A method and system for identifying defective sections of a cast-in-place pile

Abstract

The present application relates to geotechnical engineering and infrastructure engineering informatization field, specifically to a bored pile defect section identification method and system. The method comprises: obtaining historical construction data, and preprocessing the historical construction data to obtain historical construction multi-source data; dividing the historical construction multi-source data, and extracting features from the divided results to obtain a plurality of segment-level feature data; obtaining defect data by low strain detection on the pile, and matching the segment-level feature data with the defect data to obtain training samples; constructing a defect prediction model based on the training samples, and identifying the bored pile defect section using the defect prediction model. The present application solves the problems of detection lag, limited sampling coverage and difficulty in timely intervention of construction quality risk in the prior art. By correlating construction process data and pile detection waveform information, the present application improves the timeliness and accuracy of pile foundation engineering quality control.

Description

Technical Field

[0001] This invention relates to the field of information technology in geotechnical engineering and infrastructure engineering, specifically a method and system for identifying defective sections in bored piles. Background Technology

[0002] Drilled piles are widely used in municipal, building, and transportation infrastructure projects due to their advantages such as low construction noise, wide adaptability to geological formations, and high single-pile bearing capacity. However, the construction of drilled piles is characterized by its inherent concealment and procedural nature. Abnormalities in any stage, from drilling and hole cleaning to cage placement and concrete pouring, can lead to pile quality defects, thereby affecting load-bearing capacity and structural safety. Furthermore, current defect detection in drilled piles primarily relies on post-construction non-destructive testing to determine pile quality, such as low-strain reflection wave testing and ultrasonic transmission methods. While these methods are relatively mature in identifying pile defects, they suffer from issues such as detection lag and limited sampling coverage, making it difficult to promptly detect and intervene in quality risks during construction.

[0003] With the development of IoT and smart construction technologies, construction sites can now collect multi-source process data in real time, such as drilling depth, torque, current, inclination angle, and hole position deviation. However, in existing technologies, this process data is mostly used for equipment status or progress management, lacking a unified alignment mechanism and effective correlation modeling with pile inspection waveform results, making it difficult to transform it into early judgments of defect types and defect sections. Therefore, there is an urgent need for a method based on a unified benchmark to accurately align process data with waveform defect information, construct supervised samples that can be used for model training, and combine them with an identification model suitable for on-site deployment to achieve online identification and prediction of defect sections.

[0004] Therefore, there is an urgent need for a method to identify defective sections in bored piles that organically combines construction process data with pile inspection waveform defect information, in order to improve the overall quality control level of pile foundation engineering. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for identifying defective sections in bored piles. This solves the quality risk problem in the construction process of bored piles, which is difficult to detect and intervene in a timely manner in existing technologies.

[0006] To achieve the above objectives, one aspect of the present invention provides a method for identifying defective sections of bored piles. The method includes: acquiring historical construction data and preprocessing the historical construction data to obtain historical construction multi-source data; dividing the historical construction multi-source data and extracting features from the division results to obtain multiple segment-level feature data; obtaining defect data by performing low-strain detection on the pile and matching the segment-level feature data with the defect data to obtain training samples; constructing a defect prediction model based on the training samples and using the defect prediction model to identify defective sections of bored piles.

[0007] This invention preprocesses historical construction data, extracts segment-level features, and combines defect data obtained from low-strain detection to construct training samples and establish a prediction model. It can effectively correlate construction process data with pile inspection waveform information, solving the problems of lagging detection, limited sampling coverage, and difficulty in timely intervention of construction quality risks in existing technologies. It enables online identification and prediction of defective sections of bored piles, improving the timeliness and accuracy of pile foundation engineering quality control.

[0008] Optionally, the step of dividing the historical construction multi-source data and extracting features from the division results to obtain multiple segment-level feature data includes: traversing and sampling the historical construction multi-source data with a predetermined step size based on a pre-set sliding window to obtain multiple multi-source data sub-segments divided according to pile length; and extracting features from the multi-source data sub-segments to obtain multiple segment-level feature data.

[0009] By traversing and sampling historical construction multi-source data through a preset sliding window and a predetermined step size, the multi-source data sub-segments can be accurately divided according to the pile length, avoiding the randomness and one-sidedness of data division, ensuring coverage of key depth segments of the pile body, and then extracting features from the sub-segments, which can specifically capture the core information of construction parameters for each depth segment, making the segment-level feature data more consistent with the actual construction conditions and quality correlation, thus improving the practicality and accuracy of extracting information from historical construction multi-source data.

[0010] Optionally, obtaining defect data by performing low-strain testing on the pile includes: obtaining an original waveform by performing low-strain testing on the pile; performing mean removal and low-order polynomial detrending processing on the original waveform to obtain a first preprocessed waveform; applying a bandpass filter operator to the first preprocessed waveform to obtain a second preprocessed waveform; identifying abnormal reflection events on the second preprocessed waveform to determine abnormal reflection waveforms; performing time-depth transformation on the abnormal reflection waveforms to obtain segment defect data; obtaining a defect category and confidence score based on the segment defect data; and obtaining defect data based on the defect category, the confidence score, and the segment defect data.

[0011] In this embodiment, after obtaining the original waveform through low-strain detection, the waveform undergoes mean removal, low-order polynomial detrending processing, and bandpass filtering to effectively eliminate DC components, slow trend terms, and noise interference, thereby improving the quality of the waveform signal. Furthermore, abnormal reflection event identification and time-depth conversion are used to accurately locate defect segments. Combined with segment data, the defect category and confidence score are determined, thus improving the reliability of the defect data.

[0012] Optionally, the step of identifying abnormal reflection events in the second preprocessed waveform and determining the abnormal reflection waveform includes: calculating the short-time average amplitude, long-time average amplitude, and envelope amplitude using the second preprocessed waveform; and determining the abnormal reflection waveform based on the ratio of the short-time average amplitude to the long-time average amplitude, combined with the envelope amplitude.

