An ultrasonic testing device and method for internal defects in magnetic tiles

Abstract

This invention discloses an ultrasonic testing device and method for internal defects of magnetic tiles, relating to the field of magnetic tile quality inspection technology. The method includes placing the magnetic tile to be tested on the ultrasonic testing device, transmitting ultrasonic pulses to the tile, acquiring the corresponding ultrasonic echo signal, mapping the ultrasonic echo signal to a two-dimensional image, inputting the obtained two-dimensional image into a pre-trained magnetic tile defect recognition model to obtain defect recognition results, and sorting the magnetic tiles according to the defect detection results. This invention achieves low-cost, non-contact, and high-efficiency online scanning and sorting of magnetic tiles, avoiding secondary surface damage that may be caused by traditional contact testing, thereby promoting quality and efficiency improvement for magnetic tile manufacturing enterprises.

Description

Technical Field

[0001] This invention relates to the field of magnetic tile quality inspection technology, and in particular to an ultrasonic testing device and method for internal defects in magnetic tiles. Background Technology

[0002] Magnet tiles are core components in electromagnetic equipment such as motors, wind power generation, and aerospace. Their internal quality directly determines the energy conversion efficiency and lifespan of the final product. However, the production of magnet tiles involves complex processes such as pressing and high-temperature sintering. Uneven mold pressure distribution and fluctuations in heat treatment temperature control can easily lead to defects such as cracks, pores, and porosity within the magnet tiles. Statistical data shows that these invisible internal defects account for a very high percentage of all quality problems with magnet tiles, and their distribution is random and concealed, seriously threatening the operational safety of electromagnetic equipment. Therefore, comprehensive internal defect detection of magnet tiles is a crucial aspect of quality control.

[0003] Traditional methods for detecting internal defects in magnetic tiles primarily rely on manual tapping, where inspectors judge the presence of voids or cracks by listening to the differences in sound produced when the tiles are struck. This method is highly susceptible to human error, lacks quantifiable standards, and is prone to fatigue errors on high-speed production lines. While automated solutions such as acoustic emission testing have emerged in recent years, sound signals are easily affected by environmental noise, and feature extraction is challenging. Traditional ultrasonic C-scan testing, although highly accurate, suffers from expensive equipment and long scanning cycles, making it unsuitable for the online inspection needs of large-volume magnetic tile production. Therefore, developing a practical, efficient, and novel device and method for detecting internal defects in magnetic tiles is of great significance. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides an ultrasonic testing device and method for internal defects in magnetic tiles, which solves the problems of high cost, low efficiency, and difficulty in adapting to the needs of large-scale online testing of magnetic tiles.

[0005] To achieve the above-mentioned objective, the present invention provides an ultrasonic testing method for internal defects in magnetic tiles, comprising: An ultrasonic pulse is emitted toward the magnetic tile under test, and the corresponding ultrasonic echo signal of the magnetic tile under test is obtained. The ultrasonic echo signal corresponding to the magnetic tile under test is mapped into a two-dimensional image; The obtained two-dimensional image is input into a pre-trained magnetic tile defect recognition model to obtain the defect recognition result.

[0006] Secondly, the present invention also provides an apparatus for implementing an ultrasonic testing method for internal defects of magnetic tiles, including a frame; a water tank located below the frame; a movable component suspended from the frame beam; an integrated probe device installed at the bottom of the movable component; and a testing platform located below the integrated probe device and inside the water tank; the integrated probe device and the movable component constitute a defect detection mechanism. The integrated probe device includes an ultrasonic flaw detector, three parallel ultrasonic transducers, a data acquisition card, a signal preprocessing module, and a defect recognition module with a magnetic tile defect recognition model deployed thereon. The ultrasonic flaw detector is used to excite the three parallel ultrasonic transducers; the three parallel ultrasonic transducers are used to emit ultrasonic pulses towards the magnetic tile under test; the data acquisition card is used to acquire the ultrasonic echo signal corresponding to the magnetic tile under test; the signal preprocessing module is used to map the ultrasonic echo signal corresponding to the magnetic tile under test into a two-dimensional image; and the defect recognition module is used to input the obtained two-dimensional image into a pre-trained magnetic tile defect recognition model to obtain the defect recognition result.

