Intelligent extraction method and system for hole wall morphology in ultrasonic hole forming quality detection

Abstract

The application provides a hole wall morphology intelligent extraction method and system in ultrasonic hole forming quality detection, relates to the technical field of ultrasonic hole forming quality detection, and comprises the following steps: preprocessing a multi-channel original ultrasonic waveform of a current measuring point to obtain a preprocessed waveform corresponding to each channel; determining all candidate peaks; removing the candidate peaks located in a direct wave area or having pulse window energy lower than a direct wave energy threshold; determining candidate first wave positions from the removed candidate peaks based on the first wave positions of historical measuring points; inputting the preprocessed waveform into a pre-trained deep learning model to determine final first wave positions; calculating hole wall distances of the current measuring point according to the final first wave positions of each channel; re-executing the above process for the next depth measuring point until the hole wall distances of all measuring points are obtained; and smoothing the hole wall distances of all measuring points to obtain the hole wall morphology. The accuracy of the first wave picking is improved, and the high-precision identification of the hole wall morphology is ensured.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic hole forming quality inspection technology, specifically to a method and system for intelligent extraction of hole wall morphology in ultrasonic hole forming quality inspection. Background Technology

[0002] The quality of bored pile hole formation directly affects the bearing capacity of the pile foundation and the safety of the project. According to Chapter 4 of the industry standard JTG / T 3512-2020 "Technical Specification for Testing of Highway Engineering Foundation Piles", the quality testing of bored pile hole formation generally adopts an ultrasonic four-channel probe. The probe collects the reflected waveforms of the hole wall in four orthogonal directions at each measuring point along the depth direction of the pile hole. By identifying the arrival time of the first wave of the reflected wave, the radial distance from the probe to the hole wall is calculated, and then the hole wall profile is fitted.

[0003] The standard procedure for ultrasonic testing includes:

[0004] (1) A four-channel orthogonal ultrasonic probe is used, and measuring points are set up at equal intervals along the depth of the hole to transmit and receive sound wave signals synchronously;

[0005] (2) Record the round-trip time of sound waves Calculate the radial distance of the hole wall using the following formula:

[0006] (1)

[0007] in, This indicates the radial distance from the center of the probe to the wall of the borehole; Indicates the speed of sound in mud; Indicates the round-trip travel time of a sound wave;

[0008] (3) The arrival time of the first wave is obtained by the testing personnel through manual interpretation or semi-automatic means;

[0009] (4) Calculate the hole diameter and verticality at each measuring point based on the radial distance of the four orthogonal directions, evaluate the hole wall morphology, and output a hole quality inspection report.

[0010] The current mainstream processing flow typically includes the following steps:

[0011] Signal acquisition: A four-channel ultrasonic transducer is used to synchronously transmit and receive ultrasonic waves to acquire four-channel time-domain waveform data at a single measurement point;

[0012] Simple preprocessing: Amplitude correction and basic filtering are performed on the original ultrasonic waveform to eliminate some high-frequency noise and DC offset;

[0013] Reflected wave pickup: Based on a fixed threshold or a fixed template, find the position of the first wave peak in a single-point waveform as the arrival time of the reflected wave from the hole wall;

[0014] Contour output: The first wave position picked up by each measuring point is converted into radial distance, and then stitched together point by point along the depth direction. After simple smoothing, the aperture and perpendicularity are calculated and output.

[0015] Existing algorithms generally rely on fixed thresholds or inherent templates and single-point independent peak determination, which have the following problems and drawbacks:

[0016] (1) Weak anti-interference ability and easy to misjudge: In actual engineering use, the borehole wall reflection signal is greatly affected by mud properties, borehole wall roughness and formation differences, which will generate strong clutter interference and multiple reflection signals. When the reflection is weak and the signal-to-noise ratio is low, traditional methods are difficult to effectively distinguish the real borehole wall primary reflection signal from these interference signals, which can easily lead to borehole diameter measurement distortion, defect misjudgment or omission.

[0017] (2) The interference from direct waves and multiple reflections is serious: the energy of direct waves is often much greater than that of the reflected waves from the hole wall, which can easily lead to misjudgment;

[0018] (3) Poor continuity of borehole wall profile: Traditional methods analyze the echo signal at each depth point independently, failing to fully utilize the continuity and gradation of borehole wall morphology in the depth direction;

[0019] (4) Low level of automation and intelligence, strong subjectivity: The identification of the reflection points on the hole wall relies heavily on the experience of the testers for manual interpretation, which is inefficient and the interpretation results of different people may differ, affecting the consistency and objectivity of the test results. Summary of the Invention

[0020] To address the shortcomings of the existing technologies, this invention proposes an intelligent extraction method and system for hole wall morphology in ultrasonic hole forming quality inspection. This method can ensure the continuity of hole wall morphology in the depth direction and high-precision identification under complex working conditions, solving the problems of weak anti-interference, poor continuity, and low intelligence in the existing technologies.

[0021] To achieve the above objectives, the specific technical solution of the present invention is as follows:

[0022] In a first aspect, the present invention provides a method for intelligent extraction of hole wall morphology in ultrasonic hole forming quality inspection, comprising:

[0023] The original ultrasonic waveforms of the multiple channels at the current measurement point are preprocessed to obtain the preprocessed waveforms for each channel.

[0024] For each channel, the corresponding preprocessed waveform is cross-correlated with a preset transmitted wave template, and all candidate peaks are determined based on the calculation results; candidate peaks located in the direct wave region or with pulse window energy lower than the direct wave energy threshold are removed.

[0025] From the candidate peaks after elimination, the position of the first wave is determined based on the position of the first wave at historical measurement points;

[0026] The preprocessed waveform is input into a pre-trained deep learning model. If the confidence level of the model output is greater than the preset confidence threshold, the predicted position of the model output and the candidate first wave position are fused together according to the confidence level to obtain the final first wave position. Otherwise, the candidate first wave position is used as the final first wave position.

[0027] Calculate the hole wall distance at the current measuring point based on the final first wave position of each channel;

[0028] Repeat the above process for the measuring points at the next depth until the hole wall distances for all measuring points are obtained;

[0029] The hole wall distances at all measuring points are smoothed to obtain the hole wall morphology.

[0030] Furthermore, determining the candidate first wave position from the candidate peaks after elimination, based on the first wave position of historical measurement points, includes:

[0031] If there are no historical measurement points before the current measurement point, the sampling point location corresponding to the peak with the largest cross-correlation value among the candidate peaks after the elimination is adopted as the candidate first wave location;

[0032] If the number of historical measurement points before the current measurement point is less than the preset number, then the candidate peak that is closest to the first wave position of the previous depth measurement point is taken as the candidate first wave position.

[0033] If the number of historical measurement points before the current measurement point is greater than or equal to a preset number, then from the candidate peaks after the elimination, the candidate peak closest to the first wave position of the previous depth measurement point is determined; and the trend prediction position is obtained by fitting the first wave positions of multiple historical measurement points. If the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak closest to the first wave position of the previous depth measurement point is used as the candidate first wave position.

[0034] Further, determining the candidate peak closest to the first wave position of the previous depth measurement point from the candidate peaks after elimination includes:

[0035] Calculate the sampling point distance between each candidate peak after elimination and the first wave position of the previous depth measurement point;

[0036] When the distance between the sampling points is less than the pulse length, the distance between the sampling points is used as the distance value;

[0037] When the distance between the sampling points is greater than or equal to the pulse length, the distance value is determined according to the following formula:

[0038] ;

[0039] in, Indicates the distance value; Indicates the sampling point position of the current measuring point; Indicates the position of the first wave at the previous depth measurement point; Indicates pulse length;

[0040] The candidate peak with the smallest distance value after elimination is selected as the closest candidate peak.

