sample library. Every fixed time period or when the dynamic sample library accumulates a certain number of new samples, the system uses all samples in the dynamic sample library to perform incremental learning and update the model parameters of the defect signal classifier. In the data comparison of the example scenario, the detection rate of the initially trained defect signal classifier for simulated defect signals fluctuated to some extent. After running online for twelve consecutive hours and completing three incremental updates, the classification confidence scores of the defect signal classifier for newly emerging similar defect signals became more concentrated and stable, indicating that the typical representation pattern of defect signals was refined and strengthened during the learning process.

[0038] In one embodiment of the present invention, a dynamic defect feature learning framework is used to perform online scanning of continuously acquired guided wave signals, identify signal segments that match typical characterization patterns, and extract these signal segments as candidate defect feature signals. The guided wave signal stream, acquired in real-time and processed with initial signal purification, is input to an updated defect signal classifier. The defect signal classifier performs point-by-point or window-by-window feature calculation and pattern matching on the input signal stream, outputting the matching probability value for each analysis unit. A dynamic probability threshold is set; when the matching probability value of an analysis unit exceeds the dynamic probability threshold, the signal of that analysis unit and its preceding and following intervals is extracted. The extracted signals undergo boundary refinement, removing the portions of the extracted signal that are mixed with background noise at both ends, retaining the core response waveform. This process involves calculating the short-time average energy of the signal from the center of the extracted signal outwards, determining the position points where the short-time average energy drops to a certain proportion of the peak energy of the extracted signal, and using these two position points as new boundaries. Data points outside the boundaries are discarded to obtain the boundary-refined signal segments, forming clean candidate defect feature signals.

[0039] In practical implementation, a dynamic defect feature learning framework is used to perform online scanning of continuously acquired guided wave signals, identify signal segments that match typical representation patterns, and extract these signal segments as candidate defect feature signals. An example scenario involves online monitoring of the pipeline near the weld zone; the defect signal classifier in the dynamic defect feature learning framework has undergone three incremental learning updates. The guided wave signal stream, acquired in real-time and processed after initial signal purification, is input into the updated defect signal classifier, which is loaded with the latest support vector machine model weights and discrimination boundary parameters. The updated defect signal classifier performs point-by-point or window-by-window feature calculation and pattern matching on the input signal stream, and outputs the matching probability value of each analysis unit. In specific implementation, the analysis unit is set to a 20-microsecond time window of the same length as the initial training set signal segment, and slides in 10-microsecond steps. For each sliding analysis unit window, its waveform kurtosis, spectral centroid offset and cross-correlation delay with the reference channel are calculated in real time to form a feature vector and input into the updated defect signal classifier. The updated defect signal classifier outputs a matching probability value between 0 and 1, which represents the confidence level that the signal of the current analysis unit belongs to the defect signal category.

[0040] A dynamic probability threshold is set. When the matching probability value of a certain analysis unit exceeds the dynamic probability threshold, the signal of that analysis unit and its preceding and following associated intervals are extracted. Dynamic probability threshold. It is not a fixed value, but is adjusted based on the statistical characteristics of the output matching probability values ​​over a recent period. The calculation formula is as follows:

[0041] in: This represents the moving average of the matching probability values ​​of the most recent 1000 analysis units. The moving standard deviation of these matching probability values. It is an adjustable sensitivity coefficient. In a single 12-hour online scanning example, refer to Table 1 for the output records of matching probability values ​​and the dynamic threshold judgment.

[0042] Table 1: Matching probability output and threshold judgment table for some analysis units during online scanning.

[0043] When the matching probability value of the analysis unit exceeds the dynamic probability threshold, the system uses the center time of the analysis unit as a reference and extends it forward and backward by 15 microseconds to extract the original initial purified signal data with a total duration of 50 microseconds from the cache.

