An infrared thermal imaging composite material defect detection method based on deep learning

Abstract

The application discloses an infrared thermal imaging composite material defect detection method based on deep learning, and relates to the technical field of nondestructive testing, and comprises the following steps: a multi-time scale long pulse thermal excitation is used to heat a measured composite material sample in advance, the temperature change process of the sample surface is collected through an infrared thermal imager, and a plurality of groups of infrared thermal image sequences of different time scales are obtained. The application takes a pixel-level temperature time sequence as input, can directly output a pixel-level defect detection result, has excellent detection effect on carbon fiber reinforced composite material subsurface hole type, low contrast, small size and low depth-diameter ratio defects, compared with a traditional detection mode relying on artificial experience, can automatically complete defect identification, feature extraction and result generation, effectively improves detection efficiency and reduces human error, and can obviously reduce missed detection and false detection in a complex background and a weak thermal response scene.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and more specifically, to a method for detecting defects in composite materials using infrared thermal imaging based on deep learning. Background Technology

[0002] Carbon fiber reinforced polymer composites, with their outstanding advantages such as high specific strength, high specific stiffness, and good corrosion resistance, have been widely used in important fields such as aerospace, rail transportation, and energy equipment. However, these materials are prone to internal defects such as voids, delamination, and debonding during manufacturing, assembly, and use. These defects are mostly located inside the material or in subsurface areas, making them highly concealed and difficult to detect in their early stages. In severe cases, they can directly reduce the structural load-bearing capacity and threaten the safety of equipment use. Infrared thermal imaging non-destructive testing technology, with its non-contact, fast detection speed, and intuitive imaging, is the mainstream technology for defect detection in composite materials. Among them, long-pulse thermal imaging technology applies thermal excitation to the sample for a certain duration and continuously records the change of material surface temperature over time. By utilizing the difference in thermal conductivity between defective and non-defective areas, it can present the abnormal thermal response of defective areas on the thermal image, providing a basis for defect identification.

[0003] Existing infrared thermal imaging defect detection methods mostly rely on manual observation, traditional image processing, or single deep learning network structures, which have significant limitations. Manual analysis is inefficient, and the detection results are greatly affected by the operator's experience, resulting in poor consistency and stability. Traditional image processing methods have limited ability to identify weak defects, small-sized defects, and low-contrast defects, making it difficult to meet the requirements of high-precision detection. Although some existing deep learning methods can improve detection performance to a certain extent, they fail to fully integrate information from different time scales in the infrared thermal sequence and lack the ability to extract local discriminative features. This leads to frequent missed detections and false detections in complex backgrounds, low aspect ratio defects, or scenarios with weak thermal response.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a deep learning-based infrared thermal imaging method for detecting defects in composite materials. This method overcomes the shortcomings of existing infrared thermal imaging composite material defect detection technologies, such as insufficient expression of defect thermal response features, difficulty in identifying weak defects, poor robustness in complex backgrounds, and reliance on human experience in the detection process.

[0006] The technical solution of this invention is implemented as follows:

[0007] A method for detecting defects in composite materials using infrared thermal imaging based on deep learning, comprising the following steps:

[0008] The composite material sample under test was preheated with long-pulse thermal excitation at multiple time scales, and the surface temperature change process of the sample was acquired by an infrared thermal imager to obtain multiple sets of infrared thermal image sequences at different time scales.

[0009] Temperature time series are extracted pixel by pixel from infrared thermal image sequences and preprocessed to obtain multi-scale standard length input samples;

[0010] A deep learning network structure including a multi-scale temporal fusion module, a local temporal focusing module, a channel focusing module, and a classification head is constructed as a defect detection network;

[0011] The obtained multi-scale standard-length input samples are paired with pixel-level labeled labels to form training samples with pixel-level labels, which are then input into the constructed defect detection network. The defect region pixels are set to 1 and the non-defect region pixels are set to 0 as supervision labels. The cross-entropy loss function is used to iteratively optimize the network parameters to obtain the trained defect detection model.

