Method and device for detecting sub-surface defects in a continuously cast round billet

Abstract

The application provides a continuous casting round billet subsurface defect detection method and device, neural network parameters of a back propagation neural network are obtained through optimization based on a particle swarm optimization algorithm, and a defect type detection model is obtained by training the back propagation neural network by using training sample frequency domain signals including X frequency domain characteristic values and the true labels of the training sample frequency domain signal data, so that in the actual application process, after obtaining time domain signal data of a target continuous casting round billet, the time domain signal data is converted into frequency domain signal data; the frequency domain signal data is input into the defect type detection model, and the detection result of the subsurface defect of the target continuous casting round billet can be output. By converting the time domain signal to the frequency domain for more detailed analysis, the neural network model learns complex defect characteristics, and compared with manual experience flaw detection, different types of segregation defects can be more accurately identified, and misjudgment and missed detection are reduced.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing technology for metallic materials, and in particular to a method and apparatus for detecting subcutaneous defects in continuously cast round billets. Background Technology

[0002] Subsurface defects are a widespread type of defect in high-quality, large-section continuously cast round billets. They significantly impact the final product quality, firstly by reducing mechanical properties. For example, cracks and inclusions weaken the strength and toughness of the steel, making it more prone to fracture under stress. Bubble defects can lead to surface oxidation during rolling, forming surface cracks and further affecting the overall performance of the product. Secondly, they severely affect the surface quality of the steel. For instance, subsurface bubble defects cause irregularly distributed, irregularly sized, and differently shaped small protrusions on the billet surface. These protrusions are easily oxidized during rolling, forming surface cracks, thus affecting the product's appearance and quality.

[0003] In existing technologies, the classification and detection methods for subsurface defects in continuously cast round billets are mainly divided into destructive testing (DPT) and non-destructive testing (NDT). DPT primarily involves preparing low-magnification samples and observing the pickled surface, which is characterized by long sample preparation cycles and complex processes. Furthermore, if the diameter and weight are too large, continuously cast round billets cannot be automatically inspected on an assembly line using combined flaw detection equipment. NDT mainly relies on personnel using handheld flaw detection instruments to observe and analyze the detection waveforms, and the results depend heavily on the experience of the inspectors, making it highly subjective. Summary of the Invention

[0004] In view of this, this application provides a method and apparatus for detecting subsurface defects in continuously cast round billets, which realizes automated defect identification, reduces manual intervention, and improves detection efficiency and accuracy.

[0005] The first aspect of this application provides a method for detecting subsurface defects in continuously cast round billets, including:

[0006] Acquire the time-domain signal data of the target continuously cast round billet;

[0007] Convert the time-domain signal data into frequency-domain signal data;

[0008] The frequency domain signal data is input to the defect type detection model, and the detection result of the subsurface defect of the target continuous casting round billet is output. The defect type detection model is obtained by training a backpropagation neural network with a training sample set. The neural network parameters of the backpropagation neural network are optimized based on the particle swarm optimization algorithm. The training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data. The training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal.

[0009] Optionally, the method for constructing the defect type detection model includes:

[0010] Construct a training sample set; wherein the training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal;

[0011] Normalize the frequency domain signal data of the training samples and the true labels of the frequency domain signal data of the training samples to obtain the feature matrix and the true label vector of the frequency domain signal data of the training samples.

[0012] Initialize the backpropagation neural network and set the input parameters and training objective to obtain the initial backpropagation neural network; wherein, the input parameters include the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples;

[0013] Initialize the population, velocity, and fitness values ​​in the particle swarm optimization algorithm, and set the individual extreme values ​​and the population extreme values ​​to obtain the target particle swarm optimization algorithm;

[0014] Based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm, the neural network parameters of the initial backpropagation neural network are iteratively optimized to obtain the current optimal neural network parameters; wherein, the neural network parameters include the weights and biases of the neural network;

[0015] The current optimal neural network parameters are assigned to the initial backpropagation neural network to obtain the optimized backpropagation neural network.

[0016] The optimized backpropagation neural network is used as a new initial backpropagation neural network. The neural network parameters of the initial backpropagation neural network are iteratively optimized based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm to obtain the current optimal neural network parameters. The training is stopped when the training objective is met, and the current optimized backpropagation neural network is used as the defect type detection model.

[0017] Optionally, the initialization of the population, velocity, and fitness values ​​in the particle swarm optimization algorithm, and the setting of individual and population extreme values ​​to obtain the target particle swarm optimization algorithm, includes:

[0018] Calculate the number of all weights and biases in the initial backpropagation neural network;

[0019] Begin a loop that initializes each individual; where each individual is a weight or bias parameter.

[0020] Randomly generate initial values ​​for each individual and initialize a velocity vector for each individual;

[0021] Calculate the fitness value for each individual;

[0022] Find the minimum fitness value in the current population and extract the position of the globally optimal individual;

[0023] Initialize the individual's historical best position and individual's historical best fitness value for each individual;

[0024] Record the global optimal fitness value.

[0025] Optionally, calculating the fitness value for each individual includes:

[0026] Receives a vector containing neural network weights and thresholds;

[0027] Extract the input layer to hidden layer weights, hidden layer thresholds, hidden layer to output layer weights, and output layer thresholds from the vector containing neural network weights and thresholds, and assign them to the backpropagation neural network object;

[0028] The feature matrix of the training sample frequency domain signal data is input into the backpropagation neural network, and the predicted label vector of the training sample frequency domain signal data is output.

[0029] The predicted label vector of the training sample frequency domain signal data is denormalized to obtain the predicted label of the training sample frequency domain signal data.

[0030] Calculate the matching ratio between the predicted labels of the training sample frequency domain signal data and the true labels of the training sample frequency domain signal data;

[0031] The fitness value of an individual is determined based on the matching ratio between the predicted labels of the training sample frequency domain signal data and the true labels of the training sample frequency domain signal data.