[0044] The extracted signal undergoes boundary refinement to remove the portions at both ends of the extracted signal that are mixed with background noise, retaining the core response waveform to form a clean candidate defect feature signal. Specifically, boundary refinement involves calculating the short-time average energy of the signal from its center outwards. The points where the short-time average energy drops to 20% of the peak energy of the extracted signal are determined. These two points serve as new boundaries, discarding data points outside the boundaries to obtain the boundary-refined signal segment. In some embodiments, a 5-microsecond rectangular window is used to calculate the short-time average energy, with a reference energy decrease ratio set to 20%. For example, for a 50-microsecond signal segment extracted from 165 microseconds to 215 microseconds, its peak energy occurs at 179 microseconds. Calculating the short-time average energy from the center point in the direction of increasing time, it is found that the short-time average energy first drops below 20% of the peak energy at 188 microseconds. Calculating from the center point in the direction of decreasing time, the short-time average energy drops below 20% of the peak energy at 171 microseconds. Therefore, the time range of the signal segment after boundary refinement is 171 microseconds to 188 microseconds. The mixed portions at both ends of the original truncated signal are discarded. Data comparison shows that the waveform of the refined signal segment is more compact, and the signal-to-noise ratio is visually improved compared to the original truncated signal. It can be understood that the signal segment formed after boundary refinement is a clean candidate defect feature signal, ready for subsequent fusion and localization analysis.

[0045] See Figure 4This is a biaxial bar chart of the multi-sensor signal characteristics of guided wave detection for defects near the weld zone of a pipeline. The time difference of arrival (TDO) increases with the circumferential angle of the sensor; the signal amplitude decreases with the TDO. This is the signal characteristic acquisition stage before defect localization, providing basic data for subsequent "hyperbolic equation system solution and three-dimensional defect localization". The difference in TDO is the core basis for calculating the spatial location of the defect; the attenuation law of the signal amplitude can help infer the size and energy scattering characteristics of the defect. This figure clarifies the temporal and amplitude characteristics of the guided wave signal of the defect under multi-sensor channels, which is the core input data for the "spatial localization and contour delineation" of defects near the weld zone of the pipeline. Subsequently, a localization model needs to be built based on these data to achieve accurate three-dimensional localization of the defect.

[0046] In one embodiment of the invention, all candidate defect feature signals are integrated, and their temporal synchronization and spatial distribution information are combined to perform fusion localization and contour delineation of the defect source in the three-dimensional space near the weld zone of the pipeline. Temporally correlated candidate defect feature signals from different sensors in the sensor array are collected. The precise time difference between each pair of correlated candidate defect feature signals is calculated, and a set of hyperbolic equations about the defect source location is constructed based on the known sensor spatial coordinates and guided wave group velocity. A spatial grid search and consistency verification algorithm is employed to solve the hyperbolic equations and determine the two-dimensional circumferential and axial positions of the defect source on the pipe wall. The process involves unfolding the pipe surface into a two-dimensional grid with axial and circumferential coordinates. For each potential defect source point in the grid, the theoretical arrival time to each sensor is calculated. This theoretical arrival time is compared with the measured arrival times of candidate defect feature signals. The sum of squares of the time residuals from all sensors is calculated. The entire two-dimensional grid is traversed to find the grid point that minimizes the sum of squares of the time residuals. This grid point is used as the initial solution for the defect source location. Cross-validation is performed using time difference information from multiple sensors near the initial solution to eliminate inconsistent sensor data. The remaining data is used to recalculate, resulting in the final optimized defect source location coordinates. The amplitude ratio and waveform envelope shape of candidate defect feature signals from sensors in different directions are analyzed to infer the defect's orientation characteristics and possible size. All information is then integrated to generate a probability distribution cloud map of the defect in three-dimensional space.

[0047] In practical implementation, all candidate defect feature signals are integrated, and their temporal synchronization and spatial distribution information are combined to perform defect source fusion localization and contour delineation in the three-dimensional space near the weld zone of the pipeline. In an example scenario, the online scanning process from eight sensor arrays extracts time-correlated candidate defect feature signals from sensor channels numbered three, five, and seven within three consecutive detection cycles. The temporally correlated candidate defect feature signals from different sensors in the sensor array are aggregated. The correlation is determined by these candidate defect feature signals appearing within similar detection cycles and their center times falling within the possible propagation time window estimated based on wave velocity. The system aggregates these three candidate defect feature signals from different sensors along with their metadata into a processing buffer.

