Laser ultrasonic guided pixel-level imaging method and device, and storage medium

Abstract

The invention relates to a laser ultrasound guided pixel-level imaging method and device, and a storage medium, ultrasound is applied to a to-be-detected target, out-of-plane displacement data is obtained based on laser scanning, and after normalization and one-hot coding are performed, a pre-trained 1D-CNN model is utilized to obtain a pixel-level imaging result, and the training process of the model comprises the following steps: aiming at an experimental sample, carrying out one-hot coding on the obtained out-of-plane displacement data, and carrying out one-hot coding on the obtained out-of-plane displacement data; the method comprises the following steps: setting a plurality of sampling points at large intervals, carrying out global rough scanning, setting a plurality of sampling points at small intervals according to a pre-obtained region where each defect is located on an experimental sample, carrying out local fine scanning, and obtaining out-of-plane displacement data comprising defect position information after global rough scanning-local fine scanning; performing normalization and one-hot coding processing on the out-of-plane displacement data; the out-of-plane displacement data after sample balancing is divided into a plurality of defect types, labeling is completed, a data set is obtained, and the defect types are set according to the depth of defects; and training the model by using the data set.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and in particular to a laser-ultrasound-guided pixel-level imaging method, device, and storage medium. Background Technology

[0002] Laser ultrasonic testing is a non-destructive testing technique based on the interaction between laser and ultrasonic waves. It combines the advantages of both technologies, using pulsed lasers to excite ultrasonic waves on or within a material and detecting changes during their propagation. This allows for high-precision assessment of internal damage, defects, or stress states within the material. This technology offers non-contact, high-resolution, and real-time detection capabilities, making it widely used in the aerospace and industrial manufacturing sectors.

[0003] With the advancement of technology, laser ultrasonic damage detection technology is also constantly developing. On the one hand, continuous innovation in laser and ultrasonic technologies has provided more powerful tools and methods for laser ultrasonic damage detection. On the other hand, the development of computer and artificial intelligence technologies has also provided new data processing and analysis methods for laser ultrasonic damage detection. For example, algorithms such as deep learning can automatically identify and classify laser ultrasonic signals, improving the accuracy and efficiency of detection. This characteristic allows for scanning and imaging of large-area, complex-shaped internal defects in materials. However, training neural networks requires a large number of samples, and how to quickly and efficiently construct a dataset for training neural networks is a problem that urgently needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to overcome the defects of the prior art by providing a laser ultrasound-guided pixel-level imaging method, device, and storage medium, so as to solve or partially solve the problem of unsatisfactory data set construction efficiency required for deep learning in laser ultrasound imaging.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] One aspect of the present invention provides a laser-ultrasound-guided pixel-level imaging method. Ultrasound is applied to the target to be detected. Based on out-of-plane displacement data obtained from laser scanning, pixel-level imaging results are obtained using a pre-trained 1D-CNN model after normalization and one-hot encoding. The training process of the 1D-CNN model includes the following steps:

[0007] Step S1: For the experimental sample, multiple sampling points are set at large intervals to perform a global coarse scan. For each defect area on the experimental sample obtained in advance, multiple sampling points are set at small intervals to perform a local fine scan. The off-surface displacement data including defect location information after global coarse scan and local fine scan are obtained. The off-surface displacement data is then normalized and processed by one-hot encoding.

[0008] Step S2 involves dividing the out-of-plane displacement data after sample balancing in Step S1 into multiple defect types, completing the annotation, and obtaining a dataset. The defect types are set according to the depth of the defect.

[0009] Step S3: Train the 1D-CNN model using the dataset.

[0010] As a preferred technical solution, in step S1, an ultrasonic excitation signal is applied to a preset position of the experimental sample, and multiple acquisitions are performed at each sampling point at the Nyquist frequency through laser interference to obtain the out-of-plane displacement data corresponding to the sampling point.

[0011] As a preferred technical solution, the experimental sample is pre-set with multiple defects of different shapes and / or different depths.

[0012] As a preferred technical solution, in step S2, the defect types include no defect, defect with a depth of 0.5 mm, defect with a depth of 1 mm, and defect with a depth of 1.5 mm.