[0041] Further, the trend prediction position is obtained by fitting the first wave position based on multiple historical measurement points. If the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak with the closest distance is used as the candidate first wave position, including:

[0042] Linear fitting is performed using the first wave positions of a preset number of historical measuring points before the current measuring point to obtain a fitted straight line, and extrapolation is performed to determine the trend prediction position of the current measuring point.

[0043] Calculate the root mean square of the deviation between the first wave position of each historical measurement point and the fitted straight line to determine the fitting error; if the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak with the closest distance is used as the candidate first wave position.

[0044] Furthermore, the preprocessing includes: normalizing the original ultrasonic waveform and performing Gaussian filtering on the normalized waveform to obtain the preprocessed waveform.

[0045] Furthermore, for each channel, the corresponding preprocessed waveform is cross-correlated with a preset transmitted wave template, and all candidate peaks are determined based on the calculation results, including:

[0046] For each channel, the corresponding preprocessed waveform is cross-correlationly calculated with the preset transmitted wave template at each sampling point to obtain the cross-correlation value at each sampling point and generate a cross-correlation sequence.

[0047] From the cross-correlation sequence, peaks with cross-correlation values ​​greater than a preset threshold are selected, and peaks with a distance less than or equal to half the pulse length between them and any other selected peaks are removed. The sampling point positions corresponding to the final retained peaks are taken as candidate peaks.

[0048] Furthermore, the direct wave region is determined in the following manner:

[0049] Based on the equivalent direct path length of the probe and the sound velocity in the mud, the two-way propagation time of the direct wave can be calculated using the following formula:

[0050] ;

[0051] in, Indicates the two-way propagation time of the direct wave; Indicates the equivalent direct path length of the probe; Indicates the speed of sound in mud;

[0052] The product of the obtained sampling rate and the two-way propagation time of the direct wave is rounded down to obtain the cutoff sampling point. The region that is less than or equal to the cutoff sampling point is determined as the direct wave region.

[0053] Furthermore, the predicted location output by the model is fused with the candidate first-wave location according to the following formula, weighted by confidence level:

[0054] ;

[0055] in, This represents the result of the weighted fusion calculation; Indicates confidence level; This indicates the predicted location output by the model; This indicates the initial position of the candidate.

[0056] Furthermore, the pre-trained deep learning model is obtained in the following way:

[0057] Measured ultrasonic signals were collected from multiple cast-in-place piles, and the actual first wave position was marked to construct training samples;

[0058] The training samples are augmented with data, and the augmented training samples are divided into a training set, a validation set, and a test set.

[0059] For the samples in the training set, Gaussian distributed soft labels are constructed with the actual first wave location as the center;

[0060] A one-dimensional convolutional neural network is constructed and trained based on a weighted joint loss function, which includes position loss and confidence loss. Training is stopped when the validation set loss does not decrease for a preset number of consecutive rounds, and the model parameters at which the validation loss is lowest are saved, thus obtaining the pre-trained deep learning model.

[0061] Secondly, the present invention provides an intelligent extraction system for hole wall morphology in ultrasonic hole forming quality inspection, characterized in that it includes:

[0062] The signal preprocessing module is used to preprocess the original ultrasonic waveforms of the multiple channels at the current measurement point to obtain the preprocessed waveforms for each channel.

[0063] The candidate peak generation module is used to perform cross-correlation calculation between the corresponding preprocessed waveform and the preset transmitted wave template for each channel, and determine all candidate peaks based on the calculation results;

[0064] The candidate peak filtering module is used to remove candidate peaks located within the direct wave region or whose pulse window energy is lower than the direct wave energy threshold.

[0065] The peak location module is used to determine the position of the first wave from the candidate peaks after they have been eliminated, based on the position of the first wave of historical measurement points.

[0066] The deep learning correction module is used to input the preprocessed waveform into a pre-trained deep learning model. If the confidence of the model output is greater than the preset confidence threshold, the predicted position of the model output and the candidate first wave position are fused together according to the confidence weight to obtain the final first wave position; otherwise, the candidate first wave position is used as the final first wave position.

[0067] The distance calculation module is used to calculate the hole wall distance of the current measuring point based on the final first wave position of each channel;

[0068] The loop module is used to call the above module again for the measurement points at the next depth and execute the corresponding steps until the hole wall distance of all measurement points is obtained;

[0069] The morphology generation module is used to smooth the distance between the borehole walls at all measuring points to obtain the borehole wall morphology.

[0070] Compared with the prior art, the advantages of the present invention are:

[0071] This invention proposes an intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection. Through a multi-level fusion mechanism of cross-correlation detection, energy screening, historical constraints, trend prediction, and deep learning correction, it effectively suppresses noise interference and achieves adaptive recognition and suppression of direct waves, ensuring that the detection results remain spatially continuous and avoiding jumps, thereby improving the accuracy, stability, and reliability of hole wall contour recognition. At the same time, by introducing a deep learning model to intelligently correct the detection results, the system has self-learning capabilities and can continuously optimize the detection accuracy as data accumulates.

[0072] By performing amplitude normalization and Gaussian smoothing preprocessing on the original waveform, amplitude differences between different measurement points are eliminated and high-frequency noise is suppressed. Then, the preprocessed waveform is cross-correlated with a preset transmitted wave template, and candidate peaks are extracted using an overall waveform similarity metric, rather than relying on single-point amplitude determination with a fixed threshold. This addresses the problems of weak anti-interference capability and easy misjudgment / missed detection in traditional methods under low signal-to-noise ratio conditions: improving the robustness of identifying the true borehole wall primary reflection signal when mud properties are poor, borehole walls are rough, or complex reflections exist, and reducing borehole diameter measurement distortion and defect misjudgment caused by clutter and interference signals.

[0073] By employing a dual screening method based on the direct wave region and pulse window energy, the high-energy direct wave and irrelevant multiple reflection signals are effectively separated and suppressed, preventing them from masking or confusing the true aperture wall reflection signal and significantly improving the accuracy of the first wave pickup.

[0074] By using the distance constraint based on the first wave position of the previous depth measurement point and the fitting trend prediction based on the first wave positions of multiple historical measurement points, and adaptively selecting candidate first wave positions based on the fitting error, the continuity and gradation of the borehole wall morphology in the depth direction are fully utilized to improve the problem of poor continuity in borehole wall profile extraction. This further breaks through the limitations of traditional single-point independent analysis, using the continuity and gradation of the borehole wall morphology in the depth direction as a constraint to improve the smoothness and rationality of the extracted borehole wall morphology.

[0075] By inputting the preprocessed waveform into a pre-trained deep learning model, the reliability of the prediction results is evaluated using the confidence level of the model output. When the confidence level is high, the model's predicted position is weighted and fused with the regularized candidate first wave position. When the confidence level is low, the model automatically reverts to the regularized result, thus realizing the intelligent and automated first wave picking, reducing the reliance on human experience. Combined with the deep learning model, intelligent and consistent identification of the hole wall reflection point is achieved, improving detection efficiency and the objectivity of the results. Attached Figure Description

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0077] Figure 1 This is a flowchart of the intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection provided in an embodiment of the present invention;

[0078] Figure 2 This is a schematic diagram of the intelligent extraction system for hole wall morphology in ultrasonic hole forming quality inspection provided in an embodiment of the present invention. Detailed Implementation

[0079] To enable those skilled in the art to clearly and completely understand the technical solution of the present invention, the present invention will be further described in detail below with reference to embodiments. Obviously, the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0080] Existing technologies rely on single-point independent picking of the first wave with a fixed threshold, which has shortcomings such as weak anti-interference, easy misjudgment of direct waves, poor contour continuity, and strong dependence on manual intervention. Based on this, this invention proposes an intelligent extraction method and system for hole wall morphology in ultrasonic hole forming quality inspection. Through a multi-level fusion mechanism of cross-correlation detection, energy screening, historical constraints, trend prediction, and deep learning correction, it effectively suppresses noise interference, achieves adaptive recognition and suppression of direct waves, keeps the detection results spatially continuous, avoids jumps, and improves the accuracy, stability, and reliability of hole wall contour recognition. Simultaneously, by introducing a deep learning model to intelligently correct the detection results, the system possesses self-learning capabilities and can continuously optimize detection accuracy as data accumulates.

