Weld defect detection method and system based on deep learning

Abstract

The invention discloses a weld defect detection method and system based on deep learning, and relates to the field of metal welding, and the method comprises the following steps: receiving an ultrasonic signal of a weld and image information of the weld; the ultrasonic signal comprises a one-dimensional UT waveform signal; extracting time characteristics of the weld joint based on the one-dimensional UT waveform signal, and predicting a first probability of a defect category of the weld joint based on the time characteristics; extracting spatial features of the weld joint based on the image information; predicting a second probability of a defect category of the weld joint based on the spatial features; and combining the first probability and the second probability based on soft voting to calculate a third probability of the defect category, and determining the defect category of the weld joint based on the third probability.

Description

Technical Field

[0001] The present application relates to the field of metal welding, and specifically to a weld defect detection method and system based on deep learning. Background Art

[0002] Ultrasonic testing is suitable for non-destructive testing of weld defects. However, ultrasonic testing is a technology that is highly dependent on operating experience and skills. Different inspectors have different abilities and experience levels in operating equipment, interpreting signals, and judging defects, which leads to different conclusions even under the same testing conditions. Furthermore, the interpretation and judgment of ultrasonic signals often involve subjective understanding of signal characteristics. The shape, intensity, propagation speed and other characteristics of the signal are affected by internal defects, and the identification and interpretation of these characteristics require inspectors to have certain professional knowledge and experience. Different inspectors have different understandings and judgments of these characteristics, which leads to different inspection results.

[0003] In summary, due to differences in operating experience, skill level, subjective understanding, and visual perception, different inspectors will have different subjective inspection results when performing ultrasonic testing. Summary of the invention

[0004] In view of this, the purpose of this application is to provide a weld defect detection method and system based on deep learning, so as to solve the technical problem of how to avoid different subjective inspection results caused by different inspectors when conducting ultrasonic testing due to differences in operating experience, skill level, subjective understanding, and visual perception ability.

[0005] This application discloses a weld defect detection method based on deep learning, comprising the following steps: S1, receiving an ultrasonic signal of a weld and image information of the weld; the ultrasonic signal includes a one-dimensional UT waveform signal; S2, extracting the time feature of the weld based on the one-dimensional UT waveform signal, and predicting a first probability of the defect category of the weld based on the time feature; extracting the spatial feature of the weld based on the image information; and predicting a second probability of the defect category of the weld based on the spatial feature; S3. Calculate a third probability of the defect category by combining the first probability and the second probability based on soft voting, and determine the defect category of the weld based on the third probability.

[0006] In some possible implementations, the predicting a first probability of a defect category of the weld based on the time feature includes: S201, identifying and capturing a peak position and shape from the one-dimensional UT waveform signal based on the time feature; S202, predicting a first probability of a defect category of the weld based on the peak position and shape.

[0007] In some possible embodiments, predicting the first probability of the defect category of the weld based on the peak position and shape includes: sequentially repeating multiple one-dimensional convolutional layers, activation layers, pooling layers and discard layers, and obtaining the first probability of predicting the weld defect type based on a softmax function.

[0008] In some possible embodiments, the image information includes a stereoscopic scan image; and extracting the spatial features of the weld based on the image information includes: processing the image information based on a two-dimensional convolutional network model with an encoder-decoder symmetric structure to extract the spatial features of the weld; the two-dimensional convolutional network model with an encoder-decoder symmetric structure includes 23 convolutional layers, 4 downsampling layers, 4 upsampling layers, and 4 fusion layers.

[0009] In some possible implementations, predicting a second probability of a defect category of the weld based on the spatial feature includes: S211, receiving a stereoscopic scan image of the weld based on the encoder part of the two-dimensional convolutional network model; S212, sampling the scanned images of the new verification weld based on 800 pixels × 800 pixels sliding window cropping, and numbering them in sequence; S213, moving the sliding window with a step length of 600 pixels, so that different samples have an overlap of 200 pixels in length; S214 , cropping the obtained semantic segmentation image from a 100-pixel-wide region within the overlapping region, and then automatically stitching the cropped images according to the image numbers.

