Low strain defect identification method based on characteristic curve matching

Abstract

The application is specifically a low-strain defect identification method based on characteristic curve matching, relates to the technical field of pile nondestructive testing, and comprises the following steps: obtaining a standard defect characteristic curve template library and a measured low-strain reflected wave curve; performing variational modal decomposition on the measured low-strain reflected wave curve to obtain a plurality of intrinsic modal functions; selecting a target modal function located in a preset frequency range according to the center frequency of each intrinsic modal function; and superimposing and reconstructing the target modal function to obtain a denoised characteristic curve. In the application, the measured low-strain curve is matched with the standard defect characteristic curve template library to identify the defect type, and the identification conclusion is directly given by the calibrated defect label in the template library, thereby solving the technical problems of large subjective difference and difficult-to-ensure consistency in manual interpretation, and achieving the technical effect of reducing the determination difference between different testers.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology for foundation piles, and in particular to a method for identifying low-strain defects based on characteristic curve matching. Background Technology

[0002] The low-strain reflected wave method is a commonly used technique for detecting the integrity of foundation piles. Its principle is to determine the presence of defects in the pile body by analyzing the reflected wave characteristics in the pile top velocity response curve after a transient impact is applied to the pile top. When defects such as necking, widening, segregation, or pile breakage exist in the pile body, the cross-sectional impedance changes, generating reflected waves of different shapes at the defect location.

[0003] Currently, the interpretation of low-strain curves mainly relies on the manual experience of the inspectors. Inspectors infer the type and location of defects by observing the polarity, amplitude, and arrival time of the reflected waves.

[0004] However, the above-mentioned manual interpretation method has the following shortcomings: Different inspectors may arrive at different conclusions regarding the same curve, resulting in significant subjective differences. When multiple defects or minor defects exist simultaneously in the pile body, the reflected wave characteristics may overlap or become indistinct, reducing the accuracy of manual interpretation. Furthermore, the signals acquired on-site are mixed with noise from environmental vibrations and shallow soil interference, further increasing the difficulty of manual interpretation. These shortcomings make it difficult to guarantee the consistency and accuracy of low-strain testing results, which is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to propose a low-strain defect identification method based on characteristic curve matching in order to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The low-strain defect identification method based on characteristic curve matching includes at least step A and step B, wherein step A includes: Obtain a standard defect feature curve template library and a measured low-strain reflection wave curve; perform variational mode decomposition on the measured low-strain reflection wave curve to obtain multiple intrinsic mode functions; select target mode functions within a preset frequency range according to the center frequency of each intrinsic mode function, and superimpose and reconstruct the target mode functions to obtain a denoised feature curve; Step B includes: Calculate the improved dynamic time warping distance between the feature curve and each template curve in the template library, weighted by amplitude distance and derivative distance, and select the top few template curves with the smallest distance to form a candidate matching set; pair the feature curve with each template curve in the candidate matching set and input them into a Siamese neural network to obtain the similarity score of each curve pair; output the defect identification result based on the comparison result of the similarity score and the preset similarity threshold.

[0007] Preferably, in step A, after obtaining the measured low-strain reflection wave curve and the template library, the method further includes: The measured low-strain reflected wave curve and each template curve in the template library are subjected to amplitude normalization processing to ensure that the amplitude of the normalized curves is within a uniform numerical range.

[0008] Preferably, the step of selecting the target mode function within a preset frequency range based on the center frequency of each intrinsic mode function further includes: Calculate the center frequency corresponding to each intrinsic mode function; mark the mode functions with center frequencies higher than the upper frequency threshold as noise modes; mark the mode functions with center frequencies lower than the lower frequency threshold as drift modes; select the mode functions with center frequencies between the lower frequency threshold and the upper frequency threshold to form the target mode function set.

[0009] Preferably, after selecting the target mode function set, the method further includes: For modal functions whose center frequency is near the upper or lower frequency threshold, calculate their correlation coefficient with the original signal; when the correlation coefficient is higher than a preset correlation threshold, retain the modal function in the target modal function set; when the correlation coefficient is lower than the correlation threshold, exclude the modal function from the target modal function set.

[0010] Preferably, the calculation of the improved dynamic time warping distance further includes: Calculate the first-order difference for the feature curve and each template curve to obtain their respective derivative sequences; calculate the improved cost matrix based on the amplitude distance and derivative distance between the feature curve and each template curve; wherein, the value of each element in the improved cost matrix is ​​equal to the weighted sum of the amplitude distance and derivative distance between the corresponding sampling point pairs, the amplitude distance is the absolute value of the difference between the amplitude of the sampling point of the feature curve and the amplitude of the sampling point of the template curve, the derivative distance is the absolute value of the difference between the derivative value of the sampling point of the feature curve and the derivative value of the sampling point of the template curve, and the sum of the amplitude distance weight coefficient and the derivative distance weight coefficient is one.