[0081] Example 1

[0082] This invention proposes an intelligent extraction method for borehole wall morphology in ultrasonic borehole quality inspection. During inspection, a four-channel orthogonal ultrasonic probe is used to collect ultrasonic signals at preset intervals along the depth direction of the pile hole. For each depth measurement point, the four channels simultaneously transmit and receive sound waves, acquiring the original ultrasonic waveforms in four directions at that measurement point. Subsequently, the original ultrasonic waveforms of each channel are sequentially processed as follows: signal preprocessing, transmitted wave template cross-correlation detection and candidate peak extraction, direct wave suppression and energy screening, multi-strategy peak localization, deep learning intelligent correction, and borehole wall distance calculation. This yields the borehole wall distances corresponding to each channel at that measurement point, and the radial distance of the borehole wall at the current measurement point is calculated accordingly. After processing the current measurement point, the next measurement point is processed along the depth direction, and the above steps are repeated until data processing is completed for all measurement points within the entire pile hole depth range, obtaining the borehole wall distances for all measurement points arranged along the depth direction. Finally, the borehole wall distances are smoothed to obtain a continuous borehole wall morphology. (See reference...) Figure 1 The method in this embodiment specifically includes the following steps:

[0083] S1. Signal Preprocessing: The original ultrasonic waveforms of the multi-channel signals at the current measurement point are preprocessed to obtain the preprocessed waveforms for each channel. Since existing technologies typically use four-channel orthogonal ultrasonic probes for detection, the number of channels is four here. Existing technologies usually only perform simple amplitude correction and basic filtering on the original ultrasonic waveforms, failing to effectively eliminate amplitude differences between different measurement points and channels caused by fluctuations in transmitted energy and differences in probe coupling states. They also lack targeted noise suppression methods, making it difficult to uniformly apply subsequent threshold determinations, and high-frequency noise easily generates spurious peaks in the reflected wave pickup stage. Therefore, this embodiment performs two-stage processing on the original ultrasonic waveform: normalization and Gaussian smoothing. Specifically, this may include the following steps:

[0084] S11. Normalize the original ultrasonic waveform amplitude. Let the original ultrasonic waveform sampling sequence of a certain channel probe be... for:

[0085] (2)

[0086] In the above formula (2), This indicates the number of sampling points, typically 1024 or 2048.

[0087] In actual testing, the amplitude of the original ultrasonic waveform at different measuring points and channels may vary significantly due to fluctuations in emitted energy, differences in probe coupling, and changes in mud attenuation characteristics. Therefore, directly applying a fixed threshold during subsequent processing would be difficult to apply uniformly. Thus, the original ultrasonic waveform is normalized:

[0088] (3)

[0089] In the above formula (3), This represents the normalized waveform; This represents the original ultrasonic wave sampling sequence.

[0090] S12. Apply Gaussian filtering to the normalized waveform to obtain the preprocessed waveform. Utilizing the weighted averaging properties of the Gaussian kernel, high-frequency noise is effectively suppressed while preserving the original shape of the reflected wave envelope to the greatest extent possible.

[0091] (4)

[0092] In the above formula (4), This represents the waveform after preprocessing. Indicates the size of the filter window, for example, Set to 50; Indicates window offset, with a value range of - ~ ; This represents the standardized waveform; The Gaussian kernel function is expressed as follows:

[0093] (5)

[0094] In the above formula (5), This represents the Gaussian kernel standard deviation, used to control the degree of smoothing. For example, Set to 5.0.

[0095] Through the above two-stage preprocessing, on the one hand, the amplitude differences between different measurement points and different channels are eliminated, providing a unified amplitude benchmark for subsequent cross-correlation threshold screening and avoiding threshold misjudgment caused by varying signal strength; on the other hand, Gaussian smoothing effectively suppresses high-frequency noise interference and avoids excessive damage to the leading edge characteristics of the reflected wave by traditional hard cutoff filters, thereby improving the reliability of first wave pickup under low signal-to-noise ratio conditions.

[0096] S2. Based on the cross-correlation detection of the emitted wave template, all possible first waves are initially screened out to obtain candidate peaks.

[0097] Existing technologies for picking up reflected waves typically use a fixed threshold to determine the position of the first wave by directly judging the original ultrasonic waveform or the waveform after simple filtering, finding a single point amplitude peak. This method relies only on local amplitude information and does not utilize the prior waveform characteristics of the ultrasonic probe's emitted pulse. This leads to the easy misjudgment of random noise peaks or clutter peaks as the first wave of the reflected wave when there is strong noise interference, resulting in weak anti-interference capability. Therefore, in step S2, for each channel, the corresponding preprocessed waveform is cross-correlated with a preset emitted wave template, and all candidate peaks are determined based on the calculation results. Specifically, this may include the following steps:

[0098] S21. For each channel, perform cross-correlation calculation on the corresponding preprocessed waveform and the preset transmitted wave template point by point to obtain the cross-correlation value of each sampling point and generate a cross-correlation sequence.

[0099] In step S21 above, the actual transmitted waveform of the ultrasonic probe is approximately a multi-period sine wave with rising and falling envelope characteristics. To fully utilize this prior characteristic, a standard template transmitted wave is generated that matches the probe's transmitted pulse.

[0100] (6)

[0101] In the above formula (6), A sequence representing the transmitted wave template; Indicates the sampling point number within the template; This indicates the pulse length of the ultrasonic probe.

[0102] The preprocessed waveform With the emitted wave template Perform a sliding cross-correlation operation point by point to obtain the cross-correlation sequence. :

[0103] (7)

[0104] In the above formula (7), This indicates the sampling point location of the current measuring point, with a value range of 0~ ,Right now .

[0105] The physical meaning of sliding cross-correlation is: to transform the emitted wave template waveform after preprocessing Slide up point by point, each time you slide to a position The transmitted wave template and the corresponding segment of the preprocessed waveform are then used. ~ Performing the dot product, we get ; The peak value corresponds to the emitted wave template. waveform after preprocessing The most similar position, which is the possible first wave position, is identified. Compared with existing technologies that directly determine the position based on a fixed threshold of the original ultrasonic amplitude, cross-correlation detection significantly improves the ability to identify the true reflected signal and its noise robustness by measuring the overall similarity of the waveform.

[0106] S22. From the cross-correlation sequence, select the peak values ​​whose cross-correlation values ​​are greater than a preset threshold, that is:

[0107] (8)

[0108] In the above equation (8), This represents the preset peak filtering threshold coefficient, for example. The value is 0.1.

[0109] This condition only retains peak values ​​whose cross-correlation values ​​exceed a preset proportion of the global maximum value, ensuring that the candidate peaks obtained through subsequent screening have sufficient waveform similarity.

[0110] At the same time, peaks whose distance from any other selected peak is less than or equal to half the pulse length should be removed, i.e., the following should be satisfied:

[0111] (9)

[0112] In the above equation (9), and This indicates the sampling point positions corresponding to any two peak values; This represents the pulse length. Since the peak envelope width formed by a single ultrasonic pulse on a cross-correlation sequence is approximately equal to the pulse length... If the spacing constraint of equation (9) is not set, the same reflected wave envelope may be misjudged as multiple independent reflected waves due to noise or waveform distortion. This condition requires that the distance between the sampling points corresponding to any two retained peaks is greater than half the pulse length, to ensure that each retained peak represents a physically independent reflection event and to avoid the same reflected wave being misjudged as multiple different reflected waves.