[0012] The multi-scale standard length input sample obtained by processing the composite material to be tested is input into the trained defect detection model. The model classifies each pixel as defective or defect-free. The classification results are reconstructed according to the original spatial location to generate a pixel-level defect detection result map.

[0013] The acquisition of multiple sets of infrared thermal image sequences at different time scales includes infrared thermal image sequences acquired at each time scale, represented as a three-dimensional matrix:

[0014] ;

[0015] Among them, three-dimensional matrix , Indicates the spatial resolution of an infrared image. This represents the number of image frames in the time dimension.

[0016] The step of extracting the temperature time series pixel by pixel from the infrared thermal image sequence includes: extracting the temperature time series pixel by pixel. To characterize the dynamic thermal response characteristics of the tested composite material at various spatial locations under thermal excitation, for spatial locations of... The pixel and its corresponding temperature time series are represented as:

[0017] ;

[0018] In the formula, and These represent the horizontal and vertical coordinates of a pixel in the image, respectively.

[0019] The preprocessing includes sequentially performing moving average filtering for noise reduction and normalization, and then unifying temperature time series of different lengths into a standard length through interpolation resampling to obtain multi-scale standard length input samples.

[0020] The multi-scale time fusion module is used to encode the features of input sequences at multiple time scales, and then concatenate and fuse the features at each scale to comprehensively utilize the defect thermal response features at different thermal excitation stages and thermal diffusion intervals, as expressed as:

[0021] ;

[0022] In the formula, This represents input at different time scales. Representing local features, This represents the number of output channels of a one-dimensional convolution. For convolution operations, For batch normalization, For activation function, This is for splicing operations.

[0023] The local temporal focusing module is composed of ACmix units, including a local attention branch and a convolution branch. The local attention branch is used to model the feature correlation within a local time range, and the convolution branch is used to extract local temporal convolutional features. The outputs of the two branches are fused to obtain local temporal enhancement features.

[0024] For the convolution branch, the query tensor will be... Key tensors Sum tensor By splicing along the head dimension, we obtain the connection feature tensor. :

[0025] ;

[0026] In the formula, Represents one-dimensional convolution. Indicates a reshaping operation;

[0027] The connection feature tensor is then input into a two-dimensional convolutional layer for channel adjustment to obtain an intermediate feature tensor. Next, the intermediate feature tensor is reshaped into a two-dimensional tensor and coupled with a learnable tensor. Perform element-wise multiplication:

[0028] ;

[0029] In the formula, Represents two-dimensional convolution. This represents element-wise multiplication.

[0030] Finally, the result of element-wise multiplication is fed into a one-dimensional convolutional layer to restore the target channel dimension, yielding the output of the convolutional branch:

[0031] ;

[0032] Among them, for the attention branch, the key Perform filling processing and extract the kernel size through unfolding operations. Local patch The obtained tensor is reconstructed into a four-dimensional structure, with each dimension corresponding to the number of heads. Channel dimensions at each end kernel size and output time length This is composed of components that enable efficient computation of local self-attention, expressed as:

[0033] ;

[0034] In the formula, The bond representing the reconstruction into a four-dimensional structure , This indicates a fill operation. This indicates an expand operation.

[0035] The channel focusing module sequentially performs one-dimensional convolutional channel expansion, group convolutional feature extraction, batch normalization, channel attention weighting, one-dimensional convolutional channel recovery, and residual connection to achieve adaptive recalibration of channel features, enhance defect discrimination features, and suppress irrelevant interference features.

[0036] The processing procedure is as follows:

[0037] ;

[0038] In the formula, This represents the output of the local time focusing module. This indicates a group convolution operation. This indicates a batch normalization operation. This indicates a channel attention-weighted operation.

[0039] The classification head includes a first linear layer, a Dropout layer, and a second linear layer, which are used to output the probability that each pixel belongs to the defective class or the non-defective class.

