Terahertz wave-based nondestructive perspective detection method for internal defects of magnetic ring

Abstract

The invention discloses a magnet ring internal defect nondestructive perspective detection method based on terahertz waves, and relates to the technical field of electromagnetis.The method comprises the steps that 3, a network convolution module processes a terahertz wave defect image, a channel attention module and a space attention module are added to the network convolution module to reduce defect edge blurring, and the defect edge blurring is determined; an incomplete block exists between every two network convolution modules, and a channel attention module and a space attention module are incorporated into each incomplete block. Through the channel attention module and the space attention module, dynamic enhancement of defect sensitive frequency band features and concentration of space position coordinates of a defect area are realized, the problems of edge blurring and inaccurate feature extraction of a terahertz wave defect image are solved, defect edge blurring is reduced, defect features are extracted more accurately, and the detection accuracy is improved. And the defect identification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electromagnetism, in particular to a magnetic ring internal defect nondestructive perspective detection method based on terahertz waves. BACKGROUND

[0002] Magnetic rings are widely used in power systems for devices such as transformers and reactors, and the quality thereof is directly related to the stable operation of the power system. With the continuous development of the power industry, the requirements for magnetic ring quality detection are increasing. Traditional detection methods may not meet the needs of high-precision and high-reliability detection of internal defects of magnetic rings.

[0003] Ultrasonic waves are difficult to check for rough, irregular-shaped, small, thin or non-homogeneous materials, are not suitable for structures with cavities, and are not suitable for complex-shaped magnetic rings and narrow cracks, have great limitations, low magnetic flux leakage detection sensitivity, rough defect quantification, and environmental changes affect the detection process, resulting in data deviation

[0004] Therefore, it is necessary to provide a magnetic ring internal defect nondestructive perspective detection method based on terahertz waves to solve the problems in the background.

[0005] Patent document CN106353324B discloses a magnetic ring surface defect extraction method, which realizes high operation efficiency and high accuracy, strong stability and robustness, and can identify common defects.

[0006] In summary, the above-mentioned patent identifies common defects by separating magnetic ring images, has high operation efficiency and high accuracy, but has yet to solve a series of derivative problems such as defect edge blurring and whether it will damage the internal magnetic ring when detecting the internal and complex-shaped cavity structure of the magnetic ring using terahertz waves.

[0007] Therefore, the present application proposes a magnetic ring internal defect nondestructive perspective detection method based on terahertz waves, which dynamically enhances defect-sensitive frequency band features by introducing a channel attention module and a spatial attention module in a network convolution module, and accurately concentrates the spatial coordinates of the defect area. SUMMARY

[0008] The present application aims to provide a magnetic ring internal defect nondestructive perspective detection method based on terahertz waves to solve the technical problems of terahertz wave defect image edge blurring, inaccurate feature extraction, poor information transmission between different levels of networks, single scanning step unable to adapt to different defect situations, and great limitations in the above background.

[0009] To achieve the above-mentioned purpose, the present application provides the following technical solution: a magnetic ring internal defect nondestructive perspective detection method based on terahertz waves,

[0010] The detection method comprises the following steps:

[0011] Step one, the magnetic ring is fixed on the clamp, the center of the magnetic ring is aligned with the terahertz beam by adjusting the clamp, the linear movement range and the rotation angle are set;

[0012] Step two, the detection platform performs spiral scanning, the scanning step is 50 microns to 500 microns, the motor position is fed back by the encoder, the terahertz wave signal after penetration is accepted by the detector, and the signal data of each scanning point is uploaded;

[0013] Step three, the network convolution module processes the terahertz wave defect image, the network convolution module adds a channel attention module and a spatial attention module to reduce the defect edge blur, and there is a residual block between every two network convolution modules, and the channel attention module and the spatial attention module are programmed in each residual block;

[0014] Step four, two cross-scale connection paths are added at the fusion end, the relatively shallow network is transmitted to the relatively deep network by means of the cross-scale connection paths, and the deep semantic information extracted by the relatively deep network is reversely propagated to the relatively shallow network;

[0015] Step five, the C3 module is stacked to form a bottleneck layer, the gradient signs are extracted by stacking the bottleneck layer, and the contact point is led out from the edge of the C3 module to connect the residual block;

[0016] Step six, the formed terahertz wave defect gray-scale image is subjected to binarization processing, the pixel points of each defect are calculated by using a connected region marking method, and the defect area is calculated.