[0113] The sampling point locations corresponding to the final retained peak values ​​are used as candidate peaks, and all candidate peaks constitute the initial first-wave candidate set. :

[0114] (10)

[0115] In the above formula (10), This indicates the location of the sampling point corresponding to each candidate peak.

[0116] Through the above cross-correlation detection and dual screening, the global search space of 2048 sampling points in the original ultrasonic waveform is narrowed down to a local screening range containing only a few high-confidence candidate peaks. While effectively eliminating a large amount of noise interference, all possible reflection events such as direct waves, hole wall reflected waves, and multiple reflected waves are retained, providing a candidate basis for the accurate identification of the first wave position in subsequent steps.

[0117] S3. Based on energy screening, remove the influence of direct waves and output a candidate set without direct waves: remove candidate peaks that are located in the direct wave region or whose pulse window energy is lower than the direct wave energy threshold.

[0118] Existing technologies for picking up reflected waves typically rely solely on a fixed threshold to find peak values ​​in a single-point waveform, lacking prior assessment of the propagation characteristics of direct waves and failing to perform secondary energy-based screening of candidate peaks. Because direct waves have short propagation paths and low attenuation, their energy is often much greater than that of the aperture wall reflected waves. Furthermore, direct waves arrive earlier, making it difficult for existing technologies to effectively distinguish them from the actual aperture wall reflected signals. This leads to direct wave peaks being easily misidentified as the first wave, resulting in a significant underestimation of the aperture size. Therefore, in step S3, position filtering and energy filtering are performed through direct wave region determination and direct wave energy determination, respectively, to obtain the initial first wave candidate set from S2. The process involves removing direct wave interference and low-energy spurious peaks. Specifically, this may include the following steps:

[0119] S31. Direct Wave Region Determination. When ultrasound propagates in mud, some of the sound energy propagates directly to the receiver along the probe's interior or coupling medium, forming a direct wave. Since the direct wave propagation path is fixed and known, its arrival time can be estimated in advance using the equivalent direct path length of the probe. Based on the equivalent direct path length of the probe and the sound velocity in the mud, the two-way propagation time of the direct wave is calculated using the following formula:

[0120] (11)

[0121] In the above equation (11), Indicates the two-way propagation time of the direct wave; Indicates the equivalent direct path length of the probe; This indicates the speed of sound in mud.

[0122] Rounding down the product of the acquired sampling rate and the two-way propagation time of the direct wave, we obtain the cutoff sampling point (i.e., the cutoff position of the direct wave), which can be expressed as:

[0123] (12)

[0124] In the above formula (12), Indicates the cutoff sampling point; Represents the floor function; Indicates the two-way propagation time of the direct wave; This indicates the sampling rate.

[0125] The region less than or equal to the cutoff sampling point is defined as the direct wave region, i.e., the interval. The direct wave region is defined as follows: since the arrival time of the direct wave is earlier than that of the hole wall reflected wave, and its energy is usually the strongest, the candidate peaks located in this region are the peak values ​​of the direct wave. Therefore, it is necessary to remove the candidate peaks located in the direct wave region.

[0126] S32. Direct Wave Energy Determination. After location filtering based on the direct wave region determination above, there may still be multiple reflected waves or strong clutter waves that overlap with the direct wave time. Therefore, for each candidate peak obtained in S2, a window of half a pulse width is taken before and after it, and the pulse window energy within this window is calculated according to the following formula:

[0127] (13)

[0128] In the above equation (13), Indicates the energy of the pulse window; Indicates the first [waveform] on the preprocessed waveform. The amplitude of each sampling point; This represents the sampling point index, with a value range of [value range missing]. ; Indicates the pulse length.

[0129] The direct wave energy threshold is set as follows:

[0130] (14)

[0131] In the above equation (14), Indicates the direct wave energy threshold; The preset threshold coefficient is, for example, set to 0.35. This value is based on engineering experience and is determined by the physical characteristics that the direct wave has the strongest energy and the reflected wave has a significant energy decrease after propagation attenuation. A threshold coefficient of 0.35 can effectively filter low-energy noise and clutter without mistakenly deleting the actual borehole wall reflected waves. In actual engineering, this coefficient can be adjusted within the range of 0.3 to 0.5 according to different engineering scenarios (mud concentration, borehole wall medium, and probe type).

[0132] S33. Remove candidate peaks located within the direct wave region or with pulse window energy below the direct wave energy threshold to obtain a candidate set without direct waves. , can be represented as:

[0133] (15)

[0134] Through the above dual screening, the problem of excessive direct wave energy masking the real hole wall reflection signal in the existing technology is effectively solved. Direct wave interference and low-energy pseudo-peaks in the initial candidate set generated by S2 are eliminated, providing a data basis for subsequent multi-strategy peak positioning and significantly improving the accuracy of the first wave picking.

[0135] S4. Combining historical nearest distance matching and longitudinal borehole wall trend line, perform multi-strategy peak positioning: From the candidate peaks after elimination, determine the candidate first wave position based on the first wave position of historical measurement points.

[0136] Existing technologies analyze the echo signals from each depth measuring point independently, failing to consider the spatial correlation between adjacent measuring points when picking the first wave position, and also failing to utilize the continuity and gradual change of the borehole wall morphology in the depth direction for constraint. Under complex interference conditions, independent determination of a single point can easily lead to irregular jumps in the first wave position in the depth direction, resulting in poor continuity and insufficient smoothness of the extracted borehole wall morphology profile, making it difficult to accurately reflect the actual morphology of the borehole wall. Therefore, in step S4, from the candidate peaks after the elimination process, candidate first wave positions are determined based on the first wave positions of historical measuring points:

[0137] For the initial measuring point at the top of the pile hole, i.e. there are no historical measuring points before the current measuring point, since there is no first wave position of the previous depth measuring point and there is no premise for trend fitting of multiple historical measuring points, the sampling point position corresponding to the peak with the largest cross-correlation value among the candidate peaks after the elimination is directly used as the candidate first wave position of the initial measuring point.

[0138] If the number of historical measurement points before the current measurement point is less than the preset number, then the candidate peak that is closest to the first wave position of the previous depth measurement point is taken as the candidate first wave position.

[0139] If the number of historical measuring points preceding the current measuring point is greater than or equal to a preset number, extrapolation prediction is performed by combining the distance constraints of adjacent historical measuring points with the fitted curves of multiple historical measuring points. The candidate first wave position is then adaptively selected based on the fitting error, fully utilizing the continuity of the borehole wall morphology in the depth direction. This process may specifically include the following steps:

[0140] S41. From the candidate peaks eliminated in step S3, determine the candidate peak whose distance is closest to the first wave position of the previous depth measuring point. Since the borehole wall is a continuous physical entity, the radial distance between adjacent depth measuring points will not change abruptly. Therefore, based on this physical characteristic, the first wave position already determined at the previous depth measuring point can be used as a reference for distance constraint.

[0141] S411. Calculate the sampling point distance between each candidate peak after elimination and the first wave position of the previous depth measurement point;

[0142] S412. When the distance between sampling points is less than the pulse length, the distance between sampling points shall be used as the distance value.

[0143] S413. When the distance between sampling points is greater than or equal to the pulse length, the distance value is determined according to the following formula:

[0144] (16)

[0145] In the above equation (16), Indicates the distance value; Indicates the sampling point position of the current measuring point; Indicates the position of the first wave at the previous depth measurement point; Indicates the pulse length.

