A method and system for quantitative detection of material crack depth based on thermal gradient vector field

Abstract

This invention provides a method and system for quantitatively detecting material crack depth based on thermal gradient vector field, belonging to the field of nondestructive testing technology. The invention proposes a quantitative detection scheme of unilateral heat flow excitation + orthogonal gradient scanning: First, a unilateral high-temperature boundary is constructed using a portable chemical self-heating pack and insulating material, forcing the heat flow to transversely cross the crack, forming a significant temperature step; second, infrared thermographic sequences of the cooling process are acquired to construct a two-dimensional thermal potential energy field and calculate the global gradient vector field; morphological algorithms are used to extract the crack topological skeleton, and virtual rays are emitted along the local orthogonal detection direction determined by the local thermal gradient direction at the skeleton nodes. Adaptive orthogonal integration is performed to extract multidimensional spatial features under multiple time series, which are then input into a supervised learning regression model to invert the physical depth of the crack. This invention effectively eliminates the geometric projection error caused by crack curvature, and features high computational accuracy and strong anti-interference ability, making it suitable for large-scale rapid detection in engineering sites.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to a method and system for quantitatively detecting the depth of material cracks based on a thermal gradient vector field. Background Technology

[0002] Material cracking is the most common form of damage in concrete and masonry structures such as bridges, tunnels, and dams. Crack depth is a key parameter for assessing the degree of internal damage, the risk of steel corrosion, and the reduction in load-bearing capacity. Its accurate measurement is crucial for structural safety assessment and remaining life prediction.

[0003] Currently, methods for detecting crack depth mainly suffer from the following limitations:

[0004] 1. Traditional contact testing methods are inefficient and can damage structures. Although ultrasonic surface testing is widely used, it is cumbersome to operate, requires the use of a coupling agent to ensure acoustic contact, and has insufficient resolution for dense, tiny cracks. Core drilling, while able to directly measure depth, is a destructive test that can damage the structure itself and cannot be implemented over large areas, making it suitable only for localized sampling.

[0005] 2. Visible light-based visual inspection methods can only acquire surface information. Current mainstream crack image recognition technologies, including edge detection and deep learning semantic segmentation, can only accurately measure the surface width and length of cracks, but cannot penetrate the material surface to obtain information in any depth direction. Therefore, they cannot assess the severity of cracks extending into the structure.

[0006] 3. Existing infrared thermal imaging nondestructive testing technologies suffer from low quantitative accuracy, large geometric errors, and high equipment requirements. While infrared thermal imaging, due to its non-contact and large-area scanning advantages, is increasingly being applied to crack detection, its technical solutions generally suffer from the following significant shortcomings:

[0007] Lack of quantitative mapping model: Currently, most of the time, the detection personnel rely on visual observation of color difference or temperature difference in thermal images to determine whether cracks "exist" or "not exist," which only stays at the qualitative detection stage. There is a lack of a precise mathematical mapping model from "temperature grayscale" to "physical depth," and it is impossible to give a millimeter-level quantitative value of depth.

[0008] Large geometric projection error: Existing algorithms typically use fixed horizontal, vertical, or single-directional scan lines to extract temperature profiles. When the crack direction is curved or inclined, these scan lines cannot always be orthogonally tangent to the crack axis, causing the measured width of the heat-affected zone to be elongated due to the projection effect, introducing systematic geometric measurement errors and severely affecting the accuracy of depth inversion.

[0009] High barriers to entry for excitation equipment: Existing active thermal excitation methods mostly use pulsed flash lamps or high-power lasers, which are bulky, consume a lot of power, and are expensive. They are difficult to deploy in field engineering sites and complex operating environments, making it difficult to meet the needs of large-scale and high-efficiency surveys.

[0010] In summary, there is an urgent need in this field for a quantitative detection method and system for material crack depth based on thermal gradient vector field, in order to solve the technical problems of existing infrared thermal imaging crack detection technology lacking an accurate mapping model from temperature grayscale to physical depth, and its fixed-direction scanning method producing serious geometric projection errors when facing curved cracks, resulting in the inability to achieve high-precision quantitative depth inversion. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a low-cost, easy-to-operate method for quantitatively detecting material crack depth with high geometric robustness. This method achieves a rigorous nonlinear mapping of various physical depths by constructing a unilateral adiabatic boundary condition, utilizing the heat flow blocking effect of cracks, and combining it with global gradient vector field analysis.

[0012] To achieve the above objectives, this application proposes a method for quantitatively detecting the depth of material cracks based on a thermal gradient vector field, the method comprising the following steps:

[0013] Step S1, construct a one-sided heat flow excitation environment: set an active heat source on one side of the crack on the surface of the material to be tested, and use insulating material to form an insulating block at the junction of the active heat source and the crack, forcing the heat flow to pass through the crack laterally, so as to form a temperature step at the crack interface.

[0014] Step S2, Acquire transient thermal image sequence: After the thermal excitation ends, use an infrared thermal imager to acquire two-dimensional infrared radiation images containing the cooling process of the crack and the heat-affected zone on both sides, and establish a thermal image dataset containing the time dimension.

[0015] Step S3, Constructing the thermal potential energy field and preprocessing: Mapping the two-dimensional infrared radiation image in the thermal image dataset into a discrete two-dimensional thermal potential energy field, and performing smoothing and noise reduction preprocessing on the two-dimensional thermal potential energy field.

[0016] Step S4, calculate the global gradient vector field: calculate the global gradient vector of each pixel in the two-dimensional thermal potential energy field, and construct a global gradient vector field that reflects the instantaneous direction of heat flow transmission; wherein, the global gradient vector includes gradient modulus and gradient direction angle;

[0017] Step S5, extract the crack topology skeleton: Based on the global gradient vector field, use the image segmentation algorithm to extract the binary image of the crack region, and use the morphological thinning algorithm to extract the crack topology skeleton with a single pixel width as a spatial measurement reference.

[0018] Step S6, Adaptive Orthogonal Ray Integration: Traverse all nodes on the crack topology skeleton, determine the detection direction orthogonal to the local crack orientation based on the local thermal gradient direction at each node, emit a virtual detection ray along the detection direction, and dynamically determine the integration interval on the virtual detection ray according to the attenuation characteristics of the gradient modulus, and calculate the cumulative gradient energy value along the integration interval.

[0019] Step S7, Multidimensional Feature Fusion and Deep Inversion: For each node, at least the multidimensional spatial features, including the gradient energy accumulation value described in step S6, are extracted from the two-dimensional infrared radiation image of the cooling process at multiple acquisition times. These features are then concatenated to construct a high-dimensional joint feature vector, which is then input into the trained supervised learning regression model to invert the physical depth of the crack corresponding to the node.

