A method and system for locating broken wires in a prestressed steel strand

Abstract

This invention provides a method and system for locating broken wires in prestressed steel strands, comprising: preprocessing the acquired acoustic emission signals of broken steel strands with noise reduction; searching for abrupt change locations in the preprocessed acoustic emission signals based on the Hinkley criterion; constructing an analysis window based on the abrupt change location and determining the location of the first arrival point based on the AIC criterion, the corresponding sampling time being the first wave arrival time; wherein, a boundary penalty term is introduced into the AIC criterion to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result; and calculating the time difference between the arrival of the same source acoustic emission signals at different acoustic emission sensors based on the first wave arrival time to determine the specific location of the broken steel strand. This method can smooth out pulse-type interference noise in the environment, solve the problem that the first arrival picking of the AIC criterion is easily affected by the small sample effect, improve the picking accuracy of the first wave arrival time of the acoustic emission signal, and provide accurate data support for locating broken steel strands.

Description

Technical Field

[0001] This invention relates to the field of acoustic emission monitoring technology, and more specifically, to a method and system for locating broken wires in prestressed steel strands. Background Technology

[0002] Prestressed concrete box girders are widely used in high-speed railway bridge structures. The prestressed steel strands, as the core load-bearing unit, directly affect the overall safety and durability of the bridge structure due to their mechanical properties and service condition. During long-term operation, the steel strands within the ducts are susceptible to hidden damage such as corrosion, deterioration, and localized fractures due to factors including construction deviations, environmental corrosion, and repeated train loads. If the location of broken strands cannot be identified and accurately determined in a timely manner, prestress loss will accumulate, weakening the structure's load-bearing capacity and posing a potential threat to the operational safety of high-speed railways. Therefore, conducting research on the precise location of broken strand damage in prestressed steel strands has significant engineering value.

[0003] Chinese invention patent CN116973452A discloses a method for locating the cross-section of prestressed steel strands damaged by a prestressed steel strand using a waveguide and dual-sensor acoustic emission method. This method connects prestressed steel strands in different ducts using a waveguide, extending the one-dimensional linear positioning method to three-dimensional space. Only two acoustic emission sensors are needed to directly monitor the damage status of prestressed steel strands in different ducts within a bridge. The method determines the damage location by calculating the time difference between the arrival times of the same acoustic emission signals at different sensors, as well as the wave velocity parameter. Therefore, accurately picking up the arrival time of the acoustic waves is crucial for improving the actual positioning accuracy.

[0004] In his academic master's thesis, "Research on First Arrival Picking of Concrete Acoustic Emission Signal Based on EEMD," submitted in June 2024, Chen Qiang addressed the challenges of acoustic emission signals being easily contaminated by noise and the difficulty of traditional picking algorithms in accurately capturing the first arrival. He proposed a joint first arrival picking algorithm based on EEMD and wavelet analysis, which includes noise reduction preprocessing based on EEMD and wavelet analysis, preliminary determination of signal arrival time based on improved kurtosis method, and local accurate picking of signal arrival time based on AIC criterion.

[0005] The aforementioned technical research employed an improved kurtosis method to preliminarily determine the arrival time of acoustic emission signals from concrete. To capture the effective signal portion of concrete with non-Gaussian characteristics, this improved kurtosis method is highly sensitive to anomalous signals. However, the application environment of prestressed concrete box girders is complex, with numerous non-Gaussian burst pulse interference signals, leading to an extremely high false alarm rate. Furthermore, the locally accurate signal arrival time acquisition using the AIC criterion is susceptible to the small sample effect at the edge of the analysis window, resulting in false minima.

[0006] In order to solve the above problems, people have been seeking an ideal technological solution. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for accurately identifying the first wave arrival time of prestressed steel strand broken wires.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] In a first aspect, the present invention provides a method for locating broken wires in prestressed steel strands, comprising:

[0010] The collected acoustic emission signals of broken steel strands are subjected to noise reduction preprocessing.

[0011] The abrupt change location search is performed on the preprocessed acoustic emission signal based on the Hinkley criterion.

[0012] An analysis window is constructed based on the mutation location, and the location of the first arrival point is determined based on the AIC criterion. The corresponding sampling time is the arrival time of the first wave.

[0013] In the AIC criterion, a boundary penalty term is introduced to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result. Formula (1) is its expression:

[0014] (1)

[0015] In the formula, For penalty weighting; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size;

[0016] Based on the arrival time of the first wave, the time difference between the arrival of the same source acoustic emission signal at different acoustic emission sensors is calculated to determine the specific location of the broken wire in the steel strand.

[0017] First, the acoustic emission signal is preprocessed with noise reduction to provide a more stable input signal for the identification of the first arrival time. Combining the Hinkley criterion and the AIC criterion, the approximate abrupt change location of the first arrival point is quickly located using the Hinkley criterion, and then the precise location is determined within the analysis window using the AIC criterion, achieving dual protection of coarse and fine positioning and solving the problems of low positioning accuracy and poor efficiency of single methods. The Hinkley criterion focuses on the continuous change of signal energy level, rather than the abnormal fluctuation of a single point, which can smooth out non-Gaussian burst pulse interference signals and accurately respond to continuous high-energy oscillations such as steel strand breakage.

