Infrared thermal imaging detection method and system for stress cracks on surface of hybrid welding-milling workpiece

Abstract

The invention provides an infrared thermal imaging detection method and system for stress cracks on the surface of a hybrid welding-milling workpiece, and relates to the technical field of manufacturing and detection.The method comprises the steps that the junction of the milling surface of the hybrid welding-milling workpiece to be detected and a welding heat affected zone is pretreated to obtain the pretreated surface to be detected; performing uniform thermal excitation on the pretreated surface to be detected in a pulse thermal excitation mode, and acquiring a dynamic thermal image sequence in a cooling process by using an infrared thermal imager; thermal attenuation curve characteristics of each pixel point in the cooling process are extracted by performing time sequence analysis on the dynamic thermal image sequence; according to the method, accurate identification and reliable detection of the stress crack in the key area of the hybrid welding-milling workpiece are realized.

Description

Technical Field

[0001] This invention relates to the field of manufacturing inspection technology, and in particular to an infrared thermal imaging detection method and system for stress cracks on the surface of composite welded-milled parts. Background Technology

[0002] Composite welded-milled parts combine the connection strength of welded structures with the precision of milled surfaces, making them widely used in critical equipment such as automotive longitudinal beams and aero-engine casings. However, microscale stress cracks are easily generated at the junction of the milled surface and the weld heat-affected zone due to thermal cycling and mechanical stress superposition. If not detected in time, these cracks will endanger the safety of the equipment in service. Although existing infrared thermal imaging detection technology is a non-contact solution, it is difficult to effectively separate spurious signals from crack signals due to the influence of background thermal noise, stripe interference, and regular thermal signals of milling texture. When a car manufacturer uses plasma-MIG composite welding to produce automotive longitudinal beams, stress cracks appear at the interface between the heat-affected zone and the milled surface. Traditional infrared thermal imaging detection, lacking a correlation between texture and anisotropic heat conduction, misinterprets the thermal signals corresponding to regularly distributed milled textures as cracks, resulting in a large number of false positives. Additional manpower is required for ultrasonic verification, exposing the core defects of traditional technology: lack of signal separation mechanism for milled texture characteristics and low accuracy in identifying microcracks. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an infrared thermal imaging detection method and system for stress cracks on the surface of composite welded-milled parts, so as to achieve accurate identification and reliable detection of stress cracks in key areas of composite welded-milled parts.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, an infrared thermal imaging method for detecting surface stress cracks in composite welded-milled parts, the method comprising: The interface between the milled surface and the weld heat-affected zone of the composite welded-milled workpiece to be tested is pretreated to obtain the pretreated surface to be inspected. The pre-treated surface to be inspected is uniformly thermally excited by pulse thermal excitation, while an infrared thermal imager is used to collect dynamic thermal image sequences during the cooling process. By performing time-series analysis on dynamic thermal imaging sequences, the thermal decay curve features of each pixel during the cooling process are extracted. Combining the characteristics of thermal attenuation curves and the spatial information of dynamic thermal imaging sequences, and based on the regular distribution characteristics of milling textures and the random distribution characteristics of stress cracks, a correlation model between texture direction and heat conduction anisotropy is established. Principal component analysis is used to separate the regular thermal signal caused by textures from the abnormal thermal signal caused by cracks, and the separated abnormal thermal signal is obtained. The separated abnormal thermal signals and their corresponding spatial coordinate information are subjected to spatial interpolation to obtain the interpolated abnormal thermal signals and normalized spatial coordinate information. The abnormal thermal signals and normalized spatial coordinate information are then input into a pre-trained neural network model to learn the complex mapping relationship between the interpolated abnormal thermal signals and the normalized spatial distribution, and the optimized abnormal thermal signals are output. Based on the optimized abnormal thermal signal, multi-dimensional criteria are used for identification to accurately distinguish between texture pseudo-signals and real stress cracks.

[0005] Furthermore, the pre-treated surface to be inspected is uniformly thermally excited using a pulsed thermal excitation method, while a dynamic thermal image sequence during the cooling process is acquired using an infrared thermal imager, including: Based on the pre-processed surface to be inspected, a coordinated control command with specific energy and duration parameters is obtained; Based on the coordinated control command, the pulse thermal excitation device and the infrared thermal imager are synchronously triggered to enter the ready state; The pulsed thermal excitation device, which is in a ready state, performs pulsed uniform thermal excitation on the pre-treated surface to be inspected, so that the surface to be inspected is in an excited thermal state. For the surface under inspection in an excited thermal state, an infrared thermal imager in a ready state is used to record the thermal state changes of the surface during the cooling process at a set acquisition frequency to obtain the original dynamic thermal image sequence.

[0006] Furthermore, by performing time-series analysis on the dynamic thermal imaging sequence, the thermal decay curve features of each pixel during the cooling process are extracted, including: Based on the original dynamic thermal image sequence, the thermal image data of each frame in the sequence is subjected to temporal filtering for noise reduction to obtain the noise-reduced dynamic thermal image sequence. Based on the denoised dynamic thermal image sequence, temperature change data is reconstructed along the time dimension for each pixel location to obtain the initial thermal decay curve for each pixel location; The initial thermal decay curve of each pixel location is standardized to eliminate the overall temperature level difference caused by different spatial locations, resulting in a standardized thermal decay curve. Characteristic parameters representing thermal decay characteristics are extracted from the standardized thermal decay curve, including the temperature decay rate and curve shape characteristic parameters for a specific time interval, to obtain the thermal decay curve features of each pixel.

[0007] Furthermore, combining the characteristics of thermal attenuation curves and the spatial information of dynamic thermal imaging sequences, and based on the regular distribution characteristics of milling textures and the random distribution characteristics of stress cracks, a correlation model between texture direction and heat conduction anisotropy is established. Principal component analysis is used to separate the regular thermal signals caused by textures from the abnormal thermal signals caused by cracks, yielding the separated abnormal thermal signals, including: The thermal decay curve features of each pixel are integrated with the corresponding spatial location information to construct a feature matrix that includes thermal decay features and spatial distribution. Based on the feature matrix, the dominant direction of milling texture is determined by spatial gradient analysis, and a quantitative correlation model between texture direction and thermal anisotropy is established. The feature matrix is ​​input into the correlation model, and the features are decoupled through principal component analysis. The data is then projected onto the feature subspace spanned by the texture regular pattern and the orthogonal complement space. In the orthogonal complement space, signal components that are not related to the regular pattern of the texture are extracted, and these unrelated signal components are used as the separated abnormal thermal signals.

[0008] Furthermore, the separated anomalous thermal signals and their corresponding spatial coordinates are subjected to spatial interpolation to obtain the interpolated anomalous thermal signals and normalized spatial coordinates. These are then input into a pre-trained neural network model to learn the complex mapping relationship between the interpolated anomalous thermal signals and the normalized spatial distribution, and the optimized anomalous thermal signals are output, including: Based on the separated abnormal thermal signals and their corresponding spatial coordinate information, the discrete point set spatial normalization interpolation method is used for processing. The discrete point set spatial normalization interpolation method refers to converting irregularly distributed spatial coordinate points into uniformly distributed spatial coordinate points in a normalized coordinate system through interpolation calculation to obtain normalized spatial coordinate information. Based on the normalized spatial coordinate information, the abnormal thermal signal is resampled and interpolated to obtain the interpolated abnormal thermal signal. The interpolated abnormal heat signal and normalized spatial coordinate information are input into a pre-trained neural network model. The neural network model learns the complex mapping relationship between the interpolated abnormal heat signal and the normalized spatial distribution, and outputs the optimized abnormal heat signal.

[0009] Furthermore, based on the separated abnormal thermal signals and their corresponding spatial coordinate information, a spatial normalization interpolation method for discrete point sets is used. This method transforms irregularly distributed spatial coordinate points into uniformly distributed spatial coordinate points in a normalized coordinate system through interpolation calculations, thereby obtaining normalized spatial coordinate information, including: Based on the separated abnormal thermal signals and the corresponding irregularly distributed spatial coordinate information, the original spatial data point set is constructed. The distribution range of the original spatial data point set is identified by boundary, and the creation parameters of a rectangular normalized grid that completely covers all data points are calculated based on the identified boundary range. Based on the grid creation parameters, a normalized coordinate matrix is ​​instantiated in memory. Each element in the matrix represents a normalized spatial coordinate point that is uniformly distributed within the detection area. All normalized spatial coordinate points together constitute normalized spatial coordinate information.

[0010] Furthermore, based on the optimized abnormal thermal signal, multi-dimensional criteria are used for identification to accurately distinguish between texture pseudo-signals and real stress cracks, including: Based on the optimized anomalous thermal signal, a two-dimensional anomalous signal data field containing spatial distribution and signal intensity is received and constructed. Based on the two-dimensional anomalous signal data field, a multidimensional identification feature vector is extracted for each anomalous signal region. The multidimensional identification feature vector includes contour features that characterize spatial continuity, shape features that characterize morphological complexity, and amplitude features that characterize thermal response intensity. Based on multidimensional recognition feature vectors, an unsupervised clustering algorithm is used to classify all abnormal signal regions by pattern. Signal regions with regular contours, single orientation, and high correlation with texture direction are identified as texture pseudo signals, while signal regions with irregular contours, random orientation, and no relation to texture direction are identified as potential real stress cracks, so as to obtain the classification results. Based on the classification results, the confidence level of the identified potential real stress crack regions is evaluated. The matching degree between the morphology, strength and spatial characteristics and the typical stress crack characteristics is calculated to obtain the confidence score of each crack region. Based on the confidence score, the final stress crack detection result is obtained to accurately distinguish between texture pseudo signals and real stress cracks.