[0032] Optionally, the currently optimal neural network parameters include a first variable, a second variable, a first bias vector, and a second bias vector. Assigning the currently optimal neural network parameters to the initial backpropagation neural network to obtain the optimized backpropagation neural network includes:

[0033] The first variable is reshaped to obtain the first matrix; wherein the dimension of the first matrix is ​​the product of the number of hidden layer neurons and the number of input layer neurons;

[0034] Assign the first matrix to the input weights of the first layer of the initial backpropagation neural network;

[0035] The second variable is reshaped to obtain the second matrix; where the dimension of the second matrix is ​​the product of the number of neurons in the output layer and the number of neurons in the hidden layer.

[0036] Assign the second matrix to the weights from the first hidden layer to the output layer;

[0037] The first bias vector is reshaped to obtain a column vector; wherein the length of the column vector is the number of hidden layer neurons;

[0038] Assign the column vector to the bias of the first hidden layer;

[0039] The second bias vector is transposed to obtain the transposed second bias vector;

[0040] The transposed second bias vector is assigned to the bias of the output layer.

[0041] Optionally, the number of input layer nodes in the backpropagation neural network model is X, and the number of output layer nodes is the number of types of the real labels.

[0042] Optionally, the time-domain signal data of the target continuously cast round billet is obtained by collecting data in three target directions along the circumference of the target continuously cast round billet using a flaw detection device.

[0043] Optionally, the three target directions are three equally divided directions: 0°, 120°, and 240°.

[0044] Optionally, the frequency domain feature values ​​are divided into mean feature, standard deviation feature, dispersion feature, center feature, root mean square feature, and main frequency band feature.

[0045] The second aspect of this application provides a device for detecting subsurface defects in continuously cast round billets, comprising:

[0046] The acquisition unit is used to acquire the time-domain signal data of the target continuously cast round billet;

[0047] A conversion unit is used to convert the time-domain signal data into frequency-domain signal data;

[0048] The detection and analysis unit is used to input the frequency domain signal data into the defect type detection model and output the detection result of the subsurface defect of the target continuous casting round billet; wherein, the defect type detection model is obtained by training a backpropagation neural network with a training sample set; the neural network parameters of the backpropagation neural network are optimized based on the particle swarm optimization algorithm; the training sample set includes training sample frequency domain signal data and the real labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal.

[0049] Optionally, the building unit of the defect type detection model includes:

[0050] A training sample set construction unit is used to construct a training sample set; wherein, the training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal;

[0051] The normalization unit is used to normalize the frequency domain signal data of the training samples and the true labels of the frequency domain signal data of the training samples, so as to obtain the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples.

[0052] The first initialization unit is used to initialize the backpropagation neural network and set the input parameters and training objectives to obtain the initial backpropagation neural network; wherein, the input parameters include the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples;

[0053] The second initialization unit is used to initialize the population, velocity, and calculate the fitness value in the particle swarm optimization algorithm, and to set the individual extreme value and the population extreme value to obtain the target particle swarm optimization algorithm.

[0054] An iterative optimization unit is used to iteratively optimize the neural network parameters of the initial backpropagation neural network based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm, to obtain the current optimal neural network parameters; wherein, the neural network parameters include the weights and biases of the neural network;

[0055] An optimization unit is used to assign the current optimal neural network parameters to the initial backpropagation neural network to obtain an optimized backpropagation neural network.

[0056] The training unit is used to take the optimized backpropagation neural network as a new initial backpropagation neural network, and execute the step of iteratively optimizing the neural network parameters of the initial backpropagation neural network based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm to obtain the current optimal neural network parameters, until the training objective is met, then training stops, and the current optimized backpropagation neural network is used as the defect type detection model.

[0057] Optionally, the second initialization unit includes:

[0058] The first computing unit is used to calculate the number of all weights and biases in the initial backpropagation neural network;

[0059] A loop unit is used to start a loop, which initializes each individual; wherein, the individual is a weight or bias parameter;

[0060] The random generation unit is used to randomly generate initial values ​​for each individual and initialize the velocity vector for each individual.

[0061] The second calculation unit is used to calculate the fitness value of each individual;

[0062] The optimal individual position determination unit is used to obtain the minimum fitness value in the current population and extract the global optimal individual position;

[0063] The third initialization unit is used to initialize the individual's historical best position and the individual's historical best fitness value for each individual;

[0064] The recording unit is used to record the global optimal fitness value.

[0065] Optionally, the second computing unit includes:

[0066] A receiving unit is used to receive a vector containing neural network weights and thresholds;

[0067] An extraction unit is used to extract the weights from the input layer to the hidden layer, the threshold of the hidden layer, the weights from the hidden layer to the output layer, and the threshold of the output layer from the vector containing neural network weights and thresholds, and assign them to the backpropagation neural network object;

[0068] The prediction unit is used to input the feature matrix of the training sample frequency domain signal data into the backpropagation neural network and output the predicted label vector of the training sample frequency domain signal data.

[0069] The inverse normalization unit is used to inverse normalize the predicted label vector of the frequency domain signal data of the training samples to obtain the predicted label of the frequency domain signal data of the training samples.

[0070] The third calculation unit is used to calculate the matching ratio between the predicted label of the training sample frequency domain signal data and the real label of the training sample frequency domain signal data;

[0071] The fitness value determination unit is used to determine the fitness value of an individual based on the matching ratio between the predicted label of the training sample frequency domain signal data and the true label of the training sample frequency domain signal data.

[0072] Optionally, the currently optimal neural network parameters include a first variable, a second variable, a first bias vector, and a second bias vector, and the optimization unit includes:

[0073] The first shaping unit is used to reshape the first variable to obtain the first matrix; wherein the dimension of the first matrix is ​​the product of the number of hidden layer neurons and the number of input layer neurons.