[0020] As a further solution, in step S1, the active heat source is a portable chemical self-heating pack; the insulation material is thermal insulation cotton, which is closely attached to the material surface and laid on the side of the active heat source near the crack.

[0021] As a further solution, in step S3, a Gaussian kernel function is introduced. Spatial domain convolution operations are performed on two-dimensional infrared radiation images to construct a smooth thermal potential energy field. To suppress random thermal speckle noise generated by uncooled focal plane arrays.

[0022] As a further solution, the smoothed thermal potential energy field Apply the Sobel differential operator to calculate each pixel. The global gradient vector is used to characterize the gradient magnitude and gradient direction angle.

[0023]

[0024]

[0025] in, Indicates gradient modulus; This represents the global gradient vector of the thermal potential energy field; express Norms are used to calculate the Euclidean length of a vector. Indicates the gradient direction angle; This represents the arctangent function, used to transform the spatial derivative into the angle between the heat flux vector and the coordinate axis; and These represent the partial derivatives of the thermal potential energy field in the horizontal and vertical directions, respectively.

[0026] As a further solution, in step S5, the crack topology skeleton with a single pixel width is extracted through the following steps:

[0027] A gradient modulus histogram is constructed based on the global gradient vector field, and the low-value main peak and high-value long tail are determined based on the gradient modulus histogram.

[0028] The optimal inflection point between the low-value main peak and the high-value long tail is automatically found using the maximum inter-class variance algorithm and used as an adaptive threshold. Construct gradient modulus greater than binary feature set ;

[0029] Using the maximum connectivity operator Filtering out binary feature sets Discrete noise, and utilize the refinement operator Extract the topological skeleton of the crack with a width of one pixel.

[0030] As a further solution, in step S6, the cumulative gradient energy value along the integration interval is calculated through the following steps:

[0031] For any node on the crack topology skeleton The unit detection vector orthogonal to the local crack orientation is determined based on the local thermal gradient direction at the node. :

[0032]

[0033] Where i is the node number; For nodes gradient modulus, For nodes The gradient direction angle.

[0034] Define along the unit probe vector One-dimensional orthogonal ray path in direction ;in For path parameters;

[0035] One-dimensional orthogonal ray path Perform gradient energy integration to obtain the total thermal resistance. And used as the cumulative value of gradient energy:

[0036]

[0037] in, and These are the dynamically determined forward and backward truncation boundaries, respectively; k is the summation index variable; and K is the integration boundary index. For the space to be far from the walking distance, and These are discrete boundary indices determined based on forward truncation boundaries and backward truncation boundaries, respectively.

[0038] As a further solution, in step S6, the attenuation coefficient is set. And define the gradient modulus cutoff threshold. ;

[0039] The integration interval is then dynamically determined as follows:

[0040] ;

[0041] .

[0042] In the formula, Indicates a forward truncation boundary. Indicates the backward cutoff boundary. Indicates the path along a one-dimensional orthogonal ray. The gradient modulus of the specific sampling point. Represents a node gradient modulus;

[0043] The dynamically determined integration interval refers to: using the crack topology skeleton nodes Centered on, along a one-dimensional orthogonal ray path Search in both the forward and reverse directions, and when the gradient modulus of the sampling points on the path... First time less than the cutoff threshold At that time, the forward cutoff boundaries are determined respectively. and backward cutoff boundary and will This serves as the valid integration interval for that node.

[0044] As a further solution, in step S7, the multidimensional spatial features are dynamically extracted for each node at each acquisition time, using the following feature subset:

[0045] The orthogonal gradient integral characteristic is defined by the cumulative gradient energy value;

[0046] The local gradient peak characteristic is defined by the maximum gradient modulus within the integration interval;

[0047] The effective thermal resistance width characteristic is defined by the sum of the forward cutoff boundary and the backward cutoff boundary of the integration interval;

[0048] The feature subsets of the node at multiple acquisition times are concatenated and spliced ​​together to form the high-dimensional joint feature vector.

[0049] As a further solution, the supervised learning regression model trained in step S7 is a random forest regression model, the training process of which includes:

[0050] Construct a global training set consisting of multiple crack samples with known physical depths, where each sample contains a high-dimensional joint feature vector and its corresponding true depth label;

[0051] Bootstrap sampling is used to construct an independent training subset for each decision tree, and node splitting and growth are performed based on the mean square error minimization criterion.

[0052] Ultimately, the crack physical depth output by the random forest regression model is the arithmetic mean of the base prediction depths output by all decision trees.

[0053] On the other hand, the present invention also provides a quantitative detection system for material crack depth based on a thermal gradient vector field, comprising:

[0054] A single-sided active thermal excitation module, including a portable chemical self-heating pack and thermal insulation cotton, is used to construct a single-sided heat flow excitation environment;

[0055] The infrared imaging acquisition module includes a portable infrared thermal imager fixed on a tripod, used to acquire infrared thermal image sequences during the cooling process;

[0056] The image processing and inversion calculation module is connected to the infrared imaging acquisition module and is used to execute the algorithm flow of the quantitative detection method for material crack depth based on thermal gradient vector field as described in any of the above, and output the crack physical depth detection result.

[0057] Compared with related technologies, the quantitative detection method and system for material crack depth based on thermal gradient vector field provided by the present invention has the following advantages:

[0058] 1. This invention abandons the crude approach of relying on visual observation of color differences for qualitative judgment in traditional infrared detection. Based on the heat conduction equation and boundary conditions, it constructs a nonlinear mapping model from gradient energy integral to physical depth. By extracting multidimensional features such as orthogonal gradient integral values, local gradient peak sharpness, and effective thermal resistance width, and using a supervised learning regression model for high-dimensional feature fusion and inversion, it can provide millimeter-level quantitative values ​​of crack depth, significantly improving detection accuracy and reliability.

[0059] 2. To address the problem in existing technologies where fixed-direction scan lines cannot orthogonally cut through curved cracks, leading to an elongated heat-affected zone, this invention proposes an adaptive orthogonal ray integration algorithm. This algorithm uses the crack topological skeleton as a spatial reference and emits virtual probe rays at each node along an orthogonal probe direction determined by the local thermal gradient, ensuring that the integration path is always perpendicular to the local orientation of the crack. Regardless of how the crack is curved or tilted, the measurement result always represents the true normal cross-sectional thermal resistance, effectively reducing geometric projection errors.

