A pipeline defect detection method and system based on kurtosis adaptive step-by-step feature mining

Abstract

The present invention discloses a pipeline defect detection method and system based on kurtosis adaptive step-by-step feature mining, which belongs to the field of non-destructive testing processing technology. The method constructs a pipeline state recognition framework to preprocess the pipeline signal for defect recognition, wherein the features of various abnormal signals are extracted by step-by-step residual feature mining, and the abnormal signals that interfere with defect recognition are identified and eliminated in layers, so as to solve the problem that the real pipeline contains multiple abnormal signals, and the shape and size of the pipe structure signal are similar to the defect, which interferes with defect recognition and causes false detection; adaptive defect recognition is performed based on the kurtosis adjustment hyperparameter, which solves the problems of defect missed detection and low efficiency of manual parameter adjustment, and realizes fully automatic adaptive defect recognition; the odometer wheel is used for relative distance positioning to effectively reduce the cumulative error caused by the odometer wheel; at the same time, the Monte Carlo method and the Bayesian decision quantization method with minimum error rate are used to effectively solve the nonlinearity and instability of defect quantification.

Description

Technical Field

[0001] The present invention relates to the field of non-destructive testing processing technology, and in particular to a pipeline defect detection method and system based on kurtosis adaptive step-by-step feature mining. Background Art

[0002] Because pipelines are often built in complex underground environments, they are affected by factors such as soil and water quality, and the corrosive media they transport. This can lead to corrosion, such as wall thinning. Furthermore, factors such as temperature, compression, and vibration can easily cause dents and cracks. Furthermore, poor management and negligence by pipeline operators can lead to pipelines exceeding their service life, increasing safety risks. Numerous oil and gas pipeline accidents have occurred due to corrosion perforation, component damage, and extended operation. Regular pipeline health inspections are essential to prevent potential problems and promptly identify potential risks.

[0003] Eddy current testing is a nondestructive testing technology based on the theory of electromagnetic induction. When alternating current is applied to an excitation coil, an alternating magnetic field, referred to as the primary magnetic field, is generated around the coil. When this alternating magnetic field approaches a conductor, electromagnetic induction generates eddy currents in the conductor, which in turn generate an induced magnetic field, referred to as the secondary magnetic field. Defects disturb the eddy currents, which in turn affect the secondary magnetic field. Defects of varying length, width, and depth disturb the eddy currents to varying degrees, generating different secondary magnetic field intensities. The eddy current probe converts the detected secondary magnetic field into a voltage output. The acquired signal amplitude and phase carry information such as the depth of the defect, which can be used as raw data to quantify the defect.

[0004] Due to the long pipeline length and large span, the operation of the internal detector in the pipeline is complex, which brings difficulties to pipeline status / defect identification: during the pipeline manufacturing process, elbows, welds and other pipe structures appear. When the internal detector passes through these structures, the signal shape and amplitude generated are similar to the defects, which interferes with defect identification. During operation, the sensor lifting will also weaken the amplitude of the defect signal, making the defect signal more difficult to identify. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in defect recognition in the prior art and to provide a pipeline defect detection method and system based on kurtosis adaptive step-by-step feature mining. For defect recognition, a pipeline state recognition framework based on kurtosis adaptive step-by-step residual feature mining is proposed. The framework is divided into two parts. The step-by-step residual feature mining part uses feature engineering to extract the features of various abnormal state signals, hierarchically identifies and eliminates abnormal signals that interfere with defect recognition, and solves the problem that real pipelines contain a variety of abnormal signals, and the shape and size of the pipe structure signal are similar to the defects, which interfere with defect recognition and cause false detection; based on the kurtosis adaptive recognition part, the loss function with kurtosis as the framework is used to measure whether there are defects in the signal by calculating the residual signal kurtosis value and feeding it back to the model to guide the model to adjust the hyperparameters, thereby solving the problems of missed defect detection and low efficiency of manual parameter adjustment, and realizing fully automatic adaptive defect recognition.

[0006] The object of the present invention is achieved through the following technical solutions:

[0007] In a first aspect, a pipeline defect detection method based on kurtosis adaptive step-by-step feature mining is provided, the method comprising the following steps:

[0008] S1. Using an internal detector system to collect pipeline signals, the pipeline signals include eddy current signals and other abnormal signals; the other abnormal signals include IMU signals, weight wheel signals, and odometer wheel signals;

[0009] S2. Preprocessing the pipeline signal;

[0010] S3. Build a pipeline state recognition framework to perform defect recognition on the preprocessed pipeline signals. The features of various abnormal signals are extracted through step-by-step residual feature mining. Abnormal signals that interfere with defect recognition are identified and eliminated in layers. Adaptive defect recognition is performed based on kurtosis-adjusted hyperparameters.

