Deep learning-based laminated plate layering defect detection method and device
Through deep learning and infrared thermal imaging technology, a laminated plate layered defect depth prediction network is built, which solves the problems of low efficiency and high cost in the existing technology, and realizes accurate prediction and efficient detection of laminated plate layered damage depth.
Patent Information
- Application Number
- CN202510542113.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-28
- Publication Date
- 2025-07-25
AI Technical Summary
The prior art is inefficient, costly and labor-dependent in detecting layered defects of fiber-reinforced composite laminates, making it difficult to accurately predict depth information.
The laminated layered defect detection method based on deep learning is adopted, and one-dimensional convolutional network and infrared thermal imaging technology are used to construct a laminated layered defect depth prediction network, and combined with thermal image sequence and segmentation mask map for training to achieve accurate prediction of defect depth.
The precise prediction of the layered damage depth of laminated plates is achieved, which reduces detection costs, improves detection efficiency, and reduces dependence on manual detection.
Smart Images
Figure CN120374594A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of thermal imaging, machine vision and artificial intelligence, and particularly relates to a method and device for detecting delamination defects of a laminate based on deep learning. Background Art
[0002] Fiber-reinforced composite laminates are widely used in fields such as aerospace, shipbuilding, and tank armor due to their excellent mechanical properties such as high modulus, high strength, and low density. However, due to their anisotropic structure, FRCLs are prone to various complex damage forms during manufacturing and under impact, such as fiber fracture, matrix cracking, and delamination. Among them, delamination damage is one of the most important damage forms affecting the performance of composite materials. Its occurrence in composite materials will significantly reduce the structural strength of composite materials and even cause catastrophic failures. Therefore, it is of great significance to quickly and accurately detect, quantify, and evaluate the delamination defects of laminates. Currently, traditional non-destructive testing methods such as CT scanning and thermal imaging are mostly relied on. CT scanning has problems of low detection efficiency and high detection cost. Thermal imaging technology has problems such as poor visibility and difficult prediction of depth information, but the detection speed is fast. Moreover, the detection results of traditional methods all require manual measurement, resulting in problems of low detection efficiency, strong subjectivity, and high labor costs.
[0003] In recent years, the rapid development in the fields of deep learning technology and thermal imaging technology has made it possible to intelligently detect delamination defects of laminates. For example, deep learning algorithms have made breakthrough progress in fields such as object detection, language translation detection, natural language processing, and voice recognition. Artificial intelligence technology has become a key technology in the new round of scientific and technological revolution and industrial transformation. The explosive breakthroughs in deep learning and computational efficiency are the keys to promoting the progress of artificial intelligence technology. In recent years, deep learning models have been continuously deepened and formed a trend of ultra-large-scale development, making it possible to apply deep learning technology to the detection of delamination defects of laminates in this field.
[0004] Defects with different depths have different visible degrees in the defect area during the heating process. With the support of deep learning, it becomes possible to quickly and accurately identify the defect position and size from the collected thermal image sequence and predict the depth value of the defect. Summary of the Invention
[0005] To solve the above technical problems existing in the prior art, the present invention proposes a method and device for detecting delamination defects of a laminate based on deep learning, and realizes accurate prediction of the depth of delamination damage of the laminate based on a thermal image sequence.
[0006] On the one hand, to achieve the above object, the present invention provides a method for detecting delamination defects of a laminate based on deep learning, including:
[0007] Collecting the thermal image of the surface of the laminate to be processed;
[0008] Construct a prediction network for the depth of delamination defects in a laminate;
[0009] Input the thermal image of the surface of the laminate to be processed into the prediction network for the depth of delamination defects in the laminate for detection, and output the depth value of the delamination defect from the detection surface;
[0010] Among them, the prediction network for the depth of delamination defects in the laminate is a one-dimensional convolutional network, which takes the thermal image sequence dataset and the corresponding segmentation mask image as inputs and is trained with depth information as the dataset label.
[0011] Preferably, constructing the prediction network for the depth of delamination defects in the laminate includes:
[0012] Collect the thermal image sequence of the surface of the laminate with delamination defects of different sizes and depths, and perform preprocessing operations to obtain the preprocessed dataset;
[0013] Use the preprocessed dataset to train the thermal image enhancement network to obtain the enhanced thermal image sequence;
[0014] Based on the enhanced thermal image sequence, train the segmentation network to obtain the planar geometric information of the delamination area of the laminate;
[0015] Take the thermal image sequence dataset and the planar geometric information of the delamination area of the laminate as inputs, and the real depth information as the dataset label, and train the depth prediction network to obtain the prediction network for the depth of delamination defects in the laminate.
