Defect depth prediction method based on pulse infrared thermal image sequence

Abstract

The invention discloses a defect depth prediction method based on a pulse infrared thermal image sequence, relates to the technical field of non-contact nondestructive testing, solves the problems of insufficient defect depth quantitative evaluation precision, dependence on an empirical model, difficulty in adapting to different materials and the like in the existing active infrared thermal imaging detection, and fully utilizes the time information of the thermal image sequence. Features are automatically extracted through a deep learning model, accurate prediction of the burial depth of the internal defect of the material is realized, and the reliability and generalization ability of detection results under different material conditions are improved. According to the method, the surface temperature response of the material is obtained by combining pulse thermal excitation with infrared thermal imaging, and time sequence characteristics are modeled by comparing and analyzing temperature signals of a defect area and a defect-free area and introducing a long short-term memory (LSTM) neural network, so that the depth of the hidden defect is accurately predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of non-contact non-destructive testing, and in particular to a defect depth prediction method based on a relative norm ratio of defect areas of an infrared thermal image sequence and a long short-term memory neural network. Background Art

[0002] Infrared thermography is an important nondestructive testing method for detecting internal material defects. Active infrared thermography uses external thermal stimulation to induce heat flux, causing transient changes in the surface temperature of the sample being tested. Discontinuities or defects within the sample (such as voids or delamination) can lead to abnormal surface temperature distribution. This abnormality manifests as a temperature response above the defect that differs from that in undefected areas, making it possible to quantitatively assess the defect depth based on the surface temperature response.

[0003] At present, the quantitative assessment of defect depth remains a research focus and difficulty in the field of non-destructive testing. Traditionally, researchers often use characteristic parameters in thermal imaging sequences to infer defect depth, such as the maximum temperature rise (peak temperature) of the defect area relative to the background or the delay time to reach the peak. At the same time, a variety of thermal imaging data processing methods have been developed, such as independent component thermography (ICT), pulse phase thermography (PPT), thermal signal reconstruction (TSR) and principal component thermography (PCT), to extract defect features from thermal imaging sequences. These methods have improved the detection and characterization capabilities of defects to a certain extent, but there are still limitations: on the one hand, they often need to rely on accurate material thermal physical properties or manually selected features, resulting in insufficient adaptability to different materials or complex conditions; on the other hand, single features (such as peak temperature or time) are easily affected by noise and experimental conditions, resulting in limited depth prediction accuracy.

[0004] With the development of artificial intelligence and deep learning technologies, researchers have begun to explore the application of neural networks to data analysis for thermal imaging nondestructive testing. Deep learning models can automatically extract complex features from large amounts of data and have been used for tasks such as defect identification and quantitative assessment. For example, some studies have used finite element simulation to generate data to train neural networks to assess defect depth, achieving some success. However, current methods for predicting defect depth from thermal imaging sequences still face some unresolved challenges. Some methods require manual analysis of temperature curves or rely on the experience of inspectors, which is highly subjective and inefficient. In recent years, while methods have emerged that attempt to predict defect depth using machine learning techniques such as neural networks, these methods often require large amounts of representative data to train the models and lack generalizability across different workpiece types. Existing technologies struggle to achieve both accurate and applicable predictions. Therefore, a new technical solution is urgently needed to address the aforementioned issue of inaccurate defect depth prediction. Summary of the Invention

[0005] In order to solve the problems of insufficient accuracy of quantitative assessment of defect depth in existing active infrared thermal imaging detection, reliance on empirical models and difficulty in adapting to different materials, the present invention makes full use of the time information of the thermal image sequence and automatically extracts features through a deep learning model to achieve accurate prediction of the buried depth of internal defects in the material, thereby improving the reliability and generalization ability of the detection results under different material conditions.

