AI edge identification training data acquisition platform for magnetic flux leakage defect identification, construction method and system
By designing an AI edge recognition training data acquisition platform and constructing a three-dimensional data set, the problems of limited detection range and insufficient intelligence in magnetic flux leakage detection technology have been solved, and efficient detection and data analysis of internal and external defects have been achieved.
Patent Information
- Application Number
- CN202511049511.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-29
- Publication Date
- 2025-10-17
AI Technical Summary
Existing magnetic flux leakage detection technology has problems such as limited detection range, inability to take into account both internal and external defects, low data analysis efficiency, and insufficient intelligence.
An AI edge recognition training data acquisition platform was designed, including a bracket, power module, motor, detection probe, data processing module, etc., to achieve automatic transmission of the tested material, multi-dimensional signal acquisition and real-time processing. An internal/external injury dual-mode probe was used for synchronous detection, combined with the SOGI-FLL module for signal screening and filtering, to construct a three-dimensional dataset containing time domain, frequency domain and spatial coordinates.
It achieves comprehensive detection of internal and external defects in materials, improves detection accuracy and data analysis efficiency, enhances the level of intelligence, and ensures the accuracy and richness of AI training data.
Smart Images
Figure CN120801488A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of magnetic flux leakage detection, and in particular to an AI edge recognition training data acquisition platform for magnetic flux leakage defect recognition, a construction method and a system. BACKGROUND
[0002] As a non-contact detection technology, the working principle of magnetic flux leakage detection is to magnetize the measured object, when the measured object has defects, part of the magnetic force lines will be leaked from the defects to the air, and captured by the detection probe. These magnetic flux leakage signals can accurately reflect the size and topography of the defects, and are crucial to production safety. With outstanding anti-interference performance, low implementation cost and high detection accuracy, this technology has become one of the most widely used non-destructive testing methods in the industrial field. Generally, the magnetic flux leakage detection device only focuses on a single detection scene, the detection range is small and internal defects cannot be detected.
[0003] The demand for automation and intelligentization in the field of non-destructive testing is increasingly urgent. Magnetic flux leakage detection locates metal material defects by identifying magnetic field distortion, but its data analysis efficiency and accuracy have long been limited by traditional methods. The rise of artificial intelligence (AI) edge detection technology provides a new path to solve this problem, but its implementation process still faces multiple bottlenecks such as data collection, processing and system integration. SUMMARY
[0004] The technical problem to be solved by the present application is to solve the problems of limited detection range, inability to consider internal and external defects, low data analysis efficiency and insufficient intelligentization of existing magnetic flux leakage detection technology in the background art. The present application provides an AI edge recognition training data acquisition platform for magnetic flux leakage defect recognition, a construction method and a system.
[0005] The technical solution adopted by the present application to solve the technical problem is an AI edge recognition training data acquisition platform for magnetic flux leakage defect recognition, comprising: a support, a power module, a motor, a driving device, a detection probe, a data processing module, a measured material placement platform, a limit sensor, a single-chip microcomputer, a display module and an upper computer; The measured material placement platform is connected to the motor through a conveyor belt; The detection probe includes an internal injury detection probe and an external injury detection probe, each having 16 detection output ends; The data processing module includes a differential circuit, a first-order amplification circuit, a second-order active low-pass filter circuit and a second-order amplification circuit.
[0006] A complete hardware architecture is established to realize an integrated platform of automatic transmission of the measured material, multi-dimensional signal acquisition, real-time signal processing and data display, and to solve the limitation of single-point detection of traditional equipment through synchronous detection of internal / external damage dual-mode probes.
[0007] According to one embodiment of the application, the single-chip microcomputer controls the driving device through a PWM wave, a forward / reverse button and a speed adjusting knob are arranged on the support, and the limit sensors are arranged at both ends of the guide rail and connected with the single-chip microcomputer, so as to realize automatic switching of the moving direction and collision protection.
