Engineering machinery aluminum alloy internal crack visual detection method fusing infrared images

Abstract

The application relates to the technical field of visual detection, and discloses an engineering machinery aluminum profile internal crack visual detection method fusing infrared images. Through five logically progressive steps of double-mode cooperative collection, visible light semantic zoning field construction, cold frame physical correction preliminary detection, iterative self-consistent final detection and orthogonal re-sampling arbitration, the method realizes the effects of continuously building an emissivity field based on a visible light reference to eliminate coefficient jumps, stripping environment reflection by real-time closed radiation transfer equation of a cold frame and a reference piece, upgrading correction recognition to bidirectional verification iteration closed loop by removing pixels in a preliminary detection mask and reconstructing statistical values outside the mask by interpolation, jointly solving the problem of going to a blind area and inverting crack burial depth by orthogonal re-sampling and physical constraint of a thermal diffusion equation, so that the detection result is upgraded from qualitative discovery to quantitative risk assessment.

Description

Technical Field

[0001] This application relates to the field of visual inspection technology, and in particular to a visual inspection method for internal cracks in aluminum profiles for engineering machinery that integrates infrared images. Background Technology

[0002] Infrared thermal imaging technology is widely used in industrial non-destructive testing. Visual inspection methods that integrate infrared and visible light images apply active thermal excitation to the workpiece, using an infrared thermal imager to acquire surface thermal image sequences and an industrial camera to acquire visible light images. After registration and fusion, internal defects are identified. In the field of engineering machinery manufacturing and maintenance, this method is used to detect internal cracks in aluminum profile workpieces such as frame longitudinal beams, booms, and outriggers, to identify potential hazards such as fatigue cracks and delamination before assembly or during service, ensuring the structural safety of the entire machine.

[0003] However, the surfaces of aluminum profiles used in engineering machinery exhibit various states, including machined bright surfaces, anodized surfaces, oil-stained areas, and scratched / weared areas. The emissivity of each state varies significantly. Low-emissivity areas reflect ambient heat radiation, creating false bright spots. Significant apparent temperature gradients form at the boundaries of these surface states. These pseudo-thermal anomalies resemble the morphology of crack hot spots, leading to both false alarms and false negatives. Patent CN103196564A, entitled "An Infrared Thermal Imaging Temperature Measurement Method for Correcting Surface Emissivity through Image Segmentation," discloses a scheme that segments infrared images to distinguish different emissivity regions and uses an emissivity meter to measure each region to establish an emissivity distribution matrix and correct the grayscale matrix. Patent CN117726598A, entitled "A Method and System for Defect Detection of Power Equipment by Fusion of Infrared and Visible Light Images," discloses a scheme that constructs a dual-modal feature extraction network to interactively fuse registered infrared and visible light images, and then combines this with a temperature judgment method to detect defects. The former uses infrared images as the segmentation benchmark, while infrared grayscale simultaneously encodes three types of information: emissivity, true temperature, and defect thermal disturbance. Crack hot spots sever the partition boundaries, and the partition benchmark has been contaminated by the defect to be inspected, with correction errors concentrated in the crack area. The latter uses visible light only for feature supplementation, without processing the spatial heterogeneity of emissivity, and false thermal anomalies directly enter the identification process. Neither of these methods can guarantee the reliability of internal crack detection in aluminum profiles. Summary of the Invention

[0004] This application proposes a visual inspection method for internal cracks in aluminum profiles for engineering machinery that integrates infrared images. It has the advantages of in-service online inspection, dual-modal physical closure and iterative self-consistent final inspection, in order to solve the problems in the prior art such as crack truncation caused by surface emissivity jump, lack of environmental reflection reference, blind zone of unidirectional thermal excitation, and structural silent zone caused by unidirectional series correction and identification.

[0005] To achieve the above objectives, this application adopts the following technical solution: a visual inspection method for internal cracks in aluminum profiles for engineering machinery that integrates infrared images, the specific steps of which are as follows; S1. Dual-modal collaborative acquisition: Deploy orthogonal thermal excitation sources, infrared thermal imagers, industrial cameras, and reference films to acquire cold frames and trigger excitation of infrared and visible light data; S2, Semantic Partitioning Field Construction: Visible light segmentation of surface state categories, and continuous emissivity field with gradient interpolation of the normal to the boundary of adjacent blocks; S3, Cold Frame Correction Initial Inspection: Solve the reflection term for cold frames and reference films, and grow an initial inspection mask with thermal contrast in the low threshold region; S4. Iterative self-consistent final inspection: Pixels within the initial inspection mask are removed from the baseline. The parameters are corrected by interpolating the emissivity of similar blocks outside the mask. A new mask is obtained after the second correction. The process is iterated until convergence, and the union set is used for verification. S5, Orthogonal Re-mining Decision: The slope of the temperature rise curve fitted outside the mask, the deviation from the negative half power law marks the suspected defect, and the thermal contrast peak value is used to invert the crack burial depth in one-dimensional thermal conduction.