[0146] Combining the above step S412, the distance value of each candidate peak can be expressed as:

[0147] (17)

[0148] S414. The candidate peaks with the smallest distance values ​​after elimination are selected as the closest candidate peaks, i.e., the candidate set of no-direct-wave peaks obtained from step S3 above. In the middle, select distance The most recent peak point, the closest candidate peak It can be represented as:

[0149] (18)

[0150] S42. Based on the fitting of the first wave position of multiple historical measurement points, the trend prediction position is obtained. If the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak with the closest distance is used as the candidate first wave position.

[0151] S421. Using the first wave positions of K historical measuring points previously referenced by the current measuring point, perform linear fitting to obtain a fitted straight line, and extrapolate to determine the predicted trend position of the current measuring point. The fitted straight line can be expressed as:

[0152] (19)

[0153] In the above equation (19), Indicates the first The initial location of each historical measurement point; , Represents the linear fitting coefficients, obtained by using the least squares method. The data was obtained by fitting the data from historical measurement points.

[0154] Current measurement point trend prediction position for:

[0155] (20)

[0156] S422. Calculate the root mean square of the deviation between the first wave position and the fitted straight line at each historical measurement point according to the following formula (21). Determine the fitting error:

[0157] (twenty one)

[0158] In the above equation (21), Indicates the first The first wave position of each historical measurement point (the confirmed historical data); K represents the number of historical measurement points participating in the fitting.

[0159] If the fitting error is less than the fitting error threshold, it indicates that the first wave position of the first K historical measurement points shows a stable linear trend, and the borehole wall is relatively smooth in this section. The trend prediction position is used as the candidate first wave position to ensure the continuity and smoothness of the borehole wall morphology in the depth direction. If the fitting error is greater than or equal to the fitting error threshold, it indicates that the historical data fluctuates greatly and the predicted trend is unreliable. The candidate peak with the closest distance is used as the candidate first wave position. The candidate first wave position can be expressed as:

[0160] (twenty two)

[0161] In the above equation (22), Indicates the candidate first wave position of the current measuring point; This represents the fitting error threshold, for example, The value is set to 0.66K, which is an empirical engineering threshold, not a strictly statistical theoretical value. Its selection is based on the fact that K represents the number of historical measurement points involved in the fitting (for example, K is set to 50, which represents the fitting error threshold). (With 33 sampling points), this threshold sets a reasonable upper limit for the allowable fluctuations in historical data. When the fitting error is less than this upper limit, the historical trend is considered stable and reliable; otherwise, the trend is considered unreliable. In actual engineering, this threshold can be adjusted within a reasonable range according to the borehole wall geological conditions, mud properties, and detection accuracy requirements.

[0162] The multi-strategy peak localization method effectively solves the problems of poor continuity of borehole wall profile and easy random jumps caused by single-point independent analysis in existing technologies. By using the physical continuity of borehole wall morphology for local distance constraints and combining it with historical trend prediction for adaptive fusion, the overall smoothness of borehole wall morphology is ensured while retaining the ability to identify local real morphological changes (such as diameter expansion and contraction), which significantly improves the rationality of subsequent initial localization and the accuracy of borehole wall morphology reconstruction.

[0163] S5. Weighted correction is performed based on the results of deep learning intelligent picking. Existing technologies primarily rely on the experience of detection personnel for manual interpretation during the initial wave picking, or employ a semi-automatic method with fixed thresholds. This is inefficient and highly subjective, affecting the consistency and objectivity of the detection results. Therefore, this embodiment inputs the preprocessed waveform into a pre-trained deep learning model. Utilizing the probability distribution of the reflected wave position and the confidence score output by the model, the initial wave prediction result is weighted and corrected at high confidence levels, and automatically reverts to the candidate initial wave position determined in step S4 at low confidence levels.

[0164] The deep learning model used in this embodiment (the specific training method of the model will be explained in subsequent examples) is a one-dimensional convolutional neural network with a dual-branch output structure. The trained deep learning model is defined as follows: The model input is a preprocessed waveform from a single channel. The output is the probability distribution of the reflected wave location and the confidence score, which can be expressed as:

[0165] (twenty three)

[0166] In the above equation (23), The probability distribution representing the location of the first wave of the reflected wave. Indicates the sampling point number. ; Indicates the confidence level. ; This represents the preprocessed waveform determined in step S1 above.

[0167] The sampling point location corresponding to the highest probability is taken as the first wave location of the reflected wave predicted by the deep learning model:

[0168] (twenty four)

[0169] In the above equation (24), This indicates the sampling point number corresponding to the position of the first wave of the reflected wave predicted by the deep learning model.

[0170] Based on the predicted position and confidence level output by the deep learning model, the following fusion strategy is executed using the confidence level: If the confidence level output by the model is greater than the preset confidence threshold, it indicates that the model has high credibility in recognizing the current waveform. Then, the predicted position output by the model and the candidate first wave position are fused together according to the confidence level to obtain the final first wave position. The weighted fusion calculation is performed according to the following formula (25). The deep learning method is fully utilized to perform deep semantic feature mining on the waveform, while retaining the candidate first wave position determined in step S4 above, making the final first wave position more accurate.

[0171] (25)

[0172] In the above equation (25), This represents the result of the weighted fusion calculation, i.e., the final first wave position; Indicates confidence level; This indicates the predicted location output by the model; This indicates the initial position of the candidate.

[0173] If the confidence level of the model output is less than or equal to the preset confidence threshold, it indicates that the model lacks sufficient confidence in the recognition result of the current waveform. In this case, the model prediction result is abandoned, and the candidate first wave position is taken as the final first wave position.

[0174] Therefore, the final position of the first wave can be represented as:

[0175] (26)

[0176] In the above equation (26), This indicates a preset reliability threshold. Preferably, the preset reliability threshold ranges from 0.8 to 1.0 (which can be adjusted according to the actual engineering data distribution and accuracy requirements). For example, it can be set to 0.9.

[0177] S6. Hole wall distance calculation and global smoothing, output hole wall shape.

[0178] First, based on the final first wave position of each channel, the hole wall distance at the current measuring point is calculated using the following formula (27):

[0179] (27)

[0180] In the above equation (27), This indicates the distance from the borehole wall to the current measuring point; Indicates the speed of sound in mud; Indicates the round-trip time of sound waves. ,in, Indicates the sampling time interval.

[0181] For the measuring points at the next depth, repeat steps S1-S5 above and calculate the hole wall distance according to the above formula (27) until the hole wall distance of all measuring points is obtained.

[0182] Finally, the distances between all measuring points and the borehole walls are smoothed to obtain the borehole wall morphology. The smoothing method used in this embodiment is Gaussian filtering.

[0183] (28)

[0184] In the above equation (28), Indicates the first The distance between the smoothed hole walls at each depth measuring point; Indicates the first Original hole wall distance before smoothing at each depth measuring point; Let represent the Gaussian kernel function, whose expression is shown in equation (5) above; Indicates window offset, with a value range of - ~ ; Indicates the size of the filtering window.

[0185] For example, the method of the present invention is applied to the quality inspection of bored piles for highway bridges, with a pile diameter of 1.5m, a hole depth of 60m, and a mud density of 1.15g / cm³. 3 A four-channel orthogonal ultrasonic probe was used, and the testing was conducted according to JTG / T 3512-2020 "Technical Specification for Testing of Highway Engineering Foundation Piles". Key parameters were fixed as follows:

[0186] Sampling rate 1MHz;

[0187] Number of sampling points :2048;

[0188] Gaussian filter filter window size The value is 50, and the standard deviation of the Gaussian kernel is 5.0.

[0189] Peak Filtering Threshold Coefficient The value is 0.1;

[0190] Direct wave energy threshold =35% × maximum peak energy;

[0191] During linear fitting, the number of historical measurement points K is set to 50, and the fitting error threshold is... The value is 0.66K;

[0192] Preset confidence threshold for deep learning The value is 0.9;

[0193] Mud sound speed 1450m / s;

[0194] equivalent direct path length of the probe : 0.08m (80mm).