[0008] The PWM precise speed regulation and double-limit protection ensure the uniform and controllable detection process, the forward / reverse button and the limit sensor linkage mechanism realize the automation of the reciprocating motion of the detection platform.
[0009] According to one embodiment of the application, the coil output voltage of the detection probe satisfies the formula: ; Wherein, N is the number of turns, A is the effective cross-sectional area, v is the moving speed, is the magnetic induction gradient.
[0010] The mathematical relationship between the quantized output voltage and the defect characteristics provides a physical model basis for AI training, and the magnetic induction gradient is coupled with the mechanical motion parameter (v) to build a model.
[0011] According to one embodiment of the application, the cut-off frequency of the second-order active low-pass filter circuit satisfies: ; Wherein, r0=r1, c0=nc1, and n is a proportional coefficient.
[0012] Through the symmetric design of r0=r1 and c0=nc1, the stability of the cut-off frequency is improved, and the passband fluctuation is ≤±0.3dB; the proportional coefficient n is dynamically adjustable to adapt to the frequency response characteristics of different materials.
[0013] According to one embodiment of the application, a permanent magnet is further included, which is fixed to the measured material placement platform, and is used to make the measured material reach a magnetic saturation state.
[0014] The permanent magnet magnetic saturation design makes the leakage magnetic field intensity at the defect increase by 2-3 orders of magnitude, and the platform integrated magnetization scheme avoids the space interference of the traditional external magnetization device.
[0015] An AI edge recognition training data construction method for magnetic flux leakage defect recognition is also provided, which adopts the acquisition platform in the above-mentioned scheme, and includes the following steps: S1, magnetize the measured material through the permanent magnet; S2, leakage magnetic signals are collected by the 16-way internal injury detection probe and the external injury detection probe; S3, a SOGI-FLL module is used for signal screening, and a normalized transfer function expression of the SOGI-FLL module is:
[0016] wherein ω is a frequency estimation value, s is a Laplace operator, and k is an adjustment coefficient; Normalized expressions of the band-pass filter D(s) and the low-pass filter Q(s) are: (1); (2).
[0017] The normalized combination of the band-pass D(s) and the low-pass Q(s) solves the frequency band aliasing problem.
[0018] According to one embodiment of the present application, the SOGI-FLL module in step S3 completes signal frequency identification within 0.5 seconds, and the tracking error is less than 5%. The tracking error less than 5% ensures that the defect feature frequency extraction accuracy is greater than or equal to 98% According to one embodiment of the present application, the signal processing process of the data processing module includes: Sa, differential input stage: a differential amplifier is used to receive the double-ended output signal of the detection probe, to eliminate common-mode interference, and the differential rejection ratio is greater than or equal to 60dB; Sb, first-stage amplification stage: a differential signal is preliminarily amplified by a tunable gain operational amplifier, the gain range is 20-40dB, and the bandwidth is 0.1Hz-10kHz; Sc, second-order filtering stage: a second-order active low-pass filter of Sallen-Key topology is used; Sd, second-stage amplification stage: an instrument amplifier is used for accurate amplification, the gain is 40-60dB, and the common-mode rejection ratio is greater than or equal to 100dB; Se, output conditioning stage: impedance matching is performed through a voltage follower, and the output impedance is less than or equal to 100Ω; Wherein, a resistance-capacitance coupling mode is used between stages, the direct current blocking capacitance value is greater than or equal to 10μF, the overall signal link total gain is greater than or equal to 80dB, and the passband ripple is less than or equal to ±0.5dB.
[0019] According to one embodiment of the present application, the detection probe includes a replaceable coil assembly, the replaceable coil assembly has a standardized interface, and is detachably installed in the probe shell through a mechanical connecting piece; and parameter configuration of the coil assembly satisfies: The coil turns N1 of the internal injury detection probe ranges from 800±50 turns, and the wire diameter is 0.15±0.02mm; The coil turns N2 of the external injury detection probe ranges from 500 to 30 turns, and the wire diameter is 0.2+ / -0.02 mm.