[0048] The precise time difference between each pair of associated candidate defect feature signals is calculated, and a set of hyperbolic equations about the defect source location is constructed based on the known sensor spatial coordinates and waveguide group velocity. In practice, the precise time difference is obtained by fine-grained alignment of the paired candidate defect feature signal waveforms using a cross-correlation method. The sensor spatial coordinates are precisely measured and entered into the system during installation, and the waveguide group velocity... The wavegroup velocity is determined by the pipe material properties and the center frequency of the excitation guided wave. For the pipe in the example, the wavegroup velocity is... The speed is 5 kilometers per second. Taking sensor 3 and sensor 5 as an example, the calculated arrival time difference... The time difference is 2.4 microseconds. Based on this time difference and the waveguide group velocity, the location of the defect source must satisfy a hyperbolic equation where the distance difference to the two sensors is constant. The same operation was performed on sensors 3 and 7, and sensors 5 and 7, constructing three independent hyperbolic equations, forming a set of equations for solving the defect source location. The system of hyperbolic equations, where and These are the coordinates on the two-dimensional plane of the unfolded pipeline.

[0049] A spatial grid search and consistency check algorithm is used to solve the hyperbolic equations to determine the two-dimensional circumferential and axial positions of the defect source on the pipe wall. The pipe surface is unfolded into a two-dimensional grid with axial and circumferential coordinates, with a grid resolution of 1 mm axially and 1 degree circumferentially. For each potential defect source point in the grid, the theoretical arrival time to each sensor is calculated. The theoretical arrival time is compared with the measured arrival times of candidate defect characteristic signals, and the sum of squares of the time residuals of all sensors is calculated. The calculation formula is:

[0050] in: This represents the sum of squares of the time residuals of all sensors. This represents the total number of sensors involved in the positioning process, as shown in the example. , This represents the calculation up to the current grid point. The theoretical time of arrival for each sensor Representing the The arrival times of candidate defect feature signals measured by each sensor are determined. The sum of squares of the time residuals is searched by traversing the entire two-dimensional grid. The smallest grid point is used as the initial solution for the defect source location. Cross-validation is performed using time difference information from multiple sensors near the initial solution to eliminate inconsistent sensor data. In some embodiments, cross-validation calculates the time difference residuals for each sensor pair. If the absolute value of all time difference residuals from a particular sensor is significantly greater than those from other sensors, it is determined that the data from sensor number five may be affected by noise and is excluded from subsequent calculations. The remaining data is used to recalculate, resulting in the final optimized defect source location coordinates. In the example scenario data comparison, the initial solution coordinates obtained using data from all three sensors are (152 mm axially, 115 degrees circumferentially). After excluding the data from sensor number five, a new grid search is performed using data from sensors number three and seven, resulting in the final optimized coordinates (155 mm axially, 112 degrees circumferentially).

[0051] By analyzing the amplitude ratio and waveform envelope shape of candidate defect feature signals from sensors from different directions, the orientation characteristics and possible size of the defect are inferred. It can be understood that the amplitude ratio of the signals acquired by sensors 3 and 7 contains information about the directionality of the defect scatterer, while the width and diffusion characteristics of the waveform envelope can be correlated with the equivalent size of the defect. A probability distribution cloud map of the defect in three-dimensional space is generated by integrating all information. Centered on the final optimized defect source location coordinates, the probability distribution cloud map is generated through weighted diffusion based on the uncertainties of multi-source information such as time residuals, amplitude consistency, and waveform characteristics. This visually presents the most likely area where the defect exists on the pipe wall and the estimated range of its spatial contour.

