A multi-source data nondestructive testing method for stealth coating materials based on deep learning

Abstract

The present invention discloses a multi-source data nondestructive testing method for stealth coating materials based on deep learning, comprising the following steps: acquiring infrared thermal imaging images and optical images; preprocessing the infrared thermal imaging images; marking defect locations in the infrared thermal imaging images; stacking the infrared image with the optical image as a fourth channel; performing data enhancement to expand the data sample and divide the data into training and test sets; adjusting the hyperparameters of the convolutional neural network YOLO-V8 to obtain an optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images; calculating the defect area through network calculation, grading the damage level according to internal specifications, and mapping the damage to the optical image of the test piece. The present invention can solve the problems of inaccurate detection caused by single-source data and complex and time-consuming detection methods, facilitate the rapid acquisition of the location and area of ​​coating material defects, and improve the accuracy, efficiency, and robustness of detection.

Description

Technical Field

[0001] The present invention relates to the technical field of stealth coating material defect detection technology, and in particular to a multi-source data non-destructive detection method for stealth coating materials based on deep learning. Background Art

[0002] Stealth coatings are applied to aircraft surfaces to absorb electromagnetic waves emitted by radar, enhancing their radar and infrared stealth capabilities. However, due to environmental factors (such as UV radiation, temperature fluctuations, and humidity), physical impact, friction, and aging, stealth coatings on aircraft surfaces may wear, peel, or crack. This can reduce their ability to absorb or scatter radar waves, increasing the aircraft's RCS and significantly reducing its stealth capabilities. Therefore, it is crucial to be aware of the status of an aircraft's stealth coating immediately.

[0003] Currently, the main methods for detecting stealth coating defects include visual inspection, eddy current inspection, ultrasonic inspection, etc. These methods have the following problems:

[0004] (1) The visual method cannot observe the defects inside the coating and cannot directly obtain the degree of attenuation of the coating's absorption capacity. The eddy current method cannot detect non-conductive materials and is difficult to detect objects with complex shapes. The ultrasonic detection method requires the use of a coupling agent, which will have a certain impact on the coating's absorption effect.

[0005] (2) Infrared thermal imaging technology has the advantages of being non-contact and having a large detection area. However, the contrast between defects and non-defects in the thermal image is not high, and it is easy to make misjudgments with the naked eye. Therefore, the complete detection time is long and the accuracy is low.

[0006] (3) The degree of damage determined by images is mostly based on the experience of maintenance personnel, and there is currently no systematic damage degree classification. Summary of the Invention

[0007] In light of this, the present invention aims to provide a deep learning-based nondestructive testing method for stealth coating materials using multi-source data. This method uses a combination of infrared thermal imaging and optical image detection techniques to obtain preliminary data, and utilizes deep learning to develop a detection model. This method addresses the inaccurate detection caused by single-source data, as well as the complex and time-consuming detection methods. It facilitates rapid determination of the location and area of ​​coating defects, improving detection accuracy, efficiency, and robustness. The damage area and type are then graded according to internal standards, facilitating timely repairs by frontline maintenance personnel in accordance with applicable maintenance regulations.

[0008] The present invention solves the technical problem by adopting the following technical solutions:

[0009] A multi-source data nondestructive testing method for stealth coating materials based on deep learning, comprising the following steps:

[0010] Step S10: thermally excite the test piece to obtain an infrared thermal imaging image, and collect an optical image of the coating of the test piece at the same angle;

[0011] Step S20: pre-processing the infrared thermal imaging image;

[0012] Step S30: combining the acquired infrared thermal imaging image with the change of its corresponding optical image to determine the damage position and shape of the optical image of the damaged specimen and mark it;

[0013] Step S40: Calculate the mean and variance of the pixel values ​​of the four channels of the acquired thermal imaging image and optical image, and perform normalization processing on the brightness and contrast; after the processing, stack the infrared image as the fourth channel with the optical image;

[0014] Step S50: performing data enhancement on the infrared thermal imaging image and its corresponding optical image obtained in step S10 to expand the data samples, and dividing the expanded data samples into a training set and a test set in a ratio of 8:2;