[0040] The refined feature map is flattened into a one-dimensional feature vector, and then sequentially input into the first linear layer, the Dropout layer, and the second linear layer for processing. The first linear layer is used to project the input features into a low-dimensional feature space, the Dropout layer is used to reduce the risk of overfitting, and the second linear layer is used to output the scores of the two categories.

[0041] The process can be represented as follows:

[0042] ;

[0043] In the formula, This indicates the output of the channel focusing module. This indicates categorized output. This indicates a linear layer.

[0044] Another aspect of the present invention:

[0045] A deep learning-based infrared thermal imaging composite material defect detection system is used to execute the aforementioned deep learning-based infrared thermal imaging composite material defect detection method, comprising: a thermal excitation unit, an infrared image acquisition unit, a data preprocessing unit, a network construction unit, a model training unit, an inference detection unit, and a result output unit.

[0046] The beneficial effects of this invention are:

[0047] This invention constructs a multi-scale temporal fusion mechanism to comprehensively utilize the thermal response features of defects in different thermal excitation stages and thermal diffusion intervals, significantly improving the ability to characterize the dynamic thermal behavior of defects. Simultaneously, it introduces a hybrid feature extraction mechanism combining a local temporal focusing module with convolution and self-attention, greatly enhancing the network's recognition accuracy for weak defects, blurred edge defects, and local abnormal thermal response regions. Combined with a channel focusing module, it achieves adaptive recalibration of channel features, further strengthening defect-related discrimination features and suppressing invalid interference information. This method takes pixel-level temperature time series as input and can directly output pixel-level defect detection results. It exhibits excellent detection performance for subsurface pores, low contrast, small size, and low aspect ratio defects in carbon fiber reinforced composite materials. Compared with traditional detection methods that rely on human experience, it can automatically complete defect identification, feature extraction, and result generation, effectively improving detection efficiency and reducing human error. In complex backgrounds and weak thermal response scenarios, it can significantly reduce missed detections and false detections. Overall, its detection accuracy, robustness, and engineering practical value are significantly superior to traditional image processing and single deep learning methods. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of a detection scenario for an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the carbon fiber plate dimensions and defect parameters according to an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of a defect detection network for an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of a local temporal focusing module in an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram of the channel focusing module of an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention;

[0054] Figure 6 This is a pixel-level defect detection diagram of an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention, wherein (a) is a true value map of the actual defect distribution, and (b) is a map of the predicted output results of the model of the present invention.

[0055] Figure 7 This is a schematic diagram comparing the detection results of different methods of infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention, wherein (a) is the truth map, (b) is the detection result map of the method of the present invention, (c) is the detection result map of 1DCNN method, (d) is the detection result map of LSTM method, and (e) is the detection result map of Transformer method;

[0056] Figure 8 This is a schematic diagram of the ablation experiment results of an infrared thermal imaging composite material defect detection method based on deep learning according to an embodiment of the present invention. It shows a comparison of defect detection results using single-scale and multi-scale temporal fusion input, where (a) is the actual defect image, (b) is the multi-scale temporal fusion detection result image, (c) is the T1 single-scale detection result image, (d) is the T2 single-scale detection result image, (e) is the T3 single-scale detection result image, and (f) is the T4 single-scale detection result image. Detailed Implementation

[0057] 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 embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0058] According to an embodiment of the present invention, a method for detecting defects in composite materials based on infrared thermal imaging using deep learning is provided.

[0059] The infrared thermal imaging composite material defect detection method based on deep learning according to embodiments of the present invention is applicable to the detection of subsurface defects in carbon fiber reinforced composite materials, especially for defects such as pores, low contrast, small size, and low aspect ratio. The method takes pixel temperature time series as input, completes pixel classification through multi-scale temporal fusion and local feature enhancement networks, and maps to generate pixel-level defect detection results, including the following steps:

[0060] Step S1: Acquire infrared thermal image sequence. The carbon fiber reinforced composite material sample is heated using at least two time scales of long pulse thermal excitation. The infrared thermal imager simultaneously acquires the surface temperature change process during the heating stage and part of the cooling stage to obtain an infrared thermal image sequence.