[0020] Also provided is an AI edge recognition training system adopting the data set constructed by the AI edge recognition training data construction method for magnetic flux leakage defect recognition in the above scheme, wherein the data contains original detection signal data, processed feature data and spatial marker data, The original detection signal data includes: The internal injury detection 8-channel signal array has a sampling rate of greater than or equal to 100 kHz and a 16-bit resolution, and contains time domain waveform data and FFT spectrum data; The external injury detection 8-channel signal array has a sampling interval of less than or equal to 10 mu s, a signal amplitude range of plus or minus 10 V, and a time stamp marker; The processed feature data includes: The normalized signal processed by SOGI-FLL includes: The fundamental component amplitude A_f has an accuracy of 0.1 mV; The harmonic distortion THD is calculated to the 5th harmonic; The signal-to-noise ratio SNR is measured using A weighting; The phase difference data phi has a resolution of 0.1 degrees; The spatial marker data includes: The defect position three-dimensional coordinates (x, y, z), wherein: the x-axis is the motion direction of the conveying belt, the positioning accuracy is plus or minus 0.5 mm; the y-axis is the probe array direction, the positioning accuracy is plus or minus 1 mm; and the z-axis is the defect depth, which is obtained by calculating the time difference of the multi-probe signal; The defect size feature vector [L, W, D] is L, W and D, L represents length, W represents width, and D represents depth, and the measurement error is less than or equal to 5%.
[0021] The three-dimensional data fusion architecture of time domain, frequency domain and spatial coordinates enables the AI training model to be optimized, the time domain data provides the basis for CNN network training, the frequency domain features are used as the time sequence input of the LSTM network, the spatial coordinates construct 3D point cloud data for the graph neural network to process, and the three modalities are fused to greatly improve the model recognition accuracy.
[0022] The present application has the following advantages: (1) The measured material is placed on the detection platform, a permanent magnet is placed on the platform, the permanent magnet magnetizes the measured object to achieve magnetic saturation, a single-chip microcomputer is used to control a driving device to drive a motor to move the measured material, accurate control of the movement of the measured object can be realized, the detection accuracy is improved, and the measured material can be replaced as needed; (2) The magnetic flux leakage detection includes internal damage detection and external damage detection, realizes the magnetic flux leakage detection on the inner wall and surface of the material, the internal damage detection and the external damage detection both contain 16 detection probes which are distributed in an array, the probes can be freely replaced according to different detection scenes and regulations, and a large amount of magnetic flux leakage detection data which can contain multiple dimensions can be obtained; (3) The SOGI-FLL module is used for realizing the tracking of the required signal in the magnetic flux leakage detection signal and the filtering of other interference signals, realizing the screening of the magnetic flux leakage detection signal, and providing guarantee for the collection of the AI data set. BRIEF DESCRIPTION OF DRAWINGS
[0023] The application will be further described below in combination with the drawings and examples.
[0024] Figure 1 is a structural block diagram of the embodiment one of the application.
[0025] Figure 2 is a structural schematic diagram of the embodiment one of the application.
[0026] Figure 3 is a schematic diagram of the actual detection of the embodiment one of the application.
[0027] Figure 4 is a circuit diagram of the data processing module in the embodiment one of the application.
[0028] Figure 5 is an internal damage magnetic flux leakage detection experimental result diagram of the embodiment one of the application.
[0029] Figure 6 is an external damage magnetic flux leakage detection experimental result diagram of the embodiment one of the application.
[0030] Figure 7 is a principle schematic diagram of the SOGI-FLL module in the embodiment two of the application.
[0031] Figure 8 is a result diagram of the required magnetic flux leakage detection signal frequency identification in the embodiment two of the application.
[0032] Figure 9 is a result diagram of the required magnetic flux leakage detection signal frequency identification in the embodiment two of the application.