[0015] Step S60: Input the training set obtained in step S40 into the convolutional neural network YOLO-V8 for training. During the training process, input the test set into the trained convolutional neural network YOLO-V8, monitor the detection accuracy of the test set on the convolutional neural network YOLO-V8 in real time, and optimize the convolutional neural network YOLO-V8 by adjusting the hyperparameters of the convolutional neural network YOLO-V8 to obtain an optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images;

[0016] Step S70: Acquire the infrared thermal imaging image and optical image to be tested in the manner of step S10, process them through steps S20, S30, and S40, and then input them into the optimized convolutional neural network YOLO-V8 model for detecting mixed images obtained in step S60 to detect whether the mixed image contains damage. If damage is present, obtain the location coordinates and area of ​​the damage, classify the damage level according to internal specifications, and map the damage to the optical image of the specimen.

[0017] Furthermore, step S10 specifically includes the following steps:

[0018] Step S11: Setting up a light-stimulated infrared thermal imaging nondestructive testing system in the maintenance yard, using a halogen lamp as the excitation light source with a color temperature of 4500-5500K, fixing the infrared thermal imager on a lamp stand and aiming it at the area to be tested on the specimen;

[0019] Step S12: Place the optically excited infrared thermal imaging nondestructive testing system at a position 0.8 to 1.2 meters from the center of the test area in the normal direction, set the excitation current to 9A, the heating time to 2000ms, the power supply voltage to 230V, and the excitation signal to a phase-locked pulse signal to scan the area;

[0020] Step S13: using an industrial camera to collect an optical image of the coating from the normal direction of the center of the area to be measured;

[0021] Step S14: importing the acquired coating infrared thermal imaging image and optical image into the light-stimulated infrared thermal imaging nondestructive testing system.

[0022] Furthermore, in step S20, the method for preprocessing the infrared thermal imaging image is as follows: the infrared thermal imaging image and the optical image obtained in S10 are size-aligned, the resolution of the industrial camera used is (1920, 1080, 3), and the resolution of the infrared thermal imaging image is (640, 480, 1); OpenCV is used for bilinear interpolation to adjust the infrared image to (1920, 1080, 1).

[0023] Furthermore, in step S30, the method of marking the damage position and shape of the optical image is to frame the damage with a rectangular frame, and the marking format is [X min ,Y min ,X max ,Y max ], where X min , Y min is the coordinate of the upper left corner of the rectangle, X max , Y max The coordinates of the lower right corner of the rectangle.

[0024] Furthermore, in step S40, the brightness and contrast are normalized using the following formula:

[0025]

[0026] Where μ is the mean of the four channel pixel values, σ is the variance of the four channel pixel values, f(x, y) represents the adjusted image pixel, g(x, y) represents the image pixel before adjustment, and (x, y) represents the pixel coordinate position.

[0027] Furthermore, in step S50, the Mixup algorithm is used to perform data enhancement, and the formula is as follows:

[0028]

[0029] Among them, x new is the newly generated input sample, y new is the newly generated label, x Aand x B are two randomly selected input samples from the original dataset, y A and y B For x A and x B The two labels are , and λ is a random weight between 0 and 1, sampled from the Beta distribution, defined as:

[0030] λ~Beta(α,α)

[0031] Among them, α is a hyperparameter used to control the distribution of weights; if α is large, the value of λ is close to 0.5, and the generated new sample is more inclined to be a uniform mixture of the two samples; if α is small, the value of λ tends to be closer to 0 or 1, and the new sample will be closer to a certain original sample.

[0032] Furthermore, in step S60, the method for optimizing and improving the convolutional neural network YOLO-V8 is as follows: based on YOLOv8s, the C2f module in Backbone is replaced with the RepBlock module to enhance the network's feature extraction and feature expression capabilities; the ECA attention mechanism is introduced to adaptively adjust the channel weights so that the model focuses on important features, and the head part of YOLOv8s adopts a decoupled head structure; in terms of the loss function, YOLOv8 introduces the following TaskAlignedAssigner positive sample allocation strategy to improve the distribution of positive and negative samples and enhance the training effect; at the same time, a distributed focus loss is introduced to further optimize the calculation of the loss function:

[0033]

[0034] Among them, IoU(b,bgt) is the IoU of the predicted box b and the real box bgt, is the probability score of the predicted category c;

[0035]

[0036] Among them, L DFL is the distributed focal loss, w i is the corresponding weight term, which is used to adjust the importance of each interval, p i is the distribution probability predicted by the model, y i is the true distribution probability.