[0061] Specifically, infrared thermal imagers record the surface temperature changes of the sample during the heating and partial cooling phases. The infrared thermal image sequence acquired at each time scale can be represented as a three-dimensional matrix:

[0062] ;

[0063] Among them, three-dimensional matrix , Indicates the spatial resolution of an infrared image. This represents the number of image frames in the time dimension, i.e., the length of the acquired temperature sequence. The duration of thermal excitation and the length of the sampling sequence can differ at different time scales; therefore, the corresponding infrared thermal image sequences have different lengths in the time dimension. Specific experimental parameters are shown in Table 1.

[0064] Table 1. Thermal image sequence acquisition settings at different time scales

[0065] Serial Number Heating time (s) Sampling frequency (Hz) Data collection duration (s) Total frames T1 10 25 24 600 T2 15 25 30 750 T3 20 12.5 52 650 T4 25 12.5 60 750

[0066] In this embodiment, the detection device uses a Philips QVF137 2000W halogen lamp as the thermal excitation source, and an FLIRA40MWIR infrared thermal imager with a resolution of 320×240, a temperature sensitivity of 80mK, and a detection band of 7.5~13μm. The overall device layout is as follows: Figure 1 As shown; the tested carbon fiber plate measures 150mm × 100mm × 3mm, with three sets of defects of different diameters inside, and defect depths ranging from 0.5mm to 2.0mm. The sample dimensions and defect parameters are as follows. Figure 2 As shown.

[0067] Step S2 involves extracting the temperature time series pixel by pixel from the infrared thermal image sequence and performing preprocessing to obtain multi-scale standard-length input samples, as detailed below:

[0068] From the above three-dimensional matrix Extracting temperature time series pixel by pixel This is used to characterize the dynamic thermal response of the tested composite material at various spatial locations under thermal excitation. For spatial locations... The pixel and its corresponding temperature time series are represented as:

[0069] ;

[0070] In the formula, and These represent the horizontal and vertical coordinates of a pixel in the image, respectively.

[0071] Infrared thermal imaging acquisition is susceptible to environmental noise, instrument noise, and uneven emissivity of material surfaces, necessitating smoothing of the temperature time series extracted pixel by pixel. A moving average filtering method is used to denoise the temperature time series, with the filter window size set to [value missing]. , Values ​​such as 3 and 5 can be used. The smoothed temperature sequence can be represented as:

[0072] ;

[0073] In the formula, This represents the smoothed temperature sequence.

[0074] To mitigate the impact of varying acquisition conditions, initial temperatures of different pixels, and differences in absolute temperature ranges on subsequent feature extraction, the smoothed temperature time series was further normalized. The normalized temperature series is represented as follows:

[0075] ;

[0076] In the formula, This represents the normalized temperature time series. Since temperature time series obtained at different time scales have different lengths, to achieve a unified input, it is necessary to resample the temperature time series of different lengths into a standard series of consistent length using interpolation methods. Let the length of the standard series be... Then the first The standard sequence after resampling at each time scale is represented as follows:

[0077] ;

[0078] In the formula, Represents pixels In the The standard sequence obtained by interpolation and resampling at each time scale This represents the interpolation resampling function at the corresponding time scale. This indicates the length of the original sequence.

[0079] Therefore, for each pixel, multiple standard input sequences of uniform length at different time scales can be obtained:

[0080] ;

[0081] The aforementioned standard sequences at different time scales together constitute the input sample for this pixel.

[0082] Step S3: Construct a deep learning network structure that includes a multi-scale temporal fusion module, a local temporal focusing module, a channel focusing module, and a classification head as a defect detection network;

[0083] The defect detection network is input with multiple input sequences at different time scales corresponding to each pixel. The overall network structure is as follows: Figure 3 As shown, the defect detection network includes a multi-scale time fusion module, a local time focusing module, a channel focusing module, and a classification head, which are used to extract multi-scale dynamic features, local time correlation features, and channel discrimination features from the temperature time series.