[0074] The first assignment unit is used to assign the first matrix to the first layer input weights of the initial backpropagation neural network;

[0075] The second shaping unit is used to reshape the second variable to obtain the second matrix; wherein the dimension of the second matrix is ​​the product of the number of neurons in the output layer and the number of neurons in the hidden layer.

[0076] The second assignment unit is used to assign the second matrix to the weights from the first hidden layer to the output layer;

[0077] The third shaping unit is used to reshape the first bias vector to obtain a column vector; wherein the length of the column vector is the number of hidden layer neurons;

[0078] The third assignment unit is used to assign the column vector to the bias of the first hidden layer;

[0079] The transpose unit is used to transpose the second bias vector to obtain the transposed second bias vector;

[0080] The fourth assignment unit is used to assign the transposed second bias vector to the bias of the output layer.

[0081] Optionally, the number of input layer nodes in the backpropagation neural network model is X, and the number of output layer nodes is the number of types of the real labels.

[0082] Optionally, the time-domain signal data of the target continuously cast round billet is obtained by collecting data in three target directions along the circumference of the target continuously cast round billet using a flaw detection device.

[0083] Optionally, the three target directions are three equally divided directions: 0°, 120°, and 240°.

[0084] Optionally, the frequency domain feature values ​​are divided into mean feature, standard deviation feature, dispersion feature, center feature, root mean square feature, and main frequency band feature.

[0085] A third aspect of this application provides an electronic device, comprising:

[0086] One or more processors;

[0087] A storage device on which one or more programs are stored;

[0088] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for evaluating the power limit of wind turbine inertia support as described in any one of the first aspects.

[0089] The fourth aspect of this application provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the evaluation method for the power limit of wind turbine inertia support as described in any one of the first aspects.

[0090] As can be seen from the above scheme, this application provides a method and apparatus for detecting subsurface defects in continuously cast round billets. The method involves optimizing the backpropagation neural network parameters using a particle swarm optimization algorithm, and training the backpropagation neural network with training sample frequency domain signals including X frequency domain feature values ​​and the real labels of the training sample frequency domain signal data to obtain a defect type detection model. In practical applications, after acquiring the time domain signal data of the target continuously cast round billet, the time domain signal data is converted into frequency domain signal data. The frequency domain signal data is then input into the defect type detection model to output the detection result of subsurface defects in the target continuously cast round billet. By converting the time domain signal to the frequency domain for more detailed analysis, and utilizing the neural network model to learn complex defect characteristics, this method can more accurately identify different types of segregation defects compared to manual flaw detection, reducing misjudgments and missed detections. Attached Figure Description

[0091] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0092] Figure 1 A detailed flowchart of a method for detecting subcutaneous defects in continuously cast round billets provided in this application embodiment;

[0093] Figure 2 A schematic diagram showing the sampling point location for a continuously cast round billet, provided for another embodiment of this application;

[0094] Figure 3 A detailed flowchart of a method for constructing a defect type detection model provided in another embodiment of this application;

[0095] Figure 4 A schematic diagram of the structure of a backpropagation neural network is provided in another embodiment of this application;

[0096] Figure 5 A flowchart illustrating a method for calculating the fitness value of each individual, provided as another embodiment of this application;

[0097] Figure 6 A comparison chart of prediction results for a training sample set is provided for another embodiment of this application;

[0098] Figure 7 A comparison chart of prediction results for a test sample set is provided for another embodiment of this application;

[0099] Figure 8 A fitness curve of a backpropagation neural network provided in another embodiment of this application;

[0100] Figure 9 A schematic diagram of a training sample set confusion matrix provided for another embodiment of this application;

[0101] Figure 10 A schematic diagram of a test sample set confusion matrix provided for another embodiment of this application;

[0102] Figure 11 A schematic diagram of a device for detecting subsurface defects in continuously cast round billets, provided for another embodiment of this application;

[0103] Figure 12 This is a schematic diagram of an electronic device for implementing a method for detecting subsurface defects in continuously cast round billets, as provided in another embodiment of this application. Detailed Implementation

[0104] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0105] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0106] It should be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0107] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0108] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0109] This application provides a method for detecting subsurface defects in continuously cast round billets, such as... Figure 1 As shown, the specific steps include:

[0110] S101. Obtain the time-domain signal data of the target continuously cast round billet.

[0111] In the specific implementation of this application, time-domain signal data of the target continuous casting billet can be obtained by using a flaw detection instrument to collect data along the circumference of the target continuous casting billet in three target directions, but this is not limited to the following.

[0112] The three target directions can be, but are not limited to, three equally divided directions: 0°, 120°, and 240°. Specific point selection methods can be as follows: Figure 2 As shown.

[0113] In the specific implementation of this application, time-domain signal data can be acquired for steel types (H13 abrasive steel) and specifications (800mm continuous casting round billet), but not limited to these.

[0114] S102. Convert the time-domain signal data into the frequency-domain signal data.

[0115] In the specific implementation of this application, time-domain signal data can be converted into frequency-domain signal data by means of, but not limited to, Fast Fourier Transform (FFT), which is not limited here.

[0116] Taking an instrument sampling frequency of 16MHz as an example, since the frequency domain waveform is symmetrical, only one side of the spectrum is taken. The specific implementation method for converting time-domain signal data into frequency-domain signal data can be as follows:

[0117] % Get the filenames of all CSV files in the folder

[0118] file_list=dir(fullfile(folder_path,'*.csv'));

[0119] % Loop through the data in the third row of each Excel file and plot the graph.