[0007] The beneficial effects of this invention are as follows: 1. The process of detecting internal defects in magnetic tiles is fully automated by using water immersion multi-channel ultrasonic non-destructive testing. By integrating multi-transducer probes, moving components and self-unloading tilting testing tables, low-cost, non-contact and high-efficiency online scanning and sorting are achieved, solving the problems of low automation and high dependence on manual inspection in the production of magnetic tiles.

[0008] 2. By mapping the acquired multi-channel one-dimensional A-scan raw data into two-dimensional feature images representing the internal structure, and using a deep residual network for feature extraction and learning, while introducing the FocalLoss loss function to address the small sample imbalance problem, the model's ability to capture minute internal defects and its robustness are significantly enhanced. This improves the detection accuracy and the stability of industrial deployment, solves the problem of significant fluctuations in detection accuracy and speed due to human factors, and avoids secondary surface damage to magnetic tiles that may be caused by traditional contact detection, thereby promoting quality and efficiency improvement for magnetic tile manufacturing enterprises. Attached Figure Description

[0009] Figure 1 A structural diagram of the ultrasonic testing device for internal defects of magnetic tiles provided in the embodiment; Figure 2 Top view of the ultrasonic testing device for internal defects of magnetic tiles provided in the embodiment; Figure 3 Side view of the ultrasonic testing device for internal defects of magnetic tiles provided in the embodiment; Figure 4 Cross-sectional view of the ultrasonic testing device for internal defects of magnetic tiles provided in the embodiment; Figure 5Flowchart of the ultrasonic testing method for internal defects of magnetic tiles provided in the embodiment; Figure 6 The ultrasonic signal time-domain diagram of the defect location of the magnetic tile provided in the embodiment; Figure 7 The ultrasonic signal time-domain diagram of the defect-free magnetic tile provided in the embodiment; Figure 8 Example two-dimensional image of a defective magnetic tile; Figure 9 Example two-dimensional image of a qualified magnetic tile; Figure 10 A schematic diagram of the magnetic tile defect identification model structure provided in the embodiment; Figure 11 The convergence curve of the loss function for training the model; Figure 12 This is a convergence curve for the accuracy of the trained model.

[0010] The components are: 1. First conveyor belt; 2. Trapezoidal inclined plane; 3. Detection table; 4. Support mechanism; 41. First support column; 42. Second support column; 43. Third support column; 44. Fourth support column; 45. Electrically controlled telescopic column; 5. Conveying device; 6. Second conveyor belt; 7. Sorting plate; 8. Integrated probe device; 9. Moving component; 10. Frame; 11. Water tank; 12. Third conveying section. Detailed Implementation

[0011] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0012] like Figure 1 , Figure 2 , Figure 3 , Figure 4 As shown, in one embodiment of the present invention, the ultrasonic testing device for internal defects of magnetic tiles includes a frame 10; a water tank 11 located below the frame 10; a movable component 9 suspended from the crossbeam of the frame 10; an integrated probe device 8 installed at the bottom of the movable component 9; and a testing platform 3 located below the integrated probe device 8 and inside the water tank 11. The integrated probe device 8 and the movable component 9 constitute a defect detection mechanism. The integrated probe device 8 includes an ultrasonic flaw detector, three parallel ultrasonic transducers with a center frequency of 4MHz (this frequency selection is configured to ensure the sound wave penetration depth while generating a high-sensitivity response to millimeter-level pores inside the magnetic tile), a data acquisition card, a signal preprocessing module, and a defect recognition module with a magnetic tile defect recognition model deployed thereon. The ultrasonic flaw detector is used to excite the three parallel ultrasonic transducers; the three parallel ultrasonic transducers are used to emit ultrasonic pulses towards the magnetic tile under test; the data acquisition card is used to acquire the ultrasonic echo signal corresponding to the magnetic tile under test; the signal preprocessing module is used to map the ultrasonic echo signal corresponding to the magnetic tile under test into a two-dimensional image; and the defect recognition module is used to input the obtained two-dimensional image into a pre-trained magnetic tile defect recognition model to obtain the defect recognition result.