[0010] As a second aspect of the present application, a deep learning-based weld defect detection system is also provided, including: a defect data receiving module, a defect probability prediction module and a defect category determination module; the defect data receiving module is used to receive an ultrasonic signal of a weld and image information of the weld; the ultrasonic signal includes a one-dimensional UT waveform signal; the defect probability prediction module includes a time feature extraction unit, a first probability prediction unit, a spatial feature extraction unit and a second probability prediction unit, wherein the time feature extraction unit is used to extract the time feature of the weld based on the one-dimensional UT waveform signal, and the first probability prediction unit is used to predict the first probability of the defect category of the weld based on the time feature; the spatial feature extraction unit is used to extract the spatial feature of the weld based on the image information; the second probability prediction unit is used to predict the second probability of the defect category of the weld based on the spatial feature; the defect category determination module is used to calculate a third probability of the defect category based on a soft voting combination of the first probability and the second probability, and determine the defect category of the weld based on the third probability.

[0011] In some possible implementations, the first probability prediction unit includes a first subunit, a second subunit, and a third subunit; the first subunit is used to identify and capture the peak position and shape from the one-dimensional UT waveform signal based on the time feature; the second subunit is used to predict the first probability of the defect category of the weld based on the peak position and shape.

[0012] In some possible embodiments, predicting the first probability of the defect category of the weld based on the peak position and shape includes: sequentially repeating multiple one-dimensional convolutional layers, activation layers, pooling layers and discard layers, and obtaining the first probability of predicting the weld defect type based on a softmax function.

[0013] In some possible embodiments, the image information includes a stereoscopic scan image; and extracting the spatial features of the weld based on the image information includes: processing the image information based on a two-dimensional convolutional network model with an encoder-decoder symmetric structure to extract the spatial features of the weld; the two-dimensional convolutional network model with an encoder-decoder symmetric structure includes 23 convolutional layers, 4 downsampling layers, 4 upsampling layers, and 4 fusion layers.

[0014] In some possible implementations, predicting a second probability of a defect category of the weld based on the spatial feature includes: S211, receiving a stereoscopic scan image of the weld based on the encoder part of the two-dimensional convolutional network model; S212, sampling the scanned images of the new verification weld based on 800 pixels × 800 pixels sliding window cropping, and numbering them in sequence; S213, moving the sliding window with a step length of 600 pixels, so that different samples have an overlap of 200 pixels in length; S214 , cropping the obtained semantic segmentation image from a 100-pixel-wide region within the overlapping region, and then automatically stitching the cropped images according to the image numbers.

[0015] Beneficial effects: This method collects ultrasonic information through sensors and develops a multi-branch deep fusion network model. By integrating one-dimensional and two-dimensional convolutional neural networks, it simultaneously captures the temporal and spatial characteristics of the weld. It can determine whether the weld has defects and the type of defects without compromising the integrity of the finished product.

[0016] Other advantages, objectives and features of the present application will be described in the following description to some extent, and will be apparent to those skilled in the art based on the following examination and study, or can be taught from the practice of the present application to some extent. The objectives and other advantages of the present application can be achieved and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain and illustrate the present application, but should not be construed as limiting the scope of protection of the present application.

[0018] Figure 1 is the system flow chart of this application; Figure 2 It is the one-dimensional UT waveform signal diagram of qualified samples; Figure 3 It is the one-dimensional UT waveform signal diagram of the pore sample; Figure 4 It is the one-dimensional UT waveform signal diagram of the cracked sample; Figure 5 It is the one-dimensional UT waveform signal diagram of the unfused sample; Figure 6 is the feature extraction network graph; Figure 7 is the feature enhancement network graph; Figure 8 is the system structure diagram of this application; Among them: 1. Defect data receiving module; 2. Defect probability prediction module; 3. Defect category determination module. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0021] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0022] In the above description of the present application, the terms “first”, “second”, etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0023] like Figure 1 As shown, the present application discloses a weld defect detection method based on deep learning, comprising the following steps: S1. Receive an ultrasonic signal of a weld and image information of the weld; the ultrasonic signal includes a one-dimensional UT waveform signal, and the image information of the weld includes a stereoscopic scan of the weld; collect the one-dimensional UT waveform signal of the weld and the image information of the same weld through a sensor, and develop a multi-branch deep fusion network model, which integrates a one-dimensional convolutional neural network and a two-dimensional convolutional neural network to respectively receive the ultrasonic signal of the weld and the image information of the weld for processing.