[0120] for i=1:length(file_list)

[0121] % Read Excel file

[0122] file_name=fullfile(folder_path, file_list(i).name);

[0123] data = xlsread(file_name);

[0124] % Get the third row of data

[0125] row_data = data(3,:);

[0126] num_elements=length(row_data);

[0127] t=(0:num_elements-1) / 16;

[0128] % Drawing

[0129] figure;

[0130] plot(t, row_data);

[0131] % xlabel('time / μs')

[0132] % ylabel('amplitude')

[0133] title(['File', num2str(i)]);

[0134] fs=16000000; % Sampling frequency

[0135] N = length(row_data); % Signal length

[0136] f = (0:N-1)*fs / N; % Frequency axis

[0137] figure;

[0138] dft = fft(row_data, N);

[0139] stem(f, abs(dft));

[0140] a = abs(dft);

[0141] xlabel('f / MHz');

[0142] xlim([0, 8000000]);

[0143] ylabel('amplitude');

[0144] title(['File', numstr(i)]);

[0145] % title('Frequency Spectrum');

[0146] % Save image

[0147] save_name=fullfile(save_path, ['File', num2str(i), '.png']);

[0148] saveas(gcf, save_name);

[0149] end.

[0150] S103. Input the frequency domain signal data into the input defect type detection model, and output the detection results of subcutaneous defects of the target continuous casting round billet.

[0151] The defect type detection model is obtained by training the backpropagation neural network with the training sample set; the neural network parameters of the backpropagation neural network are optimized based on the particle swarm optimization algorithm; the training sample set includes training sample frequency domain signal data and the real labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal.

[0152] It should be noted that the frequency domain characteristic values ​​of a frequency domain signal can be divided into mean characteristics, standard deviation characteristics, dispersion characteristics, center characteristics, root mean square characteristics, and main frequency band characteristics, which are not limited here.

[0153] In the specific implementation of this application, the calculation method of the frequency domain eigenvalues ​​of the frequency domain signal is shown in Table 1, wherein, Let N be the frequency amplitude of the sampled data points of the signal, and N be the number of sampled data points per sample.

[0154] Table 1

[0155]

[0156] In the actual application of this application, the number of input layer nodes of the backpropagation neural network model is X, and the number of output layer nodes is the number of types of real labels, which is not limited here.

[0157] The types of genuine labels include, but are not limited to, subcutaneous blister, subcutaneous inclusion, subcutaneous crack, and normal subcutaneous tissue, etc., and are not limited here.

[0158] Optionally, in another embodiment of this application, one implementation of the defect type detection model construction method is as follows: Figure 3 As shown, it includes:

[0159] S301. Construct a training sample set.

[0160] The training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal.

[0161] In the specific implementation of this application, the data sample set includes the above 6 features and 4 sample labels (subcutaneous bubbles, subcutaneous inclusions, subcutaneous cracks, and normal subcutaneous tissue). Assuming there are 360 ​​samples, the ratio of training sample set to test sample set can be shown in Table 2.

[0162] Table 2

[0163] subcutaneous abscess 60 30 Subcutaneous inclusions 60 30 subcutaneous cracks 60 30 Normal subcutaneous tissue 60 30 total 240 120

[0164] S302. Normalize the frequency domain signal data of the training samples and the true labels of the frequency domain signal data of the training samples to obtain the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples.

[0165] In the specific implementation of this application, the purpose of normalization is to eliminate differences in feature dimensions and improve the convergence speed and performance of the model; and by vectorizing the labels, the classification labels are converted into a neural network-friendly format; when reusing parameters, the test sample set uses the normalized parameters of the training sample set to ensure data consistency.

[0166] S303. Initialize the backpropagation neural network, and set the input parameters and training target to obtain the initial backpropagation neural network.

[0167] The input parameters include the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples.

[0168] In the specific implementation of this application, the number of nodes in the backpropagation neural network is set to define the number of layer nodes in the neural network.

[0169] First, the number of input layer nodes is determined to be 6 based on the feature dimension (X) of the normalized training data. Second, the number of hidden layer nodes is determined to be 5 by manually specifying the number of neurons in the hidden layer. Finally, the number of output layer nodes is determined to be 4 based on the encoding dimension of the labels (the number of types of real labels). Figure 4 This is a diagram of the backpropagation neural network model framework. W represents the weights, and B represents the biases.

[0170] Initialize the backpropagation neural network to prepare for subsequent training. Specifically, this includes the following input parameters: p_train: normalized training sample frequency domain signal data (feature matrix); t_train: target output (true labels of vectorized training sample frequency domain signal data); hiddennum: number of neurons in the hidden layer; and output: net: the generated feedforward neural network object, containing the network structure and initial weights.

[0171] Configure the training parameters for the backpropagation neural network. First, set the maximum number of training iterations, with a maximum of 1000 epochs. If the target error is not reached, training will stop early (to prevent overfitting). Second, set the target error threshold, which will automatically terminate training when the training error (e.g., mean squared error, MSE) is ≤1e-6. Third, set the learning rate to 0.01. Finally, close the training progress window.

[0172] S304. Initialize the population, velocity, and fitness values ​​in the particle swarm optimization algorithm, and set the individual extreme values ​​and the population extreme values ​​to obtain the target particle swarm optimization algorithm.

[0173] In the specific implementation process of this application, the parameters of the particle swarm optimization algorithm are first initialized, such as: individual learning factor 4.494; social learning factor 4.494; maximum number of iterations 30; population size, i.e., number of particles 5; maximum particle velocity 1.0; minimum particle velocity -1.0; upper bound of search space 2.0; lower bound of search space -2.0.

[0174] Specifically, initializing the population, velocity, and calculating the fitness value in the particle swarm optimization algorithm includes the following:

[0175] First, the number of weights and biases in the initial backpropagation neural network is calculated. Then, a loop is started to initialize each individual. Next, initial values ​​are randomly generated for each individual, and a velocity vector is initialized for each individual. Finally, the fitness value of each individual is calculated.

[0176] Here, each individual represents a weight or bias parameter. A velocity vector is initialized for each individual to update its parameters.

[0177] In the specific implementation of this application, one method for calculating the fitness value of each individual is as follows: Figure 5 As shown, it includes:

[0178] S501, Receive a vector containing neural network weights and thresholds.