[0011] Secondly, an infrared thermal imaging detection system for stress cracks on the surface of composite welded-milled parts includes: The acquisition module is used to preprocess the interface between the milled surface and the weld heat-affected zone of the composite welded-milled workpiece to be tested, so as to obtain the preprocessed surface to be inspected. The acquisition module is used to uniformly thermally excite the pre-processed surface to be inspected using pulsed thermal excitation, while simultaneously using an infrared thermal imager to acquire dynamic thermal image sequences during the cooling process. The extraction module is used to extract the thermal decay curve features of each pixel during the cooling process by performing time-series analysis on the dynamic thermal image sequence. The separation module combines the characteristics of the thermal decay curve with the spatial information of the dynamic thermal image sequence. Based on the regular distribution characteristics of milling texture and the random distribution characteristics of stress cracks, it establishes a correlation model between texture direction and heat conduction anisotropy. Through principal component analysis, it separates the regular thermal signal caused by texture and the abnormal thermal signal caused by cracks to obtain the separated abnormal thermal signal. The optimization module is used to perform spatial interpolation on the separated abnormal heat signals and their corresponding spatial coordinate information to obtain the interpolated abnormal heat signals and normalized spatial coordinate information. The abnormal heat signals and normalized spatial coordinate information are then input into a pre-trained neural network model to learn the complex mapping relationship between the interpolated abnormal heat signals and the normalized spatial distribution, and the optimized abnormal heat signals are output. The processing module is used to identify abnormal thermal signals based on optimized signals through multi-dimensional criteria, so as to accurately distinguish between texture pseudo-signals and real stress cracks.

[0012] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0013] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0014] The above-described solution of the present invention has at least the following beneficial effects: By employing a series of technical methods—including preprocessing to optimize the surface to be inspected, simultaneous pulsed thermal excitation and infrared thermography acquisition, time-series analysis to extract thermal attenuation curve features, establishing an anisotropic correlation model between texture direction and heat conduction, separating texture pseudo-signals and crack anomaly signals through principal component analysis, combining spatial normalization interpolation and neural network optimization of abnormal thermal signals, and intelligent recognition based on multi-dimensional criteria—this technology effectively overcomes the technical problems of traditional infrared thermography in inspecting composite welded-milled parts, such as difficulty in distinguishing milling texture pseudo-signals from real stress crack signals, insufficient accuracy in micro-crack identification, and high false positive rates. This achieves accurate identification of stress cracks at the interface between the milled surface and the weld heat-affected zone of composite welded-milled parts, significantly reducing the false positive rate and manual intervention costs, while maintaining the high efficiency and safety of non-contact inspection, thus providing reliable assurance for the safe operation of critical equipment. Attached Figure Description

[0015] Figure 1This is a schematic flowchart of an infrared thermal imaging detection method for surface stress cracks in composite welded-milled parts provided by an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of an infrared thermal imaging detection system for surface stress cracks in composite welded-milled parts provided in an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, embodiments of the present invention propose an infrared thermal imaging detection method for surface stress cracks in composite welded-milled parts, the method comprising the following steps: Step 1: Pre-treat the junction between the milled surface and the weld heat-affected zone of the composite welded-milled workpiece to be tested, so as to obtain the pre-treated surface to be inspected. Step 2: The pre-treated surface to be inspected is uniformly thermally excited by pulse thermal excitation, and a dynamic thermal image sequence during the cooling process is acquired using an infrared thermal imager. Step 3: By performing time-series analysis on the dynamic thermal image sequence, the thermal decay curve features of each pixel during the cooling process are extracted; Step 4: Combining the characteristics of the thermal attenuation curve and the spatial information of the dynamic thermal image sequence, based on the regular distribution characteristics of milling texture and the random distribution characteristics of stress cracks, establish a correlation model between texture direction and heat conduction anisotropy. Through principal component analysis, separate the regular thermal signal caused by texture and the abnormal thermal signal caused by cracks to obtain the separated abnormal thermal signal. Step 5: Perform spatial interpolation on the separated abnormal heat signal and its corresponding spatial coordinate information to obtain the interpolated abnormal heat signal and normalized spatial coordinate information; input the abnormal heat signal and normalized spatial coordinate information into the pre-trained neural network model to learn the complex mapping relationship between the interpolated abnormal heat signal and the normalized spatial distribution, and output the optimized abnormal heat signal. Step 6: Based on the optimized abnormal thermal signal, identification is performed using multi-dimensional criteria to accurately distinguish between texture pseudo-signals and real stress cracks.

[0019] In this embodiment of the invention, by employing a series of technical means—including preprocessing of the key parts to be tested, synchronous acquisition of pulsed thermal excitation and dynamic thermal imaging sequences, extraction of pixel thermal attenuation curve features through time-series analysis, establishment of an anisotropic correlation model of texture direction and thermal conduction based on texture and crack distribution characteristics, signal separation through principal component analysis, optimization of abnormal thermal signals through spatial interpolation combined with neural networks, and identification based on multidimensional criteria—the invention effectively overcomes the technical problems of traditional infrared thermal imaging detection, such as difficulty in distinguishing between milling texture pseudo signals and real stress crack signals in composite welded-milled parts, insufficient accuracy in microcrack identification, and high false positive rate. This achieves accurate identification of stress cracks at the interface between the milled surface and the weld heat-affected zone of such workpieces, balancing the efficiency and safety of non-contact detection, reducing manual verification costs, and providing strong technical assurance for the service reliability of key equipment.

[0020] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Based on the 3D scanning data of the workpiece under test, determine the precise spatial location information of the boundary between the milled surface and the weld heat-affected zone. Specifically, this includes: First, starting the 3D laser scanning equipment to comprehensively scan the entire composite welded-milled workpiece under test, ensuring the scanning range covers the welded area, the milled surface, and the transition area between them, and completely acquiring the 3D point cloud data of the workpiece surface. After scanning, the acquired 3D point cloud data is preprocessed to eliminate the influence of environmental interference and equipment measurement errors by removing abnormal data points that significantly deviate from the workpiece contour. Next, point cloud registration technology is used to integrate and stitch together multiple sets of point cloud data obtained from different perspectives and positions to form a complete and coherent 3D model of the workpiece. Based on the constructed 3D model, the structural characteristics of the weld heat-affected zone are analyzed. This zone exhibits a different organizational form and surface texture due to the thermal effects of the welding process. Simultaneously, the processing texture pattern of the milled surface is identified. By comparing the differences in the structural boundary attributes of these two regions in 3D space, the spatial range of the boundary between the milled surface and the weld heat-affected zone is accurately defined, and the specific coordinate information of the boundary in the 3D coordinate system is clarified.

[0021] Step 1.2: Based on precise spatial location information, control the automated surface treatment equipment to perform preliminary cleaning of the interface area, removing surface oil, dust, and loose oxides. Specifically, this includes: using motion control software to plan the motion trajectory of the automated surface treatment equipment according to the determined precise spatial location information of the interface area, ensuring that the equipment's working range accurately covers the entire interface area and that the motion path does not interfere with other areas of the workpiece; starting the automated surface treatment equipment, which is equipped with a high-pressure airflow jet device, a neutral cleaning medium supply system, and a flexible cleaning brush assembly; first, controlling the high-pressure airflow jet device to turn on, spraying dry and clean compressed air onto the interface area to blow away the dust, loose impurities, and some floating rust adhering to the surface. Subsequently, a neutral cleaning medium is evenly sprayed onto the interface area through the media supply system. This cleaning medium has good decontamination ability and will not corrode the workpiece surface. At the same time, the flexible cleaning brush assembly is activated. The cleaning brush contacts the interface surface at an appropriate speed and pressure and performs reciprocating motion to remove the oil and stubborn oxides attached to the surface through physical wiping. During the cleaning process, the vision detection module on the equipment captures real-time images of the interface surface. The removal of surface impurities is judged by image analysis. After confirming that the oil, dust and loose oxides have been completely removed, the cleaning medium supply system and the flexible cleaning brush assembly are turned off. High-pressure airflow continues to spray for a period of time to dry the residual cleaning medium on the surface, ensuring that the interface surface is dry and clean.