[0006] To address the above issues, the present invention provides a defect depth prediction method based on the time-domain thermal signal ratio. This method uses pulsed thermal excitation combined with infrared thermal imaging to obtain the surface temperature response of the material. By comparing and analyzing the temperature signals of the defective area and the defect-free area, and introducing a long short-term memory neural network (LSTM) to model the time series characteristics, the method can accurately predict the depth of hidden defects. Specifically, the present invention includes the following steps:

[0007] Step 1: Use a pulsed thermal excitation device to instantaneously heat the sample under test, and use an infrared imaging device to non-contactly capture a sequence of infrared thermal images of the sample surface;

[0008] Step 2: During the cooling process, select a frame with a strong thermal contrast between the defect and the background (i.e., it can be seen clearly by the naked eye) when the defect outline is clear during the cooling process after pulse excitation, identify the outline of the defect area in the thermal image through the image processing algorithm, and determine the corresponding defect-free reference area.

[0009] Step 3: Extract the temperature-time signal curves for the defect area and the reference area over the entire time range. For the two sets of time-domain temperature signals obtained, calculate the L1 norm ratio of the defect area relative to the non-defect area within the heating range under different defect depth conditions (this is the process of predicting defect depth, so any defect depth that can be detected by infrared technology is acceptable). This is the ratio of the absolute value of the temperature signal of the defect area to that of the non-defect area at each moment, thus forming a time series of ratios that reflects the change in the thermal response difference between the two over time.

[0010] Step 4: The above ratio time series is input into a pre-trained LSTM (Long Short-Term Memory) neural network regression model to analyze the dynamic characteristics of the defect and predict the time when the defect is first visible. This time is defined as the time when the defect is first clearly detectable in the thermal image sequence. It reflects the time delay from the time the defect is heated to the time it becomes visible, thus avoiding the large errors caused by manual intervention.

[0011] Step 5: Perform a large number of three-dimensional finite element simulation analyses on typical defect depths (three-dimensional finite element simulation is a common technique in this field, which can be understood and easily implemented by those skilled in the art. It is a common skill of those skilled in the art). Use the parametric scanning method of the defect depth to model the curve of the "first visible moment of the defect" output by the network in step 4. And calculate (those skilled in the art can easily implement the calculation based on the above description, and only need to perform a finite element parametric scan on the depth information of the result output in step 4) to obtain the result determination coefficient R 2 , find the depth d and first visible time t d A mapping relationship between two variables (usually an exponential relationship).

[0012] Step 6: Utilize the mapping relationship from Step 5 and several defects of known depth to establish and calibrate a mapping curve between the "time of first visible defect" and defect depth for the current material. To ensure accuracy, the selected defect depths should cover as much of the detection range as possible (the range varies under different defect conditions. Typically, the minimum range is "the depth at which the defect is discovered when the pulse excitation time is greater than one order of magnitude," and the maximum range is "the maximum depth at which the defect outline remains clearly visible after pulse excitation") and be evenly distributed. Generally, selecting three defects of known depths will meet this requirement; the denser the number, the more accurate the prediction. Using pulse excitation, collect data for defects of unknown depth. The "time of first visible defect" output from the LSTM network is then fed into the mapping curve between "time of first visible defect" and defect depth to quickly and accurately predict defect depth.

[0013] As a preferred embodiment, the present invention is aimed at different isotropic and uniform materials with similar material parameters. It only needs to use more than or equal to three sets of calibration defects of known depth to perform step 6, thereby correcting the defect depth prediction curve. Specifically, the material parameters are not much different means that the thermal properties such as thermal conductivity, density and specific heat capacity are within the same order of magnitude. For example, for similar polymer materials or composite materials, the difference in their thermal parameters is small, so the aforementioned mapping relationship (such as an exponential mapping relationship) can be used to perform prediction correction without repeating the process of step five. By using the same exponential mapping relationship, a new prediction curve can be effectively constructed to perform depth prediction on the new material. This method avoids repeated complex calculations and can improve the applicability and prediction efficiency for different materials. However, for materials with complex structures (such as multilayer composite materials or non-uniform heterogeneous structures), since their thermal conductivity and diffusion characteristics may be significantly different, it is necessary to re-model through step five, establish a new mapping relationship, and then perform step six correction. Therefore, the processing of such materials is relatively more complicated, but still within the scope of protection of the method of the present invention.