[0011] As a preferred option, the extraction of features of various abnormal signals by level-by-level residual feature mining, and the hierarchical identification and elimination of abnormal signals that interfere with defect identification include:

[0012] For different abnormal signals, features that can effectively identify the abnormal signals are extracted, and the corresponding features are used to identify and eliminate interference signals step by step.

[0013] As a preferred option, the adaptive defect identification based on kurtosis-based hyperparameter adjustment includes:

[0014] By calculating the residual signal kurtosis value, it is measured whether there are defective signals in the abnormal signal and fed back to the model to guide the model to adjust the hyperparameters.

[0015] As a preferred option, a pipeline defect detection method based on kurtosis adaptive step-by-step feature mining further includes the steps of:

[0016] S4. Positioning the defect identified in step S3, wherein the positioning includes:

[0017] S41. Based on the pipeline construction drawings and the pipe structure as reference points, perform primary defect location.

[0018] S42. Perform secondary positioning based on the relative mileage distance between the defect and the reference point.

[0019] As a preferred option, a pipeline defect detection method based on kurtosis adaptive step-by-step feature mining further includes the steps of:

[0020] S5. Quantify the defects identified in step S3.

[0021] The quantification specifically includes:

[0022] S51. Extract quantitative features based on physical mechanisms;

[0023] S52, quantify defect depth based on Monte Carlo method;

[0024] S53. Correct the defect depth based on the minimum error rate Bayesian decision.

[0025] In a second aspect, a pipeline defect detection system based on kurtosis adaptive step-by-step feature mining is provided, comprising:

[0026] a signal acquisition module configured to acquire pipeline signals using an internal detector system, wherein the pipeline signals include eddy current signals and other abnormal signals;

[0027] a preprocessing module configured to preprocess the pipeline signal;

[0028] The defect recognition module is configured to perform defect recognition on pipeline signals preprocessed by building a pipeline state recognition framework. It extracts the features of various abnormal signals through step-by-step residual feature mining, hierarchically identifies and eliminates abnormal signals that interfere with defect recognition, and performs adaptive defect recognition based on kurtosis-based hyperparameter adjustment.

[0029] As a preferred option, a pipeline defect detection system based on kurtosis adaptive step-by-step feature mining further includes:

[0030] The defect localization module is configured to locate the defects identified by the defect recognition module, wherein the localization includes:

[0031] The first-level positioning of defects is carried out based on the pipeline construction drawings and pipe structure as reference points; the second-level positioning is carried out based on the relative mileage distance between the defect and the reference point;

[0032] The defect quantification module is configured to quantify the defects identified in the defect identification module.

[0033] The internal detector system includes an eddy current sensor, an inertial measurement unit (IMU), a weight wheel, and an odometer wheel.

[0034] It should be further explained that the technical features corresponding to the above options can be combined or replaced with each other to form a new technical solution if there is no conflict.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The present invention extracts the features of various abnormal signals through step-by-step residual feature mining, hierarchically identifies and eliminates abnormal signals that interfere with defect identification, and solves the problem that real pipelines contain multiple abnormal signals, and the shape and size of the pipe structure signals are similar to the defects, which interfere with defect identification and lead to false detection; based on the kurtosis adjustment hyperparameter, adaptive defect identification is performed to solve the problems of defect missed detection and low efficiency of manual parameter adjustment, and realize fully automatic adaptive defect identification. It will perform better in complex pipeline environments and the model will be more robust.

[0037] (2) In one example, the present invention uses the mileage wheel for relative distance positioning, effectively reducing the cumulative error caused by the mileage wheel, locating the defect between the pipe making structure sections of the pipeline, and facilitating the operation of the staff during defect detection.

[0038] (3) In one example, the Bayesian decision quantification method based on the Monte Carlo method and the minimum error rate of the present invention can still achieve good results for a small amount of defect sample data, and effectively solve the nonlinearity and instability of defect quantification from the perspective of statistical probability.