[0195] Deep learning models use one-dimensional convolutional neural networks (1D-CNN) as their basic architecture. The network adopts a two-branch output structure, and the specific layer order is as follows:

[0196] Input layer: dimension (2048, 1), corresponding to the pre-processed waveform of a single channel;

[0197] One-dimensional convolutional layer 1: configured with 32 convolutional kernels, each with a kernel size of 8 and a stride of 1, using ReLU as the activation function;

[0198] One-dimensional convolutional layer 2: configured with 64 convolutional kernels, kernel size of 5, stride of 1, and ReLU activation function;

[0199] Max pooling layer 1: pooling size is 2, step size is 2;

[0200] One-dimensional convolutional layer 3: configured with 128 convolutional kernels, kernel size 3, stride 1, and ReLU activation function;

[0201] Max pooling layer 2: pooling size is 2, step size is 2;

[0202] Flattening layer: Converts the multidimensional feature map after multiple convolutions and pooling into a one-dimensional feature vector, which is then used as the input to the fully connected layer;

[0203] Fully connected layer 1: It is configured with 256 neurons, uses ReLU as the activation function, and sets the Dropout ratio to 0.3 to prevent the model from overfitting on the training data.

[0204] Dual-output branch layer: The network ends in two independent output branches, each undertaking a different function.

[0205] Location branch: Configured with 2048 neurons, the activation function is Softmax, and the output is the probability distribution P(t) of each sampling point as the first wave position of the reflected wave;

[0206] Confidence branch: Configure 1 neuron, use Sigmoid as the activation function, and output confidence score S∈[0,1].

[0207] The training process of a deep learning model may specifically include the following steps:

[0208] First, measured ultrasonic signals from multiple cast-in-place piles were collected, and the actual first wave positions were marked to construct a training sample. In this example, the sample used consisted of measured ultrasonic signals from 100 cast-in-place piles. Senior engineers manually marked the actual first wave positions, creating 120,000 single-channel waveform samples as training data. The field measurement data covered different pile diameters, hole depths, mud properties, and geological conditions, ensuring the diversity and engineering representativeness of the training data.

[0209] Then, to improve the model's generalization ability and robustness to complex working conditions, data augmentation is performed on the training samples. Specifically, this may include: random scaling of amplitude by ±10% to simulate different probe coupling states and emission energy fluctuations; superimposing Gaussian noise (signal-to-noise ratio 5~20dB) to simulate low signal-to-noise ratio engineering environments; slight shift of the time axis by ±5 sampling points to simulate sampling clock jitter; and slight translation of the waveform to simulate the propagation time shift caused by changes in hole depth.

[0210] The augmented training samples are divided into training, validation, and test sets. In this example, 120,000 samples are divided into a training set (84,000 samples, 70% of the training samples), a validation set (24,000 samples, 20% of the training samples), and a test set (12,000 samples, 10% of the training samples). For the samples in the training set, Gaussian distributed soft labels are constructed centered on the true first wave location, allowing the model to learn higher probabilities near the peak and avoiding single-point hard labels.

[0211] The aforementioned one-dimensional convolutional neural network is constructed and trained based on a weighted joint loss function. This weighted joint loss function includes positional loss (using the cross-entropy loss function) and confidence loss (using the binary cross-entropy loss function), with the total loss being the sum of the positional loss and 0.2 times the confidence loss. TensorFlow 2.x or PyTorch open-source deep learning frameworks can be used as the training framework; the Adam adaptive optimizer is used, with an initial learning rate set to 1e-4; training hyperparameters are set as follows: batch size to 128, and training epochs to 50.

[0212] Early stopping strategy: Monitor the validation set loss. If the validation set loss does not decrease for a preset number of rounds (for example, 8 rounds), stop training and save the model parameters when the validation loss is the lowest, thus obtaining the pre-trained deep learning model.

[0213] Training objective: The location recognition accuracy rate is no less than 98.5%, and the confidence assessment error is no more than 0.05.

[0214] After training, the model is exported as a lightweight model file in .pb or .pth format and embedded into the detection software system to achieve real-time interface with the front-end signal acquisition module.

[0215] The specific implementation process is as follows:

[0216] (1) Perform signal preprocessing according to step S1 above: acquire the original ultrasonic waveform of a single measurement point and four channels. First, the waveform amplitude is normalized to eliminate amplitude differences; then, Gaussian smoothing filtering is used to suppress high-frequency noise caused by mud turbulence and equipment electromagnetic interference, resulting in the preprocessed waveform. This provides clean data for subsequent extraction of reflected waves.

[0217] (2) Perform cross-correlation detection of the transmitted wave template according to step S2 above: generate a standard sinusoidal transmitted template that matches the probe pulse. The preprocessed waveform With launch template Perform cross-correlation operations to obtain cross-correlation sequences Candidate peak values ​​are extracted based on threshold filtering and spacing constraints, and interfering peaks that are too close are removed to obtain the initial first wave of candidate sets. The possible location of the reflected wave was initially identified.

[0218] (3) Perform direct wave energy screening according to step S3 above: first determine the cutoff sampling point of the direct wave based on the equivalent direct path length of the probe, the mud sound velocity and the sampling rate. , demarcate The direct-wave region is defined; the pulse window energy of each candidate peak is then calculated; candidate peaks and low-energy points within the direct-wave region are removed to obtain a candidate set without direct waves. This solves the problem of excessive direct wave energy masking the reflection from the actual aperture wall.

[0219] (4) Perform multi-strategy peak positioning according to step S4 above.

[0220] Recent historical match: Using the position of the first wave at the previous depth measurement point as a reference, select the candidate peak that is closest to it to ensure continuity in the depth direction;

[0221] Linear Fitting Prediction: Using the first wave positions of the previous 50 historical measurement points, a linear fit is performed to predict the theoretical position of the current measurement point, which is then used as the predicted trend position of the current measurement point. ;

[0222] Fusion decision: When the fitting error is small, the trend prediction position is used as the candidate first wave position to ensure a smooth profile; when the error is large, the candidate peak with the closest distance is used as the candidate first wave position to solve the profile jump problem caused by traditional single-point independent analysis.

[0223] (5) Perform deep learning intelligent picking correction according to the above step S5: input the pre-processed waveform into the pre-trained deep learning model and output the probability and confidence of the reflected wave position; when the confidence is ≥0.9, use the prediction result of the model for weighted correction; when it is below 0.9, revert to the result of the previous step to reduce the subjectivity of manual interpretation and improve the recognition accuracy under low signal-to-noise ratio.

[0224] (6) Hole wall distance calculation and global smoothing: Calculate the hole wall distance of the current measuring point according to the above formula (27), and complete the processing of all measuring points according to the above steps (1)-(5) to obtain the hole wall distance of each measuring point; then perform Gaussian filtering on the hole wall distance sequence in the depth direction to output continuous and smooth hole wall shape, hole diameter, verticality and diameter expansion / diameter contraction / hole collapse defect shape.

[0225] The accuracy of the results obtained based on the above method was verified. The accuracy of the hole diameter detection was: error ≤ ±2mm, relative error ≤ 0.13%; the error of the pile hole inclination measurement was: ≤ 0.1%, the profile was continuous without jumps, and there was no local distortion.

[0226] By combining multi-level signal filtering, depth direction continuity constraints, and deep learning correction, a triple mechanism separates clutter, direct waves, and multiple reflections, accurately locking the primary hole wall reflection wave, and significantly improving measurement accuracy and weak signal identification capabilities.

