Intelligent identification method and system for pipeline weld defect, computing equipment and medium

Abstract

The invention relates to a pipeline weld defect intelligent identification method and system, a computing device and a medium, and the method comprises the steps: analyzing an AUT scanning file by using an analytic function, and determining waveform data obtained by scanning a pipeline at different scanning positions according to a scanning direction; marking each waveform data according to defect types, and determining a training set; counting the sample number corresponding to each defect type in the training set, and performing data expansion on the sample number of the defect type of which the sample number is less than a preset value to determine a first target training set; inputting the first target training set into a convolutional neural network for training, and determining an identification network; and according to the identification network, identifying the pipeline welding seam defect. The convolutional neural network is trained through the training set of the data type, the training time is short, the requirement for computer performance is low, and the recognition speed is high.

Description

Background Art

[0002] Automatic Ultrasonic Testing (AUT) is a non-destructive testing technology for weld quality based on ultrasonic testing technology. It has the advantages of high detection sensitivity, fast speed, accurate positioning and quantification of weld defects, etc. It has been applied and popularized in the weld quality inspection of industrial pipelines, improving the inspection quality and work efficiency and reducing the intensity of inspection operations. Automatic ultrasonic testing is used to detect the circumferential welds of pipelines according to the principles of ultrasonic phased array acoustic focusing and sectional scanning. The corresponding AUT scanning atlas is obtained through sectional scanning, and the weld quality is comprehensively evaluated based on the strip chart (A-scan), volume channel (B-scan) and TOFD scanning atlas.

[0003] In the existing technology based on automatic ultrasonic testing, the images obtained by AUT scanning are usually input into an identification network for classification and prediction. However, if image data is used during the training process of the identification model, the training time is long, the requirements for computer performance are high, and the identification speed is slow. Summary of the Invention

[0004] In order to overcome the problems that when using image data during the training process of the identification model, the training time is long, the requirements for computer performance are high, and the identification speed is slow, the present invention provides an intelligent identification method, system, computing device and medium for pipeline weld defects.

[0005] In the first aspect, to solve the above technical problems, the present invention provides an intelligent identification method for pipeline weld defects, including:

[0006] Using an analytical function to parse the AUT scan file to determine the waveform data obtained by scanning the pipeline at different scanning positions in the scanning direction;

[0007] Marking each waveform data according to the defect type to determine the training set;

[0008] Counting the number of samples corresponding to each defect type in the training set, and performing data augmentation on the samples of the defect types with the number of samples less than the preset value to determine the first target training set;

[0009] Inputting the first target training set into a convolutional neural network for training to determine the identification network;

[0010] Identifying pipeline weld defects according to the identification network.

[0011] In the second aspect, the present invention provides an intelligent identification system for pipeline weld defects, including:

[0012] A waveform data determination module, configured to use an analytical function to parse the AUT scan file to determine the waveform data obtained by scanning the pipeline at different scanning positions in the scanning direction;

[0013] A training set determination module, configured to label each waveform data according to the defect type and determine a training set;

[0014] A first target training set determination module, configured to count the number of samples corresponding to each defect type in the training set, perform data augmentation on the number of samples of the defect type with the number of samples less than a preset value, and determine a first target training set;

[0015] An identification network determination module, configured to input the first target training set into a convolutional neural network for training to determine an identification network;

[0016] An identification module, configured to identify pipeline weld defects according to the identification network.

[0017] In a third aspect, the present invention further provides a computing device, including a memory, a processor, and a program stored on the memory and running on the processor. When the processor executes the program, the steps of an intelligent identification method for pipeline weld defects as described above are implemented.

[0018] In a fourth aspect, the present invention further provides a computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions run on a terminal device, the terminal device is caused to execute the steps of an intelligent identification method for pipeline weld defects.

[0019] The beneficial effects of the present invention are as follows: By parsing the AUT scan file through a parsing function, waveform data at different scanning positions is obtained. Since the waveform data is a training set of data types, the convolutional neural network is trained through the training set of data types, with short training time, low requirements for computer performance, and fast identification speed. In addition, when determining the training set, sample augmentation is performed on the defect types with fewer samples to obtain a first target training set, and the convolutional neural network is trained through the first target training set, reducing the impact of sample imbalance on the identification performance of the convolutional neural network and improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following further illustrates the present invention with reference to the drawings and embodiments.