[0084] Among them, the multi-scale time fusion module is used to jointly encode and fuse features of input sequences at multiple different time scales, so as to make full use of defect thermal response information in different thermal excitation stages and different thermal diffusion intervals.

[0085] Specifically, multiple standard sequences corresponding to pixels are input into the corresponding feature extraction units to obtain feature representations at multiple time scales. These features from different time scales are then fused to obtain fused features.

[0086] ;

[0087] In the formula, This represents input at different time scales. Representing local features, This represents the number of output channels of a one-dimensional convolution. For convolution operations, For batch normalization, For activation function, This is for splicing operations.

[0088] The multi-scale time fusion processing described above can effectively enhance the model's ability to comprehensively represent thermal response information at different time scales and improve its ability to identify the dynamic thermal behavior of defects.

[0089] The local time focusing module is used to extract local time-related features from the input sequence to enhance the model's ability to perceive local thermal anomaly change patterns.

[0090] In this embodiment, the local temporal focusing module is composed of an ACmix unit, which includes a local attention branch and a convolutional branch. The local attention branch is used to model the correlation within a local temporal range, and the convolutional branch is used to extract temporal convolutional features in the local neighborhood. Its structure is as follows: Figure 4 As shown.

[0091] For the convolution branch, the query tensor will be... Key tensors Sum tensor By splicing along the head dimension, we obtain the connection feature tensor. :

[0092] ;

[0093] In the formula, Represents one-dimensional convolution. This indicates a reshaping operation.

[0094] The connection feature tensor is then input into a two-dimensional convolutional layer for channel adjustment to obtain an intermediate feature tensor. Next, the intermediate feature tensor is reshaped into a two-dimensional tensor and coupled with a learnable tensor. Perform element-wise multiplication:

[0095] ;

[0096] In the formula, Represents two-dimensional convolution. This indicates element-wise multiplication.

[0097] Finally, the result of element-wise multiplication is fed into a one-dimensional convolutional layer to restore the target channel dimension, yielding the output of the convolutional branch:

[0098] .

[0099] For the attention branch, firstly, the key... Perform filling processing and extract the kernel size through unfolding operations. Local patch The obtained tensor is reconstructed into a four-dimensional structure, with each dimension corresponding to the number of heads. Channel dimensions at each end kernel size and output time length This is achieved by combining components to enable efficient computation of local self-attention. The specific process can be expressed as the following formula:

[0100] ;

[0101] In the formula, The bond representing the reconstruction into a four-dimensional structure , This indicates a fill operation. This indicates an expand operation.

[0102] Standard self-attention mechanisms are inherently permutation-invariant, meaning they cannot perceive the sequential relationships or relative positional information between time steps. Therefore, a positional encoding mechanism is introduced to capture these important temporal dynamics, incorporating a temporal inductive bias into the standard self-attention model. Specifically, in a time dimension of length... Construct a linear space from -1 to 1 within the range to represent the bond. and query The location information is then processed using a one-dimensional convolutional layer to convert the coordinate tensor into a dimension of [dimensional value missing]. The feature representation, namely and This is to meet the dimensionality matching requirements for subsequent attention calculations. Simultaneously, the key... Position encoding Adopted and The same unfolding operation is used to extract local location patches that match the local attention mechanism. Its representation is as follows:

[0103] ;

[0104] In the formula, This represents a linear space construction operation. Represents the key after dimensional reconstruction Location encoding.

[0105] Then, a position-encoded local attention mechanism is introduced. It can be represented as follows:

[0106] ;

[0107] Among them, query and position encoding Through dimensional expansion and broadcast operations, it is mapped to a size of The tensor, and the key position They are also broadcast to the same dimension to ensure the consistency of the above operations in terms of dimension.

[0108] Finally, for the expanded values Perform a weighted summation along the kernel dimension and reshape the resulting tensor back to the original dimension. :

[0109] ;

[0110] In the formula, express and The result of weighted summation This is the output of the attention branch.