[0033] In the figure: 1. Bracket; 101. Forward and reverse buttons; 102. Speed adjustment knob; 2. Power module; 3. Motor; 4. Drive device; 5. Detection probe; 501. Internal damage detection probe; 502. External damage detection probe; 6. Data processing module; 601. Differential circuit; 602. First-stage amplifier circuit; 603. Second-order active low-pass filter circuit; 604. Second-stage amplifier circuit; 7. Measured material placement platform; 8. Limit sensor; 9. Single-chip microcomputer; 10. Display module; 11. Host computer; 12. Permanent magnet. DETAILED DESCRIPTION
[0034] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.
[0035] Example 1 like Figures 1-2 As shown, a data acquisition platform for AI edge recognition training data for magnetic flux leakage defect identification includes: a bracket 1, a power module 2, a motor 3, a drive device 4, a detection probe 5, a data processing module 6, a test material placement platform 7, a limit sensor 8, a single-chip microcomputer 9, a display module 10, a host computer 11 and a permanent magnet 12. The power module 2 is connected to the drive device 4 and the single-chip microcomputer 9, and the single-chip microcomputer 9 is also connected to the limit sensor 8 and the display module 10. The drive device 4 is connected to the motor 3, the test material placement platform 7 is connected to the motor 3 through a conveyor belt, the detection probe 5 is connected to the data processing module 6, and the data processing module 6 is connected to the host computer 11 through an acquisition card. The permanent magnet 12 is fixed to the test material placement platform 7 to make the test material reach a magnetic saturation state.
[0036] like Figure 2As shown, the internal injury detection probe 501 and the external injury detection probe 502 each have 16 detection output ends arranged in a row, which can increase the detection area; the detection signals of internal injury and external injury are small in amplitude and are affected by other clutter caused by factors such as the scene environment, and are not easy to observe. The detection probe 5 is connected with the data processing module 6, and the captured magnetic flux leakage signals are transmitted to the data processing module 6. The single-chip microcomputer 9 controls the driving device 4 through PWM waves. The support 1 is provided with a forward and reverse button 101 and a speed adjusting knob 102. The limit sensor 8 is arranged at both ends of the guide rail and is connected with the single-chip microcomputer 9, so as to realize automatic switching of the moving direction and collision protection. The collection platform of the embodiment is composed of aluminum profiles. The motor 3 is located at one end of the support 1 and is connected with the conveying belt through a shaft coupling. The measured material placing platform 7 is fixed on the conveying belt and can be adjusted in height. The internal injury detection probe 501 and the external injury detection probe 502 are located above the measured material placing platform 7 and are perpendicular to the measured material placing platform 7. The power module 2, the driving device 4, the data processing module 6, the single-chip microcomputer 9 and the display module 10 are covered with a protective cover to prevent iron filings and the like from damaging the power module 2.
[0037] The power module 2 converts 220V voltage into working voltage for the driving device 4 and the single-chip microcomputer 9, so that the driving device 4 and the single-chip microcomputer 9 work normally. The single-chip microcomputer 9 adopts an Arduino Nano model. The single-chip microcomputer 9 is connected with the driving device 4 and controls the driving device 4 through output PWM waves. The driving device 4 is connected with the motor 3 to provide torque for the motor 3 and drive the motor 3 to rotate. The single-chip microcomputer 9 can control the rotating speed and direction of the motor 3 by changing the frequency and duty cycle of the PWM wave. The forward and reverse button 101 is connected with the single-chip microcomputer 9. When the forward and reverse button 101 is pressed, the single-chip microcomputer 9 receives a level conversion signal and outputs a PWM wave to the driving device 4 to control the rotation of the motor 3. The speed adjusting knob 102 is connected with the single-chip microcomputer 9. When the speed adjusting knob 102 is rotated, the single-chip microcomputer 9 changes the duty cycle of the output PWM wave through internal analog-digital conversion to control the rotating speed of the motor 3. The display module 10 is connected with the single-chip microcomputer 9. When the rotating speed adjusting knob 102 is rotated, the rotating speed of the motor 3 is changed, and the display module 10 can display the current rotating speed. The motor 3 is connected with the conveying belt through a shaft coupling. The motor 3 drives the conveying belt to move through rotation. The measured material placing platform 7 is connected with the conveying belt. The measured material placing platform 7 moves with the conveying belt. The limit sensor 8 is connected with the single-chip microcomputer 9. When the measured material placing platform 7 moves to the limit sensor 8, the limit sensor 8 triggers a relay to send a signal to the single-chip microcomputer 9 to stop the single-chip microcomputer 9 from outputting the PWM wave, so that the motor 3 stops rotating. The motor 3 cannot continue to rotate in the current direction. When the motor 3 rotates in the reverse direction, the limit sensor 8 cannot detect the measured material placing platform 7. The relay is closed, the single-chip microcomputer 9 cannot receive the signal transmitted by the relay, and the motor 3 can rotate in the forward and reverse directions.