[0039] (4) The present invention can theoretically be applied to pipes of different diameters, and the model has a certain degree of versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a pipeline defect detection method based on kurtosis adaptive step-by-step feature mining according to an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of pipeline defect identification based on kurtosis adaptive step-by-step feature mining and Bayesian-based defect quantification method according to an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of the eddy current detection principle shown in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram showing the effects of elbow and ramp recognition and weld recognition according to an embodiment of the present invention;

[0044] Figure 5 This is a schematic diagram showing the effect of defect recognition according to an embodiment of the present invention;

[0045] Figure 6 This is an example of two types of defects, large and small, shown in an embodiment of the present invention, and illustrates the effects of the feature distribution map extracted by the present invention and the traditional feature distribution map;

[0046] Figure 7 This is the quantification result of the defect quantification model for actual pipeline defects shown in the embodiment of the present invention;

[0047] Figure 8 This is a diagram showing the effect of locating actual defects according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is described clearly and completely below with reference to the accompanying drawings. It is apparent that the embodiments described are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0049] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0050] Example 1

[0051] Reference Figure 1 In an exemplary embodiment, to address the problems existing in defect identification, a pipeline defect detection method based on kurtosis adaptive step-by-step feature mining is provided, the method comprising the following steps:

[0052] S1. Using an internal detector system to collect pipeline signals, the pipeline signals include eddy current signals and other abnormal signals; the other abnormal signals include IMU signals, weight wheel signals, and odometer wheel signals;

[0053] S2. Preprocessing the pipeline signal;

[0054] S3. Build a pipeline state recognition framework to perform defect recognition on the preprocessed pipeline signals. The features of various abnormal signals are extracted through step-by-step residual feature mining. Abnormal signals that interfere with defect recognition are identified and eliminated in layers. Adaptive defect recognition is performed based on kurtosis-adjusted hyperparameters.

[0055] Specifically, the internal detector system consists of three parts: hardware, sensors, and mechanical structure. The hardware part includes FPGA, ADC / DAC module and power supply module, which are responsible for generating the excitation signal required for eddy current detection, storing the detection data of the sensor part in real time, and powering the entire internal detector system. The sensor part includes four parts: eddy current sensor, inertial measurement unit (IMU), angle displacement sensor (weight wheel) and odometer wheel. Among them, the eddy current probe (eddy current sensor) is based on electromagnetic induction theory and is responsible for detecting defects in the pipeline, abnormal conditions such as pipe structure, and providing information for defect quantification. An internal detector generally includes front and rear end probe assemblies. Each end assembly is composed of a circumferential array of multiple eddy current probes. The number of probes is adjusted according to the different diameters of the detection pipe, and the front and rear end probes are distributed with a certain offset angle to achieve full coverage of the inner wall of the pipeline as much as possible.

[0056] The IMU is responsible for measuring the internal detector's operating posture in the pipeline to determine where bends, ramps, etc. exist in the pipeline. It includes at least two IMU sensors, which fuse the signals of the two sensors and correct the detection data. The weight wheel is responsible for measuring the rotation angle of the internal detector during operation in the pipeline, thereby determining the clock position of the defect. IMU measurement information includes three-axis acceleration, three-axis angular velocity, and quaternion. The weight wheel measurement information is the number of output pulses. The odometer wheel is responsible for measuring the operating distance of the internal detector and locating the identified defects. Depending on the diameter of the detection pipeline, it generally includes 2-4 odometer wheels, and the results of multiple odometer wheels are integrated to achieve odometer correction.

[0057] First, the internal detector system uses a variety of sensors to measure relevant raw data. Eddy current sensors detect defects, pipe structure anomalies, and other abnormalities in the pipeline, providing information for defect quantification. The IMU measures the internal detector's position within the pipeline to identify bends, ramps, and other locations. The weight wheel measures the internal detector's rotation angle during operation, determining the defect's chronograph position within the pipeline. The odometer measures the distance traveled by the internal detector to locate the defect.

[0058] The sensor data is mined step by step to identify static sections, pipeline structures (welds, elbows, ramps), and defects in layers; adaptive and accurate defect identification is achieved based on kurtosis.

[0059] Furthermore, the signals collected by the above sensors are pre-processed. Since the internal detector vibrates in the pipeline and produces different lift-off heights, the signal has a baseline offset, which affects the change trend of the eddy current signal and causes some defect signals to change unclearly. Relevant research has shown that the amplitude is greatly affected by the lift-off. Since the phase is the phase difference between the two coils, the lift-off signal has little effect on the phase differential signal. Therefore, the amplitude signal and the phase signal can be fused to suppress the effect of lift-off. The detection signal output by the eddy current probe in this article is obtained by Fourier transforming the time series signal, which includes amplitude and phase, and can be expressed as Fusion of two signals, where A is the amplitude, is the phase, α is the bias, and the amplitude and phase are used to fuse them into a timing signal.