[0013] A first conveyor belt 1 is placed on the outer side of the feed side of the water tank 11; a trapezoidal inclined surface 2 is placed on the inner side of the feed side of the water tank 11, and both the trapezoidal inclined surface 2 and the first conveyor belt 1 are in close contact with the water tank 11; the far end of the trapezoidal inclined surface 2 is on the same horizontal line as the first conveyor belt 1, and the near end of the trapezoidal inclined surface 2 is on the same horizontal line as the testing platform 3; a support mechanism 4 is provided below the testing platform 3; the support mechanism 4 includes a first support column 41, a second support column 42, a third support column 43, a fourth support column 44, and an electrically controlled telescopic column 45; the first support column 41 and the second support column 42 are located at the two ends of the testing platform 3 on the side near the trapezoidal inclined surface 2; the third support column 43 is located at the midpoint of the testing platform 3 on the side far from the trapezoidal inclined surface 2; the fourth support column 44 is parallel to the third support column 43 and located on the horizontal center line of the testing platform 3. Above; the electrically controlled telescopic column 45 has a V-shaped structure, and the intersection of its two branches is connected to the fourth support column 44 in a K-shape. The two branches extend outward obliquely from the connection point with the fourth support column 44 to the bottom of the water tank 11 and the bottom surface of the detection platform 3, respectively. To prevent it from falling during the conveying process in a water immersion environment, a conveying device 5 with a chain plate conveyor belt with positioning slots is provided on the far trapezoidal inclined surface 2 side of the detection platform 3. The conveying device 5 includes a first conveying section, a second conveying section, and a third conveying section 12. The first conveying section is a horizontal receiving section and is lower than the initial horizontal height of the detection platform 3. The second conveying section is an inclined conveying section, and the third conveying section 12 is a horizontal unloading section. A second conveyor belt 6 is provided below the third conveying section 12. A sorting paddle 7 is provided between the discharge side frame 10 of the water tank 11 and the edge of the water tank 11. When the test is completed, the electrically controlled telescopic column 45 extending to the bottom of the test platform 3 extends until the test platform 3 forms a 45-degree tilt angle with the horizontal line, so that the magnetic tile on the test platform 3 slides into the conveying device 5. When the test result is a defective magnetic tile, the sorting plate 7 rotates at a preset angle to move the defective magnetic tile that has been conveyed to the third conveyor section 12 to the second conveyor belt 6; when the test result is qualified, the sorting plate remains in its original state.

[0014] The moving component 9 includes a rotatable assembly connected to the integrated probe device 8, a telescopic column, and a drive unit mounted on the crossbeam of the frame 10. The moving component 9 dynamically adjusts the probe position according to the radius of curvature of the magnetic tile under test, ensuring that the sound beams emitted by the three ultrasonic transducers are always perpendicularly incident on the surface of the magnetic tile. The water tank 11 is filled with a coupling water medium, and the integrated probe device 8 and the magnetic tile are acoustically coupled through the water layer, thereby minimizing interface reflection dissipation and clutter interference caused by surface roughness. This results in a consistent gain and stable original echo signal, providing high signal-to-noise ratio (SNR) underlying data support for subsequent two-dimensional feature map reconstruction based on curvature compensation. The moving component 9 is configured to drive the integrated probe device 8 to move relative to the water tank 11 in the front-back and up-down directions to achieve full coverage scanning of the magnetic tile surface.

[0015] The ultrasonic flaw detector, three parallel ultrasonic transducers, data acquisition card, signal preprocessing module, and defect identification module with magnetic tile defect identification model are electrically connected in series via a coaxial cable; the transmission device 5 and the detection table 3 are controlled by a logic interlock to ensure that the magnetic tile is completely slid out before the next cycle of detection and positioning is carried out.

[0016] The method of detecting internal defects in magnetic tiles using an ultrasonic testing device is as follows: Figure 5 As shown, it includes the following steps: S1. The magnetic tile to be tested is placed on the first conveyor belt 1. When it is conveyed to the trapezoidal inclined plane 2, it slides into the testing platform 3 in the water tank 11 by its own gravity for positioning. The moving component 9 is moved above the magnetic tile to be tested and automatically inspects it according to the preset 3×5 matrix point path on the magnetic tile. The ultrasonic flaw detector excites the three transducers to synchronously emit ultrasonic pulses, which are coupled into the magnetic tile through water. The data acquisition card acquires three one-dimensional A-scan echo signals at a sampling frequency of not less than 100MHz. The time-domain diagram of the ultrasonic signal at the defect point of the magnetic tile is shown in the figure. Figure 6 As shown, the time-domain diagram of the ultrasonic signal of the corresponding defect-free magnetic tile at the same location is as follows. Figure 7 As shown.