[0227] To verify the technical contribution of each core step in the technical solution of the present invention, under the same detection conditions as the above example, individual core steps were eliminated to form comparative examples, which were then compared with the above example.

[0228] Comparative Example 1: Eliminating S4 multi-strategy peak positioning

[0229] Implementation conditions: Retain steps S1-S3, S5, and S6 as described above, delete the historical recent match and multi-point trend prediction in step S4, and directly use the candidate set of no-direct-wave output from step S3. The sampling point location corresponding to the peak value with the largest cross-correlation value is used as the candidate first wave location of the initial measurement point, and only single-point peak value determination is performed.

[0230] Test results: The borehole diameter detection error is ±8mm~±12mm, and the relative accuracy is about 0.67% (which does not meet the requirement of a relative accuracy of not less than 0.2% for borehole diameter detection in JTG / T3512-2020 standard); the pile hole inclination error is 0.42%; the borehole wall morphology shows obvious jumps, poor continuity in the depth direction, and local distortion.

[0231] Because the continuity constraint based on historical measurement points in the depth direction was removed, the first wave position of each measurement point was determined independently, failing to utilize the continuity and gradation of the borehole wall morphology in the depth direction. When the signal-to-noise ratio is low or complex reflection interference exists, the determination of the single-point peak value is easily affected by noise peaks or sidelobe interference, resulting in irregular jumps in the first wave position between adjacent measurement points and an uneven borehole wall morphology profile.

[0232] Comparative Example 2: Removing S5 Deep Learning Intelligent Correction

[0233] Implementation conditions: retain steps S1-S4 and S6 above, delete the deep learning intelligent correction step S5, and only use the candidate first wave position output by S4 as the final first wave position.

[0234] Test results: The hole diameter detection error is ±5mm~±7mm, and the relative accuracy is about 0.4% (which does not meet the requirement of a relative accuracy of not less than 0.2% for hole diameter detection in JTG / T3512-2020 standard); the pile hole inclination error is 0.3%; some low signal-to-noise ratio measuring points still need to rely on manual experience for fine-tuning and interpretation.

[0235] Results Analysis: Because the deep learning model was removed from mining and correcting the deep semantic features of the waveform, the system relied solely on rule-based algorithms (cross-correlation matching, energy filtering, and historical constraints) for initial wave picking. Although the historical constraints in step S4 ensured the continuity of the contour, under complex conditions such as weak reflected signals, low signal-to-noise ratios, or waveform distortion caused by rough aperture walls, the rule-based algorithms struggled to adequately distinguish between real reflected waves and interference signals, resulting in insufficient recognition accuracy.

[0236] Comparing Comparative Examples 1 and 2, removing either the multi-strategy peak localization or deep learning correction step alone significantly reduces detection accuracy and fails to meet industry standards. Only when historical constraints, trend prediction, and deep learning intelligent correction work in tandem can the depth-direction continuity of the hole wall morphology and high-precision recognition under complex conditions be simultaneously guaranteed, completely resolving the problems of weak anti-interference, poor continuity, and low intelligence in existing technologies. The synergistic combination of multi-strategy peak localization and deep learning correction achieves significantly better overall results than any single technique.

[0237] Compared with existing technologies, this invention effectively suppresses noise interference through a multi-level fusion mechanism including cross-correlation detection, energy screening, historical constraints, trend prediction, and deep learning correction. This enables adaptive identification and suppression of direct waves, ensuring spatial continuity of detection results, avoiding jumps, and improving the accuracy, stability, and reliability of aperture wall morphology recognition. Simultaneously, by introducing a deep learning model to intelligently correct the detection results, the system possesses self-learning capabilities, continuously optimizing detection accuracy as data accumulates. Specifically, this invention achieves the following beneficial effects:

[0238] This invention eliminates amplitude differences between different measurement points and suppresses high-frequency noise by performing amplitude normalization and Gaussian smoothing preprocessing on the original waveform. Then, it performs cross-correlation calculations between the preprocessed waveform and a preset transmitted wave template, extracting candidate peaks using an overall waveform similarity metric, rather than relying on single-point amplitude determination with a fixed threshold. This addresses the problems of weak anti-interference capability and susceptibility to misjudgment / missed detection in traditional methods under low signal-to-noise ratio conditions. It improves the robustness of identifying the true primary reflection signal from the borehole wall when mud properties are poor, borehole walls are rough, or complex reflections exist, reducing borehole diameter measurement distortion and defect misjudgment caused by clutter and interference signals.

[0239] By performing cross-correlation calculations between the preprocessed waveform and a preset transmitted wave template, candidate peaks are extracted using overall waveform similarity matching, rather than relying on fixed thresholds for single-point peak determination. This improves the ability to identify genuine aperture wall reflection signals in complex interference environments, addressing the severe interference issues caused by direct waves and multiple reflections in existing technologies. Dual screening based on the direct wave region and pulse window energy effectively separates and suppresses strong direct waves and irrelevant multiple reflection signals, preventing them from masking or confusing the genuine aperture wall reflection signal, significantly improving the accuracy of first-wave pickup.

[0240] By using the distance constraint based on the first wave position of the previous depth measurement point and the fitting trend prediction based on the first wave positions of multiple historical measurement points, and adaptively selecting candidate first wave positions based on the fitting error, the continuity and gradation of the borehole wall morphology in the depth direction are fully utilized to improve the problem of poor continuity in borehole wall profile extraction. This further breaks through the limitations of traditional single-point independent analysis, using the continuity and gradation of the borehole wall morphology in the depth direction as a constraint to improve the smoothness and rationality of the extracted borehole wall morphology.

[0241] By inputting the preprocessed waveform into a pre-trained deep learning model, the reliability of the prediction results is evaluated using the confidence level of the model output. When the confidence level is high, the model's predicted position is weighted and fused with the regularized candidate first wave position. When the confidence level is low, the model automatically reverts to the regularized result. This achieves intelligent and automated first wave picking, reduces subjective dependence, and reduces reliance on human experience. Combined with the deep learning model, it enables intelligent and consistent identification of the hole wall reflection point, improving detection efficiency and the objectivity of the results.

[0242] Example 2

[0243] Based on the same inventive concept, this invention proposes an intelligent extraction system for hole wall morphology in ultrasonic hole forming quality inspection, see reference. Figure 2 ,include:

[0244] The signal preprocessing module 101 is used to preprocess the original ultrasonic waveforms of the multiple channels at the current measuring point to obtain the preprocessed waveforms for each channel.

[0245] The candidate peak generation module 102 is used to perform cross-correlation calculation between the corresponding preprocessed waveform and the preset transmitted wave template for each channel, and determine all candidate peaks based on the calculation results.

[0246] The candidate peak screening module 103 is used to remove candidate peaks located within the direct wave region or whose pulse window energy is lower than the direct wave energy threshold.

[0247] Peak location module 104 is used to determine the position of the first wave from the candidate peaks after the elimination is completed, based on the position of the first wave of historical measurement points;

[0248] The deep learning correction module 105 is used to input the preprocessed waveform into the pre-trained deep learning model. If the confidence of the model output is greater than the preset confidence threshold, the predicted position of the model output and the candidate first wave position are fused together according to the confidence weight to obtain the final first wave position; otherwise, the candidate first wave position is used as the final first wave position.

[0249] The distance calculation module 106 is used to calculate the hole wall distance of the current measuring point based on the final first wave position of each channel.

[0250] The loop module 107 is used to call the above module again for the measurement points at the next depth and execute the corresponding steps until the hole wall distance of all measurement points is obtained.

[0251] The morphology generation module 108 is used to smooth the hole wall distances of all measuring points to obtain the hole wall morphology.

[0252] The intelligent extraction system for hole wall morphology in ultrasonic hole forming quality inspection provided in this embodiment of the invention has a similar implementation principle and technical effect to that in Embodiment 1, and will not be repeated here.