[0022] Step 1.3: Based on the pre-cleaned surface, perform surface roughness measurement and, based on the comparison of the measurement results with a preset roughness threshold, determine whether to perform fine grinding to obtain a uniform and smooth surface. Specifically, this includes: determining multiple measurement points on the pre-cleaned interface surface according to the principle of uniform distribution; the number of measurement points is reasonably set according to the area of ​​the interface, and the spacing between adjacent measurement points remains consistent to ensure that the measurement results can comprehensively reflect the roughness status of the entire interface; starting the contact roughness measuring instrument and controlling the probe of the measuring instrument to contact each measurement point sequentially along a preset path, collecting roughness data for each measurement point; after completing the data collection of all measurement points, statistically analyzing the acquired roughness data to calculate the average and variance of the roughness of all measurement points, thereby evaluating the overall level and uniformity of the surface roughness of the entire interface; comparing the calculated average roughness with a preset roughness threshold, which is determined based on the technical requirement that subsequent pulsed thermal excitation can be uniformly conducted to ensure the consistency of surface thermal response during thermal excitation. If the average roughness exceeds the preset threshold, it indicates that the surface flatness does not meet the requirements. The automated fine grinding equipment is then activated. This equipment is equipped with a diamond fine grinding wheel, and the grinding intensity and wheel speed are adjusted according to the degree of roughness exceeding the standard. The grinding equipment is controlled to uniformly grind the interface surface along a planned path. During the grinding process, the operation is paused periodically, and a roughness measuring instrument is used to sample and inspect the ground area again. The grinding parameters are adjusted in real time based on the inspection results. This grinding and inspection process is repeated until the average roughness of the interface surface meets the preset threshold, forming a uniform and smooth surface. Step 1.4: Based on a uniform and smooth surface, apply a uniform thermally enhanced coating to optimize the surface's thermal conductivity, ultimately obtaining the pre-treated surface to be inspected. Specifically, this includes: selecting a thermally enhanced coating material with high thermal emissivity, good adhesion, and stable thermal conductivity. This material effectively improves the surface's response sensitivity to thermal signals, reduces the attenuation of thermal signals during transmission, and has good compatibility with the workpiece surface material, preventing chemical reactions; starting the automated spraying equipment, adjusting the nozzle angle according to the determined spatial shape and range of the interface area to ensure the nozzle is perpendicularly aligned with the interface surface; and setting appropriate spray flow rate and spray pressure to ensure the coating material is uniformly atomized and sprayed onto the surface. Maintain a constant distance between the spray nozzle and the interface surface of the spraying equipment, and perform spraying operations according to a preset uniform speed path to ensure uniform coating thickness and avoid local accumulation or missed spraying. After the coating is applied, transfer the workpiece to be tested to a clean, dry environment with constant temperature and humidity, and allow the coating to gradually cure by natural air drying. For scenarios requiring rapid curing, low-temperature drying can be used, with the drying temperature controlled within a range that will not affect the performance of the workpiece or the coating effect. During the curing process, regularly observe the state of the coating to ensure that the coating is firmly bonded to the workpiece surface and free from defects such as bubbles, cracks, peeling, etc. After the coating is fully cured, the entire pretreatment process is completed, and finally, a test surface with stable thermal conductivity and uniform surface condition is obtained.

[0023] In this embodiment of the invention, by employing a series of technical means—including determining the precise location of the interface based on three-dimensional scanning data, automatically cleaning and removing surface impurities, finely polishing as needed through roughness measurement to obtain a uniform and smooth surface, and applying a thermosensitive enhancement coating to optimize thermal conductivity—the technical problems of inaccurate positioning of key detection areas, uneven surface impurities and roughness, uneven thermal excitation caused by inconsistent thermal conductivity characteristics, and distortion of thermal signals in traditional pretreatment are effectively overcome. This achieves the technical effects of accurately locking high-risk detection areas, ensuring the uniformity and stability of the surface to be inspected, and optimizing the quality of thermal response signals, thereby improving the overall accuracy and stability of the detection.

[0024] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Based on the pre-processed surface to be inspected, a coordinated control command with specific energy and duration parameters is obtained. Specifically, this includes: first, a comprehensive characteristic test is performed on the pre-processed surface to be inspected, including the uniformity of the surface thermal enhancement coating thickness, thermal emissivity, thermal conductivity of the surface material, and the actual area size of the inspected region; the detected surface characteristic data are transmitted to the central control unit, which performs a comprehensive analysis and calculation based on the substrate properties of the composite welded-milled part and thermal excitation test data of similar workpieces from the past; according to the surface's response to the thermal signal, specific thermal excitation energy parameters are determined that enable a uniform heat distribution on the surface without damaging the workpiece; simultaneously, based on the matching relationship between the thermal excitation energy and the cooling process, thermal excitation duration parameters are set to fully capture the thermal decay process; the central control unit integrates the determined thermal excitation energy parameters, thermal excitation duration parameters, and the synchronous trigger signal required for subsequent thermal imaging acquisition to generate a coordinated control command containing all key control information.

[0025] Step 2.2, based on the coordinated control command, synchronously trigger the pulse thermal excitation device and the infrared thermal imager to enter the ready state. Specifically, this includes: simultaneously sending the generated coordinated control command to the pulse thermal excitation device and the infrared thermal imager through a stable communication link; upon receiving the command, the pulse thermal excitation device immediately starts its internal self-test program to check the working status of the energy output module, temperature control module, and trigger response module, ensuring that each component is fault-free and performs stably. At the same time, it pre-charges the internal energy storage unit according to the energy parameters in the command, adjusting the device to a ready state that can start thermal excitation at any time; upon receiving the command, the infrared thermal imager also starts its self-test process, calibrating the lens focal length, adjusting the detector sensitivity, confirming the available space of the data storage module, preset the image acquisition mode according to the acquisition-related parameters in the command, and simultaneously calibrating its own trigger response mechanism to ensure that it can keep pace with the working rhythm of the pulse thermal excitation device; when both the pulse thermal excitation device and the infrared thermal imager complete their self-tests and reach the preset ready standards, they respectively send out ready signals. After confirming receipt of feedback from both parties, the synchronous ready state is completed.

[0026] Step 2.3 involves using the pulsed thermal excitation device, which is in a ready state, to perform pulsed uniform thermal excitation on the pre-treated surface to be inspected, thereby creating an excited thermal state on the surface. Specifically, this includes: after confirming that both the pulsed thermal excitation device and the infrared thermal imager are in a ready state, sending a formal thermal excitation start command to the pulsed thermal excitation device; the pulsed thermal excitation device, based on the energy and duration parameters in the coordinated control command, activating the energy output system and releasing pulsed thermal energy to the pre-treated surface to be inspected through an array of thermal excitation nozzles; during the thermal excitation process, the real-time temperature monitoring module on the device continuously monitors the temperature change of the surface to be inspected and dynamically adjusts the energy output intensity of each nozzle based on the temperature feedback data to ensure that the thermal energy uniformly covers the entire inspection area, avoiding local overheating or uneven thermal energy distribution; after the thermal excitation continues for the duration set by the coordinated control command, the pulsed thermal excitation device automatically stops energy output, at which point the surface to be inspected has formed a stable excited thermal state due to the absorption of uniform thermal energy.

[0027] Step 2.4: For the surface under inspection in an excited thermal state, an infrared thermal imager in a ready state records the thermal state changes of the surface during the cooling process at a set acquisition frequency to obtain the original dynamic thermal image sequence. Specifically, this includes: at the instant the pulse thermal excitation device stops outputting energy, the infrared thermal imager is immediately triggered to start image acquisition according to the synchronization mechanism in the coordinated control command; the infrared thermal imager continuously captures the thermal state changes of the surface under inspection during natural cooling at a preset acquisition frequency. The acquisition frequency setting fully considers the differences in thermal decay rate between microcrack areas and normal areas, as well as milled texture areas, to ensure accurate capture of subtle changes in thermal signals in different areas; during the acquisition process, the infrared thermal imager transmits the captured thermal image frames to the data storage unit in real time and stores them continuously in chronological order to form a complete original dynamic thermal image sequence.

[0028] In this embodiment of the invention, by employing a series of technical means—namely, determining the appropriate specific energy and duration parameters based on the characteristics of the surface to be inspected after preprocessing, synchronously triggering the pulse thermal excitation device and the infrared thermal imager to enter the ready state through coordinated control commands, performing pulsed uniform thermal excitation on the surface to be inspected, and completely recording the thermal state changes during the cooling process at a set acquisition frequency—the technical problems of poor matching between thermal excitation energy and duration, asynchronous thermal excitation and thermal image acquisition, uneven thermal excitation, and incomplete recording of thermal state changes during the cooling process leading to thermal signal distortion or missing key information, which in turn affects the separation effect of pseudo signals and crack signals, are overcome in traditional technologies. This achieves the technical effect of ensuring that the surface to be inspected obtains a uniform and stable excited thermal state, synchronously and accurately capturing the complete thermal response changes during the cooling process, obtaining high-quality original dynamic thermal image sequences, and improving the overall stability and data reliability of the inspection process.

[0029] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Based on the original dynamic thermal image sequence, perform temporal filtering on each frame of thermal image data in the sequence to obtain a denoised dynamic thermal image sequence. Specifically, this includes: First, performing inter-frame signal analysis on the original dynamic thermal image sequence to identify the signal characteristics of background thermal noise and stripe interference, which can mask the real thermal signal differences between crack areas and milled texture areas; Based on the analysis results, select an adaptive temporal moving average filtering method, which can dynamically adjust the filtering parameters according to the noise intensity; Process each frame of thermal image data in the original dynamic thermal image sequence pixel by pixel, extracting the temperature data of each pixel in the current frame and the adjacent frames before and after it, assigning different weights according to the reliability of the data in each frame, reducing the weight of frames with a high proportion of noise signal and increasing the weight of frames with a high proportion of real thermal signal, and obtaining the filtered temperature value of the pixel through weighted calculation; Performing filtering on each pixel in all frames in sequence to ensure that while removing background thermal noise and stripe interference, the real change trend of the thermal signal is fully preserved, finally obtaining a denoised dynamic thermal image sequence with reduced noise and clear thermal signal details.

[0030] Step 3.2: Based on the denoised dynamic thermal image sequence, for each pixel location, reconstruct the temperature change data along the time dimension to obtain the initial thermal decay curve for each pixel location. Specifically, this includes: firstly, performing an overall analysis of the initial thermal decay curves for all pixel locations, statistically analyzing the temperature extremes of all curves, including the highest and lowest temperature values ​​of each curve, and clarifying the temperature change range of the entire detection area; since pixels at different spatial locations may have slight differences in the amount of thermal excitation received or subtle differences in surface characteristics, the overall temperature level of the initial thermal decay curve may deviate, which will affect the comparability of the thermal decay characteristics of different pixels; based on the statistically obtained temperature extremes, determining a unified temperature standardization range, and processing the initial thermal decay curve of each pixel location point by point, converting the temperature value at each time point on the curve into a relative temperature value within the standardization range, so that the thermal decay curves of all pixel locations are under the same temperature scale, eliminating the overall temperature level differences caused by different spatial locations, allowing direct comparison of the thermal decay trends of pixels in different regions, and finally obtaining a standardized thermal decay curve with a unified standard.