[0017] Preferably, the network convolution module is sequentially connected by a convolution layer, a channel attention submodule and a spatial attention submodule, the channel attention submodule is stacked to form a channel attention module, and the spatial attention submodule is stacked to form a spatial attention module;

[0018] The channel attention submodule generates channel weights through global average pooling, and dynamically enhances defect sensitive frequency band features;

[0019] The spatial attention submodule generates spatial weights through convolution operation, and the spatial weights concentrate the spatial position coordinates of the defect area;

[0020] The input end of the residual block is the residual output of the previous layer, and the output of the residual block is fused with the main network through a jump connection.

[0021] Preferably, the relatively shallow network extracts features containing rich detail information, the relatively shallow network and the relatively deep network are set according to the bidirectional feature pyramid network concept, and the two bidirectional cross-scale connection paths set between the relatively shallow network and the relatively deep network include an up-sampling path and a down-sampling path;

[0022] The up-sampling path inputs an image guide map, and the down-sampling path inputs a binary processed terahertz wave defect gray image, the image guide map is a terahertz wave defect gray image processed by an unsharp mask algorithm, and the bidirectional cross-scale connection paths are merged and then feature extraction is performed by a relatively shallow network and a relatively deep network.

[0023] Preferably, the C3 module is composed of three convolution layers connected in series, and the output of each layer is superimposed with the input of the next layer through a residual connection;

[0024] The output end of the C3 module is connected to the input end of the defective block through a branch to form bidirectional transmission of gradient features, the C3 module stack gradually compresses the channels to extract deeper defect features, and a pooling parameter is introduced at the end of the bottleneck layer to generate a global descriptor of the defect;

[0025] The C3 module is connected through a residual connection, the input of each layer of the C3 module includes the output of the previous layer and the initial input of itself, and the input image of the acquisition unit and the output image extracted by the relatively shallow network are non-overlapping to form a residual.

[0026] Preferably, the up-sampling path transmits the feature map to a relatively deep network, and the relatively deep network improves the resolution through deconvolution and splices a relatively deep feature map;

[0027] The down-sampling path transmits the semantic information of the relatively deep layer to the relatively shallow network through a spatial attention submodule convolution, the perception features of the relatively shallow layer are repeatedly superimposed, an adaptive weight mechanism is used between the paths, and the signal-to-noise ratio of the feature map is dynamically fused.

[0028] Preferably, in the scanning process, the signal processing system is connected to the control unit of the detection platform through a data line in step two;

[0029] When the signal processing system detects that the defect area signal is weak, the control unit switches the scanning density, and the scanning step is reduced to less than 10 microns;

[0030] When the signal processing system detects that the internal integrity of the magnetic ring is 100%, the control unit switches the scanning step to more than 500 microns.

[0031] Preferably, the control unit is connected to an ultrasonic probe, the ultrasonic probe is synchronously triggered with the terahertz wave, and the signal processing system acquires an ultrasonic echo signal at the same position;

[0032] The signal processing system aligns the acquisition times of the terahertz wave signal and the ultrasonic echo signal by using an interpolation algorithm through a time stamp;

[0033] The scanning platform defines three-dimensional coordinates with the initial position of the scanning platform as the origin, records the rotation angle and linear movement coordinates, and uses a geometric transformation matrix to map the terahertz wave signal into the coordinate system created by the scanning platform according to the position of the terahertz wave detector and the beam direction, and uses a geometric transformation matrix to map the ultrasonic wave signal into the coordinate system created by the scanning platform according to the position of the ultrasonic probe and the sound beam direction, and the spatial attention module encodes the linear interpolation to align the scanning point position coordinates of the terahertz wave signal and the ultrasonic wave signal.

[0034] Preferably, the signal processing system comprises a terahertz wave detector, an ultrasonic probe, a data acquisition card and a synchronous trigger unit.

[0035] The terahertz wave detector and the ultrasonic probe transmit signals to the data acquisition card through data lines, the output end of the synchronous trigger unit is connected to the input end of the data acquisition card through data lines, and the synchronous trigger unit triggers the terahertz wave and ultrasonic wave signal acquisition.

[0036] Preferably, the main network comprises a network convolution module, a C3 module and a classifier, the network convolution module and the C3 module are arranged in the classifier, the main network weights the feature vectors of the terahertz wave and the ultrasonic wave through channel weights, and the weighted feature vectors remove redundant information through a dimension reduction algorithm.