[0016] Preferably, performing preprocessing operations to obtain the preprocessed dataset includes:
[0017] Use thermography to collect the thermal image sequence of the surface of the laminate with delamination defects of different sizes and depths, and perform normalization operations to obtain the normalized images;
[0018] Based on the data augmentation method, perform random shearing, mirroring, rotation, and adding Gaussian noise operations on the normalized images to expand the dataset and obtain the preprocessed dataset.
[0019] Preferably, the thermal image enhancement network is a convolutional autoencoder, and the segmentation network is MSAUNet.
[0020] Preferably, the processing process of the MSAUNet includes:
[0021] Obtain the enhanced thermal image sequence, acquire the high-dimensional data feature information of the image through a multi-scale self-attention convolution module and pooling operations, and use skip connections and upsampling convolution to restore the high-dimensional data feature information of the image, and output a segmentation mask image containing hierarchical position and size information.
[0022] Preferably, the multi-scale self-attention convolution module includes a multi-scale convolution unit, a channel self-attention unit, and a spatial self-attention unit connected in sequence;
[0023] Among them, the multi-scale convolution unit is used to extract multi-scale features of the input thermal image sequence, improving the adaptability to defects of different sizes;
[0024] The channel self-attention unit is used to perform adaptive cross-fusion on the extracted multi-scale features in the channel dimension;
[0025] The spatial self-attention unit is used to adaptively enhance the features in the spatial dimension, increasing the feature weight of the defect area.
[0026] On the other hand, to achieve the above object, the present invention also provides a laminated plate delamination defect detection device based on deep learning, including:
[0027] An acquisition module, including an infrared thermal imaging non-destructive testing system and an adjustable detection platform, for acquiring the surface thermal image of the laminated plate to be processed;
[0028] A model construction module: used to construct a laminated plate delamination defect depth prediction network;
[0029] A depth prediction module, for inputting the surface thermal image of the laminated plate to be processed into the laminated plate delamination defect depth prediction network for detection, outputting the depth value of the delamination defect from the detection surface, and predicting the defect depth information.
[0030] Preferably, the model construction module includes:
[0031] A data preprocessing unit, for preprocessing the acquired thermal image data;
[0032] A data transmission unit, for transmitting the preprocessed infrared thermal image to the thermal image segmentation unit;
[0033] A thermal image segmentation unit, for extracting and analyzing the infrared thermal image to obtain the position and size information of the delamination defect in the infrared thermal image.
[0034] Preferably, the infrared thermal imaging non-destructive testing system includes an infrared thermal imager, a heating device, a signal controller, and a computer, and the adjustable detection platform includes an infrared camera fixture with adjustable vertical height and horizontal angle, and a specimen placement platform with adjustable size and angle.
[0035] The present invention also provides a computer device, including a central processing unit, a graphics processing unit, registers, and a computer program stored on the registers. The registers execute the computer program to implement the steps of the method for detecting delamination defects of laminated plates based on deep learning.
[0036] Compared with the prior art, the present invention has the following advantages and technical effects:
[0037] The present invention utilizes the low cost, high efficiency, and non-contact advantages of deep learning technology and infrared thermal imaging technology, enabling the present invention to achieve effective applications in the identification and depth prediction of delamination defects of laminated plates, changing the method that the identification and quantification of most delamination defects of laminated plates rely on traditional manual detection, and realizing the accurate prediction of the depth of delamination damage of laminated plates based on thermal image sequences. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:
[0039] Figure 1 is a flowchart of a method for detecting delamination defects of laminated plates based on deep learning according to an embodiment of the present invention;
[0040] Figure 2 is a schematic diagram of a device for detecting delamination defects of laminated plates based on deep learning according to an embodiment of the present invention;
[0041] Figure 3 is a schematic diagram of an infrared thermal imaging non-destructive testing system according to an embodiment of the present invention;
[0042] Figure 4 is a schematic diagram of an adjustable detection platform according to an embodiment of the present invention;
[0043] Figure 5 is a schematic diagram of a convolutional autoencoder structure according to an embodiment of the present invention;
[0044] Figure 6 is a schematic diagram of an MSAUNet model structure according to an embodiment of the present invention;
[0045] Figure 7 is a schematic diagram of a multi-scale self-attention convolution module in the MSAUNet model according to an embodiment of the present invention;
[0046] Figure 8 is a schematic diagram of a depth prediction module according to an embodiment of the present invention;
[0047] Figure 9 is a schematic diagram of a computer device according to an embodiment of the present invention. Specific Embodiment
[0048] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will detail this application with reference to the drawings and in combination with the embodiments.