[0006] Furthermore, in step S1, two sets of orthogonal linear array thermal excitation sources, an infrared thermal imager, and an industrial camera with a polarization filter optical path are arranged at the detection station. A diffuse reflection reference film with known emissivity is fixed at the edge of the field of view. The two devices are jointly calibrated to complete pixel-level registration. The orthogonal dual-set configuration reserves heat flow switching conditions for subsequent blind zone elimination. The reference film provides boundary constraints for environmental reflection term calibration.

[0007] Furthermore, in step S1, a cold frame before excitation is first acquired, and then a thermal excitation is triggered to continuously acquire an infrared thermal image sequence. Simultaneously, a visible light image after anti-reflection is acquired. The environmental reflection term of the closed radiative transfer equation is obtained between the cold frame and the reference film, and the polarization anti-reflection path is used to eliminate specular highlight interference. The homologous registration data lays the foundation for subsequent emissivity field construction and crack identification.

[0008] Furthermore, in step S2, the visible light image after eliminating reflection is input into the semantic segmentation model, and five types of surface states are divided pixel by pixel: machining bright area, anodized area, oil stain area, scratch and wear area, and dust-covered area. The pre-calibrated mapping lookup table is called to assign values ​​pixel by pixel. Visible light and crack thermal disturbance are not coupled, the partitioning benchmark is not affected by defect contamination, and structural defects caused by defect data correction are eliminated before detection.

[0009] Furthermore, in step S2, the boundary of adjacent blocks is linearly interpolated along the normal direction with distance weighting to generate a continuous emissivity field across the entire field of view. This eliminates coefficient jumps in block-shaped assignments, ensuring that the crack morphology across the surface state boundary remains intact and uninterrupted, thus providing a continuous coefficient field without jumps for subsequent cold frame correction and temperature inversion.

[0010] Furthermore, in step S3, the environmental reflection term is solved pixel by pixel using the cold frame and the measured temperature of the reference film in step S1, the radiation receiving equation is closed, and the environmental reflection is stripped off frame by frame and the real temperature field is inverted by substituting it into the continuous emissivity field in step S2. For pixels with extremely low emissivity, the absolute temperature is replaced by time-domain difference, the pseudo-thermal anomaly is stripped off by the physical equation, and the crack thermal disturbance is completely preserved.

[0011] Furthermore, in step S3, the normalized thermal contrast is calculated for the corrected thermal image sequence, and a preliminary suspected defect mask is generated by growth discrimination in the low threshold region. This mask serves as the iterative closed-loop trigger input for step S4, connecting the correction and identification links. The environmental reflection term is acquired in real time on site.

[0012] Furthermore, in step S4, pixels within the initial detection mask are removed from the calibration benchmark. The calibration parameters are reconstructed by interpolating the emissivity of similar blocks outside the mask according to the visible light label. A new mask is obtained through secondary discrimination. The process is iterated until convergence using the area overlap rate criterion. During oscillation, the union of the marked core area is taken. The transition zone is weighted and fused based on the confidence attenuation of the transition zone. The apparent emissivity of the intact area is used to back-infer the mapping lookup table to achieve self-calibration.

[0013] Furthermore, in step S5, the second set of orthogonal excitation sources is pre-set in S1, and the thermal map is independently executed from S2 to S4. The results of the two are combined. Cracks that are parallel to the heat flow but have weak thermal response are revealed under orthogonal heat flow and move towards the blind zone to be eliminated. The closed loop is upgraded from verifying that it has been found to finding that it has been missed, and the completeness of the detection is guaranteed by the physical lower limit.

[0014] Furthermore, in step S5, the slope of the temperature rise curve is fitted pixel by pixel in logarithmic coordinates for the region outside the final mask. Connected regions that deviate from the negative half power law and show decay hysteresis are marked as suspected defects. The process is then returned to S4 for re-verification. The crack burial depth is inverted at the peak of the thermal contrast according to one-dimensional thermal conduction, and evaluated in combination with the location, orientation, and depth.