[0021] Figure 1 It is a schematic flowchart of an intelligent identification method for pipeline weld defects according to an embodiment of the present invention;

[0022] Figure 2 It is a schematic diagram of unfused defect in the filling area;

[0023] Figure 3 It is a flowchart for obtaining waveform data;

[0024] Figure 4 Schematic diagram of waveform data;

[0025] Figure 5 Schematic diagram of the process of SMOTE algorithm;

[0026] Figure 6 Flow chart of the training of convolutional neural network;

[0027] Figure 7 Flow chart of the generation of AUT scan report;

[0028] Figure 8 Schematic diagram of the process of an intelligent pipeline weld defect identification method according to another embodiment of the present invention;

[0029] Figure 9 Schematic diagram of the influence of iteration rounds, regularization, and learning rate on network performance;

[0030] Figure 10 Schematic diagram of the influence of optimization algorithms and batch normalization on network performance;

[0031] Figure 11 Curve of the convolutional neural network training process;

[0032] Figure 12 Confusion matrix of the test set;

[0033] Figure 13 Schematic diagram of the structure of an intelligent pipeline weld defect identification system according to an embodiment of the present invention. Detailed implementation manners

[0034] The following embodiments are further explanations and supplements to the present invention and do not constitute any limitation to the present invention.

[0035] The following describes an intelligent pipeline weld defect identification method, system, computing device, and medium according to an embodiment of the present invention with reference to the accompanying drawings.

[0036] As Figure 1 shown, the present invention provides an intelligent pipeline weld defect identification method, including:

[0037] S1. Use an analytical function to parse the AUT scan file to determine the waveform data obtained by scanning the pipeline in different scanning positions in the scanning direction.

[0038] The analytical function of the AUT scan file is a conventional existing technology and will not be elaborated here.

[0039] S2. Mark each waveform data according to the defect type to determine the training set.

[0040] The types of defects include lack of fusion defects, lack of penetration defects, crack defects, porosity defects, etc. Taking the lack of fusion defect in the filling area as an example, the typical AUT pattern of the lack of fusion defect in the filling area is as Figure 2 shown. This defect shows unilaterally on the FU or FD channels (F1U - F7U and F1D - F7D) of the strip chart. The main peak may appear in the F1U channel and shows as a linear image in the volume channel and the TOFD channel. Therefore, this defect is a lack of fusion defect in the filling area. After the defect is marked, it is used as a label of the waveform data and input into the neural network for training.

[0041] S3. Statistically count the number of samples corresponding to each type of defect in the training set, and perform data augmentation on the number of samples of the defect types with the number of samples less than the preset value to determine the first target training set.

[0042] S4. Input the first target training set into the convolutional neural network for training to determine the recognition network.

[0043] S5. Identify the pipeline weld defects according to the recognition network.

[0044] In this embodiment, the AUT scan file is parsed by an analytical function to obtain waveform data at different scanning positions. Since the waveform data is a training set of data types, the convolutional neural network is trained with the training set of data types, which has a short training time, low requirements for computer performance, and fast recognition speed. In addition, when determining the training set, the samples of the defect types with fewer samples are augmented to obtain the first target training set, and the convolutional neural network is trained with the first target training set, which reduces the impact of sample imbalance on the recognition performance of the convolutional neural network and improves the classification accuracy.

[0045] Optionally, use an analytical function to parse the AUT scan file to determine the waveform data obtained by scanning the pipeline at different scanning positions in the scanning direction, including:

[0046] Use an analytical function to parse the AUT scan file to determine the scan data at each preset position within the time valve range sequentially extracted by the pipeline at different scanning positions in the scanning direction;

[0047] For each scanning position, the respective scan data obtained in the scanning direction are sequentially concatenated to determine the waveform data corresponding to the scanning position.

[0048] Illustrate by way of example, such as Figure 3As shown in the figure, in this embodiment, the preset positions within the time valve range include an upstream strip chart, an upstream volume channel, a downstream volume channel, a downstream strip chart, a coupling channel, and a TOFD channel. The scanning positions include the 0 mm scanning position, 1 mm scanning position... N mm scanning position on the pipeline. After sequentially scanning each preset position within the time valve range in the scanning direction, upstream strip chart data, upstream volume channel data, downstream volume channel data, downstream strip chart data, coupling channel data, and TOFD channel data can be obtained in sequence. Finally, by connecting the data corresponding to each scanning position, the waveform data corresponding to each scanning position can be obtained, as Figure 4 shown.

[0049] Optionally, count the number of samples corresponding to each defect type in the training set, and perform data augmentation on the number of samples of the defect types with the number of samples less than the preset value to determine the first target training set, including:

[0050] Count the number of samples corresponding to each defect type in the training set, and perform data augmentation on the number of samples of the defect types with the number of samples less than the preset value to determine the first target training set, including:

[0051] For each defect type with the number of samples less than the preset value, calculate the Euclidean distance from each sample to each other sample except itself;

[0052] For each sample, after sorting the Euclidean distances from the sample in ascending order, the first preset number of other samples are used as the first target samples;

[0053] For each sample, select multiple second target samples from each of the first target samples through a preset sampling magnification;

[0054] For each sample, determine the new samples corresponding to the sample according to the sample and the corresponding second target samples;

[0055] Use each sample and each new sample as the first target training set.