[0014] As another preferred solution, in step 4 of the present invention, the LSTM network is trained using simulation data; the LSTM network finds the value of "the time when the defect is discovered" based on the time domain sequence, and in order to avoid human errors, this value is specifically quantified. Moreover, for samples of different materials, even if the thermal conduction and diffusion characteristics are different, the time delay value experienced by the output of the model from heating to manifestation is the same in principle, and no retraining is required.

[0015] The long short-term memory neural network model is obtained by pre-offline training of the data model step 3. The data model is the time series of the L1 norm ratio of the defect area relative to the reference area output by step 3 and the artificial label "time when the defect is discovered" for LSTM network regression training to obtain the value of "time when the defect is discovered" without human intervention; it contains multiple layers of long short-term memory neural units to learn the mapping relationship between the thermal image sequence and the defect depth. The model is supervised and trained using a large amount of sample data of defects of different depths, and the model parameters are continuously adjusted so that it can predict the corresponding defect burial depth based on the temperature change law over time. In the actual detection process, the thermal image sequence of the workpiece 1 to be tested is input into the trained long short-term memory neural network model. The model can output the defect depth prediction result by combining the mapping relationship curve between the "first visible moment of the defect" and the defect depth obtained in step 6.

[0016] The LSTM network is trained using simulated data. Through supervised learning, the model is trained to predict defect depths using a large amount of sample data with known defects of varying depths. During training, the model automatically optimizes parameters to ensure highly accurate predictions.

[0017] The LSTM network training process relies on established deep learning training methods. During training, a large amount of simulation data is used to continuously optimize model parameters through supervised learning to ensure high-precision predictions. The training methods and steps are well within the skill of those skilled in the art.

[0018] As another preferred solution, the present invention controls and records the emission of the pulse signal by a computer during data acquisition, and calibrates the specific time of the pulse signal emission in the time domain sequence.

[0019] As another preferred solution, for materials with high thermal conductivity, the present invention increases the power of the pulsed thermal excitation device 2 or shortens the infrared thermal image acquisition interval to capture a faster temperature decay process. For samples with low thermal conductivity, the acquisition time of the infrared camera 1 is extended to cover the entire cooling curve. For example, the excitation time is 2ms, and the instantaneous energy released is approximately 20J. For high thermal conductivity materials, the recommended temperature rise range is 20°C to 40°C, and the excitation time should always be at least two orders of magnitude less than the cooling time.

[0020] As another preferred approach, the present invention improves the versatility of prediction for samples of different materials by establishing material-specific models, incorporating material parameters into the general model, or performing linear correction based on the material's typical thermal response deviation after the general model output. Due to differences in thermal conductivity and diffusion properties, the relationship between thermal signature and depth may vary for samples of different materials. In specific implementations, the versatility of prediction is improved by establishing material-specific models. For example, a set of calibration test blocks with known defect depths is prepared for ABS plastic, carbon fiber composites, PMMA organic glass, and other materials. Training data is then obtained for each of these models to adjust or fine-tune the step model for more accurate predictions for that material. Another strategy is to incorporate material parameters into the general model: for example, using the material's thermal diffusivity as an input, or performing linear correction based on the material's typical thermal response deviation after the model output. Through these measures, the system of the present invention can maintain good prediction performance when applied to different materials. In practical applications, the operator simply selects the corresponding material settings, and the system invokes the optimized model or parameters for that material for depth calculation, eliminating the need to rebuild the detection system, demonstrating the flexibility and practical value of this method.

[0021] As another preferred solution, the pulsed thermal excitation device of the present invention adopts a pulsed laser or a high-energy flash lamp, and the infrared imaging equipment adopts an infrared thermal imager.

[0022] As another preferred embodiment, in steps 1 and 2 of the present invention, the instantaneous heating time of the sample under test (e.g., 2 ms) is at least two orders of magnitude shorter than the cooling time, thereby improving accuracy. Instantaneous heating means applying sufficient heat in a very short period of time.

