Ultrasonic C scanning abnormal image classification method based on neural network
Through the ultrasonic C scan abnormal image classification method based on neural network, the ultrasonic C scan images of titanium alloy materials are classified and trained, which solves the problem of complex signals and low automation in the existing technology of titanium alloy materials defect detection, and achieves high efficiency, Accurate defect detection.
Patent Information
- Application Number
- CN202411503141.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2025-02-11
AI Technical Summary
The existing ultrasonic non-destructive detection methods have complex signals when detecting defects in titanium alloy materials, which lead to difficulty in judgment, low degree of automation, long detection time, low efficiency, and artificial errors.
The ultrasonic C scan abnormal image classification method based on neural network is adopted, and the ultrasonic C scan image is digitally preprocessed and classified and trained through the convolutional neural network model to achieve preliminary qualitative analysis of the abnormal signals inside titanium alloy materials.
It improves the accuracy and automation of defect detection of titanium alloy materials, reduces detection time and artificial errors, and can quickly and intuitively make preliminary judgments on abnormal signals inside the material.
Smart Images
Figure CN119417795A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of nondestructive material testing and relates to an ultrasonic C-scan abnormal image classification method based on a neural network. Background Art
[0002] With the development of material technology, more and more advanced materials are being put into use. Since defects such as inclusions, holes, and cracks may occur during the production process of titanium alloy materials, it is necessary to detect defects before titanium alloy materials are put into use. At present, ultrasonic non-destructive testing is one of the important methods for quality inspection of mechanical engineering materials. Its principle is to use the acoustic properties of the material and the changes in the internal structure of the material when the ultrasonic wave propagates in the material being tested to have a certain impact on the propagation of the ultrasonic wave. The main principle is that when the ultrasonic wave enters the object and encounters a defect, a part of the sound wave will be reflected, and the receiver analyzes the reflected wave to find the hidden internal defects. The commonly used ultrasonic non-destructive testing method is to obtain A-scan images and C-scan images through the ultrasonic pulse reflection method, relying on experienced professionals to first judge the nature of the defect in combination with the A-scan image and the C-scan image, and then determine the defect area information based on the nature of the defect.
[0003] However, due to the complexity of titanium alloy materials, the ultrasonic signal is complex, which makes it difficult for experienced professionals to determine the nature of defects by combining A-scan images and C-scan images. In addition, the degree of automation of defect detection is not high, the detection time is long, the efficiency is low, and certain human errors will be introduced.
[0004] In view of this, the present invention is proposed. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a neural network-based ultrasonic C-scan abnormal image classification method, which can more effectively perform preliminary qualitative analysis of abnormal signals inside materials through non-destructive testing methods.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] The neural network-based ultrasonic C-scan abnormal image classification method comprises the following steps:
[0008] Step 1, obtaining a C-scan image of a target to be inspected that contains defects;
[0009] Step 2: digitally preprocessing the acquired C-scan image containing defects;
[0010] Step 3: construct an ultrasonic C-scan abnormal image classification model based on the convolutional neural network model;
[0011] Step 4: training and optimizing the ultrasonic C-scan abnormal image classification model;
[0012] Step 5: Input the defective C-scan image into the trained and optimized ultrasonic C-scan abnormal image classification model to obtain the classification result of the ultrasonic C-scan abnormal image.
[0013] Specifically, step 1 includes:
[0014] Use ultrasonic water immersion automatic ultrasonic testing system to scan the preset natural defects (hole defects, cracks + inclusion defects formed by preset sponge titanium defects) and abnormal structures of titanium alloy;
[0015] After the material information and testing process are confirmed, the sensitivity calibration of the comparison test block is carried out;
[0016] Finally, the bar or forging billet is subjected to ultrasonic automated inspection: the bar automated inspection requires the circumferential scanning point to be 0.5mm and the axial scanning spacing to be 2mm; the forging billet automated inspection requires the scanning spacing in the X and Y directions to be ≥1mm, and a C-scan image containing defects is obtained.
[0017] Specifically, step 2 includes: importing different C-scan images containing defects into the data preprocessing model, converting the C-scan images containing defects into the size required for training, usually setting the size to 50*50, obtaining the image and label of the data, printing and verifying, and saving the generated images; classifying the types of images used for learning into three types: "noise", "holes" and "cracks + inclusions", and the number of samples of each type is greater than 200.