[0013] As a preferred technical solution, the process of obtaining pixel-level imaging results and performing imaging using a 1D-CNN model includes:

[0014] The input features corresponding to the out-of-plane displacement data of the sampling points are obtained, and then processed through multi-level convolutional pooling and global average pooling to obtain the output defect type. Based on the defect types of multiple sampling points, pixel-level imaging is achieved through confusion matrix and / or visualization.

[0015] As a preferred technical solution, in step S1, the values ​​of discrete features are mapped to Euclidean space through one-hot encoding.

[0016] As a preferred technical solution, in step S1, normalization is performed using the following formula:

[0017]

[0018] Where x is the original signal, x min x is the minimum value in the original signal. max y is the maximum value in the original signal, and y is the signal obtained after normalization.

[0019] In another aspect, the present invention provides a laser-ultrasound-guided pixel-level imaging device for implementing the aforementioned laser-ultrasound-guided pixel-level imaging method, wherein the pixel-level imaging device comprises:

[0020] Clamping module, used to fix the target to be detected;

[0021] A signal generator, connected to the target, is used to generate an ultrasonic excitation signal and apply it to a preset position on the target;

[0022] A laser interferometer is used to collect out-of-plane displacement data at sampling points on the target.

[0023] The control module, connected to the laser interferometer, is used to control the laser interferometer to acquire out-of-plane displacement data at the sampling point, thereby achieving pixel-level imaging.

[0024] In another aspect, an electronic device is provided, characterized in that it comprises: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for performing the aforementioned laser ultrasound-guided pixel-level imaging method.

[0025] In another aspect, the present invention provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the aforementioned laser-ultrasound-guided pixel-level imaging method.

[0026] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0027] (1) Achieving defect feature classification based on deep learning technology: This invention applies ultrasound to the target to be detected, and based on the out-of-plane displacement data obtained by laser scanning, after normalization and one-hot encoding, uses a pre-trained 1D-CNN model to obtain pixel-level imaging results. By combining laser ultrasonic detection technology with deep learning technology, the automatic identification and classification of different defect features in aluminum plates can be achieved.

[0028] (2) The constructed datasets are rich in types: By setting various defects with different shapes and depths on the experimental samples, the number of training samples is increased, the constructed datasets are rich in types, thereby improving the generalization ability of the model.

[0029] (3) Sample Balancing of the Dataset: This invention balances the number of samples of different categories by combining a full-field coarse scan with a local defect fine scan, thus avoiding sample imbalance. Considering that the size of defects on aluminum plate samples is smaller than the size of the entire plate, the number of samples with defect information obtained in a single coarse scan is much smaller than the number of samples without defects. This situation will cause the model to tend to predict sample points without defects, while ignoring the sample points with defect information that this invention aims to explore, thereby affecting the model's generalization ability and prediction accuracy. To address this issue, after performing a full-field coarse scan, a local fine scan is performed on the defective local area at a shorter scan interval to increase the number of samples with defect information, ensuring that the number of samples of different types is within a reasonable ratio range.

[0030] (4) Pixel-level imaging can be achieved: Since the laser interferometer collects data by scanning point by point, theoretically, the denser the laser interferometer collects data on the defect area, the higher the precision of the final image contour. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a laser-ultrasound-guided pixel-level imaging method in one embodiment;

[0032] Figure 2 This is a schematic diagram of a laser-ultrasound-guided pixel-level imaging device in one embodiment;

[0033] Figure 3 This is a schematic diagram of the experimental sample in the examples;

[0034] Figure 4 This is a schematic diagram of the 1D-CNN model in the embodiment;

[0035] Figure 5 This is a schematic diagram of the confusion matrix of the model test set in the embodiment;

[0036] Figure 6 This is a schematic diagram of defect visualization reconstruction in the embodiment. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0038] Example 1

[0039] To address the problems existing in the prior art, this embodiment provides a laser-ultrasound-guided pixel-level imaging method, see [link to previous document]. Figure 1 The method includes the following steps:

[0040] Step S1, data acquisition.

