A method for crack diagnosis and composite reinforcement renovation of cracked wooden columns in antique-style buildings

Abstract

This invention relates to a method for crack diagnosis and composite reinforcement renovation of cracked wooden columns in antique-style buildings, belonging to the field of artificial intelligence technology. The method includes the following steps: Step 1: Acquire images of the cracked wooden column surface to generate a rectangular unfolded diagram, which is then preprocessed with grayscale conversion and edge preservation denoising. Ultrasonic detection is used to obtain the penetration depth index of each crack segment. Step 2: Acquire the column tilt angle and the eccentricity of the resultant pressure force at the column base. Step 3: Based on the most unfavorable expansion risk level of each crack segment, a repair process is matched among injection repair, wood inlay reinforcement, shape memory alloy intelligent hoop constraint, and partial unloading renovation. This method can stably identify two types of crack orientations and quantify penetration depth even under interference from wood grain texture and uneven lighting. It incorporates model uncertainty into risk decision-making, graded matching of active reinforcement processes, and forms a closed-loop safety evaluation, significantly improving the accuracy, safety margin, and traceability of diagnosis and reinforcement of cracked wooden columns in antique-style buildings.

Description

Technical Field

[0001] This disclosure relates to, but is not limited to, the field of artificial intelligence technology, and in particular to a method for crack diagnosis and composite reinforcement and renovation of cracked wooden columns in antique-style buildings. Background Technology

[0002] As a physical carrier of historical and cultural memory, antique-style buildings primarily use wood as their main structural material. Among them, wooden columns, as vertical load-bearing components, bear the axial load transmitted from the roof and superstructure for a long time. Due to the combined effects of multiple factors such as the anisotropic absorption and loss of moisture by wood itself, long-term temperature and humidity fluctuations, stress relaxation, and eccentric compression, shrinkage cracks extending along the column axis and localized ring cracks or tears appearing in the circumferential direction are commonly found on the surface of wooden columns, posing a continuous threat to the effective load-bearing capacity of the column section.

[0003] Several relatively mature technologies exist in the industry for the diagnosis and reinforcement of cracked wooden columns. Regarding crack identification, engineering practice has long relied on manual visual inspection combined with feeler gauges to measure crack surface width, while crack length is directly measured using a tape measure. In recent years, image segmentation methods based on convolutional neural networks have been introduced into the automatic identification of concrete cracks, wall cracks, and steel structure cracks. Researchers use encoder-decoder structures such as U-Net and DeepLab to perform end-to-end segmentation of crack pixels in images. For penetration depth assessment, the industry commonly uses impedance meters to drill radially into the wooden column to record resistance curves, or uses stress wave detectors to measure wave velocities at multiple points to invert the distribution of internal defects. In terms of reinforcement and repair, grouting repair, wood inlay reinforcement, and passive restraint by wrapping the column with carbon fiber cloth or steel hoops have been widely used. Some projects have also experimented with setting pre-tightened metal hoops on the outside of the column to provide more durable circumferential restraint.

[0004] However, the aforementioned existing technologies still reveal significant limitations when addressing cracked wooden columns in antique-style buildings. Manual measurement is significantly affected by ambient lighting, operational experience, and the ambiguity of the crack ends, resulting in low efficiency and poor retest stability. When traditional two-dimensional convolutional crack segmentation models are applied to wooden column images, the strong longitudinal periodic texture of the wood surface and uneven ambient lighting easily create pseudo-responses overlapping with actual cracks in the segmentation results. Furthermore, the model only provides deterministic outputs and cannot characterize its own predictive uncertainties. Impedance meters and stress wave detection rely on empirical thresholds, lacking a normalization mechanism for detection results at the same penetration depth under different tree species, moisture contents, and column diameters, leading to weak cross-sample comparability. While passive carbon fiber cloth or steel hoop reinforcement can inhibit crack propagation to some extent, its restraining force depends on the pre-tightening level during installation and cannot be actively adjusted during service life. Its restraining effect is limited for deep crack segments with significant propagation potential. Moreover, existing reinforcement methods do not form a safety level evaluation mechanism integrated with the diagnostic process, making it difficult to provide quantitative basis for subsequent secondary reinforcement decisions. Summary of the Invention

[0005] This disclosure provides a method for crack diagnosis and composite reinforcement renovation of cracked wooden columns in antique-style buildings. It can stably identify two types of cracks and quantify their penetration depth under the interference of wood grain texture and uneven lighting. It incorporates model uncertainty into risk decision-making, classifies and matches active reinforcement processes, and forms a closed-loop safety evaluation, which significantly improves the accuracy, safety margin, and traceability of crack diagnosis and reinforcement of cracked wooden columns in antique-style buildings.

[0006] To solve the above problems, the technical solution of the present invention is implemented as follows: A method for diagnosing cracks and performing composite reinforcement and renovation on cracked wooden columns in antique-style buildings includes the following steps: Step 1: Acquire images of the cracked wooden column surface to generate a rectangular unfolded map, which is then preprocessed by grayscale conversion and edge preservation denoising. Construct vertical 1D sequences and horizontal 1D sequences respectively and input them into the Mamba selective state space sequence segmentation model to output vertical crack probability maps and horizontal crack probability maps. Fuse the vertical crack probability maps and horizontal crack probability maps to obtain a fused crack segmentation mask map. Extract the crack centerline orientation angle and physical crack width distribution sequence of each crack segment from the fused crack segmentation mask map, and obtain the penetration depth index of each crack segment through ultrasonic detection. Step 2: Collect the column tilt angle and the eccentricity of the resultant pressure force at the column base. Concatenate the crack centerline orientation angle, physical crack width distribution sequence, penetration depth index, column tilt angle, and eccentricity of the resultant pressure force at the column base of each crack segment to form a feature description vector and input it into the pre-trained risk assessment model to obtain a crack propagation risk score. Use the pre-established independent calibration sample set to determine the conformal calibration boundary value using the conformal prediction method. Superimpose the conformal calibration boundary value on the crack propagation risk score to obtain the most unfavorable propagation risk level of each crack segment. Step 3: Based on the most unfavorable propagation risk level of each crack segment, match the repair process for each crack segment among the injection repair process, wood inlay reinforcement process, shape memory alloy intelligent hoop constraint process, and partial unloading renovation process; the shape memory alloy intelligent hoop constraint process triggers the phase transformation recovery shrinkage of the shape memory alloy wire wound on the outer surface of the column by heating, forming a circumferential hoop force to constrain the crack; after the reinforcement operation is completed, extract the physical crack width distribution sequence and penetration depth index of the residual crack segment of the reinforced wooden column according to the process of Step 1, combine the column tilt angle and the eccentricity of the resultant force of the column foot pressure to form a renovation safety evaluation vector and output the renovation safety level.

[0007] Furthermore, in step 1, the edge preservation and noise reduction preprocessing uses anisotropic diffusion filtering. Anisotropic diffusion filtering preserves crack edges with a low diffusion rate in areas with large gray-level gradients, and smooths wood grain textures and areas with uneven lighting with a high diffusion rate in areas with small gray-level gradients.

[0008] Furthermore, in step 1, the vertical 1D sequence is formed by extracting gray values ​​column by column along the column axis and arranging them in order from the top of the column to the bottom of the column, and the horizontal 1D sequence is formed by extracting gray values ​​row by row along the circumferential direction and arranging them in order of circumferential unfolding.

[0009] Furthermore, in step 1, the Mamba selective state space sequence segmentation model includes a first Mamba selective state space sequence segmentation model and a second Mamba selective state space sequence segmentation model. The first Mamba selective state space sequence segmentation model is input with a vertical 1D sequence and outputs a vertical crack probability map, while the second Mamba selective state space sequence segmentation model is input with a horizontal 1D sequence and outputs a horizontal crack probability map. The first Mamba selective state space sequence segmentation model and the second Mamba selective state space sequence segmentation model use the same network structure and are both trained in advance using labeled crack mask maps with pixel-level crack category labels as supervision signals. Each Mamba selective state-space sequence segmentation model maintains a fixed-dimensional hidden state vector. Elements in the sequence are input one by one. At each time step, the current input element is mapped to a selectively gated vector through the first linear projection layer. The components of the selectively gated vector take values ​​between 0 and 1 and are used to scale the retention ratio of each dimension of information passed from the previous time step in the hidden state vector element by element. At the same time, the current input element is mapped to an input driving vector through the second linear projection layer. The input driving vector is added element by element to the selectively gated and scaled hidden state vector to obtain the updated hidden state vector for the current time step. The updated hidden state vector is then mapped to the crack probability value for the current time step through the output linear layer.

[0010] Furthermore, in step 1, the process of fusing the longitudinal crack probability map and the transverse crack probability map to obtain the fused crack segmentation mask map is as follows: pixel-by-pixel, the longitudinal crack probability map and the transverse crack probability map are retrieved... Figure 2 The larger value among them is used to obtain the fused crack probability map. The fused crack probability map is binarized according to the preset segmentation threshold to obtain the binarized crack mask map. Isolated small connected component deletion and skeleton continuity filtering are performed on the binarized crack mask map to obtain the fused crack segmentation mask map. The process of extracting the crack centerline orientation angle and physical crack width distribution sequence of each crack segment from the fused crack segmentation mask map is as follows: Connected component labeling and Zhang-Suen skeleton thinning are performed on the fused crack segmentation mask map to extract the crack centerline of each crack segment. The pixel span of the fused crack segmentation mask map in the direction perpendicular to the crack centerline is measured point by point along the crack centerline and converted into physical crack width according to the cylindrical unfolding scale. The physical crack width distribution sequence is composed of all the physical crack widths of each crack segment.

[0011] Furthermore, in step 1, the process of obtaining the penetration depth index of each crack segment by ultrasonic detection is as follows: ultrasonic transmitting transducers and ultrasonic receiving transducers are arranged opposite each other along the radial direction of the column at the midpoint of each crack segment. The propagation time of the ultrasonic pulse through the cross section of the column is measured. The propagation time is compared with the reference propagation time of intact wood calibrated under the same tree species, the same column diameter grade, the same moisture content range, and the same temperature range. The ratio of the extension of propagation time to the reference propagation time of intact wood is used as the penetration depth index.

[0012] Furthermore, in step 2, the physical crack width distribution sequence in the feature description vector is spliced ​​with the maximum value of the physical crack width distribution sequence; the risk assessment model is a gradient boosting decision tree model; the process of determining the conformal calibration boundary value using the conformal prediction method is as follows: each calibration sample in the independent calibration sample set has a predicted risk score output by the same gradient boosting decision tree model and an actual risk level determined by manual review. The actual risk level is converted into a calibrated risk value with the same value range as the crack propagation risk score through a level mapping table. The absolute deviation between the predicted risk score and the corresponding calibrated risk value of each calibration sample is taken as the inconsistency score of the calibration sample. The inconsistency scores of all calibration samples are sorted from smallest to largest. The smaller of the value obtained by taking the position number equal to the total number of calibration samples plus 1, multiplying by the preset confidence level, and rounding up, and the total number of calibration samples, is taken as the conformal calibration boundary value. The upper limit of the risk prediction interval is obtained by superimposing the conformal calibration boundary value on the crack propagation risk score. The upper limit of the risk prediction interval is taken as the most unfavorable propagation risk level of each crack segment.