[0013] This invention calculates the short-time average amplitude, long-time average amplitude, and envelope amplitude using a second preprocessed waveform. It then combines the ratio of the first two to the envelope amplitude to determine the abnormal reflection waveform, enabling precise capture of abnormal signals within the waveform. This avoids the limitations of relying on a single parameter and effectively eliminates noise interference, thus improving the accuracy of abnormal reflection identification.

[0014] Optionally, the step of performing time-depth conversion on the abnormal reflection waveform to obtain the segment defect data includes: determining the excitation reference time based on the maximum slope point of the first rising edge in the second preprocessed waveform; extracting the arrival time of the pile bottom echo and the appearance time of the main peak from the second preprocessed waveform; calculating the pile wave velocity based on the excitation reference time, the arrival time of the pile bottom echo, and the pre-acquired pile length; calculating the defect center depth based on the pile wave velocity, the excitation reference time, and the appearance time of the main peak; determining the defect segment range based on the defect center depth and the abnormal reflection waveform; and obtaining the segment defect data based on the defect segment range.

[0015] This invention determines the excitation reference time, extracts the pile bottom echo and the time of the main peak from the second preprocessed waveform, calculates the pile body wave velocity by combining the pile length, calculates the depth of the defect center and determines the range of the defect section, and finally obtains the section defect data. This effectively eliminates interference factors in the waveform signal, accurately locates the defect position and range, and improves the accuracy and reliability of the section defect data.

[0016] Optionally, obtaining the defect category and confidence score based on the segment defect data includes: calculating the relative intensity of defect reflection, the time difference from the defect to the pile bottom, the envelope energy change, and the local signal-to-noise ratio using the second preprocessed waveform; determining the defect category using the relative intensity of defect reflection and the time difference from the defect to the pile bottom; and calculating the confidence score using the relative intensity of defect reflection, the local signal-to-noise ratio, and the envelope energy change.

[0017] This invention calculates multiple key parameters using a second preprocessed waveform, then determines the defect category using the relative intensity of defect reflection and the time difference between the defect and the pile bottom, and combines multiple parameters to calculate a confidence score. This approach not only accurately distinguishes defect types using two parameters, avoiding misjudgment based on a single parameter, but also improves the accuracy of defect category and confidence score calculations through comprehensive evaluation of confidence using multiple parameters.

[0018] Optionally, matching the segment-level feature data with the defect data to obtain training samples includes: calculating the crossover ratio (CRR) using the segment-level feature data and the defect data; extracting the maximum value from the CRR to obtain the maximum CRR value; setting a CRR threshold and performing binary annotation on the segment-level feature data based on the comparison between the CRR threshold and the maximum CRR value; and constructing training samples based on the binary annotation results.

[0019] This invention accurately associates construction process data with defect information by calculating the intersection-union ratio (IU) of segment-level feature data and defect data, extracting the maximum IU value, using binary labeling, and constructing training samples. This avoids the subjectivity and error of sample matching. The binary labeling clearly distinguishes between defect and non-defect samples, providing clear labels for model training and improving the accuracy of training samples.

[0020] Optionally, the training samples include defective samples and non-defective samples, and the step of constructing a defect prediction model based on the training samples includes: extracting quality bit markers from the training samples and determining confidence factors based on the quality bit markers; calculating class imbalance factors based on the defective samples and the non-defective samples; calculating the weights of the training samples using the confidence factors and the class imbalance factors; constructing a loss function using the weights; and training the defect prediction model using the training samples based on the loss function.

[0021] This invention determines the confidence factor by extracting quality bit markers and calculates the class imbalance factor, thereby obtaining the training sample weights and constructing a loss function to train the model. It can accurately quantify the sample confidence and class imbalance problem, avoid low-confidence samples interfering with model training, balance the proportion of defective and non-defective samples, and improve the performance of the defect prediction model.

[0022] Optionally, identifying defective sections of bored piles using the defect prediction model includes: calculating an optimal classification threshold; the optimal classification threshold satisfies the following formula: ,in, This represents the optimal classification threshold. The operator represents the maximum value within the interval [0,1]. Indicates at the threshold The F1 score is calculated; based on the optimal classification threshold, the defective sections of the bored pile are identified using the defect prediction model.

[0023] This invention calculates the optimal classification threshold and uses the F1 score as a metric to accurately determine the threshold. By combining this threshold with a defect prediction model to identify defective sections in bored piles, the accuracy and reliability of the defect prediction model can be further improved.

[0024] In another aspect, the present invention provides a system for identifying defective sections of bored piles, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute a method for identifying defective sections of bored piles as described in any of the preceding aspects of the present invention.

[0025] The present invention provides a borehole pile defect section identification system, which has a compact structure, stable performance, high integration and simple composition. It can stably execute the borehole pile defect section identification method provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for identifying defective sections in bored piles according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the layout of multi-source data acquisition equipment during the construction process according to an embodiment of the present invention;

[0028] Figure 3 This describes the process of acquiring and analyzing defect data in an embodiment of the present invention.

[0029] Figure 4 This is a schematic diagram of a defect section identification system for bored cast-in-place piles according to an embodiment of the present invention. Detailed Implementation

[0030] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0031] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0032] Please see Figure 1 To address the shortcomings of the prior art, in one alternative embodiment, such as Figure 1 The method for identifying defective sections in bored piles, as shown, includes the following steps:

[0033] Step S1: Obtain historical construction data and preprocess the historical construction data to obtain historical construction multi-source data.

[0034] In this embodiment, preprocessing the historical construction data to obtain multi-source historical construction data includes: detecting outliers in the historical construction data using the sliding window Z-score method, and removing the outliers to obtain first optimized historical construction data; the first optimized historical construction data includes short-term missing data and long-term missing data; interpolating the short-term missing data using a linear interpolation method to obtain second optimized historical construction data; interpolating the long-term missing data using a multiple regression model or the K-nearest neighbor algorithm to obtain third optimized historical construction data; merging the second optimized historical construction data and the third optimized historical construction data, and smoothing the merged result to obtain optimized multi-source historical construction data; and resampling the optimized multi-source historical construction data to form multi-source historical construction data.