[0037] Furthermore, the initial learning rate of the convolutional neural network YOLO-V8 is 0.01, and the momentum and weight decay are set to 0.937 and 0.0005 respectively; the optimizer used for training is SGD, the batch size is 32, and the dropout rate is set to 0.5 to reduce overfitting.

[0038] The present invention discloses a method for nondestructive testing of stealth coating materials using multi-source data based on deep learning, which has the following beneficial effects:

[0039] 1. Compared with existing infrared thermal imaging detection technology, this invention uses deep learning technology to obtain a detection model, directly mapping the damage location to the optical image, making it easier for maintenance personnel to handle, reducing detection time and improving maintenance efficiency.

[0040] 2. Compared with existing technologies that use deep learning to detect object defects, this invention can retain its original advantages and also classify damage types, allowing front-line maintenance personnel to intuitively understand the damage situation and carry out repairs accordingly. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the method of the present invention;

[0042] Figure 2 This is the schematic diagram of the optimized YOLO-V8 model. DETAILED DESCRIPTION

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0044] refer to Figure 1 The present invention discloses a method for nondestructive testing of stealth coating materials using multi-source data based on deep learning, comprising the following steps:

[0045] Step S10: thermally excite the test piece to obtain an infrared thermal imaging image, and collect an optical image of the coating of the test piece at the same angle;

[0046] Step S20: pre-processing the infrared thermal imaging image;

[0047] Step S30: combining the acquired infrared thermal imaging image with the change of its corresponding optical image to determine the damage position and shape of the optical image of the damaged specimen and mark it;

[0048] Step S40: Calculate the mean and variance of the pixel values ​​of the four channels of the acquired thermal imaging image and optical image, and perform normalization processing on the brightness and contrast; after the processing, stack the infrared image as the fourth channel with the optical image;

[0049] Step S50: performing data enhancement on the infrared thermal imaging image and its corresponding optical image obtained in step S10 to expand the data samples, and dividing the expanded data samples into a training set and a test set in a ratio of 8:2;

[0050] Step S60: Input the training set obtained in step S40 into the convolutional neural network YOLO-V8 for training. During the training process, input the test set into the trained convolutional neural network YOLO-V8, monitor the detection accuracy of the test set on the convolutional neural network YOLO-V8 in real time, and optimize the convolutional neural network YOLO-V8 by adjusting the hyperparameters of the convolutional neural network YOLO-V8 to obtain an optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images;

[0051] Step S70: Acquire the infrared thermal imaging image and optical image to be tested in the manner of step S10, process them through steps S20, S30, and S40, and then input them into the optimized convolutional neural network YOLO-V8 model for detecting mixed images obtained in step S60 to detect whether the mixed image contains damage. If damage is present, obtain the location coordinates and area of ​​the damage, classify the damage level according to internal specifications, and map the damage to the optical image of the specimen.

[0052] To further optimize the technical solution, step S10 specifically includes the following steps:

[0053] Step S11: Setting up a light-stimulated infrared thermal imaging nondestructive testing system in the maintenance yard, using a halogen lamp as the excitation light source with a color temperature of 4500-5500K, fixing the infrared thermal imager on a lamp stand and aiming it at the area to be tested on the specimen;

[0054] Step S12: Place the optically excited infrared thermal imaging nondestructive testing system at a position 0.8 to 1.2 meters from the center of the test area in the normal direction, set the excitation current to 9A, the heating time to 2000ms, the power supply voltage to 230V, and the excitation signal to a phase-locked pulse signal to scan the area;

[0055] Step S13: using an industrial camera to collect an optical image of the coating from the normal direction of the center of the area to be measured;

[0056] Step S14: importing the acquired coating infrared thermal imaging image and optical image into the light-stimulated infrared thermal imaging nondestructive testing system.

[0057] To further optimize the technical solution, in step S20, the method for preprocessing the infrared thermal imaging image is as follows: the infrared thermal imaging image and the optical image obtained in S10 are size-aligned, the industrial camera resolution used is (1920, 1080, 3), and the infrared thermal imaging image resolution is (640, 480, 1); OpenCV is used for bilinear interpolation to adjust the infrared image to (1920, 1080, 1).

