Pipeline vortex-magnetic flux leakage aliasing signal decoupling and processing method

Abstract

The invention relates to the technical field of oil and gas pipeline detection, in particular to a pipeline eddy current-magnetic leakage aliasing signal decoupling and processing method, which comprises the following steps: S101, collecting pipeline defect aliasing original signals by using a coil and a magneto-sensitive element; s102, performing frequency division through low-pass and band-pass filters, extracting direct-current and alternating-current high-frequency and low-frequency components, and inputting the components into a CNN model to extract peak values and defect type labels; s103, judging the direction and position of the defect: judging whether the defect is located in a near inner wall area or a far inner wall area of the pipeline wall according to a voltage signal of a detection coil when high and low frequency current is introduced into an excitation coil, and judging the position of the volume type defect in combination with a defect type label; and S104, establishing a correlation curve of various defects and the voltage signal intensity, comparing signal peak values, and calculating the actual depth of the defects through interpolation. According to the invention, only one coil is adopted for excitation, two detection methods of eddy current and magnetic flux leakage can be realized at the same time, the design size of the detector is reduced, and the trafficability of a complex pipeline is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas pipeline detection, and particularly relates to a pipeline eddy current-magnetic flux leakage mixed signal decoupling and processing method. BACKGROUND

[0002] In the fields of oil, natural gas, municipal water supply, etc., the pipeline is the core carrier for fluid transportation, and the inner wall and outer wall of the pipeline are prone to defects due to corrosion, wear, fatigue and other factors. If not detected in time, it may cause major safety accidents such as leakage and rupture. Therefore, accurate detection of pipeline defects is crucial to ensure the safety of industrial production and the stability of infrastructure. At present, eddy current detection and magnetic flux leakage detection are the two most widely used technologies in pipeline defect detection. The eddy current detection is used for defects near the inner and outer walls of the pipeline wall, but it has low sensitivity for defects buried deep in the pipeline wall. The magnetic flux leakage detection is based on the principle of magnetic field leakage and can realize quantitative analysis of pipeline wall thickness reduction defects. The combination of the two forms an eddy current-magnetic flux leakage composite detection technology, which can meet the requirements of defect position determination and size quantification, and has become a mainstream detection scheme.

[0003] In the prior art, at present, the German ROSEN company is a commonly used manufacturer in the field of pipeline internal detection services worldwide. It adopts a high-frequency eddy current and permanent magnet magnetic flux leakage combined method. By using double sensors, the company's self-developed signal processing algorithm can cross-verify and complementarily analyze the double-source data. At present, this technology has been widely applied to the internal detection of long-distance oil and gas pipelines worldwide.

[0004] However, in the actual use of the pipeline internal detection technology of the German ROSEN company, the following limitations still exist: it adopts a high-frequency eddy current and permanent magnet magnetic flux leakage combined method, which requires the configuration of a high-frequency eddy current excitation source and a permanent magnet magnetic flux leakage excitation device, resulting in a complex structure of the detection probe and a large overall size, poor passability in small-diameter pipelines (such as DN80 and below) or complex pipelines with many bends and valves; at the same time, the double excitation sources need to be powered independently, increasing the load pressure of the detector power system and limiting the single detection endurance; in addition, although the signals collected by the double sensors are fused by the algorithm, the signals are prone to interference due to the independent existence of the excitation sources, and characteristic mixing may occur in defect-intensive areas, affecting the detection accuracy. Therefore, it is necessary to propose a pipeline eddy current-magnetic flux leakage mixed signal decoupling and processing method to solve the above problems. SUMMARY

[0005] To solve the above problems, the present application provides a pipeline eddy current-magnetic flux leakage mixed signal decoupling and processing method, which can realize both eddy current and magnetic flux leakage detection methods by using only one AC coil excitation, reducing the design size of the detector, facilitating the integration of the electromagnetic package in the later stage, reducing the load pressure of the power system, and improving the passability of complex pipelines.

[0006] In order to achieve the above object, the technical scheme of the present application is as follows: a pipeline eddy current-magnetic flux leakage mixed signal decoupling and processing method, comprising the following steps:

[0007] S101, signal acquisition: using a coil and a magnetic sensitive element to pick up the eddy current-magnetic flux leakage mixed original signal caused by pipeline defects.

[0008] S102, signal separation: through low-pass filter and band-pass filter, the mixed original signal is processed by frequency separation, and the direct current component signal, the alternating current high frequency component signal and the alternating current low frequency component signal are extracted respectively; then, the direct current component signal, the alternating current high frequency component signal and the alternating current low frequency component signal are input into the CNN neural network model to extract the signal peak value and the defect type label of the direct current component.

[0009] S103, defect position determination: based on the skin effect principle and signal characteristics, the defect direction dimension and spatial dimension are determined.

[0010] Direction determination: if one of the circumferential type and volume type defect labels output by the CNN neural network model is greater than the preset threshold, it is determined that the defect is circumferential type or volume type.

[0011] If one of the signal peak value amplitudes of the alternating current high frequency component and the alternating current low frequency component is greater than the preset threshold, and the direct current component has no circumferential type label, it is determined that the defect is axial type.

[0012] Spatial position determination: the excitation coil and the detection coil on the detection device maintain a preset distance from the pipeline wall, the signal generator on the detection device is used to input high frequency alternating current into the excitation coil, so that a high frequency alternating electromagnetic field is generated around the excitation coil, at this time, when the detection coil picks up the voltage signal, it is determined that the defect is located in the near inner wall area of the pipeline wall.