[0017] The above steps construct a high-speed digital signal acquisition architecture based on water immersion coupling. The ultrasonic flaw detector excites the three transducers in the integrated probe device 8 to achieve non-contact, non-destructive excitation and echo conditioning of the internal structure of the magnetic tile in a constant water path environment. The ultrasonic flaw detector works in conjunction with the high-speed data acquisition card to perform synchronous multi-channel data acquisition with a single-point duration of 25μs under a 100MHz sampling frequency and 16-bit quantization accuracy configuration, ensuring high-precision acquisition of the full-range temporal path characteristics covering the surface, interior, and bottom of the magnetic tile. In the preprocessing stage, the binary data stream output by the acquisition card dynamically extracts effective signal segments with a length not less than 3 times the nominal thickness of the round-trip sound time according to the sound velocity characteristics of the magnetic tile material. This removes invalid background clutter while fully preserving key spatiotemporal feature information such as interface waves, defect scattering waves, and bottom wave attenuation required for subsequent two-dimensional image reconstruction.

[0018] S2. The signal preprocessing module is used to map the ultrasonic echo signal corresponding to the magnetic tile under test into a two-dimensional image.

[0019] The obtained two-dimensional images of defective and qualified magnetic tiles are as follows: Figure 8 , Figure 9 As shown.

[0020] Specifically, it includes: S2-1. Extract the effective echo segment from the pause time of each point in the ultrasonic echo signal corresponding to the magnetic tile to be tested, ensuring that the signal contains a complete time-domain sequence of the upper surface wave, internal defect wave and bottom surface echo of the magnetic tile. S2-2. Perform discrete signal parameterization extraction based on multidimensional physical features on the obtained complete time-domain sequence to obtain a time-domain sequence containing sound speed attenuation features, energy product features and frequency domain dominant frequency features.

[0021] Specifically, for the 3×5 matrix scanning points preset on the surface of the magnetic tile to be tested, the integrated probe device 8 collects a set of ultrasonic echo signals at each point; from the ultrasonic echo signals corresponding to each point, three types of physical parameters sensitive to defect characteristics are collected, including: The sound velocity attenuation characteristic is expressed as follows:

[0022] In the formula, Indicates the characteristics of sound velocity attenuation. This indicates the physical distance between the probe emitting the ultrasonic pulse and the ground surface of the magnetic tile being tested. This indicates the flight time of an ultrasonic wave from its emission to the capture of its peak value. The energy product characteristic is expressed as follows:

[0023] In the formula, Indicates the energy product characteristic. This represents the preset normalized balance coefficient. Indicates the total number of sampling points. Indicates the sampling point index. The amplitude of the discrete signal within the time-domain sampling window; The frequency domain dominant frequency characteristic is expressed as follows:

[0024]

[0025] In the formula, The amplitude represents the frequency domain distribution. It is a natural constant. The imaginary unit, Pi This is the index of the sampling point in the frequency domain; This represents the frequency characteristic point corresponding to the maximum amplitude value in the frequency domain distribution, i.e., the dominant frequency characteristic in the frequency domain.

[0026] S2-3. Use a path compensation algorithm based on arc curvature to correct the geometric attenuation difference of a time-domain sequence that includes sound velocity attenuation characteristics, energy product characteristics, and frequency domain dominant frequency characteristics.

[0027] Specifically, the acoustic path increment factor is calculated based on the coordinates of each scanning point on the arc surface and its corresponding radius of curvature; The time-domain sequence is corrected based on the acoustic path increment factor, and its expression is as follows:

[0028] In the formula, These are the corrected signal characteristic values. These are the characteristic values ​​of the signal before correction. The sound attenuation coefficient of the material. To follow the curvature of the magnetic tile The varying sound path increment factor.

[0029] S2-4, Perform image fusion and spatial interpolation generation. The extracted V, P, Three feature parameters are used as multi-channel pixel values ​​to fill the corresponding coordinates of the matrix. For the blank areas between points in the 3×5 matrix, ordinary kriging interpolation is used to complete the attribute information between the points, reconstructing the discrete one-dimensional waveform features into a continuous two-dimensional B-scan feature image that can intuitively represent the depth, width, and energy distribution of defects. The distribution of color intensity in the image intuitively represents the strength changes of the parameter values. Finally, morphological filtering is used to remove random noise, and grayscale linear mapping is performed to enhance the edge texture features of defects such as cracks and pores.