[0060] The quaternion output by the IMU is used to calculate the attitude change of the internal detector when it is running in the pipeline using formula (1-1). The attitude change of the internal detector can be described by Euler angles. Euler angles are a set of independent angular vectors used to determine the position of a fixed-point rotating rigid body. They are composed of roll angle (Roll), pitch angle (Pitch), and yaw angle (Yaw), represented by φ, θ, and ψ respectively. For the attitude description of the internal detector in the pipeline, the roll angle is the rotation angle of the internal detector in the pipeline; the pitch angle is the angle of the internal detector when it passes through uphill and downhill slopes when running in the pipeline; and the yaw angle is the angle of the internal detector when it passes through a bend when running.

[0061]

[0062] Every time the odometer wheel rotates one circle, it will output a fixed number of pulses. The number of pulses is recorded by the encoder built into the odometer wheel. The radius r of the odometer wheel can then be used to determine the travel distance of the internal detector system. The specific calculation formula is shown in formula (1-2).

[0063]

[0064] Where n is the number of pulses recorded by the odometer wheel, f is the number of pulses output when the odometer wheel rotates one circle, and r is the radius of the odometer wheel.

[0065] Furthermore, since the detection signal contains a variety of abnormal signals, it interferes with defect identification and causes false detection. The existing recognition algorithm relies on empirical values ​​to give the values ​​of the model hyperparameters. Non-optimal values ​​of the hyperparameters lead to missed defects and cannot achieve adaptive recognition. The present invention proposes a step-by-step residual feature mining part. The main idea is to extract features that can effectively identify different abnormal signals, use the corresponding features to identify and eliminate interference signals step by step, and identify defects at the last level. The recognition task of each level can be regarded as a two-classification task. Category one is the target signal represented by the features of this level, and category two is the remaining signal (residual). After recognition, the target signal of category one is eliminated from the signal. The later the level, the fewer types of abnormal signals it contains. The defect signal is identified at the last level to reduce the interference of other abnormal signals.

[0066] The odometer is used to record the running distance of the internal detector system. When the internal detector becomes stationary due to the "ball stuck" phenomenon, the running distance increment Δs = 0 and the change in the IMU Euler angle Δθ = 0. This feature can completely remove the stationary signal.

[0067] When the internal detector passes over a weld, the same set of sensors simultaneously generates an abnormal weld signal. Consider a probe set with N sensors, each outputting L data points. This matrix is ​​then processed using principal component analysis (PCA), with the first principal component used as the feature for identifying weld signals. This feature extraction increases the contrast between weld signals and normal signals, effectively reducing the difficulty of identification.

[0068] The IMU uses the principle of attitude calculation for the inner detector. When the inner detector passes through a bend, the IMU's yaw output will undergo a large sudden change; when passing over a slope, the IMU's pitch output will also undergo a large sudden change. The signal value is much larger than the preceding and following signal segments. The standard deviation of the sliding window of the two signals is calculated to obtain the corresponding characteristic signal. At this time, interference signals are identified and eliminated through each characteristic.

[0069] Furthermore, the adaptive defect identification based on kurtosis-based hyperparameter adjustment includes:

[0070] By calculating the kurtosis of the residual signal, we assess whether there are defects within the abnormal signal and provide feedback to the model, thereby guiding the adjustment of hyperparameters. The key idea behind this step is that the statistical distribution of normal signals is normal, while abnormal signals can cause the signal to deviate from this distribution. Kurtosis measures the degree of deviation from the normal distribution. Calculating the kurtosis of the residual signal reflects the presence and abundance of abnormal signals. This, analogous to the loss function in machine learning, allows us to evaluate the model's effectiveness and guide parameter adjustments, achieving the goal of adaptive and automatic defect identification.

[0071] Example 2

[0072] The internal detector system is equipped with an odometer, which records the operating position of the internal detector system. However, due to the complex interior of the pipeline and the high-speed operation of the internal detector inside the pipeline, the odometer slippage leads to odometer loss, resulting in significant errors in defect location. Based on Example 1, a pipeline defect detection method based on kurtosis adaptive hierarchical feature mining is provided, which also includes the following steps:

[0073] S4. Positioning the defect identified in step S3, wherein the positioning includes:

[0074] S41. Based on the pipeline construction drawings and the pipe structure as reference points, perform primary defect location.

[0075] S42. Perform secondary positioning based on the relative mileage distance between the defect and the reference point.

[0076] Specifically, defect location requires two parameters: the specific distance (axial positioning) and the specific time (circumferential positioning) of the defect in the pipeline. The odometer wheel is responsible for detecting the specific distance, while the IMU and weight wheel are responsible for detecting the specific time. A two-stage positioning method is proposed. First, a preliminary first-level positioning of the defect is performed using pipeline structural drawings and piping structures such as elbows and welds as reference points. Secondary, precise positioning is then performed using the relative odometer distance between the defect and the reference point. This method addresses the problem of defect location errors caused by slippage of a single odometer wheel.