[0011] Preferably, after obtaining the improved cost matrix, the process further includes: Local path constraints are applied to the improved cost matrix. The local path constraints are as follows: the stepping direction of the regularized path at any position is limited to three types: horizontal stepping, vertical stepping, and diagonal stepping, and the number of consecutive steps along the same direction does not exceed a preset upper limit value. The regularized path with the minimum cumulative cost is solved by dynamic programming under the constraints. The cumulative cost corresponding to the path is normalized and used as the improved dynamic time regularization distance.

[0012] Preferably, the Siamese neural network comprises two branch networks with shared weights and a similarity calculation layer; each branch network consists of a convolutional feature extraction layer and a temporal feature extraction layer connected in series. The convolutional feature extraction layer uses a convolutional neural network to extract local morphological features of the curve and outputs a morphological feature vector. The temporal feature extraction layer uses a long short-term memory network to extract temporal dependent features along the time dimension and outputs a temporal feature vector. The similarity calculation layer calculates the Euclidean distance between the temporal feature vectors output by the two branch networks and obtains the similarity score after mapping with a sigmoid function.

[0013] Preferably, the number of candidates among the first few template curves with the smallest distance is adaptively determined based on the size of the template library: When the total number of template curves in the template library does not exceed a preset scale threshold, the number of candidates is a fixed value; when the total number of template curves exceeds the scale threshold, the number of candidates is the product of the total number of template curves and a preset proportional coefficient.

[0014] Preferably, after outputting the defect identification result, the method further includes: When the defect identification result includes a defect type, the travel time of the corresponding reflected wave in the measured low-strain reflected wave curve is obtained. Combined with the longitudinal wave velocity of the pile body, half of the product of the travel time and the longitudinal wave velocity of the pile body is taken as the actual depth of the defect from the top of the pile.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention identifies defect types by matching measured low-strain curves with a standard defect characteristic curve template library. The identification conclusion is directly given by the defect labels already marked in the template library. This solves the technical problem of large subjective differences and difficulty in ensuring consistency in manual interpretation, and achieves the technical effect of reducing the judgment differences between different testing personnel.

[0016] 2. This invention uses variational mode decomposition to select intermediate frequency mode functions containing defect reflection information for reconstruction, removing high-frequency noise and low-frequency drift components, thus reducing the interference of on-site noise on the matching process. The improved dynamic time warping algorithm simultaneously introduces amplitude distance and derivative distance into the cost function and applies local path step constraints, thereby reducing the possibility of erroneous matching when waveforms are locally similar but mismatched overall. The Siamese neural network extracts local morphological features and temporal dependency features of the curve through convolutional neural networks and long short-term memory networks, respectively, achieving the technical effect of further filtering candidate matching results from both morphological and temporal dimensions and improving the accuracy of defect type identification. Attached Figure Description

[0017] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0019] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0020] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0021] In this embodiment, it includes: An engineering testing agency conducted low-strain integrity testing on a batch of precast concrete pipe piles. Time history data of the velocity response at the top of pile 001 were collected on-site, leading to a preliminary assessment that the pile body might have defects. The testing personnel imported the measured data into a computer device. The computer's memory contained a template library of 487 standard defect characteristic curves, covering four defect types: necking, widening, segregation, and broken piles. Each type included three levels: mild, moderate, and severe, with defect depths ranging from 2 meters to 18 meters. Pile 001 had a designed pile diameter of 400 mm and a designed pile length of 14 meters. The longitudinal wave velocity measured on-site was 3850 m / s.

[0022] In step S100, the computer device reads 487 template curves and their associated defect labels from the memory, and simultaneously loads the measured low-strain reflected wave curve of pile 001. The sampling frequency of the measured curve is 50 kHz, the number of sampling points is 2048, and the time length is approximately 41 milliseconds. After reading, amplitude normalization processing is uniformly performed on the measured curves and all template curves, and the amplitude range of each curve is mapped to the interval between -1 and 1.

[0023] Table 1 Sample data of the template library In step S201, the normalized measured low-strain reflection wave curve is... Input variational mode decomposition algorithm, decompose to obtain eigenmode functions to The center frequencies of each intrinsic mode function are shown in Table 2.

[0024] Table 2 Center frequencies of each intrinsic mode function In step S202, according to the preset parameters Hz, Hz, the center frequency of mode 1 is 38Hz, which is lower than The mode was determined to be a drift mode; the center frequencies of mode 5 and mode 6 were 8620Hz and 19450Hz, respectively, both higher than [the specified values]. The mode was identified as a noise mode; the center frequencies of modes 2, 3, and 4 are between 100Hz and 5000Hz, and they are initially included in the target mode function set.