[0023] As another preferred embodiment, in step 5 of the present invention, the typical defect depth is a standard defect of different depths, ranging from 0.5 mm to 2.2 mm, and these depth standards are used to establish a relationship between the LSTM network and the defect depth.

[0024] Secondly, in step 2 of the present invention, the image processing algorithm adopts a grayscale mutation analysis algorithm, a shape recognition algorithm or a threshold segmentation algorithm.

[0025] In addition, in step 3 of the present invention, the L1 norm ratio of the defective area to the reference area is calculated, that is:

[0026]

[0027] Where T(t) represents the average temperature of the corresponding area at time t. The ratio sequence R(t) is input into the neural network as a feature vector reflecting the difference in thermal diffusion.

[0028] In step 1 of the present invention, the excitation power of the pulsed thermal excitation device will not show obvious power attenuation during the multiple heatings; the pulsed excitation only needs to be heated once. This does not mean that a certain area of ​​a sample needs to be heated a certain number of times, but when other conditions such as the distance between the sample and the pulsed excitation device are the same, different defect areas must have the same sufficiently identical excitation heat source. For example, a sample is very large and needs to be divided into 100 areas for detection. This method requires that the heat sources of these 100 areas must be sufficiently consistent. As long as the pulsed heat source meets the above requirements, the goal of this method can be achieved. The reason for heating several times is that the heating area is not large. In most cases, a sample needs to be divided into more areas for detection, and the detection of each area only requires one pulsed excitation.

[0029] In step 2 of the present invention, the specific frame rate selected has little impact on the final prediction results. For example, after pulse excitation, the defect outline is relatively clear at both 10 and 12 seconds of cooling. Regardless of whether the 10-second or 12-second frame is selected, the accuracy of depth prediction after contour extraction is minimally affected.

[0030] The present invention has beneficial effects.

[0031] The prediction modeling platform of the present invention runs an LSTM deep learning model, analyzes the extracted thermal signal ratio time series, and outputs a prediction result of the first visible moment of the defect; the depth correction module calculates the defect depth based on the predicted first visible moment combined with a pre-stored time-depth mapping curve, and makes necessary corrections to the prediction model based on the thermal properties of the material.

[0032] The present invention uses the norm ratio of the thermal signal of the defect area and the non-defect area within the heating range as a feature to improve the robustness of defect signal discrimination under different environments; introduces the LSTM time series model to automatically determine the time of defect occurrence, reduces human intervention and experience judgment errors, and realizes the intelligence of defect detection and depth assessment; converts time delay into depth through the empirical mapping curve, the method is intuitive and efficient, and can be easily adapted to different materials through a small amount of calibration. In addition, with the help of finite element simulation, parameter scanning analysis of defects of different depths is carried out to verify and optimize the time-depth mapping relationship, reducing the cost of calibration of a large number of samples. In addition, the method and system of the present invention have good portability and flexibility, and the algorithm framework can be implemented on a variety of simulation and image processing platforms. It is not limited to a specific software environment and is easy to integrate and deploy in existing detection equipment. In summary, the present invention provides an engineering-practical and adaptable defect depth prediction solution. The method is suitable for low-cost and rapid detection of low-depth defects in large-area materials with the same material parameters in mass-produced items or materials. It has broad application prospects in the fields of stacking delamination problems and major equipment casing detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention is further described below with reference to the accompanying drawings and specific embodiments. The scope of protection of the present invention is not limited to the following description.

[0034] Figure 1 This is a schematic diagram of the device structure of the present invention. Here, 1 is a sample with defects, 2 is an infrared imaging device, 3 is a pulse excitation lamp, 4 is a capacitor and control components for the pulse excitation lamp, and 5 is a computer that simultaneously collects infrared thermal imaging device data and pulse timing.

[0035] Figure 2 Schematic diagram of a curve showing the ratio of the L1 norm of the defect area of ​​the example material to the reference area at different depths according to the present invention.

[0036] Figure 3 This is a schematic diagram of fitting 300 sets of simulation data sets based on the predicted time values ​​of the LSTM network.

