CN114359193A - A defect classification method and system based on ultrasonic phased array imaging - Google Patents

Abstract

The invention provides a defect classification method and system based on ultrasonic phased array imaging, belonging to the field of non-destructive testing. Focusing processing, color coding is performed according to the signal amplitude, and the image of the sample to be tested is obtained; the image of the sample to be tested is preprocessed and then input into the classification prediction model to obtain the defect classification of the sample to be tested; The method is as follows: using the finite element simulation method of ultrasonic scattering data of defects to obtain simulation data by simulating ultrasonic phased array imaging; after all-focusing processing of the simulation data, the simulation image is obtained; after preprocessing the simulation image, data enhancement is performed; The neural network extracts image features from the simulated images; the image features are input to the fully connected layer, and the classification prediction model is trained with the defect classification as the output. The present invention improves the accuracy of defect classification.

Description

A defect classification method and system based on ultrasonic phased array imaging

technical field

The invention belongs to the field of non-destructive testing, and more particularly, relates to a defect classification method and system based on ultrasonic phased array imaging.

Background technique

The development and application of ultrasonic non-destructive testing technology is based on the interaction between ultrasonic waves and the measured object; when ultrasonic waves with good guidance encounter defects during the propagation process, their propagation direction or characteristics will change. Reflection, refraction, and scattering are studied to detect and characterize workpiece defects. Compared with other non-destructive testing methods, ultrasonic non-destructive testing has unique advantages; it is widely used in non-destructive evaluation of metals, non-metals and composite materials; it has strong penetration ability, large detection depth, and good positioning ability for internal defects of workpieces; High sensitivity, low cost and no harm to human body.

In the existing ultrasonic defect detection technology, there is a lack of detection for stomatal defects, that is, the lack of parameterized data for multiple small holes with small sizes and close distances. In addition, conventional B-ultrasound imaging has low resolution and poor imaging accuracy, and does not have the ability to resolve multiple small defects (with a size less than 0.8 wavelength) that are close to each other. Among the common classification methods, dimensionality reduction algorithms such as Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) can simplify high-dimensional data to process large-scale data sets, but it is difficult to understand the meaning of the results; Naive Bayesian methods Classification is based on probability theory, so this method is limited by Bayes' theorem and conditional independence assumption, but the conditional independence assumption is often not established in practical applications, thus affecting the classification effect of Bayesian method.

SUMMARY OF THE INVENTION

In view of the defects of the prior art, the present invention provides a defect classification method and system based on ultrasonic phased array imaging, the purpose is to use all-focus imaging for the defects of the tested object, and use the convolutional neural network to perform feature extraction and then perform the defect classification. Classification to improve the accuracy of defect classification.

In order to achieve the above object, on the one hand, the present invention provides a defect classification based on ultrasonic phased array imaging, comprising the following steps:

Ultrasonic phased array imaging is used for the sample to be tested to obtain ultrasonic full-matrix data;

After the full-focus processing of the ultrasonic full-matrix data, color coding is performed according to the signal amplitude to obtain the image of the sample to be tested;

The image of the sample to be tested is preprocessed and input into the trained classification prediction model to obtain the defect classification of the sample to be tested;

Among them, the classification prediction model is obtained by adding a fully connected layer to the output layer of the convolutional neural network;

The way to train a classification prediction model is:

Using the finite element simulation method of ultrasonic scattering data of defects, set the simulation parameters, and obtain the simulation data by simulating ultrasonic phased array imaging;

After the simulation data is fully focused, the simulation image is obtained after color coding according to the signal amplitude;

Data enhancement is performed on the simulated image after preprocessing, and the enhanced simulated image is obtained;

Use the trained convolutional neural network to extract image features from the enhanced simulated image;

The image features and corresponding class labels are input to the fully connected layer to train the classification prediction model.