[0013] Furthermore, in step 3, the process of matching the repair process according to the most unfavorable expansion risk level of each crack segment is as follows: crack segments with the most unfavorable expansion risk level falling into the preset low-risk threshold range are assigned to the injection repair process; crack segments with the most unfavorable expansion risk level falling into the preset medium-risk threshold range are assigned to the wood inlay reinforcement process; crack segments with the most unfavorable expansion risk level falling into the preset high-risk threshold range are assigned to the shape memory alloy smart hoop constraint process; and crack segments with the most unfavorable expansion risk level falling into the preset extremely high-risk threshold range are assigned to the partial unloading renovation process. The preset low-risk threshold range, preset medium-risk threshold range, preset high-risk threshold range, and preset extremely high-risk threshold range are arranged continuously from low to high according to the risk value and do not overlap with each other. The injection repair process includes cleaning the crack, injecting low-viscosity wood repair adhesive into the crack, and sealing the crack surface. The wood inlay reinforcement process includes opening a repair groove along the crack direction, embedding a reinforcing wood strip of the same material, gluing and pressing, and surface coloring. The partial unloading renovation process includes setting up temporary supports to transfer the column load, removing the failed wood area, replacing wood, and restoring the finish.

[0014] Furthermore, in step 3, the execution process of the shape memory alloy intelligent hoop constraint technology is as follows: a heat-resistant flexible insulating layer is wrapped around the outer surface of the column in the crack segment area. A nickel-titanium alloy wire is spirally wound around the outer surface of the heat-resistant flexible insulating layer along the circumference of the column as a shape memory alloy wire. The nickel-titanium alloy wire is in the martensitic phase at room temperature. During winding, a constant initial tension force is applied through a tension control device. The winding range covers the entire length of the crack segment along the column axis and extends by one spiral spacing at each end. K-type thermocouples are attached at equal intervals along the spiral path on the outer surface of the nickel-titanium alloy wire as wire temperature sensors. The monitoring point is a thin-film temperature sensor placed between the heat-resistant flexible insulating layer and the surface of the column wood as a monitoring point for the wood surface temperature. Resistance heating is performed by applying DC current to both ends of the nickel-titanium alloy wire. The temperature controller simultaneously collects the wire temperature and the wood surface temperature. When the wood surface temperature reaches the preset upper limit of the wood protection temperature, the heating current is reduced to limit the temperature rise of the wood surface. When the wire temperature reaches the austenitic phase transformation termination temperature, the heating current is cut off. After the nickel-titanium alloy wire cools naturally to room temperature under the constrained state, it maintains the recovered state of the hoop and forms a continuous circumferential constraint force on the column surface.

[0015] Furthermore, in step 3, the renovation safety evaluation vector consists of the effective bearing area ratio, column tilt angle, column base pressure resultant eccentricity, and the number of residual crack segments. The effective bearing area ratio is determined by: determining the cross-sectional weakening area corresponding to each residual crack segment based on the circumferential position of all residual crack segments on the column cross-section, the physical crack width distribution sequence, and the penetration depth index; and subtracting all cross-sectional weakening areas from the nominal cross-sectional area of ​​the wooden column to obtain the effective bearing area ratio. The renovation safety level is output by: comparing each component in the renovation safety evaluation vector with its corresponding preset safety threshold; outputting a qualified renovation safety level when all components meet the corresponding preset safety threshold, and outputting a renovation safety level requiring secondary reinforcement when any one component exceeds the corresponding preset safety threshold.

[0016] This invention provides a method for crack diagnosis and composite reinforcement renovation of cracked wooden columns in antique-style buildings, which has the following beneficial effects: First, by constructing longitudinal 1D sequences and transverse 1D sequences along the column axis and circumferential direction respectively and inputting them into the Mamba selective state space sequence segmentation model, and using selective gating vectors to adaptively retain the gray-scale abrupt changes of cracks continuously distributed along the sequence direction with high retention and the low retention of short-range periodic wood grain texture, it is possible to stably separate longitudinal and transverse cracks even under strong parallel grain texture and uneven lighting interference on the wooden column surface. Both types of cracks with different orientations can always obtain long-range responses on one of the sequences, and the recognition recall and positioning accuracy are significantly better than general two-dimensional convolutional segmentation methods. Second, the penetration depth index obtained by radially opposed ultrasonic detection is used to normalize the reference propagation time, making the penetration depth assessment under different tree species, different column diameters, and different moisture contents comparable across samples. Third, by superimposing conformal calibration boundary values ​​derived from an independent calibration sample set onto the point estimates given by the risk assessment model, the safety margin resulting from "occasional underestimation by the model" is transparently incorporated into the decision-making process, preventing high-risk segments from being misjudged as low-risk segments. Fourth, based on the most unfavorable expansion risk level, four repair processes are graded and matched, and a shape memory alloy intelligent hoop constraint process is used to form a continuous active circumferential constraint force on the cylindrical surface through phase transformation recovery shrinkage triggered by heating, which has a stronger ability to suppress deep through cracks compared with traditional passive reinforcement. Fifth, after reinforcement, the renovation safety evaluation vector and multi-threshold sub-criteria are used to output the renovation safety level, providing a clear quantitative basis for whether secondary reinforcement is needed. Attached Figure Description

[0017] Figure 1 A top-view cross-sectional schematic diagram of radial ultrasonic testing of a cracked wooden column provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the interval calibration of crack propagation risk scores for eight crack segments based on conformal prediction, and the matching of repair processes according to these scores, provided in an embodiment of the present invention. Figure 3 This is a side view of the construction layout of the shape memory alloy smart hoop constraint process provided in an embodiment of the present invention on a cracked wooden column. Detailed Implementation

[0018] A method for diagnosing cracks and performing composite reinforcement and renovation on cracked wooden columns in antique-style buildings includes the following steps: Step 1: Acquire images of the cracked wooden column surface to generate a rectangular unfolded map, which is then preprocessed by grayscale conversion and edge preservation denoising. Construct vertical 1D sequences and horizontal 1D sequences respectively and input them into the Mamba selective state space sequence segmentation model to output vertical crack probability maps and horizontal crack probability maps. Fuse the vertical crack probability maps and horizontal crack probability maps to obtain a fused crack segmentation mask map. Extract the crack centerline orientation angle and physical crack width distribution sequence of each crack segment from the fused crack segmentation mask map, and obtain the penetration depth index of each crack segment through ultrasonic detection. Step 2: Collect the column tilt angle and the eccentricity of the resultant pressure force at the column base. Concatenate the crack centerline orientation angle, physical crack width distribution sequence, penetration depth index, column tilt angle, and eccentricity of the resultant pressure force at the column base of each crack segment to form a feature description vector and input it into the pre-trained risk assessment model to obtain a crack propagation risk score. Use the pre-established independent calibration sample set to determine the conformal calibration boundary value using the conformal prediction method. Superimpose the conformal calibration boundary value on the crack propagation risk score to obtain the most unfavorable propagation risk level of each crack segment. Step 3: Based on the most unfavorable propagation risk level of each crack segment, match the repair process for each crack segment among the injection repair process, wood inlay reinforcement process, shape memory alloy intelligent hoop constraint process, and partial unloading renovation process; the shape memory alloy intelligent hoop constraint process triggers the phase transformation recovery shrinkage of the shape memory alloy wire wound on the outer surface of the column by heating, forming a circumferential hoop force to constrain the crack; after the reinforcement operation is completed, extract the physical crack width distribution sequence and penetration depth index of the residual crack segment of the reinforced wooden column according to the process of Step 1, combine the column tilt angle and the eccentricity of the resultant force of the column foot pressure to form a renovation safety evaluation vector and output the renovation safety level.

[0019] When acquiring images of cracked wooden pillars in an antique-style building, one or more industrial-grade area array cameras with a resolution of at least 24 megapixels can be set up at a distance of 0.6 to 1.2 meters from the pillar surface to take pictures at equal angular intervals around the circumference of the pillar. In a typical implementation, one image of the pillar surface is acquired every 30 degrees, resulting in 12 images around the pillar. Adjacent images retain a 15% to 25% overlap in the circumferential direction to ensure the stability of subsequent stitching. For cracked wooden pillars with large diameters or long lengths, images can be acquired in multiple segments along the pillar axis. Each segment is stitched circumferentially separately before being stitched together as a whole along the pillar axis. During the acquisition process, a ring-shaped LED light source or a diffused light panel is used for supplementary lighting to avoid direct light forming strong reflective highlights on the wood grain of the pillar surface. Otherwise, the reflected highlights will form a "false crack" signal in the subsequent grayscale image, which is the opposite of the crack's grayscale depression but has a similar texture.

[0020] After obtaining multiple cylindrical images, intrinsic distortion correction is first applied to each image to eliminate radial and tangential lens distortion. Then, SIFT or ORB feature points are extracted in the overlapping areas of adjacent images, and mismatched point pairs are eliminated using the RANSAC algorithm to estimate the homography transformation between adjacent images. Considering that a cylinder is a cylindrical surface rather than a plane, directly performing planar homography stitching on adjacent images would introduce perspective distortion in areas far from the stitching seam. Therefore, each cylindrical image is first back-projected onto a cylindrical coordinate system with the cylinder axis as the central axis using a pre-measured cylinder diameter estimate. Then, the image in the cylindrical coordinate system is cut along the cylinder axis and flattened into a rectangular unfolded image. The advantage of this cylindrical parameterization is that two points equidistant along the circumferential distance on the cylinder surface also have equidistant lateral pixel distances on the rectangular unfolded image, thus providing spatial consistency for the obtained physical crack width. The lateral direction of the rectangular unfolded image corresponds to the circumference of the cylinder, and the longitudinal direction corresponds to the axial direction of the cylinder; the width of the unfolded image is equal to the cylinder circumference divided by the cylinder sampling precision, and the height is equal to the cylinder length divided by the longitudinal sampling precision. In one typical implementation, the lateral pixel sampling precision of the rectangular unfolded image is 0.2 mm to 0.5 mm per pixel, and the cylindrical unfolded scale is denoted as... The unit is millimeters per pixel, which serves as the basis for subsequent physical crack width conversion. Optionally, for situations where it is not possible to acquire images around the column in a 360-degree circumferential manner (e.g., the column is close to a wall or is obscured by structural components), only the visible arc segment is image acquired, and the unfolded image corresponding to the visible arc segment is used as a rectangular unfolded image. Subsequent steps are still performed according to the same process, with only the uncovered areas marked in the output results.