[0031] Step 3.3 involves standardizing the initial thermal decay curves for each pixel location to eliminate overall temperature level differences caused by spatial location, resulting in standardized thermal decay curves. This includes: firstly, analyzing the initial thermal decay curves for all pixel locations as a whole, statistically analyzing the temperature extremes of all curves, including the highest and lowest temperature values, to clarify the temperature variation range of the entire detection area; since pixels at different spatial locations may have slight differences in thermal excitation received or subtle differences in surface characteristics, leading to deviations in the overall temperature level of the initial thermal decay curves, these deviations affect the comparability of thermal decay characteristics of different pixels; based on the statistically obtained temperature extremes, a unified temperature standardization range is determined, and the initial thermal decay curve for each pixel location is processed point-by-point, converting the temperature value at each time point on the curve into a relative temperature value within the standardization range, ensuring that the thermal decay curves for all pixel locations are on the same temperature scale, eliminating overall temperature level differences caused by spatial location, allowing direct comparison of thermal decay trends of pixels in different regions, and ultimately obtaining standardized thermal decay curves with a unified standard.

[0032] Step 3.4: Extract characteristic parameters representing thermal decay from the standardized thermal decay curve, including the temperature decay rate and curve shape feature parameters for a specific time interval, to obtain the thermal decay curve features for each pixel. Specifically, for each standardized thermal decay curve, first divide it into two key time intervals: the rapid cooling stage after thermal excitation and the subsequent slow cooling stage. The difference in thermal decay characteristics between these two stages can effectively reflect the differences in the internal structure of the material. Within each time interval, calculate the ratio of temperature change to time change to obtain the temperature decay rate for that specific time interval and the crack region. Due to impaired heat conduction, the temperature decay rate differs significantly from that of the normal area and the milled texture area. Subsequently, the morphological characteristics of the standardized thermal decay curve are analyzed. By identifying parameters such as the inflection point position, the number of inflection points, the slope change amplitude of the rising and falling segments of the curve, and the smoothness of the curve, characteristic parameters representing the curve shape are extracted. These parameters include the time value corresponding to the curve inflection point, the maximum and minimum values ​​of the slope change rate, and the curvature of the curve. The temperature decay rate of a specific time interval is integrated with the curve shape characteristic parameters to obtain a comprehensive feature set that fully reflects the thermal decay characteristics of the pixel, i.e., the thermal decay curve characteristics of each pixel.

[0033] In this embodiment of the invention, by employing a series of technical means—namely, performing temporal filtering and noise reduction on each frame of the original dynamic thermal image sequence, reconstructing the initial thermal decay curve along the time dimension based on each pixel position in the denoised sequence, standardizing the initial thermal decay curve to eliminate the overall temperature level differences caused by spatial location, and extracting the temperature decay rate and curve morphology feature parameters for a specific time interval from the standardized curve—the technical problems of significant background thermal noise and stripe interference in the original thermal image sequence, lack of comparability of temperature change data of pixels at different spatial locations, and masking of key information on thermal decay characteristics, which in turn affect the accuracy of subsequent pseudo-signal and crack signal separation and micro-crack identification, are overcome. This achieves the technical effect of obtaining a standardized thermal decay curve with low noise and good consistency, accurately extracting the core feature parameters reflecting the differences in heat conduction in different regions, and improving the reliability of signal analysis and the accuracy of crack identification in the overall detection process.

[0034] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 integrates the thermal decay curve features of each pixel with its corresponding spatial location information to construct a feature matrix containing thermal decay features and spatial distribution. Specifically, this includes: First, uniformly numbering all pixels and establishing a unique pixel index system to ensure that all data of each pixel can be accurately associated through the index; Next, retrieving the thermal decay curve feature data of each pixel one by one. This data includes the temperature decay rate of a specific time interval extracted in the previous steps, covering the decay rate values ​​of the rapid cooling stage and the slow cooling stage after thermal excitation, as well as curve shape feature parameters, such as the time corresponding to the curve inflection point, the number of inflection points, the slope change range of each curve segment, the curvature of the curve, and other detailed parameters to ensure that no key information reflecting thermal decay characteristics is missed; Then, accurately extracting the spatial location information of each pixel in the three-dimensional coordinate system of the composite welding-milling workpiece under test, including the coordinate data of the three directions of X-axis, Y-axis, and Z-axis. This coordinate data comes from the previous three-dimensional scanning results of the workpiece under test to ensure the accuracy of the spatial location; A data association table was established, with pixel indexes as the core. Each pixel's thermal decay curve characteristic parameters were mapped one-to-one with its corresponding 3D spatial coordinates. The characteristic parameters and coordinates under each index were checked individually to identify missing, duplicate, or misaligned data, ensuring accurate matching for each set of data. Subsequently, the data was organized according to preset matrix construction rules, with each pixel as a row in the matrix. The columns of the matrix sequentially listed the pixel's 3D spatial coordinates, temperature decay rate during rapid cooling, temperature decay rate during slow cooling, curve inflection point time, number of inflection points, slope variation of each curve segment, degree of curve curvature, and all other thermal decay curve characteristic parameters. Standardize the format and precision of all data to avoid affecting subsequent analysis due to differences in data format; after the data arrangement is completed, perform a comprehensive integrity check on the entire matrix to check for blank data items, abnormal values ​​or data arrangement errors, and correct any problems found in a timely manner; at the same time, calculate the data dimensions of the matrix to ensure that the matrix fully carries the thermal attenuation characteristics and spatial distribution information of all pixels, and finally form a feature matrix with a regular structure, complete data and accurate correlation.

[0035] Step 4.2: Based on the feature matrix, the dominant direction of the milling texture is determined through spatial gradient analysis. A quantitative correlation model between the texture direction and the anisotropy of thermal conduction is established. Specifically, this includes: Based on the constructed feature matrix, firstly, the spatial distribution coordinate data of all pixels in the matrix and the key thermal decay curve feature parameters are extracted. The key feature parameters selected first are the temperature decay rate and the number of curve inflection points, which are sensitive to differences in thermal conduction. These two parameters can most directly reflect the differences in thermal conduction characteristics in different regions. Then, the spatial gradient analysis process is started. Taking each pixel as the center, multiple neighboring pixels are selected as the analysis neighborhood. The neighborhood range is reasonably set according to the pixel density of the area to be inspected to ensure that it can cover sufficient surrounding information without blurring the gradient information due to an excessively large range. For each central pixel and each neighboring pixel in the neighborhood, the difference between them in the key thermal decay feature parameters is calculated. Then, combined with the spatial distance between them, the spatial gradient value of the thermal decay feature in that direction is obtained. The spatial gradient values ​​of the central pixel in multiple different directions, such as horizontal, vertical, 45-degree, and 135-degree, are calculated in turn to ensure that the variation law of thermal decay characteristics in various spatial directions is fully captured. After calculating the spatial gradient of all pixels, all gradient values ​​are categorized and statistically analyzed by direction, and the frequency and average intensity of gradient values ​​in each direction are calculated. Since the milling texture is regularly distributed, its corresponding thermal attenuation characteristics exhibit a clear directionality in space, meaning that the gradient value in a specific direction appears more frequently than in other directions, and the average intensity remains consistently stable. This direction, with the highest gradient value proportion and the most regular variation, is identified as the dominant direction of the milling texture. To establish a quantitative correlation model between texture direction and thermal anisotropy, multiple test segments parallel to the dominant direction are selected, and multiple test segments parallel to the perpendicular direction are selected. These test segments are evenly distributed throughout the inspection area to ensure coverage of texture features at different locations. Then, the thermal attenuation rate of all pixels on each test segment is calculated. The average value and variance of the data were analyzed to determine the difference in heat conduction efficiency between the dominant direction and the direction perpendicular to the dominant direction, i.e., the difference in the magnitude and stability of the heat attenuation rate. A large amount of test data at different locations and texture densities were collected to analyze the variation of heat conduction anisotropy parameters when the dominant direction changes, such as the trend of the ratio of the heat conduction coefficient in the dominant direction to that in the perpendicular direction with the texture direction. Through statistical analysis and pattern fitting of these data, the quantitative relationship between texture direction and heat conduction anisotropy was clarified. For example, when the dominant direction of the texture is at a certain angle, the heat conduction coefficient has a specific value in that direction and a corresponding value in the perpendicular direction. Finally, a quantitative correlation model that accurately describes the heat conduction anisotropy characteristics under different texture directions was constructed.