[0111] After the weighted fusion of the results from the attention branch and the convolution branch, batch normalization is required. In addition, the normalized features need to be compared with the original input features. The residual connections are implemented by combining the results, and finally the GeLU activation function is applied to obtain the output of the ACmix module. :

[0112] ;

[0113] in and These are the weights for the self-attention branch and the convolution branch, respectively.

[0114] The channel focusing module is located after the local time focusing module and is used to adaptively recalibrate the feature responses on different channels to enhance the discriminative features related to defects and suppress irrelevant features.

[0115] Specifically, the output features of the local temporal focusing module are first expanded through one-dimensional convolution, then local features are extracted through group convolution, and batch normalization is applied to improve training stability. Next, the obtained features are input into the channel attention unit, and the responses of each channel are weighted and adjusted. Then, one-dimensional convolution is used to restore the channel dimensions to their original dimensions, and the residuals are added to the input features to obtain the enhanced features.

[0116] The processing procedure can be represented as follows:

[0117] ;

[0118] In the formula, This represents the output of the local time focusing module. Represents a group convolution operation, and represents a batch normalization operation. This indicates a channel attention-weighted operation.

[0119] The aforementioned channel focusing module effectively highlights defect-related channel features while preserving the original feature information, improving the gradient propagation process and enhancing network training stability and feature representation capabilities. Its structure is as follows: Figure 5 As shown.

[0120] Specifically, it also includes: a classification head output, which, after completing the above feature extraction, inputs the enhanced features corresponding to each pixel into the classification head for binary classification; the classification head is used to determine whether the pixel belongs to the defect-free class or the defective class.

[0121] Specifically, the refined feature map is first flattened into a one-dimensional feature vector, and then sequentially fed into a first linear layer, a Dropout layer, and a second linear layer for processing. The first linear layer projects the input features into a low-dimensional feature space, the Dropout layer reduces the risk of overfitting, and the second linear layer outputs the scores for the two classes.

[0122] The process can be represented as follows:

[0123] ;

[0124] In the formula, This indicates the output of the channel focusing module. This indicates categorized output. This indicates a linear layer.

[0125] Step S4: Pair the obtained multi-scale standard length input samples with pixel-level annotation labels to form training samples with pixel-level annotations and input them into the constructed defect detection network. Use defect region pixels as 1 and non-defect region pixels as 0 as supervision labels. Use cross-entropy loss function to iteratively optimize network parameters and obtain the trained defect detection model.

[0126] Specifically, during the training phase, pixel-level labeled training samples are used for supervised training of the defect detection network. The label of each pixel-level sample is determined based on the actual defect distribution information of the sample, wherein pixels located in defective regions are labeled as 1, and pixels located in defect-free regions are labeled as 0.

[0127] In this embodiment, the cross-entropy loss function is used to optimize the model to measure the difference between the model's predicted probability distribution and the true label. For batch sizes of... The average cross-entropy loss of the training samples can be expressed as:

[0128] ;

[0129] in, It is the first The true labels of each sample are determined. By minimizing the above loss function, the parameters in the defect detection network are iteratively updated, allowing the model to gradually learn the mapping relationship between temperature time series and defect categories at different time scales, thereby improving the ability to identify defect regions in composite materials.

[0130] Step S5: Input the multi-scale standard length input sample of the composite material to be tested, obtained after processing, into the trained defect detection model. The model classifies each pixel as defective or defect-free. The classification results are reconstructed according to the original spatial location to generate a pixel-level defect detection result map.

[0131] Specifically, the composite material to be tested is subjected to the same thermal excitation, image acquisition and preprocessing process as in steps S1 and S2 to obtain the corresponding multi-scale standard length input sample, which is then directly input into the defect detection model trained in step S4.

[0132] The defect detection model performs a binary classification of each pixel as defective / defect-free, and maps and reconstructs the classification results of all pixels according to their original spatial locations to generate a pixel-level defect detection result map. The detection results are as follows: Figure 6 As shown, (a) is the true value map of the actual distribution of defects, and (b) is the prediction output map of the model of the present invention.