[0018] To address the issue of large fluctuations in small sample data at the analysis window boundary, which easily interfere with the AIC criterion judgment results, a boundary penalty term is introduced into the AIC criterion. This term applies a smoothing penalty to the segmentation points near the window edge and where the effective sample size is insufficient, effectively suppressing the interference of small sample data fluctuations at the analysis window boundary. By setting the penalty weight, the penalty intensity can be flexibly adjusted according to the actual detection scenario and signal noise conditions, improving the adaptability and flexibility of the method. This avoids false minimum values ​​in the AIC value caused by small sample fluctuations at the boundary, ensuring the accuracy of the precise location identification of the first arrival point, thereby improving the accuracy of the first wave arrival time identification. Based on the accurate first wave arrival time, the time difference is calculated to achieve precise location of the broken wire in the prestressed steel strand, making it suitable for detecting broken wires in prestressed steel strands under different bridge working conditions.

[0019] In a preferred embodiment, the step of searching for abrupt change locations in the preprocessed acoustic emission signal based on the Hinkley criterion includes:

[0020] Based on the timing of the preprocessed acoustic emission signal, several sampling points are preset;

[0021] Choose any sampling point as the segmentation point k, and divide the M sampling points in its neighborhood signal segment into the front segment and the back segment;

[0022] Formula (2) is used to calculate the Hinkley statistic H(k) corresponding to the segmentation point k, which is used to quantify the degree of abrupt change in the acoustic emission signal at the segmentation point k;

[0023] (2)

[0024] In the formula, x t The amplitude value at sampling point t. The average amplitude value of M sampling points; This represents the average amplitude value of the sampling points in the preceding segment, including sampling points 1, 2...k. This represents the average amplitude value of the sampling points in the latter part, including k+1, k+2…M;

[0025] Iterate through all split points, and the split point k0 corresponding to the maximum value of the Hinkley statistic is the mutation position.

[0026] By pre-setting sampling points and using M sampling points within the neighborhood of a segmentation point as the analysis object, the randomness of single sampling point analysis is avoided, improving the stability of abrupt change location search. The M sampling points within the neighborhood of the segmentation point are evenly divided into a front segment and a rear segment. The degree of abrupt change in the signal at the segmentation point is quantified by the Hinkley statistic, realizing the quantitative judgment of the abrupt change location of the acoustic emission signal. By traversing all segmentation points and taking the segmentation point corresponding to the maximum value of the Hinkley statistic as the abrupt change location, the approximate location of the first arrival point of the acoustic emission signal can be accurately captured, providing a reliable benchmark for subsequent precise positioning based on the AIC criterion, reducing the computational load of subsequent fine positioning, and improving the overall positioning efficiency.

[0027] In a preferred embodiment, the step of constructing an analysis window based on the mutation location and determining the location of the first arrival point based on the AIC criterion includes:

[0028] Using the mutation location as the right boundary, a fixed-length analysis window is constructed in its neighborhood, and N sampling points are preset within the analysis window;

[0029] Choose any sampling point as the segmentation point k to divide the local segment of the acoustic emission signal within the window into the front segment and the back segment;

[0030] Formula (3) is used to calculate the AIC value corresponding to the segment k, which is used to quantify the statistical differences between the front and back segments;

[0031] (3)

[0032] In the formula, This represents the variance of the signal amplitude values ​​in the preceding segment; The variance of the signal amplitude values ​​in the latter part;

[0033] Iterate through all the dividing points, and the dividing point corresponding to the local minimum value is the location of the initial arrival point.

[0034] A fixed-length analysis window is constructed with the abrupt change location as the right boundary to focus on signal segments near the first arrival point, eliminating interference from irrelevant signals, narrowing the analysis range for precise positioning, and improving positioning efficiency. The signal within the window is divided into a pre-segment and a post-segment by a segmentation point. The statistical characteristic difference between the two segments is quantified using the AIC value, which can accurately identify the abrupt change characteristics of the signal before and after the first arrival point. Before the first arrival point, noise dominates, while after the first arrival point, effective acoustic emission signals dominate. The segmentation point corresponding to the local minimum of AIC is used as the precise location of the first arrival point, clarifying the judgment criteria for precise positioning and providing a guarantee for the accurate acquisition of the arrival time of the first wave.

[0035] In a preferred embodiment, the step of calculating the time difference between the arrival times of the same acoustic emission signals at different acoustic emission sensors based on the first wave arrival time, and determining the specific location of the broken wire in the steel strand, includes:

[0036] Two acoustic emission sensors were used to collect acoustic emission signals from the same source generated by broken steel strands in different prestressed ducts, and the characteristics of the same source acoustic emission signals were extracted.

[0037] The features include: the arrival time difference, energy ratio, and amplitude ratio of the first wave of the acoustic emission signals from the same source to the two acoustic emission sensors, as well as the dominant frequency, root mean square value, duration, rise time, and spectral entropy of the signals received by each acoustic emission sensor;

[0038] Based on the aforementioned features, a multidimensional feature vector for locating broken wires in steel strands is constructed; the multidimensional feature vector is input into a machine learning model to train the model to learn the acoustic emission signal feature distribution of broken wires in steel strands corresponding to specific prestressed ducts.

[0039] A pre-trained machine learning model is used to predict the location of broken wires in the prestressed ducts of the steel strands.

[0040] Two acoustic emission sensors are used to collect acoustic emission signals from the same source. By combining multi-dimensional features, the characteristic basis for locating broken wires in steel strands is enriched, overcoming the problem of poor reliability of single-feature positioning. By constructing multi-dimensional feature vectors, the model can more comprehensively learn the signal feature distribution of specific prestressed ducts. The machine learning model is used to predict the prestressed duct where the broken wire is located, avoiding reliance on the pre-calibrated elastic wave propagation velocity. This overcomes the systematic amplification problem of wave velocity error on positioning accuracy in traditional time-difference positioning methods, further improving the accuracy and efficiency of positioning. It is suitable for engineering scenarios with unknown wave velocities or complex propagation conditions.