[0058] To further optimize the technical solution, in step S30, the method of marking the damage position and shape of the optical image is: frame the damage with a rectangular frame, and the marking format is [X min ,Y min ,X max ,Y max ], where X min , Y min is the coordinate of the upper left corner of the rectangle, X max , Y max The coordinates of the lower right corner of the rectangle.

[0059] To further optimize the technical solution, in step S40, the brightness and contrast are normalized using the following formula:

[0060]

[0061] Where μ is the mean of the four channel pixel values, σ is the variance of the four channel pixel values, f(x, y) represents the adjusted image pixel, g(x, y) represents the image pixel before adjustment, and (x, y) represents the pixel coordinate position.

[0062] To further optimize the technical solution, in step S50, data enhancement includes methods such as random image rotation, random cropping, random scaling, adding Gaussian noise, and mosaic (mosaic data enhancement). Based on the above methods, the Mixup algorithm is used to perform data enhancement. The formula is as follows:

[0063]

[0064] Among them, x new is the newly generated input sample, y new is the newly generated label, x A and x B are two randomly selected input samples from the original dataset, y A and y B For x A and x B The two labels are , and λ is a random weight between 0 and 1, sampled from the Beta distribution, defined as:

[0065] λ~Beta(α,α)

[0066] Among them, α is a hyperparameter used to control the distribution of weights; if α is large, the value of λ is close to 0.5, and the generated new sample is more inclined to be a uniform mixture of the two samples; if α is small, the value of λ tends to be closer to 0 or 1, and the new sample will be closer to a certain original sample.

[0067] To further optimize the technical solution, in step S60, the stacked images of the training set obtained in step S50 are input into the convolutional neural network YOLO-V8, and the prediction result of the convolutional neural network YOLO-V8 is output, that is, [Xmin, Ymin, Xmax, Ymax] is compared with the true mark [Xmin, Ymin, Xmax, Ymax] of the optical image of the infrared thermal imaging image, and the loss function is calculated. The gradient of the loss function is calculated, and then the model is corrected to reduce the loss function value. At least 1000 rounds of training are performed, and after every 10 rounds of training are completed, the test set is input into the trained convolutional neural network YOLO-V8, and the detection accuracy of the test set on the convolutional neural network YOLO-V8 is monitored in real time. By adjusting the hyperparameters, an optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images is obtained.

[0068] The present invention adopts an improved YOLOv8 model. YOLOv8s is a lightweight version of the YOLOv8 model, specifically designed to perform target detection and segmentation tasks in resource-limited environments. YOLOv8s is smaller in model scale than other versions (such as YOLOv8m and YOLOv8l), but still retains strong performance and high efficiency, making it suitable for scenarios such as embedded devices and mobile devices. The improved algorithm proposed in this invention replaces the C2f module in Backbone with the RepBlock module on the basis of YOLOv8s to enhance the network's feature extraction and feature expression capabilities. The content-aware feature recombination module (CARAFE) is used to replace the upsampling method to improve the segmentation accuracy of small targets. The ECA attention mechanism is introduced to adaptively adjust the channel weights, so that the model focuses on important features and improves the ability to extract key information in complex scenes. In addition, the head part of YOLOv8s is significantly different from YOLOv5. It adopts a decoupled head structure to improve detection accuracy and efficiency. In terms of loss function, YOLOv8 introduces the following TaskAlignedAssigner positive sample allocation strategy to improve the distribution of positive and negative samples and enhance the training effect; it also introduces the distributed focal loss (DistributionFocalLoss) to further optimize the calculation of the loss function.

[0069]

[0070] Among them, IoU(b,bgt) is the IoU of the predicted box b and the real box bgt, is the probability score of the predicted category c;

[0071]

[0072] Among them, L DFL is the distributed focal loss, w i is the corresponding weight term, which is used to adjust the importance of each interval, p i is the distribution probability predicted by the model, y i is the true distribution probability (the discrete distribution generated by the true coordinate values).

[0073] To further optimize the technical solution, the initial learning rate of the convolutional neural network YOLO-V8 is 0.01, the momentum and weight decay are set to 0.937 and 0.0005 respectively; the optimizer used for training is SGD, the batch size is 32, and the dropout rate is set to 0.5 to reduce overfitting.