Further preferably, the simulation parameters include sample material, ultrasonic testing frequency, sample defect distribution, sample defect quantity, sample defect size and phased array probe parameters;

Phased array probe parameters include probe size, number of array elements, center frequency, bandwidth, and array element spacing.

Further preferably, the method for preprocessing the simulated image is:

Crop the edge image irrelevant to the defect classification according to the defect distribution position, and retain the defect area;

The cropped simulation images are divided into different folders according to the categories, and the data set production is completed;

Label the dataset with one-hot encoding.

Further preferably, the method for performing data enhancement on the simulated images in the dataset is:

Normalize the simulated images in the dataset;

Rotation and/or horizontal translation and/or vertical translation and/or random horizontal flip and/or filling in a nearest manner are performed on the normalized simulated image.

Further preferably, the method for preprocessing the image of the sample to be tested is:

The image of the sample to be tested is interactively cropped, and the defect area to be classified is retained;

The upper and lower edges and the left and right edges of the image in the defective area are filled in the height and width directions respectively by dividing into three RGB channels; the color of the filling is consistent with the image color of the defect-free area;

Combine the three channel images so that the size of the padded image is the same as the input image size of the convolutional neural network.

Further preferably, the convolutional neural network is a VGG (Visual Geometry GroupNetwork) 16 convolutional neural network after removing the top layer, including 13 convolutional layers and 5 pooling layers.

Further preferably, the activation function of the last fully connected layer of the classification prediction model adopts the softmax activation function, specifically:

为归一化系数；S_j表示经过Softmax计算后的输出结果。 _Among them, the _Softmax function _maps the _vector ( _{a 1} , a ₂ , . input value;

is the normalization coefficient; S _j represents the output result after Softmax calculation.

In another aspect, the present invention provides a defect classification system based on ultrasonic phased array imaging, comprising:

Ultrasonic phased array probe, which is used for ultrasonic phased array imaging of the sample to be tested to obtain ultrasonic full matrix data;

The all-focus module is used to perform all-focus processing on the ultrasonic full-matrix data and the simulation data, and then perform color coding according to the signal amplitude to obtain the image of the sample to be tested and the simulation image respectively;

The simulation module is used to use the finite element simulation method of ultrasonic scattering data of defects, set simulation parameters, and obtain simulation data by simulating ultrasonic phased array imaging;

an image preprocessing module, used for preprocessing the image of the sample to be tested and the simulated image;

The classification prediction module is used to input the preprocessed image of the sample to be tested into the trained classification prediction model to obtain the defect classification of the sample to be tested;

a data enhancement module, configured to perform data enhancement on the preprocessed simulation image to obtain the enhanced simulation image;

Wherein, the classification prediction model is obtained by adding a fully connected layer to the output layer of a convolutional neural network;

The training method of the classification prediction module is:

Use the trained convolutional neural network to extract image features from the enhanced simulated image;

The image features and corresponding class labels are input to the fully connected layer to train the classification prediction model.

Further preferably, the activation function of the last fully connected layer of the classification prediction model adopts the softmax activation function, and the softmax activation function is:

为归一化系数；S_j表示经过Softmax计算后的输出结果。 _Among them, the _Softmax function _maps the _vector ( _{a 1} , a ₂ , . input value;

is the normalization coefficient; S _j represents the output result after Softmax calculation.

Further preferably, the method for preprocessing the image of the sample to be tested is:

The image of the sample to be tested is interactively cropped, and the defect area to be classified is retained;

The upper and lower edges, left and right edges of the image in the defective area are filled in the height and width directions respectively by three channels of RGB; the color of the filling is consistent with the image color of the non-defective area;

Combine the three channel images so that the size of the filled image is the same as the size of the input image of the convolutional neural network;

The method of preprocessing the simulation image is as follows:

Cut out the edge image irrelevant to the defect classification for the simulated image according to the defect distribution position, and retain the defect area;

The cropped simulation images are divided into different folders according to the categories, and the data set production is completed;

Label the dataset with one-hot encoding.