[0013] When the excitation coil inputs high frequency alternating current, the detection coil has no induced voltage, then the high frequency alternating current is adjusted to low frequency alternating current, if the detection coil can pick up the voltage signal at this time, it is determined that the defect is located in the far inner wall area of the pipeline wall.

[0014] If the direct current component signal peak value and the alternating current low frequency component peak value appear synchronously, it is determined as a volume type defect in the far inner wall area of the pipeline wall; if the direct current component signal peak value and the alternating current high frequency component peak value appear synchronously, it is determined as a volume type defect in the near inner wall area of the pipeline wall.

[0015] S104, defect size quantification: using a DNN neural network model to establish a correlation curve between the circumferential type and volume type defects corresponding to the direct current component and the voltage signal strength, and a correlation curve between the pipeline wall near the inner wall area and the pipeline wall far from the inner wall area axial type defects corresponding to the alternating current component and the voltage signal strength; after the defect determination is completed, the corresponding signal peak value extracted in S102 is compared with the correlation curve, and the actual depth of the defect is determined by interpolation calculation.

[0016] Further, in S101, the magnetic sensor is one of a magnetoresistance sensor and a Hall sensor.

[0017] Further, in S102, the direct current signal component corresponds to the magnetic flux leakage detection signal; the alternating current high frequency component signal and the alternating current low frequency component signal correspond to the eddy current detection signal.

[0018] Further, in S102, the method for extracting the signal peak value by the CNN neural network model includes the following steps:

[0019] S201, data set preprocessing: 5000 groups of eddy current high frequency, eddy current low frequency and magnetic flux leakage low frequency signals of different defects of the pipeline are collected, the time stamp and amplitude of the peak value are manually labeled, and the eddy current high frequency, eddy current low frequency and magnetic flux leakage low frequency signals of different defects are divided into a training set and a verification set according to a ratio of 7:2.

[0020] S202, CNN neural network model construction: input the training set into the CNN neural network model for deep learning training, use mean square error (MSE) as the loss function during training, and adjust the dropout rate of the full connection layer and the L2 regularization coefficient; at the same time, the performance of the CNN neural network model is monitored in real time using the verification set, if the loss of the verification set does not decrease continuously for 10 epochs, the adjustment of the parameters of the CNN neural network model is stopped, and the trained CNN neural network model is obtained.

[0021] Wherein, the calculation formula of the mean square error (MSE) as the loss function is:

[0022] (1).

[0023] Wherein, is the peak value coordinate timestamp, is the peak value coordinate amplitude, is the predicted peak value coordinate timestamp, is the predicted peak value coordinate amplitude, and N is the sample number.

[0024] S203, peak extraction: the AC high frequency component signal and the AC low frequency component signal in S102 are input to the CNN neural network model, and the CNN neural network model outputs the signal peak value of the AC high frequency component signal and the AC low frequency component signal; the signal peak value includes the timestamp and amplitude value of the peak coordinate.

[0025] Further, the detection device comprises a detector body and a controller, the detector body is provided with a walking wheel in the circumference, and the detection coil and the excitation coil are fixedly connected to the detector body in the circumference.

[0026] The signal generator is embedded in the detector body, and the controller is used for controlling the signal generator to input an alternating current to the excitation coil and receiving the voltage signal picked up by the detection coil.

[0027] Further, in S103, when the detection coil picks up the voltage signal, the voltage signal is picked up in a time-sharing triggering mode of high frequency band and low frequency band.

[0028] Further, in S103, the high frequency band alternating current frequency is set to 100-200 kHz, and the low frequency band alternating current frequency is set to 1-10 kHz; after the detection coil picks up the voltage signal, the voltage signal is processed by a wavelet threshold denoising algorithm, a db6 wavelet base is selected for 3-layer decomposition, and the calculation formula is:

[0029] (2).

[0030] Wherein, T is the threshold value of the wavelet threshold denoising algorithm, N is the number of signal sampling points.

[0031] Further, in S104, the DNN neural network model is used to establish the correlation curve between the different depth defects of the near inner wall area of the pipeline wall and the far inner wall area of the pipeline wall and the corresponding voltage signal intensity, and the specific method is:

[0032] S301, data set preprocessing: collecting voltage signal samples of different depth defects of the near inner wall area of the pipeline wall and the far inner wall area of the pipeline wall, wherein the defect depth of the near inner wall area of the pipeline wall covers 0.5-5 mm, and the defect depth of the far inner wall area of the pipeline wall covers 0.5-8 mm, 3 parallel test blocks are prepared for each depth, 20 voltage signals are repeatedly collected for each parallel test block, and finally 5000 groups of samples of the near inner wall area defect signal of the pipeline wall and the far inner wall area defect signal of the pipeline wall are formed.

[0033] Meanwhile, the input and output characteristics are standardized, the data is mapped to the [0, 1] interval by using Min-Max standardization, and the calculation formula is:

[0034] (3).

[0035] Wherein, It is one of the following: defect depth and voltage signal strength. This is the minimum value of the feature. The maximum value of this feature. This is the standardized data.

[0036] S302, DNN Neural Network Model Construction: The training sets of defect signals from the near-inner wall region and the far-inner wall region of the pipe are input into two independent DNN neural network models for deep learning training. Training parameters are set for the DNN neural network models, with the root mean square error (RMSE) as the loss function. The calculation formula is as follows:

[0037] (4).

[0038] in, For loss function, The number of samples; The actual voltage signal strength, This represents the voltage signal strength predicted by the model.