[0074] Compared to existing infrared thermal imaging detection technology, this invention utilizes deep learning to generate a detection model, directly mapping damage locations to optical images. This facilitates processing by maintenance personnel, reduces detection time, and improves maintenance efficiency. While retaining the existing advantages of deep learning-based defect detection, this invention also categorizes damage types, allowing frontline maintenance personnel to intuitively understand the damage and conduct appropriate repairs.

[0075] Example

[0076] This embodiment provides a method for nondestructive testing of stealth coating materials using multi-source data based on deep learning. The specific process is as follows:

[0077] Step S10: Set up the light stimulation equipment, infrared thermal imaging equipment, and optical imaging equipment in the experimental field. The halogen lamp power is 2000W, the power supply current is 6A, the voltage is 230V, the spacing is 1 meter, and the height from the ground is 1.5 meters.

[0078] Move the test system to 1 meter in the normal direction of the test area. The excitation system outputs phase-locked excitation. The halogen lamp excitation time is 2000ms. The infrared thermal imager acquires the thermal imaging image of the measurement area in real time and stores the measurement results as the target thermal imaging data file.

[0079] Use an industrial camera to collect optical images of the coating at the same angle and store the acquired images in the system;

[0080] In this embodiment, 2,000 infrared thermal imaging images and optical images of the stealth coating are obtained, of which 1,400 are damaged and 600 are undamaged. The damage mainly includes physical scale damage and corrosion and oxidation damage. Physical scale damage manifests as shedding, cracking and other phenomena; corrosion and oxidation damage manifests as rust and blistering and other phenomena.

[0081] Step S20: Align the sizes of the infrared image and the optical image obtained in S10. The resolution of the industrial camera used is (1920, 1080, 3), and the resolution of the infrared image is (640, 480, 1). Use OpenCV for bilinear interpolation to adjust the infrared image to (1920, 1080, 1).

[0082] Step S30: Based on the stacked image obtained in step S20 and the changes in the corresponding infrared image and optical image, determine the damage position and shape of the optical image of the damaged specimen, and use a rectangular frame to frame the damage. The marking format is [X min ,Y min ,X max ,Y max ], where X min , Y min is the coordinate of the upper left corner of the rectangle, X max , Y max The coordinates of the lower right corner of the rectangle.

[0083] Step S40: Normalize the stacked image obtained in step S20: Calculate the mean μ and variance σ of the pixel values ​​of the four channels of the stacked image obtained in step S10 respectively. Let g(x, y) represent the image pixel before adjustment and f(x, y) represent the image pixel after adjustment. Use f(x, y) to normalize the brightness and contrast of the stacked image obtained in step S10:

[0084]

[0085] The image pixel value distribution is approximated to a standard normal distribution, where x and y represent the pixel coordinates. In this embodiment, the mean μ is [0.452, 0.436, 0.445] and the variance σ is [0.211, 0.217, 0.215]. The advantages of normalization are that it reduces the difficulty of model training, improves the model's generalization ability, and prevents gradient explosion.

[0086] Step S50: At the same time, data enhancement is performed on the stacked image obtained in step S20 to expand the data sample, and the number of the expanded data samples is divided into a training set and a test set in a ratio of 8:2. In this embodiment, random rotation, random cropping, random scaling, Gaussian noise addition, and mosaic data enhancement are used to enhance the sample data. Based on the above methods, the Mixup algorithm is used for data enhancement. The formula is as follows:

[0087]

[0088] Among them, x new is the newly generated input sample, y new is the newly generated label, x A and x B are two randomly selected input samples from the original dataset, y A and y B For x A and x B The two labels are , and λ is a random weight between 0 and 1, sampled from the Beta distribution, defined as:

[0089] λ~Beta(α,α)

[0090] Here, α is a hyperparameter that controls the distribution of weights. If α is large, λ is close to 0.5, and the generated new sample is more likely to be a uniform mixture of the two samples. If α is small, λ is closer to 0 or 1, and the new sample is closer to one of the original samples.

[0091] Step S60: Input the training set obtained in step S50 into the convolutional neural network YOLO-V8 for training. During the training process, input the test set into the trained convolutional neural network YOLO-V8. By adjusting the hyperparameters of the convolutional neural network YOLO-V8 (including n max and n min ), real-time monitoring of the test set detection accuracy on the convolutional neural network YOLO-V8, optimization of the convolutional neural network YOLO-V8, and the optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images are obtained. Figure 2 ;

[0092] The original convolutional neural network YOLO-V8 model has an initial learning rate of 0.01. The learning rate is changed using the "step" method. This example uses an exponential decay method to change the learning rate, which is defined as follows:

[0093] η t =η0e -λt

[0094] Where η0 is the initial learning rate, λ is the decay rate, and t is the number of training steps or epochs.