[0021] refer to Figure 1 This study presents the geometric relationship at the midpoint of a typical crack segment, formed by the ultrasonic transmitting transducer, the ultrasonic receiving transducer, the cross-section of the column under test, the crack segment, and the sound wave propagation path. The column cross-section is defined by its nominal radius. The circular outer contour is shown, with the center indicated by a cross and labeled as the column axis. Inside the outer contour are several concentric dashed rings representing the annual rings of the wooden column, illustrating the approximately isotropic stress characteristics of the wood with the column axis as its axis of symmetry. On the upper right side of the outer contour, an irregular broken line starting from the outer edge of the column surface and extending radially inwards indicates a segment of an actual crack. This broken line exhibits slight trembling along its extension direction to reflect the non-straight direction of the crack along the wood fiber direction; the radial extension length of this crack segment is approximately [missing information]. The 0.6 times that of the previous year reflects the working condition where the penetration depth is significant but has not penetrated the entire cross section.

[0022] At the midpoint of the centerline of the crack segment, a pair of ultrasonic transducers are arranged radially opposite each other along the axis of the column: an ultrasonic transmitting transducer is attached to one side of the outer edge of the crack segment, and an ultrasonic receiving transducer is attached to the other side of the outer edge along the opposite radial direction. The transducers are shown as semi-rectangles in the figure. The side in contact with the column surface is represented by a thick solid line to indicate the acoustic coupling contact surface, demonstrating that the transducer is tightly bonded to the column surface using special coupling grease, avoiding acoustic impedance mismatch introduced by air gaps. Three concentric arcs are drawn outside the ultrasonic transmitting transducer to indicate the pulsed ultrasonic signal radiated into the column when the transducer is electrically excited. The opening of the arcs faces inward to represent the incident direction of the acoustic energy.

[0023] There are two propagation paths with different physical meanings from the ultrasonic transmitting transducer through the interior of the column to the ultrasonic receiving transducer. The first path is the reference path. The radial dashed line segment connecting the two transducers represents the radial propagation time required for sound waves to pass through the column cross-section in intact (crack-free) wood under the same tree species, column diameter grade, moisture content range, and temperature range, serving as a baseline reference. Line 2 represents the actual sound ray. The broken line, representing the actual propagation path of sound waves from the transmitting transducer, first approaching the crack segment radially, then undergoing a lateral deflection near the crack segment's endpoint to bypass the crack, returning radially, and finally reaching the receiving transducer, corresponds to the actual propagation path of sound waves reaching the receiving side after being bypassed or scattered by the crack surface when the crack segment actually exists within the test column. The inflection point of this broken line near the crack segment's end characterizes the deflection and diffraction of sound energy at the interface of abrupt changes in the medium, which is the geometric root cause of the extended propagation time relative to the reference path.

[0024] exist Figure 1 The diameter of the column is indicated below by a double-headed arrow. This serves as a dimensional reference for the propagation path length, facilitating the subsequent establishment of a dimensionless ratio between the propagation time extension and the column diameter. Figure 1 The upper right and lower right corners are marked with brief text labels for the crack segment, reference path and actual sound line, respectively. The starting point of the leader line points to the corresponding geometric element, while the text body is placed in the blank area outside the column outline to avoid visual overlap with the column section, growth ring and crack segment.

[0025] Figure 1 The core technological idea it embodies lies in: using the reference propagation time of intact wood under the same testing parameters. Measured propagation time of the crack segment under test By conducting comparative measurements under the same radial direction and working conditions, the scattering effect of sound waves around the crack segment can be converted into a quantifiable extension of propagation time. Further Define a penetration depth index to normalize the test results of cracked wooden columns of different diameters to the same dimensionless scale, providing comparable input for subsequent crack propagation risk zone assessment.

[0026] The obtained rectangular unfolded image is first converted to grayscale. In one implementation, a weighted average method is used to convert the 3-channel color image into a single-channel grayscale image: ,in Represents the grayscale image in pixel coordinates grayscale value at that location , , These represent the red, green, and blue channel components of the original color image at that pixel, respectively. The coefficients 0.299, 0.587, and 0.114 represent the relative sensitivity coefficients of the human eye to red, green, and blue wavelengths of light. Weighting the grayscale image using these coefficients allows the contrast between the dark wood grain and the light sapwood on the wooden column surface to more closely approximate human subjective perception. Other optional grayscale methods include taking the maximum or minimum values ​​of the RGB channels, or taking only the G channel. For cracked wooden columns covered with a layer of light-colored varnish or tung oil, since the spectral response of the reflective layer is yellowish-green, retaining only the G channel in the grayscale image allows for a higher contrast between the exposed dark wood interior at the crack and the reflective surface layer.

[0027] After grayscale conversion, anisotropic diffusion filtering is used to denoise the grayscale image. The iterative process of anisotropic diffusion filtering can be written as follows: ,in Indicates the first Position after the next iteration grayscale value at that location This is the time step factor. Indicates by position The set of directions consisting of 4-neighborhoods or 8-neighborhoods. Indicates along direction grayscale gradient (i.e. (The gray values ​​of adjacent pixels minus the gray value of the center pixel) This represents the diffusion coefficient function. A typical choice for the diffusion coefficient function is... ,in This is a gradient threshold constant that controls the sensitivity of the diffusion behavior to the gray-level gradient. When much smaller hour, Approaching 1, the diffusion is close to classical isotropic Gaussian smoothing, thus achieving a significant smoothing effect on noise, wood grain, and lighting gradients; when Much larger hour, Approaching 0, diffusion across this direction is strongly suppressed, thus preserving the original grayscale jumps in areas with large grayscale gradients (typically corresponding to crack edges). This adaptive behavior allows the filter to preserve crack edges with a low diffusion rate in areas with large grayscale gradients, and to smooth wood grain textures and areas of uneven lighting with a high diffusion rate in areas with small grayscale gradients. This preserves the clarity of crack boundaries while suppressing grain textures, growth ring textures, and lighting gradations, preventing subsequent sequence segmentation models from misclassifying high-contrast wood grain areas as cracks. In one typical implementation, the time step coefficient... The gradient threshold constant is set between 0.20 and 0.25. The number of iterations is between 15 and 30. The number of iterations can be increased from 10 to 30. For images of wooden pillars under strong direct light or with severe local shadows, the number of iterations can be increased to more than 50. In addition, a CLAHE-limited contrast-adaptive histogram equalization should be performed on the grayscale image before each iteration to alleviate large-scale uneven lighting. Other optional edge-preserving denoising methods include bilateral filtering, guided filtering, and nonlocal mean filtering. However, when dealing with strong parallel-grain periodic textures on the surface of wooden pillars, anisotropic diffusion filtering is preferred because it uses the gradient as the only control variable and has more stable preservation characteristics for high-gradient slender structures like cracks that extend in any direction.

[0028] One-dimensional sequences are constructed along the cylindrical axis and circumferential direction for the preprocessed grayscale image. The rectangular unfolded image is considered as a single image with a height of... Width is grayscale matrix ,in This represents the number of pixels along the cylindrical axis. This represents the number of pixels in the circumferential direction. Extraction is performed column-by-column along the cylindrical axis: for the... List( The column's grayscale values ​​are extracted sequentially from top to bottom to obtain a length of [length missing]. 1-dimensional sequence As a vertical 1D sequence; extract row by row along the circumferential direction: for the first... OK( ), extracting the grayscale values ​​of the row sequentially from the starting point to the ending point in a circumferential unfolding process to obtain a length of 1D sequence This is treated as a horizontal one-dimensional sequence. The final result is... 1-dimensional vertical sequence and A transverse 1D sequence is constructed. The consideration for simultaneously constructing two types of 1D sequences in different directions is that cracks in cracked wooden pillars during drying mainly extend along the pillar axis and are elongated. However, near-transverse (circumferential) tearing or ring cracking may also occur in knots, stress concentration areas, or under load. If only one direction is constructed, for cracks approximately perpendicular to that direction, the gray-level jumps between adjacent samples in the sequence will be very frequent, and the model may easily misclassify the cracks as discrete high-frequency noise and ignore them. Constructing 1D sequences along two mutually orthogonal directions ensures that cracks of each orientation always appear as long-range, low-gray-level continuous segments in one type of sequence, forming separable statistical characteristics from the short-range periodic wood grain. Optionally, for cases where oblique splitting accounts for a high proportion, further 1D sequences can be constructed along... A separate 1D sequence is constructed to improve the detection sensitivity of diagonal cracks. However, adding the diagonal sequence will increase the overall computational load of subsequent model inference by about 2 times. Therefore, in conventional ancient building maintenance scenarios, the detection requirements can be met by using two directions: longitudinal and transverse.

[0029] All the constructed vertical 1D sequences are input into the first Mamba selective state space sequence segmentation model, and all the constructed horizontal 1D sequences are input into the second Mamba selective state space sequence segmentation model. The first and second Mamba selective state space sequence segmentation models use the same network structure. Each Mamba selective state space sequence segmentation model internally maintains a hidden state vector of fixed dimension. In a typical implementation, the dimension of the hidden state vector is... Take values ​​from 64 to 128. Let the input 1D sequence be... ,in For sequence length (for a vertical 1-dimensional sequence) For horizontal 1D sequences ), Indicates the first The input elements at the nth time step (i.e., scalars or low-dimensional vectors resulting from simple linear transformation of grayscale values). The model at the nth time step... At each time step, the current input element is first mapped to a selectively gated vector through the first linear projection layer. ,in and These are the weight matrix and bias vector of the first linear projection layer, respectively. This represents the sigmoid activation function, which makes... Each component takes a value between 0 and 1; simultaneously, the current input element is mapped to an input driving vector through the second linear projection layer. ,in and These are the weight matrix and bias vector of the second linear projection layer, respectively. Then, the hidden state vector of the previous time step... Perform selective scaling and overlay the input-driven update of the hidden state vector at the current time step. ,in This represents the element-wise Hadamard product. It can be seen that the selectively gated vector... The values ​​of each component are between 0 and 1, which is equivalent to selectively retaining or forgetting information from the previous time step element by element in each dimension of the hidden state vector: in dimensions with values ​​close to 1, information from past time steps is retained and passed to the current time step almost without loss; in dimensions with values ​​close to 0, past information is quickly discarded, and new information is... The write operation is initiated. Finally, the hidden state vector will be updated. By mapping the output linear layer to the crack probability value at the current time step. ,in and These are the weight vector and bias scalar of the output linear layer, respectively. The value ranges from 0 to 1.

[0030] After pre-training, the model adaptively learned to maintain a high information retention rate for crack gray-level abrupt changes continuously distributed along the sequence direction (i.e., the corresponding position). The component values ​​are too high, while the information retention ratio for short-range periodic wood grain texture features is relatively low (i.e., the corresponding position...). (The component values ​​are too low). This learning bias stems from the fact that wood grain exhibits short-period alternations of light and dark in the sequence, with polarity reversal occurring after several pixels. During training, the neural network tends to quickly forget such high-frequency oscillating signals to avoid accumulating false edge responses in the hidden state. In contrast, cracks in the sequence present as unidirectional gray-level depressions spanning tens to hundreds of pixels. During training, the neural network learns to perform long-term memory on such long-range low-gray-level continuous segments to accumulate evidence, thereby outputting a high crack probability value at that location. The model training uses the cross-entropy loss function, and the supervision signal comes from manually labeled pixel-level crack masks: on the acquired sample cylindrical images, annotators with experience in ancient building restoration label the cracks (label 1) and the background (label 0) pixel by pixel, and construct vertical 1D sequences and horizontal 1D sequences in a manner consistent with the inference stage, using the labels at the corresponding positions as pixel-level supervision. In one typical implementation, the training set contains approximately 300 sample images of cracked wooden pillars of different tree species, diameters, and moisture contents. The batch size for a single training run is between 32 and 64, and the initial learning rate is [value missing]. to Between these iterations, using the Adam optimizer, convergence occurs after approximately 80 to 120 training epochs. Alternatively, each Mamba selective state-space sequence segmentation model can stack 4 to 8 of the aforementioned selective state-space layers, with residual connections and layer normalization connections between adjacent layers to improve the model's ability to fit complex crack morphologies; for sequences with excessively long lengths (e.g., ... In cases where memory is tight during a single inference, the entire sequence can be slid into sub-segments of length 1024 pixels with an overlap of 128 pixels between adjacent segments for inference. The complete probability sequence can then be pieced back together by taking the average of the overlapping areas.