[0060] 3. In the orthogonal ray integration process, this invention does not use a fixed integration window. Instead, it dynamically determines the forward and backward integration cutoff boundaries based on the physical characteristic that the gradient modulus decays from its peak to a specific ratio. This mechanism can accurately locate the effective influence area of ​​thermal blockage caused by cracks, automatically eliminate interference from far-end background noise and irrelevant heat sources, avoid the cumulative error introduced by an excessively wide integration window, and ensure the purity and accuracy of feature extraction.

[0061] 4. This invention utilizes a portable chemical self-heating pack and insulation cotton to construct a unilateral active thermal excitation environment. Combined with a portable infrared thermal imager connected to a smart terminal, all data acquisition can be completed, replacing the expensive and bulky pulse flash lamps or high-power laser excitation systems of traditional solutions. The entire device is extremely low-cost, portable, and quick to deploy, making it particularly suitable for large-scale, rapid surveys of field engineering sites such as bridges, tunnels, and dams, significantly reducing the detection threshold and operating costs.

[0062] 5. This invention only requires the placement of a heat source and insulation layer on one side of the crack to form an ideal unidirectional heat conduction physical boundary, eliminating the need for symmetrical operation on both sides of the crack, thus simplifying the on-site deployment process. Combined with automated image processing and depth inversion algorithms, end-to-end processing is achieved from data acquisition to result output. A single detection can complete the depth distribution measurement of multiple points along the entire crack, significantly improving detection efficiency. Attached Figure Description

[0063] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0064] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0065] Figure 1 A schematic diagram illustrating the steps of a method for quantitatively detecting material crack depth based on a thermal gradient vector field, provided by this invention;

[0066] Figure 2 This is a schematic diagram of the layout and heat conduction of the single-sided heat flow excitation device of the present invention;

[0067] Figure 3 A schematic diagram illustrating the principle of global gradient vector field and adaptive orthogonal ray integration provided by this invention;

[0068] Figure 4 A schematic diagram of the orthogonal gradient integral characteristic distribution at time t=0 provided by the present invention;

[0069] Figure 5 This is a schematic diagram of a material crack depth quantitative detection system based on thermal gradient vector field provided by the present invention.

[0070] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0072] Example 1

[0073] like Figure 1 As shown, this embodiment provides a method for quantitatively detecting the depth of material cracks based on a thermal gradient vector field. The method includes the following steps:

[0074] Step S1, construct a one-sided heat flow excitation environment: set an active heat source on one side of the crack on the surface of the material to be tested, and use insulating material to form an insulating block at the junction of the active heat source and the crack, forcing the heat flow to pass through the crack laterally, so as to form a temperature step at the crack interface.

[0075] Step S2, Acquire transient thermal image sequence: After the thermal excitation ends, use an infrared thermal imager to acquire two-dimensional infrared radiation images containing the cooling process of the crack and the heat-affected zone on both sides, and establish a thermal image dataset containing the time dimension.

[0076] Step S3, Constructing the thermal potential energy field and preprocessing: Mapping the two-dimensional infrared radiation image in the thermal image dataset into a discrete two-dimensional thermal potential energy field, and performing smoothing and noise reduction preprocessing on the two-dimensional thermal potential energy field.

[0077] Step S4, calculate the global gradient vector field: calculate the global gradient vector of each pixel in the two-dimensional thermal potential energy field, and construct a global gradient vector field that reflects the instantaneous direction of heat flow transmission; wherein, the global gradient vector includes gradient modulus and gradient direction angle;

[0078] Step S5, extract the crack topology skeleton: Based on the global gradient vector field, use the image segmentation algorithm to extract the binary image of the crack region, and use the morphological thinning algorithm to extract the crack topology skeleton with a single pixel width as a spatial measurement reference.

[0079] Step S6, Adaptive Orthogonal Ray Integration: Traverse all nodes on the crack topology skeleton, determine the detection direction orthogonal to the local crack orientation based on the local thermal gradient direction at each node, emit a virtual detection ray along the detection direction, and dynamically determine the integration interval on the virtual detection ray according to the attenuation characteristics of the gradient modulus, and calculate the cumulative gradient energy value along the integration interval.

[0080] Step S7, Multidimensional Feature Fusion and Deep Inversion: For each node, at least the multidimensional spatial features, including the gradient energy accumulation value described in step S6, are extracted from the two-dimensional infrared radiation image of the cooling process at multiple acquisition times. These features are then concatenated to construct a high-dimensional joint feature vector, which is then input into the trained supervised learning regression model to invert the physical depth of the crack corresponding to the node.

[0081] In practical engineering inspections, the crack depth of concrete structures such as bridges, tunnels, and dams is a key parameter for assessing structural safety. However, existing inspection methods have the following shortcomings: ultrasonic flat-surface testing is cumbersome, relies on coupling agents, and has poor adaptability to curved cracks; core drilling sampling damages the structure and cannot be implemented over large areas; visible light-based visual inspection can only obtain the apparent width and length of cracks, and cannot penetrate the surface to obtain depth information; while traditional infrared thermography either remains at the qualitative stage of visually observing color differences, or uses fixed horizontal or vertical scanning lines to extract temperature profiles. Once the crack direction bends or tilts, the scanning lines cannot orthogonally tangent to the crack axis, and the measured width of the heat-affected zone is systematically elongated due to the projection effect, introducing serious geometric measurement errors.

[0082] To address the aforementioned technical issues, this embodiment proposes a quantitative detection method for material crack depth based on a thermal gradient vector field: a unilateral heat flow excitation environment is constructed using a portable chemical self-heating pack and insulation cotton, forcing the heat flow to traverse the crack laterally, forming a significant temperature step; subsequently, infrared thermographic sequences during the cooling process are acquired, a two-dimensional thermal potential energy field is constructed, and the global gradient vector field is calculated; then, the crack topological framework is extracted as a spatial reference. Under the conditions where the crack can be approximated as a smooth curve, the local thermal properties of the material are approximately uniform, and the temperature disturbance at this location is mainly dominated by the crack thermal resistance effect, the temperature changes on both sides of the crack are mainly distributed along the crack normal, while the changes along the crack tangential are relatively small. Therefore, the direction of the local thermal gradient vector on the crack surface is approximately along the local crack normal and approximately orthogonal to the local crack orientation. Based on this relationship, a detection direction orthogonal to the local crack orientation can be determined according to the local thermal gradient direction at the skeleton node. A virtual detection ray is then emitted along this detection direction, and adaptive orthogonal integration is performed to ensure that the integration path is always perpendicular to the local crack orientation. This eliminates the geometric projection error caused by crack bending in traditional methods in principle. Finally, features such as gradient energy accumulation, local gradient peak, and effective thermal resistance width are extracted from multiple time series, concatenated into a high-dimensional joint feature vector, and input into a supervised learning regression model to invert the physical depth of the crack.