[0041] In a preferred embodiment, the method further includes: using the characteristics of the homogeneous acoustic emission signals as samples, generating simulation features consistent with the statistical characteristics of real steel strand broken wire acoustic emission signals based on the TransGAN generative adversarial network;

[0042] Based on the simulation features, a simulated multidimensional feature vector for locating broken wires in steel strands is constructed; the simulated multidimensional feature vector and the multidimensional feature vector are input into a machine learning model to train the model to learn the acoustic emission signal feature distribution of broken wires in steel strands corresponding to specific prestressed ducts.

[0043] Based on the TransGAN generative adversarial network, simulated features that are consistent with the statistical characteristics of real signals are generated, which effectively expands the sample size of localization features and solves the problem of poor training effect of machine learning models due to the scarcity and imbalance of steel strand broken wire samples in actual detection.

[0044] In a preferred embodiment, the machine learning model employs an ensemble learning model, including the XGBoost algorithm, LightGBM algorithm, CATBoost algorithm, or RF algorithm.

[0045] Ensemble learning models such as XGBoost and LightGBM are selected. Compared with single machine learning models, they have stronger feature learning capabilities, anti-overfitting capabilities, and generalization capabilities, and can better adapt to the learning needs of multi-dimensional feature vectors. Ensemble learning models can effectively handle complex signal features in the localization process, reduce the impact of noise and sample imbalance on model prediction results, improve the prediction accuracy of the prestressed duct where the steel strand is broken, realize the rapid localization of the steel strand, and adapt to the working conditions of real-time on-site bridge inspection.

[0046] Secondly, the present invention provides a prestressed steel strand broken wire positioning system, comprising:

[0047] The signal processing module performs noise reduction preprocessing on the collected acoustic emission signal of broken steel strands;

[0048] The first wave arrival time identification module searches for abrupt change locations in the preprocessed acoustic emission signal based on the Hinkley criterion.

[0049] An analysis window is constructed based on the mutation location, and the location of the first arrival point is determined based on the AIC criterion. The corresponding sampling time is the arrival time of the first wave.

[0050] In the AIC criterion, a boundary penalty term is introduced to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result. Formula (1) is its expression:

[0051] (1)

[0052] In the formula, For penalty weighting; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size;

[0053] The positioning module calculates the time difference between the arrival times of the first wave and the arrival times of the same acoustic emission signals at different acoustic emission sensors to determine the specific location of the broken wire in the steel strand.

[0054] The system utilizes signal processing, first wave identification, and positioning modules to automate the location of broken strands in prestressed steel strands. The signal processing module's noise reduction and preprocessing functions, along with the first wave arrival time identification module's coarse and fine positioning and boundary penalty mechanisms, ensure the accuracy of the first wave arrival time identification. The positioning module calculates the time difference based on the precise first wave arrival time to accurately locate the specific position of the broken strand, making it stable for routine detection of broken strands in prestressed steel strands of bridges.

[0055] In a preferred technical solution, it also includes:

[0056] The signal acquisition module, which is communicatively connected to the signal processing module, includes a waveguide rod and an acoustic emission sensor;

[0057] The waveguide rod has a preset length and is used to connect prestressed steel strands located at different channel positions;

[0058] The two acoustic emission sensors are fixedly installed at the beginning and end of the waveguide rod, respectively, to collect acoustic emission signals generated by broken steel strands in different prestressed ducts.

[0059] The signal acquisition module enables autonomous acquisition of acoustic emission signals. The waveguide rod is designed with a preset length, which can be flexibly adapted to the actual distribution of prestressed ducts in the bridge, realizing the series connection of prestressed steel strands at different duct locations. This ensures that the acoustic emission signals generated by broken wires in each duct can be effectively transmitted to the sensors. Two acoustic emission sensors are fixed at the beginning and end of the waveguide rod, respectively, which can accurately acquire acoustic emission signals from the same source, ensuring the accuracy of the first wave arrival time difference calculation. This can adapt to the complex on-site environment of the bridge and improve the reliability and continuity of signal acquisition.

[0060] This invention has outstanding substantive features and significant progress compared to the prior art, specifically including:

[0061] (1) The Hinkley criterion is used to smooth the pulse-type interference noise in the environment and respond to continuous energy changes, so as to achieve coarse positioning with fast calculation and low false alarm rate; fine positioning is achieved by using the AIC criterion and the boundary penalty term is introduced to effectively solve the core defects of the AIC criterion in the prior art, which is easily affected by small sample effect and produces pseudo-minimal points; the coarse positioning and fine positioning work together to improve the picking accuracy of the first wave arrival time of the acoustic emission signal, provide accurate data support for the positioning of broken wires in steel strands, and meet the needs of accurate identification of hidden broken wire damage in high-speed railway bridges.