[0039] S303, Correlation Curve Generation: The trained DNN neural network models for the near-inner wall region and the far-inner wall region of the pipe wall are used to generate correlation curves between the corresponding defect depth and voltage signal intensity, as detailed below:

[0040] Within the defect depth range of 0.5-5mm in the near-inner wall region of the pipe wall and the defect depth range of 0.5-8mm in the far-inner wall region of the pipe wall, 100 uniformly spaced depth values ​​are selected for each region. After standardization, these values ​​are input into the corresponding DNN neural network model to obtain the predicted voltage signal intensity value for each depth value.

[0041] Using the depth value and the corresponding predicted voltage signal intensity as coordinate points, a smooth curve is plotted using a polynomial fitting method. This curve represents the correlation between defects of different depths and their corresponding voltage signal in the near-inner wall region and the far-inner wall region of the pipe.

[0042] Furthermore, in S104, the method for determining the actual depth of the defect through interpolation calculation is as follows:

[0043] S401, Correlation Curve Matching and Signal Peak Extraction: Based on the defect location determined in S103, select the correlation curve output by the DNN neural network model at the corresponding location; at the same time, extract the corresponding signal peak output by the CNN neural network model in S102.

[0044] If the defect is located in the area near the inner wall of the pipe, the amplitude of the high-frequency signal peak is extracted.

[0045] If the defect is located in the region far from the inner wall of the pipe, the amplitude of the low-frequency signal peak is extracted and used as the reference value for the voltage signal strength to be interpolated.

[0046] S402, Interpolation Interval Positioning: On the selected correlation curve, find the two coordinate points closest to the reference value of the voltage signal strength to be interpolated: Let the voltage signal strength to be interpolated be... The set of coordinate points of the associated curve Find satisfaction in

[0047] S403, Linear interpolation calculation:

[0048] .

[0049] .

[0050] Furthermore, in S104, the signal peak value is preprocessed before the interpolation calculation. Specifically, outliers are removed using the 3σ criterion, and then smoothed using a moving average filter. The formula is as follows:

[0051] (6).

[0052] in, This is the original peak sequence. This is the filtered sequence; subsequently, cubic spline interpolation is used, with the interpolation interval set as follows: interpolation function satisfy Furthermore, its second derivative is continuous within the interval, and its piecewise expression is:

[0053] (7).

[0054] in, The interpolation node spacing is set to 0.05 mm, and boundary conditions are added. .

[0055] The above approach has the following beneficial effects:

[0056] 1. This invention uses only one AC coil for excitation, which can simultaneously realize two detection methods: eddy current and leakage magnetic flux. The reduced detector design size facilitates the integration of electromagnetic packages, reduces the load on the power supply system, and improves the throughput of complex pipelines.

[0057] 2. This invention adopts a multi-frequency excitation strategy - a combination of high and low frequencies - to achieve defect location identification, with flexible frequency selection, and can assist in the quantification of defect size by leakage magnetic field detection to a certain extent. This is obviously different from the ROSEN scheme, which uses the presence or absence of a signal from an eddy current receiving coil to determine the defect location.

[0058] 3. The electromagnetic cooperative strategy proposed in this invention enables crack detection without blind spots at all angles, which is more reliable than the eddy current method of the ROSEN scheme for crack detection.

[0059] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the steps of the pipeline eddy current-leakage magnetic flux aliasing signal decoupling and processing method of the present invention.

[0061] Figure 2 This is an isometric view of the detection device in the pipeline eddy current-leakage magnetic flux aliasing signal decoupling and processing method of the present invention.

[0062] Figure 3 This is a flowchart illustrating the process of using coils and magnetic sensitive elements to pick up the original eddy current-leakage magnetic flux mixing signal caused by pipeline defects in the pipeline eddy current-leakage magnetic flux mixing signal decoupling and processing method of the present invention.

[0063] Figure 4 This is a flowchart of linear interpolation calculation in the pipeline eddy current-leakage magnetic flux aliasing signal decoupling and processing method of the present invention.

[0064] Figure 5 This is a diagram showing the arrangement of the detection coil and excitation coil in the pipeline eddy current-leakage magnetic flux aliasing signal decoupling and processing method of the present invention.

[0065] The reference numerals in the accompanying drawings of the instruction manual include: 1. Detector body; 2. Walking wheel; 3. Detection coil; 4. Excitation coil; 5. Ceramic plate; 6. Hall sensor. Detailed Implementation

[0066] The following detailed description illustrates the specific implementation method:

[0067] Implementation, for example, attached Figure 1 and Figure 3 As shown: A method for decoupling and processing pipeline eddy current-leakage magnetic flux aliasing signals, including the following steps:

[0068] S101, Signal Acquisition: A coil and a magnetic sensitive element are used to pick up the raw signal of eddy current-leakage magnetic flux mixing caused by pipe defects. The magnetic sensitive element is one of a magnetoresistive sensor and a Hall sensor 6. In this embodiment, a Hall sensor 6 is selected as the magnetic sensitive element.

[0069] S102, Signal Separation: The aliased original signal is frequency-divided using a low-pass filter and a band-pass filter to extract the DC component, AC high-frequency component, and AC low-frequency component. Specifically, a coil acquires the AC high-frequency and AC low-frequency components, while Hall sensor 6 acquires the DC component. Subsequently, the DC, AC high-frequency, and AC low-frequency components are input into a CNN neural network model to extract the signal peak values ​​and defect type labels for the DC component. The AC high-frequency and AC low-frequency components correspond to eddy current detection signals, used to reflect axial defect characteristics; the DC component corresponds to leakage magnetic flux detection signals, used to reflect circumferential or volumetric defect characteristics.

[0070] Specifically, the method for extracting signal peaks using a CNN neural network model includes the following steps:

[0071] S201, Dataset Preprocessing: Collect 5000 sets each of eddy current high-frequency, eddy current low-frequency, and leakage magnetic low-frequency signals for different defects in the pipeline. By manually labeling the timestamps and amplitudes of the peak coordinates, the eddy current high-frequency, eddy current low-frequency, and leakage magnetic low-frequency signals for different defects are divided into training set and validation set in a 7:2 ratio.