[0036] Step 4.3 involves inputting the feature matrix into the correlation model and decoupling the features through principal component analysis. The data is then projected onto the feature subspace spanned by the texture regularity pattern and the orthogonal complement space. Specifically, this includes: firstly, preprocessing the constructed feature matrix to eliminate the influence of differences in data dimensions. Although the thermal decay curve has been standardized in previous steps, the feature matrix contains both spatial coordinates and thermal decay feature parameters, which have different dimensions. Therefore, a unified normalization process is needed to adjust all data to the same numerical range to ensure the accuracy of the principal component analysis results. After preprocessing, the feature matrix is ​​fully input into the established quantitative correlation model between texture direction and thermal conduction anisotropy. The correlation model first identifies the actual dominant direction of the milling texture within the inspection area based on the spatial coordinate data and thermal decay feature parameters in the feature matrix. Then, based on the quantitative correlation relationship stored in the model, it generates the thermal conduction anisotropy feature data corresponding to this dominant direction. This data is then used to filter and enhance the original data in the feature matrix, highlighting the signal features related to the regular texture distribution and weakening irrelevant interference signals. The principal component analysis process is then initiated. First, the covariance matrix of the preprocessed feature matrix is ​​calculated. The covariance matrix reflects the degree of linear correlation between different feature parameters in the matrix. By analyzing the covariance matrix, its eigenvalues ​​and corresponding eigenvectors are solved. The magnitude of the eigenvalue represents the information contribution of the corresponding eigenvector. Based on the ranking of the eigenvalues, the first few eigenvectors with larger eigenvalues ​​are selected as principal components. These principal components can reflect the main data change patterns in the feature matrix. Since the milling texture is regularly distributed, its corresponding thermal signals have strong spatial correlation and regularity. This regularity is reflected in the first few principal components with larger eigenvalues. Therefore, a feature subspace spanned by the regular pattern of the texture is constructed based on the key principal components. This subspace is specifically used to carry all thermal signal data related to the regular distribution of the milling texture. Simultaneously, based on the principle of spatial decomposition in linear algebra, an orthogonal complementary space perpendicular to the feature subspace is constructed. These two spaces are non-overlapping and jointly cover the entire original data space. The orthogonal complementary space is specifically used to carry signal data unrelated to texture regular patterns. Finally, each data point in the feature matrix processed by the correlation model is projected onto the texture regular pattern feature subspace and the orthogonal complementary space according to the projection rules obtained from principal component analysis. During the projection process, the distribution weight of each data point in the two subspaces is determined by calculating the projection coefficients between the data points and the two subspaces. This ensures that signal components related to texture rules are mainly concentrated in the feature subspace, while potential anomalous signal components unrelated to texture fall entirely into the orthogonal complementary space, achieving effective separation of regular and anomalous signals at the data space level.

[0037] Step 4.4: Extract signal components unrelated to the regular texture pattern in the orthogonal complement space. These unrelated signal components are treated as separated anomalous thermal signals. Specifically, this includes: First, performing a comprehensive feature analysis on all signal components projected into the orthogonal complement space, considering both spatial distribution and temporal variation. In terms of spatial distribution, by drawing a spatial distribution map of the signal components, observe whether their distribution pattern is random, and calculate parameters such as the contour regularity and orientation consistency of the signal region. In terms of temporal variation, analyze whether the changes in the thermal decay curves corresponding to the signal components conform to the characteristics of impeded heat conduction in the crack region, i.e., whether there are abnormal temperature decay rates or abrupt changes in curve shape. Since the orthogonal complement space theoretically has stripped away most of the regular texture signals, a small number of regular signal components weakly correlated with the texture direction may still remain. The spatial distribution of these residual signals usually still has a certain regularity, such as exhibiting a striped distribution parallel to the dominant texture direction, or having a highly consistent orientation. By comparing their spatial distribution parameters with the previously determined dominant milling texture direction, select signal components with an orientation consistent with the dominant texture direction and high contour regularity, determine them as residual texture-related signals, and remove them. Based on the random distribution characteristics of stress cracks mentioned in the background technology, this study focuses on retaining signal components with irregular spatial distribution, random orientation, and irregular contours in the orthogonal complement space. Further analysis of the thermal response intensity of these retained signals is conducted, identifying and eliminating signal components with thermal response intensities significantly lower than normal noise levels as invalid noise, ensuring that the remaining signal components are all anomalous signals with actual physical significance. To further improve the purity of the anomalous signals, spatial continuity verification is performed on the retained signal components. Anomalous signals in crack regions typically exhibit some spatial continuity, while isolated random noise signals show a discrete distribution. By analyzing the spatial connectivity of the signal components, completely isolated discrete signal points are eliminated, retaining signal regions with continuous distribution characteristics. Finally, the signal components after multiple rounds of screening and verification are integrated. These signal components are completely unrelated to the regular patterns of milling textures and possess typical characteristics of crack signals such as random spatial distribution, irregular contours, abnormal thermal response, and spatial continuity, thus identifying them as separated anomalous thermal signals.

[0038] In this embodiment of the invention, a series of technical means are employed to overcome the technical problems of traditional technologies, such as lack of signal separation mechanism for milling texture characteristics, inability to establish correlation between texture and thermal conductivity anisotropy, difficulty in effectively separating regular thermal signals corresponding to milling textures from abnormal thermal signals corresponding to cracks, resulting in high false positive rates and low accuracy in microcrack identification. These means include integrating the thermal attenuation curve features of each pixel with corresponding spatial location information to construct a feature matrix, determining the dominant direction of milling texture through spatial gradient analysis and establishing a quantitative correlation model between texture direction and thermal conductivity anisotropy, inputting the feature matrix into the correlation model and then performing feature decoupling through principal component analysis and projecting it onto the texture regular pattern feature subspace and orthogonal complement space, and extracting signal components unrelated to texture regular patterns in the orthogonal complement space. This achieves the technical effect of accurately removing the interference of texture pseudo signals on crack detection, obtaining high purity and strong identification of crack abnormal thermal signals, and improving the accuracy and reliability of surface stress crack detection in composite welded-milled parts.

[0039] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Based on the separated abnormal heat signals and their corresponding spatial coordinate information, the discrete point set spatial normalization interpolation method is used for processing. This method involves converting irregularly distributed spatial coordinate points into uniformly distributed spatial coordinate points in a normalized coordinate system through interpolation calculations to obtain normalized spatial coordinate information. Specifically, this includes: First, comprehensively organizing the separated abnormal heat signals and their corresponding spatial coordinate information, verifying the correspondence between each abnormal heat signal data point and the original spatial coordinates to ensure accurate matching, and identifying and eliminating invalid data points with missing data, misaligned coordinates, or duplicate signals to form a complete and accurately correlated original discrete data set. Next, boundary identification is performed on the spatial coordinate information in the original discrete data set. The spatial coordinates of all valid data points are traversed, and the maximum and minimum values ​​in each coordinate dimension are recorded to determine the spatial boundary range of the original abnormal heat signal distribution area. Based on the identified spatial boundary range and detection accuracy requirements, the creation parameters for the normalized mesh are determined. Considering the actual size of the area to be inspected and the precision requirements for microcrack detection, the number of rows and columns of the mesh are reasonably set to ensure that the mesh node density can adequately capture the detailed changes in abnormal thermal signals, while avoiding excessive computation due to overly dense meshes. The uniform step size of the mesh in each coordinate dimension is calculated by dividing the difference between the maximum and minimum values ​​of the boundary range by the number of mesh rows or columns, ensuring that the mesh nodes are evenly distributed in the normalized coordinate system. Then, the normalized coordinate matrix is ​​instantiated in memory, and the mesh is generated from the boundary according to the set number of rows and columns. Starting with the minimum value, normalized spatial coordinates for each grid node are generated sequentially at calculated uniform step sizes. Each grid node corresponds to a unique normalized coordinate. All nodes together form a uniform normalized spatial coordinate system covering the entire detection area, ensuring that the system can fully cover the distribution range of the original discrete data points, and that each original data point can correspond to a corresponding position in the normalized grid. Finally, a mapping relationship between the original irregular spatial coordinates and the normalized spatial coordinates is established. By calculating the spatial distance from each original data point to the normalized grid node, the corresponding position of each original data point in the normalized coordinate system is determined.

[0040] Step 5.2: Based on the normalized spatial coordinate information, the abnormal thermal signal is resampled and interpolated to obtain the interpolated abnormal thermal signal. Specifically, this includes: First, based on the generated normalized spatial coordinate information, a scientific resampling strategy is formulated. According to the node density of the normalized grid and the variation characteristics of the abnormal thermal signal, the distribution of sampling points is determined to ensure that the sampling points can uniformly cover the entire normalized coordinate system. At the same time, in areas where the abnormal thermal signal changes drastically, the sampling points are densified to retain the detailed features of the signal and avoid the loss of weak signals from microcracks due to insufficient sampling. Next, according to the formulated sampling strategy, the separated abnormal thermal signal is resampled. For each node in the normalized grid, the original abnormal thermal signal data points within a certain range around it are found. By calculating the spatial distance between the node and the surrounding original data points, the influence weight of each original data point on the sampling point is determined. The closer the original data point is, the higher its weight, and the farther the distance is, the lower its weight, ensuring that the sampling result truly reflects the signal strength at that location. Then, an interpolation method suitable for the continuity characteristics of the thermal signal is selected for interpolation calculation. Considering that the abnormal thermal signal has a certain continuity in space and that the signal change corresponding to the microcrack has a gradual characteristic, an interpolation algorithm that can take into account both signal smoothness and detail preservation is adopted. Based on the signal value and corresponding weight of the original data points around each sampling point, the interpolated signal value of the sampling point is obtained by weighted calculation. The interpolation calculation of all nodes in the normalized grid is completed in sequence to form the preliminary interpolated abnormal thermal signal. During the interpolation process, the signal gradient is monitored in real time to verify the reasonableness of the interpolation results. If the signal gradient in a certain region changes abnormally drastically, exceeding the range corresponding to normal heat conduction, the original data points and interpolation parameters for that region need to be rechecked to avoid signal distortion caused by the interpolation algorithm. Simultaneously, the interpolated signal is smoothed to eliminate any minor fluctuations that may be introduced during interpolation, ensuring signal continuity and stability. Finally, the integrity and consistency of the interpolated abnormal thermal signal are checked to confirm that all regularized grid nodes have obtained valid signal values, with no blank or abnormal values. The signal distribution characteristics before and after interpolation are compared to ensure that key features in the original abnormal thermal signal are fully preserved and clearer, transforming the originally irregularly distributed abnormal thermal signal, which may have data gaps, into a spatially uniform and continuous interpolated abnormal thermal signal.