[0062] (2) A method for locating broken steel strands by identifying initial arrival, expanding features and predicting location was constructed. The time difference was obtained by reasonably deploying waveguide rods and dual sensors. The TransGAN generative adversarial network was used to solve the problem of scarce and unbalanced broken steel strand samples. Combined with an integrated learning model, the method can accurately locate broken steel strands in different ducts of the bridge and identify the ducts. This method can effectively avoid the difficulty in accurately calibrating wave velocity in the engineering field, ensure the operational safety and durability of high-speed railway bridges, and has important engineering application value. Attached Figure Description

[0063] Figure 1 This is a flowchart of the method of the present invention;

[0064] Figure 2 This is a schematic diagram of the signal acquisition arrangement structure of the method of the present invention;

[0065] Figure 3 This is the result of the first wave arrival time identification of the partial prestressed duct wire breakage signal in Example 2;

[0066] Figure 4 These are the comparative test results of the four first wave arrival time identification methods in Example 2;

[0067] Figure 5 This is a structural diagram of the TransGAN network in Example 2;

[0068] Figure 6 The Pearson correlations of the 13 features selected in Example 2;

[0069] Figure 7 These are the marginal distribution diagrams of the four typical feature columns selected in Example 2;

[0070] Figure 8 It is the classification confusion matrix of the test set of XGBoost, LightGBM, CATBoost, and RF algorithms in Example 2. Detailed Implementation

[0071] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0072] To facilitate understanding of the technical solutions provided in this application, the technical terms involved in the embodiments of this application are explained below.

[0073] The Hinkley criterion, proposed by DV Hinkley in his 1971 paper "Inference about the Change-Point from Cumulative Sum Tests" published in the statistics journal Biometrika, is a statistical method for detecting the onset time of acoustic emission signals.

[0074] The Akaike Information Criterion (AIC) is based on the non-stationary nature of acoustic emission signals. It divides the observed acoustic emission waveform sequence into multiple fixed-length waveform segments to approximate the original signal, with each segment considered as an autoregressive process. Simultaneously, the acoustic emission signal record can be separated into two steady-state process sequences (a noise sequence and a signal sequence), with the signal's initial arrival point serving as the boundary point separating these two steady-state process sequences.

[0075] TransGAN is a generative adversarial model that replaces the traditional convolutional network with a Transformer structure. It consists of a generator and a discriminator and captures the feature distribution of real data through an adversarial learning mechanism to achieve the generation of high-fidelity synthetic samples.

[0076] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0077] Example 1

[0078] This application proposes a method for locating broken wires in prestressed steel strands, including:

[0079] The collected acoustic emission signals of broken steel strands are subjected to noise reduction preprocessing.

[0080] The abrupt change location search is performed on the preprocessed acoustic emission signal based on the Hinkley criterion.

[0081] An analysis window is constructed based on the mutation location, and the location of the first arrival point is determined based on the AIC criterion. The corresponding sampling time is the arrival time of the first wave.

[0082] In the AIC criterion, a boundary penalty term is introduced to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result. Formula (1) is its expression:

[0083] (1)

[0084] In the formula, For penalty weighting; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size;

[0085] Based on the arrival time of the first wave, the time difference between the arrival of the same source acoustic emission signal at different acoustic emission sensors is calculated to determine the specific location of the broken wire in the steel strand.

[0086] Acoustic emission signals generated by broken steel strands are easily affected by environmental interference, sensor noise, and scattering from the concrete medium when propagating in the bridge structure, leading to signal distortion and interfering with subsequent arrival point identification. Noise reduction processing filters out irrelevant noise signals while retaining the core characteristics of the effective acoustic emission signals generated by broken steel strands. This provides a high-quality signal foundation for subsequent arrival point search and first wave arrival time identification, avoiding identification errors caused by noise.

[0087] The acoustic emission signal generated by broken steel strands is a continuous abrupt change signal, and the first arrival point is the key node where the signal abruptly changes from noise to effective signal. The Hinkley criterion captures the location of the signal abrupt change by quantifying the degree of abrupt change in the acoustic emission signal at different sampling points and constructing a statistical measure. It divides the signal segment into two parts, calculates the difference in average amplitude between the two parts, quantifies the abrupt change intensity at the segmentation point, and after traversing all segmentation points, takes the segmentation point with the largest abrupt change intensity, i.e., the maximum value of the statistical measure, as the approximate location of the first arrival point. This achieves rapid coarse localization of the first arrival point, narrows the analysis range for subsequent fine localization, and improves localization efficiency.

[0088] Using the coarse localization abrupt change location determined by the Hinkley criterion as a benchmark, a fixed-length analysis window is constructed to focus on signal segments near the first arrival point, further eliminating irrelevant signal interference. The core function of the AIC criterion is to quantify the difference in statistical characteristics between the two segments before and after the signal segmentation point. The first segment is dominated by noise, while the second segment is dominated by effective acoustic emission signals. When the segmentation point is the true first arrival point, the difference in statistical characteristics between the two segments is the greatest, corresponding to the minimum AIC value. However, small sample data fluctuations exist at the boundary of the analysis window, which can easily lead to pseudo-minimum points in the AIC value, interfering with the precision localization accuracy. Therefore, a boundary penalty term is introduced to smooth the segmentation points near the window edge with insufficient effective sample size, suppressing the influence of small sample fluctuations and ensuring that the segmentation point corresponding to the local minimum of the AIC value is the true first arrival point, and the corresponding sampling time is the accurate first wave arrival time.

[0089] Acoustic emission signals generated by a broken wire in the same steel strand propagate at a fixed speed within the bridge structure or propagation medium, but their arrival times at different acoustic emission sensors differ. Based on the accurately identified arrival time of the first wave, this time difference is calculated. Combined with the propagation speed of the acoustic emission signal in the bridge medium and the placement relationship of each sensor, the source of the homogeneous acoustic emission signal, i.e., the specific location of the broken wire in the steel strand, is deduced, thus completing the wire breakage localization. For scenarios where the wave velocity is unknown in the engineering site, the statistical characteristics of the homogeneous acoustic emission signal, including the time difference, can be used to predict the location of the broken wire in the prestressed duct using a pre-trained machine learning model.