[0037] The feature training of the terahertz wave is set as a first classifier, the feature training of the ultrasonic wave is set as a second classifier, the two classifiers respectively output defect types, and the main network outputs a defect type report when the output results of the first classifier and the second classifier are the same.

[0038] When the output results of the first classifier and the second classifier are different, the bottleneck layer calls the defect type probability distribution of each C3 module, the network convolution module allocates weights, the spatial attention module analyzes the consistency of the spatial coordinates of the terahertz wave signal and the ultrasonic wave signal, the position deviation is controlled within ± 50 microns, and the main network outputs a detection result in combination with the difference value of the spatial position coordinates and the weighted defect type probability distribution.

[0039] Preferably, the step three detection platform respectively acquires the first time point acquisition image and the second time point acquisition image of the magnetic ring, the judgment module compares the first time point acquisition image and the second time point acquisition image of the terahertz wave defect gray scale image with the gray scale values of the image guide map respectively, calculates the gray scale difference value, and when the gray scale difference value is within the threshold range, the terahertz wave defect gray scale image is transmitted to the network convolution module.

[0040] Compared with the prior art, the present application has the following advantages:

[0041] 1. The application realizes dynamic enhancement of defect sensitive frequency band features and concentration of spatial position coordinates of defect areas through a channel attention module and a spatial attention module, solves the problems of edge blur of terahertz wave defect images and inaccurate feature extraction, reduces defect edge blur, more accurately extracts defect features, and improves the accuracy of defect recognition;

[0042] 2. The application realizes automatic switching of scanning density by the control unit according to actual conditions through spiral scanning, solves the problem that a single scanning step cannot adapt to different defect conditions, improves detection efficiency, performs targeted scanning on different defect conditions, and ensures that internal information is obtained without damaging the internal part of the magnetic ring;

[0043] 3. The application realizes cross-scale fusion of defect features of a relatively shallow network and a relatively deep network through a bidirectional cross-scale connection path, solves the problem of poor information transmission of different levels of networks, fully utilizes the detailed information of the shallow layer and the semantic information of the deep layer, and improves the accuracy of detection;

[0044] 4. The application realizes multi-modal fusion detection through a synchronous trigger unit, solves the problem of large limitations of single detection, combines the advantages of terahertz waves and ultrasonic waves, and more accurately judges the defect type. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a schematic diagram of the magnetic ring scanning structure of the application;

[0046] Figure 2 is a schematic diagram of the cross-scale connection path of the application;

[0047] Figure 3 is a schematic diagram of the image preprocessing and judgment process of the application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0049] Please refer to Figure 1 , Figure 2 and Figure 3The application provides a kind of embodiment: a kind of based on terahertz wave's magnetic ring internal defect nondestructive perspective detection method, the detection method includes the following steps: step one, magnetic ring is fixed on fixture, adjust fixture and make the center of magnetic ring with terahertz wave beam alignment, set linear movement range and rotation angle;Step two, detection platform carries out spiral scanning, scanning step length takes 50 microns to 500 microns, encoder feedback motor position, detector receives the terahertz wave signal after penetration, upload each scanning point signal data;Step three, network convolution module processes terahertz wave defect image, network convolution module adds channel attention module and spatial attention module to reduce defect edge blur, there is residual block between every two network convolution modules, channel attention module and spatial attention module are encoded in each residual block;

[0050] The step three detection platform respectively acquires the first time point acquisition image and the second time point acquisition image of the magnetic ring, and the judgment module compares the first time point acquisition image and the second time point acquisition image of the terahertz wave defect gray image with the gray scale value of the image guide map, calculates the gray scale difference value, and the terahertz wave defect gray image is transmitted to the network convolution module when the gray scale difference value is within the threshold range.

[0051] Further, a terahertz time-domain spectroscopy system is used, equipped with a 0.1-3THz broadband pulse source with a peak power of 20mW and a time resolution of ≤100fs, a six-axis high-precision motion control platform is used for the scanning platform, which integrates a rotary motor and a linear slide, and a low-temperature cooled ZnTe crystal terahertz detector is used for the detector with a signal-to-noise ratio of ≥80dB.