[0095] The training of the convolutional neural network model YOLOv8 is as follows: the infrared thermal imaging image of the training set is input into the convolutional neural network YOLOv8, and the prediction result of the convolutional neural network YOLOv8 is output, that is, [x min ,y min ,x max ,y max ](predict) the true mark of the optical image corresponding to the infrared thermal imaging image [x min ,y min ,x max ,y max ] for comparison, calculate the loss function, calculate the gradient of the loss function, and then modify the model to reduce the loss function value. Perform at least 1000 rounds of training, and input the test set into the trained convolutional neural network YOLOv8 after every 10 rounds of training. Monitor the detection accuracy of the test set on the convolutional neural network YOLOv8 in real time, and obtain the optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images by adjusting the hyperparameters.

[0096] YOLOv8, the convolutional neural network, is the latest version of the real-time object detection algorithm YOLO. Its basic framework integrates CSPNet and other lightweight structures, enabling efficient feature extraction. The YOLOv8 algorithm used in this experiment utilizes an improved backbone network, replacing the C2f module in the original Backbone with the RepBlock module to enhance the network's feature extraction and representation capabilities. The SPPF (Spatial Pyramid Pooling-Fast) module extracts multi-scale features, preserving efficient feature representation while reducing computational overhead. YOLOv8 downsamples through convolutions with a stride of 2, undergoing multiple downsampling steps. It also incorporates PAN (Path Aggregation Network) for feature fusion and performs multiple detections within the network architecture. The Content-Aware Feature Restructuring Module (CARAFE) replaces upsampling to improve segmentation accuracy for small objects. The ECA attention mechanism, introduced through adaptive adjustment of channel weights, focuses on important features, improving the ability to extract key information in complex scenes. YOLOv8 also uses an anchor-free detection mechanism, simplifying network training and inference and reducing hyperparameter complexity. Furthermore, YOLOv8 improves the loss function, introducing the TaskAlignedAssigner positive sample assignment strategy and distributed focal loss to further optimize the loss function calculation and achieve more accurate bounding box regression.

[0097] During the training of the YOLOv8 convolutional neural network model, using the SiLU (Swish) activation function can effectively prevent vanishing gradients and accelerate network training. By multiplying the input by the output of the sigmoid function, the SiLU function smoothes the activation function while maintaining nonlinear expression capabilities, thereby improving model accuracy.

[0098] The SiLU activation function is defined as:

[0099]

[0100] The loss function uses the TaskAlignedAssigner positive sample allocation strategy to improve the distribution of positive and negative samples and enhance the training effect. At the same time, the distribution focal loss is introduced to further optimize the calculation of the loss function as shown in the following formula:

[0101]

[0102] Among them, IoU(b,bgt) is the IoU of the predicted box b and the real box bgt, is the probability score of the predicted category c;

[0103]

[0104] Among them, L DFL is the distributed focal loss, w i is the corresponding weight term, which is used to adjust the importance of each interval, p i is the distribution probability predicted by the model, y i is the true distribution probability (the discrete distribution generated by the true coordinate values).

[0105] The purpose of network training is to find the parameter value that minimizes the loss function. In the present invention, the parameter value is updated, and the updating method is shown in the following formula:

[0106]

[0107] Where v t is the estimate of the first-order momentum of the gradient at step t; s t is the estimate of the second-order momentum of the gradient at step t; v t-1 is the estimated value of the first-order momentum of the gradient at step t-1; s t-1 is the estimate of the second-order momentum of the gradient at step t-1; β1 is the exponentially weighted average coefficient of the first-order momentum estimate, usually set to 0.9, which affects the smoothness of the gradient accumulation; β2 is the exponentially weighted average coefficient of the second-order momentum estimate, usually set to 0.999, which affects the smoothness of the mean square gradient; vt corret is the first-order momentum estimate of the corrected gradient; s t corret is the second-order momentum estimate of the corrected gradient; ω t is the model parameter at step t; ω t+1 is the parameter at step t+1; η is the learning rate, which controls the step size of parameter update; ∈ is a very small positive number (take ∈=10-8), which is used to prevent division by zero during update. It ensures that the denominator always has a certain value to avoid numerical instability; λ is the weight attenuation coefficient.