[0253] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent extraction of hole wall morphology in ultrasonic hole forming quality inspection, characterized in that, include: The original ultrasonic waveforms of the multiple channels at the current measurement point are preprocessed to obtain the preprocessed waveforms for each channel. For each channel, the corresponding preprocessed waveform is cross-correlated with a preset transmitted wave template, and all candidate peaks are determined based on the calculation results. Candidate peaks located in the direct wave region or with pulse window energy lower than the direct wave energy threshold are removed, and the product of the maximum value of the pulse window energy of each candidate peak and the preset threshold coefficient is used as the direct wave energy threshold. From the candidate peaks after elimination, the candidate first wave position is determined based on the first wave position of historical measurement points. Specifically, if there are no historical measurement points before the current measurement point, the sampling point position corresponding to the peak with the largest cross-correlation value among the candidate peaks after elimination is used as the candidate first wave position. If the number of historical measurement points before the current measurement point is less than a preset number, the candidate peak closest to the first wave position of the previous depth measurement point is used as the candidate first wave position. If the number of historical measurement points before the current measurement point is greater than or equal to a preset number, the candidate peak closest to the first wave position of the previous depth measurement point is determined from the candidate peaks after elimination. A trend prediction position is obtained by fitting the first wave positions of multiple historical measurement points. If the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak closest to the first wave position of the previous depth measurement point is used as the candidate first wave position. The preprocessed waveform is input into a pre-trained deep learning model. If the confidence level of the model output is greater than the preset confidence threshold, the predicted position of the model output and the candidate first wave position are fused together according to the confidence level to obtain the final first wave position. Otherwise, the candidate first wave position is used as the final first wave position. Calculate the hole wall distance at the current measuring point based on the final first wave position of each channel; Repeat the above process for the measuring points at the next depth until the hole wall distances for all measuring points are obtained; The hole wall distances at all measuring points are smoothed to obtain the hole wall morphology.

2. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1, characterized in that, The step of determining the candidate peak closest to the first wave position of the previous depth measurement point from the candidate peaks after elimination includes: Calculate the sampling point distance between each candidate peak after elimination and the first wave position of the previous depth measurement point; When the distance between the sampling points is less than the pulse length, the distance between the sampling points is used as the distance value; When the distance between the sampling points is greater than or equal to the pulse length, the distance value is determined according to the following formula: ； in, Indicates the distance value; Indicates the sampling point position of the current measuring point; Indicates the position of the first wave at the previous depth measurement point; Indicates pulse length; The candidate peak with the smallest distance value after elimination is selected as the closest candidate peak.

3. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1 or 2, characterized in that, The trend prediction position is obtained by fitting the first wave position based on multiple historical measurement points. If the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak with the closest distance is used as the candidate first wave position, including: Linear fitting is performed using the first wave positions of a preset number of historical measuring points before the current measuring point to obtain a fitted straight line, and extrapolation is performed to determine the trend prediction position of the current measuring point. Calculate the root mean square of the deviation between the first wave position of each historical measurement point and the fitted straight line to determine the fitting error; if the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak with the closest distance is used as the candidate first wave position.

4. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1, characterized in that, The preprocessing includes: normalizing the original ultrasonic waveform and performing Gaussian filtering on the normalized waveform to obtain the preprocessed waveform.

5. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1, characterized in that, For each channel, the corresponding preprocessed waveform is cross-correlated with a preset transmitted wave template. Based on the calculation results, all candidate peaks are determined, including: For each channel, the corresponding preprocessed waveform is cross-correlationly calculated with the preset transmitted wave template at each sampling point to obtain the cross-correlation value at each sampling point and generate a cross-correlation sequence. From the cross-correlation sequence, peaks with cross-correlation values ​​greater than a preset threshold are selected, and peaks with a distance less than or equal to half the pulse length between them and any other selected peaks are removed. The sampling point positions corresponding to the final retained peaks are taken as candidate peaks.

6. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1, characterized in that, The direct wave region is determined in the following manner: Based on the equivalent direct path length of the probe and the sound velocity in the mud, the two-way propagation time of the direct wave can be calculated using the following formula: ； in, Indicates the two-way propagation time of the direct wave; Indicates the equivalent direct path length of the probe; Indicates the speed of sound in mud; The product of the obtained sampling rate and the two-way propagation time of the direct wave is rounded to the nearest integer to obtain the cutoff sampling point. The region that is less than or equal to the cutoff sampling point is determined as the direct wave region.

7. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1, characterized in that, The predicted location output by the model is fused with the candidate first-wave location according to the following formula, weighted by confidence level: ； in, This indicates the result of the weighted fusion calculation; Indicates confidence level; This indicates the predicted location output by the model; This indicates the initial position of the candidates.

8. The intelligent extraction method for hole wall morphology in ultrasonic hole forming quality inspection according to claim 1, characterized in that, The pre-trained deep learning model is obtained in the following way: Measured ultrasonic signals were collected from multiple cast-in-place piles, and the actual first wave position was marked to construct training samples; The training samples are augmented with data, and the augmented training samples are divided into a training set, a validation set, and a test set. For the samples in the training set, Gaussian distributed soft labels are constructed with the actual first wave location as the center; A one-dimensional convolutional neural network is constructed and trained based on a weighted joint loss function, which includes position loss and confidence loss. Training is stopped when the validation set loss does not decrease for a preset number of consecutive rounds, and the model parameters at which the validation loss is lowest are saved, thus obtaining the pre-trained deep learning model.

9. A smart extraction system for hole wall morphology in ultrasonic hole forming quality inspection, characterized in that, include: The signal preprocessing module is used to preprocess the original ultrasonic waveforms of the multiple channels at the current measurement point to obtain the preprocessed waveforms for each channel. The candidate peak generation module is used to perform cross-correlation calculation between the corresponding preprocessed waveform and the preset transmitted wave template for each channel, and determine all candidate peaks based on the calculation results; The candidate peak screening module is used to remove candidate peaks that are located within the direct wave region or whose pulse window energy is lower than the direct wave energy threshold. The direct wave energy threshold is the product of the maximum value of the pulse window energy of each candidate peak and a preset threshold coefficient. The peak location module is used to determine the candidate first wave position from the candidate peaks after elimination, based on the first wave position of historical measurement points. Specifically, if there are no historical measurement points before the current measurement point, the sampling point position corresponding to the peak with the largest cross-correlation value among the candidate peaks after elimination is used as the candidate first wave position; if the number of historical measurement points before the current measurement point is less than a preset number, the candidate peak closest to the first wave position of the previous depth measurement point is used as the candidate first wave position; if the number of historical measurement points before the current measurement point is greater than or equal to a preset number, the candidate peak closest to the first wave position of the previous depth measurement point is determined from the candidate peaks after elimination; and a trend prediction position is obtained by fitting the first wave positions of multiple historical measurement points. If the fitting error is less than the fitting error threshold, the trend prediction position is used as the candidate first wave position; otherwise, the candidate peak closest to the first wave position of the previous depth measurement point is used as the candidate first wave position. The deep learning correction module is used to input the preprocessed waveform into a pre-trained deep learning model. If the confidence of the model output is greater than the preset confidence threshold, the predicted position of the model output and the candidate first wave position are fused together according to the confidence weight to obtain the final first wave position; otherwise, the candidate first wave position is used as the final first wave position. The distance calculation module is used to calculate the hole wall distance of the current measuring point based on the final first wave position of each channel; The loop module is used to call the above module again for the measurement points at the next depth and execute the corresponding steps until the hole wall distance of all measurement points is obtained; The morphology generation module is used to smooth the distance between the borehole walls at all measuring points to obtain the borehole wall morphology.

Images