[0037] Figure 4 The present invention uses prediction curves to predict the depth of the result values ​​output by the LSTM network for the thermal imaging detection result sequences of different defect depths. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, the defect depth prediction system of the present invention includes: an excitation device 3 for pulse heating (such as a pulse laser or a high-energy flash lamp), an infrared imaging device 2 for thermal signal acquisition, a computer 5 for synchronously recording the pulse excitation timing and thermal image sequence, and a matching image processing and deep learning model analysis module. The pulse excitation device 3 provides stable and uniform heating power through a pulse excitation power supply and a control module 4 (the control module 4 can be a power supply control switch of the battery of the flash lamp. The computer 5 controls the battery to power the flash lamp and records when the flash starts, that is, the moment when the data set starts to heat), ensuring the consistency of thermal excitation in multiple samplings. The computer 5 accurately calibrates the excitation start time point through a synchronous control mechanism to ensure that a complete and time-aligned infrared thermal image sequence is collected.

[0040] Step 1: Pulsed thermal excitation is applied to the sample 1. Infrared imaging device 2 continuously captures the sample's surface temperature distribution at a fixed frame rate, forming a sequence of thermal images. The short duration and high energy of the thermal excitation stimulate the sample's thermal response in a short period of time.

[0041] Step 2: At the end of the cooling phase of the thermal imaging sequence, when the temperature difference is significant, extract the outline of the defect area from the thermal image. This step can be performed manually by selecting obvious defects, or automatically using image processing algorithms (such as grayscale mutation analysis, shape recognition, and threshold segmentation). The extracted defect area and adjacent non-defective reference areas will serve as the basis for subsequent analysis.

[0042] Step 3: Extract the average temperature change curves of the defect area and the reference area in the entire thermal image sequence. Then calculate the L1 norm ratio of the temperature signals of the two areas in each frame, that is:

[0043]

[0044] Where T(t) represents the average temperature of the corresponding area at time t. The ratio sequence R(t) is input into the neural network as a feature vector reflecting the difference in thermal diffusion.

[0045] Step 4: Input the ratio sequence R(t) into the trained LSTM neural network model, perform time series analysis, and output the “first visible time point” of the defect t d This moment corresponds to the position where the defect first becomes clearly visible in the thermal map, reflecting its dynamic response in thermal diffusion.

[0046] Step 5: For typical materials (such as ABS) and defects of multiple known depths (equivalent to the anchor point of the mapping relationship between "defect discovery time" and "defect depth"), finite element simulation is used to perform parametric modeling to obtain the first visible time t corresponding to different depths. d By fitting means such as the least square method, the depth d and t are established. d The mapping relationship between Figure 2 ). Usually an exponential relationship:

[0047]

[0048] Where a and b are parameters; after fitting, a high-precision fitting curve (R 2 It can reach above 0.998), providing a mathematical basis for predicting the depth of defects.

[0049] Step 6: Repeat steps 1-4 for samples with unknown depth defects to obtain the first visible time t of the defect d The depth can be directly predicted by fitting the mapping curve obtained in step 5. Figure 3 As shown in , for large-scale simulation data sets, the present invention establishes a stable and high-precision fitting relationship between the LSTM prediction value and the actual defect depth; Figure 4 The prediction effect of this method on defects of different depths is demonstrated on experimental thermal image sequences, verifying the accuracy of this method.

[0050] Material Adaptation and Correction Mechanism: For homogeneous materials with similar parameters (such as similar polymers or composites), the present invention requires only three or more calibration samples at different depths, and rapid adaptation can be achieved by adjusting the mapping curve in step 6. For heterogeneous or complex materials, step 5 can be repeated to create a new model. The present invention's LSTM network inherently exhibits strong cross-material adaptability, eliminating the need for retraining in most cases.

[0051] System Integration: The entire method can be embedded in an industrial infrared inspection system. The system comprises a heat source, a thermal imager, a data synchronization and control module, image processing software, and a defect prediction platform. All modules can be integrated within a single software interface, making it suitable for both production line inspection and batch evaluation.