[0052] The magnetic ring to be measured is installed on a vacuum suction clamp, the center axis of the magnetic ring is adjusted to coincide with the terahertz wave beam axis through a laser-assisted centering device, the linear movement range is ±100mm, so as to cover the full size of the magnetic ring in the axial direction, the rotation angle is 0-360°, the reference scanning is performed for the first time according to the spiral trajectory and the detector to generate a reference image I1 by collecting the terahertz time-domain waveform after penetration, the step is compressed to 50 microns in the same trajectory in the second verification scanning, the scanning speed is reduced to 5mm / s to improve the signal-to-noise ratio, and a verification image I2 is generated, the judgment module compares the gray scale values of the same pixel points (i, j) of I1 and I2, pre-processes I1 using a de-sharpening mask algorithm to enhance the edge contrast and generate an image guide map I guide , the terahertz wave defect gray image is superimposed with the corresponding region of I guide , and input to the network convolution module.

[0053] The input layer receives a 512*512 pixel fusion image, a channel attention module generates channel weights through global average pooling, suppresses background noise of 0.5-1 THz, a spatial attention module generates a spatial weight map by using a 7*7 convolution kernel, focuses on cracks and pores, a residual block is inserted between every two convolution modules to prevent gradient disappearance, the smallest detectable defect is 50 mu m microcrack, the signal-to-noise ratio is greater than or equal to 15 dB, an interface is reserved to support data fusion with ultrasonic and infrared data, random noise is eliminated through two scans, and the defect recognition stability is improved by 40%.

[0054] Referring to Figure 1 and Figure 3 , the present application provides an embodiment: a kind of based on terahertz wave's magnetic ring internal defect nondestructive perspective detection method, step four, two cross-scale connection paths are added in fusion end, shallow network is transmitted to deeper network by cross-scale connection path, and deep semantic information extracted by deeper network is reversely propagated to shallow network;Step five, C3 module is stacked to form bottleneck layer, and gradient signs are extracted by bottleneck layer stacking;Contact point is led out at the edge of C3 module to connect residual block;Step six, the formed terahertz wave defect gray scale image is binarized, the pixel points of each defect are calculated using connected region labeling method, and the area of the defect is calculated;

[0055] The C3 module is composed of three convolution layers connected in series, and the output of each layer is stacked through residual connection to the input of the next layer;The output end of C3 module leads out a contact point and is connected to the input end of residual block through branch connection, forming bidirectional transmission of gradient features, C3 module is stacked to gradually compress channels and extract deeper defect features, and a pooling parameter is introduced at the end of the bottleneck layer to generate a global descriptor of the defect;The input of each layer of C3 module includes the output of the previous layer and the initial input of itself, and the non-overlapping points of the input image of the acquisition unit and the output image extracted by the shallow network form a residual;

[0056] The input end of the residual block is the residual of the output of the previous layer, and the output of the residual block is fused with the main network through jump connection;

[0057] Further, the detection system scans 10 samples, and finds that 2 of them have internal defects, including pores and cracks, and the defect areas are 3.2 mm² and 1.8 mm² respectively, the gradient features extracted by the C3 module show that the edge blur of the pore defect is lower, while the edge blur of the crack defect is higher;

[0058] The magnetic ring to be detected is fixed on the clamp, the clamp is adjusted so that the center of the magnetic ring is aligned with the terahertz beam, the linear movement range is set to ±25mm, the rotation angle is set to 0° to 360°, the detection platform is started for spiral scanning, the scanning step is set to 100 microns, the encoder feeds back the motor position in real time, the detector receives the terahertz wave signal after penetrating the magnetic ring, the signal data of each scanning point is uploaded to the signal processing unit through the data acquisition card, and during the scanning process, if the signal processing system detects that the signal of the defect area is weak and the signal-to-noise ratio is lower than 10dB, the control unit automatically reduces the scanning step to less than 10 microns, and if it is detected that the internal integrity of the magnetic ring is 100%, the scanning step is switched to more than 500 microns;

[0059] The collected terahertz wave signal is processed through a network convolution module, the network convolution module is sequentially connected by a convolution layer, a channel attention submodule and a spatial attention submodule, the channel attention submodule generates channel weights through global average pooling to dynamically enhance defect-sensitive frequency band features; the spatial attention submodule generates spatial weights through convolution operation to concentrate the spatial position coordinates of the defect area, the C3 module is composed of three convolution layers connected in series, the output of each layer is stacked through residual connection to the input of the next layer, the formed terahertz wave defect grayscale image is binarized, the pixel points of each defect are calculated by using the connected region marking method, and the defect area = defect pixel point number x (0.1mm) 2 .