[0072] S202, CNN Neural Network Model Construction: Input the training set into the CNN neural network model for deep learning training. During training, the mean squared error (MSE) is used as the loss function, and the dropout rate and L2 regularization coefficient of the fully connected layers are adjusted. At the same time, the performance of the CNN neural network model is monitored in real time using the validation set. If the validation set loss does not decrease for 10 consecutive epochs, the adjustment of the CNN neural network model parameters is stopped, and the trained CNN neural network model is obtained.

[0073] The mean squared error (MSE) is the formula for calculating the loss function:

[0074] (1).

[0075] in, For the peak coordinate timestamp, The peak coordinate amplitude, For the predicted peak coordinate timestamp, N represents the predicted peak coordinate amplitude, and N is the number of samples.

[0076] S203, Peak Extraction: The high-frequency AC component signal and the low-frequency AC component signal in S102 are input into the CNN neural network model. The CNN neural network model outputs the peak values ​​of the high-frequency AC component signal and the low-frequency AC component signal. The peak values ​​include the timestamp and amplitude of the peak coordinates.

[0077] S103, Defect Location Determination: Based on the skin effect principle and signal characteristics, the defect direction dimension and spatial dimension are determined.

[0078] Direction determination: If the CNN neural network model outputs either a circumferential or volumetric defect label, and the peak amplitude of the DC component signal is greater than a preset threshold, the defect is determined to be either circumferential or volumetric.

[0079] If either the peak amplitude of the AC high-frequency component or the AC low-frequency component exceeds a preset threshold, and the DC component has no circumferential label, the defect is determined to be axial.

[0080] Spatial location determination: The excitation coil 4 and detection coil 3 on the detection device are kept at a preset distance from the pipe wall. In this embodiment, a ceramic plate 5 is placed between the excitation coil 4 and detection coil 3 and the pipe wall. The ceramic plate 5 can improve the wear resistance of the detection device and is non-magnetic. A high-frequency alternating current is passed to the excitation coil 4 by the signal generator on the detection device, so that a high-frequency alternating electromagnetic field is generated around the excitation coil 4. At this time, when the detection coil 3 picks up the voltage signal, it is determined that the defect is located in the area near the inner wall of the pipe.

[0081] When a high-frequency alternating current is applied to the excitation coil 4, and the detection coil 3 has no induced voltage, the high-frequency alternating current is adjusted to a low-frequency alternating current. If the detection coil 3 can pick up a voltage signal at this time, it is determined that the defect is located in the far inner wall region of the pipe.

[0082] If the peak value of the DC component signal appears simultaneously with the peak value of the AC low-frequency component, it is determined to be a volumetric defect in the far inner wall region of the pipe wall; if the peak value of the DC component signal appears simultaneously with the peak value of the AC high-frequency component, it is determined to be a volumetric defect in the near inner wall region of the pipe wall.

[0083] like Figure 2 and Figure 5 As shown, specifically, the detection device includes a detector body 1 and a controller. The detector body 1 is circumferentially provided with wheels 2, and the detection coil 3 and excitation coil 4 are both circumferentially fixedly connected to the detector body 1. In this embodiment, the excitation coil 4 is a U-shaped excitation coil (opening towards the pipe wall), and 12 groups are evenly arranged along the circumference of the detector body 1; the plane of the detection coil 3 is perpendicular to the opening plane of the excitation coil 4, and each group of excitation coil 4 corresponds to one detection coil 3.

[0084] The signal generator is embedded in the detector body 1. The controller is used to control the signal generator to supply alternating current to the excitation coil 4 and to receive the voltage signal picked up by the detection coil 3.

[0085] When detection coil 3 picks up a voltage signal, it uses a time-division triggering method for high-frequency and low-frequency bands to acquire the signal. The high-frequency alternating current frequency is set to 100-200kHz, and the low-frequency alternating current frequency is set to 1-10kHz. After the voltage signal is acquired, detection coil 3 processes it using a wavelet threshold noise reduction algorithm, selecting a db6 wavelet basis for 3-level decomposition. The calculation formula is as follows:

[0086] (2).

[0087] Where T is the threshold value of the wavelet thresholding algorithm. N is the number of signal sampling points.

[0088] Specifically, this embodiment takes a gas pipeline as an example. The pipeline is made of X80 steel, with a diameter of 300mm and a wall thickness of 12mm. Previous inspections revealed micro-cracks and localized corrosion defects in the pipe wall, which need to be accurately located and quantified using this method. The Hall sensor 6 uses an A1324 linear Hall element with a sensitivity of 1.3mV / Gs and a measurement range of ±800Gs.

[0089] like Figure 5 As shown, during installation, the Hall sensor 6 is installed between the detection coil 3 and the excitation coil 4. The planes of the detection coil 3 and the Hall sensor 6 are perpendicular to the opening plane of the excitation coil 4. Each group of excitation coils 4 corresponds to one detection coil 3 and one Hall sensor 6, and 12 groups are evenly arranged circumferentially. The detector body 1 is placed inside the gas pipeline. By introducing positive pressure into the gas pipeline, a pressure difference is generated at both ends of the detector body 1, thereby pushing the detector body to move inside the gas pipeline for detection. At this time, the traveling wheel 2 can travel along the inner wall of the pipeline. Meanwhile, in some other embodiments, the detector body 1 is provided with a drive structure for driving the traveling wheel 2 to rotate. When the traveling wheel 2 is in contact with the inner wall of the pipeline, the rotation of the traveling wheel 2 can drive the detector body 1 to move and detect inside the pipeline.