[0056] In this embodiment, the sample augmentation uses the SMOTE (Synthetic Minority Oversampling Technique) algorithm, which is an improvement on the random sampling method and is used to solve the problem of data class imbalance. It is based on the k nearest neighbor sample points of each sample point, randomly selects N neighboring points for difference multiplication by a threshold in the range of [0, 1] to achieve the purpose of synthesizing data. The core of this algorithm is that points adjacent in the feature space have similar features, and the automatic augmentation of samples can be completed through the training model. Specifically as follows:

[0057] As Figure 5As shown, the waveform data to be expanded is divided into a training set and a test set. The training set is input into the network model for training, and the Euclidean distance from each sample x in the minority class samples to other samples is calculated. Then, the K nearest neighbor points (the closest Euclidean distance) of sample x are taken. The sampling magnification is set to N, and N points are selected from the K nearest neighbor sample points x of each sample x. New samples are generated according to the formula to obtain a new training set. Then, the network model parameters are set according to the new training set and trained to obtain the required network model. Finally, the required network model is tested with the test set, the classification result is output, and it is checked whether the classification result meets the preset requirements, and then the training can be completed. i to other samples, and then the K nearest neighbor points (the closest Euclidean distance) of sample x are taken. i Set the sampling magnification to N, and from each sample x i of the K nearest neighbor sample points x ij select N points, generate new samples according to the formula to obtain a new training set, then set the network model parameters according to the new training set and train to obtain the required network model. Finally, test the required network model with the test set, output the classification result, and check whether the classification result meets the preset requirements to complete the training.

[0058] Optionally, for each sample, determine the new sample corresponding to the sample according to the sample and each second target sample. The formula is as follows:

[0059] x new = x i + rand(0, 1) × (x ij - x i );

[0060] where x new represents the new sample corresponding to the i-th sample, x i represents the i-th sample, and x ij represents the j-th second target sample corresponding to the i-th sample.

[0061] Optionally, as Figure 6 shown, input the first target training set into the convolutional neural network for training to determine the recognition network, including:

[0062] S11. Input the first target training set;

[0063] S12. Divide the first target training set into a second target training set and a test set;

[0064] S13. Determine the size of the input layer of the convolutional neural network according to the data dimension (the dimension of the waveform data in the second target training set);

[0065] S14. Determine the number of convolutional layers and the size of the convolutional kernels;

[0066] S15. Determine the network parameters, including the number of iteration rounds, regularization value, learning rate, and optimization algorithm, etc.;

[0067] S16. Input the second target training set into the convolutional neural network, and train and test the convolutional neural network by modifying the number of convolutional layers, the size of the convolutional kernels, and the initial network parameters until the recognition accuracy of the convolutional neural network reaches the preset defect recognition rate.

[0068] S17. When the recognition accuracy reaches the preset defect recognition rate, the number of layers of the convolutional layer, the size of the convolutional kernel, and the network parameters corresponding thereto are respectively used as the target number of layers, the target size, and the target network parameters. Then, the convolutional neural network is trained on the second target training set using the target number of layers, the target size, and the target network parameters to determine the target network model, and the target network model is used to test the defect types of the test set:

[0069] S18. Output the classification result to obtain the recognition model.

[0070] In this embodiment, the convolutional neural network adopted is used for the training of the recognition network. The basic principle of the convolutional neural network is to use convolutional operations to extract and transform matrix data to achieve tasks such as classification, detection, and recognition. The convolutional layer is one of its core components. It performs convolutional operations on the input data through a group of convolutional kernels to obtain a feature map, thereby extracting the important features of the input data. The convolutional kernel is a matrix and can be regarded as a feature extractor. It slides on the input data, performs convolutional operations on each local area with the convolutional kernel, and obtains an output value. This output value is the feature representation of the local area. The convolutional operation formula is as follows:

[0071] S(i, j, k) = (K * I)(i, j, k) = ∑I(i, j, k) * K(i - m, j - n, k - p)

[0072] In the formula, i, j, and k are three-dimensional vector indices; m, n, and p are the sizes of the convolutional kernel; I(i, j, k) is the input three-dimensional vector; K(i, j, k) is the three-dimensional convolutional kernel; and S(i, j, k) is the three-dimensional feature.