[0035] like Figure 2As shown, in order to cover the key working conditions of bored pile construction, various monitoring devices, including depth sensors, tilt sensors, Beidou positioning and orientation modules, current sensors, and torque sensors, are deployed on the drilling rig to collect multi-source construction process data such as drilling depth, X / Y tilt angle, north / south deviation, current, and torque in real time. All monitoring data are identified by a unique pile number (pile_id) as the core index and a unified timestamp accurate to the millisecond level (preferred accuracy DATETIME(3)) to achieve synchronous acquisition and associated storage across devices and parameters, and report to the data access layer. In order to ensure that the data from different devices can be accurately aligned by time and depth in subsequent analysis, the time reference and coordinate system of the acquisition end and the data receiving end need to be unified.

[0036] At construction sites, due to complex environmental conditions, sensors may experience short-term or long-term signal interruptions, drift, or failures. Communication links may also experience data delays or loss due to interference, leading to issues such as missing timestamps, missing values, or outliers, thus affecting the continuity and integrity of monitoring data. Therefore, a systematic preprocessing procedure must be performed before data is entered into the database: First, outliers are detected using the sliding window Z-score method to obtain the first optimized historical construction data. When collecting values... If the following condition is met, mark it as an anomaly and remove it.

[0037] The first optimized historical construction data satisfies the following formula:

[0038] ,

[0039] in, For the first Record number in The value of the parameter channel; This represents the mean of the channel within the sliding window. Indicates the channel threshold coefficient. This represents the standard deviation of the channel within the sliding window.

[0040] Secondly, for data points that are missing for a short period of time (lasting no more than a certain number of sampling intervals), linear interpolation is used to complete them and obtain the second optimized historical construction data. The second optimized historical construction data satisfies the following formula:

[0041] ,

[0042] in, Indicates interpolation point estimation, Indicates the previous valid observations. Represents the valid observations after the observations are taken. These represent the corresponding timestamps.

[0043] For long-term missing data segments, the third optimized historical construction data is obtained by combining historical data of similar piles or other construction parameters at adjacent depths and using a multiple regression model or K-nearest neighbor algorithm. A quality identifier field (e.g., ...) is then added to the database record for this data. This allows for weight adjustment or data removal in subsequent analyses. To suppress high-frequency glitches, Savitzky-Golay smoothing is applied to the depth / tilt channels before resampling; its key parameters are window length and polynomial order.

[0044] ,

[0045] in, Indicates the length of the smoothing window (if odd, the number of sampling points); This indicates the order of the fitted polynomial.

[0046] Subsequently, each channel was sampled at a uniform frequency. Resampling is performed to generate a consistent "time-depth" master sequence (i.e., historical construction multi-source data):

[0047] ,

[0048] in, Represents the timestamp (s) of the i-th sampling time. This indicates the drilling depth (m) at that moment, with zero at the top of the pile and positive downwards; Indicates drilling speed (m / s); These represent the drill pipe at that moment. Axis tilt angle, These represent the plane coordinates (m) of the drill bit in the construction coordinate system at that moment, and the difference between them and the design coordinates is used to calculate the hole position deviation. These are expressed as current (A) and torque (N·m), respectively.

[0049] To support the collaborative management of high-frequency process data and engineering master data, a hybrid architecture of "relational database + time series database" is adopted.

[0050] The same observation retains both the original and processed values, and the processing rules and parameters are recorded in the ETL log. Finally, a process data sequence with the station number as the primary key and "time-depth" as the unified index is obtained, which is then firmly linked with the project master data, equipment calibration information, and quality position.

[0051] Step S2: Divide the historical construction multi-source data into segments, and extract features from the division results to obtain multiple segment-level feature data.

[0052] The process of dividing the historical construction multi-source data and extracting features from the division results to obtain multiple segment-level feature data specifically includes the following sub-steps:

[0053] Step S201: Based on a pre-set sliding window, the historical construction multi-source data is traversed and sampled at a predetermined step size to obtain multiple multi-source data sub-segments divided according to pile length.

[0054] In this embodiment, historical construction multi-source data is reparameterized from the time domain into a depth domain sequence, and sliding window segmentation and segment-level feature extraction are performed on a unified depth axis to form "feature-label" samples that can be used for training.

[0055] First, establish equidistant windows with the depth axis defined as zero at the pile top and positive downwards. Let the window length be... (m), step size is (m), then the first A depth window is defined as:

[0056] ,

[0057] in, Represents a window. Indicates the first The starting depth of a depth window. Indicates the window length.

[0058] ,

[0059] in, This represents the starting depth of the nth depth window. Indicates the starting point of the analysis. Indicates the window number. Indicates the step size.

[0060] in, As the starting point for analysis, the window overlap rate is... . Let be the starting depth (m) of the nth window. To align the channels in the depth domain, let . The inverse function of the main sequence, for any channel Resampling In each window Inside, by depth spacing (m) Sampled , It is in a depth window The number of sampling points within, .

[0061] Step S202: Feature extraction is performed on the multi-source data sub-segments to obtain multiple segment-level feature data.

[0062] In this embodiment, each sub-segment contains multi-source construction parameters (drilling rate) within that depth range. Torque Current Verticality Features such as hole position deviation δ(⋅) require the calculation of three core features (statistical, dynamic, and interactive) for these parameters to reflect the construction status and quality correlation of this depth segment.

[0063] The segment-level feature data includes window mean, window standard deviation, coefficient of variation, window maximum, window minimum, peak-to-valley difference, least squares linear fitting slope, average gradient magnitude, verticality threshold overshoot ratio, and torque-drilling speed Pearson correlation coefficient.