[0133] In addition, in another embodiment, a deep learning-based infrared thermal imaging composite material defect detection system is disclosed, which is used to execute the deep learning-based infrared thermal imaging composite material defect detection method described above, including: a thermal excitation unit, an infrared image acquisition unit, a data preprocessing unit, a network construction unit, a model training unit, an inference detection unit, and a result output unit.

[0134] Using the above method, the experimental results are as follows:

[0135] After training, the model achieved an average F1 score of 0.84775 and an average IoU of 0.75213 in the training region, and an average F1 score of 0.80698 and an average IoU of 0.68438 in the independent test region.

[0136] To further verify the effectiveness of the method of the present invention, the prediction results of the present invention on the test area were compared with the detection results of one-dimensional convolutional neural network (1D-CNN), Transformer encoder architecture, and long short-term memory network (LSTM). The comparison results are as follows: Figure 7 As shown, (a) is the ground truth image, (b) is the method of this invention, (c) is 1DCNN, (d) is LSTM, and (e) is Transformer. The results show that the method of this invention exhibits good performance in terms of overall detection accuracy, noise suppression capability, and boundary false alarm control, especially in challenging defect scenarios with low aspect ratios and weak thermal features, where the number of missed detections is significantly reduced.

[0137] Meanwhile, to verify the effectiveness of the multi-timescale feature fusion strategy in this invention, further ablation experiments were conducted. The results of the ablation experiments are as follows: Figure 8As shown, the defect detection results are compared using single-scale and multi-scale temporal fusion inputs, where (a) is the ground truth image, (b) is the multi-scale temporal fusion image, (c) is T1, (d) is T2, (e) is T3, and (f) is T4. Experimental results show that compared to single-scale input, the multi-scale fusion method significantly improves detection performance, with an average F1 score improvement of 0.11 and an average IoU improvement of 0.12. Furthermore, the convolutional branch and self-attention branch in the local temporal focusing module have a significant complementary effect; their synergy further enhances the identification ability of difficult-to-detect defects and effectively reduces boundary false alarms.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art, upon considering the disclosure in the specification and embodiments, will readily conceive of other embodiments of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0139] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A deep learning-based infrared thermography composite material defect detection method, characterized in that, Includes the following steps: The composite material sample under test was preheated with long-pulse thermal excitation at multiple time scales, and the surface temperature change process of the sample was acquired by an infrared thermal imager to obtain multiple sets of infrared thermal image sequences at different time scales. Temperature time series are extracted pixel by pixel from infrared thermal image sequences and preprocessed to obtain multi-scale standard length input samples; A deep learning network structure including a multi-scale temporal fusion module, a local temporal focusing module, a channel focusing module, and a classification head is constructed as a defect detection network; The obtained multi-scale standard-length input samples are paired with pixel-level labeled labels to form training samples with pixel-level labels, which are then input into the constructed defect detection network. The defect region pixels are set to 1 and the non-defect region pixels are set to 0 as supervision labels. The cross-entropy loss function is used to iteratively optimize the network parameters to obtain the trained defect detection model. The multi-scale standard length input sample obtained by processing the composite material to be tested is input into the trained defect detection model. The model classifies each pixel as defective or defect-free. The classification results are reconstructed according to the original spatial location to generate a pixel-level defect detection result map.

2. The deep learning-based infrared thermography composite defect detection method of claim 1, wherein, The acquisition of multiple sets of infrared thermal image sequences at different time scales includes infrared thermal image sequences acquired at each time scale, represented as a three-dimensional matrix: ； wherein the three-dimensional matrix , denotes the spatial resolution of the infrared image, denotes the number of image frames in the time dimension.

3. The method for detecting defects in composite materials based on deep learning using infrared thermal imaging according to claim 2, characterized in that, The pixel-by-pixel extraction of the temperature time sequence from the infrared thermal image sequence comprises pixel-by-pixel extraction of the temperature time sequence In order to characterize the dynamic thermal response characteristics of each spatial position of the measured composite material under the action of thermal excitation, for a pixel point with a spatial position The corresponding temperature time sequence is represented as: ； In the formula, and respectively represent the horizontal coordinate and the vertical coordinate of the pixel point in the image.