[0049] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.
[0050] This embodiment proposes a method for detecting delamination defects of laminated plates based on deep learning, as Figure 1 , including:
[0051] Collect the surface thermal image of the laminated plate to be processed;
[0052] Construct a depth prediction network for delamination defects of laminated plates;
[0053] Input the surface thermal image of the laminated plate to be processed into the depth prediction network for delamination defects of laminated plates for detection, and output the depth value of the delamination defect from the detection surface;
[0054] Among them, the depth prediction network for delamination defects of laminated plates is a one-dimensional convolutional network. The depth prediction network for delamination defects of laminated plates takes the thermal image sequence dataset and the corresponding segmentation mask image as inputs and is trained with depth information as the dataset label.
[0055] Specifically, this embodiment utilizes the low cost, high efficiency and non-contact advantages of deep learning technology and infrared thermal imaging technology, enabling the present invention to achieve effective applications in the identification and depth prediction of delamination defects of laminated plates, changing the method that most delamination defect identification and quantification of laminated plates rely on traditional manual detection, and realizing the accurate prediction of the depth of delamination damage of laminated plates based on thermal image sequences.
[0056] Further, constructing a depth prediction network for delamination defects of laminated plates includes:
[0057] Collect the surface thermal image sequence of laminated plates with delamination defects of different sizes and depths, and perform preprocessing operations to obtain a preprocessed dataset;
[0058] Use the preprocessed dataset to train a thermal image enhancement network to obtain an enhanced thermal image sequence;
[0059] Based on the enhanced thermal image sequence, train a segmentation network to obtain the planar geometric information of the delamination area of the laminated plate;
[0060] Taking the thermal image sequence dataset and the planar geometric information of the delaminated area of the laminate as inputs, and the true delamination depth information as the dataset label, train a depth prediction network to obtain the delamination defect depth prediction network for the laminate. The true delamination depth information refers to the distance from the delamination defect to the detection surface.
[0061] Specifically, the delamination defect depth prediction network for the laminate is a one-dimensional convolutional network. The input is the thermal image sequence and the corresponding segmentation mask, and the output part includes a classification module and a regression module. The classification module is used to output which layer of the laminate the delamination defect is in. The regression module can directly output the distance from the delamination defect to the surface, without being limited by the layer thickness, and is suitable for laminates of different thicknesses.
[0062] In this embodiment, the laminate material is a ultra-high molecular weight polyethylene fiber reinforced material.
[0063] Furthermore, perform preprocessing operations to obtain the preprocessed dataset, including:
[0064] Use thermography to collect the thermal image sequence of the surface of the laminate with delamination defects of different sizes and depths, and perform normalization operations to obtain the normalized images;
[0065] Based on the data augmentation method, perform random shearing, mirroring, rotation, and adding Gaussian noise operations on the normalized images to expand the dataset and obtain the preprocessed dataset.
[0066] Furthermore, the thermal image enhancement network is a convolutional autoencoder, and the segmentation network is MSAUNet. Obtain the geometric information of the delamination defect area in the thermal image sequence through MSAUNet.
[0067] Specifically, use the preprocessed dataset to train the thermal image enhancement network to obtain the enhanced thermal image sequence, including:
[0068] The enhancement network compresses the input and restores the compressed data, retains the main information, and removes irrelevant information such as noise, etc., and can be used to output a thermal image sequence with stronger visibility and better signal-to-noise ratio, that is, the enhanced thermal image sequence.
[0069] Specifically, based on the enhanced thermal image sequence, train the segmentation network to obtain the planar geometric information of the delaminated area of the laminate, including:
[0070] Use the delamination defect mask output by the segmentation network for analysis and processing to obtain the planar geometric information of the delaminated area of the laminate to achieve delamination defect detection.
[0071] Furthermore, the processing process of MSAUNet includes:
[0072] Obtain the enhanced thermal image sequence, obtain the high-dimensional data feature information of the image through the multi-scale self-attention convolution module and pooling operation, and use skip connection and upsampling convolution to restore the high-dimensional data feature information of the image, and output a segmentation mask image containing hierarchical position and size information.
[0073] Among them, the multi-scale self-attention convolution module includes a multi-scale convolution unit, a channel self-attention unit, and a spatial self-attention unit connected in sequence;
[0074] The multi-scale convolution unit is used to extract the multi-scale features of the input thermal image sequence to improve the adaptability to defects of different sizes.
[0075] The channel self-attention unit is used to perform adaptive cross-fusion on the extracted multi-scale features in the channel dimension, so that the model focuses on the key features that can effectively distinguish the defect area and the background area, suppresses the expression of invalid features, and improves the detection accuracy of the model.
