Strain image-based pile body damage identification method and system

Abstract

The invention provides a strain image-based pile body damage identification method and a strain image-based pile body damage identification system, which are characterized in that step-by-step progressive links such as space-time adaptive gradient normalization, dynamic suppression weight-guided inter-frame registration, multi-scale form-driven regional aggregation and strain transfer rule-based direction verification are constructed; enabling the weak abnormal response to obtain scale unified expression in a gradient domain, achieving phase alignment in a time dimension, and completing structure condensation and geometric filtering in a space dimension; the finally generated pile body damage confidence coefficient field and depth affiliation sequence can stably represent the spatial position, the extension orientation and the functional depth section of the real damage, so that the weak localization strain abnormal area is accurately reserved under the scene of low signal-to-noise ratio and strong interference, the false alarm caused by global disturbance is avoided, and the reliability of the system is improved. And early-stage, shallow-layer or low-amplitude damage response is not missed, and robust extraction of the weak localization strain abnormal region is realized.

Description

Technical Field

[0001] This invention relates to the field of pipe pile monitoring technology, specifically to a method and system for identifying pile damage based on strain images. Background Technology

[0002] In pile foundation health monitoring, distributed strain sensing technology is widely used to acquire continuous strain responses along the pile length and organize them into a raw strain image sequence. Existing methods typically perform global thresholding, inter-frame differencing, or principal component analysis directly on the raw strain image sequence to identify abnormal regions. These methods rely on significant jumps in strain amplitude as damage criteria and generally employ fixed thresholds, global statistical models, or preset template matching mechanisms. When the pile is under low load, damage is in its early initiation stage, or the monitoring environment is affected by strong vibration interference and temperature drift, the strain response caused by the actual damage often exhibits weak (amplitude less than 0.3 times the overall standard deviation of the image), localized (spatial span only a few pixels), and non-steady-state (only appearing in some loaded frames) structural characteristics, which are severely masked by high-frequency speckle noise, sensor coupling jitter, and slow trend terms.

[0003] Under conditions of low signal-to-noise ratio and strong noise interference, existing methods are unable to stably separate weak and localized abnormal strain regions of the pile body from the original strain image sequence. They are prone to misjudging global disturbances as damage, or completely missing the real defects due to insufficient response amplitude. Summary of the Invention

[0004] This invention aims to provide a method and system for pile damage identification based on strain images, which enables robust extraction of weak localized strain anomaly regions.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for identifying pile damage based on strain images, comprising: Spatiotemporal adaptive gradient normalization is performed on the original strain image sequence to generate a gradient intensity field sequence. A spatiotemporal consistency mask is constructed based on the gradient intensity field sequence to generate a dynamic suppression weight field sequence. Structure-guided inter-frame registration is performed on the dynamic suppression weight field sequence along the pile axis to generate a calibrated intensity field sequence. The calibrated intensity field sequence is then aggregated in the reference frame coordinate system to generate a pile damage confidence field. Multi-scale morphological closure and connected domain separation are performed on the pile damage confidence field to generate a preliminary set of damage regions. Based on the pile axis geometric constraints and strain transmission law, the directional consistency check is performed on the initial set of damage areas to generate a set of effective damage areas. Within each effective damage area, a strain anomaly intensity profile is reconstructed to generate a pile damage depth attribution sequence. Based on the pile damage depth attribution sequence and pile design parameters, a pile structural integrity status label is generated.

[0006] Preferably, performing spatiotemporal adaptive gradient normalization on the original strain image sequence includes: A square neighborhood is extracted centered on each pixel, and the horizontal and vertical gradient components within the neighborhood are calculated. Calculate the median of the absolute values ​​of the gradient components in the neighborhood to obtain the horizontal median and the vertical median. The original gradient components are normalized using the median in the horizontal direction and the median in the vertical direction. The intensity values ​​of the normalized gradient components are calculated to generate a gradient intensity field sequence.

[0007] Preferably, the construction of the spatiotemporal consistency mask includes: Extract the gradient intensity time series for each pixel location and perform sliding window mid-value filtering to obtain a robust background intensity estimate; Calculate the relative deviation between the original intensity and the robust background intensity estimate, statistically analyze the upper quartile of the relative deviation sequence, and set a spatiotemporal consistency criterion; A binary mask is generated based on the spatiotemporal consistency criterion, forming a spatiotemporal consistency mask sequence.

[0008] Preferably, the implementation of structure-guided inter-frame registration includes: Extract the set of pixel coordinates with a value of 1 from the mask sequence to form a candidate point cloud; The first frame is selected as the reference frame. The nearest matching point is searched in the candidate point cloud of other frames. The displacement vector is recorded. A piecewise linear displacement function along the pile axis is fitted based on the matching point. Vertical resampling is performed only at candidate point cloud locations for the gradient intensity field to generate a calibrated intensity field sequence.

[0009] Preferably, the generation of the pile damage confidence field includes: Count the frequency at which each pixel position is assigned a value across all frames; Calculate the upper quartile of the defined intensity value for each pixel location, and calculate the average frequency of pixels in the 8-neighborhood to obtain the spatial connectivity weight; By combining the strength upper quartile, frequency and spatial connectivity weights, a pile damage confidence field is generated.

[0010] Preferably, the multi-scale morphological closure and connected component separation includes: Adaptive dual-threshold binarization is applied to the pile damage confidence field to distinguish between strong and weak response points; Perform small-size closing operations on strong response point sets and large-size opening operations on weak response point sets; After merging the strong and weak response regions, connective domain labeling is performed to generate a preliminary set of damaged regions.