Further preferably, the simulation parameters include sample material, ultrasonic testing frequency, sample defect distribution, sample defect quantity, sample defect size and phased array probe parameters;

The phased array probe parameters include probe size, array element number, center frequency, bandwidth and array element spacing.

Further preferably, the method for performing data enhancement on the simulated images in the data set is:

normalizing the simulated images in the data set;

Rotation and/or horizontal translation and/or vertical translation and/or random horizontal flip and/or filling in a nearest manner are performed on the normalized simulated image.

Further preferably, the convolutional neural network is a VGG16 convolutional neural network with the top layer removed, including 13 convolutional layers and 5 pooling layers.

In general, compared with the prior art, the above technical solutions conceived by the present invention have the following beneficial effects:

When training the classification prediction model, the present invention uses the finite element simulation method of ultrasonic scattering data for defects to simulate any defect, and obtains a simulated image after simulating ultrasonic phased array imaging and then using full focus processing. In order to construct the training set of the classification prediction model, the finite element simulation model takes into account the multiple scattering of ultrasonic waves between multiple defects, so the constructed simulation image can accurately reflect the defect problem, and the trained classification prediction model is effective for defect classification. High accuracy.

The invention adopts the advanced data post-processing method for imaging, that is, the full-focusing method is used to perform delay and superposition processing on the ultrasonic full-matrix data collected by the ultrasonic phased array, so as to obtain the internal imaging results of the measured object, and can make full use of all the collected data. signal, therefore, the imaging resolution is much higher than conventional ultrasound B-scan imaging.

In the classification prediction model of the present invention, a fully connected layer is added to the output layer of the convolutional neural network, and the trained convolutional neural network is used to extract image features. Since only the network parameters of the classification part in the classification prediction model need to be trained, the classification prediction model is greatly shortened. computation and training time. And the classification prediction model provided by the present invention has a self-updating function, that is, model training and updating can be performed when new data is added.

Description of drawings

1 is a flowchart of a defect classification method based on ultrasonic phased array imaging provided by an embodiment of the present invention;

2 is a schematic diagram of random distribution of defect positions provided by an embodiment of the present invention;

3 is a schematic diagram of an ultrasound full focus imaging provided by an embodiment of the present invention;

4 is a schematic diagram of a classification prediction model provided by an embodiment of the present invention;

FIG. 5 is a schematic diagram of accuracy changes in a training process provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

As shown in FIG. 1 , the present invention provides a defect classification method based on ultrasonic phased array imaging, and the overall implementation process is as follows: taking the size and density of circular holes as defect classification parameters, designing four types of circular hole defects and processing actual samples; Combine the detection frequency, bandwidth percentage, number of ultrasonic phased array elements and other parameters to obtain simulation data and experimental data; perform all-focus imaging processing on the simulation data and experimental data; Convolutional Neural Networks and Fully Connected Neural Networks to classify imaging results;

Example

As shown in FIG. 1 , on the one hand, this embodiment provides a defect classification method based on ultrasonic phased array imaging, including the following steps:

Step 1: Set simulation parameters during simulation, and obtain simulation data by simulating ultrasonic phased array imaging;

Simulation parameters include: selected material to be tested, ultrasonic testing frequency, hole size and density;

In this embodiment, the selected material is aluminum; the frequency of the ultrasonic phased array probe in the simulation is set to 10MHz; four groups of stomatal defects with different parameters are obtained through the shape modeling of pore-shaped defects, specifically: the first type of defect circle The hole radius is 0.5mm, and the number of defects is 3; the radius of the second type of defect is 0.45mm, and the number of defects is 5; the radius of the third type of defect is 0.4mm, and the number of defects is 7; the radius of the fourth type of defect is 7. is 0.35mm, and the number of defects is 11; the distribution of defects adopts random distribution, as shown in Figure 2; the simulation data is stored in the simul_X.mat file;