[0052] The present invention extracts the relative norm ratio between defective areas and non-defective areas and uses this ratio as a training feature to truly quantify the "time when the defect is discovered". Simulation is also used to find the mapping relationship between "time when the defect is discovered" and "defect depth".

[0053] The present invention reasonably quantifies the "defect identification time" and uses batch simulation to reasonably map this indicator to the depth of the defect; and has correction strategies for materials with different properties.

[0054] Example:

[0055] A defect depth prediction method based on infrared thermal image sequences and LSTM technology was validated in a thermal imaging test setup. The specimen used was an ABS thermoplastic sample manufactured via SLA 3D printing, with a total thickness of 3.0 mm. Multiple artificial defects (cylindrical cavities) ranging in depth from 0.5 mm to 2.2 mm were embedded within the sample, spaced in increments of 0.1 mm. Each defect had a diameter of approximately 3 mm and was coated with a matte finish to enhance infrared absorption.

[0056] Step 1: Dual flash lamp pulses (each with an output energy of 6.4 kJ and a pulse width of approximately 2 ms) were used to transiently heat the sample surface. An infrared thermal imager (bandwidth 1.5–5.4 μm, resolution 640 × 512 pixels, frame rate 50 Hz) was used to capture a 30-second sequence of 1500 infrared thermal images during the cooling process. All thermal image data and pulse initiation signals were controlled and recorded by a connected computer.

[0057] Step 2: After acquisition, the 1200th thermal image frame, where the defect outline is most clearly visible in the temperature field, is selected as a reference image for defect area identification. The experimenter uses the Hough circle algorithm tool to select each defect area and manually selects a reference area in an adjacent defect-free area. The average temperature series of each area over the entire time domain are extracted.

[0058] Step 3: Calculate the L1 norm of the temperature series for the defect area and the reference area, and then generate a ratio sequence reflecting the difference in thermal diffusion based on the time point ratio. This ratio sequence is input into the trained LSTM neural network model, which outputs a predicted "first visible time" for each data set. This value is the time, in seconds, when the defect first becomes noticeable in the thermal image.

[0059] Step 4: Through finite element simulation, the corresponding ABS material model was constructed in COMSOL Multiphysics software to simulate the heating response behavior of defects at different depths. A total of 300 sets of simulation data were collected. The first visible time predicted by the LSTM network was compared with the known depth to establish a mapping relationship and evaluate the R under various mapping conditions. 2 Value, select exponential mapping and fit:

[0060]

[0061] Where d is the defect depth, t is the time point of first visibility, and the fitting parameters a=0.21, b=0.65 are obtained. The fitting coefficient R 2 As high as 0.998.

[0062] Step 5: Input the predicted time of the experimental sample into the mapping curve to obtain the predicted depth value of each defect. Compare the predicted value with the actual defect depth of the sample. For the case of using three defects with known depths to predict the depths of another 15 defects, at a 95% confidence level, the depth range corresponding to each time point is: 1 second, the depth range is [0.39, 0.45]; at 2 seconds, the depth range is [0.67, 0.76]; at 4 seconds, the depth range is [1.08, 1.23]; at 6 seconds, the depth range is [1.38, 1.58]; at 8 seconds, the depth range is [1.62, 1.84]. The accuracy within the corresponding depth range is controlled within ±0.11 mm, verifying the accuracy of this method.

[0063] It can be understood that the above specific description of the present invention is only used to illustrate the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those skilled in the art should understand that the present invention can still be modified or replaced by equivalents to achieve the same technical effects; as long as the use requirements are met, they are within the scope of protection of the present invention.