[0064] The window mean satisfies the following formula:

[0065] ,

[0066] The window standard deviation satisfies the following formula:

[0067] ,

[0068] in, Represents the window mean. Indicates the number of channels at the sampling points. Indicates the current depth window Inner Channel values ​​at each sampling point This represents the standard deviation of the window.

[0069] The maximum window size satisfies the following formula:

[0070] ,

[0071] The minimum window size satisfies the following formula:

[0072] ,

[0073] in, Indicates the maximum value of the internal signal. This indicates that the maximum value is taken for the k-th sampling point within the window. Indicates the current depth window Inner Channel values ​​at each sampling point Indicates the minimum value of the internal signal. Indicates the first [item] in the window The minimum value is taken from each sampling point.

[0074] The coefficient of variation satisfies the following formula:

[0075] ,

[0076] The peak-to-valley difference satisfies the following formula:

[0077] ,

[0078] in, Represents the coefficient of variation. Indicates the standard deviation of the window. Represents the window mean. This represents a small constant used to prevent the denominator from being zero. Indicates the peak-to-valley difference. Indicates the maximum value of the internal signal. This represents the minimum value of the internal signal.

[0079] Dynamic features are used to characterize the trend of change with depth, employing least squares linear slope and average gradient.

[0080] The slope of the least squares linear fit satisfies the following formula:

[0081] ,

[0082] in, Display window Internal channel data With depth The slope of the least squares linear fit. Indicates the number of channels at the sampling points. Indicates depth Channel value at that location, This represents the average depth within the window. Indicates the current depth window Inner Channel values ​​at each sampling point This represents the window mean.

[0083] ,

[0084] in, This represents the average depth within the window. Indicates the number of channels at the sampling points. Indicates depth The channel value at that location.

[0085] The average gradient magnitude satisfies the following formula:

[0086] ,

[0087] in, This represents the average gradient magnitude of adjacent sampling points within the window. Indicates the number of channels at the sampling points. Indicates the current depth window Inner Channel values ​​at each sampling point Indicates the current depth window Inner Channel values ​​at each sampling point Indicates the sampling depth interval.

[0088] In addition, combined with standardized thresholds (e.g., verticality) ), which defines the percentage of threshold crossings.

[0089] The percentage of verticality exceeding the threshold satisfies the following formula:

[0090] ,

[0091] in, Display window Inner channel value Exceeding the threshold proportion, Indicates the number of channels at the sampling points. Display window Inner channel value, Indicates depth Channel value at that location, This indicates the threshold for the channel to exceed its limits.

[0092] ,

[0093] in, This indicates the threshold for the channel to exceed its limits.

[0094] Cross-channel interaction features reflect correlation coefficients within the acceptable window for working condition coordination, such as torque-drilling speed correlation.

[0095] The torque-drilling rate Pearson correlation coefficient satisfies the following formula:

[0096] ,

[0097] in, Display window Internal torque With drilling speed Pearson correlation coefficient, Indicates the number of channels at the sampling points. Display window Internal torque, This represents the average torque within the window. Indicates drilling speed. This represents the average drilling speed within the window.

[0098] Step S3: Defect data is obtained by performing low-strain testing on the piles, and the segment-level feature data is matched with the defect data to obtain training samples.

[0099] The process of obtaining defect data through low-strain testing of piles includes the following sub-steps:

[0100] Step S301: Obtain the original waveform by performing low-strain testing on the pile.

[0101] In this embodiment, after the pile construction (drilling, hole cleaning, lowering of the reinforcing cage, and concrete pouring) is completed and the concrete reaches the design strength, transient excitation (such as hammering or impact) is applied at a designated position on the pile top (such as the center of the pile top or a symmetrical point) to generate elastic waves propagating along the depth direction in the pile body. At the same time, high-precision acceleration or velocity sensors are deployed near the excitation point, and the original waveform is formed by real-time acquisition of the reflected wave signals generated by defects (narrowing, interlayer, etc.) encountered during the propagation of the elastic waves or generated at the bottom of the pile.

[0102] Step S302: Perform mean removal processing and low-order polynomial detrending processing on the original waveform to obtain the first preprocessed waveform.

[0103] In this embodiment, the original waveforms obtained by low strain detection mostly contain interference such as DC components caused by sensor zero drift and slow trend terms caused by signal baseline drift. First, the mean of the original waveform is calculated and then the mean is subtracted from the original waveform to eliminate DC interference. Then, a low-order polynomial trend term is fitted to the waveform after the mean is removed and the low-order polynomial trend term is removed from the waveform to achieve low-order polynomial detrending processing. Finally, a first preprocessed waveform that effectively weakens DC components and slow trend interference is obtained.

[0104] Step S303: Apply a bandpass filter operator to the first preprocessed waveform to obtain the second preprocessed waveform.

[0105] The second preprocessed waveform satisfies the following formula:

[0106] ,

[0107] in, Indicates the first The amplitude of the second preprocessed waveform at each sampling point This represents the bandpass filter operator. Represents the original waveform. express The mean, Indicates to The low-order polynomial trend term.

[0108] Step S304: Perform abnormal reflection event identification on the second preprocessed waveform to determine the abnormal reflection waveform.

[0109] The process of identifying abnormal reflection events in the second preprocessed waveform and determining the abnormal reflection waveform specifically includes the following sub-steps:

[0110] Step S30401: Calculate the short-time average amplitude, long-time average amplitude, and envelope amplitude using the second preprocessed waveform.

[0111] The short-time average amplitude satisfies the following formula:

[0112] ,

[0113] in, Indicates the short-time average amplitude. Indicates the length of the short window. The number of sampling points. Indicates the first The amplitude of the second preprocessed waveform at each sampling point.

[0114] The long-term average amplitude satisfies the following formula:

[0115] ,

[0116] in, Indicates the long-term average amplitude. Indicates the length of the long window (where ), The number of sampling points. Indicates the first The amplitude of the second preprocessed waveform at each sampling point.