Step 2: Design the parameters of the experimental sample during the experiment, and process the experimental sample;

Sample parameters include: hole size and density, detection frequency and bandwidth percentage; use wire cutting to process test samples;

In this example, 6061 aluminum material is used to make the sample, and the cylindrical hole is processed by WEDM to simulate defects. The design method of the size and random distribution position of the round hole is consistent with that in step 1, and the position of the round hole is distributed in the distance detection method. The contact surface is within the range of 300±5mm;

Step 3: Use the ultrasonic phased array probe to obtain the ultrasonic full matrix data for the experimental sample, that is, the experimental data;

In this embodiment, an ultrasonic phased array probe with 64 array elements linearly arranged is used, the array element spacing is 0.3mm, the frequency is 10MHz, and the array element length is 5mm; 64 independent transmit-receive channels corresponding to the number of probe array elements can be controlled in parallel; rectangular pulses are used to excite each array element in turn, and all array elements in each transmit are used as receivers, thus obtaining each transmit- Receive the full matrix data of the time domain signal of the array element group;

Step 4: Perform full focus processing on the collected simulation data and experimental data, and display and store corresponding images (simulation images and experimental images) after color coding according to the amplitude, as shown in Figure 3;

In this embodiment, the array elements are excited in sequence, the excitation delay time of each array element is calculated according to the principle of geometric acoustics, and the echo signals received by each channel of the phased array are delayed and superimposed, so as to realize the beam at the specified time. focus of the position point;

In the process of data processing, there are two ways to accelerate the calculation, one is to use the graphics accelerator to perform parallel processing; the other is to use the triangular matrix to replace the original full matrix data according to the reciprocity theorem of sound field to each pixel. Superposition is simplified; the final signal amplitude at the focal point is given by:

Among them, S _ref is the amplitude of the signal received on an acoustic wave propagation path passing through the focal point; t _p and t _q are the time required for the acoustic wave to reach the focal point from the transmitting array element and from the focal point to the receiving array element, respectively; δ _pq is the weighting coefficient, which is used to control the number of times that each data in the whole matrix participates in the calculation, so that it satisfies the reciprocity theorem; when p=q, δ _pq =1; otherwise, δ _pq =0; the obtained image is saved as classX_XXX.png Format;

Step 5: Preprocess the simulated image, including automatic cropping and retaining defect areas, classification according to defect design, and one-hot encoding;

In this embodiment, edge images irrelevant to defect classification are automatically cropped and removed according to defect distribution positions, defect areas are retained, and they are divided into folders according to categories to complete data set production; one-hot encoding is used for label setting, which is set to (1, 0,0,0), (0,1,0,0), (0,0,1,0) and (0,0,0,1);

Step 6: Perform data enhancement on the simulated images in the dataset;

In this embodiment, first normalize the simulated image, then rotate 0-40°, horizontally translate 0-0.2 image width, vertically translate 0-0.2 image height, randomly flip the general image horizontally and use The nearest method fills and other methods for data enhancement;

Step 7: Use the trained convolutional neural network to extract the features of the simulated image after data enhancement;

In this embodiment, as shown in FIG. 4 , the trained convolutional neural network refers to the VGG16 convolutional neural network; the pre-trained VGG16 convolutional neural network with the top layer removed is used to perform feature extraction on the simulated image after data enhancement, which is used for Classification prediction model training;

The VGG16 convolutional neural network after removing the top layer contains 13 convolutional layers, namely conv3-XXX and five pooling layers, namely maxpool; among them, the convolutional layer adopts a 3*3 convolution kernel; 20 training samples are used as One group, use the picture generator to generate the picture group inputs_batch and label group labels_batch cyclically, and input the VGG16 convolutional neural network for feature extraction;

Step 8: Use the convolutional neural network to establish a defect image classification prediction model, and use the simulated image features obtained in step 7 for training of the classification prediction model, as shown in Figure 5.