[0090] During the signal acquisition phase, the signal generator embedded in the detector body 1 is activated by the controller, and an initial high-frequency current (150kHz) is supplied to the excitation coil 4. At this time, eddy currents are generated at the pipe defect due to electromagnetic induction, and the pipe's own magnetic field leaks due to the defect, forming an aliased original signal. The Hall sensor 6 and the detection coil 3 convert this magnetic signal into a voltage signal (amplitude range 0-5V), which is sampled by a 16-bit ADC analog-to-digital converter (sampling frequency set to 1MHz to ensure the capture of high-frequency signal details). The analog signal is converted into a digital signal and temporarily stored. The single acquisition time is set to 10s, and 5 sets of aliased original signals are acquired to avoid random errors.

[0091] In the S102 signal separation step, the temporarily stored aliased raw signal is first input to a low-pass filter and a band-pass filter. The low-pass filter is used to extract the DC component signal, the high-frequency band-pass filter has a passband of 80-250kHz and is used to extract the AC high-frequency component signal corresponding to eddy current detection, and the low-frequency band-pass filter has a passband of 0.5-15kHz and is used to extract the AC low-frequency component signal corresponding to leakage flux detection.

[0092] After filtering, the amplitude of the DC component signal is concentrated in the range of -0.5V to +0.5V, the amplitude of the AC high-frequency component signal is concentrated in the range of 0.8-2.2V, and the amplitude of the AC low-frequency component signal is concentrated in the range of 0.3-1.1V. The waveforms of the three components do not overlap significantly, thus achieving effective separation.

[0093] The separated DC component signal, AC high-frequency signal, and AC low-frequency signal were then input into the trained CNN neural network model. The CNN neural network model had previously been trained with 5000 sets of X80 steel pipe defect signals (3500 sets for training and 1000 sets for validation, including 0.5-5mm defects near the inner wall of the pipe and 0.5-8mm defects far from the inner wall). The validation set loss was stable at 0.09, the peak timestamp prediction error was ±2 sampling points, and the amplitude error was ±0.06V.

[0094] After inputting the AC high-frequency signal in this embodiment, the CNN neural network model outputs a peak timestamp of "sampling point 489231" (corresponding to the axial position of the pipe at 12.3m) and an amplitude of 1.92V; after inputting the AC low-frequency signal in this embodiment, the output peak timestamp is "sampling point 651784" (corresponding to the axial position of the pipe at 16.5m) and an amplitude of 0.87V, indicating that there are two defects at different locations.

[0095] Next, the S103 defect location determination is performed: The controller control signal generator first supplies a 150kHz high-frequency alternating current (within the high-frequency range) to the excitation coil 4. At this time, the detection coil 3 starts to pick up the voltage signal. If a voltage signal is picked up, it means that the high-frequency electromagnetic field penetrates the surface of the pipe and induces a defect at the defect location. Combined with the skin effect, it is determined to be a defect in the area near the inner wall of the pipe. If no voltage signal is picked up, it switches to a 5kHz low-frequency alternating current (within the low-frequency range). If a voltage signal is detected at this time, because the low-frequency current has a large skin depth, it can act on the entire range of the pipe wall thickness, and it is determined to be a defect in the area far from the inner wall of the pipe.

[0096] In this embodiment, when a 150kHz high-frequency current is applied, the detection coil 3 picks up a 1.2V voltage signal at the location corresponding to the "489231st sampling point" (after 3-layer decomposition and noise reduction using a db6 wavelet base, the signal-to-noise ratio is improved from 20dB to 35dB, and the waveform is smooth and free of noise), and this location is determined to be a defect in the near-inner wall region of the pipe wall; while at the location corresponding to the "651784th sampling point", when a 150kHz high-frequency current is applied, the detection coil 3 has no voltage signal output. After switching to a 5kHz low-frequency current, a 0.7V voltage signal is picked up (also after wavelet noise reduction processing), and this location is determined to be a defect in the far-inner wall region of the pipe wall.

[0097] In this embodiment of X80 steel gas pipeline inspection, the specific process of DC component signal detection for identifying circumferential and volumetric defects is as follows:

[0098] The DC component signal is extracted from the aliased signal using a low-pass filter with a cutoff frequency of 0.1 kHz, and then input into the CNN neural network model.

[0099] In the defect determination process, if the peak amplitude of the DC component signal is ≥0.5V (the preset threshold for X80 steel pipe) and the CNN neural network model outputs the label "circumferential", combined with the corresponding axial position of the pipe at 14.7m, the defect is determined to be circumferential. If the peak amplitude is 0.7V and the label is "volume", and it appears simultaneously with the peak of the AC high-frequency component, it is determined to be a volume defect in the area near the inner wall of the pipe.

[0100] S104, Defect Size Quantification: Establish correlation curves between circumferential and volumetric defects and voltage signal intensity corresponding to DC component signals, as well as correlation curves between voltage signal intensity and near-inner wall region and far-inner wall region of pipe wall corresponding to AC component signals; after completing defect judgment, compare the corresponding signal peak value extracted in S102 with the correlation curve, and determine the actual depth of the defect through interpolation calculation.

[0101] The specific method for establishing the correlation curves between defects at different depths in the near-inner wall region and the far-inner wall region of the pipe wall and the corresponding voltage signal intensity using a DNN neural network model is as follows:

[0102] S301, Dataset Preprocessing: Collect voltage signal samples of defects at different depths in the near-inner wall region and the far-inner wall region of the pipe wall. The defect depth in the near-inner wall region of the pipe wall covers 0.5-5mm, and the defect depth in the far-inner wall region of the pipe wall covers 0.5-8mm. Prepare 3 parallel test blocks for each depth. Repeat the voltage signal acquisition 20 times for each parallel test block. Finally, 5000 sets of samples are formed for the defect signals in the near-inner wall region and the far-inner wall region of the pipe wall.