[0015] The beneficial effects of this invention are as follows: This application provides a visual inspection method for internal cracks in aluminum profiles for engineering machinery that integrates infrared images. Through five logically progressive steps—dual-modal collaborative acquisition, visible light semantic partitioning field construction, cold frame physical correction for initial inspection, iterative self-consistent final inspection, and orthogonal re-acquisition adjudication—it achieves the following effects: constructing a continuous emissivity field based on visible light to eliminate coefficient jumps; real-time closure of the radiative transfer equation with cold frames and reference films to remove environmental reflections; pixel elimination within the initial inspection mask and statistical value interpolation reconstruction outside the mask to upgrade the correction and identification to a two-way verification iterative closed loop; and joint resolution of blind zones by orthogonal re-acquisition and the physical constraints of the thermal diffusion equation to invert the crack burial depth. This upgrades the detection results from qualitative discovery to quantitative risk assessment. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort: Figure 1 This is a flowchart illustrating the method steps of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] However, existing methods have fundamental limitations in the specific scenario of in-service online inspection of aluminum profiles for engineering machinery. Aluminum profile surfaces typically exhibit various surface states, including machined bright areas, anodized areas, oily areas, scratched and worn areas, and dust-covered areas, with significant differences in emissivity across these states. Existing methods often employ the assumption of global average emissivity or simple block-like partitioning, leading to emissivity jumps at surface state boundaries being misinterpreted as thermal anomalies during temperature inversion. This results in crack morphology being truncated or distorted at cross-zone boundaries. Furthermore, the strong specular reflection on aluminum profile surfaces and the unknown reflection components of environmental radiation prevent the radiative transfer equation from being closed, resulting in a lack of temperature inversion benchmarks and drift in inversion results due to changes in the on-site environment. In addition, unidirectional thermal excitation has an inherent detection blind spot for cracks parallel to the heat flow direction. Cracks exert weak disturbances on the thermal field, resulting in insignificant thermal responses and missed detections. Moreover, the identification results rely entirely on model performance, lacking a safety net of physical laws. Therefore, there is an urgent need for a visual detection method for internal cracks that integrates infrared and visible light dual modes, is based on closed physical equations, and is backed by iterative self-consistency and orthogonal resampling.

[0019] This application provides a visual inspection method for internal cracks in aluminum profiles used in engineering machinery, which integrates infrared images. Please refer to [link / reference]. Figure 1 The specific steps are as follows: S1. Dual-modal collaborative acquisition: Two sets of orthogonal linear array thermal excitation sources, an infrared thermal imager, and an industrial camera with a polarization filter are arranged at the inspection station, and a diffuse reflection reference sheet with known emissivity is fixed at the edge of the field of view. The two sets of orthogonal linear array thermal excitation sources are arranged orthogonally. The first set of thermal excitation sources is arranged longitudinally along the aluminum profile, and the second set is arranged transversely along the aluminum profile. The heat flow directions of the two sets are perpendicular to each other. This orthogonal dual-set configuration reserves the condition for switching heat flow directions to eliminate blind spots laterally. Each set of thermal excitation sources consists of multiple halogen lamps or laser diode arrays, with a preset power of 500W to 2000W. The emission wavelength covers the mid-infrared band from 1μm to 3μm, matching the absorption peak of the aluminum profile surface. The pulse width can be set from 1ms to 10ms or a continuous modulation mode can be used. The infrared thermal imager employs an uncooled focal plane detector with a resolution of at least 640×512 pixels, a spectral response range of 7.5μm to 14μm, and a frame rate set at at least 60Hz for continuous acquisition of thermal image sequences. The industrial camera uses a CMOS sensor with a resolution of at least 2048×2048 pixels. A linear or circular polarizing filter is placed in front of the lens, with adjustable polarization direction to eliminate specular reflections from the aluminum profile surface. The camera's frame rate is synchronized with the infrared thermal imager or set as an integer multiple thereof. A diffuse reflection reference plate, made of polytetrafluoroethylene or gold-plated, with its emissivity calibrated to a known value, is placed at the four corners of the field of view to provide known boundary constraints for calibrating the environmental reflection term in the radiative transfer equation. The two devices undergo joint calibration to achieve pixel-level registration with a registration accuracy of at least one pixel, ensuring strict spatial alignment between the infrared thermal image and the visible light image.