In this embodiment, the classification prediction model is a DNN neural network model, which includes a four-layer structure. Except for the last fully-connected layer, the activation functions of other fully-connected layers use relu, and the formula is as follows:

The activation function of the last fully connected layer adopts the softmax activation function, and the formula is as follows:

为归一化系数；S_j表示经过Softmax计算后的输出结果；Softmax函数用于多分类任务的概率计算；dropout设为0.5，随机失活一部分神经元以防止过拟合；通过随机排列索引的方式打乱数据集，取其中的30％作为验证集，并在每一回合训练完成后打印出训练集和验证集对应的准确率与categorical_crossentropy损失；The Softmax function maps a vector (a ₁ , a ₂ , ..., an ) to a vector (S ₁ , S ₂ , ..., S _n ), where _n is the number of categories; a _j represents the input value of the jth output node ;

is the normalization coefficient; S _j represents the output result after Softmax calculation; the Softmax function is used for the probability calculation of multi-classification tasks; dropout is set to 0.5, and some neurons are randomly inactivated to prevent overfitting; The data set is scrambled in this way, 30% of which is taken as the validation set, and the accuracy and categorical_crossentropy loss corresponding to the training set and the validation set are printed out after each round of training is completed;

Step 9: Preprocess the experimental image;

In this embodiment, a mouse is used to interactively crop and retain the defect area to be classified; the upper and lower edges and the left and right edges of the experimental image are respectively filled with equal height or width in three RGB channels; and then the three channel images are combined to make the filling The size of the experimental image after is the same as that of the input image of the VGG16 convolutional neural network in step 7; the filling color is consistent with the image color at the defect-free position, which is [0, 0, 143] in this embodiment;

Step 10: Use the classification prediction model to predict the experimental image preprocessed in Step 9.

In another aspect, the present invention provides a defect classification system based on ultrasonic phased array imaging, comprising:

Ultrasonic phased array probe, which is used for ultrasonic phased array imaging of the sample to be tested to obtain ultrasonic full matrix data;

The all-focus module is used to perform all-focus processing on the ultrasonic full-matrix data and the simulation data, and then perform color coding according to the signal amplitude to obtain the image of the sample to be tested and the simulation image respectively;

The simulation module is used to use the finite element simulation method of ultrasonic scattering data of defects, set simulation parameters, and obtain simulation data by simulating ultrasonic phased array imaging;

an image preprocessing module, used for preprocessing the image of the sample to be tested and the simulated image;

The classification prediction module is used to input the preprocessed image of the sample to be tested into the trained classification prediction model to obtain the defect classification of the sample to be tested;

a data enhancement module, configured to perform data enhancement on the preprocessed simulation image to obtain the enhanced simulation image;

Wherein, the classification prediction model is obtained by adding a fully connected layer to the output layer of a convolutional neural network;

The training method of the classification prediction module is:

Use the trained convolutional neural network to extract image features from the enhanced simulated image;

The image features and corresponding class labels are input to the fully connected layer to train the classification prediction model.

Further preferably, the activation function of the last fully connected layer of the classification prediction model adopts the softmax activation function, and the softmax activation function is:

为归一化系数；S_j表示经过Softmax计算后的输出结果。 _Among them, the _Softmax function _maps the _vector ( _{a 1} , a ₂ , . input value;

is the normalization coefficient; S _j represents the output result after Softmax calculation.

Further preferably, the method for preprocessing the image of the sample to be tested is:

The image of the sample to be tested is interactively cropped, and the defect area to be classified is retained;

The upper and lower edges, left and right edges of the image in the defective area are filled in the height and width directions respectively by three channels of RGB; the color of the filling is consistent with the image color of the non-defective area;

Combine the three channel images so that the size of the filled image is the same as the size of the input image of the convolutional neural network;

The method of preprocessing the simulation image is as follows:

Cut out the edge image irrelevant to the defect classification for the simulated image according to the defect distribution position, and retain the defect area;

The cropped simulation images are divided into different folders according to the categories, and the data set production is completed;

Label the dataset with one-hot encoding.