[0018] Specifically, the construction process of the ultrasonic C-scan abnormal image classification model in step 3 is as follows:
[0019] Establish a convolutional neural network model: first set the convolutional layer*2 and pooling layer*2 of the convolutional neural network, then build a fully connected layer, which uses 128 fully connected neurons, and finally establish a classification layer to complete the establishment of the ultrasonic C-scan abnormal image classification model.
[0020] Specifically, step 4, training and optimizing the ultrasonic C-scan abnormal image classification model, includes:
[0021] Step 4.1, input the classified three types of training sample images into the ultrasound C-scan abnormal image classification model for training;
[0022] Step 4.2, testing the ultrasonic C-scan abnormal image classification model again, and calculating the accuracy of the ultrasonic C-scan abnormal image classification model;
[0023] Step 4.3, repeating steps 4.1 and 4.2 until the accuracy of the ultrasound C-scan abnormal image classification model is higher than 98%;
[0024] In step 4.4, apply loss to define the loss function and apply the optimization method so that the rate of change of the loss function is less than 0.001, and then end the training.
[0025] Furthermore, the three training samples were read into the ultrasonic C-scan abnormal image classification model: the CNN model in Python was used for image classification training, with 200 to 500 training iterations and a learning rate of 0.002 to establish three types of model libraries: "noise", "holes" and "cracks + inclusions". The training data labels were image width, image height, batch size, and maximum capacity of the queue.
[0026] Compared with the prior art, the technical solution provided by the present invention includes the following beneficial effects: the present invention uses the ultrasonic amplitude obtained by the pulse reflection method to form a C-scan image, adopts an image classification method based on a convolutional neural network, and combines the python programming language to train different types of C-scan images, preliminarily determine the type of abnormal image, and then effectively distinguish "noise (generated by tissue reflection signals)", "hole" defects and "crack + inclusion" defects, which solves the current lack of accurate qualitative methods based on C-scan images in ultrasonic testing, and more effectively performs quality control and fault analysis on bars or forgings.
[0027] In addition, the method provided by the present invention can perform qualitative classification analysis on abnormal ultrasonic C-scan images of materials, has a clear process, strong operability, and fast and intuitive analysis results, which reduces the difficulty of qualitative analysis of abnormal signal images by ultrasonic testing. No special materials are required, and it can more effectively perform preliminary judgment on abnormal signals inside bars or forgings through non-destructive testing methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] The accompanying drawings are incorporated in and constitute a part of this specification and, together with the description, serve to explain the principles of the present invention.
[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0030] Figure 1 A flowchart of a neural network-based ultrasonic C-scan abnormal image classification method provided by the present invention;
[0031] Figure 2is an ultrasonic detection C-scan image of defect I in Example 1;
[0032] Figure 3 is a metallographic anatomical image of defect I in Example 1;
[0033] Figure 4 is an ultrasonic C-scan image of defect II in Example 2;
[0034] Figure 5 is the metallographic anatomical image of defect II in Example 2;
[0035] Figure 6 is an ultrasonic detection C-scan diagram of abnormal signal 3 in Example 3;
[0036] Figure 7 This is the metallographic anatomical image of abnormal signal 3 in Example 3. DETAILED DESCRIPTION
[0037] Here, exemplary embodiments will be described in detail, and the embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are only examples consistent with some aspects of the present invention described in detail in the appended claims.
[0038] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.
[0039] Explanation of relevant terms:
[0040] 1. Ultrasonic C-scan detection: Ultrasonic waves are often used in industry to detect defects below the surface of bars or forging materials. The reflected signal waves can be used to obtain a more intuitive defect image using ultrasonic imaging technology. Ultrasonic C-scan detection is an ultrasonic detection method that images a cross section parallel to the workpiece surface at a certain depth inside the workpiece. Using automated detection equipment, motion-controlled ultrasonic probes are scanned on the workpiece, and the ultrasonic echo signal intensity inside the workpiece is stored and analyzed. A digital image of the cross section of the internal defect of the workpiece is formed through different threshold ranges. The defect imaging is intuitive and the test accuracy is high.
[0041] 2. Neural network is an intelligent engineering system that uses mathematical methods to artificially simulate the structure and behavior of the human brain in the study of its organizational structure and operating mechanism. The basic component of a neural network is a neuron, and the commonly used mechanism is generally that multiple inputs correspond to one output. A neural network is understood as layers of interconnected neuron cells. The neurons in the input layer receive information from the outside world, and then transmit the information to the next layer of neurons through synapses, and finally obtain the information of the output layer after passing through several layers. The main functions of a neural network are learning and execution. Through learning, the neural network can be trained to have some useful information.