[0103] Simultaneously, the input and output features are standardized. Min-Max standardization is used to map the data to the [0,1] interval. The calculation formula is as follows:

[0104] (3).

[0105] in, It is one of the following: defect depth and voltage signal strength. This is the minimum value of the feature. The maximum value of this feature. This is the standardized data.

[0106] S302, DNN Neural Network Model Construction: The training sets of defect signals from the near-inner wall region and the far-inner wall region of the pipe are input into two independent DNN neural network models for deep learning training. Training parameters are set for the DNN neural network models, with the root mean square error (RMSE) as the loss function. The calculation formula is as follows:

[0107] (4).

[0108] in, For loss function, The number of samples; The actual voltage signal strength, This represents the voltage signal strength predicted by the model.

[0109] S303, Correlation Curve Generation: The trained DNN neural network models for the near-inner wall region and the far-inner wall region of the pipe wall are used to generate correlation curves between the corresponding defect depth and voltage signal intensity, as detailed below:

[0110] Within the defect depth range of 0.5-5mm in the near-inner wall region of the pipe wall and the defect depth range of 0.5-8mm in the far-inner wall region of the pipe wall, 100 uniformly spaced depth values ​​are selected for each region. After standardization, these values ​​are input into the corresponding DNN neural network model to obtain the predicted voltage signal intensity value for each depth value.

[0111] Using the depth value and the corresponding predicted voltage signal intensity as coordinate points, a smooth curve is plotted using a polynomial fitting method. This curve represents the correlation between defects of different depths and their corresponding voltage signal in the near-inner wall region and the far-inner wall region of the pipe.

[0112] like Figure 4 As shown, the specific method for determining the actual depth of a defect through interpolation calculation is as follows:

[0113] S401, Correlation Curve Matching and Signal Peak Extraction: Based on the defect location determined in S103, select the correlation curve output by the DNN neural network model at the corresponding location; at the same time, extract the corresponding signal peak output by the CNN neural network model in S102.

[0114] If the defect is located in the area near the inner wall of the pipe, the amplitude of the high-frequency signal peak is extracted.

[0115] If the defect is located in the region far from the inner wall of the pipe, the amplitude of the low-frequency signal peak is extracted and used as the reference value for the voltage signal strength to be interpolated.

[0116] S402, Interpolation Interval Positioning: On the selected correlation curve, find the two coordinate points closest to the reference value of the voltage signal strength to be interpolated: Let the voltage signal strength to be interpolated be... The set of coordinate points of the associated curve Find satisfaction in

[0117] S403, Linear interpolation calculation:

[0118] .

[0119] .

[0120] Before interpolation calculation, the signal peak value is preprocessed, specifically by removing outliers using the 3σ criterion and then smoothing it using a moving average filter. The formula is as follows:

[0121] (6).

[0122] in, This is the original peak sequence. This is the filtered sequence; subsequently, cubic spline interpolation is used, with the interpolation interval set as follows: interpolation function satisfy Furthermore, its second derivative is continuous within the interval, and its piecewise expression is:

[0123] (7).

[0124] in, The interpolation node spacing is set to 0.05 mm, and boundary conditions are added. .

[0125] Specifically, in S104, defect size quantification is performed by first selecting the corresponding DNN neural network model correlation curve based on the location determination result: for defects in the near-inner wall region of the pipe wall, the "defect depth - high frequency signal peak amplitude" correlation curve of the near-inner wall region DNN neural network model of the pipe wall is selected; for defects in the far-inner wall region of the pipe wall, the "defect depth - low frequency signal peak amplitude" correlation curve of the far-inner wall region DNN neural network model of the pipe wall is selected.

[0126] The DNN neural network model was previously trained using test blocks with defects of 0.5-5 mm (0.1 mm interval) in the near-inner wall region of the pipe and defects of 0.5-8 mm (0.1 mm interval) in the far-inner wall region of the pipe. Three parallel test blocks were used for each depth, and signals were collected 20 times for each block, for a total of 5000 sets of samples. After training, the root mean square error (RMSE) was 0.08 mm, and the goodness of fit of the correlation curve R² was greater than 0.98.

[0127] For defects in the area near the inner wall of the pipe, the peak amplitude of the high-frequency signal output by the CNN neural network model is extracted as 1.92V. First, outliers are removed by the 3σ criterion (in this embodiment, the amplitude is within the normal range and does not need to be removed). Then, the signal is smoothed by moving average filtering (window size is set to 5) to obtain the filtered amplitude of 1.91V.

[0128] On the correlation curve of "defect depth in the near-inner wall region of the pipe wall - peak amplitude of high-frequency signal", find the two coordinate points closest to 1.91V: (2.8mm, 1.89V) and (2.9mm, 1.93V) respectively, and substitute them into formula (5) to calculate:

[0129]

[0130] That is, the depth of the defect in the area near the inner wall of the pipe is 2.85mm.

[0131] For defects in the far inner wall region of the pipe wall, the peak amplitude of the low-frequency signal output by the CNN neural network model is extracted as 0.87V, which is then filtered by the 3σ criterion and moving average to obtain 0.86V. On the correlation curve of "defect depth in the far inner wall region of the pipe wall - peak amplitude of low-frequency signal", the two coordinate points closest to 0.86V are found: (4.2mm, 0.84V) and (4.3mm, 0.88V). These are then substituted into formula (5) for calculation.