[0024] S2. Extracting the time characteristics of the weld based on the one-dimensional UT waveform signal, predicting the first probability of the defect category of the weld based on the time characteristics, the first probability is obtained through a one-dimensional CNN model, the input parameter of the CNN model is the ultrasonic waveform data, that is, the ultrasonic waveform data of the defect is learned through the one-dimensional CNN model to obtain a defect classification model based on the ultrasonic waveform, and the time-varying characteristics such as oscillation and sparks are extracted through the one-dimensional CNN model. The vector is normalized by Softmax before output to obtain the probability of each type of crack, incomplete penetration, incomplete fusion and pore; extracting the spatial characteristics of the weld based on the image information; predicting the defects of the weld based on the spatial characteristics The second probability of the category is obtained through a two-dimensional CNN model. The input parameter is image data. The spatial features, such as position and shape, are extracted through the model. That is, the image data of the defect is learned through the two-dimensional CNN model to obtain a defect classification model based on the image data. The spatial features are extracted through the two-dimensional CNN model, and the vector is normalized with Softmax before output to obtain the probability of each type of crack, incomplete penetration, incomplete fusion and pore; that is, the first probability is the probability of predicting each defect category of crack, incomplete penetration, incomplete fusion and pore through time features, and the second probability is the probability of predicting each defect category of crack, incomplete penetration, incomplete fusion and pore through spatial features.

[0025] In some embodiments, predicting a first probability of a defect category of the weld based on the time feature includes: S201. Identify and capture the peak position and shape from the one-dimensional UT waveform signal based on the time feature. External inclusions such as pores or slag along the ultrasonic path, as well as improper welding conditions that lead to incomplete structures, will change the phase and amplitude of the reflected ultrasonic signal. These changes appear as peaks in the one-dimensional UT waveform. Therefore, different weld defects have one-dimensional UT waveforms of different shapes. Therefore, it is possible to determine which defect type the current UT waveform belongs to by learning the waveform data. The one-dimensional convolutional network is used to extract time features, and can identify and capture the peak position and shape from the one-dimensional UT waveform data; through the peak position and shape and Analyze the specific defect location and defect type, among which, , is the defect depth, is the propagation distance of the ultrasonic wave, is the detection angle, is the propagation speed of ultrasound, It is the propagation time of ultrasound. Ultrasonic information is collected through the probe and pulse receiver, and the ultrasonic signal is converted into digital data through the analog-to-digital converter. After the conversion, the horizontal axis is the current signal transmission time, and the vertical axis is the signal amplitude; Figure 2-5(a) 1D waveform of a qualified sample, the peak signal appears at about 65 On the time axis, the formula in the article can be used to calculate that it is about 7.6 mm deep. The peak value is at a certain depth + the amplitude is within a certain range. If it is not within this range, it is an abnormal situation. (b) is the waveform of the pore sample, showing the peak value caused by the defect at 50 (about 5.8mm deep); (c) is a crack, from 50 (d) is the lack of fusion, showing a large peak at 55 On the left and right sides (approximately 6.4 mm deep).

[0026] S202, predicting the first probability of the defect category of the weld based on the peak position and shape. In two-dimensional image data, because the original image is too large, the deep learning model cannot directly process the image and needs to be cropped first. In this embodiment, a sliding window image cropping method is proposed to obtain training and test samples of the model. First, the weld and different welding defects are marked on the entire scanned image to increase the speed of label image acquisition. Then, a sliding window of 800 pixels × 800 pixels is used to sample along the center of the original image, and the window is moved to the right 100 pixels each time to obtain several sample images. Finally, the same sampling method is used on the labeled image to obtain the label of the corresponding sample. In order to better train the segmentation model, according to the characteristics of the sampled image, the image is horizontally flipped, vertically flipped, and horizontally flipped and then vertically flipped for data enhancement, and the same processing method is applied to the obtained labeled sample image to obtain more and richer data. The method of predicting the first probability of the defect category of the weld based on the peak position and shape includes: along the time characteristics of the effective peak on the time axis of the one-dimensional UT waveform signal, normalizing the one-dimensional ultrasonic waveform to the range of 0 to 1, and introducing random noise to further normalize the signal to improve the generalization ability of the model, and extracting the global features, local features and sub-local features of the one-dimensional UT waveform through multiple one-dimensional convolutional blocks including one-dimensional convolutional layers, activation layers, pooling layers and discard layers to obtain feature vectors of three branches, which are then spliced ​​together, and after normalization + activation layer + discard layer, softmax is used to obtain the first probability of the type predicted for the current weld, wherein the global feature is the time feature of the entire ultrasonic signal, the local feature is the time feature of the ultrasonic signal of the defect segment, and the sub-local feature is the time feature of the signal before and after the defect segment. In this embodiment, the filter size of the one-dimensional convolutional block of the deep learning model is , using ReLU as the activation function, using maximum pooling, the parameter of the dropout layer is set to 0.5, and the number of filters is , , and There is a residual connection between the input and output of the one-dimensional convolutional block to solve the gradient vanishing problem.