[0041] Step 5.3 involves inputting the interpolated abnormal thermal signal and normalized spatial coordinates into a pre-trained neural network model. The neural network model learns the complex mapping relationship between the interpolated abnormal thermal signal and the normalized spatial distribution, and outputs an optimized abnormal thermal signal. Specifically, this includes: First, clarifying the foundation for the pre-trained neural network model. The model uses a large amount of inspection data from composite welded-milled parts containing real microcracks and crack-free components as training samples. The sample data includes the obtained interpolated abnormal thermal signal, the corresponding normalized spatial coordinates, and the real crack locations and signal characteristics manually labeled with ultrasonic testing results. The model employs a U-Net structure combined with an attention gating mechanism. The encoder part of t extracts multi-scale spatial features of the signal through convolutional layers, such as the local signal intensity distribution and edge contour of microcracks. The decoder part recovers signal details through upsampling. The attention gating mechanism is used to focus on the signal features at the junction of the weld heat-affected zone and the milled surface, suppressing invalid signals in other areas and improving the sensitivity to microcrack signals. During model training, the mean squared error loss function is used to measure the difference between the predicted signal and the true label. The network parameters are iteratively adjusted through a stochastic gradient descent optimizer. The training rounds are set to 200 rounds. After each round, the performance is evaluated on the validation set until the signal reconstruction accuracy on the validation set reaches more than 96%, ensuring that the model has the ability to accurately optimize weak anomalous signals. The data input to the model is preprocessed, and the interpolated abnormal thermal signals are converted into a three-dimensional tensor, with each element corresponding to the signal intensity of a grid node. Simultaneously, the normalized spatial coordinates are converted into relative coordinates with the grid's starting point as the origin, serving as auxiliary input features to enhance the model's ability to learn the correlation between spatial location and signal distribution. The two types of data are concatenated along the channel dimension to form the model's input tensor, ensuring that the data dimension matches the model's input layer. Then, the neural network model is launched for inference computation. The input tensor first enters the U-Net encoder, where multiple convolutional kernels extract multi-scale features of the signal. Small-sized convolutional kernels capture local signal abrupt changes in microcracks, while large-sized kernels capture the overall distribution trend of the signal. Pooling layers retain key features while reducing dimensionality, avoiding information loss. An attention gating mechanism, through learning a weight matrix, assigns higher weights to the signal features at the interface between the weld heat-affected zone and the milled surface, strengthening the focus on microcrack signals in this area. The model's decoder upsamples the features extracted by the encoder back to the original spatial dimension, and combines skip connections to fuse feature information at different scales. It learns the complex mapping relationship between the interpolated anomalous thermal signal and the normalized spatial distribution, including the intensity attenuation law of the microcrack signal at different spatial locations and the signal difference characteristics with the surrounding normal area. A nonlinear activation function is used to simulate the nonlinear changes of the signal. Finally, the output layer reconstructs and optimizes the signal, enhancing the weak anomalous signal corresponding to the microcrack, suppressing residual random noise, and improving the feature contrast of the signal. Finally, the optimized anomalous thermal signal output by the model is post-processed: a signal intensity threshold is set, invalid areas below the threshold are removed, and high signal areas that may contain cracks are retained; Gaussian filtering is used to eliminate possible local fluctuations in the model output, ensuring the spatial continuity of the signal, and finally obtaining the optimized anomalous thermal signal.

[0042] In this embodiment of the invention, a series of technical means are employed, including using a spatial normalization interpolation method based on discrete point sets to convert irregular spatial coordinate points corresponding to abnormal thermal signals into uniformly distributed normalized spatial coordinate information; resampling and interpolating the abnormal thermal signals based on these normalized coordinates; and inputting the interpolated signals and normalized coordinate information into a pre-trained neural network model to learn complex mapping relationships and optimize the signals. This overcomes the problems of irregular spatial distribution, missing or discontinuous signal data, and masking of weak abnormal features corresponding to microcracks in the separated abnormal thermal signals. Simultaneously, it enhances the abnormal signal features corresponding to microcracks, suppresses residual noise interference, and obtains high-quality optimized abnormal thermal signals. This provides strong support for accurately distinguishing between texture pseudo-signals and real stress cracks through multi-dimensional criteria, further improving the accuracy and reliability of detection.

[0043] In a preferred embodiment of the present invention, step 5.1 above may include: Step 5.11: Based on the separated abnormal thermal signals and their corresponding irregularly distributed spatial coordinate information, an original spatial data point set is constructed. This specifically includes: First, collecting all separated abnormal thermal signal data. These signals are the thermal anomaly signals related to potential stress cracks, retained only after the milling texture pseudo-signals have been removed in the previous signal separation step. Simultaneously, retrieving the original spatial coordinate information corresponding to each abnormal thermal signal. This coordinate information is based on the three-dimensional scanning results of the composite welded-milled workpiece under test, accurately reflecting the actual position of each signal point on the workpiece surface. Furthermore, due to the randomness of crack distribution and the discreteness of the original detection data... These coordinates are irregularly distributed. Subsequently, a data association mechanism is established to match the intensity value, thermal response characteristics, and other parameters of each abnormal thermal signal with its corresponding three-dimensional spatial coordinates one by one, ensuring that each signal data can find a unique spatial location identifier. During the matching process, problems such as signal-coordinate misalignment, duplicate records, null values, or abnormal values ​​that may occur during data transmission or processing are checked one by one, and defective data is removed. Finally, all accurately matched, complete, and valid abnormal thermal signals and their corresponding irregular spatial coordinate information are integrated and organized into an original spatial data point set according to a unified data format.

[0044] Step 5.12 involves identifying the distribution range of the original spatial data point set and calculating the creation parameters of a rectangular normalized grid that completely covers all data points based on the identified boundary range. Specifically, this includes: First, a complete traversal of the constructed original spatial data point set is performed to extract the three-dimensional spatial coordinate values ​​of each data point, focusing on the X-axis and Y-axis coordinates. Through traversal statistics, the maximum and minimum coordinate values ​​in the X-axis direction, the Y-axis direction, and the Z-axis direction are determined to clarify the overall distribution range of the original spatial data point set in three-dimensional space. Next, based on the identified three-dimensional spatial boundary range, the creation target of the rectangular normalized grid is determined. The grid must completely enclose all original data points, and the grid nodes must be evenly distributed in each axis direction to meet the accuracy requirements of subsequent signal interpolation and feature analysis. Considering the microscale characteristics of microcracks in the background technology, the mesh density needs to be adapted to the microcrack detection requirements to avoid signal detail loss due to excessive node spacing. Then, the mesh creation parameters are calculated: the length of the mesh in the X-axis direction is determined by the difference between the maximum and minimum coordinates in the X-axis direction, the length in the Y-axis direction is determined by the difference in coordinates in the Y-axis direction, and the thickness in the Z-axis direction is determined by the difference in coordinates in the Z-axis direction. Then, the mesh node spacing in each axis direction is determined according to the preset detection accuracy, and the number of mesh nodes in each axis direction is calculated by dividing the length or thickness by the step size. At the same time, the start coordinates and end coordinates of the mesh are set to ensure that the mesh can completely cover the distribution range of the original data point set. Finally, the start coordinates, end coordinates, number of nodes in each axis direction, node spacing and other parameters of the mesh are integrated to form a complete rectangular regularized mesh creation parameter set.

[0045] Step 5.13: Based on the mesh creation parameters, instantiate a normalized coordinate matrix in memory. Each element in the matrix represents a normalized spatial coordinate point uniformly distributed within the detection area. All normalized spatial coordinate points together constitute normalized spatial coordinate information. Specifically, this includes: First, allocating corresponding storage space in the memory of the manufacturing detection system; initializing a three-dimensional normalized coordinate matrix according to the determined mesh creation parameters. The matrix dimension is consistent with the number of mesh nodes, i.e., the number of rows corresponds to the number of nodes in the X-axis direction, the number of columns corresponds to the number of nodes in the Y-axis direction, and the number of layers corresponds to the number of nodes in the Z-axis direction. Each element in the matrix will be used to store the three-dimensional coordinate information of a normalized spatial coordinate point; then, according to the set mesh starting coordinates and node intervals, sequentially generate the normalized spatial coordinate points corresponding to each matrix element. Starting from the X-axis starting coordinates, calculate the X-axis coordinates of each node sequentially according to a preset step size; For each X-axis coordinate, the Y-axis coordinate of each node is calculated sequentially according to the starting Y-axis coordinate and step size; for each combination of X and Y coordinates, the Z-axis coordinate of each node is calculated sequentially according to the starting Z-axis coordinate and step size. Through ordered iterative calculation, it is ensured that each element in the matrix corresponds to a coordinate point uniformly distributed in three-dimensional space, and the grid formed by all coordinate points completely covers the distribution range of the original spatial data point set. During the coordinate point generation process, the numerical accuracy of each coordinate point is verified in real time to ensure that it meets the grid creation parameter requirements and that there are no cases of exceeding the boundary range or uneven spacing. After generation, the integrity of the entire normalized coordinate matrix is ​​checked to confirm that all matrix elements have correctly stored the corresponding coordinate information and that there are no null values ​​or erroneous values. Finally, all uniformly distributed normalized spatial coordinate points contained in the normalized coordinate matrix in memory are integrated to form normalized spatial coordinate information.