[0030] S3. Input the obtained two-dimensional image into the pre-trained magnetic tile defect recognition model through the defect recognition module to obtain the defect recognition result.

[0031] The magnetic tile defect identification model has good cross-platform compatibility and can run on computers running Linux embedded systems, Windows industrial operating systems, or high-performance server environments. It also supports the real-time distribution of judgment results to the production line PLC execution system via industrial Ethernet.

[0032] like Figure 10 As shown, the magnetic tile defect identification model includes a normalization layer, a front-end convolutional layer, a batch normalization layer, a ReLU activation function layer, a max pooling layer, a first residual module group, a second residual module group, a global average pooling layer, a regularization layer, a fully connected layer, and a softmax activation function layer connected in sequence. The first residual module group includes three residual modules connected in sequence; the second residual module group includes four residual modules connected in sequence. The residual module includes a main path and a shortcut path; the main path sequentially includes a 3×3 convolutional layer, a batch normalization layer, a ReLU activation function layer, a 3×3 convolutional layer, and a batch normalization layer; the shortcut path includes a 1×1 convolutional layer; the output of the residual module is the result obtained by adding the output of the main path and the output of the shortcut path element by element and applying the ReLU activation function.

[0033] The model employs a deep residual network (ResNet) that incorporates the Focal Loss loss function. It extracts minute texture features and amplitude variations in the depth direction from the image through multi-layer residual mapping, identifies pores or cracks caused by uneven density of the magnetic tiles, and outputs binary classification results with confidence in real time.

[0034] The specific processing logic using this model during training is as follows: First, data input and dimensional reconstruction are performed. The model first receives the original one-channel one-dimensional A-scan signal obtained by matrix point scanning, and reconstructs its spatiotemporal features according to its spatial coordinate mapping relationship on the magnetic tile arc surface. The sound velocity, energy, and frequency domain feature vectors extracted from each acquisition point are mapped according to their physical location into a two-dimensional feature matrix containing magnetic tile depth information and arc surface spatial topology information, and then transformed into a normalized resolution two-dimensional B-scan feature image as network input, realizing the dimensional transformation from one-dimensional time series to two-dimensional spatial image.

[0035] Secondly, the two-dimensional image is input into the magnetic tile defect recognition model, where shallow feature extraction and residual mapping are performed. The input image first extracts primary acoustic edge features through a front-end convolutional layer, followed by pooling operations to extract local strong response features through spatial dimensionality reduction. This initially suppresses background noise while reducing the feature map size, preparing for subsequent deep feature extraction. Then, the model enters a core module group composed of cascaded residual units. Each residual unit establishes an identity mapping channel to propagate the input features across layers and add them to the deep residual information extracted by the convolutional path, thereby capturing subtle signal disturbances caused by minute pores or cracks inside the magnetic tile. Each residual module, by establishing an identity mapping channel, ensures the stability of the deep texture features of the ultrasonic signal during transmission, effectively preventing the gradient vanishing problem in deep networks.

[0036] Its residual mapping logic formula is:

[0037] in, The output characteristics of the residual module, It is the input two-dimensional feature vector. Residual mapping transformation performed for network layers, This represents the set of all learnable weight parameters within the residual module.

[0038] Furthermore, global feature integration and spatial dimensionality reduction are performed. To integrate the global correlations generated by the 3×5 matrix scanning trajectory, a global average pooling layer (GAP) is set at the end of the feature extraction layer. By performing spatial dimensional mean processing on the final-level feature map, the abnormal echo amplitude and frequency fluctuation characteristics of the overall magnetic tile detection area are integrated to extract a global abstract feature vector reflecting the complete state of the magnetic tile. The calculation formula is as follows:

[0039] in, The height of the two-dimensional image. The width of the two-dimensional image. Coordinates in a two-dimensional image The corresponding response value is obtained. While preserving spatial topological information, the number of model parameters is greatly reduced, effectively improving the model's generalization ability under small sample conditions.