[0179] In the practical application of this application, a function named `fun` can be defined to receive the following five parameters and return an error value: `error`. These parameters are: `pop`: a vector containing the weights and thresholds of the neural network; `hiddennum`: the number of neurons in the hidden layer; `net`: an untrained neural network object; `p_train`: the input data of the training sample set (the frequency domain signal data of the training samples); and `t_train`: the target data of the training sample set (the true labels of the frequency domain signal data of the training samples).

[0180] S502. Extract the weights from the input layer to the hidden layer, the threshold of the hidden layer, the weights from the hidden layer to the output layer, and the threshold of the output layer from the vector containing the weights and thresholds of the neural network, and assign them to the backpropagation neural network object.

[0181] Specifically, the weights from the input layer to the hidden layer, the threshold of the hidden layer, the weights from the hidden layer to the output layer, and the threshold of the output layer are assigned to the corresponding layers of the neural network object net.

[0182] In the specific implementation of this application, the reshape function can be used, but is not limited to, to convert the vector into an appropriate matrix shape; no limitation is made here.

[0183] S503. Input the feature matrix of the training sample frequency domain signal data into the backpropagation neural network, and output the predicted label vector of the training sample frequency domain signal data.

[0184] Specifically, the sim function can be used, but is not limited to, to simulate a neural network and obtain the predicted label vector of the frequency domain signal data of the training samples.

[0185] S504. Perform inverse normalization on the predicted label vector of the training sample frequency domain signal data to obtain the predicted label of the training sample frequency domain signal data.

[0186] In the specific implementation of this application, the method for inverse normalizing the predicted label vector of the training sample frequency domain signal data and the real label vector of the training sample frequency domain signal data can be, but is not limited to, the vec2ind function, and is not limited here.

[0187] The category index (label) is obtained by inverse normalizing the output vector of the neural network.

[0188] S505. Calculate the matching ratio between the predicted labels of the training sample frequency domain signal data and the true labels of the training sample frequency domain signal data.

[0189] In the specific implementation of this application, the matching ratio between the predicted label of the training sample frequency domain signal data and the real label of the training sample frequency domain signal data can be determined by, but is not limited to, cosine similarity. No limitation is made here.

[0190] S506. Determine the fitness value of an individual based on the matching ratio between the predicted label of the training sample frequency domain signal data and the true label of the training sample frequency domain signal data.

[0191] In the specific implementation of this application, the fitness value, also known as the error value, can be obtained by subtracting the matching ratio from 1. This error value is used to evaluate the quality of the current parameter set (i.e., pop). The smaller the error value, the better the parameter set.

[0192] Specifically, the methods for setting individual and group extreme values ​​include the following:

[0193] First, obtain the minimum fitness value in the current population and extract the global best individual position. Then, initialize the historical best position and historical best fitness value of each individual. Finally, record the global best fitness value for subsequent drawing of convergence curves or termination condition judgment.

[0194] S305. Based on the feature matrix of the frequency domain signal data of the training samples, the true label vector of the frequency domain signal data of the training samples, and the target particle swarm optimization algorithm, the neural network parameters of the initial backpropagation neural network are iteratively optimized to obtain the current optimal neural network parameters.

[0195] The neural network parameters include the weights and biases of the neural network.

[0196] In the specific implementation of this application, the specific implementation of the target particle swarm optimization (PSO) algorithm for iterative optimization can be, but is not limited to, searching for the optimal parameter set of the backpropagation neural network step by step through velocity updates, position adjustments, adaptive mutations, etc. The process of step-by-step searching for the optimal parameter set of the backpropagation neural network includes, but is not limited to, velocity updates, population updates, adaptive mutations, fitness value calculations, individual optimality, and population optimality updates, which are not limited here.

[0197] Specifically, based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm, the neural network parameters of the initial backpropagation neural network are iteratively optimized to obtain an optimal solution vector zbest. Then, the weights and bias parameters of the neural network are extracted from zbest and decomposed into specific matrices and vectors according to the network structure to improve model performance. This includes the first variable w1: the weight matrix from the input layer to the hidden layer (dimension: inputnum × hiddennum); the first bias vector B1: the bias vector of the hidden layer (dimension: hiddennum × 1); the second variable w2: the weight matrix from the hidden layer to the output layer (dimension: hiddennum × outputnum); and the second bias vector B2: the bias vector of the output layer (dimension: outputnum × 1).

[0198] S306. Assign the current optimal neural network parameters to the initial backpropagation neural network to obtain the optimized backpropagation neural network.

[0199] Optionally, in another embodiment of this application, one implementation of step S306 includes:

[0200] The first variable is reshaped to obtain the first matrix; the dimension of the first matrix is ​​the product of the number of hidden layer neurons (hiddennum) and the number of input layer neurons or features (inputnum); the first matrix is ​​assigned to the first layer input weights net.Iw{1,1} of the initial backpropagation neural network; the second variable is reshaped to obtain the second matrix; the dimension of the second matrix is ​​the product of the number of output layer neurons (outputnum) and the number of hidden layer neurons (hiddennum); the second matrix is ​​assigned to the weights from the first hidden layer to the output layer (net.Lw{2,1}); the first bias vector is reshaped to obtain a column vector; the length of the column vector is the number of hidden layer neurons (hiddennum); the column vector is assigned to the bias of the first hidden layer (net.b{1}); the second bias vector (assuming it is a row vector or matrix) is transposed to obtain the transposed second bias vector; the transposed second bias vector is assigned to the output layer bias net.b{2}.

[0201] S307. Use the optimized backpropagation neural network as the new initial backpropagation neural network.

[0202] S308. When the training objective is met, stop training and use the current optimized backpropagation neural network as the defect type detection model.