[0060] Referring to Figure 1 and Figure 2 , an embodiment provided by the present application provides a kind of based on terahertz wave's magnetic ring internal defect nondestructive perspective detection method, the network convolution module is sequentially connected by convolution layer, channel attention submodule and spatial attention submodule, channel attention submodule is stacked to form channel attention module, spatial attention submodule is stacked to form spatial attention module;Channel attention submodule generates channel weights through global average pooling to dynamically enhance defect-sensitive frequency band features;Spatial attention submodule generates spatial weights through convolution operation, and spatial weights concentrate the spatial position coordinates of the defect area;

[0061] The main network includes a network convolution module, a C3 module and a classifier, the network convolution module and the C3 module are arranged in the classifier, the main network weights the feature vectors of the terahertz wave and the ultrasonic wave through the channel weights, and the weighted feature vectors remove redundant information through a dimension reduction algorithm;

[0062] The upsampling path delivers the feature map to a deeper network, and the deeper network improves the resolution through deconvolution and splices the deeper feature map; the downsampling path delivers the deeper semantic information to a shallower network through spatial attention sub-module convolution, the perception features of the shallower semantic are repeatedly superimposed, an adaptive weight mechanism is used between the paths, and the signal-to-noise ratio of the feature map is dynamically fused;

[0063] Further, the scanning platform is started to perform spiral scanning, the terahertz beam is incident from the surface of the magnetic ring, is received by the detector after penetrating the sample, the encoder feeds back the motor position in real time, the detector collects the terahertz wave signal of each scanning point, at the same time, the ultrasonic probe and the terahertz wave are synchronously triggered, the ultrasonic echo signal of the same position is collected, the signal processing system aligns the collection time of the two kinds of signals through an interpolation algorithm, and ensures spatial position consistency;

[0064] The collected terahertz wave signal is input into a network convolution module after pretreatment, the network convolution module is sequentially connected by a convolution layer, a channel attention sub-module and a spatial attention sub-module, a time domain peak amplitude and a phase delay of a penetrating waveform are extracted, a two-dimensional amplitude-phase matrix is constructed, an ultrasonic detection point is mapped to a terahertz coordinate system by using a geometric transformation matrix, a 3*3 convolution kernel, a step of 1, padding of 1, 64 output channels, global average pooling, two layers of full connection compression ratio of 16, a 7*7 convolution generates a spatial weight map, an upsampling path is input into a C3 module after splicing with a deep feature, and a downsampling path is weighted and fused to a shallow layer through a spatial attention sub-module.

[0065] Referring to Figure 2 and Figure 3 An embodiment provided by the application: a kind of based on terahertz wave's magnetic ring internal defect nondestructive perspective detection method, the characteristic training of terahertz wave is set as No. 1 classifier, the characteristic training of ultrasonic wave is set as No. 2 classifier, two classifiers respectively output defect type, when the output result of No. 1 classifier and No. 2 classifier is the same, main network produces defect type report; when the output result of No. 1 classifier and No. 2 classifier appears difference, bottleneck layer calls the defect type probability distribution of each layer C3 module, network convolution module allocates weight, spatial attention module analyzes the consistency of terahertz wave signal and ultrasonic signal spatial coordinates, position deviation is controlled in ± 50 microns, main network combines the difference of spatial position coordinates and the defect type probability distribution after weighting and outputs detection result;

[0066] The shallow network extracts features containing rich detail information, and the shallow network and the deep network are provided with two bidirectional cross-scale connection paths including an up-sampling path and a down-sampling path according to a bidirectional feature pyramid network concept; an image guide map is input to the up-sampling path, and a binary processed terahertz wave defect gray image is input to the down-sampling path, the image guide map is a terahertz wave defect gray image processed by a de-sharpening mask algorithm, and the bidirectional cross-scale connection paths are merged and then feature extraction is performed by the shallow network and the deep network;

[0067] The signal processing system comprises a terahertz wave detector, an ultrasonic probe, a data acquisition card and a synchronous trigger unit; the terahertz wave detector and the ultrasonic probe transmit signals to the data acquisition card through data lines, the output end of the synchronous trigger unit is connected to the input end of the data acquisition card through a data line, and the synchronous trigger unit triggers the acquisition of terahertz wave and ultrasonic signals;