[0041] See Figure 3 The experiment involved designing two aluminum plate samples, each with two different shapes and three different depths of defects. By arranging multiple defect types, including two different shapes and three different depths (0.5 mm, 1 mm, and 1.5 mm), on a single aluminum plate, the variety of training samples was increased, allowing for better training of the network model and enabling the analysis of multi-defect samples.

[0042] Figure 3 In the diagram, (a), (b), and (c) are schematic diagrams of the defect location and shape of the first sample, the surface where the defect is located, and the other side, respectively; and (d), (e), and (f) are schematic diagrams of the defect location and shape of the second sample, the surface where the defect is located, and the other side, respectively.

[0043] Step S2, data processing.

[0044] Considering the negative impact of sample imbalance on model training, this embodiment adopts a single excitation source to excite at a preset position. A laser interferometer is used to perform a full-field full-field out-of-plane displacement data by combining global coarse scanning and local defect fine scanning. Normalization processing technology and one-hot encoding technology are used to convert the data labels and establish an experimental dataset.

[0045] Specifically, the overall coarse scan involves scanning the entire 200mm×200mm surface of the aluminum plate point by point at a scanning interval of 2.5mm; the local defect fine scan involves setting a 10.2mm square area for the defective local area and scanning it point by point at a scanning interval of 0.3mm.

[0046] Step S3, model training.

[0047] A one-dimensional convolutional neural network model (1D-CNN) was constructed and trained to establish a mapping relationship between laser ultrasound signal data and the location and depth of defects, thereby achieving accurate identification of defects.

[0048] Specifically, the collected signals are normalized to accelerate the convergence speed of the network, improve the efficiency of subsequent feature extraction, and reduce the dynamic transformation range of the signals to a certain extent. At the same time, before the model is trained, the labels of the dataset are converted into one-hot encoding, and the values ​​of discrete features are converted into Euclidean space, making the distance calculation between features more reasonable.

[0049] Step S4, model prediction.

[0050] The shape of the defect is reconstructed using the classification results predicted by the model.

[0051] Specifically, the one-dimensional convolutional neural network takes the preprocessed laser and ultrasonic signal data of the target to be detected as the input of the neural network model, takes the depth information label value of the defect as the output of the neural network model, and reconstructs the shape of the defect.

[0052] The following is a specific implementation process to illustrate this method.

[0053] 1. Experimental data collection.

[0054] (1) Experimental procedure.

[0055] like Figure 2 As shown, an excitation signal generated by a signal generator is connected to the sample on the sample holder via a PZT piezoelectric element, generating a laser ultrasonic signal on the sample surface. The waveform of the excitation signal is displayed on an oscilloscope. A piezoelectric element with an excitation frequency of 0.2 MHz is selected, and the waveform is modulated using a Hanning window to generate a 5-cycle excitation signal. The laser scanning control system controls the movement of the sample holder in the X and Y directions and controls the laser interferometer to scan the sample. The acquired signals are displayed on the system interface and saved in the background. The Nyquist sampling frequency is set to 25 MHz, and 10,000 samples are taken at each sampling point on the aluminum plate.

[0056] (2) Experimental specimen.

[0057] This experiment used 7075 aluminum plate as the test specimen. The aluminum plate specimen was 300mm long and wide, and 2mm thick. Two aluminum plate specimens were prepared for the experiment, and six defects of different shapes and depths were machined into each specimen. For example... Figure 3 As shown, the first aluminum plate has two types of blind hole defects: square and equilateral triangle, both 10mm in size, with depths of 0.5mm, 1mm, and 1.5mm respectively from top to bottom. Similarly, the second aluminum plate also has these two types of blind hole defects.

[0058] (3) Experimental operation steps.

[0059] The experiment selected a 200mm × 200mm area at the center of the aluminum plate sample as the scanning area. First, a global coarse scan was performed, scanning the entire 200mm × 200mm surface of the aluminum plate point-by-point at 2.5mm intervals. Next, a detailed scan of local defects was performed, with a 10.2mm square area defined for defective regions and scanned point-by-point at 0.3mm intervals. Finally, the dataset obtained according to the above experimental steps was divided into four categories: no defects, defects with a depth of 0.5mm, defects with a depth of 1mm, and defects with a depth of 1.5mm. For ease of subsequent model training, these four types were labeled 0, 1, 2, and 3, respectively.