[0025] The center frequency of mode 2 is 285Hz. Near Hz, further calculations are needed to determine its relationship with the original signal. The correlation coefficient between mode 2 and mode 3 is shown in the calculation results. The correlation coefficient is 0.71, which is higher than the preset correlation threshold of 0.60, therefore mode 2 is retained. The index set of the final target mode function set is... .

[0026] In step S203, modes 2, 3, and 4 are superimposed and summed to obtain the denoised feature curve. The denoised characteristic curves retain the complete shape of the incident wave at the pile top and the reflected wave from the defect, and both high-frequency burrs and low-frequency baseline shifts are suppressed.

[0027] In step S301, for Calculate the first-order difference for each of the 487 template curves in the template library to obtain the derivative sequence. and Since all curves have undergone amplitude normalization, the dimensions of the derivative sequence are consistent with those of the amplitude sequence, and they can be directly used in subsequent weighted summation.

[0028] In step S302, the amplitude distance weight is taken. Derivative distance weight The improved cost matrix is ​​calculated for each feature curve and each template curve. Taking the matching of the feature curve and template T0124 as an example, the sampling points... , The cost element at that location is: ; Substitution , ,have to .

[0029] In step S303, set the upper limit value of the step. After applying local path constraints to each improved cost matrix, dynamic programming is performed to obtain the normalized path with the minimum cumulative cost, which is then normalized to obtain the improved dynamic time normalized distance corresponding to each template curve. .

[0030] In step S304, 487 The values ​​are sorted from smallest to largest. The template library size of 487 exceeds the preset size threshold of 400, therefore the candidate number parameter... Pick That is, the top 10 template curves with the smallest distance are selected to form a candidate matching set.

[0031] Table 3 Candidate Match Set (First 6 Matches) In step S401, the characteristic curve is... Each curve is paired with one of the ten template curves in the candidate matching set to form ten curve pairs.

[0032] In step S402, the 10 curve pairs are sequentially input into the pre-trained Siamese neural network. The two curves of each pair are fed into two branch networks sharing weights. Each branch network extracts local morphological feature vectors via a convolutional neural network, and then outputs temporal feature vectors via a long short-term memory network. and The similarity calculation layer calculates the Euclidean distance. The similarity score is obtained by mapping using the sigmoid function. The calculation results for each candidate curve pair are shown in Table 4.

[0033] Table 4. Similarity Calculation Results of Siamese Neural Networks In step S403, the maximum similarity score among the 10 curve pairs is This corresponds to template T0124. A preset similarity threshold is set. , Therefore, the defect label associated with T0124 is output as the defect identification result: the defect type is necking and the defect reference depth is 5.20 meters.

[0034] In step S404, the travel time of the necked-down reflection wave is extracted from the measured low-strain reflection wave curve. Milliseconds, combined with the longitudinal wave velocity of the pile body Meters per second, calculate the actual depth of the defect: rice; The actual depth of the defect from the top of the pile is 5.20 meters, which matches the defect depth marked on template T0124.

[0035] In step S405, the normalized amplitude of the defect reflection wave in the measured curve is extracted. Query the table of correspondence between amplitude and defect severity. This amplitude corresponds to the moderate level of necking defect, which is consistent with the defect severity label of template T0124.

[0036] In steps S406 and S407, the defect depth (5.20 meters), defect severity quantification value, similarity score (0.42), and three-dimensional spatial distribution data of all measuring points of pile 001 are summarized, and a structured inspection report is generated according to the preset report template.

[0037] Table 5 Summary of Final Inspection Results for Pile 001 The entire data stream originates from the original measured time history data of pile 001. After normalization, it enters variational mode decomposition. The six intrinsic mode functions obtained from the decomposition are filtered by center frequency and verified by correlation coefficient. Modes 2, 3, and 4 are retained and superimposed to reconstruct feature curves. The waveform information carried by the feature curves then enters the improved dynamic time warping calculation stage. Through the weighted cost matrix of amplitude distance and derivative distance and local path constraints, 10 candidate templates are selected from 487 templates. The candidate templates enter the Siamese neural network fine matching stage. The temporal feature vectors extracted by convolution and long short-term memory networks are mapped by Euclidean distance and sigmoid, and a similarity score is output. Finally, template T0124 is confirmed as the optimal match, corresponding to a moderate necking defect located 5.20 meters from the pile top. The output data of each step is directly used as the input of the next step, ensuring a complete data link from the original acquired signal to the final detection report.