[0027] In some embodiments, the extracting the spatial features of the weld based on the image information includes: processing the image information based on a two-dimensional convolutional network model with an encoder-decoder symmetric structure to extract the spatial features of the weld; the two-dimensional convolutional network model with an encoder-decoder symmetric structure includes 23 convolutional layers, 4 downsampling layers, 4 upsampling layers and 4 fusion layers, and the number of optimized convolution kernels is 64, 128, 256, 512, 512, 256, 128 and 64 respectively. Figure 6-7 In the figure, the left half of the two-dimensional convolutional network model is the encoder part, which belongs to the feature extraction network. The encoder part downsamples the input image four times, and the step size and zero padding value of the convolution layer are 1, so the feature size remains unchanged; the right half is the decoder part, which belongs to the feature enhancement network. Through four upsampling operations, the feature map is gradually restored to the size of the input image. The feature extraction network receives the image input and uses The convolutional network + ReLU activation function of the convolution kernel is expanded to 64 dimensions, and then used Downsampling is performed by using a filter with a step size of 1 and a maximum pooling with a step size of 1 to obtain a vector with the same number of channels but reduced height and width. Then, the vector is expanded to a dimension of 128 through the same processing steps. This process is repeated until four downsamplings are completed. The feature enhancement network uses the output of the previous network as the original input and uses The transposed convolution is used to upsample to obtain the first result, and the first results of the same layer in the extraction network are spliced ​​together, using two layers The convolution network of the convolution kernel + ReLU activation function processing obtains the second result, and so on, until 4 upsamplings are completed; finally, the network will obtain a 64-dimensional vector, which is normalized by the Softmax function to obtain the classification result.

[0028] When upsampling, the feature layers are stacked with the copied ones from the same layer position to better reconstruct the details of the feature map and ensure that the corresponding spatial information dimension remains unchanged. In the last layer of the 2D convolutional network model, The convolution operation maps each 64-component feature vector to different defect categories. The two-dimensional convolutional network model uses the cross entropy loss function to measure the error between the predicted second probability and the true sample label. The formula is as follows: ; Where gt represents the true label, p represents the predicted second probability, is the cross entropy loss function.

[0029] In some embodiments, predicting a second probability of a defect category of the weld based on the spatial feature includes: S211, receiving a stereoscopic scan image of the weld based on the encoder part of the two-dimensional convolutional network model; S212, sampling the scanned images of the new verification weld based on 800 pixels × 800 pixels sliding window cropping, and numbering them in sequence; S213, moving the sliding window with a step length of 600 pixels, so that different samples have an overlap of 200 pixels in length; S214, cropping the obtained semantic segmentation image from the 100-pixel-wide area in the overlapping area, and then automatically stitching the cropped image according to the image number. The stitched image obtained based on steps S211-S214 is consistent with the size of the original scanned image, which can effectively solve the segmentation problem of the stitched edge of the cropped image.

[0030] After training, the two-dimensional convolutional network model can process input images of various sizes. The size of the obtained semantic segmentation image is also 800 pixels × 800 pixels. In order to obtain the classification result of the complete image, the image fusion method based on the segmentation image features is used to process and analyze the verification weld after the model training is completed. The scanned image of the new verification weld is sampled by using a sliding window cropping of 800 pixels × 800 pixels and numbered in sequence; the sliding window is moved with a step size of 600 pixels, so that there is an overlap of 200 pixels between different samples; the obtained semantic segmentation image is cropped from the 100-pixel wide area in the overlapping area, and then the cropped image is automatically stitched according to the image number. The obtained stitched image is made consistent with the size of the original scanned image, which can effectively solve the segmentation problem of the stitched edge of the cropped image.