Claims

1. A defect depth prediction method based on pulsed infrared thermal imaging sequence, characterized in that The following steps are involved: Step 1: Use a pulsed thermal excitation device to instantaneously heat the sample under test, and use an infrared imaging device to non-contactly capture a sequence of infrared thermal images of the sample surface; Step 2: During the cooling process, select a frame with a strong thermal contrast between the defect and the background during the cooling process after pulse excitation when the defect outline is clear. Use the image processing algorithm to identify the outline of the defect area in the thermal image and determine the corresponding defect-free reference area. Step 3: Extract the temperature-time signal curves of the defect area and the reference area over the entire time range; for the two sets of time-domain temperature signals obtained, calculate the L1 norm ratio of the defect area to the non-defect area within the heating range under different defect depth conditions, that is, the ratio of the absolute value of the temperature signal of the defect area to that of the non-defect area at each moment, thereby forming a time series of ratios that reflects the change in the thermal response difference between the two over time; Step 4: Input the above ratio time series into the pre-trained LSTM network regression model to perform dynamic feature analysis on the defect and predict the first visible time of the defect. The first visible time is defined as the time point when the defect sign is first clearly detectable in the thermal image sequence, reflecting the time delay from the time the defect is heated to the time it becomes visible. Step 5: Perform a large number of three-dimensional finite element simulations on typical defect depths, and use a parametric scan of the defect depth to model the curve of the "first visible moment of the defect" output by the network in step 4 with respect to the defect depth; and calculate the result determination coefficient R 2 , find the depth d and first visible time t d The mapping relationship between two variables; Step 6: Using the mapping relationship in step 5 and defects of known depth, establish and calibrate the mapping relationship curve between the "first visible moment of the defect" and the defect depth of the current material; through pulse excitation, collect data on defects of unknown depth, input the infrared time series sequence into the LSTM network, and input the "first visible moment of the defect" output into the mapping relationship curve between the "first visible moment of the defect" and the defect depth to quickly and accurately predict the defect depth.

2. A defect depth prediction method based on a pulsed infrared thermal imaging sequence according to claim 1, characterized in that For different materials that are isotropic and uniform with thermal properties such as thermal conductivity, density, and specific heat capacity within the same order of magnitude, it is only necessary to perform step 6 using three or more sets of calibration defects with known depths to correct the defect depth prediction curve without repeating the process of step 5.

3. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that In step 4, the LSTM network is trained using simulation data. The LSTM network finds the value of "defect discovery time" based on the time domain sequence and quantifies this value. Moreover, for samples of different materials, even if the thermal conduction and diffusion characteristics are different, the time delay value from heating to manifestation is the same in principle, and no retraining is required.

4. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that The LSTM network model is obtained by pre-offline training of the data model in step 3. The data model, i.e., the time series of the L1 norm ratios of the defect area relative to the reference area output in step 3, is used for LSTM network regression training with the human label "time when the defect was discovered" to obtain the value of "time when the defect was discovered" without human intervention. The LSTM network model includes multiple layers of long-short-term memory neural units to learn the mapping relationship between thermal image sequences and defect depths. The model is supervised and trained using sample data of defects of different depths, and the model parameters are continuously adjusted so that the model can predict the corresponding defect burial depth based on the temperature variation pattern over time.

5. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that For materials with high thermal conductivity, increase the power of the pulsed thermal excitation device or shorten the time interval for infrared thermal image acquisition to capture a faster temperature decay process; for samples with low thermal conductivity, extend the acquisition time of the infrared camera to cover the complete cooling curve.

6. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that For samples of different materials, the versatility of prediction can be improved by establishing material-specific models, adding material parameters to the general model, or performing linear correction based on the typical thermal response deviation of the material after the general model output.

7. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that In step 1 and step 2, the instantaneous heating time of the sample to be tested is more than two orders of magnitude shorter than the cooling time.

8. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that In step 5, the typical defect depths are standard defects of different depths, ranging from 0.5 mm to 2.2 mm. These depth standards are used to establish the relationship between the LSTM network and the defect depth.

9. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that In step 2, the image processing algorithm adopts a grayscale mutation analysis algorithm, a shape recognition algorithm or a threshold segmentation algorithm.

10. A defect depth prediction method based on pulsed infrared thermal imaging sequence according to claim 1, characterized in that In step 3, the L1 norm ratio of the defect area to the reference area is calculated, that is: Where T(t) represents the average temperature of the corresponding area at time t; the ratio sequence R(t) is input into the neural network as a feature vector reflecting the difference in thermal diffusion.

Images