[0117] The envelope amplitude satisfies the following formula:

[0118] ,

[0119] in, Indicates the envelope amplitude; For Hilbert transform, Indicates the first The amplitude of the second preprocessed waveform at each sampling point.

[0120] Step S30402: Determine the abnormal reflection waveform based on the ratio of the short-time average amplitude to the long-time average amplitude, combined with the envelope amplitude.

[0121] In this embodiment, the ratio of average amplitudes satisfies the following formula: , ( (A preset ratio threshold is used to identify significant abrupt changes in the short-time signal relative to the long-time background), and Candidate abnormal reflection event windows are selected based on the criterion of local maxima. Then, by combining the minimum interval limit and non-maximum suppression with the overlapping event window, the abnormal reflection waveform is finally determined.

[0122] Step S305: Perform time-depth conversion on the abnormal reflection waveform to obtain section defect data.

[0123] The specific steps for obtaining segment defect data by performing time-depth conversion on the abnormal reflection waveform include the following:

[0124] Step S30501: Determine the excitation reference time based on the maximum slope point of the first rising edge in the second preprocessed waveform.

[0125] In this embodiment, the first wave (the first effective wave of elastic wave propagation) and its rising edge from the start to the peak amplitude are located in the second preprocessed waveform. Then, the derivative is calculated at each moment of the rising edge of the first wave to find the point with the largest derivative (i.e. the point with the largest slope of the rising edge of the first wave). The time corresponding to this point is the excitation reference time.

[0126] The excitation reference time satisfies the following formula:

[0127] ,

[0128] in, Indicates the excitation reference time (unit: s). This represents finding the function that maximizes the expression. Value operators, Indicates the first The amplitude of the second preprocessed waveform at each sampling point.

[0129] Step S30502: Extract the arrival time of the pile bottom echo and the appearance time of the main peak from the second preprocessed waveform.

[0130] In this embodiment, the characteristic signal of the corresponding pile bottom reflection is identified in the second preprocessed waveform. The reflected wave signal with significant amplitude and conforming to the propagation law of the pile bottom position is found. The time corresponding to this signal is the arrival time of the pile bottom echo. Then, for the previously identified abnormal reflection event window, the point with the largest signal amplitude is located within the event window. The time corresponding to this point is the time when the main peak appears.

[0131] The timing of the main peak's appearance satisfies the following formula:

[0132] ,

[0133] in, Indicates the time when the main peak appears. Indicates the first The amplitude of the second preprocessed waveform at each sampling point Indicates the start time of the abnormal reflection waveform. This indicates the end time of the abnormal reflection waveform.

[0134] Step S30503: Calculate the pile wave velocity based on the excitation reference time, the arrival time of the pile bottom echo, and the pre-acquired pile length.

[0135] The wave velocity of the pile body satisfies the following formula:

[0136] ,

[0137] in, Indicates the wave velocity of the pile body. Indicates the pile length. Indicates the arrival time of the echo at the pile bottom. This indicates the excitation reference time.

[0138] Step S30504: Calculate the depth of the defect center based on the pile wave velocity, the excitation reference time, and the time when the main peak appears.

[0139] The depth of the defect center satisfies the following formula:

[0140] ,

[0141] in, Depth of the defect center For the wave velocity of the pile body, Indicates the time when the main peak appears. This indicates the excitation reference time.

[0142] Step S30505: Determine the range of the defect section based on the defect center depth and the abnormal reflection waveform.

[0143] The defective section range satisfies the following formula:

[0144] ,

[0145] in, Indicates the defective section. Indicates the depth of the defect center. Indicates the wave velocity of the pile body. Indicates the time when the main peak appears. Indicates the start time of the abnormal reflection waveform. This indicates the end time of the abnormal reflection waveform.

[0146] Step S30506: Obtain segment defect data based on the defect segment range.

[0147] In this embodiment, the core information corresponding to the defect section is further integrated according to the defect section range to form section defect data. This data not only includes the start and end depths of the defect section and the depth of the defect center, but also associates key parameters in the calculation process (such as pile wave velocity and the time when the main peak appears) to ensure data traceability and provide structured depth dimension information support for subsequent calculation of intersection-union ratio with section-level feature data and construction of training samples.

[0148] Step S306: Obtain the defect category and confidence score based on the defect data of the section.

[0149] The specific steps for obtaining the defect category and confidence score based on the defect data of the aforementioned section include the following:

[0150] Step S30601: Calculate the relative intensity of defect reflection, time difference from defect to pile bottom, envelope energy change, and local signal-to-noise ratio using the second preprocessed waveform.

[0151] The relative intensity of defect reflections satisfies the following formula:

[0152] ,

[0153] in, Indicates the relative intensity of the defect reflection; Indicates the peak time The amplitude of the second preprocessed waveform, Indicates the arrival time of the echo at the pile bottom. The amplitude of the second preprocessed waveform.

[0154] The time difference from the defect to the bottom of the pile satisfies the following formula:

[0155] ,

[0156] in, This represents the time difference (s) from the defect to the bottom of the pile. Indicates the arrival time of the echo at the pile bottom. Indicates the moment when the main peak appears.

[0157] When calculating the envelope energy change, a Hilbert transform is performed on the second preprocessed waveform to extract the envelope curve. A local time window centered on the peak time of the defect is defined, and the energy integral of the envelope within the window is calculated. Simultaneously, a reference time window of equal length is selected near the arrival time of the pile bottom echo, and its envelope energy is calculated. The difference between the two is the envelope energy change. When calculating the local signal-to-noise ratio, the mean square value of the signal amplitude (signal energy) is calculated within the local time window of the defect, and the mean square value of the noise amplitude is calculated in the noise time window in the stable segment next to the time window. The ratio of the two is the local signal-to-noise ratio.

[0158] Step S30602: Determine the defect category using the relative intensity of the defect reflection and the time difference from the defect to the pile bottom.