[0060] 2. Data processing.

[0061] (1) Normalization process.

[0062] To accelerate network convergence, improve the efficiency of subsequent feature extraction, and reduce the dynamic range of the signal to some extent, we consider normalizing the acquired signal. The normalization formula used is:

[0063]

[0064] Where x is the original signal, x min x is the minimum value in the original signal. max y is the maximum value in the original signal, and y is the signal obtained after normalization.

[0065] (2) One-hot encoding technology.

[0066] Before inputting data into the network model, the dataset labels are converted to one-hot encoded form. This converts discrete feature labels into binary form, mapping them to a high-dimensional space where only one bit is 1 and the others are 0. In classification problems, the calculation of distance and similarity between features relies on Euclidean space, and one-hot encoding makes the distance calculation between features more reasonable by transforming the values ​​of discrete features to Euclidean space.

[0067] 3. Training and prediction of 1D-CNN models.

[0068] A 1D-CNN structure is proposed for feature extraction of laser ultrasonic signals from aluminum plate structures, such as... Figure 4 As shown, it mainly consists of a data input layer, a one-dimensional convolutional layer, a pooling layer, a global average pooling layer, and an output layer.

[0069] This model comprises eight convolutional layers for extracting input features, four max-pooling layers for reducing dimensionality and preserving salient features, and one global average pooling layer for integrating features. The convolutional layers consist of 16 kernels of length 8×1, 64 kernels of length 4×1, 256 kernels of length 4×1, and 512 kernels of length 4×1, respectively, using the ReLU activation function. A pooling layer of length 2×1 is placed after every two convolutional layers, employing max pooling to reduce the dimensionality of the data. The final layer uses global average pooling instead of a fully connected layer. The model completes the classification task using the Softmax function, employing cross-entropy loss and the Adam optimizer. Specifically, the Adam optimizer is configured with a learning rate of 0.0002, a Dropout layer parameter of 0.3, and 100 iterations.

[0070] In this embodiment, the training set to test set ratio is 7:3. The test set data, after normalization and one-hot encoding, is input into the trained network model to classify the unlabeled data and output the results. Figure 5 The image shows the confusion matrix of the model's test set. As can be seen from the figure, the model's classification prediction results are quite good. Furthermore, to more intuitively analyze defects, this invention reconstructs the model-classified defect signals onto an aluminum plate for visual display, as shown below. Figure 6 As shown, the defect in (a) was scanned point by point with a laser interferometer at 1mm intervals, with a scanning range of 30mm × 30mm. The collected data was processed using the methods mentioned above and input into the trained model for prediction. The prediction results were reconstructed using a visualization method, as shown in (b). The basic shape of the defect can be seen from the figure, but there are a small number of scan points in the surrounding area that were misjudged by the model, although the number is relatively small. It can be seen that the model obtained by the method of this embodiment has good automatic defect identification and classification capabilities, as well as good generalization performance.

[0071] This method combines laser ultrasonic testing technology with deep learning technology to automatically identify and classify different defect features in various targets such as aluminum plates. In sample preparation, multiple defects are processed on an aluminum plate sample to increase the number of training samples and improve the generalization ability of the model. A combination of full-field coarse scanning and local defect fine scanning is used to balance the number of samples of different categories and avoid sample imbalance.

[0072] Example 2

[0073] Based on Example 1, this example provides a laser-ultrasound-guided pixel-level imaging device to achieve the pixel-level imaging method in Example 1. See [link to example]. Figure 2 The equipment includes:

[0074] Clamping module, used to fix the target to be detected;

[0075] A signal generator, connected to the target, is used to generate an ultrasonic excitation signal and apply it to a preset position on the target;

[0076] A laser interferometer is used to acquire out-of-plane displacement data at sampling points on a target.