[0020] In terms of acquisition timing, a cold frame before excitation is acquired first, with an acquisition time of no less than 2 seconds. Multiple frames are averaged to reduce detector noise and record background radiation information under non-thermal excitation conditions. Then, thermal excitation is triggered, and a sequence of infrared thermal images is continuously acquired, with an acquisition time of 10 to 30 seconds, covering the entire process of thermal excitation application, thermal diffusion, and attenuation. Simultaneously, an industrial camera acquires visible light images after anti-reflection. The cold frame, in conjunction with the reference image, can close the environmental reflection term in the radiative transfer equation, achieving real-time calibration of the environmental reflection component. The polarization filter optical path eliminates the interference of specular highlights on the state partitioning of the aluminum profile surface through a polarization anti-reflection mechanism, keeping the surface state information in the visible light image pure. Thus, the dual-modal data with co-registration provides a pollution-free data foundation for subsequent emissivity field construction and crack thermal disturbance identification.

[0021] Through the above steps, the equation closure of the environmental reflection term and the geometric reservation of the direction blind zone are completed at the acquisition end, without the need for additional coatings or external measurement equipment, laying a data foundation for in-service online deployment. In existing technologies, infrared detection usually relies on a single thermal excitation and bare infrared image acquisition. Surface reflection and unknown environmental reflection lead to the lack of a reference for temperature field inversion, and unidirectional thermal excitation has an inherent detection blind zone for cracks parallel to the heat flow direction. This step fundamentally solves the problems of missing reference and blind zone reservation at the data acquisition end through the synergy of orthogonal dual-group excitation, cold frame self-calibration, and polarization anti-reflection.

[0022] S2. Semantic Partitioning and Field Construction: Input the visible light image after anti-reflection into the semantic segmentation model, divide the surface into five categories: machined bright area, anodized area, oil stain area, scratch and wear area, and dust-covered area, pixel by pixel. Call the mapping lookup table pre-calibrated according to the state category and ambient temperature, convert the label image into emissivity value pixel by pixel, and perform distance-weighted linear gradient interpolation along the normal direction of the boundary of adjacent blocks to generate a continuous emissivity field across the entire field of view.

[0023] The semantic segmentation model employs an encoder-decoder structure. The encoder extracts multi-scale features through multi-layer convolution and downsampling, while the decoder restores pixel-level spatial resolution through upsampling and skip connections, outputting pixel-level label maps for five surface states. These five surface states cover common surface types encountered by aluminum profiles in engineering machinery during service. Each surface state exhibits differentiated emissivity characteristics: the emissivity of machined bright areas is as low as 0.05 to 0.15, anodized areas 0.6 to 0.8, oily areas 0.85 to 0.95, and scratched / weared areas and dust-covered areas show spatially random emissivity distribution. Using a global average emissivity assumption would lead to systematic biases in temperature inversion; using simple block assignment would result in emissivity jumps at block boundaries. A mapping lookup table is pre-established through calibration, representing a three-dimensional mapping relationship between state category, ambient temperature, and emissivity. Ambient temperature is acquired in real-time by temperature sensors located at the detection station, and the corresponding emissivity value is obtained by interpolation based on the current ambient temperature. Linear gradient interpolation is performed on the boundaries of adjacent blocks along the normal direction with distance weighting. The gradient width is preset to 5 to 15 pixels, so that the emissivity transitions continuously along the normal direction and eliminates the coefficient jump of block assignment.

[0024] Through the above steps, visible light imaging is decoupled from the physical characteristics of internal crack thermal disturbances. The zoning benchmark is unaffected by defect contamination, and structural defects caused by the removal of contaminated data are detected after correction at the source. The continuous emissivity field ensures that the crack thermal disturbance morphology across surface state boundaries remains intact and uninterrupted, providing a continuous and boundary-continuous coefficient field for subsequent cold frame correction and temperature inversion. Existing technologies often use block emissivity assignment or global average emissivity assumptions, and emissivity jumps at surface state boundaries cause truncation or distortion of crack thermal signals. This step combines visible light semantic zoning with boundary gradient interpolation to ensure a physically continuous transition of the emissivity field, guaranteeing the morphological integrity of crack thermal disturbances at cross-zone boundaries.