[0159] In this embodiment, a two-dimensional classification rule is constructed using the relative intensity of defect reflection and the time difference from the defect to the pile bottom. First, through statistical analysis of engineering samples, the time difference from the defect to the pile bottom and the reflection intensity threshold of the relative intensity of defect reflection are defined. Then, the defect characteristics are matched according to the combination of the two: shallow + strong reflection corresponds to narrowing (abrupt cross-section leads to strong reflection), middle + medium reflection corresponds to segregation (medium reflection due to discontinuous medium), and deep + weak reflection corresponds to widening (cross-section expansion causes wave diffusion and energy dispersion, resulting in weak reflection). In this way, the defect category is quantitatively determined.

[0160] Step S30603: Calculate the confidence score using the relative intensity of the defect reflection, the local signal-to-noise ratio, and the envelope energy change.

[0161] The confidence score satisfies the following formula:

[0162] ,

[0163] ,

[0164] in, Indicates the confidence level (0-1). This represents the normalized amount of the relative intensity of the defect reflection. This represents the normalized value of the envelope energy change. This represents the normalized value of the local signal-to-noise ratio. , , These represent the weights.

[0165] Step S307: Obtain defect data based on the defect category, the confidence score, and the segment defect data.

[0166] In this embodiment, based on the determined defect category, combined with a quantitative confidence score (calculated by the matching degree between defect reflection parameters and typical patterns, and the degree of feature fluctuation, numerically representing the reliability of the judgment, such as the low confidence interval corresponding to suspected defects), and the depth interval, defect center depth, and associated pile wave velocity contained in the segment defect data, it is integrated into defect data through structured mapping. This data simultaneously records the defect type, confidence value spatial depth range, and wave velocity parameters, forming a complete result of attribute-confidence-space-physical association, which not only supports defect level assessment but also provides accurate annotation basis for the construction of training samples for supervised learning.

[0167] The process of matching the segment-level feature data with the defect data to obtain training samples specifically includes the following sub-steps:

[0168] Step S311: Calculate the intersection-union ratio using the segment-level feature data and the defect data.

[0169] The intersection-union ratio satisfies the following formula:

[0170] ,

[0171] in, Represents depth window With the The intersection-union ratio of each defective segment, where... Indicates the initial depth of the window. Indicates the window length. Indicates the first One defective section.

[0172] Step S312: Extract the maximum value from the crossover-union ratio to obtain the maximum crossover-union ratio value.

[0173] The maximum intersection-union ratio satisfies the following formula:

[0174] ,

[0175] in, This represents the maximum intersection-union ratio of the window with all defective sections, where, This represents the maximum value of the intersection-union ratio (IUU) for all defective sections. Represents depth window With the The intersection-combination ratio of each defective section.

[0176] Step S313: Set the intersection-union ratio threshold, and perform binary labeling on the segment-level feature data based on the comparison between the intersection-union ratio threshold and the maximum intersection-union ratio.

[0177] Binary annotations satisfy the following formula:

[0178] ,

[0179] in, Display window binary tags, This represents the maximum intersection-union ratio between the window and all defective sections. This represents the intersection-union ratio threshold, with a value between 0 and 1.

[0180] like Then inheritance and The corresponding defect type is used as a secondary label:

[0181] ,

[0182] in, Indicates the first The defect type of each window, Indicates the first Category labels for class defects Indicates the relationship with the current sample The defect category index with the largest IoU satisfies the following formula:

[0183] , Represents depth window With the The intersection-combination ratio of each defective section.

[0184] Step S314: Construct training samples based on the results of the binary annotation.

[0185] In this embodiment, the binary annotation results (binary labels and defect types) are associated with segment-level feature data to form structured training samples. , The number of samples is used to define the segment-level feature data, which includes the mean, standard deviation, maximum, minimum, coefficient of variation, and peak-to-valley difference extracted from historical construction multi-source data (drilling speed, torque, current, inclination angle, hole position deviation, etc.). Finally, the training set and validation set are divided according to a preset ratio (e.g., 8:2) to provide standardized and labeled input data for training machine learning models.

[0186] Step S4: Construct a defect prediction model based on the training samples, and use the defect prediction model to identify defective sections of the bored pile.

[0187] The training samples include defective samples and non-defective samples. The construction of the defect prediction model based on the training samples specifically includes the following sub-steps:

[0188] Step S401: Extract the quality bit flag from the training samples and determine the confidence factor based on the quality bit flag.

[0189] The confidence factor satisfies the following formula:

[0190] ,

[0191] in, This represents the confidence factor corresponding to the quality bit. Indicates the first The quality bit flag for each window.

[0192] The quality flag originates from the quality detection of waveform data within the window: by analyzing the signal continuity of the waveform (such as whether there are abrupt changes or truncation), the sufficiency of effective points, and the fitting extrapolation features (extrapolation segment label), each window sample is pre-graded in terms of quality (outputting discrete labels of 0-5, with different values ​​corresponding to high quality, medium-low quality, abnormal / extrapolation, etc.), which is used to quantify the reliability of the data.

[0193] Step S402: Calculate the class imbalance factor based on the defective samples and the non-defective samples.

[0194] The class imbalance factor satisfies the following formula:

[0195] ,

[0196] in, For class imbalance factor, The number of non-defect samples. This represents the number of defective samples.

[0197] Step S403: Calculate the weights of the training samples using the confidence factor and the class imbalance factor.

[0198] ,

[0199] in, Indicates weight, Represents the class imbalance factor, when Time to take Otherwise, take 1. This represents the confidence factor corresponding to the quality bit. Indicates the first The quality bit flag for each window.

[0200] The above formula gates the data quality within the window. Let... For window Internal mass position (Extrapolation / Abnormal) Sample Proportion, when ( The threshold for the proportion of low-quality samples within a window (a preset parameter based on engineering experience or statistical derivation) is used to directly remove samples from the window if there are insufficient valid points. Sample weights are assigned to the retained samples. This is due to a combination of factors, including class imbalance and data confidence.