[0011] Preferably, the consistency check of the execution direction includes: For each initially screened damage area, fit the principal inertial axis, calculate the cosine value of the angle between the principal direction and the pile axis direction, and retain the area where the cosine value of the angle is greater than the preset threshold. The confidence sequence is extracted along the main direction of the retained region, the proportion of negative values ​​in the first-order difference sequence is calculated, and the region with a negative value proportion less than the preset proportion is retained to generate a set of effective damage regions.

[0012] Preferably, the reconstructed strain anomaly intensity profile includes: The mean time series value of each pixel in the effective damage area in the original strain image sequence is extracted to obtain the average strain profile. Interpolate the mean strain profile and calculate the second derivative; The lower boundary of the damage is determined by searching for inflection points with significantly downward curvature, and the upper boundary of the damage is determined by searching for inflection points with significantly upward curvature. The midpoint of the calculated depth interval is used as the representative depth of the damage to generate a sequence of pile damage depths.

[0013] Preferably, the generation of the pile structure integrity status label includes: Based on the pile design parameters, the pile length is divided into the pile top area, the main body area, and the pile end area; Calculate the mean value of each effective damage area in the confidence field, and determine the damage level by looking up the table based on the functional area category and the mean value; A three-level structural integrity status label is generated by combining all damage levels.

[0014] On the other hand, the present invention proposes a pile damage identification system based on strain images, comprising: The gradient normalization module is used to perform spatiotemporal adaptive gradient normalization on the original strain image sequence to generate a gradient intensity field sequence. The mask construction module is used to construct a spatiotemporally consistent mask based on the gradient intensity field sequence and generate a dynamic suppression weight field sequence. The inter-frame registration module is used to perform structure-guided inter-frame registration on the dynamic suppression weight field sequence along the pile axis direction to generate a calibrated intensity field sequence. The confidence aggregation module is used to aggregate the calibrated intensity field sequence in the reference frame coordinate system to generate a pile damage confidence field. The region separation module is used to perform multi-scale morphological closure and connected domain separation on the pile damage confidence field to generate a preliminary set of damage regions. The orientation verification module is used to verify the orientation consistency of the initial set of damaged areas based on the pile axis geometric constraints and strain transmission laws, and to generate a set of valid damaged areas. The depth attribution module is used to reconstruct the strain anomaly intensity profile within each effective damage area and generate a pile damage depth attribution sequence. The status label module is used to generate status labels for the structural integrity of the pile body based on the pile damage depth attribution sequence and pile design parameters.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a progressively layered approach, including spatiotemporal adaptive gradient normalization, dynamically suppressed weight-guided inter-frame registration, multi-scale morphology-driven region aggregation, and direction verification based on strain propagation laws. This enables weak anomaly responses to achieve a scale-uniform representation in the gradient domain, phase alignment in the temporal dimension, and structural cohesion and geometric filtering in the spatial dimension. The resulting pile damage confidence field and depth attribution sequence stably characterize the spatial location, extension orientation, and functional depth segment of the true damage. This allows for accurate preservation of weak, localized strain anomaly regions even in low signal-to-noise ratio and high-interference scenarios, avoiding false alarms caused by global disturbances and preventing the omission of early, shallow, or low-amplitude damage responses. This robust extraction of weak, localized strain anomaly regions is achieved. Attached Figure Description

[0016] Figure 1 This is a flowchart of the pile damage identification method based on strain images according to the present invention; Figure 2 This is a block diagram of the pile damage identification system based on strain images according to the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] like Figure 1 As shown, this invention proposes a pile damage identification method based on strain images. Its core lies in constructing an end-to-end signal-structure-semantic mapping path. Using the spatiotemporal-frequency joint characteristics of the image sequence as anchors, it progressively suppresses unstructured disturbances, enhances deformation geometric consistency, calibrates multi-frame correlation deviations, and ultimately completes pixel-level attribution determination. Specifically, it includes the following steps: Spatiotemporal adaptive gradient normalization is performed on the original strain image sequence to generate a gradient intensity field sequence. Specifically, this includes: extracting a square neighborhood centered on each pixel and calculating the horizontal and vertical gradient components within the neighborhood; calculating the median of the absolute values ​​of the gradient components within the neighborhood to obtain the horizontal and vertical medians; normalizing the original gradient components using the horizontal and vertical medians; calculating the intensity value of the normalized gradient components to generate the gradient intensity field sequence.

[0019] This enables the gradient responses at different spatial locations to be comparable, eliminates the scale bias introduced by material constitutive differences and uneven sensing gain, thereby highlighting local deformation geometric abrupt changes and providing a scale-uniform basic expression for subsequent identification of weak strain anomalies.

[0020] Based on the gradient intensity field sequence, a spatiotemporal consistency mask is constructed, generating a dynamic suppression weight field sequence. Specifically, this includes: extracting the gradient intensity time series for each pixel location and performing sliding window mid-range filtering to obtain a robust background intensity estimate; calculating the relative deviation between the original intensity and the robust background intensity estimate, statistically analyzing the upper quartile of the relative deviation sequence, and setting a spatiotemporal consistency criterion; generating a binary mask based on the spatiotemporal consistency criterion to form a spatiotemporal consistency mask sequence.

[0021] It can effectively suppress non-uniform noise, enhance the spatiotemporal coherence of the real gradient signal, provide a reliable weight field sequence for accurately identifying and tracking subtle changes in the dynamic process, and improve the accuracy and robustness of subsequent analysis.

[0022] The dynamic suppression weight field sequence is subjected to structure-guided inter-frame registration along the pile axis to generate a calibrated intensity field sequence. Specifically, this includes: extracting the set of pixel coordinates with a value of 1 from the mask sequence to form a candidate point cloud; selecting the first frame as the reference frame, searching for the nearest matching point in the candidate point cloud of other frames, recording the displacement vector, fitting a piecewise linear displacement function along the pile axis based on the matching point; and performing longitudinal resampling on the gradient intensity field only at the candidate point cloud position to generate a calibrated intensity field sequence.