[0068] Further, when the output difference between the first classifier and the second classifier exceeds a threshold value, a three-dimensional refocusing scanning mode is triggered;

[0069] First, a ring-shaped scanning track is generated according to a spatial coordinate difference value, a spiral radius compression scan is performed with the controversial point as the center, and the scanning step is adaptively adjusted to 1 / 10 of the defect feature size, then the polarization state modulation of the terahertz wave is synchronously activated, the transmission signals in four orthogonal polarization directions are collected to construct a dielectric tensor feature matrix, and the time-frequency features of the ultrasonic echo signals and the terahertz polarization features are input into a spatial fusion module, the spatial fusion module comprises cascaded 3D convolution kernels and bidirectional LSTM, cross-modal defect dynamic evolution features are extracted, a constitutive relationship verification channel of the internal defect of the magnetic ring is constructed, the measured dielectric tensor is projected and matched in a subspace with a magnetic ring material database, a material abnormality index is calculated, a controversial sample is injected into a pre-trained twin network through transfer learning, multi-scale similarity measurement is performed with a standard defect sample library, and the distribution of the final defect type is output;

[0070] A three-dimensional defect topology map is generated according to the defect type distribution and the spatial fusion module result.

[0071] Please refer to Figure 1 , Figure 2 and Figure 3 , an embodiment provided by the application: a kind of based on terahertz wave's magnetic ring internal defect nondestructive perspective detection method, the step two in scanning process, signal processing system is connected to the control unit of detection platform by data line;Signal processing system detects weak signal in defect area, and control unit switches scanning density, and scanning step is reduced to below 10 microns;Signal processing system detects 100% when the internal integrity of magnetic ring, control unit switches scanning step 500 microns or more;

[0072] The control unit is connected with the ultrasonic probe, the ultrasonic probe is triggered synchronously with the terahertz wave, and the signal processing system acquires the ultrasonic echo signal at the same position; the signal processing system uses an interpolation algorithm to align the acquisition time of the terahertz wave signal and the ultrasonic echo signal through a time stamp; the scanning platform defines a three-dimensional coordinate with the initial position of the scanning platform as the origin, records the rotation angle and the linear movement coordinate, uses a geometric transformation matrix to map the terahertz wave signal into the coordinate system created by the scanning platform according to the position and beam direction of the terahertz wave detector, and uses the geometric transformation matrix to map the ultrasonic wave signal into the coordinate system created by the scanning platform according to the position and beam direction of the ultrasonic wave probe, and the spatial attention module encodes the linear interpolation to align the scanning point position coordinates of the terahertz wave signal and the ultrasonic wave signal.

[0073] Further, a terahertz wave-based detection platform is prepared, which includes a terahertz wave detector, a control unit, a signal processing system and an ultrasonic probe, the ultrasonic probe is connected to the control unit to ensure that it can be triggered synchronously with the terahertz wave, the initial position of the scanning platform is set as the origin to define a three-dimensional coordinate, and the control unit of the detection platform is connected with the signal processing system through a data line for signal transmission and instruction transmission.

[0074] The magnetic ring to be detected is fixed on the clamp, the clamp is adjusted so that the center of the magnetic ring is aligned with the terahertz beam, the scanning step is set to 200 microns, and the detection platform starts spiral scanning. During the scanning process, the terahertz wave detector receives the terahertz wave signal after penetrating the magnetic ring, the ultrasonic probe synchronously acquires the ultrasonic echo signal at the same position, and the signal processing system analyzes the received terahertz wave signal in real time. When a weak signal is detected in the defect area, an instruction is sent to the control unit, the control unit reduces the scanning step to less than 10 microns, and when the internal integrity of the magnetic ring is 100%, the control unit increases the scanning step to more than 500 microns. The signal processing system uses an interpolation algorithm to align the acquisition time of the terahertz wave signal and the ultrasonic echo signal according to a time stamp, to ensure the synchronization of the data of the two.

[0075] The scanning platform uses a geometric transformation matrix to map the terahertz wave signal into the created coordinate system according to the position and beam direction of the terahertz wave detector; similarly, the ultrasonic wave signal is also mapped into the coordinate system according to the position and beam direction of the ultrasonic wave probe, and a linear interpolation algorithm is encoded in the spatial attention module to further align the scanning point position coordinates of the terahertz wave signal and the ultrasonic wave signal.