Further preferably, the simulation parameters include sample material, ultrasonic testing frequency, sample defect distribution, sample defect quantity, sample defect size and phased array probe parameters;

The phased array probe parameters include probe size, array element number, center frequency, bandwidth and array element spacing.

Further preferably, the method for performing data enhancement on the simulated images in the data set is:

normalizing the simulated images in the data set;

Rotation and/or horizontal translation and/or vertical translation and/or random horizontal flip and/or filling in a nearest manner are performed on the normalized simulated image.

Further preferably, the convolutional neural network is a VGG16 convolutional neural network with the top layer removed, including 13 convolutional layers and 5 pooling layers.

To sum up, compared with the prior art, the present invention has the following advantages:

When training the classification prediction model, the present invention uses the finite element simulation method of ultrasonic scattering data for defects to simulate any defect, and obtains a simulated image after simulating ultrasonic phased array imaging and then using full focus processing. In order to construct the training set of the classification prediction model, the finite element simulation model takes into account the multiple scattering of ultrasonic waves between multiple defects, so the constructed simulation image can accurately reflect the defect problem, and the trained classification prediction model is effective for defect classification. High accuracy.

The invention adopts the advanced data post-processing method for imaging, that is, the full-focusing method is used to perform delay and superposition processing on the ultrasonic full-matrix data collected by the ultrasonic phased array, so as to obtain the internal imaging results of the measured object, and can make full use of all the collected data. signal, therefore, the imaging resolution is much higher than conventional ultrasound B-scan imaging.

In the classification prediction model of the present invention, a fully connected layer is added to the output layer of the convolutional neural network, and the trained convolutional neural network is used to extract image features. Since only the network parameters of the classification part in the classification prediction model need to be trained, the classification prediction model is greatly shortened. computation and training time. And the classification prediction model provided by the present invention has a self-updating function, that is, model training and updating can be performed when new data is added.

Those skilled in the art can easily understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (10)

1. A defect classification method based on ultrasonic phased array imaging is characterized by comprising the following steps:

acquiring ultrasonic full matrix data by adopting ultrasonic phased array imaging on a sample piece to be detected;

after the ultrasonic full matrix data is subjected to full focusing processing, color coding is carried out according to the signal amplitude value, and an image of a sample piece to be detected is obtained;

preprocessing the image of the sample to be tested, inputting the preprocessed image into a trained classification prediction model, and obtaining the defect classification of the sample to be tested;

the classification prediction model is obtained by adding a full connection layer to the output layer of the convolutional neural network;

the method for training the classification prediction model comprises the following steps:

setting simulation parameters by using a finite element simulation method of defect ultrasonic scattering data, and acquiring simulation data through simulation ultrasonic phased array imaging;

after full focusing processing is carried out on the simulation data, color coding is carried out according to the signal amplitude, and then a simulation image is obtained;

carrying out data enhancement on the preprocessed simulation image to obtain an enhanced simulation image;

carrying out image feature extraction on the enhanced simulation image by adopting the trained convolutional neural network;

and inputting the image characteristics and the corresponding class labels into a full connection layer, and training a classification prediction model.

2. The defect classification method of claim 1, wherein the simulation parameters include sample material, ultrasonic inspection frequency, sample defect distribution, sample defect number, sample defect size, and phased array probe parameters;

the phased array probe parameters comprise probe size, array element number, center frequency, bandwidth and array element spacing.

3. The defect classification method according to claim 1 or 2, wherein the method for preprocessing the simulation image is as follows:

cutting an edge image irrelevant to defect classification according to the defect distribution position of the simulation image, and reserving a defect area;

dividing the cut simulation image into different folders according to categories to finish the data set manufacturing;

and performing label setting on the data set by adopting thermal coding.