[0046] In this embodiment of the invention, by employing a series of technical means—namely, constructing an original spatial data point set based on the separated abnormal thermal signals and corresponding irregular spatial coordinate information, identifying the boundaries of the distribution range of the data point set and calculating the parameters for creating a rectangular normalized grid that completely covers all data points, and instantiating a normalized coordinate matrix in memory based on the parameters to obtain uniformly distributed normalized spatial coordinate information—the technical problems of irregular spatial coordinate distribution and scattered data points corresponding to the separated abnormal thermal signals, which lead to difficulty in accurately associating spatial positions with signal features and easy data loss or matching deviations during subsequent signal interpolation, feature analysis, and model calculation, are effectively overcome. This achieves the technical effect of normalizing and homogenizing the spatial distribution of abnormal thermal signals, clarifying the precise spatial association of each signal point, and ensuring the accuracy and coherence of signal processing.

[0047] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1: Based on the optimized abnormal thermal signal, receive and construct a two-dimensional abnormal signal data field containing spatial distribution and signal intensity. Specifically, this includes: First, receiving the abnormal thermal signal optimized by the neural network model. The signal has effectively enhanced the weak features corresponding to the microcracks and suppressed noise, and corresponds to the normalized spatial coordinate information. Next, determining the spatial range of the two-dimensional abnormal signal data field: using the extreme values ​​of the X-axis and Y-axis coordinates covered by the optimized abnormal thermal signal as boundaries, ensuring that the data field completely includes the interface between the welding heat-affected zone and the milled surface, and distributing the X-axis and Y-axis coordinates according to the normalized network... The grid is divided into continuous spatial units with uniform step size, each unit corresponding to a two-dimensional coordinate point (x, y). Then, the optimized abnormal thermal signal intensity value is associated with the corresponding two-dimensional coordinate point: the signal intensity value of each coordinate point (x, y) is directly taken from the value of that position in the optimized abnormal thermal signal, forming a two-dimensional data field with spatial coordinates (x, y) as the horizontal and vertical axes and signal intensity as the third dimension. Finally, the data field is smoothed to eliminate local small fluctuations, ensure the continuity of signal intensity in space, and retain the local high-intensity signal area that may be formed by microcracks.

[0048] Step 6.2: Based on the two-dimensional anomalous signal data field, extract the multi-dimensional identification feature vector for each anomalous signal region. The multi-dimensional identification feature vector includes contour features representing spatial continuity, shape features representing morphological complexity, and amplitude features representing thermal response intensity. Specifically, this includes: First, segmenting the two-dimensional anomalous signal data field by setting a signal intensity threshold. Regions in the data field with signal intensity exceeding the threshold are divided into independent anomalous signal regions, each considered as a potential object to be identified. The segmented regions are then labeled to ensure each region has a unique identifier. Next, extracting the contour features representing spatial continuity: For each anomalous signal region, extract its contour boundary using an edge detection algorithm, and calculate the continuous length, number of breaks, and closure of the contour. For example, the pseudo-signal region corresponding to milling texture has a long continuous length, few breaks, and low closure, while the contour of the crack region may have more branches. Multiple fractures and high closure are identified. Then, shape features characterizing morphological complexity are extracted, and shape parameters for each region are calculated, including the tortuosity of the contour, the ratio of the actual contour length to the perimeter of the region's minimum bounding rectangle, the aspect ratio, the ratio of the long side to the short side of the region's minimum bounding rectangle, and the ratio of area to perimeter, reflecting shape compactness. Texture pseudo-signals, due to their regular distribution, have low tortuosity, high aspect ratio, and compact shape; crack regions, due to their irregular shape, have high tortuosity, large aspect ratio fluctuations, and loose shape. Finally, amplitude features characterizing thermal response intensity are extracted: the peak value, mean value, and gradient rate of change of signal intensity within each region are calculated. Real cracks, due to impeded heat conduction, typically exhibit localized high-intensity peaks and more drastic intensity gradient changes; texture pseudo-signals have a more uniform intensity distribution, less prominent peaks, and gentler gradient changes. Contour features, shape features, and amplitude features are integrated in a preset order to form a multi-dimensional identification feature vector for each abnormal signal region.

[0049] Step 6.3: Based on the multidimensional recognition feature vectors, an unsupervised clustering algorithm is used to classify all anomalous signal regions by pattern. Signal regions with regular contours, a single orientation, and high correlation with texture direction are identified as texture pseudo-signals, while signal regions with irregular contours, random orientation, and no relation to texture direction are identified as potential real stress cracks, thus obtaining the classification results. Specifically, this includes: First, collecting multidimensional recognition feature vectors of all anomalous signal regions to construct a feature vector dataset. The dataset covers parameters such as the continuous length of the contour, the number of fractures, the tortuosity, the aspect ratio, and the peak intensity of each region. These parameters can reflect the core differences between texture pseudo-signals and real cracks. Next, selecting an unsupervised clustering algorithm suitable for high-dimensional feature classification, which does not require a preset number of categories and can automatically group based on feature similarity, and setting the algorithm... The core parameters, neighborhood radius and minimum sample size, enable the algorithm to effectively distinguish between two types of regions with different features. Then, a clustering algorithm is launched to classify the feature vector dataset. By calculating the Euclidean distance between the feature vectors of different regions, regions with similar features are grouped into one category. The feature vectors of one type of region are characterized by continuous contours, low tortuosity, large aspect ratio, and uniform intensity distribution. Furthermore, by comparing with the dominant direction of the milling texture, it is found that the angle between the direction and the texture direction is less than 10 degrees. This type is identified as a texture pseudo-signal region. The feature vectors of the other type of region are characterized by a high number of contour breaks, high tortuosity, irregular aspect ratio, prominent intensity peaks, and an angle between the direction and the dominant texture direction greater than 60 degrees or no fixed direction. This type is identified as a potential real stress crack region. After clustering, the classification results containing the identifiers of the two types of regions are output.

[0050] Step 6.4: Based on the classification results, a confidence assessment is performed on the identified potential real stress crack regions. The matching degree between the morphological, strength, and spatial characteristics and typical stress crack characteristics is comprehensively calculated to obtain a confidence score for each crack region. Specifically, this includes: focusing on the confidence assessment of potential real stress crack regions, firstly, based on historical real crack data verified by ultrasound, a typical stress crack feature library containing morphological, strength, and spatial characteristics is constructed as an assessment benchmark; then, for each potential crack region, its morphological matching degree, strength matching degree, and spatial matching degree with the feature library are calculated separately, and weighted summation is performed with a spatial matching degree of 0.4 and morphological and strength matching degrees of 0.3 each, to obtain a confidence score of 100 points; finally, by setting two key nodes of 40 and 60 points, regions with scores below 40 points are re-checked for misclassification, regions with scores between 40 and 60 points are marked as requiring review, and regions with scores above 60 points are confirmed as high-confidence potential cracks, ensuring the authenticity and reliability of the scores.

[0051] Step 6.5: Based on the confidence score, the final stress crack detection result is obtained to accurately distinguish between texture pseudo-signals and real stress cracks. Specifically, this includes: outputting the final detection result based on the confidence score. First, a threshold of 50 is set. Potential crack areas with scores above 50 are identified as real stress cracks, while potential crack areas with scores below 50 are classified as suspected noise or pseudo-signals. Texture pseudo-signals in the original classification are directly confirmed. Then, detailed parameters such as the spatial coordinates and contour dimensions of real cracks are extracted, and areas to be reviewed are listed separately and ultrasonic confirmation is recommended. Finally, pseudo-signals are removed, and real cracks are marked on the workpiece's two-dimensional drawing according to the confidence level using visualization technology. The final result, including the number of cracks, parameters, score, and visualization, is output.

[0052] In this embodiment of the invention, a series of technical means are employed to overcome the technical problems of traditional infrared thermal imaging detection technology, such as difficulty in accurately distinguishing milling texture pseudo signals from real stress crack signals in composite welded-milled parts, high false positive rate, and low accuracy of microcrack identification. These means include constructing a two-dimensional abnormal signal data field containing spatial distribution and signal intensity based on optimized abnormal thermal signals, extracting contour features representing spatial continuity, shape features representing morphological complexity, and amplitude features representing thermal response intensity, constructing a multi-dimensional identification feature vector, classifying by signal region feature patterns using an unsupervised clustering algorithm to distinguish texture pseudo signals from potential real stress cracks, evaluating the confidence of morphological intensity and spatial feature matching degree of potential crack regions, and outputting the final result based on the score. This achieves the technical effect of accurately distinguishing texture pseudo signals from real stress cracks, reducing the false positive rate, improving the accuracy and reliability of microcrack identification, reducing manual review costs, and providing strong protection for the service safety of composite welded-milled parts.

[0053] like Figure 2 As shown, embodiments of the present invention also provide an infrared thermal imaging detection system for surface stress cracks in composite welded-milled parts, comprising: The acquisition module is used to preprocess the interface between the milled surface and the weld heat-affected zone of the composite welded-milled workpiece to be tested, so as to obtain the preprocessed surface to be inspected. The acquisition module is used to uniformly thermally excite the pre-processed surface to be inspected using pulsed thermal excitation, while simultaneously using an infrared thermal imager to acquire dynamic thermal image sequences during the cooling process. The extraction module is used to extract the thermal decay curve features of each pixel during the cooling process by performing time-series analysis on the dynamic thermal image sequence. The separation module combines the characteristics of the thermal decay curve with the spatial information of the dynamic thermal image sequence. Based on the regular distribution characteristics of milling texture and the random distribution characteristics of stress cracks, it establishes a correlation model between texture direction and heat conduction anisotropy. Through principal component analysis, it separates the regular thermal signal caused by texture and the abnormal thermal signal caused by cracks to obtain the separated abnormal thermal signal. The optimization module is used to perform spatial interpolation on the separated abnormal heat signals and their corresponding spatial coordinate information to obtain the interpolated abnormal heat signals and normalized spatial coordinate information. The abnormal heat signals and normalized spatial coordinate information are then input into a pre-trained neural network model to learn the complex mapping relationship between the interpolated abnormal heat signals and the normalized spatial distribution, and the optimized abnormal heat signals are output. The processing module is used to identify abnormal thermal signals based on optimized signals through multi-dimensional criteria, so as to accurately distinguish between texture pseudo-signals and real stress cracks.