[0076] The spatial self-attention unit is used to adaptively enhance the features in the spatial dimension, increase the feature weight of the defect area, and suppress the feature values of the background area to achieve more accurate defect segmentation.
[0077] This embodiment also provides a laminated board delamination defect detection device based on deep learning, such as Figure 2 , including:
[0078] An acquisition module, including an infrared thermal imaging non-destructive testing system and an adjustable detection platform, for acquiring the surface thermal image of the laminated board to be processed;
[0079] A model construction module: used to construct a laminated board delamination defect depth prediction network;
[0080] A depth prediction module, used to input the surface thermal image of the laminated board to be processed into the laminated board delamination defect depth prediction network for detection, output the depth value of the delamination defect from the detection surface, and predict the defect depth information.
[0081] Furthermore, the model construction module includes:
[0082] A data preprocessing unit, used to preprocess the collected thermal image data, including image amplification, enhancement, and data normalization processing in this embodiment.
[0083] A data transmission unit, used to transmit the preprocessed infrared thermal image to the thermal image segmentation unit;
[0084] A thermal image segmentation unit, used to extract and analyze the infrared thermal image to obtain the position and size information of the delamination defect in the infrared thermal image.
[0085] Furthermore, the infrared thermal imaging non-destructive testing system includes an infrared thermal imager, a heating device, a signal controller, and a computer, as Figure 3 .
[0086] Specifically, the purpose of the infrared thermal imaging non-destructive testing system is to collect the thermal image sequence on the surface of the specimen during the heating process. The heating device consists of two 1000w halogen lamps placed on a liftable bracket.
[0087] Furthermore, the adjustable detection platform includes an infrared camera fixture that can adjust the vertical height and horizontal angle, and a specimen placement platform that can adjust the size and angle, as Figure 4 shown.
[0088] Specifically, the purpose of the adjustable detection platform is to standardize the thermal image acquisition process. During the acquisition process, the infrared camera is adjusted to be perpendicular to the specimen, and the infrared camera is adjusted to a certain position so that the specimen fills the entire camera field of view.
[0089] The image information collected by the infrared camera is a 256×256 grayscale image at 10 frames per second. Starting from the beginning of heating, a total of 20s are collected, including 10s of heating time and 10s of cooling time. Each sample obtains a thermal image sequence containing 200 thermal images in total.
[0090] Furthermore, the processing model of the data preprocessing unit is a convolutional autoencoder.
[0091] The calculation process of the convolutional autoencoder is divided into two parts: a convolutional compression encoder and a convolutional restoration encoder. First, the input thermal image sequence is compressed within the numerical range of [-1,1] through a normalization operation. The convolutional compression encoder compresses the 256×256×200 thermal image sequence into a latent feature of 64×64×8. The convolutional restoration encoder restores it to a 256×256 thermal image, as Figure 5 shown. The self-supervised training is used during the training process, and the label is the input thermal image. The trained model outputs the latent feature as the enhanced thermal image data.
[0092] Furthermore, the analysis and processing model of the thermal image segmentation unit is MSAUNet.
[0093] The calculation process of the MSAUNet model is divided into 4 steps: First, the latent feature output by the convolutional encoder is used as the input, and a multi-scale convolutional module is used for feature extraction. Second, the maximum pooling is used to process the features to reduce the computational cost and accelerate the model convergence. Third, the extracted features are restored and fused through multiple transposed convolution modules and skip connections. Fourth, the predicted hierarchical defect mask map is output through a fully connected layer, as Figures 6 - 7 shown.
[0094] Further, the analysis and processing model of the depth prediction module is a one-dimensional convolutional network.
[0095] The calculation process of the depth prediction module is divided into 4 steps:
[0096] In the first step, a dot product operation is performed on the thermal image sequence of 256*256*200 and the mask image of the delamination defect. In the second step, the operation result is subjected to global average pooling to obtain an average thermal signal with a size of 1*200. In the third step, the thermal signal data is input into multiple one-dimensional convolutional layers and one-dimensional pooling layers for feature extraction. In the fourth step, the extracted features are input into the classification module and regression module composed of fully connected layers for outputting the delamination defect depth information, as Figure 8 shown.
[0097] This embodiment also provides a computer device, including a central processing unit, a graphics processing unit, registers, and a computer program stored in the registers. The registers execute the computer program to implement the steps of the method for detecting delamination defects of a laminated board based on deep learning, and its device structure is as Figure 9 shown.