[0025] S3. Initial inspection of cold frame correction: Using the cold frame from step S1 and the measured temperature of the reference film, the ambient reflectance radiance is solved pixel by pixel. The ambient reflection term in the radiation receiving equation is closed. The ambient reflection component is stripped from the continuous emissivity field in step S2 and the true surface temperature field is inverted. For pixels with extremely low emissivity, the absolute temperature is replaced by the time-domain difference signal. The normalized thermal contrast is calculated for the corrected thermal image sequence. The region growth discrimination is performed with a preset low threshold to generate a mask for initial inspection of suspected defects.

[0026] The cold frame provides a background radiation reference under unexcited conditions, while the reference image provides a radiance reference with known emissivity. Together, they establish a set of radiative transfer equations, solving for the environmental reflected radiance pixel-by-pixel, thus closing the environmental reflection term in the radiation receiving equation. Substituting the continuous emissivity field into the closed radiation equation, the environmental reflection component is stripped away frame-by-frame and pixel-by-pixel to invert the true surface temperature field. Since the absolute temperature inversion of extremely low emissivity surfaces (such as machined bright areas) is extremely sensitive to emissivity errors, even small emissivity deviations can lead to severe distortion of the temperature inversion value. Therefore, for pixels with emissivity below 0.2, a time-domain difference signal is used instead of absolute temperature. The time-domain difference signal characterizes thermal perturbations through the temperature change between adjacent frames, effectively avoiding the stringent requirements for emissivity accuracy in absolute temperature inversion. This allows pseudo-thermal anomalies to be accurately stripped away as optical attribute components by the physical equation, while crack thermal perturbations are completely preserved as true temperature components. For the corrected thermal image sequence, the difference between each frame and the cold frame is calculated and normalized to obtain a normalized thermal contrast map. Using a preset low threshold as the seed point determination condition, region growth is performed starting from the seed point using the 8-neighbor connectivity criterion. The growth terminates when the thermal contrast of adjacent pixels is lower than the threshold or the preset maximum area is reached, generating an initial suspected defect mask. This mask serves as the trigger input for the iteration loop in step S4, forming the connecting hub between the correction and recognition stages.

[0027] Through the above steps, the environmental reflectance term is acquired in real time according to the field conditions, without the need for coating samples and external measuring equipment, making it suitable for in-service online inspection scenarios. In existing technologies, the environmental reflectance term usually relies on empirical estimation or offline calibration. Changes in the field environment cause temperature inversion drift, and the reliability of temperature inversion in extremely low emissivity regions is poor. This step uses real-time collaborative calibration of cold frames and reference films, and simultaneously uses time-domain difference to replace absolute temperature to solve the ill-posedness problem of inversion in extremely low emissivity regions, enabling the physical correction process to have field adaptive capabilities.

[0028] S4. Iterative Self-Consistent Final Inspection: Remove pixels within the mask from the emissivity correction benchmark in the initial inspection of step S3. Based on the visible light state label of the pixels within the mask, reconstruct the correction parameters by interpolating the emissivity statistics of similar blocks outside the mask. Obtain a new mask through secondary correction. Iterate until convergence with the overlap rate of the areas of two adjacent masks not being lower than the convergence threshold as the criterion. During oscillation, take the union to mark the verification area. Perform weighted fusion based on the transition zone and the confidence attenuation of the reconstructed pixels. The final inspection segmentation outputs the crack contour. The apparent emissivity is back-inferred from the intact area confirmed by the final inspection to form a measured sample pair. Rolling update of the mapping lookup table realizes online self-calibration of the model.

[0029] The initial inspection mask covers areas containing suspected cracks and associated phenomena (such as discoloration and ablation marks). If the emissivity of this area is directly used as the correction benchmark, the crack-related discoloration will cause local interference with the regional benchmark, leading to the correction erasing of cracks or misidentification. Pixels within the mask are removed from the correction benchmark, and based on the visible light state label of each pixel within the mask, similar surface state blocks are found outside the mask. The statistical value (such as mean or median) of the emissivity of these blocks is calculated. Using this statistical value as a known point, spatial interpolation reconstruction is performed on the area within the mask. The interpolation method can use the inverse distance weighting method, where the distance weight decreases as the distance from the mask boundary increases. The reconstructed emissivity field is substituted back into the radiation equation for secondary correction to obtain the secondary corrected temperature field. The thermal contrast is recalculated on the secondary corrected temperature field, and region growth discrimination is performed to generate a new mask. The area overlap rate of two adjacent mask rounds, i.e., the ratio of the intersection area to the union area, is used as the spatial stability criterion. Iterative convergence is determined when the overlap rate is not lower than a preset convergence threshold. If the mask area oscillates during the iteration process, the union of multiple mask rounds is taken as the verification area marker. Mask shrinkage indicates that the suspected area has shrunk after secondary correction and the original area is a false anomaly. Mask expansion indicates that the suspected area has expanded after secondary correction and the original area is a real crack. Mask convergence indicates that the correction and recognition achieve spatial self-consistency. The transition zone is located in a certain width area inside and outside the mask boundary. The reconstructed pixels are the interpolated pixels. The confidence of both is lower than that of the directly measured pixels. Weighted fusion is performed by the confidence attenuation amount to keep the final inspection result conservative in the uncertain area and accurate in the certain area. The apparent emissivity of the intact area confirmed by the final inspection is used to form a measured sample pair. This sample pair is added to the training set of the mapping lookup table. The lookup table is updated continuously so that the emissivity model is updated synchronously with the service evolution of the workpiece surface condition, and the correction accuracy is self-maintained in the time domain.