4. The deep learning-based infrared thermography composite defect detection method of claim 3, wherein, The preprocessing includes sequentially performing moving average filtering for noise reduction and normalization, and then unifying temperature time series of different lengths into a standard length through interpolation resampling to obtain multi-scale standard length input samples.

5. The deep learning-based infrared thermography composite defect detection method of claim 1, wherein, The multi-scale time fusion module is used to encode the features of the input sequences at multiple time scales, and then concatenate and fuse the features of each scale to comprehensively utilize the defect thermal response features of different thermal excitation stages and thermal diffusion intervals, as expressed as: ； wherein, denotes different time scale inputs, denotes local features, denotes the output channel number of one-dimensional convolution, is a convolution operation, is batch normalization, is an activation function, is a concatenation operation.

6. The deep learning-based infrared thermography composite defect detection method of claim 5, wherein, The local temporal focusing module is composed of ACmix units, including a local attention branch and a convolution branch. The local attention branch is used to model the feature correlation within a local time range, and the convolution branch is used to extract local temporal convolution features. The outputs of the two branches are fused to obtain local temporal enhancement features. wherein, for the convolution branch, the query tensor , the key tensor and the value tensor are concatenated along the head dimension to obtain the concatenated feature tensor : ； In the formula, Represents one-dimensional convolution. Indicates a reshaping operation; The connection feature tensor is then input into a two-dimensional convolutional layer for channel adjustment to obtain an intermediate feature tensor. Next, the intermediate feature tensor is reshaped into a two-dimensional tensor and coupled with a learnable tensor. Perform element-wise multiplication: ； In the formula, Represents two-dimensional convolution. This indicates element-wise multiplication. Finally, the result of element-wise multiplication is fed into a one-dimensional convolutional layer to restore the target channel dimension, yielding the output of the convolutional branch: ； Among them, for the attention branch, the key Perform filling processing and extract the kernel size through unfolding operations. Local patch The obtained tensor is reconstructed into a four-dimensional structure, with each dimension corresponding to the number of heads. Channel dimensions at each end kernel size and output time length This is composed of components that enable efficient computation of local self-attention, expressed as: ； In the formula, The bond representing the reconstruction into a four-dimensional structure , This indicates a fill operation. This indicates an expand operation.

7. The infrared thermal imaging composite material defect detection method based on deep learning according to claim 6, characterized in that, The channel focusing module sequentially performs one-dimensional convolutional channel expansion, group convolutional feature extraction, batch normalization, channel attention weighting, one-dimensional convolutional channel recovery, and residual connection to achieve adaptive recalibration of channel features, strengthen defect discrimination features, and suppress irrelevant interference features. The processing procedure is as follows: ； In the formula, This represents the output of the local time focusing module. This indicates a group convolution operation. This indicates a batch normalization operation. This indicates channel attention weighting.

8. The infrared thermal imaging composite material defect detection method based on deep learning according to claim 7, characterized in that, The classification head includes a first linear layer, a Dropout layer, and a second linear layer, which are used to output the probability that each pixel belongs to the defective class or the non-defective class. The refined feature map is flattened into a one-dimensional feature vector, and then sequentially input into the first linear layer, the Dropout layer, and the second linear layer for processing. The first linear layer is used to project the input features into a low-dimensional feature space, the Dropout layer is used to reduce the risk of overfitting, and the second linear layer is used to output the scores of the two categories. The process can be represented as follows: ； In the formula, This indicates the output of the channel focusing module. This indicates the output by category. This indicates a linear layer.

9. The infrared thermal imaging composite material defect detection system based on deep learning according to claim 1, used to execute the infrared thermal imaging composite material defect detection method based on deep learning according to any one of claims 1-8, characterized in that, include: The system includes a thermal excitation unit, an infrared image acquisition unit, a data preprocessing unit, a network construction unit, a model training unit, an inference detection unit, and a result output unit.

Images