[0073] The convolutional kernel slides on each dimension of the input tensor. The distance of this movement is called the stride, and the size of the stride will affect the size of the output tensor of the convolutional layer. For the size of the output tensor in a certain dimension, it is jointly determined by the size of the input tensor, the stride, and the size of the convolutional kernel in that dimension. The formula is as follows:

[0074]

[0075] Tout represents the size of the output tensor, Tin represents the size of the input tensor, K represents the size of the convolutional kernel, and stride represents the stride.

[0076] In addition to the convolutional layer, a convolutional neural network also includes a pooling layer. The commonly used pooling methods are max pooling and average pooling. The pooling layer can help the convolutional neural network perform more effective feature extraction on the input data, and at the same time can also improve the invariance of the model to the input data. The convolutional neural network includes a fully connected layer and an activation function, which are used for tasks such as classifying or regressing features. The role of the fully connected layer is to convert the feature map into a one-dimensional vector, and then perform tasks such as classification or regression. The activation function is used to introduce non-linearity and increase the expressive power of the model.

[0077] In addition, in this embodiment, regularization is introduced into the network parameters to reduce network overfitting, and batch normalization and learning rate decay strategies are used to accelerate the network convergence speed and improve the defect recognition accuracy.

[0078] Optionally, according to the recognition network, the pipeline weld defects are recognized, including:

[0079] Use an analytical function to parse the AUT scan file to be detected, and obtain the waveform data to be detected obtained by scanning the pipeline at different scanning positions in the scanning direction;

[0080] Input each waveform data to be detected into the recognition network to determine the predicted defect type corresponding to each scanning position;

[0081] Statistically screen each scanning position corresponding to each predicted defect type to determine the defect length, defect start point, and defect end point corresponding to each predicted defect type;

[0082] According to the defect length, defect start point, and defect end point corresponding to each predicted defect type, the pipeline weld defects are recognized.

[0083] In this embodiment, through the output result of the recognition model, different defect types can be classified, and the positions where each defect type appears can be statistically counted respectively, so as to identify the defect length, defect start point, and defect end point corresponding to the same type of defect, and then recognize the pipeline weld.

[0084] Optionally, according to the defect length, defect start point, and defect end point corresponding to each predicted defect type, the recognition of pipeline weld defects further includes:

[0085] Obtain the calibration drawing of the AUT scan file to be detected;

[0086] According to the calibration drawing and the defect quantification calculation criterion, determine the defect main partition corresponding to each predicted defect type;

[0087] According to the amplitudes of each channel in the defect main partition corresponding to each predicted defect type, determine the defect depth and defect height corresponding to each predicted defect type;

[0088] Determine the AUT inspection report according to the defect depth, defect height, defect length, defect starting point, and defect ending point corresponding to each predicted defect.

[0089] Identify the pipeline weld defects according to the defect length, defect starting point, and defect ending point corresponding to each predicted defect type, including:

[0090] Identify the pipeline weld defects according to the AUT inspection report.

[0091] In this embodiment, the method for obtaining the defect depth is as follows:

[0092] Each scanning position in the AUT scanning file to be detected corresponds to multiple channels, such as volume channel, coupling channel data, and TOFD channel data, etc. After counting the positions corresponding to each predicted defect type, the defect interval corresponding to each predicted defect type can be obtained. According to the calibration diagram and the defect quantification calculation criterion, calculate the amplitude of each channel within the defect interval corresponding to each predicted defect type to obtain the maximum amplitude B m As the main defect partition corresponding to this predicted defect type, its B m On both sides are adjacent partitions, and the amplitudes are B p1 and B p2 ; Take the distance between the main defect partition and the CAP channel as the defect depth.

[0093] In this embodiment, the method for obtaining the defect height is as follows:

[0094] After determining the main defect partition according to the above method, determine the strip chart of the main defect partition corresponding to each predicted defect type. The height of the strip chart is H, then determine the amplitude of each channel in the strip chart, and further determine the maximum amplitude B ms and the maximum amplitude B of the adjacent partition ps1 and B ps2 , then the defect height H1 is:

[0095] H 1 = HB m + H(B p1 - B ps1 B m / B ms ) + H(B p2 - B ps2 B m / B ms ).

[0096] Optionally, the process of identifying the pipeline weld by another embodiment is described as follows:

[0097] As Figure 7As shown in the figure, obtain the weld file to be recognized, input it into the neural network (recognition model), obtain the defect types at each weld position, determine the position intervals of each defect type, and determine the defect positions and lengths. At the same time, calculate the amplitudes of each channel at each position, and screen out the channel amplitudes corresponding to the position intervals to obtain the maximum value of each channel in the defect (the largest amplitude). Screen and extract the channels with amplitudes (the largest amplitudes) greater than 40% of the total channels. If the screened channels are continuous, there is only one main defect partition. If the screened channels are discontinuous, there are multiple main defect partitions. Finally, determine the defect height and depth of the main defect partition, and obtain the AUT scan report based on the defect position and length and the defect height and depth.