[0077] Example 3

[0078] Only the amplitude and phase of the detection signal can represent depth information, which contains relatively little information. In real pipelines, the amplitude and phase are affected not only by depth but also by factors such as sensor liftoff and defect shape. This results in a highly nonlinear relationship between the amplitude and phase of the defect and depth, making it difficult to quantify the depth of the defect. Therefore, based on Example 1, a pipeline defect detection method based on kurtosis adaptive step-by-step feature mining is provided, which also includes the following steps:

[0079] S5. Quantify the defects identified in step S3.

[0080] The quantification specifically includes:

[0081] S51. Extract quantitative features based on physical mechanisms;

[0082] S52, quantify defect depth based on Monte Carlo method;

[0083] S53. Correct the defect depth based on the minimum error rate Bayesian decision.

[0084] Specifically, using differential amplitude and differential phase as two-dimensional features, a KNN algorithm was used to determine the defect depth in a single simulation, and a Monte Carlo method was used to determine the defect depth in multiple simulations. Finally, a Bayesian decision-making algorithm with minimum error rate was used to correct and fuse the depths of pipes with different diameters to obtain the final defect depth.

[0085] Because eddy current sensors provide relatively few effective raw signals that can characterize defect depth, and available labeled datasets are limited, various random factors cause the amplitude and phase signals to become complex, multi-source mixed signals. This paper first analyzes the physical mechanism of eddy current testing. Based on the support of the internal detector system, it extracts features to reduce the interference of multiple factors. From a probabilistic reasoning perspective, it fully utilizes labeled datasets from laboratory simulated pipelines and uses the Monte Carlo method to determine the depth of defects in different datasets. The results from these different datasets are then fused based on prior information in the datasets to form the defect depth result.

[0086] Step S51 includes:

[0087] The eddy current sensor consists of two excitation coils and one detection coil. Let A1 and A2 be the voltage amplitude generated by the detection coil due to the change of magnetic flux of the excitation coil 1 and the excitation coil 2. is the corresponding phase value, are the voltage amplitudes of the eddy currents generated on the specimen and acting on the detection coil, is the corresponding phase value. Then the sensor outputs the amplitude A and the phase As shown in formulas (1-3) and (1-4).

[0088]

[0089]

[0090] Taking the baseline values ​​on both sides of the flaw signal as reference values, the differential peak is the difference between the flaw signal peak and the reference value. The same applies to the differential phase. Using differential amplitude-differential phase as the characteristic for flaw quantification reduces interference such as eddy current probe differences and lift-off.

[0091] Step S52 includes:

[0092] The artificially simulated pipelines consist of various diameters, each containing defects of varying lengths, widths, and depths. Each diameter is considered a scenario, and multiple simulated inspection experiments are conducted using an internal detector. During these repeated simulations, each defect is randomly inspected multiple times by different sensors in different ways. The resulting amplitude-phase distribution reflects the coupling of multiple random factors, including sensor variability and the effective inspection area of ​​both the sensor and the defect. In actual inspections, these factors, in addition to depth, can also affect the amplitude and phase. When a sufficient number of simulations are performed, the resulting amplitude-phase distribution can be considered an approximate distribution of the amplitude and phase with respect to depth, taking into account these multiple factors. This distribution serves as the basis for quantifying the depth. Using KNN, the amplitude and phase of the defect to be quantified are used to determine the defect depth for each simulation. Monte Carlo analysis is then used to determine the defect depth for each simulation.

[0093] Step S53 includes:

[0094] When using the Bayesian decision with minimum error rate to solve specific problems, let P(w i |x) is class w i The posterior probability density curve, p(x|w i ) is class w i Class conditional probability density curve of . Classification errors include two cases.

[0095] (1) Sample x∈w1 is classified as w2.

[0096] (2) Sample x∈w2 is classified as w1.

[0097] Let P1(e) be the probability of a defect being misclassified, as shown in formula (1-5); P1(c) be the probability of a defect being correctly classified, as shown in formula (1-6). i ) is class w i The class conditional probability density curve of .

[0098]

[0099] Assuming that the amplitude and phase of the same type of deep defects obey the two-dimensional normal distribution, after determining the form of the probability density function, the maximum likelihood estimation can be used to calculate the unknown parameter values ​​of the probability density function. Then, p(x|w i ) expression. Suppose that the depth quantification results of the defect by the Monte Carlo method in M ​​simulation experiments are a set d = {d 1 ,d 2 ,…,d M}, then formula (1-7) can be used to fuse the M results, and d is the final defect of the depth.