[0038] As shown in Figure 3 , the measured material is adsorbed by the permanent magnet 12 and placed on the measured material placement platform 7, the measured material is connected with the permanent magnet 12, and the permanent magnet 12 is used to magnetize and reach magnetic saturation of the measured material; when the material has a defect cutting the magnetic line, the defect or the change of the organization state of the material surface will change the magnetic permeability, because the magnetic permeability of the defect is very small and the magnetic resistance is very large, the magnetic flux in the magnetic circuit is distorted, and the flow direction of the magnetic induction line is changed, in addition to part of the magnetic flux directly passing through the defect or the inside of the material to bypass the defect, part of the magnetic flux will leak to the space above the material surface, bypass the defect through the air and enter the material, forming a leakage magnetic field at the defect. The detection probe 5 includes an internal injury detection probe 501 and an external injury detection probe 502, and the number of turns of the coil in the detection probe 5 is freely replaced according to the national standard, the standard in the industry and different use scenarios to meet different use scenarios. When the measured material placement platform 7 passes under the detection probe 5 at a certain speed, the leakage magnetic field of the magnetized measured material changes when passing through the coil of the detection probe 5, and the changing magnetic flux generates an electromotive force in the coil to obtain a signal, and the signal is proportional to the number of turns of the coil and the magnetic flux, and the internal injury and external injury signals can be obtained through a suitable coil.
[0039] As shown in Figure 4 , the data processing module 6 adopts a TL084 chip, which includes a differential circuit 601, a first-order amplification circuit 602, a second-order active low-pass filter circuit 603 and a second-order amplification circuit 604. By differentially inputting the leakage magnetic signal, the interference of external disturbance on the leakage magnetic detection signal can be effectively reduced, and the anti-interference ability is improved. The detection signal is amplified through the first-order amplification circuit 602, in order to reduce the interference of high-frequency signals generated by amplification, the second-order active low-pass filter circuit 603 is used for filtering, and finally the filtered leakage magnetic detection signal is output through the second-order amplification circuit 604 to realize the output of the leakage magnetic detection signal, and the precision of the leakage magnetic detection signal is improved. The data processing module 6 is connected with the acquisition card, the acquisition card is connected with the upper computer 11, and the upper computer 11 is used to display the leakage magnetic detection signal.
[0040] As shown in Figures 5-6 , the measured material is placed on the measured material placement platform 7 and magnetized by the permanent magnet 12, and a defect leakage magnetic field is generated at the defect position after the measured material reaches magnetic saturation. The leakage magnetic field signal at the defect is obtained through the internal injury detection probe 501 and the external injury detection probe 502. The detection probe 5 contains a coil and is arranged in a column, which increases the detection area and realizes the first 8-way output of internal injury detection and the first 8-way output of external injury detection. The obtained signal is transmitted to the data processing module 6 for signal processing, and the processed signal is transmitted to the upper computer 11 through the acquisition card for analysis. Figure 5 and Figure 6It can be seen that: initially, the detection voltage remains stable, when the measured material passes through the detection probe 5, the detection voltage signal changes suddenly under the action of the magnetic leakage field at the defect to form a defect detection signal, but due to the influence of the material surface due to processing or external environment, other interference signals may be formed, which will affect the recognition effect. When the detection voltage changes suddenly, it indicates that the magnetic field of the measured material surface has changed, indicating that there is damage or defect at this place.