[0090] Example 2

[0091] like Figure 1 As shown, this embodiment, based on a standard precast high-speed railway box girder with a length of 32.6m, provides a specific implementation method for locating broken wires in prestressed steel strands, including the following steps.

[0092] Step S1: Signal acquisition and preprocessing.

[0093] The beam structure is a post-tensioned prestressed concrete simply supported box girder with a concrete strength grade of C50. The total length of the girder is 32.6m, the calculated span is 31.5m, the top slab width is 12.6m, the bottom slab width is 5.4m, and the beam height is 3.0m. The box girder has a total of 17 prestressing ducts, and each duct contains prestressed steel strands.

[0094] like Figure 2 As shown, in the cross-sectional direction of the box girder, the 17 prestressed ducts are connected in series via black hole waveguide rods (ABH waveguide rods) to achieve long-distance transmission and location monitoring of the acoustic emission signal of broken wires. The waveguide rods are connected to the anchors using a hose clamp method. RS-2A type acoustic emission sensors are deployed at the acoustic black hole structures at both ends, and high-vacuum silicone grease is used as the coupling medium between the acoustic emission sensors (AE sensors) and the waveguide rods. The acoustic emission signal acquisition system uses the AE-DS5 series full-information acoustic emission analyzer manufactured by Beijing Ruandao Technology Co., Ltd., with a threshold value set to 40dB, a sampling frequency of 3MHz, and synchronous recording of the acoustic emission waveform signal. The PDT, HDT, and HLT parameters are set to 300µs, 600µs, and 1000µs, respectively.

[0095] Step S2: Two-stage first arrival time (TOA) identification method.

[0096] First, the Hinkley statistic is used to roughly locate the arrival time of the first wave on the denoised signal to obtain the initial estimated location of the arrival time. Then, using this location as the right boundary of the analysis window, the variance-type AIC objective function with a boundary penalty term is minimized to achieve a fine location of the arrival time of the first wave.

[0097] In the first stage, the Hinkley statistic is used to quickly search for signal abrupt change locations to determine the initial value of the arrival time of the first wave. The Hinkley statistic is a mutation detection method based on the characteristics of the cumulative mean change in a time series. Its basic idea is to identify abrupt change points by analyzing the trend of the signal mean evolution over time.

[0098] For the acoustic emission signal after wavelet denoising, based on the time sequence of the acoustic emission signal, several sampling points are preset; one sampling point is selected as the segmentation point k, and the M sampling points in its neighborhood signal segment are divided into the front segment and the back segment; using formula (2), the Hinkley statistic H(k) corresponding to the segmentation point k is calculated to quantify the degree of change of the acoustic emission signal at the segmentation point k;

[0099] (2)

[0100] In the formula, x t The amplitude value at sampling point t. The average amplitude value of M sampling points; This represents the average amplitude value of the sampling points in the preceding segment, including sampling points 1, 2...k. This represents the average amplitude value of the sampling points in the latter part, including k+1, k+2…M;

[0101] Iterate through all split points, and the split point k0 corresponding to the maximum value of the Hinkley statistic is the mutation position.

[0102] In the second stage, using the mutation location obtained by the Hinkley method as a benchmark, an analysis window is constructed in its neighborhood, and the AIC criterion is introduced within this window to finely search and determine the first wave arrival time. The AIC method is a statistical detection method that automatically identifies the first wave arrival time by minimizing the weighted logarithmic sum of the variances of the signals before and after the segmentation point. For time-series signals containing first wave arrival characteristics, the signal amplitude fluctuation characteristics and statistical variance will change significantly at the actual arrival location. Therefore, by segmenting the signal sequence at different candidate times and calculating the variances of the two segments before and after the segmentation point, the weighted logarithmic sum of the variances on both sides can be used as the objective function to evaluate the probability of the first wave arrival corresponding to each candidate segmentation point.

[0103] Specifically, an analysis window of fixed length is constructed within the neighborhood of the mutation location as the right boundary, and N sampling points are preset within the analysis window;

[0104] Choose any sampling point as the segmentation point k, and divide the local segment of the acoustic emission signal within the window into a front segment and a rear segment; the lengths of the front segment and the rear segment are divided asymmetrically or symmetrically based on the segmentation point k.

[0105] Formula (3) is used to calculate the AIC value corresponding to the segment k, which is used to quantify the statistical differences between the front and back segments;

[0106] (3)

[0107] In the formula, This represents the variance of the signal amplitude values ​​in the preceding segment; The variance of the signal amplitude values ​​in the latter part;

[0108] Iterate through all the dividing points, and the dividing point corresponding to the local minimum value is the location of the initial arrival point.

[0109] However, the two ends of the analysis window are susceptible to small sample effects, leading to spurious minima, especially in short time windows and high-noise environments. To overcome this problem, a boundary penalty term is introduced into the AIC criterion, and formula (1) is its expression:

[0110] (1)

[0111] In the formula, The preferred value for the penalty weight is 0.1; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size.

[0112] Specifically, to verify the effectiveness of the two-stage first arrival time (TOA) identification method, 100 real steel strand fracture acoustic emission samples were collected for each prestressed duct, and a total of 1,700 real steel strand fracture acoustic emission samples were collected for 17 prestressed ducts. Then, 34 wire breakage signal samples were randomly selected.