[0100] Furthermore, the effectiveness of the model is verified by performing defect identification, location, and quantification on pipeline data obtained from simulated pipelines and actual pipelines in different scenarios. The model's predicted results are compared with the actual results to verify the effectiveness of the model.

[0101] Example 4

[0102] Reference Figure 2-3 Based on Examples 1-3, a pipeline defect identification method based on kurtosis adaptive step-by-step feature mining and a Bayesian-based defect quantification method are provided. Furthermore, for multi-sensor signal preprocessing, during signal preprocessing, a single odometer wheel will cause a large cumulative error in positioning, and the data of multiple odometer wheels need to be fused. The fusion of odometer wheel data is based on the fact that the internal detector is generally equipped with multiple odometer wheels evenly installed along the circumference of the pipeline. The detection results of multiple odometer wheels are fused to reduce part of the cumulative error. When calculating the distance based on the number of pulses, the odometer wheel with the largest increase in the number of pulses is selected for calculation to achieve the fusion of odometer wheel data. Next, the preprocessed signal is subjected to sliding window processing. The parameters that need to be paid attention to are the number of points included in the window and the coverage rate of the window.

[0103] Furthermore, a pipeline status judgment model is constructed:

[0104] For the curve and ramp identification model, the heading angle is the angle of the internal detector when it passes through a bend, and the pitch angle is the angle of the internal detector when it passes through an uphill or downhill slope. By calculating the standard deviation of these two angles, the isolation forest algorithm can be used to identify bends and ramps. Figure 4 Demonstrates the effect of elbow and ramp recognition.

[0105] For the weld seam discrimination model, the remaining signal after removing the signal identified in the previous step is used as the input signal for this layer. Assume a probe set has N sensors, each outputting L data points. This matrix is ​​considered an N×L matrix. Principal Component Analysis (PCA) is used to process this matrix, and the first principal component is used as the characteristic for identifying weld seam signals. The triple standard deviation method can be used to identify weld seams. Figure 4 Demonstrates the effect of weld recognition.

[0106] For the defect discrimination model, the target signal identified in the previous step is removed and the remaining signal is used as the input signal of this layer. The signal is windowed and the triple standard deviation method is used to identify some defects. Figure 5 (a) is the input signal of the defect layer after the abnormal signals except defects are identified and eliminated by the step-by-step feature framework. Figure 5(b) shows the initial defect recognition results. In the previous step, the adaptive defect recognition model was able to identify some defects with larger signal amplitudes, but missed defects with smaller signal amplitudes. In the above model, the multiple of the standard deviation is a hyperparameter. The identified signals are removed, and the kurtosis of the remaining signals is calculated. If the kurtosis K is greater than 3, the model hyperparameter is adjusted to further identify defects with smaller amplitudes. Figure 5 (c) shows the change of kurtosis value during the model hyperparameter adjustment process. Figure 5 (d) shows the final defect recognition effect after the kurtosis-guided model performs hyperparameter adjustment, and all defects are identified.

[0107] Furthermore, a defect depth quantification model is constructed:

[0108] Extraction of quantitative features: The amplitude and phase of the defect signal can be used as the raw data for quantifying the defect depth. However, the two are a complex mixed signal from multiple sources and are interfered by multiple factors. Based on the hardware conditions of the present invention, the two influencing factors that can be reduced are sensor difference and lift-off height. The sensor of the present invention includes a detection coil and two excitation coils. When the voltage amplitudes of the eddy currents generated on the test piece acting on the detection coils are equal, the amplitude A and phase A of the sensor output are equal. As shown in formulas (1-3) and (1-4).

[0109] A=A1-A2 (1-8)

[0110]

[0111] Sensor variability manifests itself in the fact that the asymmetric manufacturing of the two excitation coils results in unequal outputs from different eddy current sensors when there are no defects. Liftoff also manifests itself in the fact that when the same sensor is inspecting a defect-free specimen, the sensor output varies with liftoff. Using the baseline values ​​on both sides of the defect signal as reference values, the differential peak is the difference between the peak value of the defect signal and the reference value. The same applies to the differential phase. The differential amplitude-differential phase is used as the characteristic for defect quantification. Figure 6 The results of the feature distribution diagrams extracted by the present invention and the traditional feature distribution diagram for two types of defects, large and small, are shown. The features proposed by the present invention make the distribution of the two types of defects have a clear inter-class interval, which is beneficial for subsequent quantification.