[0083] The following detailed description, using specific experimental data and operational procedures, illustrates the present invention's method and system for quantitative detection of material crack depth based on a thermal gradient vector field. It should be clarified that the parameters involved in this embodiment, such as heat source temperature, heating time, imaging interval, and specific algorithm thresholds, are optimized configurations for this specific scenario, intended to better illustrate the invention, and not to limit the scope of protection of the invention. In practical engineering applications, operators can adaptively adjust the above parameters according to the material type, ambient temperature, and required detection accuracy.

[0084] Please refer to Figure 2 To implement the detection method of this embodiment, a complete detection system was first built on-site. In the specific implementation process, to construct a physical environment conforming to the unilateral heat conduction model, the following low-cost experimental system was built in this embodiment:

[0085] Material cracks with different orientations were selected as test objects, including approximately straight cracks, inclined cracks, curved cracks and partially bifurcated cracks. The crack lengths were set into four tiers: 5cm, 10cm, 15cm and 20cm, and covered different depths.

[0086] In actual setup, to ensure the stability and geometric consistency of the imaging field of view, the infrared thermal imager was fixed on an adjustable tripod. The lens optical axis was adjusted to be perpendicular to the material plane where the crack is located (90° downward angle), and the vertical height of the lens in front of the image from the material surface was maintained at 40cm. This height setting aims to balance imaging resolution and field of view, ensuring that the crack and the complete heat-affected zones on both sides are located in the center of the image.

[0087] After the system was set up, the operators began to construct the core physical excitation environment. The purpose of this step was to create a controllable, unidirectional heat conduction boundary, forcing the heat flow to pass laterally through the crack under test, thereby forming a temperature step at the crack interface that could be quantified and analyzed.

[0088] The operator takes out the industrial-grade chemical self-heating pack, which, after activation, reaches an average temperature of 90°C and can continuously generate heat for 30 minutes as an active heat source. In this embodiment, 15 minutes of the stable heating phase is selected as the thermal excitation time, along with 5mm thick and 5cm wide insulation cotton. During on-site operation, the insulation cotton is tightly attached to the material surface and laid on the side of the self-heating pack closest to the crack, forming a thermal insulation barrier. This physically blocks the direct heat radiation from the heat source to the other side of the crack (the low-temperature side), ensuring that heat is conducted laterally only through the internal medium of the material and the crack cross-section.

[0089] After the heating phase, operators removed the chemical self-heating pack and all insulation materials to prevent them from interfering with the subsequent cooling process's thermal radiation. The infrared thermal imager was then activated in continuous acquisition mode to record the cooling process in the crack area.

[0090] Specifically, this embodiment sets three key data acquisition moments:

[0091] The data acquisition in this embodiment strictly follows the following timing protocol to capture the thermal inertial characteristics of the crack:

[0092] Thermal excitation stage: Place the self-heating pack tightly against one side of the crack and insulate it with thermal insulation cotton, and continue heating for 15 minutes; during this process, heat accumulates inside the material and forms a significant temperature jump at the crack interface.

[0093] Remove the heat source: Then remove the self-heating pack and insulation cotton, and immediately begin the cooling process.

[0094] Transient image acquisition: Two-dimensional infrared radiation images were acquired at the thermal shock moment, the first decay moment, and the second decay moment to establish a thermal image dataset containing the time dimension.

[0095] t=0min (thermal shock moment): Remove the heating pack and insulation cotton, and immediately take the first picture; at this time, the temperature difference is the largest, the gradient signal is the strongest, and it can reflect the instantaneous heat resistance of the crack.

[0096] t=5min (first decay time): After standing and cooling for 5 minutes, the second image is taken; at this time, the surface heat begins to dissipate, and the thermal resistance effect of the deep cracks begins to dominate the temperature distribution.

[0097] t=10min (second decay time): After 10 minutes of static cooling, a third shot is taken; this is used to record the long-term decay characteristics of the thermal field.

[0098] Furthermore, the two-dimensional infrared radiation images in the thermal imaging dataset are mapped to discrete two-dimensional thermal potential energy fields. Let the discrete data matrix corresponding to the two-dimensional infrared radiation image at any time in the thermal imaging dataset be... .

[0099] To suppress random speckle noise generated by the uncooled focal plane array in an infrared thermal imager, a Gaussian filter is introduced to perform spatial domain convolution on the original thermal image to construct a smooth thermal potential energy field. :

[0100]

[0101] In the above formula, the meanings of each parameter are as follows:

[0102] The coordinates in the original infrared thermal image are: The pixel grayscale value at that location physically corresponds to the original radiative energy distribution on the material surface;

[0103] This represents the smoothed thermal potential energy field after noise reduction, which serves as the reference field for subsequent gradient vector calculations.

[0104] This represents a two-dimensional Gaussian kernel function, used as a low-pass filter to remove high-frequency thermal noise.

[0105] Represents the local relative coordinates inside the convolution kernel, defining the spatial offset during pixel smoothing;

[0106] The standard deviation of the Gaussian distribution (in this example, we take...) The physical properties of this property determine the strength of the smoothness. The larger the value, the stronger the suppression of speckle noise, but it is necessary to preserve the characteristic details of the crack edges.

[0107] The range of the window neighborhood representing the convolution integral determines the size of the surrounding pixels participating in the calculation of a single point;

[0108] This represents the natural constant, used to construct a normal distribution weight allocation model.

[0109] Subsequently, the smooth thermal potential energy field Apply the Sobel differential operator to calculate each pixel. global gradient vector To construct a global gradient vector field that reflects the instantaneous direction of heat flow transmission.

[0110] global gradient vector They are respectively characterized as gradient modulus With gradient direction angle :

[0111]

[0112]

[0113] The gradient modulus represents the local thermal resistance intensity. For crack regions, the thermal resistance effect of air gaps will form a significant gradient modulus peak.

[0114] This represents the gradient direction angle, used to determine the detection direction orthogonal to the local orientation of the crack;

[0115] It represents the global gradient vector of the thermal potential energy field, used to reflect the direction of the most drastic temperature change;

[0116] and Let represent the partial derivatives of the thermal potential energy field in the horizontal (x-axis) and vertical (y-axis) directions, respectively. Obtained by convolution with the Sobel operator;

[0117] express Norms are used to calculate the Euclidean length of a vector.