[0030] Through the above steps, the correction and identification process is upgraded from a unidirectional serial connection to a bidirectional verification iterative closed loop. Scratch-associated cracks are no longer eliminated, structural silent regions are eliminated, and the mapping model is adaptively updated as the service process progresses. In existing technologies, correction and identification are usually in a unidirectional serial connection. Once the correction parameters are determined, they are fixed for identification. Local interference from crack-associated discoloration or ablation marks on the partition reference can lead to correction erasing cracks or misidentification of corrections. Furthermore, the mapping model is static and fixed, unable to adapt to the evolution of surface conditions during service. This step achieves bidirectional verification of correction and identification and online self-calibration of the mapping model through pixel elimination within the mask and statistical value interpolation reconstruction outside the mask.

[0031] S5. Orthogonal Re-examination Decision: Trigger the second set of orthogonal excitation sources preset in step S1, re-examine the heat map sequence and independently execute steps S2 to S4, take the union of the two identification results, for the area in the corrected heat map sequence located outside the final mask, fit the slope of the temperature rise time curve in logarithmic coordinates pixel by pixel, any connected region that deviates from the negative half power law and shows decay hysteresis is forcibly upgraded to a suspected defect, reinject step S4 iterative closed loop re-verification, and finally invert the crack burial depth according to the one-dimensional heat conduction relationship based on the peak time of thermal contrast, and perform grade evaluation in combination with location, orientation and depth.

[0032] In this system, the second group of orthogonal excitation sources is arranged orthogonally to the first group. When the crack direction is parallel to the heat flow direction of the first group, the heat flow is conducted along the crack surface, and the crack's disturbance to the thermal field is weak, resulting in an insignificant thermal response and a detection blind zone. Under orthogonal heat flow, the heat flow is conducted perpendicular to the crack surface, and the crack's disturbance to the thermal field is significantly enhanced, thus eliminating the direction blind zone. The detection completeness is improved after the union of the two identification results. For the region outside the final mask in the corrected thermal image sequence, the temperature values ​​at multiple moments after thermal excitation are extracted pixel by pixel, and the slope is obtained by linear fitting in the logarithmic coordinate system. The theoretical solution of the one-dimensional thermal diffusion equation shows that the temperature rise of an ideal semi-infinite body surface under pulsed thermal excitation has a negative half power law relationship with time. In the logarithmic coordinate system, this relationship is expressed as a straight line with a slope of negative half. As a thermal resistance, the temperature rise decay of the region above the crack deviates from this power law and exhibits decay hysteresis characteristics. However, the emissivity artifact is not constrained by the thermal diffusion equation, and its temperature rise decay does not exhibit this characteristic. Therefore, using the negative half power law as the physical criterion, the thermal response of cracks and emissivity artifacts can be physically distinguished. The self-consistent closed loop upgrades from verifying detected cracks to detecting missed ones, eliminating the initial detection deadlock and providing a physical lower limit guarantee for detection completeness independent of the recognition model's performance. For the cracked areas confirmed in the final inspection, the moment when the thermal contrast reaches its peak is extracted. According to the one-dimensional heat conduction theory, there is a correspondence between the crack burial depth and the peak thermal contrast moment. This relationship is used to infer the burial depth of the crack within the aluminum profile from the surface thermal signal. Combined with the crack's planar location, orientation, and depth, a risk level assessment is performed, generating an inspection report with a verification zone marker.