[0042] 3. Image classification technology is often applied to convolutional neural network technology (CNN). CNN is a deep feedforward neural network. Compared with other common neural networks, convolutional neural networks have some special layers, including fully connected layers, convolutional layers, pooling layers, and batch normalization layers. Among them, the convolutional layer is different from the fully connected layer. It retains the spatial features of the input image; the convolution kernel is applied to slide on the input image in sequence, and the sliding direction is left and right, up and down. The number of pixels moved is set by the step size. After completing the one-dimensional convolution calculation, continue to learn the two-dimensional convolution calculation.
[0043] When the neural network-based ultrasonic C-scan abnormal image classification method provided by the present invention is used for defect recognition, the specific implementation steps are as follows:
[0044] ①Material and tool preparation
[0045] 1) Ultrasonic water immersion automatic ultrasonic detection system;
[0046] 2) Natural defects detected by ultrasonic C-scan (hole defects, cracks + inclusion defects formed by pre-set sponge titanium defects) and abnormal structures of titanium alloys.
[0047] ②Method steps, see Figure 1
[0048] Step 1: Get scan data:
[0049] The ultrasonic water immersion automatic ultrasonic testing system is used to preset the natural defects and abnormal structures of titanium alloys;
[0050] After the material information and testing process are confirmed, the sensitivity calibration of the comparison test block is carried out;
[0051] Finally, the bar or forging billet is subjected to ultrasonic automated inspection: the bar automated inspection requires the circumferential scanning point to be 0.5mm and the axial scanning spacing to be 2mm; the forging billet automated inspection requires the scanning spacing in the X and Y directions to be ≥1mm, and a C-scan image containing defects is obtained;
[0052] Step 2: Digital preprocessing of image classification model: Import different C-scan images with defects into the data preprocessing model, and convert the C-scan images with defects into the size required for training; classify the images used for learning into three types: "noise", "holes" and "cracks + inclusions", and use them as training sample images;
[0053] Step 3: construct an ultrasonic C-scan abnormal image classification model based on the convolutional neural network model;
[0054] Step 4: Read the three training sample images into the training model: Apply the CNN model in Python to perform image classification training. After the training, the model can accurately identify three types of anomalies;
[0055] Step 5: Load the C-scan image to be predicted and obtain the classification result of the abnormal ultrasound C-scan image.
[0056] Example 1
[0057] This embodiment provides a neural network-based ultrasonic C-scan abnormal image classification method for identifying defects in Gr.5 titanium alloy bars. It is necessary to first pre-set titanium sponge defects to form holes or cracks for ultrasonic testing. The specific operations of the method are as follows:
[0058] Step 1: For Gr.5 titanium alloy bars, C-scan detection is performed at a sensitivity of Φ0.8mm flat-bottom hole. The circumferential scanning point of the bar is 0.5mm, and the axial scanning interval is 2mm. After the detection, the image of defect I is found in the C-scan image (see Figure 2 ), the maximum equivalent is Φ0.8-1dB;
[0059] Step 2: Import more than 600 C-scan images of three types, namely "noise", "hole" and "crack + inclusion", into the data preprocessing model, and convert the more than 600 C-scan images into the size required for training; classify the images used for learning (i.e., training samples) into three types, namely "noise", "hole" and "crack + inclusion";
[0060] Step 3: Based on Python, a convolutional neural network model (CNN model) is used to build an ultrasonic C-scan abnormal image classification model;
[0061] Step 4: Read the three training samples into the training model: Apply the CNN model in Python to perform image classification training. After the training, the three types of anomalies can be accurately identified;
[0062] Step 5: After the ultrasonic C-scan image is detected, an image of defect I is found, and the image is loaded into the aforementioned ultrasonic C-scan abnormal image classification model to obtain a classification result of the ultrasonic C-scan abnormal image, and the abnormal image is predicted to be a hole defect;
[0063] Step 6: Perform anatomical analysis on the located defects to obtain the specific morphology of the defects (see Figure 3 ), the defect is actually a hole.
[0064] Example 2
[0065] On the basis of Example 1, this example provides another neural network-based ultrasonic C-scan abnormal image classification method for identifying forging billet hole defects, and its operating principle is the same as that of the example:
[0066] For the Gr.5 forging billet, C-scan detection was performed at the sensitivity of Φ1.2mm flat bottom hole, and the scanning interval between X and Y directions was 1.2mm. After the detection, the image of defect II was found in the C-scan image (see Figure 4 The image was loaded into the trained ultrasound C-scan abnormal image classification model based on convolutional neural network. The anatomical result showed that the defect was a hole (see Figure 5 ).