[0118] The arctangent function is used to transform the spatial derivative into the angle between the heat flux vector and the coordinate axis, which is the key angle basis for the subsequent realization of "adaptive orthogonal integration".

[0119] To establish a spatial benchmark for integral measurements, a crack topological skeleton with a single pixel width is extracted through morphological operations based on a global gradient vector field. It includes the following steps:

[0120] 1. The global gradient vector obtained from the previous calculation is statistically analyzed, and a gradient modulus histogram is constructed to analyze the distribution characteristics of local thermal resistance. This histogram exhibits a typical long-tailed distribution under thermal excitation.

[0121] The low-value main peak corresponds to a large area of ​​stable heat transfer background.

[0122] High-value long tails correspond to real cracks that produce temperature steps.

[0123] 2. Based on this physical distribution difference, the Otsu's algorithm is used to automatically find the optimal inflection point between the low-value main peak and the high-value long tail, which serves as an adaptive threshold. This filters out environmental thermal noise and constructs a pure binary feature set. :

[0124]

[0125] 3. Introduce the maximum connected component operator Filtering out binary feature sets Discrete noise points in the data are then processed using morphological thinning operators. Extract the topological skeleton of the crack with a width of one pixel:

[0126]

[0127] At this point, each element in the crack topological skeleton set The emission center of the integral scan is precisely the result of the accurate removal of background noise by the pre-formed histogram. This skeleton effectively avoids the interference of pseudo-branches and ensures the accuracy of the subsequent orthogonal ray emission nodes and normal angles.

[0128] like Figure 3 As shown, Figure 3 The schematic diagram provided by this invention illustrates the principle of integration of the global gradient vector field and the adaptive orthogonal ray at time t=0min, showing the integration path distributed along the skeleton normal.

[0129] Traverse all nodes on the crack topology skeleton S, for any node Perform the following operations:

[0130] 1. Determine the unit detection vector orthogonal to the local fracture orientation based on the local thermal gradient direction at the node. :

[0131]

[0132] Where i is the node number; For nodes gradient modulus, For nodes The gradient direction angle.

[0133] It should be noted that the direction of the local thermal gradient refers to the node. The direction of the most significant temperature change can be determined by the gradient direction angle. and its corresponding unit probe vector The local direction of a crack refers to the extension direction of the crack skeleton along the line at that node, which can be determined by adjacent pixels on the skeleton. , and The orientation of the cracks directly represents the direction of temperature change. In other words, the direction of the local thermal gradient represents the direction of the most significant temperature change, and the local orientation of the crack represents the direction in which the crack extends along its line. Therefore, the two are approximately orthogonal.

[0134] 2. Define along the unit probe vector One-dimensional orthogonal ray path in direction :

[0135] ; ;

[0136] 3. Dynamically determine the integration interval: To eliminate background noise and interference from unrelated heat sources, and The boundary is not a fixed value, but is dynamically determined based on the decay characteristics of the gradient modulus to determine the effective range of the crack thermal resistance:

[0137] set up For the attenuation coefficient, in this embodiment, we take... Define the gradient modulus cutoff threshold:

[0138] ;

[0139] Traversing along a one-dimensional orthogonal ray path to both sides, when the gradient modulus first falls below the cutoff threshold, it is determined that the region has escaped the effective thermal resistance influence zone, and the forward cutoff boundary of the integration interval is dynamically determined accordingly. and backward cutoff boundary :

[0140] ;

[0141] ;

[0142] Specifically: using crack topology skeleton nodes Centered on, along a one-dimensional orthogonal ray path Search in both the forward and reverse directions, and when the gradient modulus of the sampling points on the path... First time less than the cutoff threshold At that time, the forward cutoff boundaries are determined respectively. and backward cutoff boundary and will This serves as the valid integration interval for that node.

[0143] This mechanism ensures that the integration domain only covers the effective area affected by thermal blockage caused by the crack, avoiding the cumulative error introduced by an excessively wide integration window.

[0144] 4. One-dimensional orthogonal ray path Perform gradient energy integration to obtain the total thermal resistance. And used as the cumulative value of gradient energy:

[0145]

[0146] In the formula, and These are the dynamically determined forward and backward truncation boundaries, respectively; k is the summation index variable; and K is the integration boundary index. The space is far from the walking distance; Represents the spatial physical integral distance along the direction of orthogonal rays; This is the index of the forward discrete boundary determined based on the forward truncation boundary and the backward truncation boundary. This is a backward discrete boundary index determined based on the forward truncation boundary and the backward truncation boundary;

[0147] Represents skeleton nodes The unit probe vector at that location,

[0148] Geometrically, it is strictly perpendicular to the local orientation of the crack.

[0149] In a physical sense, it indicates the shortest effective path through which instantaneous heat flow is blocked.

[0150] Furthermore, in order to accurately invert the physical depth of the crack, this embodiment establishes a nonlinear mapping relationship between high-dimensional spatial features and physical depth.

[0151] Dynamic extraction and standardization of 3D spatial feature subsets: Considering that heat conduction inside a material is a dynamic diffusion process, the image acquisition time series is set as follows: ( (Number of time points for continuous acquisition across multiple frames). For each valid sampling point on the skeleton. At any specific moment On the thermal image, based on the physical attenuation law of local thermal resistance, the system performs the following dynamic extraction process to obtain a subset of three-dimensional spatial features:

[0152] Extraction Action 1, Normal Orientation and Ray Emission: Only at the extreme moment of thermal shock ( Extracting the main skeleton And solidify each node The orthogonal normal vector. At subsequent decay times ( ), continued Rays are emitted from a time-determined normal vector path. This ensures multi-temporal characteristics. It describes the change in thermal resistance on the same physical cross-section.

[0153] Extraction Action 2, Subpixel Stepping and Adaptive Truncation: Setting the Adaptive Attenuation Cutoff Coefficient (In this embodiment, 15% of the local peak value is taken). The system reads the gradient modulus along the probe ray to both sides. When the sampling gradient value is lower than the cutoff threshold When it is determined that the system has moved out of the effective thermal resistance zone, the system immediately cuts off the ray, thereby dynamically locking onto the target area. Forward truncation boundary at time and backward cutoff boundary .