[0040] Next, classification decision-making and loss optimization based on difficulty-weighted samples are implemented. The feature vector output from the GAP layer is mapped through a fully connected layer and then input into the Softmax function to calculate the normalized classification probability distribution:

[0041] In the formula, This indicates that the input features are At that time, the tested magnetic tile was determined to be the first The probability of a class; For the predicted category label, , Indicates the pair of categories output by the fully connected layer. The original score, Indicates the pair of categories output by the fully connected layer. The original score; To address the severe class imbalance caused by the far greater number of qualified products than defective products in a magnetic tile production line, the FocalLoss loss function is introduced. This function dynamically adjusts the weight ratios of samples with different levels of difficulty, automatically weakening the loss contribution of easily classified normal samples, forcing the model to focus on minute defect samples with weak echo characteristics and difficult-to-identify defects during iteration. Its expression is:

[0042] In the formula, This represents the Focal Loss value. This is a balancing factor used to balance the proportion of positive and negative samples. To predict probabilities, , This is a regulatory factor used to adjust the weights of easy and difficult samples; The magnetic tile defect recognition model is trained based on Focal Loss until the Focal Loss value no longer decreases. In this embodiment, according to... Figure 11 and Figure 12 The trained model performed well on the training set with a training loss of 0.0191, a validation loss of 0.0238, and an accuracy of 0.9867. In testing with 850 samples, the trained magnetic tile defect recognition model achieved a precision of 98.32%, a recall of 0.98, an F1 score of 0.98, and a loss rate of only 0.0251, as shown in Table 1. Experiments demonstrate that even in scenarios with a small number of labeled defect samples, by adjusting the Focal Loss function, the model can achieve a recognition rate of at least 95% for qualified magnetic tiles, effectively preventing excessive false detections in production.

[0043] Table 1

[0044] S4. After the test is completed, the electrically controlled telescopic column 45 extending to the bottom of the test platform 3 extends until the test platform 3 forms a 45-degree angle with the horizontal line, causing the magnetic tile on the test platform 3 to slide into the conveying device 5 and be conveyed out of the water tank 11. After the magnetic tile to be tested has completely slid out of the conveying device 5, the electrically controlled telescopic column 45 extending to the bottom of the test platform 3 retracts to its original length for the next cycle of testing. When the test result is a defective magnetic tile, the sorting plate 7 rotates a preset angle of 180°, moving the defective magnetic tile conveyed to the third conveying section 12 to the second conveyor belt 6, which then sends it to the waste recycling area. When the test result is qualified, the sorting plate remains in its original state, and the qualified product is sent to the discharge port by the third conveying section 12 of the conveying device 5, realizing the online automatic separation of qualified and defective products.

[0045] This invention aims to address the problems of low automation in internal defect detection during the production of magnetic tiles, high reliance on manual inspection, and significant fluctuations in detection accuracy and speed due to human factors. It also avoids secondary surface damage to magnetic tiles that may be caused by traditional contact inspection, thereby promoting quality and efficiency improvements for magnetic tile manufacturers. In terms of the detection device, a fully automated water-immersion multi-channel ultrasonic non-destructive testing method is used to improve the detection process of internal defects in magnetic tiles. By integrating multi-transducer probes, a moving component 9, and a self-unloading tilting inspection stage 3, low-cost, non-contact, and high-efficiency online scanning and sorting are achieved. Regarding the detection method, the deep learning classification model is optimized and adjusted. By mapping the acquired multi-channel one-dimensional A-scan raw data into two-dimensional feature images representing the internal structure, a deep residual network is used for feature extraction and learning. Simultaneously, a Focal Loss loss function is introduced to address the small sample imbalance problem, thereby significantly enhancing the model's ability to capture minute internal defects and its robustness, thus improving detection accuracy and the stability of industrial deployment.

Claims

1. A method for ultrasonic testing of internal defects in magnetic tiles, characterized in that, include: An ultrasonic pulse is emitted toward the magnetic tile under test, and the corresponding ultrasonic echo signal of the magnetic tile under test is obtained. The ultrasonic echo signal corresponding to the magnetic tile under test is mapped into a two-dimensional image; The obtained two-dimensional image is input into a pre-trained magnetic tile defect recognition model to obtain the defect recognition result.