[0098] In order to verify the effectiveness of a method and device for detecting delamination defects of a laminated board based on deep learning proposed in this example, a verification experiment was designed. 28 samples were selected for 5 repeated experiments, and at the same time, a comparison was made with the standard measurement parameters. The maximum error of depth prediction was 5.56%, which fully demonstrated the effectiveness of the technical solution of this embodiment.
[0099] The above is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A method for detecting delamination defects in laminated plates based on deep learning, characterized in that, Including: Collecting the surface thermal image of the laminate to be processed; Constructing a depth prediction network for laminate delamination defects; Inputting the surface thermal image of the laminate to be processed into the depth prediction network for laminate delamination defects for detection, and outputting the depth value of the delamination defect from the detection surface; Among them, the depth prediction network for laminate delamination defects is a one-dimensional convolutional network, which takes the thermal image sequence dataset and the corresponding segmentation mask image as inputs and is trained with depth information as the dataset label.
2. The method for detecting delamination defects of a laminate based on deep learning according to claim 1, wherein Constructing the depth prediction network for laminate delamination defects includes: Collecting a sequence of surface thermal images of laminates with delamination defects of different sizes and depths, and performing preprocessing operations to obtain a preprocessed dataset; Using the preprocessed dataset to train a thermal image enhancement network to obtain an enhanced thermal image sequence; Training a segmentation network based on the enhanced thermal image sequence to obtain the planar geometric information of the laminate delamination area; Taking the thermal image sequence dataset and the planar geometric information of the laminate delamination area as inputs, and the real depth information as the dataset label, training a depth prediction network to obtain the depth prediction network for laminate delamination defects.
3. The method for detecting delamination defects of a laminated plate based on deep learning according to claim 2, wherein Performing preprocessing operations to obtain a preprocessed dataset, including: Using thermal imaging to collect a sequence of surface thermal images of laminates with delamination defects of different sizes and depths, and performing normalization operations to obtain normalized images; Based on data augmentation methods, randomly shearing, mirroring, rotating, and adding Gaussian noise to the normalized images to expand the dataset, and obtaining the preprocessed dataset.
4. The method for detecting delamination defects of a laminate based on deep learning according to claim 2, wherein The thermal image enhancement network is a convolutional autoencoder, and the segmentation network is MSAUNet.
5. The method for detecting delamination defects of a laminated plate based on deep learning according to claim 4, characterized in that, The processing process of the MSAUNet includes: Obtaining the enhanced thermal image sequence, obtaining the high-dimensional data feature information of the image through a multi-scale self-attention convolution module and pooling operations, and using skip connections and upsampling convolutions to restore the high-dimensional data feature information of the image, and outputting a segmentation mask image containing delamination position and size information.
6. The method for detecting delamination defects of a laminate based on deep learning according to claim 1, wherein The multi-scale self-attention convolution module includes a multi-scale convolution unit, a channel self-attention unit, and a spatial self-attention unit connected in sequence; Among them, the multi-scale convolution unit is used to extract the multi-scale features of the input thermal image sequence to improve the adaptability to defects of different sizes; The channel self-attention unit is used to perform adaptive cross-fusion on the extracted multi-scale features in the channel dimension; The spatial self-attention unit is used to adaptively enhance the features in the spatial dimension to improve the feature weight of the defect area.
7. A delamination defect detection device for laminated plates based on deep learning, characterized in that, Including: A collection module, including an infrared thermal imaging non-destructive testing system and an adjustable testing platform, for collecting the surface thermal image of the laminate to be processed; A model construction module: for constructing a depth prediction network for laminate delamination defects; A depth prediction module, for inputting the surface thermal image of the laminate to be processed into the depth prediction network for laminate delamination defects for detection, and outputting the depth value of the delamination defect from the detection surface to predict the defect depth information.
8. The laminate delamination defect detection device based on deep learning according to claim 7, wherein, The model construction module includes: A data preprocessing unit for preprocessing the collected thermal image data; A data transmission unit for transmitting the preprocessed infrared thermal image to the thermal image segmentation unit; A thermal image segmentation unit for extracting and analyzing the infrared thermal image to obtain the position and size information of the delamination defects in the infrared thermal image.
9. The laminate delamination defect detection device based on deep learning according to claim 7, characterized in that, The infrared thermal imaging non-destructive testing system includes an infrared thermal imager, a heating device, a signal controller and a computer, and the adjustable detection platform includes an infrared camera fixture with adjustable vertical height and horizontal angle, and a specimen placement platform with adjustable size and angle.
10. A computer device, comprising a central processing unit, a graphics processing unit, registers, and a computer program stored on the registers, characterized in that, The register executes the computer program to implement the steps of the delamination defect detection method for laminated plates based on deep learning according to any one of claims 1-6.