[0154] Action 3: Spatial Feature Quantization Calculation: Within the dynamically locked ray path range, the three-dimensional spatial features of the node are quantified using the following mathematical expression. Wherein,

[0155] Orthogonal gradient integral features The calculation formula is:

[0156]

[0157] Local gradient peak characteristics The calculation formula is:

[0158]

[0159] Effective thermal resistance width characteristics The calculation formula is:

[0160]

[0161] The specific physical and mathematical meanings of each parameter in the above formula are as follows:

[0162] Indicates time The extracted orthogonal gradient integral features are used to quantify the total resistance energy encountered by heat flow through the node;

[0163] Represents skeleton nodes exist The total value of the orthogonal gradient integrals at time step 1;

[0164] Indicates in At time , along a one-dimensional orthogonal ray path The gradient modulus of a specific sampling point;

[0165] It represents the spatial distance between sampling points (i.e., the physical distance between adjacent sampling points, such as the equivalent of a single pixel).

[0166] Indicates time The extracted local gradient peak features are used to reflect the absolute temperature step extreme value at the moment when the heat flow is cut off.

[0167] This indicates the operation of finding the maximum value within the valid cutoff interval;

[0168] This indicates that at the k-th acquisition time, based on the forward truncation boundary... and backward cutoff boundary Determined forward discrete boundary index;

[0169] This indicates that at the k-th acquisition time, based on the forward truncation boundary... and backward cutoff boundary Determined backward discrete boundary index;

[0170] Summation interval to Represents the topological skeleton node of the crack Centered on the front effective thermal resistance boundary, integrate along a one-dimensional orthogonal ray path from the back effective thermal resistance boundary to the front effective thermal resistance boundary;

[0171] Indicates time The extracted effective thermal resistance width feature is used to decouple the geometric interference of crack apparent width;

[0172] and They represent in The forward and backward truncation boundaries are calculated at each time step.

[0173] Then, the node exist The feature subsets extracted at consecutive time points are concatenated and spliced ​​together to construct the final feature set. The joint feature vector is x = [x1,1,x1,2,x1,3,…,xK,1,xK,2,xK,3]T. Finally, the input data is standardized. In the formula, This represents the specific value in the j-th dimension of the feature vector x. and They represent the first two elements in the pre-built global training set. The arithmetic mean and standard deviation of the dimensional features.

[0174] Regression Model Architecture, Applications, and Hyperparameter Optimization: Constructing Nonlinear Mapping Functions This embodiment uses a random forest regression model as the core inversion algorithm, which consists of multiple independent decision trees. The ensemble learning architecture is constructed, where M is the total number of decision trees.

[0175] Model specific structure and splitting mechanism:

[0176] During the training and construction phase, in order to ensure the overfitting resistance and structural differences of each decision tree in the ensemble model, the system uses Bootstrap sampling technology to construct an absolutely independent training subset for each decision tree.

[0177] Specifically, assuming a globally standardized feature dataset was constructed through preliminary experimental data collection and extraction. This global dataset contains a total of Historical sampling nodes along the route at known depths, each sample is composed of... The dimensional input feature vector is composed of the corresponding true depth label. For the 3rd dimension... Each decision tree is used to construct a dedicated training subset. At that time, the system will retrieve data from the global dataset. The process performs a "random selection with replacement" operation, and the total number of selections is strictly set to a certain limit. Second-rate.

[0178] The training subset generated through this independent sampling mechanism The total data size remains unchanged. A subset of samples. However, in its internal data structure, due to the "with replacement" characteristic, this training subset... The dataset must contain a certain proportion of completely duplicated feature samples; at the same time, theoretically, about 36.8% of the original sampling nodes in the global dataset will not be selected even once during the entire extraction process. These unselected independent samples constitute the out-of-bag data, which is directly used for the internal cross-validation and generalization error evaluation of the decision tree.

[0179] It is precisely because of the training subset obtained by each decision tree The distinct distribution of duplicate and missing data in the samples ensures that each tree experiences a unique "local data morphology," guaranteeing structural diversity for each tree during subsequent growth. Then, when splitting internal nodes within each tree, the algorithm only performs splitting within its respective training subset. Based on the mean square error minimization (MSE) criterion, the optimal splitting features and spatial partitioning threshold are selected, and the final leaf nodes are recursively generated.

[0180] Model application and basis prediction depth output:

[0181] In the application inference stage, high-dimensional joint feature vectors Parallel crossing Each decision tree. ,feature The node splitting rules are applied layer by layer, and the node eventually falls into a specific leaf node. The base prediction depth output by this leaf node. In mathematical terms, it is equal to the arithmetic mean of the true depth values ​​of all historical training samples that fall into that leaf node during the model training phase. Its specific calculation formula is:

[0182]

[0183] In the formula, Indicates the first Each decision tree represents the current input features. Output the base prediction depth value;

[0184] This indicates that during the model training phase, the data was partitioned and fell into that leaf node. The set of indices of all training samples;

[0185] This set The total number of training samples included;

[0186] Indicates the first The true depth label (i.e., known prior depth) corresponding to each historical training sample.

[0187] Finally, the model integrates the results from all subtrees and uses an arithmetic average mechanism to output the final predicted physical depth value for that spatial node. ,in, M represents the total number of decision trees in the model, M=100.

[0188] Hyperparameter adaptive optimization mechanism:

[0189] The model's core hyperparameters are rigorously optimized globally using a combination of grid search and V-fold cross-validation to determine the optimal combination that minimizes the mean squared error (MSE) criterion. The definition of MSE is given below:

[0190] During model training, define the target loss function. The mean squared error between the predicted and actual values:

[0191]

[0192] Where N is the total number of training samples, Indicates the first The true depth labels corresponding to each historical training sample Indicates the first A high-dimensional joint feature vector of historical training samples, Indicates the model input number. The prediction depth output after considering the features of each sample;

[0193] Iteratively update parameters using a greedy splitting algorithm until the loss function The model converges to the global minimum on the validation set, at which point the optimal combination of model parameters is obtained. Model training is completed. During the application deployment phase, the features of the crack to be tested are extracted and standardized, and then input end-to-end into the frozen optimal model. The absolute physical depth quantization value of the crack node is directly output at the output end. .

[0194] To clearly illustrate the entire data flow mechanism of this invention, from physical image acquisition and feature dataset construction to quantitative output of unknown depth, the entire embodiment is divided into two stages: "model calibration and training" and "unknown crack detection inference".

[0195] Phase 1: Construction of Multi-Depth Crack Feature Dataset and Model Training

[0196] 1. Experimental sample preparation: Prepare or select multiple standard test samples of material cracks with different known physical depths in advance (e.g., multiple cracks with actual depths of 5.0 mm, 10.0 mm, 15.0 mm, and 20.0 mm).