[0098] Optionally, another embodiment is used to illustrate a method for intelligent recognition of pipeline weld defects, which is as follows:

[0099] As Figure 8 shown in the figure, it is divided into a network training process, a network prediction process, and a defect position and size calculation process. Among them, obtain the AUT evaluation drawing software and the parsing function, parse the scan drawing of the AUT evaluation drawing software through the parsing function to obtain the waveform data corresponding to each defect position, and obtain the defect database, which contains, for example, the waveform data corresponding to each scan position in 11 - 33mm with marks of lack of fusion, incomplete penetration, slag inclusion, porosity, no defect, and crack. Similarly, add the same marks to the waveform data corresponding to each scan position in 245 - 249mm and 370 - 407mm. Then, input the data in the defect database into the convolutional neural network. The data passes through the input layer (1×2893×1), convolutional layer 1 (1×575×16), max pooling layer 1 (1×287×16), convolutional layer 1 (1×141×32), max pooling layer 1 (1×70×32), fully connected layer (1×2240×1), and output layer (1×1×6) to obtain the classification result.

[0100] After counting the positions corresponding to each classification result for the classification result, for example, no defect includes 0mm, 1mmN - 1mm, and Nmm, lack of fusion includes 182mm, 183mm, 203mm, and 204mm, etc., determine the defect start point, defect end point, and defect length. Finally, it is also necessary to obtain the calibration drawing file of the test block corresponding to the AUT scan file, and obtain the data matrix of each channel through the parsing function, and extract the maximum amplitude of each channel, calculate the defect depth and defect height, and generate the AUT detection report in combination with the obtained defect start point, defect end point, and defect length.

[0101] Optionally, a specific simulation experiment corresponding to this embodiment is as follows:

[0102] This algorithm is written in MATLAB R2021b version and requires the use of the Deep Learning Toolbox included in the software.

[0103] The training data uses 14 existing weld inspection files. According to the inspection reports of the AUT atlas, the defect data in each file is extracted. There are four types of defects in total, namely lack of fusion, incomplete penetration, crack and porosity, and at the same time, some defect-free data is extracted. It should be noted that the defect-free data is not all composed of low-amplitude data, but a part of the interference data and defect data showing the root contour are added to restore the discrimination under real conditions to the greatest extent. The quantities of various types of data are shown in Table 3.

[0104] In the original data, the sample quantities of incomplete penetration, porosity and crack are small. Therefore, the SMOTE algorithm is used to expand the samples of incomplete penetration, porosity and crack to reduce the imbalance problem of various types of defect data. The data before expansion is shown in Table 1, and the data after expansion is shown in Table 2, as follows:

[0105] Table 1

[0106]

[0107] Table 2

[0108]

[0109] After the data expansion is completed, labels are added to the defects. Labels 1 to 5 are used to represent defect-free, lack of fusion, incomplete penetration, porosity and crack respectively. Each type of sample is divided into a training set, a validation set and a test set according to the ratio of 8:1:1. The size of the input layer of the network is determined according to the waveform data dimension, the number of convolutional layers and the convolutional kernel size of the network are determined, and the network is trained and tested by modifying the number of layers and the convolutional kernel size to obtain the highest defect recognition rate. The convolutional layer number and convolutional kernel size under the highest recognition rate are selected to determine the network parameters, such as the number of iterations, regularization, learning rate, optimization algorithm and other parameters. The influence of different hyperparameters on the network performance is as Figure 9 and Figure 10 shown, where, Figure 9 includes the recognition rates obtained with the passage of running time with different numbers of iterations, the recognition rates obtained with the passage of running time with different L2 regularization values, and the recognition rates obtained with the passage of time with different learning rates; Figure 10 includes the recognition rates obtained with the passage of running time using the Adam algorithm, RMSprop algorithm and SGDM algorithm as optimization algorithms, and the recognition rates obtained with the passage of running time with batch normalization added and without batch normalization added.

[0110] By studying the values of different influencing parameters, the number of iteration rounds is finally determined to be 20, the L2 regularization value is 0.001, and the learning rate is 0.0001. According to the analysis of network influencing factors, the SGDM optimization algorithm is determined to be used, and a batch normalization layer is added to the network structure.