[0108] Step S70: Acquire the infrared thermal imaging image and optical image to be detected in the manner of step S10, process them through steps S20, S30, and S40, and then input them into the optimized convolutional neural network YOLO-V8 model for coating detection obtained in step S60 to detect whether the stealth coating contains damage. If it contains damage, obtain the location coordinates and damage area of ​​the damage, mark the damage on the optical image of the coating, and classify the damage level according to the degree of damage according to the internal grading specifications, so that the damage location and degree can be visualized on the optical image.

[0109] The present invention combines infrared thermal imaging technology with optical image detection technology to develop a multi-source data non-destructive testing solution based on a deep learning model. The detection process first uses a halogen lamp to perform phase-locked excitation on the coating area, uses an infrared thermal imager to collect thermal imaging images of the coating in real time, and uses an industrial camera to obtain optical images of the corresponding angles. The collected infrared thermal imaging and optical images are normalized for contrast, and a variety of data enhancement methods (such as random rotation, cropping, scaling, adding Gaussian noise and mosaic data enhancement) are used to expand the data samples. The processed data is divided into training set and test set in a ratio of 8:2, and then input into the convolutional neural network YOLO-V8 for training. During the model training process, the test set is input in real time and its detection accuracy is monitored. The model's hyperparameters (such as initial learning rate, momentum and weight decay, etc.) are adjusted for optimization, thereby gradually improving the accuracy of detection and the robustness of the model. In order to further improve the detection effect, the present invention introduces the distributed focal loss (Distribution FocalLoss) and the TaskAlignedAssigner positive sample allocation strategy in the training process of YOLO-V8 to optimize the distribution of positive and negative samples and improve the accuracy of bounding box regression. The optimized YOLO-V8 model is used to detect newly acquired infrared thermal imaging images and optical images, and directly outputs the coordinates of the defect area and its corresponding area through the deep learning model. The detection results map the location of the damage to the optical image and classify the damage according to the preset internal standards, thereby helping front-line maintenance personnel to intuitively judge the severity of the defect and then formulate corresponding repair plans.

[0110] Compared with traditional detection methods, the present invention has significant advantages. Traditional detection methods such as visual inspection, eddy current detection, and ultrasonic detection all have their own limitations. The present invention effectively overcomes these technical obstacles through deep learning technology and multi-source data fusion, greatly improving detection accuracy and efficiency. At the same time, by combining infrared thermal imaging images with internal grading specifications, the present invention realizes damage grading of different types of defects (such as coating peeling, cracks, and corrosion, etc.), which not only simplifies the maintenance process, but also significantly improves maintenance efficiency and reduces the possibility of human subjective misjudgment.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-source data nondestructive testing method for stealth coating materials based on deep learning, characterized in that: The steps include: Step S10: thermally excite the test piece to obtain an infrared thermal imaging image, and collect an optical image of the coating of the test piece at the same angle; Step S20: pre-processing the infrared thermal imaging image; Step S30: combining the acquired infrared thermal imaging image with the change of its corresponding optical image to determine the damage position and shape of the optical image of the damaged specimen and mark it; Step S40: Calculate the mean and variance of the pixel values ​​of the four channels of the acquired thermal imaging image and optical image, and perform normalization processing on the brightness and contrast; After processing, the infrared image is used as the fourth channel and stacked with the optical image; Step S50: performing data enhancement on the infrared thermal imaging image and its corresponding optical image obtained in step S10 to expand the data samples, and dividing the expanded data samples into a training set and a test set in a ratio of 8:2; Step S60: Input the training set obtained in step S40 into the convolutional neural network YOLO-V8 for training. During the training process, input the test set into the trained convolutional neural network YOLO-V8, monitor the detection accuracy of the test set on the convolutional neural network YOLO-V8 in real time, and optimize the convolutional neural network YOLO-V8 by adjusting the hyperparameters of the convolutional neural network YOLO-V8 to obtain an optimized convolutional neural network YOLO-V8 model for detecting infrared thermal imaging images; Step S70: Acquire the infrared thermal imaging image and optical image to be tested in the manner of step S10, process them through steps S20, S30, and S40, and then input them into the optimized convolutional neural network YOLO-V8 model for detecting mixed images obtained in step S60 to detect whether the mixed image contains damage. If damage is present, obtain the location coordinates and area of ​​the damage, classify the damage level according to internal specifications, and map the damage to the optical image of the specimen.

2. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 1, characterized in that: Step S10 specifically includes the following steps: Step S11: Setting up a light-stimulated infrared thermal imaging nondestructive testing system in the maintenance yard, using a halogen lamp as the excitation light source with a color temperature of 4500-5500K, fixing the infrared thermal imager on a lamp stand and aiming it at the area to be tested on the specimen; Step S12: Place the optically excited infrared thermal imaging nondestructive testing system at a position 0.8 to 1.2 meters from the center of the test area in the normal direction, set the excitation current to 9A, the heating time to 2000ms, the power supply voltage to 230V, and the excitation signal to a phase-locked pulse signal to scan the area; Step S13: using an industrial camera to collect an optical image of the coating from the normal direction of the center of the area to be measured; Step S14: importing the acquired coating infrared thermal imaging image and optical image into the light-stimulated infrared thermal imaging nondestructive testing system.

3. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 2, characterized in that: In step S20, the method for preprocessing the infrared thermal imaging image is as follows: the infrared thermal imaging image and the optical image obtained in S10 are size-aligned. The resolution of the industrial camera used is (1920, 1080, 3), and the resolution of the infrared thermal imaging image is (640, 480, 1); bilinear interpolation is performed using OpenCV to adjust the infrared image to (1920, 1080, 1).

4. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 3, characterized in that: In step S30, the method of marking the damage position and shape of the optical image is to frame the damage with a rectangular frame, and the marking format is [X min ,Y min ,X max ,Y max ], where X min , Y min is the coordinate of the upper left corner of the rectangle, X max , Y max The coordinates of the lower right corner of the rectangle.

5. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 4, characterized in that: In step S40, the brightness and contrast are normalized using the following formula: Where μ is the mean of the four channel pixel values, σ is the variance of the four channel pixel values, f(x, y) represents the adjusted image pixel, g(x, y) represents the image pixel before adjustment, and (x, y) represents the pixel coordinate position.

6. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 5, characterized in that: In step S50, the Mixup algorithm is used to perform data enhancement, and the formula is as follows: Among them, x new is the newly generated input sample, y new is the newly generated label, x A and x B are two randomly selected input samples from the original dataset, y A and y B For x A and x B The two labels are , and λ is a random weight between 0 and 1, sampled from the Beta distribution, defined as: λ~Beta(α,α) Among them, α is a hyperparameter used to control the distribution of weights; if α is large, the value of λ is close to 0.5, and the generated new sample is more inclined to be a uniform mixture of the two samples; if α is small, the value of λ tends to be closer to 0 or 1, and the new sample will be closer to a certain original sample.

7. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 6, characterized in that: In step S60, the method for optimizing and improving the convolutional neural network YOLO-V8 is as follows: based on YOLOv8s, the C2f module in Backbone is replaced with the RepBlock module to enhance the network's feature extraction and feature expression capabilities; the ECA attention mechanism is introduced to adaptively adjust the channel weights so that the model focuses on important features, and the head part of YOLOv8s adopts a decoupled head structure; in terms of the loss function, YOLOv8 introduces the following TaskAlignedAssigner positive sample allocation strategy to improve the distribution of positive and negative samples and enhance the training effect; at the same time, a distributed focus loss is introduced to further optimize the calculation of the loss function: Among them, IoU(b,bgt) is the IoU of the predicted box b and the real box bgt, is the probability score of the predicted category c; Among them, L DFL is the distributed focal loss, w i is the corresponding weight term, which is used to adjust the importance of each interval, p i is the distribution probability predicted by the model, y i is the true distribution probability.

8. The method for nondestructive testing of stealth coating materials using multi-source data based on deep learning according to claim 7, characterized in that: The initial learning rate of the convolutional neural network YOLO-V8 is 0.01, and the momentum and weight decay are set to 0.937 and 0.0005 respectively. The optimizer used for training is SGD, the batch size is 32, and the dropout rate is set to 0.5 to reduce overfitting.

Images