4. The defect classification method according to claim 1 or 2, wherein the method for preprocessing the image of the sample to be tested comprises:

interactively cutting the image of the sample piece to be detected, and reserving a defect area to be classified;

filling the upper edge, the lower edge, the left edge and the right edge of the image with the defect area in the height and width directions respectively by three RGB channels; wherein the filled color is consistent with the image color of the non-defective area;

and combining the three channel images to ensure that the size of the filled image is consistent with the size of the input image of the convolutional neural network.

5. The defect classification method of claim 4, wherein the method for enhancing the simulation image in the data set comprises:

normalizing the simulation image in the data set;

and (3) performing rotation and/or horizontal translation and/or vertical translation and/or random horizontal overturning and/or filling in a nearest mode on the simulation image after the normalization processing.

6. The defect classification method of claim 1, wherein the convolutional neural network is a VGG16 convolutional neural network with the top layer removed, and comprises 13 convolutional layers and 5 pooling layers.

7. The defect classification method according to claim 1 or 6, characterized in that the activation function of the last fully-connected layer of the classification prediction model adopts a softmax activation function, and the softmax activation function is:

wherein the Softmax function transforms the vector (a)₁，a₂，…，a_n) Mapping as a vector (S)₁，S₂，…，S_n) Wherein n is the number of categories; a is_jRepresents the input value of the jth output node;

is a normalized coefficient; s_jShowing the output after Softmax calculation.

8. A defect classification system based on ultrasonic phased array imaging, comprising:

the ultrasonic phased array probe is used for imaging a sample piece to be detected by adopting an ultrasonic phased array to acquire ultrasonic full matrix data;

the full-focusing module is used for carrying out full-focusing processing on the ultrasonic full-matrix data and the simulation data, carrying out color coding according to the signal amplitude, and respectively obtaining an image and a simulation image of the sample piece to be detected;

the simulation module is used for setting simulation parameters by using a finite element simulation method of the defect ultrasonic scattering data and acquiring simulation data by simulating ultrasonic phased array imaging;

the image preprocessing module is used for preprocessing the image and the simulation image of the sample piece to be detected;

the data enhancement module is used for enhancing the data of the preprocessed simulation image to obtain an enhanced simulation image;

the classification prediction module is used for inputting the preprocessed image of the sample piece to be detected into the trained column prediction model to obtain the defect classification of the sample piece to be detected;

the classification prediction model is obtained by adding a full connection layer to the output layer of the convolutional neural network;

the training method of the classification prediction module comprises the following steps:

carrying out image feature extraction on the enhanced simulation image by adopting the trained convolutional neural network;

and inputting the image characteristics and the corresponding class labels into a full connection layer, and training a classification prediction model.

9. The defect classification system of claim 8, wherein the last fully-connected layer activation function of the classification prediction model is a softmax activation function, and the softmax activation function is:

wherein the Softmax function transforms the vector (a)₁，a₂，…，a_n) Mapping as a vector (S)₁，S₂，…，S_n) Wherein n is the number of categories; a is_jRepresents the input value of the jth output node;

is a normalized coefficient; s_jShowing the output after Softmax calculation.

10. The defect classification system of claim 8 or 9, wherein the method for preprocessing the image of the sample to be tested comprises:

interactively cutting the image of the sample piece to be detected, and reserving a defect area to be classified;

filling the upper edge, the lower edge, the left edge and the right edge of the image with the defect area in the height and width directions respectively by three RGB channels; wherein the filled color is consistent with the image color of the non-defective area;

merging the three channel images to ensure that the size of the filled image is consistent with the size of an input image of the convolutional neural network;

the method for preprocessing the simulation image comprises the following steps:

cutting an edge image irrelevant to defect classification according to the defect distribution position of the simulation image, and reserving a defect area;

dividing the cut simulation image into different folders according to categories to finish the data set manufacturing; and performing label setting on the data set by adopting thermal coding.

Images