[0031] After the model inference is completed, the crack probability values ​​at each time step output by the first Mamba selective state space sequence segmentation model for each column of the vertical 1D sequence are restored to a matrix of the same size as the rectangular unfolded diagram according to the original column number and time step number, thus obtaining the vertical crack probability map. The crack probability map is obtained by restoring the crack probability values ​​at each time step of the horizontal 1D sequence output of the 2nd Mamba selective state-space sequence segmentation model to a matrix of the same size as the rectangular unfolded diagram, based on the original row number and time step number. .

[0032] Probability map of longitudinal cracks Probability map of transverse cracks Perform pixel-by-pixel fusion, for each pixel position The fusion crack probability map is obtained by taking the larger probability value from the two probability maps. The reason for using a larger value for each pixel instead of averaging or multiplying is that most actual cracks only exhibit long-range gray-scale concavity features in one of the two directions—longitudinal cracks produce a strong response in the longitudinal 1D sequence along the column axis but a weaker response in the transverse 1D sequence along the circumferential direction, and transverse cracks are exactly the opposite. If an arithmetic or geometric mean is used, the longitudinal cracks will be significantly diluted by the low probability value of the transverse output, resulting in a fused response that is weaker than the output in a single direction. Taking a larger value ensures that as long as the model in any one direction confirms that a pixel belongs to a crack, that pixel will be retained in the fused crack probability map.

[0033] Fusion crack probability map The probability value of each pixel and the preset segmentation threshold Compare and generate binarized crack mask images :when season Otherwise In one typical implementation, the segmentation threshold... Take 0.5; for worm-eaten microcracks with weak gray-scale transitions at the crack edges, the value can be... The recall rate was lowered to 0.35 to 0.4 to improve the recall rate, and then false positives were suppressed by subsequent screening steps. Then, the binarized crack mask image was analyzed. To delete isolated small connected components: first... Perform 8-neighbor connected component labeling, count the number of pixels covered by each connected component, and select the component with a pixel count less than a preset minimum connected component threshold. The entire connected component is set to 0. How much to take based on the scale of the cylindrical surface? In one typical implementation, the number of pixels corresponding to a length of 8 mm to 15 mm on the actual cylindrical surface was selected. This selection aims to eliminate isolated wormholes, knot carbonization points, surface stains, and other small spots that are mistakenly identified as cracks. While these targets do form local depressions in grayscale, their geometric scale is much smaller than that of actual cracks. Eliminating them according to a scale threshold significantly reduces false detections without sacrificing crack recall. Subsequently, a skeleton continuity selection process is performed: the remaining connected components are refined using the Zhang-Suen skeleton to obtain their centerlines. The direction angles at each point along the centerline are calculated, and the variance of these direction angles is also calculated. Connected components with a direction angle variance exceeding a preset threshold are deemed to lack the elongated continuity characteristic of cracks (typically corresponding to irregular patches or randomly distributed wood damage), and their values ​​are set to 0. Finally, a fused crack segmentation mask is obtained. .

[0034] Mask image for fusion crack segmentation The 8-neighbor connected component labeling is performed again, with each independent connected component being numbered as a crack segment. Then, Zhang-Suen skeleton thinning is performed on each crack segment, iteratively deleting pixels on the boundaries of connected components that satisfy the preset topology preservation conditions, until the crack segment converges to a crack centerline with a width of one pixel. The Zhang-Suen algorithm is suitable for this scenario because it deletes pixels by judging the topology conditions of the pixel's 8-neighborhood in each iteration, which can gradually converge to the geometric center while maintaining connectivity and branching structure, avoiding the problem that ordinary morphological erosion can easily cut narrow cracks in the middle. For the crack centerline of each crack segment, the pixels on it are topologically sorted in order from one end to the other to obtain the centerline point sequence of the crack segment. ,in This is the pixel length of the centerline of the crack segment. Least squares linear fitting or principal component analysis is performed on the point sequence, and the angle between the fitted line (or the first principal direction) and the horizontal axis (i.e., the circumferential axis) of the rectangular unfolded graph is taken as the crack centerline orientation angle of the crack segment. In one typical implementation, the orientation angle is limited to... arrive between, This indicates that the crack segment extends circumferentially. This indicates that the crack segment extends completely along the column axis. For curved crack segments with significant changes in orientation, windows can be slid open along the centerline (the window length is typically 50 to 100 pixels). The orientation angle is calculated in each window, and the arithmetic mean of all windows is taken as the orientation angle of the crack centerline of the crack segment.

[0035] Next, the physical crack width of the crack segment is measured point by point along the crack centerline. For each point on the centerline... First, the local tangent direction of the centerline at a point is obtained by fitting several pixels before and after the point. Then, the search is performed on both sides along the direction perpendicular to the tangent (i.e., the crack normal), and the results are statistically analyzed on both sides. The pixel span at a given point is obtained by summing the number of pixels with consecutive values ​​of 1 and adding 1. For applications requiring high pixel span accuracy, sub-pixel interpolation (typically quadratic curve fitting based on grayscale gradient) can be used at both boundaries to... Precision improved to sub-pixel level. Combined with a cylindrical unfolded scale. Convert the pixel span to the physical crack width The unit is millimeters. The physical crack widths of each crack segment, obtained along its centerline, are arranged according to the centerline number to form the physical crack width distribution sequence for that crack segment. This method of distributing the crack along its centerline preserves more structural information than a single maximum or average width indicator: actual cracks often exhibit a spindle-shaped morphology, converging at the tip and widest in the middle. The sequence morphology itself carries the basis for judging whether the crack segment has a stress concentration end and whether it is at the propagation front, which is convenient for subsequent risk assessment. Optionally, for conditions where the resolution of the cylindrical surface acquisition images is insufficient to stably distinguish sub-millimeter crack widths, high-resolution 3D scanning and remeasurement of the centerline region of each crack segment can be performed using structured light projection or laser line scanning. The physical crack width obtained from the remeasurement can then replace the image measurement results to improve the accuracy of the physical crack width distribution sequence.

[0036] For each crack segment, a pair of ultrasonic transducers are arranged radially along the column (i.e., within the column's cross-section, along the direction passing through the column axis) at the midpoint of the crack segment's centerline: an ultrasonic transmitting transducer is attached to the outer side of the column where the crack segment is located, and an ultrasonic receiving transducer is attached to the opposite radial side of the outer side of the column. The transducer center frequency is chosen to be between 50 kHz and 150 kHz. Because wood is an anisotropic porous medium, excessively high frequencies of ultrasonic waves attenuate drastically in wood, limiting penetration depth; excessively low frequencies lack sufficient spatial resolution to distinguish local time delay changes caused by the crack. 50 kHz to 150 kHz is a commonly used frequency band that balances penetration and resolution. The transducers are coupled to the column surface using a special coupling grease to avoid acoustic impedance mismatch introduced by air gaps. The ultrasonic transmitting transducer emits a pulsed ultrasonic signal, and the ultrasonic receiving transducer receives the signal after it passes through the column cross-section, recording the propagation time from the emission time to the reception time. The unit is microseconds. Comparison with reference propagation time of intact timber under the same working conditions: During the test phase, reference propagation time was pre-calibrated for intact (uncracked) similar timber column samples. During calibration, the samples and field measurement points were strictly controlled to be of the same tree species, same column diameter grade, same moisture content range, and same temperature range. The reason for strictly controlling these four parameters is that the propagation speed of ultrasound in wood strongly depends on these four variables: tree species determines the cell wall structure and density gradient, thus determining the longitudinal sound velocity baseline; column diameter grade determines the propagation path length corresponding to different tissue layers from the bark to the core; moisture content significantly affects the attenuation coefficient and sound velocity of sound waves in the free water of the cell cavity; and temperature alters the mechanical relaxation properties of lignin. If these four parameters are not strictly matched, the divergence of the reference propagation time itself will be greater than the propagation time extension caused by cracks, rendering subsequent comparisons meaningless.

[0037] Define penetration depth index The ratio of the extended propagation time to the reference propagation time. ,in This represents the propagation time obtained by performing ultrasonic testing on the current crack segment. This represents the reference propagation time pre-calibrated on a healthy timber sample that matches the tree species, diameter grade, moisture content range, and temperature range of the timber column where the crack segment is located. and Units are consistent (microseconds). A ratio is used instead of an absolute propagation time extension. The considerations for this indicator are: the propagation path length of cracked wooden pillars varies significantly with different pillar diameters, and the relative path proportions corresponding to the same absolute elongation differ greatly under different pillar diameters; through analysis of... Normalization, make The dimensionless nature of the input features, which fall within a roughly comparable range, facilitates horizontal comparisons between cracked wooden pillars of different diameters and sizes. It also makes it easier for subsequent risk assessment models to normalize the input feature. The larger the value, the more intense the ultrasonic waves are circled or scattered by the crack at the cross section where the crack is located, and the closer the radial penetration depth of the crack is to the column diameter. A value close to 0 indicates that the propagation time is almost equivalent to that of intact wood, suggesting that the crack is only a shallow surface crack. In one typical implementation, the upper limit of the penetration depth index is limited to 3.0 to avoid abnormal values ​​calculated under extreme conditions where signal reflection is severe. In an alternative implementation, multiple sets of ultrasonic transmitting transducers and ultrasonic receiving transducers can be evenly spaced along the centerline of the crack segment (e.g., one set every 200 mm to 400 mm along the centerline). The propagation time of each set is measured, and the penetration depth index of each set is calculated according to the above definition. The arithmetic mean of the penetration depth indices of each set is then taken as the final penetration depth index of the crack segment to mitigate measurement errors caused by the randomness of the midpoint location.

[0038] Thus, for each crack segment, three types of geometric and penetration characteristic parameters were obtained: crack centerline orientation angle, physical crack width distribution sequence, and penetration depth index, providing input for subsequent crack propagation risk zone assessment.