[0203] In the specific implementation of this application, during the network training process, a training monitoring interface can be enabled to display the following key indicators in real time: training error (MSE), validation error, gradient change, iteration progress, etc., without limitation here.

[0204] In the specific implementation of this application, the backpropagation neural network is iteratively optimized using the feature matrix of the training sample frequency domain signal data and the true label vector of the training sample frequency domain signal data, and the weights and bias parameters in the network are automatically adjusted.

[0205] In the specific implementation of this application, the `sim` function can also be called to perform forward propagation calculations on the input data `p_train` and `p_test` using the pre-trained neural network `net`, generating predicted outputs. The model's ability to fit known data is verified through predictions on the training sample set (`t_sim1`); the model's generalization performance on unknown data is evaluated through predictions on the test sample set (`t_sim2`).

[0206] In the specific implementation of this application, T_train and T_test can be sorted, and the sorted indices can be obtained. Secondly, these indices can be used to rearrange T_sim1 and T_sim2, and finally, it is ensured that they are consistent with the sorted T_train and T_test.

[0207] Figure 6 and Figure 7 This section compares the prediction results for the training and testing sample sets. It calculates the percentage accuracy of the model on each set. Stars represent the true values, and circles represent the predicted values. When they overlap (stars and circles overlap),... Figure 6 and Figure 7 The underlined portion (in the diagram) indicates a correct prediction, and vice versa. The accuracy rate on the training sample set is 97.9%, and the accuracy rate on the test sample set is 92.3%.

[0208] Figure 8 As shown in the fitness curve, the fitness value (error rate) of the backpropagation neural network gradually decreases with the increase of the number of ion swarm iterations.

[0209] Figure 9 , Figure 10 These are the confusion matrices for the training and testing sample sets, respectively. The accuracy rates of the training sample set for the four types of subcutaneous defect labels are 98.4%, 96.7%, 98.2%, and 98.3%, respectively, while the accuracy rates of the testing sample set for the four types of subcutaneous defect labels are 92.3%, 100%, 84.2%, and 96.6%, respectively.

[0210] Classification performance comparison. Traditional convolutional neural network (CNN) and long short-term memory (LSTM) models with time-series data as input were used as comparison objects, and softmax classifier was used as output. Test sample data was selected and the model was run 5 times consecutively. The classification accuracy of the model is shown in Table 3.

[0211] Table 3

[0212]

[0213] As shown in Table 3, the average accuracy of the prediction results of the CNN model and the LSTM model is 76.87% and 83.80%, respectively. The average accuracy of the prediction results of the BP-PS (the defect type detection model in this invention) on the test sample set is 92.31%. The proposed BP-PSO method model improves the classification accuracy by about 15.44% and 8.51% compared with the CNN and LSTM method models, respectively.

[0214] As can be seen from the above scheme, this application provides a method for detecting subsurface defects in continuously cast round billets. The method optimizes the backpropagation neural network parameters using a particle swarm optimization algorithm, and trains the backpropagation neural network with training sample frequency domain signals including X frequency domain feature values ​​and the real labels of the training sample frequency domain signal data to obtain a defect type detection model. In practical applications, after acquiring the time domain signal data of the target continuously cast round billet, the time domain signal data is converted into frequency domain signal data. The frequency domain signal data is then input into the defect type detection model, which outputs the detection result of subsurface defects in the target continuously cast round billet. By converting the time domain signal to the frequency domain for more detailed analysis, and utilizing the neural network model to learn complex defect features, this method can more accurately identify different types of segregation defects compared to manual flaw detection, reducing misjudgments and missed detections.

[0215] Another embodiment of this application provides a device for detecting subsurface defects in continuously cast round billets, such as... Figure 11 As shown, it specifically includes:

[0216] Acquisition unit 1101 is used to acquire time-domain signal data of the target continuously cast round billet.

[0217] The conversion unit 1102 is used to convert time-domain signal data into frequency-domain signal data.

[0218] The detection and analysis unit 1103 is used to input frequency domain signal data into the defect type detection model and output the detection results of subcutaneous defects of the target continuous casting round billet.

[0219] The defect type detection model is obtained by training the backpropagation neural network with the training sample set; the neural network parameters of the backpropagation neural network are optimized based on the particle swarm optimization algorithm; the training sample set includes training sample frequency domain signal data and the real labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal.

[0220] For details on the specific working process of the units disclosed in the above embodiments of this application, please refer to the corresponding method embodiments, such as... Figure 1 As shown, it will not be elaborated further here.

[0221] Optionally, in another embodiment of this application, one implementation of the construction unit of the defect type detection model includes:

[0222] Training sample set construction unit, used to construct the training sample set.

[0223] The training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal.

[0224] The normalization unit is used to normalize the frequency domain signal data of the training samples and the true labels of the frequency domain signal data of the training samples, so as to obtain the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples.

[0225] The first initialization unit is used to initialize the backpropagation neural network and set the input parameters and training target to obtain the initial backpropagation neural network.

[0226] The input parameters include the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples.

[0227] The second initialization unit is used to initialize the population, velocity, and calculate the fitness value in the particle swarm optimization algorithm, and to set the individual extreme value and the population extreme value to obtain the target particle swarm optimization algorithm.

[0228] The iterative optimization unit is used to iteratively optimize the neural network parameters of the initial backpropagation neural network based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm, so as to obtain the current optimal neural network parameters.

[0229] The neural network parameters include the weights and biases of the neural network.

[0230] The optimization unit is used to assign the current optimal neural network parameters to the initial backpropagation neural network to obtain the optimized backpropagation neural network.

[0231] The training unit is used to take the optimized backpropagation neural network as a new initial backpropagation neural network, and perform iterative optimization of the neural network parameters of the initial backpropagation neural network based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm, until the training objective is met, the training stops, and the current optimized backpropagation neural network is used as the defect type detection model.