[0052] See Figure 5This is a bar chart showing the amplitude distribution of multi-sensor signals from guided wave detection of defects near the weld zone in a pipeline. The signal amplitudes of sensors 3, 5, and 7 are significantly higher than other channels, exhibiting a distribution characteristic of "locally high amplitude and otherwise low amplitude," reflecting the closer spatial distance between the defect and these sensors. This is the initial defect signal screening stage, used to identify sensor channels more sensitive to defect responses, focusing effective data for subsequent "defect feature extraction and localization analysis." High-amplitude channels (3, 5, and 7) are the core reference channels for defect localization, containing richer defect information; low-amplitude channels can serve as background noise references, helping to improve the signal-to-noise ratio of the defect signal. This chart clarifies the differences in response of different sensor channels to defect guided wave signals, providing the foundation for "effective signal screening and high signal-to-noise ratio feature extraction" of defects near the weld zone in pipelines. Subsequent defect localization and contour analysis will primarily focus on signals from high-amplitude channels.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline, characterized in that, include: Acquire the raw guided wave echo signal collected by a sensor array arranged near the weld zone of the pipeline; Initial signal purification processing is performed to model and separate the inherent weld structure echo components in the original guided wave echo signal, resulting in the separated weld structure echo model and the initial purified signal data. The signal data of the initial purification is compared and analyzed in parallel with the echo model of the weld structure. Through iterative optimization, potential defect response signals are located and marked in the signal data of the initial purification, and a set of signal data marked with potential defect response signals is generated. Based on the signal data set labeled with potential defect response signals, a dynamic defect feature learning framework is constructed. The dynamic defect feature learning framework uses the weld structure echo model as a suppression reference to continuously learn and update the typical representation pattern of defect signals from the newly input data stream. The dynamic defect feature learning framework is used to perform online scanning of continuously acquired guided wave signals, identify signal segments that match the typical characterization pattern, and extract these signal segments as candidate defect feature signals. By integrating all the candidate defect feature signals and combining their temporal synchronization and spatial distribution information, the defect source is located and outlined in the three-dimensional space near the weld zone of the pipeline.

2. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 1, characterized in that, The initial signal purification process involves modeling and separating the inherent weld structure echo components in the original guided wave echo signal to obtain the separated weld structure echo model and the initially purified signal data. Specifically, this includes: Signal samples from defect-free historical sections are selected from the original guided wave echo signals as the training basis; Multi-cycle alignment and averaging are performed on the signal samples of the defect-free historical section to generate a standard reference waveform representing the geometric characteristics of an ideal weld. The energy ratio and phase delay characteristics of the standard reference waveform under different guided wave modes are analyzed, and a mathematical model describing the scattering of the weld structure is established. The mathematical model is the weld structure echo model. The original guided wave echo signal is input into the weld structure echo model. Through adaptive filtering technology, the weld structure echo component predicted by the model is subtracted from the original signal, and the residual signal is output. The residual signal is the signal data of the initial purification.

3. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 2, characterized in that, The process of performing parallel comparative analysis between the initial cleaned signal data and the weld structure echo model, and locating and marking potential defect response signals in the initial cleaned signal data through an iterative optimization method, thereby generating a signal data set marked with potential defect response signals, specifically includes: The signal data from the initial purification is divided into continuous signal analysis windows; Within each signal analysis window, the time-frequency energy distribution of the initial purified signal data is calculated and compared with the theoretical attenuation background of the weld structure echo model within the same time window; Design an optimization function with the significance of signal distortion as the objective, and find the signal region where the optimization function value reaches a local maximum by iteratively adjusting the detection threshold and window function parameters; Signal regions that meet the optimization criteria are identified as potential defect response signals, and markers are added to the corresponding time positions and sensor channels of the initial purified signal data to form a set of signal data marked with potential defect response signals.

4. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 3, characterized in that, Based on the signal data set labeled with potential defect response signals, a dynamic defect feature learning framework is constructed. This dynamic framework utilizes the weld structure echo model as a suppression reference to continuously learn and update the typical representation patterns of defect signals from newly input data streams. Specifically, it includes: From the signal data set labeled with potential defect response signals, a segment of the labeled signal is extracted as the initial training set; Extract multi-dimensional features for each signal segment, including but not limited to waveform kurtosis, spectral centroid shift, and cross-correlation delay between multiple sensors; Using the initial training set and its multi-dimensional features, an initial defect signal classifier is trained, which is used to distinguish defect signals from residual noise. During the online detection phase, the initial defect signal classifier interacts with the newly input signal that has been suppressed by the weld structure echo model. New defect signal samples with high classification confidence are automatically collected and used to periodically update the feature weights and discrimination boundaries of the defect signal classifier, thereby realizing the dynamic evolution of the typical representation mode.

5. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 4, characterized in that, The process of using the dynamic defect feature learning framework to perform online scanning of continuously acquired guided wave signals, identifying signal segments that match the typical characterization pattern, and extracting these signal segments as candidate defect feature signals specifically includes: The guided wave signal stream, which is acquired in real time and purified by the initial signal, is input into the updated defect signal classifier. The defect signal classifier performs feature calculation and pattern matching on a point-by-point or window-by-window basis on the input signal stream and outputs the matching probability value of each analysis unit. A dynamic probability threshold is set. When the matching probability value of a certain analysis unit exceeds the dynamic probability threshold, the signal of the analysis unit and its preceding and following associated intervals is extracted. The extracted signal is refined at the boundary to remove the parts at both ends of the extracted signal that are mixed with background noise, while retaining the core response waveform to form a clean candidate defect feature signal.

6. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 5, characterized in that, The process of integrating all the candidate defect feature signals, combining their temporal synchronization and spatial distribution information, and performing defect source fusion localization and contour delineation in the three-dimensional space near the weld zone of the pipeline specifically includes: Collect temporally correlated candidate defect feature signals from different sensors in the sensor array; Calculate the precise time difference between each pair of associated candidate defect feature signals, and construct a set of hyperbolic equations about the defect source location based on the known sensor spatial coordinates and waveguide group velocity. The spatial grid search and consistency verification algorithm is used to solve the hyperbolic equation system to determine the two-dimensional circumferential and axial positions of the defect source on the pipe wall. By analyzing the amplitude ratio and waveform envelope shape of candidate defect feature signals from sensors from different directions, the orientation characteristics and possible size of the defect are inferred, and all information is combined to generate a probability distribution cloud map of the defect in three-dimensional space.

7. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 2, characterized in that, The step of selecting signal samples from defect-free historical sections in the original guided wave echo signal as the training basis specifically includes: Retrieve and archive historical inspection data of the pipeline in the area near the weld when it was initially defect-free; Select multiple consecutive detection cycle data from the historical detection data archive that have a signal-to-noise ratio higher than a preset standard and no abnormal alarm records; The selected data from multiple consecutive detection cycles are overlaid in time to generate a signal sample set for the defect-free historical segment.

8. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 3, characterized in that, The design of an optimization function with signal distortion significance as the objective involves iteratively adjusting the detection threshold and window function parameters to find the signal region where the optimization function value reaches a local maximum. Specifically, this includes: The significance of signal distortion is defined as a function of the ratio of time-frequency energy to background model energy within the signal analysis window, and a penalty term for instantaneous phase changes in the signal is added. Initialize a broad detection threshold and a default rectangular window function; During the iteration process, the detection threshold is gradually narrowed and the shape and length of the window function are changed, and the signal distortion significance of each signal analysis window is recalculated; The system monitors changes in the significance value of signal distortion. When the significance value of signal distortion in a certain signal region continues to increase in several consecutive iterations and eventually stabilizes at a high level, the iteration is stopped and the signal region with a significance value higher than the preset threshold is locked.

9. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 5, characterized in that, The process of refining the boundaries of the truncated signal, removing the portions at both ends of the truncated signal that are mixed with background noise, and retaining the core response waveform, specifically includes: Calculate the short-time average energy of the signal from the center of the extracted signal outwards to both ends; Determine the location point where the short-time average energy drops to a certain proportion of the peak energy of the intercepted signal; Using these two locations as the new boundaries, discard the data points outside the boundaries to obtain the signal segments after boundary refinement.

10. The method for extracting guided wave characteristic signals of defects near the weld zone of a pipeline according to claim 6, characterized in that, The method employs a spatial grid search and consistency verification algorithm to solve the hyperbolic equation system, determining the two-dimensional circumferential and axial positions of the defect source on the pipe wall. Specifically, this includes: The pipe surface is unfolded into a two-dimensional grid with axial and circumferential coordinates; For each potential defect source point in the grid, calculate its theoretical time of arrival to each sensor; The theoretical arrival time is compared with the measured arrival time of the candidate defect characteristic signal, and the sum of squares of the time residuals of all sensors is calculated. Traverse the entire two-dimensional grid to find the grid point that minimizes the sum of squares of the time residuals, and use the grid point as the initial solution for the location of the defect source. Cross-validation is performed using time difference information from multiple sensors near the initial solution to eliminate inconsistent sensor data. The remaining data is then used to recalculate and obtain the final optimized defect source location coordinates.

Images