[0031] S3. Calculate the third probability of the defect category based on the combination of the first probability and the second probability by soft voting, and determine the defect category of the weld based on the third probability. In order to obtain the final decision, soft voting is used to combine the probabilities of the predicted defect categories output by the two models. The soft voting method is a combination strategy for classification problems in ensemble learning. The variance is reduced by integrating multiple models, thereby improving the robustness of the model. In this embodiment, the third probability is the weighted sum of the first probability and the second probability, that is, the third probability = a*first probability + b*second probability, where a+b=1, and the third probability is the result of combining the first probability and the second probability by soft voting. For example, the prediction result of a one-dimensional convolutional neural network for a certain position is 89% for the normal type and 11% for the crack defect. The result of a two-dimensional convolutional neural network for the same position is 75% for the normal type and 9% for the crack defect. The probability of the normal type for the position is 82% on average, and the probability of the crack defect is 10%.

[0032] Ultrasonic information is collected through sensors and a multi-branch deep fusion network model is developed. By integrating one-dimensional and two-dimensional convolutional neural networks, the temporal and spatial characteristics of the weld are captured at the same time. It can determine whether the weld has defects and the type of defects without compromising the integrity of the finished product. It also avoids the one-dimensional UT that transmits and receives pulse signals through a single sound wave with a fixed angle and frequency, which is prone to signal distortion due to various aspects such as probe-workpiece contact, probe position, and direction.

[0033] like Figure 8 As shown, as a second aspect of the present application, a deep learning-based weld defect detection system is also provided, including: a defect data receiving module 1, a defect probability prediction module 2 and a defect category determination module 3; the defect data receiving module is used to receive an ultrasonic signal of a weld and image information of the weld; the ultrasonic signal includes a one-dimensional UT waveform signal; the defect probability prediction module includes a time feature extraction unit, a first probability prediction unit, a spatial feature extraction unit and a second probability prediction unit, wherein the time feature extraction unit is used to extract the time feature of the weld based on the one-dimensional UT waveform signal, and the first probability prediction unit is used to predict the first probability of the defect category of the weld based on the time feature; the spatial feature extraction unit is used to extract the spatial feature of the weld based on the image information; the second probability prediction unit is used to predict the second probability of the defect category of the weld based on the spatial feature; the defect category determination module is used to calculate a third probability of the defect category based on a soft voting combination of the first probability and the second probability, and determine the defect category of the weld based on the third probability.

[0034] In some embodiments, the first probability prediction unit includes a first subunit and a second subunit; the first subunit is used to identify and capture the peak position and shape from the one-dimensional UT waveform signal based on the time feature; the second subunit is used to predict the first probability of the defect category of the weld based on the peak position and shape.

[0035] In some embodiments, the first probability of predicting the defect category of the weld based on the peak position and shape includes: sequentially repeating multiple convolution modules including one-dimensional convolution layers, activation layers, pooling layers and discard layers, and finally using a softmax function to obtain the final output.

[0036] In some embodiments, the image information includes a stereoscopic scan image; the extracting the spatial features of the weld based on the image information includes: processing the image information based on a two-dimensional convolutional network model with an encoder-decoder symmetric structure to extract the spatial features of the weld; the two-dimensional convolutional network model with an encoder-decoder symmetric structure includes 23 convolutional layers, 4 downsampling layers, 4 upsampling layers and 4 fusion layers.

[0037] In some embodiments, predicting a second probability of a defect category of the weld based on the spatial feature includes: S211, receiving a stereoscopic scan image of the weld based on the encoder part of the two-dimensional convolutional network model; S212, sampling the scanned images of the new verification weld based on 800 pixels × 800 pixels sliding window cropping, and numbering them in sequence; S213, moving the sliding window with a step length of 600 pixels, so that different samples have an overlap of 200 pixels in length; S214 , cropping the obtained semantic segmentation image from a 100-pixel-wide region within the overlapping region, and then automatically stitching the cropped images according to the image numbers.

[0038] Those skilled in the art will appreciate that although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features from different embodiments is meant to be within the scope of the present application and to form different embodiments.

[0039] Those skilled in the art will appreciate that the description of each embodiment has its own emphasis, and for parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0040] Although the embodiments of the present application are described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations all fall within the scope defined by the attached claims. The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed by the present invention, and these modifications or substitutions should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A weld defect detection method based on deep learning, characterized in that: The steps include: S1, receiving an ultrasonic signal of a weld and image information of the weld; the ultrasonic signal includes a one-dimensional UT waveform signal; S2, extracting the time feature of the weld based on the one-dimensional UT waveform signal, and predicting the first probability of the defect category of the weld based on the time feature; extracting the spatial feature of the weld based on the image information; Predicting a second probability of a defect category of the weld based on the spatial feature; S3. Calculate a third probability of the defect category by combining the first probability and the second probability based on soft voting, and determine the defect category of the weld based on the third probability.