[0039] When collecting the tilt angle of a column, a dual-axis tilt sensor can be installed at the top of the cracked wooden column, and another dual-axis tilt sensor can be installed at the base of the column. Each sensor has two mutually orthogonal measurement axes, denoted as the east-west axis and the north-south axis, respectively. Let the tilt angle read by the column top tilt sensor on the east-west axis be... The inclination angle read from the north-south axis is The tilt angle read by the column base tilt sensor on the east-west axis is The inclination angle read from the north-south axis is If the units are all degrees, then the angle of inclination of the column is defined. for ,in This represents the overall tilt angle of the cracked wooden column deviating from the vertical direction. The reason for using the difference between the readings at the top and bottom of the column, rather than just taking the readings at either end, is that the foundation of the antique-style building may experience long-term settlement or slight overturning of the platform. Taking only the reading at the top would include the platform's own tilt in the column's tilt, while the difference calculation can deduct the overall overturning component of the platform, retaining only the actual tilt degree of the column itself relative to the column base section. In one typical implementation, the tilt sensor is selected with a resolution of not less than 0.01 degrees and a range of... A MEMS digital tilt sensor with a sampling frequency of 1 Hz was used, and the arithmetic mean was taken after 60 seconds of continuous sampling. , , , This is to suppress transient disturbances introduced by on-site vibration and electrical noise. In an alternative implementation, where it is not possible to install sensors at the column base (typically when the column base is covered by a stone foundation or ground), only one biaxial tilt sensor can be installed at the top of the column, and the column base can be used as a vertical reference. However, it should be noted in the evaluation description that the foundation overturning is not deducted.

[0040] When collecting the eccentricity of the resultant pressure force at the column base, a thin-film pressure sensor array is embedded between the bottom surface of the column base and the contact surface of the stone foundation. This array consists of several pressure-sensitive units arranged in rows and columns, covering the entire circular area of ​​the column base bottom. Assume the array has a total of... Only pressure-sensitive unit, the first The center of the pressure-sensitive unit is located in the coordinate system of the column base. (The origin of the coordinate system is taken as the center of the circle at the base of the column) The axis points east. (Axis pointing north), the pressure value read by the pressure-sensitive unit at the time of acquisition is The unit is Newton, where Values ​​range from 1 to Coordinates of the point of application of the resultant force of the column base pressure. The calculation is performed by weighted averaging of the readings of each pressure-sensitive unit. , ,in , These represent the x-coordinate and y-coordinate of the point of application of the resultant force of the column base pressure in the coordinate system of the column base bottom surface, respectively. The eccentricity of the resultant force of the column base pressure. Defined as the distance from the point of application of the resultant force to the center of the circle at the base of the column. The unit is millimeters. This index reflects the degree of eccentricity of the vertical load on the column base section. The larger the eccentricity, the further the axial load at the column base deviates from the center of the section, and the greater the bending moment caused by eccentric compression inside the column, which significantly promotes the extension of cracks along the column axis. A scalar form, the distance from the point of application of the resultant force to the center of the circle, is used instead of a direct two-dimensional eccentricity vector. The reason for using this as input for subsequent risk assessment is that the wood column material has approximately isotropic properties with the column axis as the axis of symmetry. The influence of the eccentric direction on the risk of crack propagation can be attributed to the scalar parameter of the eccentric modulus, thereby controlling the input dimension of the subsequent risk assessment model and improving the sample utilization efficiency. In a typical implementation, the thin-film pressure sensor is selected as a resistive or capacitive pressure film with a range of 0 to 10 MPa and a resolution of not less than one-thousandth of the full scale. The spacing between the pressure-sensitive units is 10 mm to 20 mm, and the effective acquisition time for a single session is not less than 30 seconds. The arithmetic mean of multiple frames of readings within this time period is taken as the input. To suppress interference from short-term live loads (pedestrian movement, wind load fluctuations).

[0041] For each crack segment, three parameters—the angle of the crack centerline, the physical crack width distribution sequence, and the penetration depth index—are concatenated with two parameters—the column inclination angle and the eccentricity of the resultant pressure force at the column base—to form the feature description vector for that crack segment. Considering that the length of the physical crack width distribution sequence varies from crack segment to crack segment and cannot be directly fixed for modeling, the sequence is compressed into a fixed-length subvector through statistical summarization before being incorporated into the feature description vector. In one typical implementation, the statistical summary of the physical crack width distribution sequence consists of five items: the maximum value, the average value, the median, the standard deviation, and the coefficient of variation (i.e., the standard deviation divided by the average value). The coefficient of variation reflects the non-uniformity of the crack width distribution along its centerline—a larger coefficient of variation indicates significant local widening of the crack segment at certain locations, corresponding to the existence of stress concentration endpoints. Let... This represents the maximum value in the physical crack width distribution sequence. This represents the average value. This represents the median. Indicates standard deviation, This represents the coefficient of variation, with units of millimeters (except for...). (Dimensionless). The above five summary quantities are compared with the angle of the crack centerline. Penetration depth indicators Column tilt angle Column base pressure resultant force eccentricity The feature description vector of the crack segment is obtained by splicing. There are a total of 9 components. In an optional implementation, if the actual engineering is not sensitive to computational overhead and there are sufficient training samples, the physical crack width distribution sequence can be directly resampled to a uniform length (typically 64 or 128) as part of the feature description vector and input into the model to preserve the morphological information along the crack centerline; otherwise, if the number of training samples is limited, the above statistical summarization method is recommended to avoid overfitting caused by excessive feature dimension.

[0042] Feature description vector The crack propagation risk score for the crack segment is obtained by inputting the pre-trained risk assessment model. In one typical implementation, the risk assessment model employs a gradient boosting decision tree model. The model consists of several regression decision trees connected sequentially. Each decision tree uses the residual of the previous tree as the fitting target. The final risk score is obtained by summing the output values ​​of all leaf nodes of the decision trees and then performing a sigmoid transformation. The values ​​are mapped to the range of 0 to 1. The gradient boosting decision tree was chosen as the risk assessment model because: in this scenario, the components of the feature description vector have different dimensions (angle, millimeter, dimensionless ratio), and there is significant nonlinearity between them and the crack propagation risk, as well as interactions between components (for example, there is a cross-effect between the column tilt angle and the crack centerline direction angle—when the tilt direction is close to the crack direction, the eccentric bending caused by the tilt will significantly exacerbate the crack opening). The gradient boosting decision tree naturally handles the dimensional differences through nonlinear segmentation and automatically models the cross-effect, and has moderate requirements for training sample size and strong interpretability. The samples collected during model training are from past engineering archives: each training sample contains a feature description vector of an observed crack segment and the actual risk level given by manual verification. Let the number of training samples be... , No. The feature description vector of each training sample is denoted as . The corresponding actual risk level is recorded as ( The values ​​are set to preset discrete levels, such as low risk, medium risk, high risk, and extremely high risk (4 levels, mapped to 0.2, 0.4, 0.6, and 0.8 respectively). The model minimizes during the training phase. ,in The model represents the first The predicted risk score is output from each training sample. In one typical implementation, the number of decision trees is between 300 and 800, the maximum depth of each tree is between 4 and 6, the learning rate is between 0.05 and 0.1, and the above hyperparameters are determined using 5-fold cross-validation. In optional implementations, the risk assessment model can also use gradient boosting variants such as random forest, XGBoost, or LightGBM. For engineering data with more than 100,000 training samples, a shallow multilayer perceptron regression network can also be considered as the risk assessment model.

[0043] The risk assessment model only provides a point estimate of the crack propagation risk score. The inherent uncertainty of the rating cannot be directly represented. Considering that the liability risk level of restored ancient buildings is much higher than that of ordinary buildings, making the "most unfavorable estimate" of crack propagation risk is more in line with the engineering safety concept than the "most likely estimate". Therefore, a conformal calibration boundary value derived from the conformal prediction method is superimposed on the point estimate output of the risk assessment model, widening the point estimate to the higher risk side to an upper bound with uncertainty guarantee. The key to the conformal prediction method is that, without assuming that the prediction residuals of the risk assessment model follow any specific distribution (such as a normal distribution), it can provide a prediction interval coverage guarantee with a target confidence level in the long run, relying only on the empirical distribution of inconsistency scores observed on an independent calibration sample set.

[0044] refer to Figure 2 , Figure 2 The horizontal axis represents the crack segment number, sequentially numbered from 1 to 8, covering the 8 independent crack segments to be assessed. The vertical axis represents the crack propagation risk score, extending from 0.00 to 1.00, corresponding to the value range of the risk assessment model after sigmoid transformation. The vertical axis between 0.00 and 1.00 is divided by four consecutive and non-overlapping preset risk threshold intervals: 0.00 to 0.30 is the preset low-risk threshold interval, 0.30 to 0.55 is the preset medium-risk threshold interval, 0.55 to 0.80 is the preset high-risk threshold interval, and 0.80 to 1.00 is the preset extremely high-risk threshold interval. Each risk threshold interval is shown in the figure as a horizontally extending rectangular band. Each band is filled with distinct dense patterns to differentiate them. The band extends horizontally from near the starting point of the horizontal axis to the right end of the adjacent data column. On the right outer edge of each band, the text "low risk, medium risk, high risk, and very high risk" is centered and labeled respectively, avoiding any data points and error bars.

[0045] For each of the eight crack segments, a set of data elements is plotted along the horizontal axis corresponding to its number in the figure. This is based on the crack propagation risk score point estimated by the risk assessment model. Indicated by hollow dots, derived from the conformal upper bound value The points are represented by solid triangles, connected by a thick vertical line segment to represent the point estimates superimposed with conformal calibration boundary values. The resulting uncertainty interval is represented by two horizontal end-cap line segments drawn at each end of the vertical line segment, making the entire set of data primitives appear as asymmetric error bars with end caps. The point estimates corresponding to the eight crack segments are approximately monotonically ascending between 0.18 and 0.84, covering the complete working condition range from low risk to extremely high risk; the corresponding upper bound value is shifted upwards by a fixed step size based on each point estimate, and at the eighth crack segment... Exceeded 1.00, press The capping mechanism is set to 1.00. This capping process ensures that the upper bound value always falls within the 0 to 1 range designed for the risk score, preventing it from losing its interpretable meaning after exceeding the limit.

[0046] Figure 2 The core technological significance lies in the fact that the repair process for each crack segment is not based on the point estimate given by the risk assessment model. The risk threshold range it falls into is not determined by the upper bound derived from the conformal prediction method. The risk threshold range within which the crack falls is determined. In other words, the repair process matching for each crack segment is based on a safety decision made according to the "most unfavorable estimate" rather than the "most likely estimate." For example, the point estimation of the fourth crack segment. It is within the preset medium-risk threshold range, but due to the upper limit value After crossing the boundary between medium and high risk, the crack falls into a preset high-risk threshold range. The repair process for this crack segment will be upgraded from wood-embedded reinforcement to shape memory alloy smart hoop constraint technology. This upper-boundary-value-based process matching method transparently incorporates the safety margin resulting from "occasional model underestimation" into the decision-making process, ensuring the overall repair plan maintains a pre-set confidence level in the long term. Ensure that no high-risk segments are missed. Figure 2 The text labels, threshold interval names, and data elements are arranged in strict layers in the horizontal direction to avoid text obscuring data points or strip areas.