[0111] After determining the network architecture and hyperparameters, the training set data is trained using the optimal number of convolutional layers, convolutional kernel sizes, and network parameters. Among them, the optimal number of convolutional layers and convolutional kernel sizes are shown in Table 3:

[0112] Table 3

[0113]

[0114] The network parameters are shown in Table 4:

[0115] Table 4

[0116]

[0117] Finally, the network training process curve is as Figure 11 shown. After 10,000 iteration times, the accuracy of the training set stabilizes at 95%, and the loss value remains at 0.1;

[0118] The trained network model is used to predict the defect types of the test set, and the confusion matrix of the test set is as Figure 12 shown. The accuracy of the test set is obtained as 93.38%, and the training time is 292 seconds.

[0119] For no defect, 107 sample data are misjudged as lack of fusion; for incomplete penetration, 3 samples are misjudged as no defect; for crack, 1 sample is misjudged as no defect and 1 sample is misjudged as lack of fusion; while for lack of fusion and porosity, the recognition accuracy of the network reaches 100%.

[0120] The network's recognition of no-defect data is relatively poor. 107 are misjudged as lack of fusion. The reason for this result is that some data in no defect has a high similarity to the defect data in lack of fusion, resulting in the network being unable to accurately identify this part of the data.

[0121] Another set of weld inspection data files that have not participated in training is used as the file to be recognized. The waveform data obtained at each inspection position is sequentially input into the trained network for prediction to obtain a defect prediction label set. Taking lack of fusion as an example, the start and end points of each defect interval are determined, and the defect length, defect height, and defect depth are calculated. Finally, the detection report is output as shown in Table 5.

[0122] Table 5

[0123]

[0124]

[0125] After organizing into the AUT inspection report, Table 6 is obtained. In the inspection report given in Table 6, the number of defects with a defect height exceeding 1 mm is 38. In the neural network prediction report of Table 5, the number of defects with a defect height exceeding 1 mm is 35. For defect 11, its defect interval is the same as that of defect 12, and there is an interval intersection. This algorithm only detects defect 12; for defects 14 and 18, this algorithm does not recognize these two defect intervals; for defect 22, its defect interval is the same as that of defect 23, and this algorithm only detects defect 23; for defect 29, its defect interval is the same as that of defect 28, and this algorithm only detects defect 28; for defect 34, this algorithm can detect it, but the defect on the right is not recognized.

[0126] Table 6

[0127]

[0128]

[0129] From the perspective of the defect distribution interval, two intervals in the defect distribution interval given in the inspection report are not predicted by the neural network, and the defect detection rate is 94.7%; from the perspective of the number of defects (defect numbers), the defects with defect numbers 14 and 18 are missed, and the defects with defect numbers 11, 22, and 29 are not recognized because they are in the same position as other defects, and the defect detection rate is 84.2%.

[0130] As Figure 13 shown, the present invention provides an intelligent identification system for pipeline weld defects, including:

[0131] A waveform data determination module, configured to parse the AUT scan file using an analytical function to determine the waveform data obtained by scanning the pipeline in different scanning positions according to the scanning direction;

[0132] A training set determination module, configured to mark each waveform data according to the defect type to determine a training set;

[0133] A first target training set determination module, configured to count the number of samples corresponding to each defect type in the training set, and perform data augmentation on the number of samples of the defect type with the number of samples less than a preset value to determine a first target training set;

[0134] An identification network determination module, configured to input the first target training set into a convolutional neural network for training to determine an identification network;

[0135] An identification module, configured to identify pipeline weld defects according to the identification network.

[0136] Optionally, the waveform data determination module is specifically configured to:

[0137] Use a parsing function to parse the AUT scan file, and determine the scan data at each preset position within the time valve range sequentially extracted in the scan direction at different scan positions of the pipeline;

[0138] For each scan position, splice the respective scan data obtained in the scan direction in sequence to determine the waveform data corresponding to the scan position.

[0139] Optionally, the first target training set determination module is specifically configured to:

[0140] Count the number of samples corresponding to each defect type in the training set, and perform data augmentation on the number of samples of the defect types with the number of samples less than the preset value to determine the first target training set, including:

[0141] For each defect type with the number of samples less than the preset value, calculate the Euclidean distance from each sample to each other sample except itself;

[0142] For each sample, after sorting the Euclidean distances from the sample in ascending order, take the first preset number of other samples as the first target samples;

[0143] For each sample, select multiple second target samples from each of the first target samples through a preset sampling magnification;

[0144] For each sample, determine the new sample corresponding to the sample according to the sample and the respective second target samples;

[0145] Use each of the samples and each of the new samples as the first target training set.