[0113] To reduce background noise and smooth the signal waveform, threshold denoising preprocessing based on Discrete Wavelet Transform (DWT) was performed on the acquired broken wire acoustic emission signal before TOA extraction. Specifically, the sym8 wavelet basis was selected and a 6-level decomposition was performed. The noise level was estimated using the median absolute deviation (MAD) of the highest level detail coefficients, expressed as follows:

[0114] (4)

[0115] in, This represents the detail coefficients obtained from the highest-level wavelet decomposition.

[0116] Based on the noise estimation results, a soft threshold is then constructed: (5)

[0117] And for each layer of detail coefficients Shrinkage is performed using a soft threshold function:

[0118] (6)

[0119] The denoised signal is then obtained through wavelet reconstruction, which provides a more stable input signal for subsequent feature curve calculation and TOA recognition.

[0120] Specifically, four TOA detection methods were selected as comparison schemes: manual picking results, traditional AIC method, STA-LTA method, and Hinkley criterion. The detection performance of different methods was compared and analyzed.

[0121] It should be noted that the setting of the long and short time window parameters in the STA-LTA method and the parameter values ​​in the Hinkley criterion have a significant impact on the final detection results. However, there is still a lack of mature and unified parameter selection standards for the acoustic emission signal of broken wires in prestressed steel strands. Therefore, through trial and error, the relevant parameters were adjusted and screened in multiple rounds to finally determine a relatively stable and high-performance parameter combination. The corresponding hyperparameter settings are shown in Table 1.

[0122] Table 1 Hyperparameter settings for different methods

[0123]

[0124] Figure 3 The table shows the waveform changes of selected prestressed duct wire breakage signal samples before and after threshold denoising processing, as well as the identification results of the proposed method. The test results of the traditional AIC method, Hinkley criterion, STA-LTA method, and the method proposed in this embodiment are shown in Table 2 and... Figure 4 As shown.

[0125] Table 2. Absolute differences between the four methods and manual annotation

[0126]

[0127] Depend on Figure 4 As shown in (a), the average first arrival time obtained by the proposed method in this embodiment is 39199 μs, which is basically consistent with the manually labeled result of 39198 μs. In contrast, the traditional AIC method (42216 μs) and the Hinkley criterion (43489 μs) are generally too large, and the STA-LTA method (39632 μs) also has some deviation. This result shows that the proposed method is superior to the above-mentioned traditional methods in terms of first arrival time identification accuracy. In addition, the AIC and Hinkley methods showed more outliers in the detection results, while the number of outliers in the proposed method was significantly less, further indicating that the proposed method has better stability and robustness under complex acoustic emission signal conditions.

[0128] Figure 4 Figure (b) shows the absolute error distribution of different methods at the sample level. It can be seen that the error curve of the proposed method is generally closely distributed near zero, with small fluctuations across the entire sample range and no significant deviation. Combined with Table 2, the maximum error between the proposed method and the manually labeled results is only 14 μs, and the average error in the randomly selected 34 samples is 3.66 μs. If converted to the commonly used elastic wave propagation speed of 3200 m / s, the corresponding spatial positioning error is approximately 1.17 cm, which is sufficiently applicable to the engineering scale requirement of a minimum spacing of 40 cm for prestressed ducts in high-speed railway box girders. In contrast, the AIC detection method showed an error as high as 10 μs in some samples. 5 The proposed method exhibits an abnormal error peak of μs, with an overall error mean of 3018.4 μs, while STA-LTA and Hinkley show persistent systematic bias. The magnified local plot further reveals that the proposed method maintains the smallest error amplitude and the most stable fluctuation range at the detailed scale.

[0129] Step S3: Construction and data augmentation based on multidimensional feature vector dataset.

[0130] A stratified random sampling method was used to divide the 1700 real steel strand fracture acoustic emission samples in the original dataset into a training set of 1190 samples and a test set of 510 samples at a ratio of 7:3. Based on the arrival times of the first wave at both ends of the 1190 samples in the training set obtained in step S2, the arrival time difference characteristics of the broken wire acoustic emission signal were calculated, and further time-domain and frequency-domain features such as energy ratio, amplitude ratio, dominant frequency, root mean square value, duration, rise time, and spectral entropy were extracted from the acoustic emission signals at both ends. A total of 13 features were selected to construct a multi-dimensional feature vector to reflect the differences in acoustic emission characteristics between different prestressed ducts when wires break.

[0131] A TransGAN data augmentation strategy with category-conditional constraints was introduced on the training set. Based on the 13 features corresponding to the 1190 samples in the training set, the TransGAN generative model was used to expand the data, generating 3570 samples and corresponding simulation features that are consistent with the statistical characteristics of the 13 features. Simulation multidimensional feature vectors were then constructed based on the simulation features.

[0132] During the training of TransGAN, the network parameters were fine-tuned in multiple rounds using a trial-and-error approach. The main parameters that were finally determined were Epochs=300, Batchsize=256, Generatorlr=2e-4, and Discriminatorlr=2e-4.

[0133] like Figure 5 As shown, TransGAN is a generative adversarial model that replaces the traditional convolutional network with a Transformer structure. It consists of a generator and a discriminator, and captures the feature distribution of real data through an adversarial learning mechanism to generate high-fidelity synthetic samples. Unlike traditional CTGAN, which mainly relies on convolutional structures, TransGAN introduces multi-head self-attention (MHSA) mechanisms in both the generator and discriminator to model long-range dependencies between features, thereby maintaining statistical consistency among complex features even with limited sample conditions.