[0047] The pre-established independent calibration sample set is strictly disjoint from the training samples, comes from the same sample collection process as the training samples, and is subject to the same manual review process to determine the actual risk level. Let the total number of samples in the independent calibration sample set be... , No. One calibration sample ( The feature description vector of ) is denoted as The risk assessment model is used to assess The output predicted risk score is denoted as The actual risk level of the calibration sample, as determined by manual verification, is recorded as follows: Since the actual risk level is a discrete value while the predicted risk score is a continuous scalar, directly subtracting the two has no clear dimensional meaning. Therefore, each level value is first converted into a calibrated risk value with the same value range (i.e., 0 to 1) as the crack propagation risk score using a preset level mapping table. For example, low risk is mapped to 0.2, medium risk to 0.4, high risk to 0.6, and very high risk to 0.8. This grading mapping table is strictly consistent with the grading mapping method used in the training phase to ensure that prediction and calibration are based on the same scale.

[0048] Definition of the first The inconsistency score for each calibration sample is the absolute deviation between the predicted risk score and the corresponding calibrated risk value for that calibration sample. The reason for using absolute deviation as the inconsistency score is that the inconsistency score should simultaneously consider both directions of deviation: underestimating and overestimating the actual level. The absolute value form penalizes the deviations in both directions equally, preventing the conformal calibration boundary value from expanding only in one direction. Alternatively, if the engineering practice clearly focuses only on the direction of model underestimation (i.e., only compensating for the uncertainty of underestimation risk towards the high-risk side), one-sided deviation can be used. As an inconsistency score; if the sample sizes of each level are extremely unbalanced (the number of extremely high-risk samples is much smaller than that of low-risk samples), a weighting factor that is inversely frequency-weighted by level can be introduced into the inconsistency score.

[0049] Will The inconsistency scores of the calibration samples are sorted in ascending order to obtain an ordered sequence. Given a preset confidence level (In one typical implementation) (Set to 0.1, i.e., confidence level of 0.9), conformal calibration boundary value Take the permutation position index of the ordered sequence as The inconsistency scores, among which This represents the floor operation. It introduces the concepts of "add 1 and then floor" and "AND". The design of "taking the smaller value" is to cope with Finite-time finite-sample correction: when Exceed Time (typically occurs in) In cases where the value is small and the confidence level is close to 1, the directly corresponding ordered sequence number will go out of bounds, therefore... As an upper limit cap, the conformal calibration boundary value is within Even with finite time constraints, the algorithm still selects the maximum value from the ordered sequence, thus ensuring that it does not report errors on small calibration sets while still providing a conservative coverage guarantee. In one typical implementation, the total number of samples in the independent calibration sample set... Take between 300 and 500; too little will cause... Large variance and poor stability of results between different calibrations can lead to excessive calibration costs and limited marginal improvement.

[0050] Obtain conformal calibration boundary values Then, for each crack segment currently being assessed, first obtain the crack propagation risk score given by the risk assessment model. The upper bound of the risk prediction interval for the crack segment is then obtained by superimposing the crack propagation risk score with the conformal calibration boundary value. The cap of 1 is used because the risk score is designed to range from 0 to 1; values ​​outside this range are no longer interpretable. Finally, This represents the most unfavorable risk level for the propagation of the crack segment. The physical meaning of this value is: in the long term, at least... The probability ensures that the true risk value of the crack segment does not exceed [a certain value]. Therefore, use As a basis for decision-making, the safety margin provided by "occasional underestimation by the model" can be transparently incorporated into the decision. In an alternative implementation, if the liability risk for replica ancient buildings is more stringent, the confidence level can be further increased to 0.95 or 0.99, correspondingly... Taking 0.05 or 0.01 incurs a cost of... Enlarging and repair techniques are matched to a more conservative approach, which will increase the overall repair cost.

[0051] After obtaining the most unfavorable propagation risk level for all crack segments, a repair process is matched among four repair techniques based on the most unfavorable propagation risk level for each crack segment. Four preset risk threshold intervals are pre-defined, arranged consecutively from low to high and without overlap: a preset low-risk threshold interval, a preset medium-risk threshold interval, a preset high-risk threshold interval, and a preset extremely high-risk threshold interval. A closed-open convention is used at the adjacent boundaries of the four intervals (e.g., taking...). , , , Therefore, for any most unfavorable expansion risk level, the value can fall into one and only one interval. Cracks with the most unfavorable expansion risk level falling into the preset low-risk threshold interval are assigned to injection repair technology; those falling into the preset medium-risk threshold interval are assigned to wood inlay reinforcement technology; those falling into the preset high-risk threshold interval are assigned to shape memory alloy smart hoop constraint technology; and those falling into the preset extremely high-risk threshold interval are assigned to partial unloading renovation technology. The considerations behind this tiered matching are as follows: Low-risk crack segments mainly manifest as shallow, fine surface cracks with limited impact on load-bearing capacity, requiring only sealing and moisture protection, making injection repair the least costly option; Medium-risk crack segments have clear openings but limited penetration depth, requiring the addition of wood to restore local stiffness, and wood inlay reinforcement can restore mechanical continuity without significantly altering the column's appearance; High-risk crack segments have large penetration depths or are under eccentric compression conditions with high expansion potential energy, necessitating the use of actively applied circumferential constraints to hold the column together and prevent further crack opening, corresponding to shape memory alloy intelligent hoop constraint technology; Extremely high-risk crack segments already exhibit significant cross-sectional weakening, and passive constraints are insufficient to guarantee safety, requiring temporary unloading and removal and replacement of the failed area to fundamentally restore the column's load-bearing capacity, corresponding to partial unloading and renovation technology.

[0052] The specific implementation process of the injection repair technique is as follows: First, use a special crack cleaning knife and compressed air to remove sawdust, dust, and mold from the inside of the crack along the section to be repaired. If necessary, spray the inner wall of the crack with a wood preservative after cleaning. After the inner wall of the crack dries, use an injection nozzle to slowly inject low-viscosity wood repair adhesive (typically epoxy resin or polyurethane wood repair adhesive with a viscosity of less than 500 mPa·s at 25 degrees Celsius) from one end of the crack. During the injection process, observe the overflow from the other end to determine whether the adhesive has filled the crack properly. After the adhesive is injected, use wood putty of the same color or excess repair adhesive to seal and level the surface of the crack. After natural curing for 24 to 72 hours, sand it smooth with fine sandpaper. The low-viscosity repair adhesive can penetrate into the narrow gap at the end of the crack through capillary action, avoiding leaving cavities in the deep part of the crack that will become stress concentration points later.

[0053] The specific implementation process of the inlay wood reinforcement technique is as follows: A rectangular cross-section repair groove is cut along the direction of the crack to be repaired using a woodworking electric planer or grooving machine. The groove width is slightly larger than the maximum value of the physical crack width distribution sequence (typically 1.5 to 2 times the maximum value). The groove depth is 1.2 to 1.5 times the actual depth corresponding to the crack penetration depth index to ensure that the reinforcing wood strip can cover the actual penetration range of the original crack segment. Select high-quality wood strips that are of the same species as the wood column to be repaired, have matching moisture content (difference not exceeding 2%), and consistent grain direction. The reinforcing strips are processed to fit the repair groove with an interference fit (approximately 0.2 mm to 0.5 mm). Structural adhesive (typically E1 grade wood structural epoxy) is evenly applied to the sides of the reinforcing strip and the inner wall of the repair groove. The reinforcing strip is then embedded into the repair groove and pressed firmly using woodworking clamps perpendicular to the insertion direction. The adhesive layer is allowed to cure for at least 24 hours. Finally, the end face of the reinforcing strip is colored with the same wood wax oil or wood staining agent to ensure visual harmony between the reinforced area and the surrounding wood. The consideration of using the same tree species and moisture content is important because wood exhibits significant anisotropic characteristics in hygroscopic expansion and shrinkage. If the reinforcing strip and the original log are mismatched in their coefficients of expansion and contraction, seasonal humidity changes will repeatedly generate shear stress at the adhesive interface, ultimately leading to the reinforcing strip separating from the original log.

[0054] The specific implementation process of the partial unloading renovation process is as follows: First, temporary supports (typically using jacks in conjunction with I-beams) are set between adjacent structural components of the wooden column to be renovated, transferring the vertical load originally borne by the wooden column to the temporary supports, so that the wooden column to be renovated enters a near-unloaded state; then, a wood chisel, chainsaw, or band saw is used to remove the entire area along the boundary of the failed wood area, retaining healthy wood that is well bonded to the original wooden column during removal, and forming a transition slope with a reasonable gradient (typically 1:3 to 1:4) on the removed section to facilitate the mechanical force transmission of the replacement wood; healthy wood of the same tree species and moisture content as the original wooden column is selected as the replacement wood, and the replacement wood is processed according to the geometry of the removed section and made to fit tightly against the transition slope of the original wooden column, with structural adhesive applied to the joint surface and reinforced with hidden pins or tie rods to enhance the shear resistance of the joint surface; after the adhesive layer on the joint surface has cured, the temporary supports are gradually released, allowing the load to slowly return to the wooden column, and finally, the exposed surface of the replacement wood is repaired and restored with color and finish.

[0055] For crack segments falling within a preset high-risk threshold range, the following operations are performed using the shape memory alloy intelligent hoop constraint process. The shape memory alloy intelligent hoop constraint process utilizes the mechanical property that shape memory alloys, after being heated to the austenitic phase transformation termination temperature, will recover and shrink along their trained shape direction. A spiral hoop made of shape memory alloy wire is pre-laid tightly against the cylindrical surface in a martensitic phase state. Then, by applying electricity and heating, the martensite transforms into austenite, recovering and shrinking along the wire length direction. This shrinkage is converted into a circumferential hooping force on the cylindrical surface, actively constraining the crack to prevent further opening.

[0056] A heat-resistant flexible insulating layer is first applied to the outer surface of the column in the area of ​​the crack. This heat-resistant flexible insulating layer is typically made of silicone cloth or fiberglass cloth with a thickness of 0.3 mm to 0.6 mm, and has a continuous temperature resistance of not less than 200 degrees Celsius. The insulating layer physically separates the subsequent high-temperature shape memory alloy wires from the wood surface to prevent direct contact and burning of the wood, and also provides electrical insulation to prevent accidental short circuits to ground caused by heating current passing through the moist wood.

[0057] A nickel-titanium alloy wire is spirally wound along the circumference of a column on the outer surface of the heat-resistant flexible insulating layer as a shape memory alloy wire. In one typical embodiment, the diameter of the nickel-titanium alloy wire is 0.8 mm to 1.5 mm, with a nickel content of approximately 50.8 atomic percent and a titanium content of approximately 49.2 atomic percent. The austenitic transformation initiation temperature is between 55°C and 70°C, and the austenitic transformation termination temperature is between 70°C and 85°C. The wire is a straight segment with a shrinkage rate of 4% to 6% pre-set according to the trained shape. At room temperature, the nickel-titanium alloy wire is in the martensitic phase, where its deformation resistance is low. It can be spirally wound along the circumference of the column with a preset spiral spacing (typically 5 mm to 15 mm) under a constant initial tension applied by a tension control device (typically 5% to 10% of the yield strength of the nickel-titanium alloy wire, approximately 5 Newtons to 15 Newtons). The winding range covers the entire length of the crack segment along the column axis, and extends by one helical spacing at each end of the crack segment to ensure that the clamping effect is not limited to the middle of the crack segment and to prevent the ends from becoming new crack initiation points due to lack of constraint.