[0112] The depth information of defects in different pipe diameters is obtained based on the Monte Carlo method: each pipe diameter is used as a scenario, and multiple simulation detection experiments are performed on each scenario using the internal detector. After a single detection of the simulated pipeline, the differential amplitude-differential phase information of the pipeline defect is obtained. The KNN algorithm can be used to find the k defects closest to the defect to be quantified on the amplitude-phase distribution plane. The depth category of the defect is obtained by the KNN algorithm, and the average of the k defect depths is used as the depth value of the defect obtained in this simulation experiment. Repeat the simulation experiment multiple times, and assume that the internal detector system performs N pulling tests on the pipe diameter. Assume N wi For multiple classification defects, the defects are classified as w i The number of times of the class, the category of the defect can be determined by formula (1-11), and the average value of the depth obtained by N experiments is used as the depth result of the defect simulation in the pipe diameter. The formula is as follows: m is the depth value obtained in the mth pipe diameter.

[0113] Then x∈ω i (1-11)

[0114] Repeat the above quantification steps in pipes with different diameters to obtain the quantified depth of the defect in different pipe diameters.

[0115] Fusion of defect quantification results based on Bayesian decision making: The Monte Carlo method will obtain different depths of defects in data sets with different pipe diameters. The following describes how to fuse different depths to obtain the final defect depth. Since the amplitude and phase distributions of each type of defect are mixed together in the two-dimensional amplitude and phase distribution diagram obtained from a simulation experiment, there is no obvious boundary between classes. The amplitude and phase distributions of each type of defect have a certain overlap. The overlapping part indicates that the depth of the defect of this amplitude may belong to either a large class or a medium class. According to the Monte Carlo method, the category and depth value of the defect are uniquely determined, so there will be a certain probability of misclassification. This misclassification rate can be characterized by the minimum error rate Bayesian decision. Assume that in the previous step, the defect was classified as w i , then the probability that the defect is classified correctly is p i (c) Assume that the depth quantization results of the defect in M ​​scenarios are a set d = {d 1 ,d 2 ,…,d M}, then formula (1-7) can be used to fuse the M results. Figure 7 The quantification results of the defect quantification model on actual pipeline defects are shown.

[0116]

[0117] Furthermore, for each identified defect and pipe structure such as welds and elbows, the mileage wheel fusion data will give an absolute distance relative to the starting point of the pipe, including the cumulative error. This absolute distance is used in conjunction with the pipe construction drawings to determine the elbows and ramps as reference points to locate the welds, and then the welds are used as reference points to locate the defects. The distance between two adjacent welds is a pipe section, and a pipe section is generally about 10m. Let the axial distance to the left girth weld of the pipe section where the defect is located be S weld1 The axial distance of the right side ring weld is S weld2 , the axial distance of the defect is S flaw The relative distance S between the defect and the left girth weld can be calculated. relative .

[0118] S relative =S weld1 -S flaw (1-14)

[0119] For circumferential positioning of defects, a weight wheel is used for detection. Given the hardware support of the internal detector system in this article, the clock position of the sensor that detects the defect on the pipeline is used as the circumferential positioning of the defect. During pipeline inspection, the probe number at the 12 o'clock position, or directly above the detector system, is recorded when the internal detector system is placed in the pipeline. Before testing, the weight wheel's detection value is reset to zero. During operation, the weight wheel's detection angle is the angle of the probe at 12 o'clock. The angle values ​​of other probes are the product of the number of probes separated from the probe by the number of adjacent probe angles. Figure 8 The effect diagram of field defect positioning is shown.

[0120] Example 5

[0121] In another exemplary embodiment, based on the same technical concept as in Example 1, a pipeline defect detection system based on kurtosis adaptive level-by-level feature mining is provided, comprising:

[0122] a signal acquisition module configured to acquire pipeline signals using an internal detector system, wherein the pipeline signals include eddy current signals and other abnormal signals;

[0123] a preprocessing module configured to preprocess the pipeline signal;

[0124] The defect recognition module is configured to perform defect recognition on pipeline signals preprocessed by building a pipeline state recognition framework. It extracts the features of various abnormal signals through step-by-step residual feature mining, hierarchically identifies and eliminates abnormal signals that interfere with defect recognition, and performs adaptive defect recognition based on kurtosis-based hyperparameter adjustment.