[0134] Specifically, to verify the authenticity and practical value of the TransGAN-generated data, this embodiment conducts a comprehensive evaluation from two aspects: statistical distribution consistency and applicability to downstream tasks. These two types of indicators are also core evaluation dimensions commonly used in the current research field of synthetic data. Statistical similarity is mainly analyzed by comparing the distribution characteristics of real and synthetic data through data value range, mean, standard deviation, and SDMetrics. The SDMetrics quality assessment system, proposed by MIT, aims to quantify the degree of fit between synthetic data and the attribute characteristics of real data and has been widely used in data augmentation effect evaluation research. This assessment system systematically analyzes the quality of synthetic data from multiple dimensions, mainly including two types of indicators: column distribution consistency and inter-column correlation. Column distribution consistency is used to characterize the degree of distribution matching of a single feature between real and synthetic data, while inter-column correlation is used to assess the ability of synthetic data to maintain the statistical relationship between different features. Therefore, the SDMetrics report can provide reliable quantitative support for the authenticity and effectiveness of synthetic data.

[0135] As shown in Table 3 and Figure 6 As shown, the synthetic data are highly consistent with the real data in terms of statistical characteristics such as value range, mean, and standard deviation.

[0136] Table 3 Statistical Information of Real and Synthetic Data

[0137]

[0138] like Figure 7 As shown, based on the above analysis, to further visually compare the differences between synthetic and real data in the distribution of single-column features, four representative features were selected for comparative analysis: feature 7 (Rise time-1), feature 4 (Dominant frequency-1), feature 12 (Rise time-2), and feature 9 (Dominant frequency-2), and corresponding column marginal distribution comparison curves were plotted. The distributions of these four features in the real and synthetic data show high consistency. Overall, the two sets of data exhibit highly overlapping characteristics in terms of distribution shape, peak position, and probability density variation trend, indicating that the generative model can accurately learn the statistical distribution patterns of real samples and achieve effective approximation of the real data distribution. Simultaneously, this result also demonstrates that the discriminator based on the Transformer architecture has good modeling capabilities in capturing data distribution characteristics.

[0139] Step S4: Training and validation of the four localization models.

[0140] Multidimensional feature vectors and simulated multidimensional feature vectors were used as inputs to machine learning models. XGBoost, LightGBM, CatBoost, and RF machine learning models were trained to transform the problem of locating broken wires in prestressed steel strands into a multi-class classification problem. The machine learning models learned the differences in acoustic emission signal characteristics when wires broke in different prestressed ducts, thereby locating the prestressed duct number where the broken wire was located. All four machine learning models used 10-fold cross-validation. To determine the optimal performance of the models, Bayesian optimization was used for hyperparameter optimization. The optimal hyperparameter tuning results are shown in Table 4.

[0141] Table 4 Optimal hyperparameter settings for different models

[0142]

[0143] This embodiment introduces four commonly used evaluation metrics: Accuracy, Recall, Precision, and F1-score, to evaluate the performance of four models. Their calculation formulas are shown below:

[0144] (7)

[0145] (8)

[0146] (9)

[0147] (10)

[0148] Among them, TP (True Positive): the number of positive classes predicted as positive; TN (True Negative): the number of negative classes predicted as negative; FP (False Positive): the number of negative classes incorrectly predicted as positive; and FN (False Negative): the number of positive classes incorrectly predicted as negative.

[0149] Tables 5 and 6 present the comparison results of the classification performance of the four learning models before and after data augmentation on the training and test sets, respectively. Table 5 shows that all models achieved high recognition accuracy on the training set, but their generalization performance on the test set varied significantly. XGBoost had the highest accuracy on the test set at 94.31%; CATBoost and RF followed, at 92.94% and 92.35%, respectively; LightGBM had a relatively low accuracy on the test set at 90.39%. These results indicate that under conditions of limited original sample size and insufficient distribution coverage, different models exhibit varying degrees of adaptability to unknown samples.

[0150] Based on this, to further improve the generalization stability of the model, a data augmentation strategy was introduced only on the training set. The performance results after augmentation are shown in Table 6. Figure 8 The corresponding confusion matrices for the test set are provided. After data augmentation, all four models maintained high performance on the training set, with XGBoost achieving Precision, Recall, and F1 scores all exceeding 99.6%. On the test set, the accuracy of XGBoost, CATBoost, and LightGBM improved to 96.67%, 95.10%, and 94.31%, respectively, representing increases of 2.36, 2.16, and 3.92 percentage points compared to before augmentation. In contrast, RF's accuracy only slightly improved to 92.55%, an increase of 0.20 percentage points.

[0151] Combination Figure 8 The confusion matrix shown reveals a clear concentration of predictions across the models along the main diagonal, indicating that most samples are correctly classified, with only a few misclassifications occurring between adjacent prestressed ducts. This phenomenon primarily stems from the small spacing between adjacent prestressed ducts, resulting in high similarity in the feature distribution of the acoustic emission signals received by the sensors, leading to some feature overlap among boundary samples. Furthermore, this result demonstrates that the overall prediction performance of each model on the test set is more stable after introducing data augmentation strategies. A comprehensive comparison of the classification performance of different models reveals that XGBoost exhibits the best performance in feature representation, decision boundary partitioning, and generalization, enabling more accurate differentiation of acoustic emission signals corresponding to different ducts. Therefore, XGBoost is selected as the final classification model for wire breakage localization. The feature system and modeling strategy constructed based on this model demonstrate good separability and stability, providing reliable technical support for the intelligent localization of wire breakage in prestressed ducts of high-speed railway box girders.