[0058] Several K-type thermocouples are attached at equal intervals along the spiral path on the outer surface of the wound nickel-titanium alloy wire as wire temperature monitoring points. The K-type thermocouples have a range of 0°C to 400°C and a response time of less than 1 second, enabling stable tracking of the transient temperature of the nickel-titanium alloy wire during the heating process. A thin-film temperature sensor is placed between the heat-resistant flexible insulating layer and the surface of the columnar wood as a wood surface temperature monitoring point. The thickness of the thin-film temperature sensor is typically less than 0.2 mm to avoid altering the bonding state of the insulating layer, and its range covers 0°C to 150°C. The consideration for simultaneously monitoring both the wire temperature and the wood surface temperature is that the wire temperature required to trigger a phase change (above 70°C) is close to the safe operating temperature boundary of the wood. Wire temperature monitoring alone cannot determine whether the wood itself has entered the damaged temperature zone; therefore, the wood surface temperature must be directly monitored by the thin-film temperature sensor below the insulating layer, and the wood surface temperature is used as the basis for deciding whether to reduce the heating current.

[0059] Heating is achieved by applying a direct current to both ends of the nickel-titanium alloy wire via a temperature controller, resulting in resistance heating (commonly known as Joule heating). The initial value of the heating current is typically 2 to 5 amperes, corresponding to a wire power density of 2 to 5 watts / cm². The temperature controller simultaneously acquires the wire temperature from each type K thermocouple and the wood surface temperature from each thin-film temperature sensor. Let the current monitored value of the wire temperature be... The current monitored value of the wood surface temperature is All units are in degrees Celsius. The preset austenitic phase transformation initiation temperature is... The preset austenitic phase transformation termination temperature is The preset upper limit of the wood protection temperature is (In one typical implementation) A temperature of 70 degrees Celsius is chosen, as this value balances the upper limit of the adhesive layer curing temperature and the threshold for wood thermal decomposition. Throughout the heating process, the temperature controller dynamically adjusts the heating current according to the following control rules: when... arrive Immediately reduce the heating current according to the preset step size to limit the temperature rise on the wood surface; when from Rise to During the process, the nickel-titanium alloy wire undergoes a martensitic-to-austenitic phase transformation, resulting in recovery shrinkage along its length. This recovery shrinkage causes the nickel-titanium alloy wire wound around the column to generate a circumferential clamping force on the outer surface of the isolation layer with each turn of the spiral, and this force is evenly transmitted to the surface of the wooden column through the isolation layer, thus restraining the continued opening of cracks. arrive After the heating current is cut off, the nickel-titanium alloy wire cools naturally to room temperature under constrained conditions. During the cooling process, because the austenitic phase transformation end temperature of the nickel-titanium alloy wire is higher than the ambient temperature, and its martensitic recovery deformation under constrained conditions is prevented by the geometric constraints of the column, the wire segment remains in a recovered constricted state under constrained conditions after cooling, continuously forming a circumferential constraint force on the surface of the column. In one typical embodiment, the clamping force of the recovered shape memory alloy wire along the surface of the column can reach 50 Newtons to 150 Newtons per wire, and the equivalent linear constraint force along the column axis, converted according to the helix pitch, can reach 5 kN to 30 kN per meter.

[0060] In optional implementations, for cases with large column diameters or long crack sections, a single nickel-titanium alloy wire can be replaced with two parallel nickel-titanium alloy wires wound simultaneously (electrically connected in parallel to the heating circuit) to increase the recovery shrinkage force per unit length; alternatively, the heat-resistant flexible insulating layer can be replaced with a composite insulating layer embedded with a graphene thermally conductive film to accelerate heat conduction between the wire and the closed-loop acquisition point of the temperature controller, thereby improving the temperature control response speed. After the shape memory alloy intelligent hoop constraint process is completed, a layer of wood wax oil coating or a matte topcoat of the same color as the original column surface can be applied to the outside of the nickel-titanium alloy wire to restore the visual appearance of the original antique building.

[0061] The partial unloading and renovation process for crack segments that fall within the preset extremely high risk threshold range has been described above.

[0062] After reinforcing all crack segments according to their respective matching repair techniques, the process proceeds to the post-renovation safety level output stage. Images of the reinforced wooden columns are re-acquired, and the physical crack width distribution sequence and penetration depth index of the remaining crack segments are extracted following the aforementioned image acquisition and crack feature segmentation extraction process. The column tilt angle and the eccentricity of the resultant pressure force at the column base are then acquired again using the same acquisition method as before reinforcement to ensure comparability.

[0063] Next, we calculate the effective load-bearing area ratio of the reinforced timber column cross-section. Let the cross-section of the reinforced timber column be its nominal radius. The approximately circular cross-section, its nominal cross-sectional area The unit is the square of millimeters. For each residual crack segment, based on its circumferential position on the column section (determined by the angle record during column surface image acquisition) and the physical crack width distribution sequence (its maximum value),... The maximum opening width of the segment in the cylindrical direction and the penetration depth index (its value) are represented by the following: (representing the relative penetration depth of the segment along the radial direction), approximating the weakened area of ​​the cross-section corresponding to the crack segment as having the cylindrical surface as the outer side and the penetration depth as the inner side. Extending radially into the interior of the column, corresponding to the circumferential span The sector-shaped or rectangular region is approximated as the circumferential span in one typical implementation. Radial depth The rectangular region is used to obtain the cross-sectional weakening area corresponding to the crack segment. subscript Pointing to the The remaining crack segment. Assume the remaining crack segment has a total of... The effective bearing area of ​​the reinforced wooden column cross-section is obtained by subtracting all weakened areas from the nominal cross-sectional area of ​​the wooden column. ,Will and Divide to obtain the effective bearing area ratio The value of this indicator ranges from 0 to 1. A value closer to 1 indicates a smaller degree of weakening of the wooden column cross-section by cracks, while a value closer to 0 indicates almost no effective load-bearing section and that reinforcement has not met expectations. The use of a percentage rather than the absolute value of the effective bearing area is chosen because horizontal comparisons are needed between cracked wooden columns of different diameters, and the percentage, as a dimensionless indicator, provides cross-sample comparability. In an optional implementation, when the circumferential positions of residual crack segments overlap on the column cross-section (typically multiple crack segments distributed along similar angles and with overlapping penetration depths), the union of the weakened cross-section areas corresponding to all residual crack segments is first taken to obtain a merged weakened cross-section area to avoid duplicate deductions. Then, the merged weakened cross-section area is used in the calculation of the effective bearing area.

[0064] The effective carrying area ratio , Reinforced column tilt angle 1. Eccentricity of the resultant force of the column base pressure after reinforcement and the number of residual crack segments These four indicators, arranged in a preset order, constitute the renovation safety evaluation vector. A preset safety threshold is set for each component of the renovation safety evaluation vector: The corresponding preset minimum effective carrying area ratio threshold (typically 0.85). The corresponding maximum column tilt angle threshold (typically 0.5 degrees). The corresponding preset maximum column base pressure resultant force eccentricity threshold (typically 5% of the column diameter). The corresponding preset threshold for the maximum number of residual crack segments (typically 2 segments). Each component in the renovation safety evaluation vector is compared with its corresponding preset safety threshold: when all four components meet their respective preset safety thresholds (i.e., ... Not lower than its threshold Not higher than its threshold Not higher than its threshold When the safety level of the refurbishment is not higher than its threshold, the output refurbishment safety level is qualified; when any component exceeds the corresponding preset safety threshold, the output refurbishment safety level is required for secondary reinforcement, and the component that exceeds the preset safety threshold is marked and output to guide the next round of reinforcement scheme to adjust the process type or process parameters in a targeted manner.

[0065] The system employs a multi-threshold criterion where all four components pass independently, requiring secondary reinforcement if any one fails. This differs from a composite criterion that compares a weighted sum of the four components to a single threshold. This is because exceeding the limit for any single indicator signifies that the corresponding potential failure mode (insufficient cross-sectional bearing capacity, overturning, increased eccentric compression, excessive residual cracks) is approaching or exceeding the engineering tolerance. Under a composite criterion, good performance in other indicators might mask the danger of a single indicator. The individual criterion, however, independently exposes each failure mode and clearly indicates the specific exceeding indicator in the non-compliance output, allowing on-site engineers to adjust subsequent repair plans more effectively. In an optional implementation, for cases where the renovation safety level output indicates the need for secondary reinforcement, the corresponding repair process can be automatically iterated based on the component marked as exceeding a preset safety threshold—for example… Exceeding the limit triggers the upgrade of the relevant cracked sections to wood inlay reinforcement or partial unloading and renovation processes. After the excessive triggering of the column base stone foundation is leveled and corrected, the crack diagnosis and reinforcement process is executed again, thus forming a closed-loop iterative renovation process until the renovation safety level output is qualified.

[0066] refer to Figure 3 , Figure 3The central outline of the cracked wooden pillar to be repaired is defined by two parallel vertical lines and ellipses at the top and bottom. The upper ellipse is fully visible, while the lower ellipse is drawn with dashed lines to represent the geometric perspective of the lower end facing the observer. Inside the outer outline, several thin, slanted lines indicate the internal grain of the wood extending along the pillar's axis, reflecting the directional characteristics of the wood fibers. A slightly wavy solid line segment extending along the pillar's axis in the middle section indicates the crack to be reinforced, spanning approximately five pillar diameters axially across the middle section of the pillar.

[0067] A heat-resistant flexible insulating layer is shown as a dashed rectangle enveloping the side view of the column in the region of the crack. This insulating layer covers the entire length of the crack in the axial direction and has a redundant extension at each end of the crack to prevent abrupt changes in the subsequent clamping force zone at the crack end and the formation of new stress concentration points. Nickel-titanium alloy wires are spirally wound around the outer surface of the insulating layer along the circumference of the column. The wire segments are shown as a continuous sine curve in the figure: the first half of the curve is drawn as a solid line to represent the spiral wire segments along the outside of the column facing the observer, and the second half of the curve is drawn as a dashed line to represent the spiral wire segments along the outside of the column facing away from the observer, thus fully reflecting the spiral direction and equal pitch distribution of the wire segments along the column surface.

[0068] Four K-type thermocouples are evenly spaced along the spiral wire segment as temperature monitoring points. Solid square marks are drawn only on the side facing the observer to avoid overlapping with the dashed lines of the wire segment facing away from the observer. Simultaneously, three thin-film temperature sensors are attached between the insulating layer and the wooden surface of the column as surface temperature monitoring points, indicated by hollow diamond marks on the outer right edge of the column. A lead wire connects to the upper right and lower left sides of the column, and the two wires converge at the dashed vertical segment drawn on the right side of the column to represent the closed heating circuit formed by the temperature controller and the DC heating power supply.