[0232] For details on the specific working process of the units disclosed in the above embodiments of this application, please refer to the corresponding method embodiments, such as... Figure 3 As shown, it will not be elaborated further here.

[0233] Optionally, in another embodiment of this application, one implementation of the second initialization unit includes:

[0234] The first computational unit is used to calculate the number of all weights and biases in the initial backpropagation neural network.

[0235] The loop unit is used to start a loop, which initializes each individual.

[0236] Here, each individual is a weight or bias parameter.

[0237] The random generation unit is used to randomly generate initial values ​​for each individual and initialize the velocity vector for each individual.

[0238] The second calculation unit is used to calculate the fitness value of each individual.

[0239] The optimal individual position determination unit is used to obtain the minimum fitness value in the current population and extract the global optimal individual position.

[0240] The third initialization unit is used to initialize the individual's historical best position and the individual's historical best fitness value for each individual.

[0241] The recording unit is used to record the global optimal fitness value.

[0242] For details on the specific working process of the units disclosed in the above embodiments of this application, please refer to the corresponding method embodiments, which will not be repeated here.

[0243] Optionally, in another embodiment of this application, one implementation of the second computing unit includes:

[0244] The receiving unit is used to receive a vector containing the weights and thresholds of the neural network.

[0245] The extraction unit is used to extract the input layer to hidden layer weights, hidden layer thresholds, hidden layer to output layer weights, and output layer thresholds from the vector containing neural network weights and thresholds, and assign them to the backpropagation neural network object.

[0246] The prediction unit is used to input the feature matrix of the frequency domain signal data of the training samples into the backpropagation neural network and output the predicted label vector of the frequency domain signal data of the training samples.

[0247] The inverse normalization unit is used to inverse normalize the predicted label vector of the frequency domain signal data of the training samples to obtain the predicted label of the frequency domain signal data of the training samples.

[0248] The third calculation unit is used to calculate the matching ratio between the predicted labels of the training sample frequency domain signal data and the true labels of the training sample frequency domain signal data.

[0249] The fitness value determination unit is used to determine the fitness value of an individual based on the matching ratio between the predicted label of the training sample frequency domain signal data and the true label of the training sample frequency domain signal data.

[0250] For details on the specific working process of the units disclosed in the above embodiments of this application, please refer to the corresponding method embodiments, such as... Figure 5 As shown, it will not be elaborated further here.

[0251] Optionally, in another embodiment of this application, the currently optimal neural network parameters include a first variable, a second variable, a first bias vector, and a second bias vector. One implementation of the optimization unit includes:

[0252] The first shaping unit is used to reshape the first variable to obtain the first matrix.

[0253] The dimension of the first matrix is ​​the product of the number of neurons in the hidden layer and the number of neurons in the input layer.

[0254] The first assignment unit is used to assign the first matrix to the first layer input weights of the initial backpropagation neural network.

[0255] The second shaping unit is used to reshape the second variable to obtain the second matrix.

[0256] The dimension of the second matrix is ​​the product of the number of neurons in the output layer and the number of neurons in the hidden layer.

[0257] The second assignment unit is used to assign the second matrix to the weights from the first hidden layer to the output layer.

[0258] The third shaping unit is used to reshape the first bias vector to obtain a column vector.

[0259] The length of the column vector is the number of neurons in the hidden layer.

[0260] The third assignment unit is used to assign column vectors to the bias of the first hidden layer.

[0261] The transpose unit is used to transpose the second bias vector to obtain the transposed second bias vector.

[0262] The fourth assignment unit is used to assign the transposed second bias vector to the bias of the output layer.

[0263] For details on the specific working process of the units disclosed in the above embodiments of this application, please refer to the corresponding method embodiments, which will not be repeated here.

[0264] Optionally, in another embodiment of this application, the number of input layer nodes of the backpropagation neural network model is X, and the number of output layer nodes is the number of types of the real labels.

[0265] Optionally, in another embodiment of this application, the time-domain signal data of the target continuous casting billet is obtained by collecting data in three target directions along the circumferential surface of the target continuous casting billet using a flaw detection device.

[0266] Optionally, in another embodiment of this application, the three target directions are three equally divided directions: 0°, 120°, and 240°.

[0267] Optionally, frequency domain features can be categorized into mean features, standard deviation features, dispersion features, center features, root mean square features, and main frequency band features.

[0268] As can be seen from the above scheme, this application provides a detection device for subsurface defects in continuously cast round billets. The device optimizes the backpropagation neural network parameters using a particle swarm optimization algorithm, and trains the backpropagation neural network with training sample frequency domain signals including X frequency domain feature values ​​and the real labels of the training sample frequency domain signal data to obtain a defect type detection model. In practical applications, after acquiring the time domain signal data of the target continuously cast round billet, the time domain signal data is converted into frequency domain signal data. The frequency domain signal data is then input into the defect type detection model, which outputs the detection result of subsurface defects in the target continuously cast round billet. By converting the time domain signal to the frequency domain for more detailed analysis, and utilizing the neural network model to learn complex defect features, it can more accurately identify different types of segregation defects compared to manual flaw detection, reducing misjudgments and missed detections.

[0269] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0270] Another embodiment of this application provides an electronic device, such as... Figure 12 As shown, it includes:

[0271] One or more processors 1201.

[0272] Storage device 1202, on which one or more programs are stored.

[0273] When the one or more programs are executed by the one or more processors 1201, the one or more processors 1201 implement the method for detecting subsurface defects in continuously cast round billets as described in the above embodiments.

[0274] Another embodiment of this application provides a computer storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the method for detecting subsurface defects in continuously cast round billets as described in the above embodiments.

[0275] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0276] It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0277] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0278] Another embodiment of this application provides a computer program product, which, when executed, is used to perform the above-described method for detecting subsurface defects in continuously cast round billets.

[0279] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments of this application.

[0280] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in this application is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely exemplary forms of implementing this application.