[0023] Achieving precise alignment of abnormal responses along the pile axis direction allows weak signals of the same physical damage to be superimposed in phase across multiple frames, significantly enhancing their detectability in the time dimension and avoiding response cancellation caused by inter-frame misalignment.

[0024] The calibrated intensity field sequence is aggregated in the reference frame coordinate system to generate a pile damage confidence field. Specifically, this includes: counting the frequency of each pixel position being assigned a value in all frames; calculating the upper quartile of the defined intensity value for each pixel position and calculating the average frequency of pixels in the 8-neighborhood to obtain the spatial connectivity weight; and combining the upper quartile of intensity, frequency and spatial connectivity weight to generate the pile damage confidence field.

[0025] This aggregation method takes into account time repeatability, strength robustness, and spatial extensibility, enabling the real damage response to form a stable, continuous, and significantly higher high-value region in the confidence field than the background. This effectively suppresses isolated noise points and pseudo-anomalies, improving the reliability of damage localization and structural interpretability.

[0026] The pile damage confidence field is subjected to multi-scale morphological closure and connected component separation to generate a preliminary set of damage regions. Specifically, this includes: performing adaptive dual-threshold binarization on the pile damage confidence field to distinguish between strong and weak response points; performing a small-size closure operation on the set of strong response points and a large-size opening operation on the set of weak response points; merging the processed strong and weak response regions, marking the connected components, and generating a preliminary set of damage regions.

[0027] While preserving the connectivity of the actual damage space, discrete noise points and fracture artifacts are eliminated, generating a preliminary screening area with clear boundaries, complete structure, and conforming to the geometric characteristics of pile damage, providing a reliable geometric carrier for subsequent physical consistency verification.

[0028] Based on the geometric constraints of the pile axis and the strain transmission law, the direction consistency check is performed on the initial screening damage area set to generate an effective damage area set; specifically, this includes: fitting the principal inertial axis to each initial screening damage area, calculating the cosine value of the angle between the principal direction and the pile axis direction, and retaining areas where the cosine value of the angle is greater than a preset threshold; extracting the confidence sequence along the principal direction for the retained areas, calculating the proportion of negative values ​​in the first-order difference sequence, and retaining areas where the proportion of negative values ​​is less than a preset proportion to generate an effective damage area set.

[0029] This verification process eliminates false areas that deviate from the pile axis and have chaotic internal responses, ensuring that the retained effective damage areas conform to the physical laws of pile stress in terms of spatial extension and strain change trends, thus significantly improving the engineering reliability of the identification results.

[0030] Within each effective damage region, a strain anomaly intensity profile is reconstructed to generate a pile damage depth attribution sequence. Specifically, this includes: extracting the time series mean of each pixel within the effective damage region from the original strain image sequence to obtain an average strain profile; interpolating the average strain profile and calculating its second derivative; searching for inflection points with significantly downward curvature to determine the lower boundary of the damage, and searching for inflection points with significantly upward curvature to determine the upper boundary of the damage; calculating the midpoint of the depth interval as the representative depth of the damage, and generating a pile damage depth attribution sequence.

[0031] By generating damage depth attribution sequences, this method provides engineers with clear and intuitive damage location information, helping to improve the targeting and efficiency of maintenance work. Furthermore, this method can effectively distinguish multiple damage areas, avoiding misjudgments caused by confusing different damage types.

[0032] Based on the pile damage depth attribution sequence and pile design parameters, a pile structural integrity status label is generated; specifically, this includes: dividing the pile length into the pile top area, the main pile body area, and the pile end area according to the pile design parameters; calculating the mean value of each effective damage area in the confidence field; determining the damage level based on the functional area category and the mean value; and generating a three-level structural integrity status label by combining all damage levels.

[0033] By mapping the location and severity of damage to the functional zones of the pile body, a hierarchical status label that can directly support engineering decisions is output, which upgrades the identification results from pixel-level response to structural-level identification, significantly improving the practicality and operability of monitoring conclusions.

[0034] On the other hand, this invention proposes a pile damage identification system based on strain images, such as... Figure 2 As shown, it includes: The gradient normalization module is used to perform spatiotemporal adaptive gradient normalization on the original strain image sequence to generate a gradient intensity field sequence. The mask construction module is used to construct a spatiotemporally consistent mask based on the gradient intensity field sequence and generate a dynamic suppression weight field sequence. The inter-frame registration module is used to perform structure-guided inter-frame registration on the dynamic suppression weight field sequence along the pile axis direction to generate a calibrated intensity field sequence. The confidence aggregation module is used to aggregate the calibrated intensity field sequence in the reference frame coordinate system to generate a pile damage confidence field. The region separation module is used to perform multi-scale morphological closure and connected domain separation on the pile damage confidence field to generate a preliminary set of damage regions. The orientation verification module is used to verify the orientation consistency of the initial set of damaged areas based on the pile axis geometric constraints and strain transmission laws, and to generate a set of valid damaged areas. The depth attribution module is used to reconstruct the strain anomaly intensity profile within each effective damage area and generate a pile damage depth attribution sequence. The status label module is used to generate status labels for the structural integrity of the pile body based on the pile damage depth attribution sequence and pile design parameters.