[0069] At four axially evenly distributed heights inside the column, a pair of horizontal arrows pointing towards the crack surface are drawn. The arrows point from the outer edge of the column to the crack surface, representing the circumferential clamping force acting on the wood surface of the column after being transmitted through the isolation layer. The mechanism of the circumferential clamping force is as follows: under electrical heating, the nickel-titanium alloy wire transforms from martensitic to austenitic phase and undergoes recovery shrinkage along its length. This recovery shrinkage causes the wire segment on each spiral turn to acquire a circumferential tension component pointing towards the column axis. This tension is then evenly transmitted to the column surface through the isolation layer, forming an active constraint on the continued opening of the crack. Figure 3 The lower right and lower left sides are marked with short, centered text to indicate the K-type thermocouple, thin-film temperature sensor, and the front solid line and rear dashed line, respectively. All markings are placed in the blank area outside the outline of the column to avoid any text overlapping with curves, arrows, or rectangular edges visually.

[0070] The preferred embodiments of this disclosure have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of this disclosure. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of this disclosure shall be within the scope of the claims of this disclosure.

Claims

1. A method for diagnosing cracks and performing composite reinforcement and renovation on cracked wooden columns in antique-style buildings, characterized in that, Includes the following steps: Step 1: Acquire images of the cracked wooden column surface to generate a rectangular unfolded map, which is then preprocessed by grayscale conversion and edge preservation denoising. Construct vertical 1D sequences and horizontal 1D sequences respectively and input them into the Mamba selective state space sequence segmentation model to output vertical crack probability maps and horizontal crack probability maps. Fuse the vertical crack probability maps and horizontal crack probability maps to obtain a fused crack segmentation mask map. Extract the crack centerline orientation angle and physical crack width distribution sequence of each crack segment from the fused crack segmentation mask map, and obtain the penetration depth index of each crack segment through ultrasonic detection. Step 2: Collect the column tilt angle and the eccentricity of the resultant pressure force at the column base. Concatenate the crack centerline orientation angle, physical crack width distribution sequence, penetration depth index, column tilt angle, and eccentricity of the resultant pressure force at the column base of each crack segment to form a feature description vector and input it into the pre-trained risk assessment model to obtain a crack propagation risk score. Use the pre-established independent calibration sample set to determine the conformal calibration boundary value using the conformal prediction method. Superimpose the conformal calibration boundary value on the crack propagation risk score to obtain the most unfavorable propagation risk level of each crack segment. Step 3: Based on the most unfavorable propagation risk level of each crack segment, match the repair process for each crack segment among the injection repair process, wood inlay reinforcement process, shape memory alloy intelligent hoop constraint process, and partial unloading renovation process; the shape memory alloy intelligent hoop constraint process triggers the phase transformation recovery shrinkage of the shape memory alloy wire wound on the outer surface of the column by heating, forming a circumferential hoop force to constrain the crack; after the reinforcement operation is completed, extract the physical crack width distribution sequence and penetration depth index of the residual crack segment of the reinforced wooden column according to the process of Step 1, combine the column tilt angle and the eccentricity of the resultant force of the column foot pressure to form a renovation safety evaluation vector and output the renovation safety level.

2. The method according to claim 1, characterized in that, In step 1, the edge preservation and noise reduction preprocessing uses anisotropic diffusion filtering. Anisotropic diffusion filtering preserves crack edges with a low diffusion rate in areas with large gray-level gradients, and smooths wood grain textures and areas with uneven lighting with a high diffusion rate in areas with small gray-level gradients.

3. The method according to claim 1, characterized in that, In step 1, the vertical 1D sequence is formed by extracting gray values ​​column by column along the column axis and arranging them in the order from the top of the column to the bottom of the column, while the horizontal 1D sequence is formed by extracting gray values ​​row by row along the circumferential direction and arranging them in the order of circumferential unfolding.

4. The method according to claim 1, characterized in that, In step 1, the Mamba selective state space sequence segmentation model includes a first Mamba selective state space sequence segmentation model and a second Mamba selective state space sequence segmentation model. The first Mamba selective state space sequence segmentation model inputs a vertical 1D sequence and outputs a vertical crack probability map; the second Mamba selective state space sequence segmentation model inputs a horizontal 1D sequence and outputs a horizontal crack probability map. Both the first and second Mamba selective state space sequence segmentation models use the same network structure and are pre-trained using labeled crack masks with pixel-level crack category labels as supervision signals. Each Mamba selective state space sequence segmentation model... The Ambam selective state-space sequence segmentation model internally maintains a fixed-dimensional hidden state vector. Elements in the sequence are input one by one. At each time step, the current input element is mapped to a selectively gated vector through the first linear projection layer. The components of the selectively gated vector take values ​​between 0 and 1 and are used to scale the retention ratio of each dimension information passed from the previous time step in the hidden state vector element by element. At the same time, the current input element is mapped to an input driving vector through the second linear projection layer. The input driving vector is added element by element to the selectively gated and scaled hidden state vector to obtain the updated hidden state vector for the current time step. The updated hidden state vector is then mapped to the crack probability value for the current time step through the output linear layer.

5. The method according to claim 1, characterized in that, In step 1, the process of fusing the longitudinal crack probability map and the transverse crack probability map to obtain the fused crack segmentation mask map is as follows: the larger value between the longitudinal crack probability map and the transverse crack probability map is taken pixel by pixel to obtain the fused crack probability map; the fused crack probability map is binarized according to a preset segmentation threshold to obtain a binarized crack mask map; isolated small connected component deletion and skeleton continuity filtering are performed on the binarized crack mask map to obtain the fused crack segmentation mask map; the process of extracting the crack centerline orientation angle and physical crack width distribution sequence of each crack segment from the fused crack segmentation mask map is as follows: connected component labeling and Zhang-Suen skeleton thinning are performed on the fused crack segmentation mask map to extract the crack centerline of each crack segment; the pixel span of the fused crack segmentation mask map in the direction perpendicular to the crack centerline is measured point by point along the crack centerline and converted into physical crack width according to the cylindrical unfolding scale; the physical crack width distribution sequence is composed of all the physical crack widths of each crack segment.

6. The method according to claim 1, characterized in that, In step 1, the process of obtaining the penetration depth index of each crack segment by ultrasonic detection is as follows: ultrasonic transmitting transducers and ultrasonic receiving transducers are arranged opposite each other along the radial direction of the column at the midpoint of each crack segment. The propagation time of the ultrasonic pulse through the cross section of the column is measured. The propagation time is compared with the reference propagation time of intact wood calibrated under the same tree species, the same column diameter grade, the same moisture content range, and the same temperature range. The ratio of the extension of propagation time to the reference propagation time of intact wood is used as the penetration depth index.

7. The method according to claim 1, characterized in that, In step 2, the physical crack width distribution sequence in the feature description vector is spliced ​​with the maximum value of the physical crack width distribution sequence; the risk assessment model is a gradient boosting decision tree model; the process of determining the conformal calibration boundary value using the conformal prediction method is as follows: each calibration sample in the independent calibration sample set has a predicted risk score output by the same gradient boosting decision tree model and an actual risk level determined by manual review. The actual risk level is converted into a calibrated risk value with the same value range as the crack propagation risk score through a level mapping table. The absolute deviation between the predicted risk score and the corresponding calibrated risk value of each calibration sample is taken as the inconsistency score of the calibration sample. All calibration samples are sorted from smallest to largest inconsistency scores. The smaller of the inconsistency score obtained by taking the position number equal to the total number of calibration samples plus 1, multiplied by the preset confidence level and rounded up, and the total number of calibration samples is taken as the conformal calibration boundary value. The upper limit of the risk prediction interval is obtained by superimposing the conformal calibration boundary value on the crack propagation risk score. The upper limit of the risk prediction interval is taken as the most unfavorable propagation risk level of each crack segment.

8. The method according to claim 1, characterized in that, In step 3, the process of matching the repair process according to the most unfavorable expansion risk level of each crack segment is as follows: the crack segment with the most unfavorable expansion risk level falling into the preset low risk threshold range is assigned the grouting repair process; the crack segment with the most unfavorable expansion risk level falling into the preset medium risk threshold range is assigned the wood inlay reinforcement process; the crack segment with the most unfavorable expansion risk level falling into the preset high risk threshold range is assigned the shape memory alloy smart hoop constraint process; and the crack segment with the most unfavorable expansion risk level falling into the preset extremely high risk threshold range is assigned the partial unloading renovation process. The preset low-risk threshold range, preset medium-risk threshold range, preset high-risk threshold range and preset extremely high-risk threshold range are arranged consecutively from low to high according to risk value and do not overlap with each other; The injection repair process includes cleaning the crack, injecting low-viscosity wood repair adhesive into the crack, and sealing the crack surface; The wood inlay reinforcement process includes creating a repair groove along the crack direction, embedding reinforcing wood strips of the same material, gluing and pressing them together, and surface coloring; the partial unloading renovation process includes setting up temporary supports to transfer the column load, removing the failed wood areas, replacing wood and restoring the finish.

9. The method according to claim 1, characterized in that, In step 3, the execution process of the shape memory alloy smart hoop constraint technology is as follows: a heat-resistant flexible insulating layer is wrapped around the outer surface of the column in the crack segment area. A nickel-titanium alloy wire is spirally wound around the outer surface of the heat-resistant flexible insulating layer along the circumference of the column as a shape memory alloy wire. The nickel-titanium alloy wire is in the martensitic phase at room temperature. During winding, a constant initial tension force is applied through a tension control device. The winding range covers the entire length of the crack segment along the column axis and extends by one spiral spacing at each end. K-type thermocouples are attached at equal intervals along the spiral path on the outer surface of the nickel-titanium alloy wire as wire temperature monitoring. A thin-film temperature sensor is placed between the heat-resistant flexible insulating layer and the surface of the column wood as a monitoring point for the wood surface temperature. Resistance heating is performed by applying DC current to both ends of the nickel-titanium alloy wire. The temperature controller simultaneously collects the wire temperature and the wood surface temperature. When the wood surface temperature reaches the preset upper limit of the wood protection temperature, the heating current is reduced to limit the temperature rise of the wood surface. When the wire temperature reaches the austenite phase transformation termination temperature, the heating current is cut off. After the nickel-titanium alloy wire cools naturally to room temperature under the constrained state, it maintains the recovered state of the hoop and forms a continuous circumferential constraint force on the column surface.

10. The method according to claim 1, characterized in that, In step 3, the renovation safety evaluation vector consists of the effective bearing area ratio, column tilt angle, column base pressure resultant eccentricity, and the number of residual crack segments. The effective bearing area ratio is determined by: determining the cross-sectional weakening area corresponding to each residual crack segment based on the circumferential position of all residual crack segments on the column cross-section, the physical crack width distribution sequence, and the penetration depth index; and subtracting all cross-sectional weakening areas from the nominal cross-sectional area of ​​the wooden column to obtain the effective bearing area ratio. The renovation safety level is output by: comparing each component in the renovation safety evaluation vector with its corresponding preset safety threshold; outputting a qualified renovation safety level when all components meet the corresponding preset safety threshold, and outputting a renovation safety level requiring secondary reinforcement when any one component exceeds the corresponding preset safety threshold.

Images