[0041] Embodiment two An AI edge recognition training data construction method for magnetic flux leakage defect recognition, using the collection platform of embodiment one, comprising the steps of: S1, magnetizing the measured material by the permanent magnet 12; S2, collecting the magnetic flux leakage signal by the 16-way internal injury detection probe 501 and the external injury detection probe 502; S3, using the SOGI-FLL module to perform signal screening, and the normalized transfer function expression of the SOGI-FLL module is:
[0042] Wherein, ω is the frequency estimation value, s is the Laplace operator, and k is the adjustment coefficient; The expression of the band-pass filter D(s) and the low-pass filter Q(s) after normalization is: (1); (2).
[0043] In step S3, the SOGI-FLL module completes signal frequency recognition within 0.5 seconds, and the tracking error is less than 5%.
[0044] The signal processing process of the data processing module 6 includes: Sa, differential input stage: using a differential amplifier to receive the double-ended output signal of the detection probe 5, eliminating common-mode interference, and the differential rejection ratio is ≥60dB; Sb, first amplification stage: using an adjustable gain operational amplifier to perform primary amplification on the differential signal, the gain range is 20-40dB, and the bandwidth is 0.1Hz-10kHz; Sc, second-order filtering stage: using a second-order active low-pass filter of Sallen-Key topology; Sd, secondary amplification stage: using an instrument amplifier for accurate amplification, the gain is 40-60dB, and the common-mode rejection ratio is ≥100dB; Se, output conditioning stage: impedance matching is performed through a voltage follower, and the output impedance is ≤100Ω; Among them, the resistance-capacitance coupling mode is adopted between stages, the direct current blocking capacitor has a capacitance value of greater than or equal to 10 mu F, the overall signal link total gain is greater than or equal to 80 dB, and the passband ripple is less than or equal to plus or minus 0.5 dB.
[0045] The cutoff frequency of the second-order active low-pass filter circuit 603 satisfies: ; Among them, r0=r1, c0=nc1, and n is a proportional coefficient.
[0046] Through the symmetric design of r0=r1 and c0=nc1, the cutoff frequency stability is improved, and the passband fluctuation is less than or equal to plus or minus 0.3 dB; the proportional coefficient n is dynamically adjustable to adapt to the frequency response characteristics of different materials.
[0047] The detection probe 5 includes a replaceable coil assembly having a standardized interface and being detachably mounted to the probe housing through a mechanical connector; the parameter configuration of the coil assembly satisfies: The coil of the internal injury detection probe 501 has a number of turns N1 in the range of 800±50 turns and a wire diameter of 0.15±0.02 mm; The coil of the external injury detection probe 502 has a number of turns N2 in the range of 500±30 turns and a wire diameter of 0.2±0.02 mm.
[0048] The coil output voltage of the detection probe 5 satisfies the formula: ; Among them, N is the number of turns, A is the effective cross-sectional area, v is the moving speed, is the magnetic induction intensity gradient. The mathematical relationship between the quantized output voltage and the defect characteristics provides a physical model basis for AI training, and the magnetic induction intensity gradient is coupled with the mechanical motion parameter (v) to build a model.
[0049] As shown in Figure 7 , s is the input signal; s' is the tracking signal of the input; alpha is the gain in the algorithm, and its size determines the convergence of the frequency estimation; f' is the tracking value of the input signal frequency. In the algorithm, the upper channel outputs the same tracking signal s' as the input signal s by using the band-pass characteristic, and the lower channel realizes the conversion of the input signal into its orthogonal component by constructing a closed-loop feedback system and a double-integration link, and realizes synchronous extraction. The frequency error is extracted by using the complex plane relationship formed by the input signal and the orthogonal component, and the f' value is corrected in real time after the gain alpha and the integrator, until it converges to the true frequency.