[0035] Furthermore, the modules in the above system also perform other steps to implement the above-mentioned method for pile damage identification based on strain images, as follows: Step 1: Perform spatiotemporal adaptive gradient normalization on the original strain image sequence to generate a gradient intensity field sequence. The original strain image sequence is denoted as ,in Indicates the first Frame acquisition time, For image plane coordinates, Characterize the location at The axial strain value measured at each moment is expressed in microstrain. Due to the different confining pressure conditions of pile segments at different depths, the strain amplitude varies by order of magnitude under the same load. Simultaneously, the inherent gain non-uniformity of fiber optic gratings or distributed strain sensing devices leads to shifts in response sensitivity between adjacent pixels, causing the same physical strain gradient in the original image to exhibit different pixel gradient magnitudes in different regions. If the spatial gradient is directly calculated, strong gradient regions can easily mask the true trend of weak gradient regions, causing subsequent anomaly localization to be biased towards high-amplitude areas while overlooking shallow microcracks or interface delamination—low-gradient but highly hazardous defects. Therefore, it is necessary to first establish a local gradient scale recalibration mechanism that does not rely on global statistical assumptions, ensuring that the gradient response in the neighborhood of each pixel matches the intrinsic strain transfer capability of its location.

[0036] Step 1.1: For each pixel Centered on, cut off at a size of square neighborhood ,in A value of 3 corresponds to a 7×7 pixel window; the strain value of all pixels within that neighborhood. Construct a second-order central difference matrix to obtain the horizontal gradient components. gradient components in the vertical direction This operation occurs in every frame. The gradient calculations are performed independently to ensure that the gradient calculations do not overlap across time dimensions.

[0037] Step 1.2: For the same neighborhood All and Calculate the median of the absolute values ​​of each value, denoted as . and The median reflects the typical intensity level of the gradient response in the local region and is not sensitive to isolated spikes in the neighborhood, thus effectively resisting single-pixel impulse noise interference. Since the real structural gradient in the strain image is usually clustered, while the noise gradient is random and discrete, the median is better at characterizing the local dominant gradient scale than the mean.

[0038] Step 1.3: Perform pixel-by-pixel division normalization on the original gradient components using the median mentioned above, and define the normalized gradient components as follows: ,

[0039] In the formula To prevent zero values ​​and ensure that the denominator is always positive, this normalization operation makes the gradient values ​​at different spatial locations within the same frame comparable. For example, the concrete section at the top of the pile has a large gradient modulus due to its high elastic modulus, while the soft soil contact section at the bottom of the pile has a small gradient modulus due to its low modulus. After this processing, if the two undergo the same degree of strain distortion in the normalization domain, their normalized gradient responses will be nearly identical.

[0040] Step 1.4: For each frame All pixels Calculate its normalized gradient strength

[0041] This intensity value That is, the first in the gradient intensity field sequence Frame image; since the normalization process has been completed in step 1.3, no additional scaling is introduced here; The range of values ​​is Its peak no longer corresponds to the location of the maximum strain rate in the original image, but rather to the region of most significant orientation change relative to the local gradient background; For example, if a strain band that was originally changing smoothly suddenly shows an inflection point, then its The value will jump, even if the absolute value of the original gradient is still small at that inflection point; therefore, This becomes the first-layer feature map characterizing the geometric abrupt changes in local deformation, providing a scale-uniform basic expression for subsequent identification of weak anomalies under strong noise.

[0042] Step 2: Construct a spatiotemporal consistent mask based on the gradient intensity field sequence to generate a dynamically suppressed weight field sequence. The gradient intensity field sequence output in step one Spatial gradient scale bias has been eliminated, but the problem of non-stationary disturbances in the temporal dimension remains unresolved. In actual monitoring, the pile body is affected by vehicle vibration, wind-induced swaying, or precipitation infiltration, which can cause a full-map gradient intensity increase within a range of several frames. This increase is unrelated to damage, but it can overshadow the weak gradient enhancement truly induced by local damage. Furthermore, the slow trend term caused by sensor temperature drift will... The continuous unidirectional shift over dozens of frames creates a pseudo-anomaly band. If directly... Thresholding will inevitably introduce a large number of false alarms. Therefore, it is necessary to construct a discrimination criterion that can distinguish between intensity fluctuations driven by global disturbances and intensity fixation driven by local damage, and generate dynamic weights per pixel and per frame based on this criterion to suppress the former and retain the latter.

[0043] Step 2.1: For each pixel position Extract it from all Gradient intensity time series on frames Perform sliding window mid-range filtering on the sequence, with a window width set to... Frame, that is, for each calculate

[0044] This operation preserves the dominant level that persists throughout the time series while removing spikes in isolated frames; since once an impairment occurs, its corresponding gradient anomaly will stably reproduce over multiple consecutive frames (at least 3 frames), while transient interference typically lasts only 1–2 frames, therefore… This can be considered a robust estimate of the background intensity at that pixel location.

[0045] Step 2.2: For the same pixel Calculate its original strength With a stable background relative deviation

[0046] In the formula To prevent small amounts from being zero; this deviation degree It represents the degree of fluctuation in the gradient intensity of a pixel in the current frame relative to its historical normal value; when a real damage occurs at a certain location, its It will remain higher than in multiple consecutive frames Thus Maintaining a high positive value during this period; while global perturbations can simultaneously increase the value of all pixels. However, because it is also captured by median filtering... Therefore During the disturbance, it tends to approach zero; therefore, Spatial decoupling of disturbance sources has been achieved.

[0047] Step 2.3: For each Statistics The sequence in all Upper quartiles in the frame This value represents the maximum tolerable level of normal deviation for this pixel over long-term observation; subsequently, the spatiotemporal consistency criterion is defined: if a certain frame middle ,in If the pixel deviates abnormally from its normal fluctuation range within the frame, it is determined that the pixel has deviated abnormally from its normal fluctuation range. This criterion does not use a fixed threshold, but rather uses the statistical dispersion of each pixel itself as a benchmark, adapting to the natural differences in noise levels in different regions—for example, in the loose soil area along the pile. The natural threshold is higher, and the threshold for judging abnormalities is also raised accordingly to avoid oversensitivity.