[0201] Step S404: Construct a loss function using the weights.

[0202] In this embodiment, an XGBoost decision tree is used to make binary (defect / normal) distinctions, with logical loss as the objective and sample weights entering the loss function.

[0203] The loss function satisfies the following formula:

[0204] ,

[0205] in, Represents the loss function. Represents the total number of samples. Indicates the first The weights of each training sample, Indicates the first The binary labels of each training sample. The output of the model represents the first... The defect probability of each training sample. Indicates the first A tree, Regularization for tree complexity.

[0206] , ,

[0207] in, This represents the defect probability of the nth depth window output by the model. This represents the sigmoid function, which outputs the model. Mapped to the interval (0,1), This represents the score of the nth sample in the additive decision tree model. Indicates the learning rate. Indicates the first A tree, This represents the feature vector of the nth sample.

[0208] Step S405: Based on the loss function, a defect prediction model is trained using the training samples.

[0209] In this embodiment, training a defect prediction model using the training samples based on the loss function includes: dividing the training samples into a training set and a validation set using a stub-based K-fold cross-validation strategy; training a defect prediction model using the loss function based on the training set; and validating the defect prediction model using the validation set.

[0210] The defect prediction model is built based on XGBoost decision trees.

[0211] To avoid data leakage within the same stake, training and evaluation employ K-fold cross-validation grouped by stake. Let the stake set be... Construct K-fold grouping And ensure that the window sample of any stump appears in only a subset.

[0212] The training set satisfies the following formula:

[0213] ,

[0214] in, Indicates the first The training dataset in cross-validation Indicates the first The set of sample IDs for the compromise training set. Indicates the first The feature vector of each sample Indicates the first Binary labels for each sample, Indicates the first The weights of each sample.

[0215] The validation set satisfies the following formula:

[0216] ,

[0217] in, Indicates the first The validation dataset in folded cross-validation Indicates the first The set of sample IDs for the compromise validation set. Indicates the first The feature vector of each sample Display window binary tags, This represents the weight of the nth sample.

[0218] The identification of defective sections in bored piles using the aforementioned defect prediction model includes:

[0219] Step S411: Calculate the optimal classification threshold.

[0220] The optimal classification threshold satisfies the following formula:

[0221] ,

[0222] in, This represents the optimal classification threshold. The operator represents the maximum value within the interval [0,1]. Indicates at the threshold The F1 score is calculated below.

[0223] Step S412: Based on the optimal classification threshold, the defective sections of the bored pile are identified using the defect prediction model.

[0224] In this embodiment, identifying defective sections of bored piles using the defect prediction model based on the optimal classification threshold includes: obtaining real-time sliding window features based on pre-obtained real-time construction data of the bored piles; inputting the real-time sliding window features into the defect prediction model to obtain a defect probability set; calculating an upper threshold and a lower threshold based on the optimal classification threshold; determining the defect state corresponding to the real-time sliding window features based on the upper threshold, the lower threshold, and the defect probability set; merging adjacent real-time sliding window features based on the defect state to obtain a defect cluster; and calculating the depth segment and defect confidence level based on the defect cluster.

[0225] After completing model training and threshold calibration, the trained classification model is deployed to the data service at the construction site to perform segmented prediction, hysteresis determination and segment merging output on real-time process data, and the results are visualized and stored in the database to form a closed-loop management system.

[0226] Data access is based on the "time-depth" master sequence established in step one. The system uses a fixed depth step size. Advance, to the latest coverage depth window:

[0227] ,

[0228] in, Represents a window. This represents the starting depth of the nth depth window. Indicates the window length.

[0229] ,

[0230] in, This represents the starting depth of the nth depth window. Indicates the window number. This indicates a fixed depth step size.

[0231] Calculate segment-level feature vectors from the cleaned multi-channel data (drilling speed, torque, current, verticality, hole position deviation, etc.). If the percentage of quality units within the window exceeds the threshold (e.g., ... If the proportion is >20%, then skip this window; otherwise, feed it into the trained model and output the defect probability.

[0232] ,

[0233] in, This represents the defect probability of the nth depth window output by the model. This represents the trained model. This represents the segment-level feature vector.

[0234] To determine the hysteresis threshold, two thresholds, one upper and one lower, are used for state determination, and the predicted state is recorded. (1 represents the "defect window")

[0235] ,

[0236] in, Indicates the predicted state. This represents the defect probability of the nth window after smoothing or correction. This represents the upper threshold for defect determination. This indicates the prediction status of the previous window. This represents the lower threshold for defect determination.

[0237] in, Typical , , .

[0238] Will satisfy The windows are clustered into several consecutive clusters according to depth order, assuming that the gap between the right / left boundaries of two adjacent defect windows does not exceed [a certain value]. (m) are considered to be in the same cluster and merged. For the merged... For each defect cluster, output its depth range and defect confidence level:

[0239] ,

[0240] in, This represents the initial depth of the g-th defect. Indicates in set Take the minimum value from the middle. Indicates belonging to a category The set of window indices This represents the starting depth of the nth depth window.

[0241] ,

[0242] in, This represents the ending depth of the g-th defect. Indicates in set Take the maximum value from the middle. This represents the starting depth of the nth depth window. Indicates belonging to a category The set of window indices Indicates the window length.

[0243] ,

[0244] in, Indicate category The confidence level of the defect Indicates in set Take the maximum value from the middle. Indicates belonging to a category The set of window indices This represents the defect probability of the nth window after smoothing or correction.

[0245] The profile interface overlays and displays the location, length, type, and confidence level of each defect cluster, labeled with the pile axis as the reference. and according to Hierarchical coloring; exceeding engineering thresholds (e.g., length ≥ and This triggers an alarm and displays the process characteristics of that segment for the operator and supervisor to review.