[0281] While several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0282] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described application concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. A method for detecting subsurface defects in continuously cast round billets, characterized in that, include: Acquire the time-domain signal data of the target continuously cast round billet; Convert the time-domain signal data into frequency-domain signal data; The frequency domain signal data is input into the defect type detection model, and the detection result of the subsurface defect of the target continuously cast round billet is output. The defect type detection model is obtained by training a backpropagation neural network using a training sample set. The neural network parameters of the backpropagation neural network are optimized based on a target particle swarm optimization algorithm. The training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data. The training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal. The target particle swarm optimization algorithm is obtained by initializing the population, velocity, and fitness value in the particle swarm optimization algorithm, and setting individual extreme values ​​and population extreme values. The process of initializing the population, velocity, and fitness values ​​in the particle swarm optimization algorithm, and setting individual and population extrema to obtain the target particle swarm optimization algorithm includes: Calculate the number of all weights and biases in the initial backpropagation neural network; Begin a loop that initializes each individual; where each individual is a weight or bias parameter. Randomly generate initial values ​​for each individual and initialize a velocity vector for each individual; Calculate the fitness value for each individual; Find the minimum fitness value in the current population and extract the position of the globally optimal individual; Initialize the individual's historical best position and individual's historical best fitness value for each individual; Record the global optimal fitness value.

2. The method for detecting subsurface defects in continuously cast round billets according to claim 1, characterized in that, The method for constructing the defect type detection model includes: Construct a training sample set; wherein the training sample set includes training sample frequency domain signal data and the true labels of the training sample frequency domain signal data; the training sample frequency domain signal data includes X frequency domain feature values ​​of the training sample frequency domain signal; Normalize the frequency domain signal data of the training samples and the true labels of the frequency domain signal data of the training samples to obtain the feature matrix and the true label vector of the frequency domain signal data of the training samples. Initialize the backpropagation neural network and set the input parameters and training objective to obtain the initial backpropagation neural network; wherein, the input parameters include the feature matrix of the frequency domain signal data of the training samples and the true label vector of the frequency domain signal data of the training samples; Initialize the population, velocity, and fitness values ​​in the particle swarm optimization algorithm, and set the individual extreme values ​​and the population extreme values ​​to obtain the target particle swarm optimization algorithm; Based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm, the neural network parameters of the initial backpropagation neural network are iteratively optimized to obtain the current optimal neural network parameters; wherein, the neural network parameters include the weights and biases of the neural network; The current optimal neural network parameters are assigned to the initial backpropagation neural network to obtain the optimized backpropagation neural network. The optimized backpropagation neural network is used as a new initial backpropagation neural network. The neural network parameters of the initial backpropagation neural network are iteratively optimized based on the feature matrix of the training sample frequency domain signal data, the true label vector of the training sample frequency domain signal data, and the target particle swarm optimization algorithm to obtain the current optimal neural network parameters. The training is stopped when the training objective is met, and the current optimized backpropagation neural network is used as the defect type detection model.

3. The method for detecting subsurface defects in continuously cast round billets according to claim 1, characterized in that, The calculation of the fitness value for each individual includes: Receives a vector containing neural network weights and thresholds; Extract the input layer to hidden layer weights, hidden layer thresholds, hidden layer to output layer weights, and output layer thresholds from the vector containing neural network weights and thresholds, and assign them to the backpropagation neural network object; The feature matrix of the training sample frequency domain signal data is input into the backpropagation neural network, and the predicted label vector of the training sample frequency domain signal data is output. The predicted label vector of the training sample frequency domain signal data is denormalized to obtain the predicted label of the training sample frequency domain signal data. Calculate the matching ratio between the predicted labels of the training sample frequency domain signal data and the true labels of the training sample frequency domain signal data; The fitness value of an individual is determined based on the matching ratio between the predicted labels of the training sample frequency domain signal data and the true labels of the training sample frequency domain signal data.

4. The method for detecting subsurface defects in continuously cast round billets according to claim 2, characterized in that, The currently optimal neural network parameters include a first variable, a second variable, a first bias vector, and a second bias vector. Assigning the currently optimal neural network parameters to the initial backpropagation neural network to obtain the optimized backpropagation neural network includes: The first variable is reshaped to obtain the first matrix; wherein the dimension of the first matrix is ​​the product of the number of hidden layer neurons and the number of input layer neurons; Assign the first matrix to the input weights of the first layer of the initial backpropagation neural network; The second variable is reshaped to obtain the second matrix; where the dimension of the second matrix is ​​the product of the number of neurons in the output layer and the number of neurons in the hidden layer. Assign the second matrix to the weights from the first hidden layer to the output layer; The first bias vector is reshaped to obtain a column vector; wherein the length of the column vector is the number of hidden layer neurons; Assign the column vector to the bias of the first hidden layer; The second bias vector is transposed to obtain the transposed second bias vector; The transposed second bias vector is assigned to the bias of the output layer.

5. The method for detecting subsurface defects in continuously cast round billets according to claim 1, characterized in that, The number of input layer nodes in the backpropagation neural network model is X, and the number of output layer nodes is the number of types of the real labels.

6. The method for detecting subsurface defects in continuously cast round billets according to claim 1, characterized in that, The time-domain signal data of the target continuously cast round billet is obtained by collecting data in three target directions along the circumference of the target continuously cast round billet using a flaw detection device.

7. The method for detecting subsurface defects in continuously cast round billets according to claim 6, characterized in that, The three target directions are divided into three equal parts: 0°, 120°, and 240°.

8. The method for detecting subsurface defects in continuously cast round billets according to claim 1, characterized in that, The frequency domain features are divided into mean features, standard deviation features, dispersion features, center features, root mean square features, and main frequency band features.

9. A device for detecting subsurface defects in continuously cast round billets, characterized in that, The method for detecting subcutaneous defects in continuously cast round billets according to any one of claims 1-8.

Images