[0048] Step 2.4: Based on the above criteria, for each Generate binary mask values

[0049] The mask Constructing a spatiotemporally consistent mask sequence Its physical meaning is: a pixel is only assigned the identity of a potential anomaly candidate when it shows a significant deviation in a frame that exceeds its own fluctuation inertia; all other spatiotemporal points are marked as background credibility, and their corresponding weights will be reduced in the next step; the mask does not directly output the anomalous region, but serves as a logical switch for subsequent weighted fusion to ensure that only responses that pass spatiotemporal dual verification can participate in deep structure reconstruction.

[0050] Step 3: Perform structure-guided inter-frame registration on the dynamic suppression weight field sequence along the pile axis to generate the calibrated intensity field sequence. The mask sequence output in step two While the selection of spatiotemporal points worth preserving has been clearly identified, the inter-frame misalignment issue remains unresolved. Due to the micrometer-level axial expansion and contraction and lateral sway of the pile under dynamic loads, strain images between adjacent frames exhibit significant differences. There is subpixel-level translation in the direction (pile length direction), and this translation amount varies non-linearly with depth—for example, the displacement amplitude at the pile top is large, and at the pile bottom it approaches zero; if directly applied to… Calculating the mean or variance by stacking pixel coordinates will cause misalignment, leading to anomalous responses canceling each other out between frames, making them invisible in the statistics. Therefore, precise registration must be performed only on candidate pixel groups under mask guidance to ensure strict alignment of responses at the same physical location in different frames, thereby releasing their cumulative effect.

[0051] Step 3.1: For each frame Extract its mask The set of all pixel coordinates with a median value of 1 This set represents the candidate point cloud of anomalies that passed the spatiotemporal consistency test in this frame; because real damage has spatial connectivity, It usually appears as along The point cloud consists of narrow bands or discrete patches extending in a directional direction, rather than being completely randomly distributed; this point cloud provides a sparse but high-confidence set of control points for subsequent registration.

[0052] Step 3.2: Select the first frame As a reference frame, its candidate point cloud Each point in In the remaining frames Candidate point cloud Search for the nearest matching point in Euclidean distance. The search radius is limited to Pixels; if a unique matching point exists and the distance is less than 1000 pixels; Then record the displacement vector. This operation does not rely on interpolation or global deformation models; it establishes inter-frame correspondence solely based on the geometric proximity between high-confidence points, thus avoiding registration deviations caused by interference from low-confidence points.

[0053] Step 3.3: For each frame Based on all successful matches Fitting along the pile axis direction ( Piecewise linear displacement function of axis Specifically, The axis is divided into Given several equal-length intervals, for each interval... of all matching points within and Coordinates, calculate its average longitudinal offset ,in For falling in the first The set of matching point pairs in the interval; the piecewise average displacement function. Reflects the different depth sections of the pile body at different depths The actual stretching state of the frame relative to the reference frame conforms to the physical segmentation characteristics of pile deformation.

[0054] Step 3.4: For each frame gradient intensity field Only its candidate point clouds Perform vertical resampling on each pixel: Its intensity value Give a new position The new position is located in the reference frame coordinate system, and because... Obtained by fitting from high confidence points, the resampling error is constrained to the sub-pixel level; after resampling, the anomalous candidate responses of all frames are aligned to the same spatial reference frame, forming a calibrated intensity field sequence. ,in exist The value at this location indicates the position within the reference frame. exist The gradient intensity observed in the frame after displacement compensation; thus, the responses of the same physical damage in multiple frames are superimposed in phase, providing the geometric premise for its manifestation in subsequent steps.

[0055] Step 4: Aggregate the calibrated strength field sequence in the reference frame coordinate system to generate the pile damage confidence field. The calibrated intensity field sequence output in step three While spatiotemporal alignment has been achieved, the intensity of a single frame is still modulated by residual noise, and the actual damage often only reaches observable levels in some frames (e.g., it appears during the loading phase and attenuates after unloading). If only the maximum or average value is taken, the former is easily misled by the peak noise of a single frame, while the latter is lowered by the low values ​​of inactive frames, which drag down the overall response. Therefore, it is necessary to design an aggregation method that can reflect the degree of cohesion of anomalous responses in the temporal dimension, so that the confidence level depends not only on the intensity amplitude, but also on its stable frequency of occurrence and spatial coherence across multiple frames.

[0056] Step 4.1: For each pixel in the reference frame The number of frames assigned a value in the calibrated sequence is counted. This count reflects the position in the total. The frequency of frames covered by anomalous candidate point clouds; since both the mask selection in step two and the registration in step three are driven by high confidence points, A high value indicates that the anomaly criterion is repeatedly triggered at this location in multiple independent observation windows, demonstrating temporal repeatability.

[0057] Step 4.2: For each In all of its defined Calculate the upper quartile of the value. This value represents a robust high level of intensity response at that location, excluding extreme noise interference in a single frame; compared to the maximum value, the upper quartile is more sensitive to the persistence of anomalies—if a spike occurs only in one frame, its... Almost unaffected; however, if there is a moderate response in all 4 frames, then It will rise significantly.

[0058] Step 4.3: Define spatial connectivity weights: using Centered on the data, examine all pixels within its 8-neighborhood. Value, calculate the average frequency of the neighborhood

[0059] in It is an 8-neighbor set; the average frequency reflects the overall abnormal activity level of the region surrounding the pixel; the real damage is spatially expansive, and its neighboring pixels... Typically, noise levels rise synchronously; while isolated noise points... It must be far lower than Therefore, it can be identified and suppressed.