[0246] The online prediction results are written into the online prediction result table, and the window-level probability and state trajectory are recorded simultaneously. When a pile exhibits a continuous cluster of high-confidence defects, the system automatically generates a "process review" link: locating the process parameter curves and construction logs for the corresponding depth segment, forming a "defect-process" evidence set. After subsequent pile inspection and retesting, the final test results are returned and entered into the database for continuous model calibration and performance evaluation.

[0247] In summary, such as Figure 3 As shown, in pile foundation construction and quality inspection, on the one hand, multi-source data of the pile foundation construction process is collected by construction equipment equipped with multi-source sensing modules such as GNSS antenna, dual-axis inclinometer, torque sensor, and depth sensor. On the other hand, following the process of historical construction data preprocessing → segmented feature extraction → combining pile low strain detection defect data to build training samples → training defect prediction model, a data-driven analysis system is constructed to ultimately realize the identification and prediction of defect sections of bored cast-in-place piles.

[0248] like Figure 4 As shown, in another aspect, the present invention also provides a system for identifying defective sections of bored piles, including: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the relevant steps of a relevant embodiment of the method for identifying defective sections of bored piles of the present invention.

[0249] This invention provides a defect section identification system for bored piles. The functional components can be integrated into a single processing unit, or each component can exist independently, or two or more components can be integrated into one unit. The integrated components can be implemented in hardware or software.

[0250] 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. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for identifying a defective section of a cast-in-place pile, characterized by, The method comprises: acquiring historical construction data and preprocessing the historical construction data to obtain historical construction multi-source data; dividing the historical construction multi-source data and extracting features from the divided results to obtain a plurality of segment-level feature data; obtaining defect data by low-strain detection of a pile and matching the segment-level feature data with the defect data to obtain training samples; the defect data obtained by low-strain detection of the pile comprises: obtaining an original waveform by low-strain detection of the pile; performing mean value removal processing and low-order polynomial detrending processing on the original waveform to obtain a first preprocessed waveform; applying a band-pass filtering operator to the first preprocessed waveform to obtain a second preprocessed waveform; identifying abnormal reflection events from the second preprocessed waveform to determine an abnormal reflection waveform; performing time-depth conversion on the abnormal reflection waveform to obtain section defect data; obtaining defect categories and confidence scores from the section defect data; obtaining defect data based on the defect categories, the confidence scores and the section defect data; constructing a defect prediction model based on the training samples and using the defect prediction model to identify defect sections of a cast-in-place pile.

2. The method according to claim 1, wherein the division of the historical construction multi-source data and the feature extraction from the divided results to obtain a plurality of segment-level feature data comprises: traversing and sampling the historical construction multi-source data based on a pre-set sliding window with a predetermined step size to obtain a plurality of multi-source data sub-segments divided by pile length; extracting features from the multi-source data sub-segments to obtain a plurality of segment-level feature data.

3. The method according to claim 1, wherein the identification of abnormal reflection events from the second preprocessed waveform to determine an abnormal reflection waveform comprises: calculating short-time average amplitude, long-time average amplitude and envelope amplitude from the second preprocessed waveform; determining abnormal reflection waveforms based on the ratio of short-time average amplitude to long-time average amplitude and the envelope amplitude.

4. The method according to claim 1, wherein the time-depth conversion of the abnormal reflection waveform to obtain section defect data comprises: determining a reference time of excitation based on the maximum slope point of the rising edge of the first wave in the second preprocessed waveform; extracting the arrival time of the pile bottom echo and the occurrence time of the main peak in the second preprocessed waveform; calculating the wave velocity of the pile body based on the reference time of excitation, the arrival time of the pile bottom echo and the pre-acquired pile length of the pile; calculating the center depth of the defect based on the wave velocity of the pile body, the reference time of excitation and the occurrence time of the main peak; determining the range of the defect section based on the center depth of the defect and the abnormal reflection waveform; obtaining section defect data based on the range of the defect section.

5. The method according to claim 1, wherein the obtaining of defect categories and confidence scores from the section defect data comprises: calculating defect reflection relative intensity, defect-to-pile-bottom time difference, envelope energy change and local signal-to-noise ratio from the second preprocessed waveform; determining defect categories based on the defect reflection relative intensity and the defect-to-pile-bottom time difference; calculating confidence scores based on the defect reflection relative intensity, the local signal-to-noise ratio and the envelope energy change.

6. The method according to claim 1, wherein The matching of the segment-level feature data and the defect data obtains training samples, and the matching includes: Calculating an intersection-over-union ratio by using the segment-level feature data and the defect data; Extracting a maximum value in the intersection-over-union ratio to obtain a maximum intersection-over-union ratio value; Setting an intersection-over-union ratio threshold, and performing binary labeling on the segment-level feature data based on a comparison result of the intersection-over-union ratio threshold and the maximum intersection-over-union ratio value; Constructing training samples based on a result of the binary labeling.

7. The method according to claim 1, wherein the method is characterized by, The training samples include defect samples and non-defect samples, and constructing a defect prediction model based on the training samples includes: Extracting a quality bit flag in the training samples, and determining a confidence factor according to the quality bit flag; Calculating a class imbalance factor based on the defect samples and the non-defect samples; Calculating a weight of the training samples by using the confidence factor and the class imbalance factor; Constructing a loss function by using the weight; Training a defect prediction model by using the training samples based on the loss function.

8. The method according to claim 1, wherein the method is characterized by, The recognition of a defect section of a cast-in-place bored pile by using the defect prediction model includes: Calculating an optimal classification threshold; The optimal classification threshold satisfies the following formula: ， wherein, denotes the optimal classification threshold, denotes the operator that takes the maximum value in the interval [0, 1], denotes the F1 score computed at the threshold value. Recognizing a defect section of a cast-in-place bored pile by using the defect prediction model based on the optimal classification threshold.

9. A bored pile defect section identification system characterized by, It includes: A processor, an input device, an output device and a memory, which are connected to each other, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions to execute a method for recognizing a defect section of a cast-in-place bored pile according to any one of claims 1 to 8.

Images