[0125] Furthermore, the system also includes:

[0126] The defect localization module is configured to locate the defects identified by the defect recognition module, wherein the localization includes:

[0127] The first-level positioning of defects is carried out based on the pipeline construction drawings and pipe structure as reference points; the second-level positioning is carried out based on the relative mileage distance between the defect and the reference point;

[0128] The defect quantification module is configured to quantify the defects identified in the defect identification module.

[0129] The internal detector system includes an eddy current sensor, an inertial measurement unit (IMU), a weight wheel, and an odometer wheel.

[0130] The above specific implementation methods are detailed descriptions of the present invention. It cannot be considered that the specific implementation methods of the present invention are limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions and substitutions without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A pipeline defect detection method based on kurtosis adaptive step-by-step feature mining, characterized in that: The method comprises the following steps: S1. Using an internal detector system to collect pipeline signals, the pipeline signals include eddy current signals and other abnormal signals; the other abnormal signals include IMU signals, weight wheel signals, and odometer wheel signals; S2. Preprocessing the pipeline signal; S3. Constructing a pipeline state recognition framework to perform defect recognition on the preprocessed pipeline signal, wherein the features of various abnormal signals are extracted through step-by-step residual feature mining, abnormal signals that interfere with defect recognition are identified and eliminated in layers, and adaptive defect recognition is performed based on kurtosis adjustment hyperparameters. The step-by-step residual feature mining to extract the features of various abnormal signals and the step-by-step residual feature mining to identify and eliminate abnormal signals that interfere with defect recognition include: For different abnormal signals, extract the features that can effectively identify the abnormal signal, and use the corresponding features to identify and eliminate the interference signal layer by layer; The adaptive defect identification based on kurtosis-based hyperparameter adjustment includes: By calculating the residual signal kurtosis value, it is measured whether there are defective signals in the abnormal signal and fed back to the model to guide the model to adjust the hyperparameters.

2. The pipeline defect detection method based on kurtosis adaptive step-by-step feature mining according to claim 1 is characterized in that: Also includes the steps: S4. Positioning the defect identified in step S3, wherein the positioning includes: S41. Based on the pipeline construction drawings and the pipe structure as reference points, perform primary defect location. S42. Perform secondary positioning based on the relative mileage distance between the defect and the reference point.

3. The pipeline defect detection method based on kurtosis adaptive step-by-step feature mining according to claim 1 is characterized in that: Also includes the steps: S5. Quantify the defects identified in step S3.

4. The pipeline defect detection method based on kurtosis adaptive step-by-step feature mining according to claim 3 is characterized in that: The quantification specifically includes: S51. Extract quantitative features based on physical mechanisms; S52, quantify defect depth based on Monte Carlo method; S53. Correct the defect depth based on the minimum error rate Bayesian decision.

5. A pipeline defect detection system based on kurtosis adaptive step-by-step feature mining, characterized in that: include: a signal acquisition module configured to acquire pipeline signals using an internal detector system, wherein the pipeline signals include eddy current signals and other abnormal signals; The other abnormal signals include IMU signals, weight wheel signals and mileage wheel signals; a preprocessing module configured to preprocess the pipeline signal; The defect recognition module is configured to perform defect recognition on pipeline signals preprocessed by constructing a pipeline state recognition framework, wherein the features of various abnormal signals are extracted through level-by-level residual feature mining, abnormal signals that interfere with defect recognition are identified and eliminated in layers, and adaptive defect recognition is performed based on kurtosis adjustment hyperparameters. The step-by-step residual feature mining is used to extract the features of various abnormal signals, and the step-by-step residual feature mining is used to identify and eliminate abnormal signals that interfere with defect recognition in layers, including: For different abnormal signals, extract the features that can effectively identify the abnormal signal, and use the corresponding features to identify and eliminate the interference signal layer by layer; The adaptive defect identification based on kurtosis-based hyperparameter adjustment includes: By calculating the residual signal kurtosis value, it is measured whether there are defective signals in the abnormal signal and fed back to the model to guide the model to adjust the hyperparameters.

6. The pipeline defect detection system based on kurtosis adaptive step-by-step feature mining according to claim 5 is characterized in that: Also includes: The defect localization module is configured to locate the defects identified by the defect recognition module, wherein the localization includes: The first-level positioning of defects is carried out based on the pipeline construction drawings and pipe structure as reference points; the second-level positioning is carried out based on the relative mileage distance between the defect and the reference point; The defect quantification module is configured to quantify the defects identified in the defect identification module.

7. The pipeline defect detection system based on kurtosis adaptive step-by-step feature mining according to claim 5 is characterized in that: The internal detector system includes an eddy current sensor, an inertial measurement unit (IMU), a weight wheel, and an odometer wheel.

Images