[0146] Optionally, the first target training set determination module is specifically configured to:

[0147] For each sample, determine the new sample corresponding to the sample according to the sample and the respective second target samples, and the formula is as follows:

[0148] x new = x i + rand(0,1) × (x ij - x i ) ;

[0149] Wherein, x new represents the new sample corresponding to the i-th sample, x i represents the i-th sample, and x ij represents the j-th second target sample corresponding to the i-th sample.

[0150] Optionally, the recognition network determination module is specifically configured to:

[0151] Divide the first target training set into a second target training set and a test set;

[0152] Determine the size of the input layer of the convolutional neural network according to the dimension of the waveform data in the second target training set;

[0153] Obtain the number of convolutional layers, the size of the convolutional kernels, and the network parameters, where the network parameters include the number of iterations, the regularization value, the learning rate, and the optimization algorithm;

[0154] Input the second target training set into the convolutional neural network, and train and test the convolutional neural network by modifying the number of convolutional layers, the size of the convolutional kernels, and the initial network parameters until the recognition accuracy of the convolutional neural network reaches the preset defect recognition rate;

[0155] Take the number of convolutional layers, the size of the convolutional kernels, and the network parameters corresponding to when the recognition accuracy reaches the preset defect recognition rate as the target number of layers, the target size, and the target network parameters respectively;

[0156] Train the second target training set using the convolutional neural network with the target number of layers, the target size, and the target network parameters to determine the target network model;

[0157] Use the target network model to test the defect types of the test set, output the classification results, and obtain the recognition model.

[0158] Optionally, the recognition module is specifically used for:

[0159] Use the parsing function to parse the AUT scan file to be detected, and obtain the waveform data to be detected obtained by scanning the pipeline in different scanning positions in the scanning direction;

[0160] Input each waveform data to be detected into the recognition network to determine the predicted defect type corresponding to each scanning position;

[0161] Statistically screen each scanning position corresponding to each predicted defect type to determine the defect length, the defect starting point, and the defect ending point corresponding to each predicted defect type;

[0162] Identify the pipeline weld defects according to the defect length, the defect starting point, and the defect ending point corresponding to each predicted defect type.

[0163] Optionally, the system further includes an AUT detection report determination module, specifically used for:

[0164] Obtain the calibration drawing of the reference block corresponding to the AUT scan file to be detected;

[0165] Determine the main defect partition corresponding to each predicted defect type according to the calibration drawing and the defect quantification calculation criterion;

[0166] Determine the defect depth and defect height corresponding to each predicted defect type according to the amplitudes of each channel in the main defect partition corresponding to each predicted defect type;

[0167] Determine the AUT inspection report according to the defect depth, defect height, defect length, defect start point and defect end point corresponding to each predicted defect;

[0168] Then the recognition module is specifically used for:

[0169] Identify the pipeline weld defects according to the AUT inspection report.

[0170] A computing device according to an embodiment of the present invention includes a memory, a processor, and a program stored on the memory and running on the processor. When the processor executes the program, it implements some or all of the steps of the above-mentioned intelligent identification method for pipeline weld defects.

[0171] Among them, the computing device can be a computer. Correspondingly, its program is computer software. And the above parameters and steps in a computing device of the present invention can refer to the parameters and steps in the embodiments of the intelligent identification method for pipeline weld defects in the foregoing text, which will not be elaborated here.

[0172] Those skilled in the art know that the present invention can be implemented as a system, a method, or a computer program product. Therefore, the present disclosure can be specifically implemented in the following forms: it can be completely hardware, can be completely software (including firmware, resident software, microcode, etc.), or can be a combination of hardware and software, which is generally referred to as "circuit", "module" or "system" in this article. In addition, in some embodiments, the present invention can also be implemented in the form of a computer program product in one or more computer-readable media, and the computer-readable media contains computer-readable program code. The computer-readable storage medium can be, for example, but not limited to - an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above.

[0173] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0174] Although embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. An intelligent identification method for pipeline weld defects, characterized in that, it includes: Using an analytical function to parse the AUT scan file to determine the waveform data obtained by scanning the pipeline at different scanning positions in the scanning direction; Marking each of the waveform data according to the defect type to determine a training set; Counting the number of samples corresponding to each defect type in the training set, and performing data augmentation on the number of samples of the defect type with the number of samples less than a preset value to determine a first target training set; Inputting the first target training set into a convolutional neural network for training to determine an identification network; Identifying pipeline weld defects according to the identification network.

2. The method according to claim 1, characterized in that, The using an analytical function to parse the AUT scan file to determine the waveform data obtained by scanning the pipeline at different scanning positions in the scanning direction includes: Using an analytical function to parse the AUT scan file to determine the scan data at each preset position within the time valve range sequentially extracted at different scanning positions of the pipeline in the scanning direction; For each of the scanning positions, splicing the respective scan data obtained in the scanning direction in sequence to determine the waveform data corresponding to the scanning position.