[0197] 2. Time-series data acquisition: The same unilateral heat flux excitation (e.g., heating at 90°C for 15 minutes) was applied to multiple crack samples of known depths. After removing the heat source, infrared thermal imagers were used to collect data under the same cooling time series (e.g.,...). Two-dimensional infrared radiation images of the cooling process were acquired to establish a thermal image dataset that includes the time dimension.

[0198] 3. Batch Sampling and Feature Extraction Along the Crack: For each crack of known depth, the algorithm extracts the crack topological skeleton with a single pixel width, and then selects an effective sampling node at fixed physical distances (e.g., 1 cm) along the crack direction. For each node, the algorithm emits a probe ray along the unit probe vector, dynamically truncates it, and extracts its three-dimensional spatial features at three sampling times.

[0199] 4. Constructing the Training Matrix and Model Growth: For each sampling node, a 3K-dimensional joint feature vector is generated (K=3 in this embodiment, corresponding to a 9-dimensional joint feature vector), and its corresponding true depth label is used as the supervision label. After traversing all samples, a global high-dimensional feature dataset covering rich thermal resistance states is constructed. Then, the Bootstrap sampling technique is used to complete the independent growth and training of the random forest model.

[0200] Phase Two: Quantitative Detection and Inference of Unknown Crack Depth

[0201] 1. Unknown Target Feature Extraction: Suppose we are detecting a crack of unknown depth in a real-world engineering project. Image acquisition and skeleton extraction are performed under the same boundary conditions as described above. At the first moment... min (thermal shock extreme value), in conjunction with the appendix of this invention Figure 4 (Schematic diagram of orthogonal gradient integral feature distribution) The actual running algorithm data, with representative nodes capturing high signal-to-noise ratio extrema on the skeleton. For example, its orthogonal gradient integral value is 89.9 (corresponding to the attached figure). Figure 4 The highest point), the local gradient peak is 8.5, and the effective thermal resistance width is 94.0 pixels.

[0202] Subsequently, to fully demonstrate the multi-temporal feature input mechanism of the model, based on the physical thermal diffusion attenuation law, it is reasonable to deduce that the features extracted by this node at the subsequent two attenuation times exhibit the following trend: at the second acquisition time... At min (initial heat diffusion), due to the decrease in temperature difference and the diffusion of the heat-affected zone, the orthogonal gradient integral decreases to 65.2, the local gradient peak decreases to 5.3, and the effective thermal resistance width expands to 110.0 pixels. (At the third acquisition time...) The minimum thermal flux (deep thermal diffusion) is further reduced to 42.1, the local gradient peak is reduced to 3.1, and the effective thermal resistance width is expanded to 135.0 pixels.

[0203] 2. Deep Inversion Output: The features of this node at three time points are concatenated into a 9-dimensional joint feature vector. Input this feature into the already trained random forest regression model from stage one. Assume that in the first decision tree, the leaf node where this vector falls outputs the base prediction value. mm; the predicted value of the output base in the second decision tree is mm. The system integrates the base prediction depths of all decision trees and calculates the arithmetic mean. Finally, the system directly outputs the absolute physical depth prediction quantization value corresponding to the spatial node at the output end. mm. This completes the entire quantitative detection process from the original thermal image sequence to the absolute value of physical depth.

[0204] Example 2

[0205] Please see Figure 5 The present invention also provides a quantitative detection system for material crack depth based on a thermal gradient vector field, comprising:

[0206] A single-sided active thermal excitation module, including a portable chemical self-heating pack and thermal insulation cotton, is used to construct a single-sided heat flow excitation environment;

[0207] The infrared imaging acquisition module includes a portable infrared thermal imager fixed on a tripod, used to acquire infrared thermal image sequences during the cooling process;

[0208] The image processing and inversion calculation module is connected to the infrared imaging acquisition module and is used to execute the algorithm flow of the material crack depth quantitative detection method based on thermal gradient vector field and output the crack physical depth detection result.

[0209] The following detailed explanation, using specific methodologies, elaborates on the functions, implementation principles, and collaborative mechanisms of each module. Those skilled in the art should understand that there is a clear correspondence between the steps in the method embodiments and the modules in the system embodiments, and that the method steps can be implemented by the system modules executing corresponding computer program instructions.

[0210] Specifically, during on-site testing, the single-sided active thermal excitation module first constructs a single-sided thermal flow excitation environment on one side of the crack on the surface of the material under test: a portable chemical self-heating pack is activated and attached tightly to one side of the crack, and thermal insulation cotton is laid on its side near the crack to form a thermal insulation barrier, forcing the heat flow to pass laterally through the crack, forming a significant temperature step. Subsequently, the heat source and insulation material are removed, and the infrared imaging acquisition module acquires two-dimensional infrared radiation images of the cooling process of the crack area according to a preset time series, establishing a thermal image dataset containing the time dimension. Finally, the image processing and inversion calculation module receives the thermal image dataset and sequentially performs: two-dimensional thermal potential energy field mapping and preprocessing, global gradient vector field calculation, crack topology skeleton extraction, adaptive orthogonal ray integration, multi-dimensional feature fusion and depth inversion, and finally outputs the physical depth detection results corresponding to each node of the crack.

[0211] The collaborative work among the above modules achieves a complete closed loop from physical excitation and data acquisition to quantitative inversion, enabling automated and high-precision detection of crack depth without human intervention.