[0038] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0039] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0040] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A low-strain defect identification method based on characteristic curve matching, characterized in that, It includes at least steps A and B, wherein step A includes: Obtain a standard defect feature curve template library and a measured low-strain reflection wave curve; perform variational mode decomposition on the measured low-strain reflection wave curve to obtain multiple intrinsic mode functions; select target mode functions within a preset frequency range according to the center frequency of each intrinsic mode function, and superimpose and reconstruct the target mode functions to obtain a denoised feature curve; Step B includes: Calculate the improved dynamic time warping distance between the feature curve and each template curve in the template library, weighted by amplitude distance and derivative distance, and select the top few template curves with the smallest distance to form a candidate matching set; pair the feature curve with each template curve in the candidate matching set and input them into a Siamese neural network to obtain the similarity score of each curve pair; output the defect identification result based on the comparison result of the similarity score and the preset similarity threshold. Calculating the improved dynamic time warp distance also includes: Calculate the first-order difference for the feature curve and each template curve to obtain their respective derivative sequences; calculate the improved cost matrix based on the amplitude distance and derivative distance between the feature curve and each template curve; wherein, the value of each element in the improved cost matrix is ​​equal to the weighted sum of the amplitude distance and derivative distance between the corresponding sampling point pairs, the amplitude distance is the absolute value of the difference between the amplitude of the sampling point of the feature curve and the amplitude of the sampling point of the template curve, the derivative distance is the absolute value of the difference between the derivative value of the sampling point of the feature curve and the derivative value of the sampling point of the template curve, and the sum of the amplitude distance weight coefficient and the derivative distance weight coefficient is one; The Siamese neural network comprises two branch networks sharing weights and a similarity calculation layer. Each branch network consists of a concatenated convolutional feature extraction layer and a temporal feature extraction layer. The convolutional feature extraction layer uses a convolutional neural network to extract local morphological features of the curve and outputs a morphological feature vector. The temporal feature extraction layer uses a long short-term memory network to extract temporal dependent features along the time dimension and outputs a temporal feature vector. The similarity calculation layer calculates the Euclidean distance between the temporal feature vectors output by the two branch networks and obtains the similarity score after mapping with a sigmoid function.

2. The low-strain defect identification method based on characteristic curve matching according to claim 1, characterized in that, In step A, after obtaining the measured low-strain reflection wave curve and the template library, the method further includes: The measured low-strain reflected wave curve and each template curve in the template library are subjected to amplitude normalization processing to ensure that the amplitude of the normalized curves is within a uniform numerical range.

3. The low-strain defect identification method based on characteristic curve matching according to claim 1, characterized in that, Selecting target mode functions within a preset frequency range based on the center frequencies of each intrinsic mode function also includes: Calculate the center frequency corresponding to each intrinsic mode function; mark the mode functions with center frequencies higher than the upper frequency threshold as noise modes; mark the mode functions with center frequencies lower than the lower frequency threshold as drift modes; select the mode functions with center frequencies between the lower frequency threshold and the upper frequency threshold to form the target mode function set.

4. The low-strain defect identification method based on characteristic curve matching according to claim 3, characterized in that, After selecting the target modal function set, the method further includes: For modal functions whose center frequency is near the upper or lower frequency threshold, calculate their correlation coefficient with the original signal; when the correlation coefficient is higher than a preset correlation threshold, retain the modal function in the target modal function set; when the correlation coefficient is lower than the correlation threshold, exclude the modal function from the target modal function set.

5. The low-strain defect identification method based on characteristic curve matching according to claim 1, characterized in that, After obtaining the improved cost matrix, the process also includes: Local path constraints are applied to the improved cost matrix. The local path constraints are as follows: the stepping direction of the regularized path at any position is limited to three types: horizontal stepping, vertical stepping, and diagonal stepping, and the number of consecutive steps along the same direction does not exceed a preset upper limit value. The regularized path with the minimum cumulative cost is solved by dynamic programming under the constraints. The cumulative cost corresponding to the path is normalized and used as the improved dynamic time regularization distance.

6. The low-strain defect identification method based on characteristic curve matching according to claim 1, characterized in that, The number of candidates is adaptively determined based on the size of the template library, selecting the top few template curves with the smallest distance. When the total number of template curves in the template library does not exceed a preset scale threshold, the number of candidates is a fixed value; when the total number of template curves exceeds the scale threshold, the number of candidates is the product of the total number of template curves and a preset proportional coefficient.

7. The low-strain defect identification method based on characteristic curve matching according to claim 1, characterized in that, After outputting the defect identification result, the following is also included: When the defect identification result includes a defect type, the travel time of the corresponding reflected wave in the measured low-strain reflected wave curve is obtained. Combined with the longitudinal wave velocity of the pile body, half of the product of the travel time and the longitudinal wave velocity of the pile body is taken as the actual depth of the defect from the top of the pile.

Images