3. The method according to claim 1, characterized in that, The counting the number of samples corresponding to each defect type in the training set, and performing data augmentation on the number of samples of the defect type with the number of samples less than a preset value to determine a first target training set includes: For each defect type with the number of samples less than a preset value, calculating the Euclidean distance from each sample to each other sample except itself; For each sample, after sorting the Euclidean distances from the sample in ascending order, the first preset number of other samples are used as the first target samples; For each sample, selecting a plurality of second target samples from each of the first target samples through a preset sampling magnification; For each sample, determining the new samples corresponding to the sample according to the sample and the respective second target samples corresponding thereto; Using each of the samples and each of the new samples as the first target training set.

4. The method according to claim 3, characterized in that, For each sample, determining the new samples corresponding to the sample according to the sample and the respective second target samples corresponding thereto, the formula is as follows: x new = x i + rand(0,1) × (x ij - x i ); where, x new represents the new sample corresponding to the i-th sample, x i represents the i-th sample, x ij represents the j-th second target sample corresponding to the i-th sample.

5. The method according to claim 1, characterized in that, The inputting the first target training set into a convolutional neural network for training to determine an identification network includes: Dividing the first target training set into a second target training set and a test set; Determining the size of the input layer of the convolutional neural network according to the dimension of the waveform data in the second target training set; Obtaining the number of convolutional layers, the size of the convolutional kernels and network parameters, where the network parameters include the number of iteration rounds, the regularization value, the learning rate and the optimization algorithm; Input the second target training set into the convolutional neural network, and train and test the convolutional neural network by modifying the number of convolutional layers, the size of the convolutional kernels, and the initial network parameters until the recognition accuracy of the convolutional neural network reaches the preset defect recognition rate; Take the number of convolutional layers, the size of the convolutional kernels, and the network parameters corresponding to the situation where the recognition accuracy reaches the preset defect recognition rate as the target number of layers, the target size, and the target network parameters respectively; Train the second target training set with the convolutional neural network using the target number of layers, the target size, and the target network parameters to determine the target network model; Use the target network model to test the defect types of the test set, output the classification results, and obtain the recognition model.

6. The method according to any one of claims 1-5, characterized in that, The identification of pipeline weld defects according to the identification network includes: Use an analytical function to parse the AUT scan file to be detected, and obtain the waveform data to be detected obtained by scanning the pipeline at different scanning positions in the scanning direction; Input each piece of the waveform data to be detected into the recognition network to determine the predicted defect type corresponding to each scanning position; Statistically screen each scanning position corresponding to each predicted defect type to determine the defect length, defect start point, and defect end point corresponding to each predicted defect type; Identify pipeline weld defects according to the defect length, defect start point, and defect end point corresponding to each predicted defect type.

7. The method according to claim 6, characterized in that, further comprising: Obtain the calibration drawing of the test block corresponding to the AUT scan file to be detected; Determine the main defect area corresponding to each predicted defect type according to the calibration drawing and the defect quantification calculation criterion; Determine the defect depth and defect height corresponding to each predicted defect type according to the amplitudes of each channel in the main defect area corresponding to each predicted defect type; Determine the AUT detection report according to the defect depth, defect height, defect length, defect start point, and defect end point corresponding to each predicted defect; The identification of pipeline weld defects according to the defect length, defect start point, and defect end point corresponding to each predicted defect type includes: Identify pipeline weld defects according to the AUT detection report.

8. An intelligent pipeline weld defect identification system, characterized in that, comprising: A waveform data determination module, configured to use an analytical function to parse the AUT scan file and determine the waveform data obtained by scanning the pipeline at different scanning positions in the scanning direction; A training set determination module, configured to mark each piece of the waveform data according to the defect type and determine the training set; A first target training set determination module, configured to count the number of samples corresponding to each defect type in the training set and perform data augmentation on the samples of the defect types with the number of samples less than the preset value to determine the first target training set; An identification network determination module, configured to input the first target training set into the convolutional neural network for training to determine the identification network; An identification module, configured to identify pipeline welds according to the identification network.

9. A computing device, comprising a memory, a processor, and a program stored on the memory and running on the processor, characterized in that, when the processor executes the program, it implements the steps of an intelligent identification method for pipeline weld defects according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, the computer-readable storage medium stores instructions, and when the instructions run on a terminal device, the terminal device is caused to execute the steps of an intelligent identification method for pipeline weld defects according to any one of claims 1 to 7.

Images