2. The method according to claim 1, characterized in that, The ultrasonic echo signal corresponding to the magnetic tile under test is mapped into a two-dimensional image, including: Extract the effective echo segment from the pause time of each point in the ultrasonic echo signal corresponding to the magnetic tile under test, and obtain a complete time-domain sequence containing the surface wave, internal defect wave and bottom echo of the magnetic tile. The obtained complete time-domain sequence is subjected to discrete signal parameterization extraction based on multidimensional physical features to obtain a time-domain sequence containing sound velocity attenuation features, energy product features, and frequency domain dominant frequency features; A path compensation algorithm based on arc curvature is used to correct the geometric attenuation differences of time-domain sequences that include sound velocity attenuation characteristics, energy product characteristics, and frequency domain dominant frequency characteristics. The kriging interpolation algorithm is used to reconstruct the corrected time-domain sequence into a two-dimensional feature image; Random noise in the two-dimensional feature image is removed by morphological filtering, and grayscale linear mapping is performed to enhance the edge texture features of the defect, thus obtaining a two-dimensional image.

3. The method according to claim 2, characterized in that, Perform discrete signal parameterization extraction based on multidimensional physical features on the obtained complete time-domain sequence, including: For the pre-defined a×b matrix scanning points on the surface of the magnetic tile to be tested, a set of ultrasonic echo signals is collected at each point. Three types of physical parameters sensitive to defect characteristics were collected from the ultrasonic echo signal corresponding to each point, including: The sound velocity attenuation characteristic is expressed as follows: In the formula, Indicates the characteristics of sound velocity attenuation. This indicates the physical distance between the probe emitting the ultrasonic pulse and the ground surface of the magnetic tile being tested. This indicates the flight time of an ultrasonic wave from its emission to the capture of its peak value. The energy product characteristic is expressed as follows: In the formula, Indicates the energy product characteristic. This represents the preset normalized balance coefficient. Indicates the total number of sampling points. Indicates the sampling point index. The amplitude of the discrete signal within the time-domain sampling window; The frequency domain dominant frequency characteristic is expressed as follows: In the formula, The amplitude represents the frequency domain distribution. It is a natural constant. The imaginary unit, Pi This is the index of the sampling point in the frequency domain; This represents the frequency feature point corresponding to the maximum amplitude value in the frequency domain distribution.

4. The method according to claim 3, characterized in that, The geometric attenuation difference of a time-domain sequence, including sound velocity attenuation characteristics, energy product characteristics, and frequency domain dominant frequency characteristics, is corrected using a path compensation calculation based on the curvature of an arc surface. The acoustic path increment factor is calculated based on the coordinates of each scanning point on the arc surface and its corresponding radius of curvature. The time-domain sequence containing sound velocity attenuation characteristics, energy product characteristics, and frequency domain dominant frequency characteristics is corrected based on the sound path increment factor, and its expression is as follows: In the formula, These are the corrected signal characteristic values. These are the characteristic values ​​of the signal before correction. The sound attenuation coefficient of the material. To follow the curvature of the magnetic tile The varying sound path increment factor.

5. The method according to claim 4, characterized in that, The magnetic tile defect identification model consists of a normalization layer, a front-end convolutional layer, a batch normalization layer, a ReLU activation function layer, a max pooling layer, a first residual module group, a second residual module group, a global average pooling layer, a regularization layer, a fully connected layer, and a softmax activation function layer, connected in sequence. The first residual module group includes three residual modules connected in sequence; the second residual module group includes four residual modules connected in sequence. The residual module includes a main path and a shortcut path; the main path includes a 3×3 convolutional layer, a batch normalization layer, a ReLU activation function layer, a 3×3 convolutional layer, and a batch normalization layer in sequence; the shortcut path includes a 1×1 convolutional layer; the output of the residual module is obtained by adding the output of the main path and the output of the shortcut path element by element and applying the ReLU activation function.

6. The method according to claim 5, characterized in that, The training process for the magnetic tile defect identification model specifically includes: Map the ultrasonic echo signal corresponding to the magnetic tile with known defect state into a two-dimensional image; The two-dimensional image is input into the magnetic tile defect recognition model to obtain the defect prediction probability of the tested magnetic tile, which is expressed as: In the formula, This indicates that the input features are At that time, the tested magnetic tile was determined to be the first The probability of a class; For the predicted category label, , Indicates the pair of categories output by the fully connected layer. The original score, Indicates the pair of categories output by the fully connected layer. The original score; The Focal Loss is calculated based on the predicted probability, and its expression is as follows: In the formula, This represents the Focal Loss value. This is a balancing factor used to balance the proportion of positive and negative samples. To predict probabilities, , This is a regulatory factor used to adjust the weights of easy and difficult samples; The magnetic tile defect recognition model is trained based on Focal Loss until the Focal Loss value no longer decreases.