[0132]

[0133] That is, the defect depth in the far inner wall area of ​​the pipe is 4.25mm.

[0134] To verify the accuracy, ultrasonic testing was used to re-inspect the two defects. The results showed that the defect depth in the near-inner wall area of ​​the pipe was 2.9 mm and the defect depth in the far-inner wall area of ​​the pipe was 4.3 mm. The detection error of this method was less than 0.05 mm, which meets the accuracy requirements for industrial pipeline defect detection (allowable error ±0.1 mm). Moreover, the detection coil 3 position does not need to be adjusted throughout the process, and the detection is completed only through signal processing. The detection efficiency is 40% higher than the existing pipeline internal inspection technology of German company ROSEN, and the probe weight is reduced by 35%, which is suitable for the continuous detection requirements of the gas pipeline over a distance of 10 km.

[0135] This invention solves the problems of large probe size, heavy weight, and frequent mechanism switching caused by mechanical structures by using pure signal processing logic to determine the location of defects through signal frequency division, neural network feature extraction, and skin effect. No interruption or hardware adjustment is required during the detection process, enabling continuous scanning across the entire length of the pipeline, significantly improving the efficiency and stability of long-distance pipeline inspection, and avoiding the risk of detection interruption due to mechanical failure.

[0136] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for decoupling and processing pipeline eddy current-leakage magnetic flux aliasing signals, characterized in that, Includes the following steps: S101, Signal Acquisition: The original signal of eddy current-leakage magnetic flux mixing caused by pipeline defects is picked up using coils and magnetic sensitive elements; S102, Signal Separation: The aliased original signal is frequency-divided by a low-pass filter and a band-pass filter to extract the DC component signal, AC high-frequency component signal and AC low-frequency component signal respectively; then, the DC component signal, AC high-frequency component signal and AC low-frequency component signal are input into a CNN neural network model to extract the signal peak and the defect type label of the DC component respectively. S103, Defect Location Determination: Based on the skin effect principle and signal characteristics, the defect direction dimension and spatial dimension are determined. Direction determination: If the CNN neural network model outputs either a circumferential or volumetric defect label, and the peak amplitude of the DC component signal is greater than a preset threshold, the defect is determined to be either circumferential or volumetric. If either the peak amplitude of the AC high-frequency component or the AC low-frequency component is greater than a preset threshold, and the DC component has no circumferential label, the defect is determined to be axial. Spatial position determination: Using the excitation coil (4) and detection coil (3) on the detection device to maintain a preset distance from the pipe wall, the signal generator on the detection device is used to pass a high-frequency alternating current to the excitation coil (4) to generate a high-frequency alternating electromagnetic field around the excitation coil (4). At this time, when the detection coil (3) picks up the voltage signal, it is determined that the defect is located in the area near the inner wall of the pipe wall. When the excitation coil (4) is supplied with a high-frequency alternating current, the detection coil (3) has no induced voltage. The high-frequency alternating current is then adjusted to a low-frequency alternating current. If the detection coil (3) can pick up a voltage signal at this time, the defect is determined to be located in the area far from the inner wall of the pipe. If the peak value of the DC component signal and the peak value of the AC low-frequency component signal appear simultaneously, it is determined to be a volumetric defect in the far inner wall region of the pipe wall; if the peak value of the DC component signal and the peak value of the AC high-frequency component signal appear simultaneously, it is determined to be a volumetric defect in the near inner wall region of the pipe wall. S104, Defect Size Quantification: Establish correlation curves between circumferential and volumetric defects and voltage signal intensity corresponding to the DC component, and correlation curves between axial defects in the near-inner wall region and far-inner wall region of the pipe wall and voltage signal intensity corresponding to the AC component; after completing the defect judgment, compare the corresponding signal peak value extracted in S102 with the correlation curve, and determine the actual depth of the defect by interpolation calculation.

2. The method for decoupling and processing pipeline eddy current-leakage magnetic flux mixing signals according to claim 1, characterized in that, In S101, the magnetic sensitive element is either a magnetoresistive sensor or a Hall sensor (6).

3. The method for decoupling and processing pipeline eddy current-leakage magnetic flux mixing signals according to claim 2, characterized in that, In S102, the DC signal component corresponds to the leakage flux detection signal; the AC high-frequency component signal and the AC low-frequency component signal correspond to the eddy current detection signal.

4. The method for decoupling and processing pipeline eddy current-leakage magnetic flux aliasing signals according to claim 3, characterized in that, In S102, the method for extracting signal peaks using the CNN neural network model includes the following steps: S201, Dataset Preprocessing: Collect 5000 sets each of eddy current high frequency, eddy current low frequency, and leakage magnetic low frequency signals of different defects in the pipeline. By manually annotating the timestamps and amplitudes of the peak coordinates, the eddy current high frequency, eddy current low frequency, and leakage magnetic low frequency signals of different defects are divided into training set and validation set in a 7:2 ratio. S202, CNN Neural Network Model Construction: Input the training set into the CNN neural network model for deep learning training. During training, the mean squared error (MSE) is used as the loss function, and the dropout rate and L2 regularization coefficient of the fully connected layers are adjusted. At the same time, the performance of the CNN neural network model is monitored in real time using the validation set. If the validation set loss does not decrease for 10 consecutive epochs, the adjustment of the CNN neural network model parameters is stopped, and the trained CNN neural network model is obtained. The mean squared error (MSE) is the formula for calculating the loss function: （1）； in, For the peak coordinate timestamp, The peak coordinate amplitude, For the predicted peak coordinate timestamp, for The predicted peak coordinate amplitude, where N is the number of samples; S203, Peak Extraction: The high-frequency AC component signal and the low-frequency AC component signal in S102 are input into the CNN neural network model. The CNN neural network model outputs the peak values ​​of the high-frequency AC component signal and the low-frequency AC component signal. The peak values ​​include the timestamp and amplitude of the peak coordinates.