[0060] Step 4.4: Generate a pile damage confidence field by integrating frequency, strength robustness, and spatial connectivity. :

[0061] In the formula This is a scaling parameter used to adjust the contribution of frequency differences to the confidence level. The function ensures that when Slightly higher When, the confidence level is slightly improved; when Significantly higher than (That is, if this point is a local frequency peak), then the term approaches 2, and the confidence level nearly doubles; while The points were directly set to zero because they lacked evidence of time repetition and did not constitute a reliable anomaly; ultimately This is a real-valued image of the same size as the original image. Each pixel value represents the overall confidence level of the damage characteristics at that location in all observations. The higher the value, the more likely the location is to be the actual pile damage. This confidence field is the final output result of robustly extracting weak localized strain anomaly areas under low signal-to-noise ratio and strong interference conditions in this embodiment. It can be directly used for pixel-level damage localization and spatial morphology delineation.

[0062] Step 5: Perform multi-scale morphological closure and connected domain separation on the pile damage confidence field to generate a preliminary set of damage regions. The confidence field output in step four Spatiotemporal aggregation and spatial weighting have been completed, and the numerical distribution reflects the overall credibility of the damage, but a clearly identifiable region with distinct geometric boundaries has not yet been formed. This is because real pile damage (such as vertical microcracks, local voids, and strain concentration zones caused by steel reinforcement rust expansion) appears in images as a continuous structure extending along the pile axis, with a limited width (typically 3–15 pixels) and gradually changing edges. The system still retains discrete high-value points caused by registration residuals or neighborhood statistical fluctuations, and direct thresholding will lead to fragmented responses. Therefore, morphological operations are needed to eliminate isolated noise responses, bridge damage bands interrupted by intensity attenuation, and perform preliminary classification based on spatial extension characteristics, while preserving the original structural topology.

[0063] Step 5.1: For the confidence field Implement adaptive dual-threshold binarization - first calculate the entire image global median Then set the next threshold. Upper threshold For each pixel ,like Mark as a strong response point; if Mark as a weak response point; if The points are marked as background points. This dual-threshold strategy does not rely on fixed percentiles, but uses the median as the anchor point, so that the threshold automatically scales with the overall confidence level to adapt to the differences in damage manifestation intensity under different working conditions.

[0064] Step 5.2: Perform structuring on the set of strong response points with a struct size of [size missing]. The closing operation (dilation followed by erosion) fills the internal holes and connects adjacent strong-response pixels; the structuring element used for closing is a square, covering the center and all 8 neighborhoods; this operation only acts on strong-response points, avoiding blindly expanding weak-response areas into pseudo-connected regions; after this processing, the fractured high-strength damage band can be reconstructed into a single-connected structure, while isolated strong points are completely eliminated in the erosion stage because they cannot form stable neighborhood support.

[0065] Step 5.3: Apply a size of [size] to the weak response point set. The opening operation (erosion followed by dilation) removes small noise spots while preserving weak response regions of a certain area. In the opening operation, the erosion step shrinks all weak response regions by one ring. Only if a weak response region still has at least one pixel that has not been erased after shrinkage is it fully restored in the subsequent dilation. This design ensures that only weak response structures with the minimum spatial scale (≥3×3 pixels) are preserved, eliminating substructure artifacts caused by interpolation errors or registration jitter.

[0066] Step 5.4: Perform a logical union between the closed strong response region and the weak response region preserved by the opening operation to obtain a unified binary mask. Subsequently, connected component labeling is performed on the mask to obtain the set of all non-overlapping maximum connected pixels. Each of them Indicates a preliminary screening damage area; each It contains at least 9 pixels, and its pixel coordinate set satisfies the following: there exists a 4-neighborhood path consisting of pixels with the same value between any two points; this set is the initial screening damage region set, denoted as . ; At this point, The dispersed confidence responses have been aggregated into candidate entities with clear spatial boundaries and internal connectivity, providing a geometric basis for subsequent deep structural analysis.

[0067] Step Six: Based on the pile axis geometric constraints and strain transfer laws, perform directional consistency verification on the initially screened damage areas to generate a set of effective damage areas. The initial set of damaged areas output in step five While spatial integrity has been achieved, two types of false alarms have not yet been ruled out: one type is short transverse bands caused by transverse construction joints or sensor installation scratches, whose direction is perpendicular to the pile axis, contradicting the longitudinal mechanical response characteristics of true damage; the other type is L-shaped or angular pseudo-regions formed by image edge distortion or inter-frame registration boundary effects, whose shape does not conform to the basic law of gradual strain along the axial direction under pile stress. Therefore, it is necessary to introduce the inherent geometric and mechanical priors of the pile body—that is, the true damage response must exhibit longitudinal... The dominant extension in the direction (pile length direction), and the variation of its internal strain gradient should obey the stress diffusion trend in a one-dimensional continuous medium - for each Perform dual verification of directionality and gradient monotonicity.

[0068] Step 6.1: For each initial screening area Extract the coordinates of all pixels it covers. Fit its principal inertial axes: Construct the covariance matrix

[0069] In the formula for The centroid coordinates of the matrix are used to find the eigenvectors of the matrix, and the unit eigenvectors corresponding to the larger eigenvalues ​​are selected. As the main direction of this region; this vector represents The longest extension orientation in the plane.

[0070] Step 6.2: Calculate the principal direction and the pile axis direction (i.e., The cosine of the angle between the positive axis and the unit vector (0,1)

[0071] This value measures The degree of consistency between the extension direction and the pile length direction; the cosine value of the principal direction of the actual damage should be close to 1 due to the axial load driving it; set the criterion: if If the direction of the area deviates significantly from the pile axis, it is determined that the effective set is removed. This threshold corresponds to a deflection angle of approximately 31.8°, which is sufficient to accommodate construction tilt and imaging perspective deviation, but excludes typical lateral interference.