[0212] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for quantitatively detecting the depth of material cracks based on a thermal gradient vector field, characterized in that, The method includes the following steps: Step S1, construct a one-sided heat flow excitation environment: set an active heat source on one side of the crack on the surface of the material to be tested, and use insulating material to form an insulating block at the junction of the active heat source and the crack, forcing the heat flow to pass through the crack laterally, so as to form a temperature step at the crack interface. Step S2, Acquire transient thermal image sequence: After the thermal excitation ends, use an infrared thermal imager to acquire two-dimensional infrared radiation images containing the cooling process of the crack and the heat-affected zone on both sides, and establish a thermal image dataset containing the time dimension. Step S3, Constructing the thermal potential energy field and preprocessing: Mapping the two-dimensional infrared radiation image in the thermal image dataset into a discrete two-dimensional thermal potential energy field, and performing smoothing and noise reduction preprocessing on the two-dimensional thermal potential energy field. Step S4, calculate the global gradient vector field: calculate the global gradient vector of each pixel in the two-dimensional thermal potential energy field, and construct a global gradient vector field that reflects the instantaneous direction of heat flow transmission; wherein, the global gradient vector includes gradient modulus and gradient direction angle; Step S5, extract the crack topology skeleton: Based on the global gradient vector field, use the image segmentation algorithm to extract the binary image of the crack region, and use the morphological thinning algorithm to extract the crack topology skeleton with a single pixel width as a spatial measurement reference. Step S6, Adaptive Orthogonal Ray Integration: Traverse all nodes on the crack topology skeleton, determine the detection direction orthogonal to the local crack orientation based on the local thermal gradient direction at each node, emit a virtual detection ray along the detection direction, and dynamically determine the integration interval on the virtual detection ray based on the attenuation characteristics of the gradient modulus, and calculate the cumulative gradient energy value along the integration interval. The gradient energy accumulation value along the integration interval is calculated through the following steps: For any node on the crack topology skeleton The unit detection vector orthogonal to the local crack orientation is determined based on the local thermal gradient direction at the node. : Where i is the node number; For nodes gradient modulus, For nodes The gradient direction angle, and These represent the partial derivatives of the thermal potential energy field in the horizontal and vertical directions, respectively; Define along the unit probe vector One-dimensional orthogonal ray path in direction ;in For path parameters; One-dimensional orthogonal ray path Perform gradient energy integration to obtain the total thermal resistance. And used as the cumulative value of gradient energy: in, and These are the dynamically determined forward and backward truncation boundaries, respectively; k is the summation index variable; and K is the integration boundary index. For the space to be far from the walking distance, Represents skeleton nodes The unit probe vector at that location, This is the index of the forward discrete boundary determined based on the forward truncation boundary and the backward truncation boundary. This is a backward discrete boundary index determined based on the forward truncation boundary and the backward truncation boundary; Set attenuation coefficient And define the gradient modulus cutoff threshold. ; The integration interval is then dynamically determined as follows: ； ； In the formula, Indicates a forward truncation boundary. Indicates the backward cutoff boundary. Indicates the path along a one-dimensional orthogonal ray. The gradient modulus of the specific sampling point. Represents a node gradient modulus; The dynamically determined integration interval refers to: using the crack topology skeleton nodes Centered on, along a one-dimensional orthogonal ray path Search in both the forward and reverse directions, and when the gradient modulus of the sampling points on the path... First time less than the cutoff threshold At that time, the forward cutoff boundaries are determined respectively. and backward cutoff boundary and will This serves as the valid integration interval for that node; Step S7, Multidimensional Feature Fusion and Deep Inversion: For each node, at least the multidimensional spatial features, including the gradient energy accumulation value described in step S6, are extracted from the two-dimensional infrared radiation image of the cooling process at multiple acquisition times. These features are then concatenated to construct a high-dimensional joint feature vector, which is then input into the trained supervised learning regression model to invert the physical depth of the crack corresponding to the node.

2. The method for quantitatively detecting material crack depth based on thermal gradient vector field according to claim 1, characterized in that, In step S1, the active heat source is a portable chemical self-heating pack; the insulation material is thermal insulation cotton, which is closely attached to the material surface and laid on the side of the active heat source near the crack.

3. The method for quantitatively detecting material crack depth based on thermal gradient vector field according to claim 1, characterized in that, In step S3, a Gaussian kernel function is introduced. Spatial domain convolution operations are performed on two-dimensional infrared radiation images to construct a smooth thermal potential energy field. To suppress random thermal speckle noise generated by uncooled focal plane arrays.

4. The method for quantitatively detecting material crack depth based on thermal gradient vector field according to claim 3, characterized in that, For the smooth thermal potential energy field Apply the Sobel differential operator to calculate each pixel. The global gradient vector is used to characterize the gradient magnitude and gradient direction angle. in, Indicates gradient modulus; This represents the global gradient vector of the thermal potential energy field; express Norms are used to calculate the Euclidean length of a vector. Indicates the gradient direction angle; This represents the arctangent function, used to transform the spatial derivative into the angle between the heat flux vector and the coordinate axis; and These represent the partial derivatives of the thermal potential energy field in the horizontal and vertical directions, respectively.

5. The method for quantitatively detecting material crack depth based on thermal gradient vector field according to claim 1, characterized in that, In step S5, the crack topology skeleton with a single pixel width is extracted through the following steps: A gradient modulus histogram is constructed based on the global gradient vector field, and the low-value main peak and high-value long tail are determined based on the gradient modulus histogram. The optimal inflection point between the low-value main peak and the high-value long tail is automatically found using the maximum inter-class variance algorithm and used as an adaptive threshold. Construct gradient modulus greater than binary feature set ; Using the maximum connectivity operator Filtering out binary feature sets Discrete noise, and utilize the refinement operator Extract the topological skeleton of the crack with a width of one pixel.

6. The method for quantitatively detecting material crack depth based on thermal gradient vector field according to claim 1, characterized in that, In step S7, for each node at each acquisition time, the following feature subset is dynamically extracted from the multidimensional spatial features: The orthogonal gradient integral characteristic is defined by the cumulative gradient energy value; The local gradient peak characteristic is defined by the maximum gradient modulus within the integration interval; The effective thermal resistance width characteristic is defined by the sum of the forward cutoff boundary and the backward cutoff boundary of the integration interval; The feature subsets of the node at multiple acquisition times are concatenated and spliced ​​together to form the high-dimensional joint feature vector.

7. The method for quantitatively detecting material crack depth based on thermal gradient vector field according to claim 1, characterized in that, The supervised learning regression model trained in step S7 is a random forest regression model, and its training process includes: Construct a global training set consisting of multiple crack samples with known physical depths, where each sample contains a high-dimensional joint feature vector and its corresponding true depth label; Bootstrap sampling is used to construct an independent training subset for each decision tree, and node splitting and growth are performed based on the mean square error minimization criterion. Ultimately, the crack physical depth output by the random forest regression model is the arithmetic mean of the base prediction depths output by all decision trees.

8. A quantitative detection system for material crack depth based on thermal gradient vector field, characterized in that, include: A single-sided active thermal excitation module, including a portable chemical self-heating pack and thermal insulation cotton, is used to construct a single-sided heat flow excitation environment; The infrared imaging acquisition module includes a portable infrared thermal imager fixed on a tripod, used to acquire infrared thermal image sequences during the cooling process; The image processing and inversion calculation module is connected to the infrared imaging acquisition module and is used to execute the algorithm flow of the quantitative detection method for material crack depth based on thermal gradient vector field as described in any one of claims 1 to 7, and output the crack physical depth detection result.

Images