2. The weld defect detection method based on deep learning according to claim 1, characterized in that: The predicting a first probability of a defect category of the weld based on the time feature comprises: S201, identifying and capturing a peak position and shape from the one-dimensional UT waveform signal based on the time feature; S202, predicting a first probability of a defect category of the weld based on the peak position and shape.

3. The weld defect detection method based on deep learning according to claim 2 is characterized in that: The first probability of predicting the defect category of the weld based on the peak position and shape includes: sequentially repeating multiple one-dimensional convolutional layers, activation layers, pooling layers and discard layers, and obtaining the first probability of predicting the weld defect type based on a softmax function.

4. The weld defect detection method based on deep learning according to claim 3 is characterized in that: The image information includes a stereoscopic scan image; the extracting the spatial features of the weld based on the image information includes: processing the image information based on a two-dimensional convolutional network model with an encoder-decoder symmetric structure to extract the spatial features of the weld; the two-dimensional convolutional network model with an encoder-decoder symmetric structure includes 23 convolutional layers, 4 downsampling layers, 4 upsampling layers and 4 fusion layers.

5. The weld defect detection method based on deep learning according to claim 4 is characterized in that: Predicting a second probability of a defect category of the weld based on the spatial feature comprises: S211, receiving a stereoscopic scan image of the weld based on the encoder part of the two-dimensional convolutional network model; S212, sampling the scanned images of the new verification weld based on 800 pixels × 800 pixels sliding window cropping, and numbering them in sequence; S213, moving the sliding window with a step length of 600 pixels, so that different samples have an overlap of 200 pixels in length; S214 , cropping the obtained semantic segmentation image from a 100-pixel-wide region within the overlapping region, and then automatically stitching the cropped images according to the image numbers.

6. A weld defect detection system based on deep learning, characterized in that: include: Defect data receiving module, defect probability prediction module and defect category determination module; The defect data receiving module is used to receive the ultrasonic signal of the weld and the image information of the weld; the ultrasonic signal includes a one-dimensional UT waveform signal; the defect probability prediction module includes a time feature extraction unit, a first probability prediction unit, a space feature extraction unit and a second probability prediction unit, wherein the time feature extraction unit is used to extract the time feature of the weld based on the one-dimensional UT waveform signal, and the first probability prediction unit is used to predict the first probability of the defect category of the weld based on the time feature; the space feature extraction unit is used to extract the space feature of the weld based on the image information; the second probability prediction unit is used to predict the second probability of the defect category of the weld based on the space feature; the defect category determination module is used to calculate the third probability of the defect category by combining the first probability and the second probability based on soft voting, and determine the defect category of the weld based on the third probability.

7. The deep learning-based weld defect detection system according to claim 6, characterized in that: The first probability prediction unit includes a first subunit and a second subunit; the first subunit is used to identify and capture the peak position and shape from the one-dimensional UT waveform signal based on the time feature; the second subunit is used to predict the first probability of the defect category of the weld based on the peak position and shape.

8. The deep learning-based weld defect detection system according to claim 7, characterized in that: The first probability of predicting the defect category of the weld based on the peak position and shape includes: sequentially repeating multiple one-dimensional convolutional layers, activation layers, pooling layers and discard layers, and obtaining the first probability of predicting the weld defect type based on a softmax function.

9. The deep learning-based weld defect detection system according to claim 8, characterized in that: The image information includes a stereoscopic scan image; the extracting the spatial features of the weld based on the image information includes: processing the image information based on a two-dimensional convolutional network model with an encoder-decoder symmetric structure to extract the spatial features of the weld; the two-dimensional convolutional network model with an encoder-decoder symmetric structure includes 23 convolutional layers, 4 downsampling layers, 4 upsampling layers and 4 fusion layers.

10. The deep learning-based weld defect detection system according to claim 9, characterized in that: Predicting a second probability of a defect category of the weld based on the spatial feature comprises: S211, receiving a stereoscopic scan image of the weld based on the encoder part of the two-dimensional convolutional network model; S212, sampling the scanned images of the new verification weld based on 800 pixels × 800 pixels sliding window cropping, and numbering them in sequence; S213, moving the sliding window with a step length of 600 pixels, so that different samples have an overlap of 200 pixels in length; S214 , cropping the obtained semantic segmentation image from a 100-pixel-wide region within the overlapping region, and then automatically stitching the cropped images according to the image numbers.

Images