[0076] The working principle is that the magnetic ring is fixed on the clamp, the center of the magnetic ring is aligned with the terahertz beam, the linear movement range and the rotation angle are set, the detection platform is helically scanned at a scanning step of 50-500 microns, the encoder feeds back the motor position during scanning, the detector receives the terahertz wave signal after penetrating the magnetic ring and uploads the data, the signal processing system is connected with the control unit, and the scanning step is dynamically adjusted according to the signal strength;

[0077] The control unit is connected with the ultrasonic probe, so that the ultrasonic probe is triggered synchronously with the terahertz wave to obtain the ultrasonic echo signal at the same position, the interpolation algorithm and the geometric transformation matrix are used to align the terahertz wave signal and the ultrasonic echo signal in time and space, and the terahertz wave signal and the ultrasonic echo signal are mapped into the three-dimensional coordinate system created by the scanning platform, meanwhile, the detection platform obtains the acquisition images of the magnetic ring at different time points, the gray value comparison is performed between the terahertz wave defect gray image and the image guide image through the judgment module, the network convolution module processes the terahertz wave defect image, the feature extraction effect is enhanced through the channel attention module and the spatial attention module, the residual block between every two network convolution modules receives the residual of the previous layer and is fused with the main network, the relatively shallow network and the relatively deep network perform feature transmission and fusion through the up-sampling and down-sampling paths according to the bidirectional feature pyramid network concept, the C3 module is composed of three serial convolution layers, is stacked into a bottleneck layer through residual connection, extracts deep defect features and generates a global descriptor, and also forms gradient feature bidirectional transmission with the residual block.

[0078] The main network weights and reduces the dimension of the feature vectors of the terahertz wave and the ultrasonic wave, respectively trains the features by two classifiers, and finally performs binaryzation processing on the terahertz wave defect gray image, and calculates the defect pixel points and the area.

[0079] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and that the present application can be implemented in other particular forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, and the scope of the present application should be defined by the appended claims rather than the above description, and it is intended to encompass all changes falling within the meaning and scope of equivalents of the claims. Any reference signs in the claims should not be considered as limiting the claims involved.

Claims

1. A non-destructive imaging method for detecting internal defects in a magnetic ring based on terahertz waves, characterized in that: The detection method includes the following steps: Step 1: Fix the magnetic ring on the clamp, adjust the clamp to align the center of the magnetic ring with the terahertz beam, and set the linear movement range and rotation angle; Step 2: The detection platform performs a spiral scan with a scanning step size of 50 micrometers to 500 micrometers. The encoder provides feedback on the motor position, the detector receives the terahertz wave signal after penetration, and uploads the signal data of each scanning point. Step 3: The network convolution module processes the terahertz wave defect image. The network convolution module adds channel attention module and spatial attention module to reduce the blurring of defect edges. There is a missing block between every two network convolution modules. The channel attention module and spatial attention module are programmed into each missing block. Step 4: Add two cross-scale connection paths at the fusion end. The shallower network uses the cross-scale connection paths to pass to the deeper network, and the deep semantic information extracted by the deeper network is back propagated to the shallower network. Step 5: Stack C3 modules to form a bottleneck layer. Stack the bottleneck layers to form gradient feature extraction. Connect the missing blocks by leading out contact points at the edge of C3 modules. Step 6: Binarize the resulting terahertz wave defect grayscale image, calculate the pixel points of each defect using the connected component labeling method, and determine the defect area.

2. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 1, characterized in that: The network convolutional module consists of convolutional layers, channel attention submodules, and spatial attention submodules connected in sequence. Channel attention submodules are stacked to form a channel attention module, and spatial attention submodules are stacked to form a spatial attention module. The channel attention submodule generates channel weights through global average pooling, dynamically enhancing the features of defect-sensitive frequency bands; The spatial attention submodule generates spatial weights through convolution operations, and the spatial weights concentrate the spatial coordinates of the defect region. The input of the incomplete block is the residual of the previous layer's output, and the output of the incomplete block is fused with the main network through a skip connection.

3. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 1, characterized in that: The shallower network extracts features containing rich details. The shallower and deeper networks are based on the concept of a bidirectional feature pyramid network. The two bidirectional cross-scale connection paths set between the shallower and deeper networks include an upsampling path and a downsampling path. The upsampling path inputs the image guide map, and the downsampling path inputs the binarized grayscale image of the terahertz wave defect. The image guide map is the grayscale image of the terahertz wave defect processed by the anti-sharpening mask algorithm. After the bidirectional cross-scale connection paths merge, features are extracted by the shallower and deeper networks.

4. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 1, characterized in that: The C3 module consists of three cascaded convolutional layers, with the output of each layer being superimposed on the input of the next layer via residual connections. The output of the C3 module leads to a contact point that is connected to the input of the incomplete block through a branch, forming a bidirectional transmission of gradient features. The C3 modules are stacked to gradually compress the channels and extract deeper defect features. At the end of the bottleneck layer, pooling parameters are introduced to generate a global descriptor of the defect. The C3 module uses residual connections. The input of each layer of the C3 module includes the output of the previous layer and its own initial input. The non-overlapping points between the input image of the acquisition unit and the output image extracted by the shallower network form residuals.

5. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 3, characterized in that: The upsampling path passes the feature map to a deeper network, and the deeper network improves the resolution through deconvolution and concatenates it with the deeper feature map; The downsampling path passes deeper semantic information to shallower network layers through convolution of the spatial attention submodule. The perceptual features of shallower semantics are repeatedly superimposed, and an adaptive weighting mechanism is used between paths. The signal-to-noise ratio of the feature maps is dynamically fused.

6. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 1, characterized in that: In step two, during the scanning process, the signal processing system is connected to the control unit of the detection platform via a data cable; When the signal processing system detects a weak signal in the defect area, the control unit switches the scanning density and reduces the scanning step size to less than 10 micrometers. When the signal processing system detects that the internal integrity of the magnetic ring is 100%, the control unit switches to a scan step size of 500 micrometers or more.

7. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 6, characterized in that: The control unit is simultaneously connected to the ultrasonic probe, which is synchronously triggered with the terahertz wave, and the signal processing system acquires the ultrasonic echo signal from the same location. The signal processing system uses an interpolation algorithm to align the acquisition time of the terahertz wave signal and the ultrasonic echo signal using timestamps; The scanning platform defines three-dimensional coordinates with its initial position as the origin, and records the rotation angle and linear movement coordinates. The scanning platform uses a geometric transformation matrix to map the terahertz wave signal to the coordinate system created by the scanning platform according to the position and beam direction of the terahertz wave detector. The scanning platform uses a geometric transformation matrix to map the ultrasonic signal to the coordinate system created by the scanning platform according to the position and beam direction of the ultrasonic probe. The spatial attention module incorporates linear interpolation to align the scanning point position coordinates of the terahertz wave signal and the ultrasonic signal.

8. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 6, characterized in that: The signal processing system includes a terahertz wave detector, an ultrasonic probe, a data acquisition card, and a synchronization triggering unit; The terahertz wave detector and ultrasonic probe transmit signals to the data acquisition card via a data cable. The output of the synchronous trigger unit is connected to the input of the data acquisition card via a data cable, and the synchronous trigger unit triggers the acquisition of terahertz wave and ultrasonic signals.

9. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 2, characterized in that: The main network includes a network convolution module, a C3 module, and a classifier. The classifier contains a network convolution module and a C3 module. The main network weights the feature vectors of terahertz waves and ultrasonic waves through channel weights. The weighted feature vectors are then subjected to a dimensionality reduction algorithm to remove redundant information. The features of terahertz waves are trained to classify the first classifier, and the features of ultrasound are trained to classify the second classifier. The two classifiers output the defect type respectively. When the output of the first classifier and the second classifier are the same, the main network generates a defect type report. When the output results of classifier 1 and classifier 2 differ, the bottleneck layer retrieves the defect type probability distribution of each C3 module, the network convolution module assigns weights, the spatial attention module analyzes the consistency of spatial coordinates of terahertz wave signal and ultrasonic signal, the position deviation is controlled within ±50 micrometers, and the main network combines the difference in spatial position coordinates and the weighted defect type probability distribution to output the detection result.

10. The non-destructive testing method for internal defects of a magnetic ring based on terahertz waves according to claim 1, characterized in that: In step three, the detection platform acquires images of the magnetic ring at the first and second time points, respectively. The judgment module compares the first and second time point images of the terahertz wave defect grayscale image with the grayscale values ​​of the image guide map, calculates the grayscale difference, and if the grayscale difference is within the threshold range, the terahertz wave defect grayscale image is transmitted to the network convolution module.

Images