[0072] Step 6.3: For each of the remaining... Arrange its pixels in the main direction Projecting onto one-dimensional coordinate axes: Defining projected coordinates and for all according to Sort in ascending order; then extract the confidence value sequence at the corresponding position. Calculate its first-order forward difference. ; Calculate the percentage of negative values ​​in the difference sequence. ,in This is an indicator function; true damage often exhibits a single peak or a gradually rising-decreasing shape in the main direction, and the proportion of negative values ​​in its difference sequence should be lower than the proportion of positive values; if This indicates that the confidence level oscillates frequently along the main direction, which does not conform to the smooth transmission characteristics of strain diffusion, and therefore it is rejected.

[0073] Step 6.4: Define the effective damage region set by combining the directional consistency and gradient monotonicity criteria.

[0074] This set Each All conditions are met: the spatial extension axis is close to the pile axis, the confidence distribution changes gently along the axis, and there is no high-frequency reversal; its geometric shape is highly consistent with the physical manifestation of pile damage, making it a reliable input source for subsequent positioning and assessment.

[0075] Step 7: Reconstruct the strain anomaly intensity profile within each effective damage area to generate a sequence of pile damage depth assignments. Step 6 outputs the set of effective damage areas The spatial accuracy has been confirmed, but the question of at what depth the damage is located within the pile remains unanswered. The original strain image uses pixel rows to correspond to pile depth; however, due to imaging perspective, fiber optic bending, and the non-uniformity of lateral soil constraints, the same pixel row may cover several centimeters of pile, and the depth mapping relationship varies locally across different areas. Therefore, it is necessary to perform a depth mapping analysis on each... Internally, based on its internal pixels in the original strain image sequence The temporal response characteristics presented in the data are used to inversely determine the most likely corresponding pile depth range.

[0076] Step 7.1: For each Extract all its pixels In all Time series mean in the original strain image of the frame

[0077] That is: for each fixed Coordinates, collection The strain values ​​of all pixels located in that row across all frames are averaged over time to obtain the strain along the line. Average strain profile in the direction This profile reflects the average strain level of the damaged area at different depths, eliminating single-frame transient disturbances.

[0078] Step 7.2: For Perform cubic spline interpolation to generate a continuous function. Domain coverage Location Minimum coordinates To the maximum value Then, its second derivative is calculated. and in the interval Internal search for all that meet the criteria The point, among which This condition identifies the inflection point where the curvature drops significantly downward, corresponding to the depth at which strain changes from growth to decay, i.e., the lower boundary of the damage's influence range.

[0079] Step 7.3: For the same Search all that meet the criteria The point marks the inflection point where the curvature significantly increases, corresponding to the depth at which strain changes from decay to increase, i.e., the upper boundary of the damage influence range; the minimum value among all upper inflection points is taken. Value as upper bound The largest of all downward inflection points value as lower bound If no inflection point satisfying the conditions is found, the process degenerates into taking... The original Boundary; this operation ensures that depth attribution is always based on the inherent curvature characteristics of the strain profile, rather than being artificially defined.

[0080] Step 7.4: Define the depth assignment interval of the effective damage region as a closed interval. and its midpoint This serves as the representative depth of the damage; ultimately, a sequence of pile damage depth assignments is generated.

[0081] Each item in the sequence consists of a scalar depth value and a set of pixels, which correspond one-to-one and form a traceable spatial-depth joint description, providing a positioning benchmark for subsequent structural integrity assessment.

[0082] Step 8: Generate pile structural integrity status labels based on the damage depth attribution sequence and pile design parameters. The depth attribution sequence output in step seven While precise location of individual damages has been achieved, the final engineering assessment must focus on whether the damage jeopardizes the pile foundation's bearing capacity. Different depth sections of the pile bear different mechanical roles: the top section bears bending moment and shear force, the middle section mainly bears axial force, and the end section controls settlement and end resistance. Therefore, it is necessary to determine the representative depth of each damage. The structural integrity status is determined by mapping the data to the functional zones of the pile body and combining the strain amplitude level of the zone.

[0083] Step 8.1: According to the pile design drawings, determine the total pile length. Divided into three functional zones: pile top zone ( ), main pile body area ( ), pile tip area ( ); for each Determine its functional area category. This classification does not rely on actual measurements but is uniquely determined by design parameters, ensuring that the criteria are objective and reproducible.

[0084] Step 8.2: For each Calculate the confidence field of all pixels within it. mean This value reflects the overall significance level of the damage across all observations; because It already integrates frequency, intensity and spatial connectivity. It can be regarded as a dimensionless representation of the severity of the injury.

[0085] Step 8.3: Based on functional area category with the mean Checking the preset grading table: the threshold for the pile top zone criterion is... The main body area of ​​the pile is The pile tip area is This difference stems from the varying tolerances to anomalous responses in different regions—the pile top allows for a certain degree of plastic deformation, while the pile tip is extremely sensitive to even minor voids; if Mark it as "needs attention"; otherwise mark it as "acceptable".

[0086] Step 8.4: Based on the combined marking results of all damages, generate a structural integrity status label for the pile body: If all If all are marked as acceptable, the overall label is Grade A: structural integrity is not affected; If there is at least one area that requires attention and its functional area does not include the pile end area, then the label is B: local performance degradation, and a review is recommended. If any of the structures requiring attention exists and its functional area is the pile end area, or if there are two or more structures requiring attention that span more than two functional areas, then the label is C: there is a risk to structural integrity, and intervention is recommended. This label does not output probability or confidence values, but instead directly corresponds the three-level discrete state to the engineering decision chain, completing a closed-loop deduction from the original strain image to structural safety assessment.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for identifying pile damage based on strain images, characterized in that, include: Spatiotemporal adaptive gradient normalization is performed on the original strain image sequence to generate a gradient intensity field sequence. A spatiotemporal consistency mask is constructed based on the gradient intensity field sequence to generate a dynamic suppression weight field sequence. Structure-guided inter-frame registration is performed on the dynamic suppression weight field sequence along the pile axis to generate a calibrated intensity field sequence. The calibrated intensity field sequence is then aggregated in the reference frame coordinate system to generate a pile damage confidence field. Multi-scale morphological closure and connected domain separation are performed on the pile damage confidence field to generate a preliminary set of damage regions. Based on the pile axis geometric constraints and strain transmission law, the directional consistency check is performed on the initial set of damage areas to generate a set of effective damage areas. Within each effective damage area, a strain anomaly intensity profile is reconstructed to generate a pile damage depth attribution sequence. Based on the pile damage depth attribution sequence and pile design parameters, a pile structural integrity status label is generated.