[0050] As can be seen from Figure 8 , the identification of the collected magnetic flux leakage frequency can be completed in about 0.5 s. As can be seen from Figure 9It can be known that the curve smoothing in the tracked signal eliminates other interference signals, and the signal is used as a data set for AI edge recognition training, so as to ensure the accuracy of the magnetic flux leakage detection information and the AI edge recognition training effect.
[0051] Embodiment three An AI edge recognition training system adopts the data set constructed by the construction method of embodiment two, wherein the data contains original detection signal data, processed feature data and spatial marker data, The original detection signal data includes: An 8-channel signal array for internal injury detection, each channel has a sampling rate of greater than or equal to 100 kHz and a resolution of 16 bits, and contains time domain waveform data and FFT spectrum data; An 8-channel signal array for external injury detection, with a sampling interval of less than or equal to 10 microseconds and a signal amplitude range of ±10V, and with a time stamp marker; The processed feature data includes: Normalized signals processed by SOGI-FLL, including: Fundamental component amplitude A_f, accuracy 0.1mV; Harmonic distortion THD, calculated to the 5th harmonic; Signal-to-noise ratio SNR, measured by A-weighting; Phase difference data φ, resolution 0.1°; The spatial marker data includes: Defect position three-dimensional coordinates (x, y, z), wherein: the x-axis is the motion direction of the conveyor belt, with a positioning accuracy of ±0.5mm; the y-axis is the direction of the probe array, with a positioning accuracy of ±1mm; and the z-axis is the defect depth, obtained by calculating the time difference of the multi-probe signals; Defect size feature vector [L, W, D], L represents length, W represents width, and D represents depth, with a measurement error of less than or equal to 5%.
[0052] The three-dimensional data fusion architecture of time domain, frequency domain and spatial coordinates optimizes the AI training model, the time domain data provides the basis for CNN network training, the frequency domain features are used as the time sequence input of the LSTM network, and the spatial coordinates construct 3D point cloud data for the graph neural network processing, and the three modalities fusion greatly improves the model recognition accuracy.
[0053] Based on the above ideal embodiments according to the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the contents in the specification, and must be determined according to the scope of the claims.
Claims
1. A platform for collecting AI edge recognition training data for magnetic flux leakage defect identification, characterized by: include: A bracket (1), a power module (2), a motor (3), a driving device (4), a detection probe (5), a data processing module (6), a test material placement platform (7), a limit sensor (8), a single chip microcomputer (9), a display module (10) and a host computer (11); The platform (7) for placing the material to be tested is connected to the motor (3) via a conveyor belt; The detection probe (5) comprises an internal injury detection probe (501) and an external injury detection probe (502), each having 16 detection output terminals; The data processing module (6) comprises a differential circuit (601), a first-stage amplification circuit (602), a second-order active low-pass filter circuit (603) and a second-stage amplification circuit (604).
2. The AI edge recognition training data acquisition platform for magnetic flux leakage defect identification according to claim 1 is characterized in that: The single-chip microcomputer (9) controls the driving device (4) through PWM waves. The bracket (1) is provided with forward and reverse buttons (101) and a speed adjustment knob (102). The limit sensors (8) are provided at both ends of the guide rail and connected to the single-chip microcomputer to achieve automatic switching of the moving direction and collision protection.
3. The AI edge recognition training data acquisition platform for magnetic flux leakage defect identification according to claim 1 is characterized in that: The coil output voltage of the detection probe (5) satisfies the formula: ； Among them, N is the number of coil turns, A is the effective cross-sectional area, v is the moving speed, is the magnetic induction intensity gradient.
4. The AI edge recognition training data acquisition platform for magnetic flux leakage defect identification according to claim 1 is characterized in that: The cut-off frequency of the second-order active low-pass filter circuit (603) satisfies: ； Among them, r0=r1, c0=nc1, and n is the proportional coefficient.