[0077] The control module, connected to the laser interferometer, is used to control the laser interferometer to acquire out-of-plane displacement data at the sampling point and achieve pixel-level imaging.

[0078] Laser ultrasonic damage detection technology, as a non-destructive testing method with unique advantages, has significant application value in material damage detection, defect assessment, and safety assurance. With the continuous development and improvement of related technologies, laser ultrasonic damage detection technology will play an even more important role in the future.

[0079] Example 3

[0080] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing a laser ultrasound-guided pixel-level imaging method as described in Embodiment 1.

[0081] Example 4

[0082] This embodiment provides a computer-readable storage medium, characterized in that it includes one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing a laser ultrasound-guided pixel-level imaging method as described in Embodiment 1.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A laser ultrasound-guided pixel-level imaging method, characterized in that: Ultrasound is applied to the target to be detected, and the off-plane displacement data obtained based on laser scanning is normalized and one-hot encoded, and then the pixel-level imaging result is obtained using the pre-trained 1D-CNN model, wherein the training process of the 1D-CNN model includes the following steps: Step S1, for the experimental sample, multiple sampling points are set at large intervals to perform a global rough scan, and for each defect area on the experimental sample obtained in advance, multiple sampling points are set at small intervals to perform a local fine scan, to obtain off-plane displacement data including defect location information after global rough scan-local fine scan, and the off-plane displacement data is normalized and one-hot encoded; Step S2, dividing the off-plane displacement data after sample balancing in step S1 into multiple defect types, completing labeling, and obtaining a data set, wherein the defect type is set according to the depth of the defect; Step S3, using the data set to train the 1D-CNN model.

2. The laser ultrasound guided pixel-level imaging method according to claim 1, characterized in that: In the step S1, an ultrasonic excitation signal is applied to a preset position of the experimental sample, and multiple acquisitions are performed at each sampling point at the Nyquist frequency through laser interference to obtain off-plane displacement data corresponding to the sampling point.

3. The laser ultrasound guided pixel-level imaging method according to claim 1, characterized in that: The experimental sample is pre-arranged with a plurality of defects of different shapes and / or depths.

4. The laser ultrasound guided pixel-level imaging method according to claim 1, characterized in that: In the step S2, the defect types include no defect, defect with a depth of 0.5 mm, defect with a depth of 1 mm, and defect with a depth of 1.5 mm.

5. The laser ultrasound guided pixel-level imaging method according to claim 1, characterized in that: The process of using the 1D-CNN model to obtain pixel-level imaging results and perform imaging includes: The input features corresponding to the off-plane displacement data of the sampling points are obtained, and multi-level convolution pooling and global average pooling are performed in sequence to obtain the output defect type. Based on the defect types of multiple sampling points, pixel-level imaging is achieved through confusion matrix and / or visualization.

6. The laser ultrasound guided pixel-level imaging method according to claim 1, characterized in that: In the step S1, the values ​​of discrete features are mapped to the Euclidean space through one-hot encoding.

7. The laser ultrasound guided pixel-level imaging method according to claim 1, characterized in that: In the step S1, the normalization is performed using the following formula: Among them, x is the original signal, x min is the minimum value in the original signal, x max is the maximum value in the original signal, and y is the signal obtained after normalization.

8. A laser ultrasound-guided pixel-level imaging device, characterized in that: A laser ultrasound-guided pixel-level imaging method according to any one of claims 1 to 7, wherein the pixel-level imaging device comprises: A clamping module, used to fix the target to be detected; A signal generator, connected to the target, for generating an ultrasonic excitation signal and applying the ultrasonic excitation signal to a preset position of the target; A laser interferometer, used to collect off-plane displacement data at sampling points on the target; The control module is connected to the laser interferometer and is used to control the laser interferometer to collect off-plane displacement data at sampling points to achieve pixel-level imaging.

9. An electronic device, characterized in that: include: One or more processors and a memory, wherein the memory stores one or more programs, wherein the one or more programs include instructions for executing a laser ultrasound-guided pixel-level imaging method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: It comprises one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs comprise instructions for executing a laser ultrasound-guided pixel-level imaging method as described in any one of claims 1-8.

Images