[0033] Through the above steps, the completeness of the detection is guaranteed by physical laws, and the detection results are upgraded from qualitative discovery to quantitative risk assessment. In existing technologies, unidirectional thermal excitation has a detection blind spot for cracks parallel to the heat flow direction, and the identification results depend entirely on model performance, lacking the backing of physical laws. This step, through the combination of orthogonal multiple sampling and the physical constraints of the thermal diffusion equation, provides a physical lower limit guarantee for the completeness of the detection, independent of the performance of the identification model, thereby achieving quantitative risk assessment of internal cracks.

[0034] The five-step progressive logic can be summarized as follows: Step S1 provides uncontaminated data and equation closure conditions; Step S2 constructs a continuous emissivity field based on visible light as the reference, solving the problem of separation; Step S3 performs physical correction and generates a mask to connect correction and identification; Step S4 iteratively verifies and self-calibrates to solve the problems of accurate and reliable calibration; Step S5 orthogonal multiple sampling and temporal decision-making as a fallback to solve the problem of comprehensive detection. The three-layer optimization shares the same visible light reference, the same emissivity field, and the same iterative closed loop, with each step interconnected and no absolutely independent steps. False alarms and false negatives decrease synchronously with the progression of the layers.

[0035] Based on Implementation Method 1, the industrial cameras in step S1 can be configured with binocular stereo vision. Two cameras are symmetrically arranged above the inspection station with a baseline distance of 100mm to 200mm. The three-dimensional morphology of the aluminum profile surface is reconstructed using a stereo matching algorithm, which assists in surface state classification in semantic segmentation and improves the accuracy of distinguishing between oil-stained areas and dust-covered areas. Through the above binocular stereo vision configuration, the three-dimensional morphology information of the surface is introduced. The liquid film thickness in the oil-stained area and the particle height in the dust-covered area have discernible depth differences in stereo vision, further improving the accuracy of surface state classification and reducing the error in subsequent emissivity field construction.

[0036] Building upon Implementation Method 1, the semantic segmentation model in step S2 can employ an edge enhancement branch. An edge detection auxiliary task is added to the encoder-decoder structure, constraining the precision of the segmentation boundaries through an edge loss function. This improves the distinguishability between scratches / wear areas and actual cracks at the segmentation boundaries, reducing the probability of scratches being misclassified as independent surface states. Through this edge enhancement branch, the semantic segmentation model becomes more sensitive to the geometric differences between linear scratches and regional cracks. The segmentation boundaries of the scratch / wear area more closely match the actual scratch contour, preventing actual cracks from being scratched into the scratch / wear area and thus avoiding missed detections.

[0037] Based on Implementation Method 1, in step S4, the inward interpolation reconstruction can use Kriging interpolation instead of linear interpolation. Optimal unbiased interpolation is performed using the spatial variation structure function of emissivity in similar blocks outside the mask. Spatial correlation is modeled through a semi-variogram function, ensuring that the reconstruction parameters maintain optimal estimation accuracy within the mask region. Through this Kriging interpolation, the emissivity reconstruction process considers the directionality and anisotropy of spatial variation, maintaining the spatial continuity of the reconstructed field better than linear interpolation, and improving the temperature inversion accuracy of the second correction.

[0038] Building upon Implementation Method 1, in step S5, the temperature rise curve slope fitting can employ the least squares method to perform block-by-block fitting on the region outside the mask, replacing pixel-by-pixel fitting. This divides the region outside the mask into several sub-blocks, uniformly fitting the temperature rise curve slope within each sub-block, reducing computational complexity and improving the real-time performance of online detection. Through this block-by-block fitting, computational resource consumption is significantly reduced, while the statistical averaging effect within the sub-blocks suppresses the interference of random noise on the slope estimation, enhancing the stability of the physical fallback criterion.

[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A visual inspection method for internal cracks in aluminum profiles used in engineering machinery, incorporating infrared images, characterized in that: The specific steps are as follows; S1. Dual-modal collaborative acquisition: Deploy orthogonal thermal excitation sources, infrared thermal imagers, industrial cameras, and reference films to acquire cold frames and trigger excitation of infrared and visible light data; S2, Semantic Partitioning Field Construction: Visible light segmentation of surface state categories, and continuous emissivity field with gradient interpolation of the normal to the boundary of adjacent blocks; S3, Cold Frame Correction Initial Inspection: Solve the reflection term for cold frames and reference films, and grow an initial inspection mask with thermal contrast in the low-threshold region; S4. Iterative self-consistent final inspection: Pixels within the initial inspection mask are removed from the baseline. The parameters are corrected by interpolating the emissivity of similar blocks outside the mask. A new mask is obtained after the second correction. The process is iterated until convergence, and the union set is used for verification. S5, Orthogonal Re-mining Decision: The slope of the temperature rise curve fitted outside the mask, the deviation from the negative half power law marks the suspected defect, and the thermal contrast peak value is used to invert the crack burial depth in one-dimensional thermal conduction.

2. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 1, characterized in that, In step S1, two sets of orthogonal linear array thermal excitation sources, an infrared thermal imager, and an industrial camera with a polarization filter optical path are arranged at the detection station. A diffuse reflection reference film with known emissivity is fixed at the edge of the field of view. The two devices are jointly calibrated to complete pixel-level registration. The orthogonal dual-set configuration reserves heat flow switching conditions for subsequent blind zone elimination. The reference film provides boundary constraints for environmental reflection calibration.

3. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 2, characterized in that, In step S1, the cold frame before excitation is first acquired, and then the infrared thermal image sequence is continuously acquired after thermal excitation. The visible light image after anti-reflection is acquired simultaneously. The environmental reflection term of the closed radiative transfer equation is obtained between the cold frame and the reference film. The polarization anti-reflection path eliminates specular highlight interference. The homologous registration data lays the foundation for subsequent emissivity field construction and crack identification.

4. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 3, characterized in that, In step S2, the visible light image after eliminating reflection is input into the semantic segmentation model. The model divides the surface into five types of states: machined bright area, anodized area, oil stain area, scratch and wear area, and dust-covered area, pixel by pixel. The model calls the pre-calibrated mapping lookup table to assign values ​​pixel by pixel. Visible light and crack thermal disturbance are not coupled. The partitioning benchmark is not affected by defect contamination. The structural defects caused by the elimination of defect data correction are detected again.

5. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 4, characterized in that, In step S2, the boundary of adjacent blocks is linearly interpolated along the normal direction with distance weighting to generate a continuous emissivity field across the entire field of view. This eliminates coefficient jumps in block-shaped assignments, and the crack morphology across the surface state boundary remains intact and uninterrupted, providing a continuous coefficient field without jumps for subsequent cold frame correction and temperature inversion.

6. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 5, characterized in that, In step S3, the environmental reflection term is solved pixel by pixel using the cold frame and the measured temperature of the reference film in step S1. The radiation receiving equation is closed and substituted into the continuous emissivity field in step S2 to strip away the environmental reflection and invert the real temperature field. For pixels with extremely low emissivity, the absolute temperature is replaced by time-domain difference. The pseudo-thermal anomaly is stripped away by the physical equation, and the crack thermal disturbance is completely preserved.

7. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 6, characterized in that, In step S3, the normalized thermal contrast is calculated for the corrected thermal image sequence, and a preliminary suspected defect mask is generated by growth discrimination in low-threshold areas. This mask serves as the iterative closed-loop trigger input for step S4, connecting the correction and identification links. The environmental reflection term is acquired in real time on site.

8. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 7, characterized in that, In step S4, pixels within the initial detection mask are removed from the calibration reference. The calibration parameters are reconstructed by interpolating the emissivity of similar blocks outside the mask according to the visible light label. A new mask is obtained through secondary discrimination. The process is iterated until convergence using the area overlap rate criterion. During oscillation, the union of the marked core area is taken. The transition zone is weighted and fused based on the confidence attenuation amount. The apparent emissivity of the intact area is used to update the mapping lookup table to achieve self-calibration.

9. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 8, characterized in that, In step S5, the second set of orthogonal excitation sources is pre-set in S1. The thermal map is re-sampled and executed independently from S2 to S4. The results of the two are combined. Cracks that are parallel to the heat flow but have weak thermal response are revealed under orthogonal heat flow and move towards the blind zone to be eliminated. The closed loop is upgraded from verifying that it has been found to finding that it has been missed. The completeness of the detection is guaranteed by the physical lower limit.

10. The method for visually detecting internal cracks in aluminum profiles for engineering machinery by fusing infrared images according to claim 9, characterized in that, In step S5, the slope of the temperature rise curve is fitted pixel by pixel in logarithmic coordinates for the area outside the final mask. Connected regions that deviate from the negative half power law and show decay hysteresis are marked as suspected defects. The process is then returned to S4 for re-verification. The crack burial depth is inverted at the peak of the thermal contrast according to one-dimensional thermal conduction, and evaluated in combination with the location, orientation and depth.

Images