[0054] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method of infrared thermographic detection of surface stress cracks in a composite welded-milled workpiece, characterized in that, The method comprises: The milling surface of the composite welding-milling workpiece to be tested is pretreated at the junction with the welding heat-affected zone to obtain a pretreated surface to be inspected; The pretreated surface to be inspected is uniformly excited by a pulse thermal excitation method, and an infrared thermal imager is used to collect a dynamic thermal image sequence during cooling; Through time series analysis of the dynamic thermal image sequence, the thermal decay curve characteristics of each pixel point during cooling are extracted; Based on the regular distribution characteristics of the milling texture and the random distribution characteristics of the stress cracks, an association model of the texture direction and the thermal anisotropy is established by combining the thermal decay curve characteristics with the spatial information of the dynamic thermal image sequence, the regular thermal signals caused by the texture are separated from the abnormal thermal signals caused by the cracks through principal component analysis, and the separated abnormal thermal signals are obtained; The separated abnormal thermal signals and the corresponding spatial coordinate information are subjected to spatial interpolation processing to obtain interpolated abnormal thermal signals and regularized spatial coordinate information; the abnormal thermal signals and the regularized spatial coordinate information are input into a pre-trained neural network model to obtain a complex mapping relationship between the learned interpolated abnormal thermal signals and the regularized spatial distribution, and the optimized abnormal thermal signals are output; Based on the optimized abnormal thermal signals, multi-dimensional criteria are used for identification to accurately distinguish between texture false signals and real stress cracks.

2. The method of infrared thermographic testing for surface stress cracking of a composite welded-milled article of claim 1, wherein, The pretreated surface to be inspected is uniformly excited by a pulse thermal excitation method, and an infrared thermal imager is used to collect a dynamic thermal image sequence during cooling, comprising: Based on the pretreated surface to be inspected, a cooperative control instruction of specific energy and time length parameters is obtained; Based on the cooperative control instruction, the pulse thermal excitation device and the infrared thermal imager are synchronously triggered to enter a ready state; The pulse thermal excitation device in the ready state performs pulse uniform thermal excitation on the pretreated surface to be inspected, so that the surface to be inspected forms an excited thermal state; For the surface to be inspected in the excited thermal state, the infrared thermal imager in the ready state records the thermal state changes of the surface during cooling at a set acquisition frequency to obtain an original dynamic thermal image sequence. Through time series analysis of the dynamic thermal image sequence, the thermal decay curve characteristics of each pixel point during cooling are extracted, comprising:

3. The method of infrared thermographic testing for surface stress cracking of a composite welded-milled article of claim 2, wherein, Based on the original dynamic thermal image sequence, each frame of thermal image data in the sequence is subjected to time domain filtering denoising processing to obtain a denoised dynamic thermal image sequence; Based on the denoised dynamic thermal image sequence, for each pixel position, temperature change data is reconstructed along the time dimension to obtain an initial thermal decay curve for each pixel position; The initial thermal decay curve for each pixel position is subjected to data standardization processing to eliminate overall temperature level differences caused by different spatial positions, and a standardized thermal decay curve is obtained; Feature parameters representing thermal decay characteristics, including temperature decay rates and curve shape feature parameters in specific time intervals, are extracted from the standardized thermal decay curve to obtain thermal decay curve characteristics for each pixel point. Based on the regular distribution characteristics of the milling texture and the random distribution characteristics of the stress cracks, 4. The method of infrared thermographic testing for surface stress cracking of a composite welded-milled article of claim 3, wherein, ​ A correlation model between texture direction and thermal conduction anisotropy is established, and regular thermal signals caused by texture and abnormal thermal signals caused by cracks are separated by principal component analysis to obtain separated abnormal thermal signals, including: Integrate the thermal decay curve features of each pixel point with the corresponding spatial position information to construct a feature matrix containing thermal decay features and spatial distribution; Based on the feature matrix, the dominant direction of the milling texture is determined by spatial gradient analysis method, and a quantitative correlation model between texture direction and thermal conduction anisotropy is established; The feature matrix is input into the correlation model, and the feature decoupling is carried out by principal component analysis, and the data is projected into the feature subspace and orthogonal complement space spanned by the regular pattern of texture; In the orthogonal complement space, the signal components unrelated to the regular pattern of texture are extracted, and the unrelated signal components are taken as the separated abnormal thermal signals.

5. The method of infrared thermographic testing for surface stress cracking of a composite welded-milled part according to claim 4, wherein, The separated abnormal thermal signals and the corresponding spatial coordinate information are subjected to spatial interpolation processing to obtain interpolated abnormal thermal signals and normalized spatial coordinate information; the abnormal thermal signals and the normalized spatial coordinate information are input into the pre-trained neural network model to obtain the complex mapping relationship between the interpolated abnormal thermal signals and the normalized spatial distribution, and output the optimized abnormal thermal signals, including: Based on the separated abnormal thermal signals and the corresponding spatial coordinate information; the spatial regularization interpolation method of discrete point set is used for processing, wherein the spatial regularization interpolation method of discrete point set refers to converting irregularly distributed spatial coordinate points into uniformly distributed spatial coordinate points in a regularized coordinate system through interpolation calculation to obtain regularized spatial coordinate information; Based on the regularized spatial coordinate information, the abnormal thermal signals are resampled and interpolated to obtain the interpolated abnormal thermal signals; The interpolated abnormal thermal signals and the regularized spatial coordinate information are input into the pre-trained neural network model, the complex mapping relationship between the interpolated abnormal thermal signals and the regularized spatial distribution is learned through the neural network model, and the optimized abnormal thermal signals are output.

6. The method of infrared thermographic testing for surface stress cracking of a composite welded-milled article of claim 5, wherein, Based on the separated abnormal thermal signals and the corresponding spatial coordinate information; the spatial regularization interpolation method of discrete point set is used for processing, wherein the spatial regularization interpolation method of discrete point set refers to converting irregularly distributed spatial coordinate points into uniformly distributed spatial coordinate points in a regularized coordinate system through interpolation calculation to obtain regularized spatial coordinate information, including: Based on the separated abnormal thermal signals and the corresponding irregularly distributed spatial coordinate information, an original spatial data point set is formed; The distribution range of the original spatial data point set is identified, and the creation parameters of a rectangular regular grid completely covering all data points are calculated based on the identified boundary range; Based on the grid creation parameters, a regularized coordinate matrix is instantiated in memory, and each element in the matrix represents a uniformly distributed regularized spatial coordinate point in the detection area, and all regularized spatial coordinate points together constitute the regularized spatial coordinate information.

7. The method of infrared thermographic testing for surface stress cracking of a composite welded-milled article of claim 6, wherein, Based on the optimized abnormal thermal signals, multi-dimensional criteria are used for identification to accurately distinguish texture pseudo signals from real stress cracks, including: Based on the optimized abnormal heat signal, a two-dimensional abnormal signal data field containing spatial distribution and signal intensity is received and constructed; Based on the two-dimensional abnormal signal data field, a multi-dimensional recognition feature vector of each abnormal signal region is extracted, including contour features representing spatial continuity, shape features representing morphological complexity, and amplitude features representing thermal response intensity; Based on the multi-dimensional recognition feature vector, an unsupervised clustering algorithm is used to classify all abnormal signal regions, identifying signal regions with regular contours, single trends, and high correlation with texture direction as texture false signals, and signal regions with irregular contours, random trends, and no correlation with texture direction as potential real stress cracks, to obtain a classification result; Based on the classification result, a confidence evaluation is performed on the identified potential real stress crack regions, and the matching degree of morphological, intensity, and spatial features with typical stress crack features is comprehensively calculated to obtain a confidence score for each crack region; Based on the confidence score, a final stress crack detection result is obtained to accurately distinguish texture false signals from real stress cracks.

8. Infrared thermographic detection system of surface stress cracks in composite welded-milled pieces, which implements the method according to any one of claims 1 to 7, characterized in that, The method comprises: An acquisition module is configured to preprocess a milling surface and a welding heat-affected zone joint of a composite welding-milling workpiece to obtain a preprocessed detection surface; An acquisition module is configured to uniformly heat the preprocessed detection surface by pulse thermal excitation, and simultaneously collect a dynamic thermal image sequence during cooling using an infrared thermal imager; An extraction module is configured to extract thermal decay curve features of each pixel point during cooling by performing time series analysis on the dynamic thermal image sequence; A separation module is configured to combine the thermal decay curve features and the spatial information of the dynamic thermal image sequence, establish an association model of the texture direction and the thermal anisotropy based on the regular distribution characteristics of the milling texture and the random distribution characteristics of the stress cracks, separate the regular thermal signals caused by the texture and the abnormal thermal signals caused by the cracks through principal component analysis, and obtain separated abnormal thermal signals; An optimization module is configured to perform spatial interpolation on the separated abnormal thermal signals and corresponding spatial coordinate information to obtain interpolated abnormal thermal signals and regularized spatial coordinate information, input the abnormal thermal signals and regularized spatial coordinate information into a pre-trained neural network model, obtain a complex mapping relationship between the learned interpolated abnormal thermal signals and the regularized spatial distribution, and output optimized abnormal thermal signals; A processing module is configured to identify texture false signals and real stress cracks based on the optimized abnormal thermal signals and multi-dimensional criteria.

9. A computing device, comprising: One or more processors; A storage device is configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-7. The computer readable storage medium stores a program which, when executed by a processor, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, ​

Images