[0067] Example 3
[0068] On the basis of Example 1, this example provides another neural network-based ultrasonic C-scan abnormal image classification method for identifying forging blank tissue reflections, and its operating principle is the same as that of the example:
[0069] For Gr.5 bars, C-scan detection was performed at a sensitivity of Φ0.8mm flat-bottom hole. The circumferential scanning point of the bars was 0.5mm, and the axial scanning interval was 2mm. After the detection, the C-scan image of abnormal signal 3 was found (see Figure 6 The image is loaded into the trained ultrasound C-scan abnormal image classification model based on convolutional neural network, and the abnormal image is classified as a noise signal. The anatomical result shows that the abnormal signal image is a noise signal caused by tissue orientation (see Figure 7 ).
[0070] In summary, the neural network-based ultrasonic C-scan abnormal image classification method provided by the present invention adopts a neural network-based image classification method, combined with the python programming language, to train different types of C-scan images, and preliminarily determine the type of abnormal image; it reduces the difficulty of qualitative analysis of abnormal information of ultrasonic detection of bars and forgings, and does not require special materials. It can more effectively and fully perform preliminary qualitative analysis of internal abnormal signals through non-destructive testing methods, especially it can quickly perform preliminary qualitative analysis of abnormal signals without destroying Gr.5 materials, providing a fast and reliable qualitative method for Gr.5 ultrasonic detection technology.
[0071] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention.
[0072] It should be understood that the present invention is not limited to what has been described above and that various modifications and changes may be made without departing from its scope. The scope of the present invention is limited only by the appended claims.
Claims
1. A neural network-based ultrasonic C-scan abnormal image classification method, characterized in that: The steps include: Step 1, obtaining a C-scan image of a target to be inspected that contains defects; Step 2: digitally preprocess the acquired C-scan image containing defects; Step 3: construct an ultrasonic C-scan abnormal image classification model based on the convolutional neural network model; Step 4: training and optimizing the ultrasonic C-scan abnormal image classification model; Step 5: Input the defective C-scan image into the trained and optimized ultrasonic C-scan abnormal image classification model to obtain the classification result of the ultrasonic C-scan abnormal image.
2. The method for classifying abnormal ultrasonic C-scan images based on a neural network according to claim 1, characterized in that: The target to be detected in step 1 is a bar or a forging billet.
3. The method for classifying abnormal ultrasonic C-scan images based on a neural network according to claim 2, characterized in that: In the step 1, an ultrasonic water immersion automatic ultrasonic testing system is used to perform a C-scan image test.
4. The method for classifying abnormal ultrasonic C-scan images based on a neural network according to claim 3, characterized in that: When performing C-scan image inspection: the bar automated inspection requires the circumferential scanning point to be 0.5mm and the axial scanning spacing to be 2mm; the forging automated inspection requires the X and Y direction scanning spacing to be ≥1mm.
5. The method for classifying abnormal ultrasonic C-scan images based on a neural network according to claim 1, characterized in that: The step 2 is specifically as follows: Import the C-scan image containing defects into the data preprocessing model, and convert the C-scan image containing defects into the size required for training; The types of images used for learning are divided into three types: "noise", "holes" and "cracks + inclusions".
6. The method for classifying abnormal ultrasonic C-scan images based on a neural network according to claim 1, characterized in that: The process of building the convolutional neural network model in step 3 is: first set the convolution layer and pooling layer of the convolutional neural network, then build the fully connected layer, and finally establish the classification layer; wherein the fully connected layer uses 128 fully connected neurons.
7. The method for classifying abnormal ultrasonic C-scan images based on a neural network according to claim 1, characterized in that: The step 4 specifically includes: Step 4.1, input the classified three types of training sample images into the ultrasound C-scan abnormal image classification model for training; Step 4.2, testing the ultrasonic C-scan abnormal image classification model again, and calculating the accuracy of the ultrasonic C-scan abnormal image classification model; Step 4.3, repeating steps 4.1 and 4.2 until the accuracy of the ultrasound C-scan abnormal image classification model is higher than 98%; In step 4.4, apply loss to define the loss function and apply the optimization method so that the rate of change of the loss function is less than 0.001, and then end the training.
8. The neural network-based ultrasonic C-scan abnormal image classification method according to claim 7, characterized in that: The number of training iterations in step 4.3 is 200 to 500, and the learning rate is 0.
002.
9. The neural network-based ultrasonic C-scan abnormal image classification method according to claim 7, characterized in that: The training data labels in step 4.1 are image width, image height, batch size and maximum capacity of the queue.