5. The AI edge recognition training data acquisition platform for magnetic flux leakage defect identification according to claim 1 is characterized in that: It also includes a permanent magnet (12), which is fixed to the platform (7) for placing the material to be measured and is used to make the material to be measured reach a magnetic saturation state.
6. A method for constructing AI edge recognition training data for magnetic flux leakage defect identification, using the acquisition platform described in any one of claims 1 to 5, characterized in that: Including steps: S1, magnetizing the material to be tested by a permanent magnet (12); S2, collecting magnetic flux leakage signals through 16 internal damage detection probes (501) and external damage detection probes (502); S3. Use the SOGI-FLL module to perform signal screening. The normalized transfer function expression of the SOGI-FLL module is: Among them, ω is the frequency estimate, s is the Laplace operator, and k is the adjustment coefficient; The normalized expressions of the bandpass filter D(s) and the lowpass filter Q(s) are: （1）； （2）。 7. The method for constructing AI edge recognition training data for magnetic flux leakage defect identification according to claim 6, characterized in that: In step S3, the SOGI-FLL module completes signal frequency identification within 0.5 seconds, and the tracking error is less than 5%.
8. The method for constructing AI edge recognition training data for magnetic flux leakage defect identification according to claim 6, characterized in that: The signal processing flow of the data processing module (6) includes: Sa, differential input stage: using a differential amplifier to receive the double-ended output signal of the detection probe (5) to eliminate common-mode interference, with a differential rejection ratio of ≥60dB; Sb, primary amplification stage: the differential signal is amplified by an adjustable gain operational amplifier with a gain range of 20-40dB and a bandwidth of 0.1Hz-10kHz; Sc, second-order filtering stage: a second-order active low-pass filter using Sallen-Key topology; Sd, secondary amplification stage: using instrumentation amplifier for precise amplification, gain 40-60dB, common mode rejection ratio ≥100dB; Se, output conditioning stage: impedance matching is performed through a voltage follower, and the output impedance is ≤100Ω; Among them, the resistance-capacitance coupling method is adopted between each stage, the capacitance of the DC blocking capacitor is ≥10μF, the total gain of the overall signal chain is ≥80dB, and the passband ripple is ≤±0.5dB.
9. The method for constructing AI edge recognition training data for magnetic flux leakage defect identification according to claim 6, characterized in that: The detection probe (5) includes a replaceable coil assembly having a standardized interface and being detachably mounted on the probe housing via a mechanical connector; the parameter configuration of the coil assembly satisfies: The number of turns N1 of the coil of the internal injury detection probe (501) is in the range of 800±50 turns, and the wire diameter is 0.15±0.02 mm; The number of turns N2 of the coil of the trauma detection probe (502) is in the range of 500±30 turns, and the wire diameter is 0.2±0.02 mm.
10. An AI edge recognition training system, characterized in that: A data set constructed using the method for constructing AI edge recognition training data for magnetic flux leakage defect identification according to any one of claims 6 to 9, wherein the data includes original detection signal data, processed feature data, and spatial marker data. Raw heartbeat data includes: Internal injury detection 8-channel signal array, each channel sampling rate ≥ 100kHz, 16-bit resolution, including time domain waveform data and FFT spectrum data; Trauma detection 8-channel signal array, sampling interval ≤ 10μs, signal amplitude range ±10V, with time stamp mark; The processed feature data includes: The normalized signal after SOGI-FLL processing includes: Fundamental frequency component amplitude A_f, accuracy 0.1mV; Harmonic distortion THD, calculated up to the 5th harmonic; Signal-to-noise ratio (SNR), measured using A-weighting; Phase difference data φ, resolution 0.1°; Spatial labeling data includes: The three-dimensional coordinates (x, y, z) of the defect position, where the x-axis is the conveyor belt movement direction, with a positioning accuracy of ±0.5mm; the y-axis is the probe array direction, with a positioning accuracy of ±1mm; and the z-axis is the defect depth, obtained by calculating the time difference of multi-probe signals. Defect size characteristic vector [L, W, D], L represents length, W represents width, and D represents depth. The measurement error is ≤5%.