7. An apparatus for implementing the ultrasonic testing method for internal defects of magnetic tiles according to any one of claims 1 to 6, characterized in that, Includes a frame; a water tank located below the frame; a movable component suspended from the frame beam; an integrated probe device installed at the bottom of the movable component; a detection platform located below the integrated probe device and inside the water tank; the integrated probe device and the movable component constitute a defect detection mechanism. The integrated probe device includes an ultrasonic flaw detector, three parallel ultrasonic transducers, a data acquisition card, a signal preprocessing module, and a defect recognition module with a magnetic tile defect recognition model deployed thereon. The ultrasonic flaw detector is used to excite the three parallel ultrasonic transducers; the three parallel ultrasonic transducers are used to emit ultrasonic pulses towards the magnetic tile under test; the data acquisition card is used to acquire the ultrasonic echo signal corresponding to the magnetic tile under test; the signal preprocessing module is used to map the ultrasonic echo signal corresponding to the magnetic tile under test into a two-dimensional image; and the defect recognition module is used to input the obtained two-dimensional image into a pre-trained magnetic tile defect recognition model to obtain the defect recognition result.

8. The apparatus according to claim 7, characterized in that, The inner wall of the water tank is coated with an anti-rust layer, and its bottom has a sealed channel for data cables. A first conveyor belt is placed on the outer side of the water tank's inlet side. A trapezoidal inclined surface is placed on the inner side of the water tank's inlet side, and both the trapezoidal inclined surface and the first conveyor belt are in close contact with the water tank. The far end of the trapezoidal inclined surface is on the same horizontal line as the first conveyor belt, and the near end of the trapezoidal inclined surface is on the same horizontal line as the testing platform. A support mechanism is provided below the testing platform. The support mechanism includes a first support column, a second support column, a third support column, a fourth support column, and an electrically controlled telescopic column. The first and second support columns are located at opposite ends of the testing platform on the side near the trapezoidal inclined surface, respectively. The third support column is located at the midpoint of the testing platform on the side far from the trapezoidal inclined surface. The fourth support column is parallel to the first... The three support columns are located on the horizontal center line of the testing platform; the electrically controlled telescopic column has a V-shaped structure, and the intersection of its two branches is connected to the fourth support column in a K-shape. The two branches extend outward obliquely from the connection point with the fourth support column to the bottom of the water tank and the bottom surface of the testing platform, respectively; a conveying device with a chain plate conveyor belt with positioning slots is set on the far trapezoidal inclined side of the testing platform; the conveying device includes a first conveying section, a second conveying section, and a third conveying section. The first conveying section is a horizontal receiving section and is lower than the initial horizontal height of the testing platform. The second conveying section is an inclined conveying section, and the third conveying section is a horizontal unloading section; a second conveyor belt is set below the third conveying section; a sorting plate is set between the machine frame on the discharge side of the water tank and the edge of the water tank; When the test is completed, the electrically controlled telescopic column branch extending to the bottom of the test platform extends until the test platform forms a 45-degree tilt angle with the horizontal line, so that the magnetic tile on the test platform slides into the conveying device. When the test result is a defective magnetic tile, the sorting plate rotates at a preset angle to move the defective magnetic tile that has been conveyed to the third conveyor section to the second conveyor belt; when the test result is qualified, the sorting plate remains in its original state.

9. The apparatus according to claim 8, characterized in that, The moving component includes a rotatable assembly connected to the integrated probe device, a telescopic column, and a drive unit mounted on the frame beam. The moving component dynamically adjusts the probe position according to the radius of curvature of the magnetic tile under test, ensuring that the sound beams emitted by the three ultrasonic transducers are always perpendicularly incident on the surface of the magnetic tile. The moving component is configured to drive the integrated probe device to move relative to the water tank in the front-back and up-down directions to achieve full coverage scanning of the surface of the magnetic tile. The water tank is filled with a coupling water medium, and the integrated probe device and the magnetic tile are acoustically coupled through the water layer.

10. The apparatus according to claim 9, characterized in that, The ultrasonic flaw detector, three parallel ultrasonic transducers, data acquisition card, signal preprocessing module, and defect identification module with magnetic tile defect identification model are electrically connected in series via coaxial cable; the transmission device and the testing table are controlled by logic interlock to ensure that the magnetic tile is completely slid out before the next cycle of testing and positioning is carried out.

Images