2. The pile damage identification method based on strain images according to claim 1, characterized in that, The spatiotemporal adaptive gradient normalization of the original strain image sequence includes: A square neighborhood is extracted centered on each pixel, and the horizontal and vertical gradient components within the neighborhood are calculated. Calculate the median of the absolute values ​​of the gradient components in the neighborhood to obtain the horizontal median and the vertical median. The original gradient components are normalized using the median in the horizontal direction and the median in the vertical direction. The intensity values ​​of the normalized gradient components are calculated to generate a gradient intensity field sequence.

3. The pile damage identification method based on strain images according to claim 2, characterized in that, The construction of the spatiotemporal consistency mask includes: Extract the gradient intensity time series for each pixel location and perform sliding window mid-value filtering to obtain a robust background intensity estimate; Calculate the relative deviation between the original intensity and the robust background intensity estimate, statistically analyze the upper quartile of the relative deviation sequence, and set a spatiotemporal consistency criterion; A binary mask is generated based on the spatiotemporal consistency criterion, forming a spatiotemporal consistency mask sequence.

4. The pile damage identification method based on strain images according to claim 1, characterized in that, The implementation of structure-guided inter-frame registration includes: Extract the set of pixel coordinates with a value of 1 from the mask sequence to form a candidate point cloud; The first frame is selected as the reference frame. The nearest matching point is searched in the candidate point cloud of other frames. The displacement vector is recorded. A piecewise linear displacement function along the pile axis is fitted based on the matching point. Vertical resampling is performed only at candidate point cloud locations for the gradient intensity field to generate a calibrated intensity field sequence.

5. The pile damage identification method based on strain images according to claim 1, characterized in that, The generated pile damage confidence field includes: Count the frequency at which each pixel position is assigned a value across all frames; Calculate the upper quartile of the defined intensity value for each pixel location, and calculate the average frequency of pixels in the 8-neighborhood to obtain the spatial connectivity weight; By combining the strength upper quartile, frequency and spatial connectivity weights, a pile damage confidence field is generated.

6. The pile damage identification method based on strain images according to claim 1, characterized in that, The implementation of multi-scale morphological closure and connected component separation includes: Adaptive dual-threshold binarization is applied to the pile damage confidence field to distinguish between strong and weak response points; Perform small-size closing operations on strong response point sets and large-size opening operations on weak response point sets; After merging the strong and weak response regions, connective domain labeling is performed to generate a preliminary set of damaged regions.

7. The pile damage identification method based on strain images according to claim 1, characterized in that, The consistency check of the implementation direction includes: For each initially screened damage area, fit the principal inertial axis, calculate the cosine value of the angle between the principal direction and the pile axis direction, and retain the area where the cosine value of the angle is greater than the preset threshold. The confidence sequence is extracted along the main direction of the retained region, the proportion of negative values ​​in the first-order difference sequence is calculated, and the region with a negative value proportion less than the preset proportion is retained to generate a set of effective damage regions.

8. The pile damage identification method based on strain images according to claim 1, characterized in that, The reconstructed strain anomaly intensity profile includes: The mean time series value of each pixel in the effective damage area in the original strain image sequence is extracted to obtain the average strain profile. Interpolate the mean strain profile and calculate the second derivative; The lower boundary of the damage is determined by searching for inflection points with significantly downward curvature, and the upper boundary of the damage is determined by searching for inflection points with significantly upward curvature. The midpoint of the calculated depth interval is used as the representative depth of the damage to generate a sequence of pile damage depths.

9. The pile damage identification method based on strain images according to claim 1, characterized in that, The generated pile structure integrity status label includes: Based on the pile design parameters, the pile length is divided into the pile top area, the main body area, and the pile end area; Calculate the mean value of each effective damage area in the confidence field, and determine the damage level by looking up the table based on the functional area category and the mean value; A three-level structural integrity status label is generated by combining all damage levels.

10. A pile damage identification system based on strain images for implementing the method as described in any one of claims 1-9, characterized in that, include: The gradient normalization module is used to perform spatiotemporal adaptive gradient normalization on the original strain image sequence to generate a gradient intensity field sequence. The mask construction module is used to construct a spatiotemporally consistent mask based on the gradient intensity field sequence and generate a dynamic suppression weight field sequence. The inter-frame registration module is used to perform structure-guided inter-frame registration on the dynamic suppression weight field sequence along the pile axis direction to generate a calibrated intensity field sequence. The confidence aggregation module is used to aggregate the calibrated intensity field sequence in the reference frame coordinate system to generate a pile damage confidence field. The region separation module is used to perform multi-scale morphological closure and connected domain separation on the pile damage confidence field to generate a preliminary set of damage regions. The orientation verification module is used to verify the orientation consistency of the initial set of damaged areas based on the pile axis geometric constraints and strain transmission laws, and to generate a set of valid damaged areas. The depth attribution module is used to reconstruct the strain anomaly intensity profile within each effective damage area and generate a pile damage depth attribution sequence. The status label module is used to generate status labels for the structural integrity of the pile body based on the pile damage depth attribution sequence and pile design parameters.

Images