5. The method for decoupling and processing pipeline eddy current-leakage magnetic flux mixing signals according to claim 4, characterized in that, In S103, the detection device includes a detector body (1) and a controller. The detector body (1) is provided with a walking wheel (2) in the circumferential direction. The detection coil (3) and the excitation coil (4) are both fixedly connected to the detector body (1) in the circumferential direction. The signal generator is embedded in the detector body (1). The controller is used to control the signal generator to pass alternating current to the excitation coil (4) and receive the voltage signal picked up by the detection coil (3).

6. The method for decoupling and processing pipeline eddy current-leakage magnetic flux aliasing signals according to claim 5, characterized in that, In S103, when the detection coil (3) picks up the voltage signal, the voltage signal is picked up by time-division triggering of high frequency and low frequency bands.

7. The method for decoupling and processing pipeline eddy current-leakage magnetic flux aliasing signals according to claim 6, characterized in that, In S103, the frequency of the high-frequency alternating current is set to 100-200kHz, and the frequency of the low-frequency alternating current is set to 1-10kHz; after the detection coil (3) picks up the voltage signal, it is processed by the wavelet threshold noise reduction algorithm, and the db6 wavelet basis is selected for 3-level decomposition. The calculation formula is as follows: （2）； Where T is the threshold value of the wavelet thresholding algorithm. N is the number of signal sampling points.

8. The method for decoupling and processing pipeline eddy current-leakage magnetic flux mixing signals according to claim 7, characterized in that, In S104, the specific method for establishing the correlation curves between defects at different depths and corresponding voltage signal intensities in the near-inner wall region and the far-inner wall region of the pipe using a DNN neural network model is as follows: S301, Dataset Preprocessing: Collect voltage signal samples of defects at different depths in the near-inner wall region and the far-inner wall region of the pipe wall. The defect depth in the near-inner wall region of the pipe wall covers 0.5-5mm, and the defect depth in the far-inner wall region of the pipe wall covers 0.5-8mm. Prepare 3 parallel test blocks for each depth. Repeat the voltage signal acquisition 20 times for each parallel test block. Finally, 5000 sets of samples are formed for the defect signals in the near-inner wall region and the far-inner wall region of the pipe wall. Simultaneously, the input and output features are standardized. Min-Max standardization is used to map the data to the [0,1] interval. The calculation formula is as follows: （3）； in, It is one of the following: defect depth and voltage signal strength. This is the minimum value of the feature. The maximum value of this feature. The data is standardized. S302, DNN Neural Network Model Construction: The training sets of defect signals from the near-inner wall region and the far-inner wall region of the pipe are input into two independent DNN neural network models for deep learning training. Training parameters are set for the DNN neural network models, with the root mean square error (RMSE) as the loss function. The calculation formula is as follows: （4）； in, For loss function, The number of samples; The actual voltage signal strength, The voltage signal strength predicted by the model; S303, Correlation Curve Generation: The trained DNN neural network models for the near-inner wall region and the far-inner wall region of the pipe wall are used to generate correlation curves between the corresponding defect depth and voltage signal intensity, as detailed below: Within the defect depth range of 0.5-5mm in the near-inner wall region of the pipe wall and the defect depth range of 0.5-8mm in the far-inner wall region of the pipe wall, 100 uniformly spaced depth values ​​are selected for each region. After standardization, these values ​​are input into the corresponding DNN neural network model to obtain the predicted voltage signal intensity value for each depth value. Using the depth value and the corresponding predicted voltage signal intensity as coordinate points, a smooth curve is plotted using a polynomial fitting method. This curve represents the correlation between defects of different depths and their corresponding voltage signal in the near-inner wall region and the far-inner wall region of the pipe.

9. The method for decoupling and processing pipeline eddy current-leakage magnetic flux mixing signals according to claim 8, characterized in that, In S104, the method for determining the actual depth of the defect through interpolation calculation is as follows: S401, Correlation curve matching and signal peak extraction: Based on the defect location determined in S103, select the correlation curve output by the DNN neural network model at the corresponding location; at the same time, extract the corresponding signal peak output by the CNN neural network model in S102. If the defect is located in the area near the inner wall of the pipe, the amplitude of the high-frequency signal peak is extracted. If the defect is in the region far from the inner wall of the pipe, the amplitude of the low-frequency signal peak is extracted and used as the reference value of the voltage signal strength to be interpolated. S402, Interpolation Interval Positioning: On the selected correlation curve, find the two coordinate points closest to the reference value of the voltage signal strength to be interpolated: Let the voltage signal strength to be interpolated be... The set of coordinate points of the associated curve Find satisfaction in ; S403, Linear interpolation calculation: ; （5）； 。 10. The method for decoupling and processing pipeline eddy current-leakage magnetic flux mixing signals according to claim 9, characterized in that, In S104, the signal peak value is preprocessed before interpolation calculation. Specifically, outliers are removed using the 3σ criterion, and then smoothed using a moving average filter. The formula is as follows: （6）； in, This is the original peak sequence. This is the filtered sequence; subsequently, cubic spline interpolation is used, with the interpolation interval set as follows: interpolation function satisfy Furthermore, its second derivative is continuous within the interval, and its piecewise expression is: （7）； in, The interpolation node spacing is set to 0.05 mm, and boundary conditions are added. .

Images