[0152] Table 5 compares the performance of the four models on the training and test sets without data augmentation.

[0153]

[0154] Table 6 compares the performance of the four models on the training and test sets after data augmentation.

[0155]

[0156] Example 3

[0157] This embodiment provides a specific implementation of a prestressed steel strand broken wire positioning system, including:

[0158] The signal processing module performs noise reduction preprocessing on the collected acoustic emission signal of broken steel strands;

[0159] The first wave arrival time identification module searches for abrupt change locations in the preprocessed acoustic emission signal based on the Hinkley criterion.

[0160] An analysis window is constructed based on the mutation location, and the location of the first arrival point is determined based on the AIC criterion. The corresponding sampling time is the arrival time of the first wave.

[0161] In the AIC criterion, a boundary penalty term is introduced to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result. Formula (1) is its expression:

[0162] (1)

[0163] In the formula, For penalty weighting; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size;

[0164] The positioning module calculates the time difference between the arrival times of the first wave and the arrival times of the same acoustic emission signals at different acoustic emission sensors to determine the specific location of the broken wire in the steel strand.

[0165] In one alternative embodiment, it further includes:

[0166] The signal acquisition module, which is communicatively connected to the signal processing module, includes a waveguide rod and an acoustic emission sensor;

[0167] The waveguide rod has a preset length and is used to connect prestressed steel strands located at different channel positions;

[0168] The two acoustic emission sensors are fixedly installed at the beginning and end of the waveguide rod, respectively, to collect acoustic emission signals generated by broken steel strands in different prestressed ducts.

[0169] This embodiment also provides an electronic device, which may be a server. The electronic device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities; the memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database, while the internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium; the database stores all data required for the prestressed strand broken wire location method; and the network interface is used for communication with external terminals via a network connection. The processor is capable of executing the prestressed strand broken wire location method described in embodiments 1-2.

[0170] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the prestressed steel strand broken wire positioning method described in embodiments 1-2.

[0171] 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 preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for locating broken wires in prestressed steel strands, characterized in that, include: The collected acoustic emission signals of broken steel strands are subjected to noise reduction preprocessing. The abrupt change location search is performed on the preprocessed acoustic emission signal based on the Hinkley criterion. An analysis window is constructed based on the mutation location, and the location of the first arrival point is determined based on the AIC criterion. The corresponding sampling time is the arrival time of the first wave. In the AIC criterion, a boundary penalty term is introduced to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result. Formula (1) is its expression: （1） In the formula, For penalty weighting; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size; Based on the arrival time of the first wave, the time difference between the arrival of the same source acoustic emission signal at different acoustic emission sensors is calculated to determine the specific location of the broken wire in the steel strand.

2. The method for locating broken wires in prestressed steel strands according to claim 1, characterized in that, The step-change location search of the preprocessed acoustic emission signal based on the Hinkley criterion includes: Based on the timing of the preprocessed acoustic emission signal, several sampling points are preset; Choose any sampling point as the segmentation point k, and divide the M sampling points in its neighborhood signal segment into the front segment and the back segment; Formula (2) is used to calculate the Hinkley statistic H(k) corresponding to the segmentation point k, which is used to quantify the degree of abrupt change in the acoustic emission signal at the segmentation point k; （2） In the formula, x t The amplitude value at sampling point t. The average amplitude value of M sampling points; This represents the average amplitude value of the sampling points in the preceding segment, including sampling points 1, 2...k. This represents the average amplitude value of the sampling points in the latter part, including k+1, k+2…M; Iterate through all split points, and the split point k0 corresponding to the maximum value of the Hinkley statistic is the mutation position.

3. The method for locating broken wires in prestressed steel strands according to claim 2, characterized in that, The process of constructing an analysis window based on the mutation location and determining the location of the first arrival point based on the AIC criterion includes: Using the mutation location as the right boundary, a fixed-length analysis window is constructed in its neighborhood, and N sampling points are preset within the analysis window; Choose any sampling point as the segmentation point k to divide the local segment of the acoustic emission signal within the window into the front segment and the back segment; Formula (3) is used to calculate the AIC value corresponding to the segment k, which is used to quantify the statistical differences between the front and back segments; （3） In the formula, This represents the variance of the signal amplitude values ​​in the preceding segment; The variance of the signal amplitude values ​​in the latter part; Iterate through all the dividing points, and the dividing point corresponding to the local minimum value is the location of the initial arrival point.

4. The method for locating broken wires in prestressed steel strands according to claim 3, characterized in that, The method of calculating the time difference between the arrival times of the same acoustic emission signals at different acoustic emission sensors based on the arrival time of the first wave, and determining the specific location of the broken wire in the steel strand, includes: Two acoustic emission sensors were used to collect acoustic emission signals from the same source generated by broken steel strands in different prestressed ducts, and the characteristics of the same source acoustic emission signals were extracted. The features include: the arrival time difference, energy ratio, and amplitude ratio of the first wave of the acoustic emission signals from the same source to the two acoustic emission sensors, as well as the dominant frequency, root mean square value, duration, rise time, and spectral entropy of the signals received by each acoustic emission sensor; Based on the aforementioned features, a multidimensional feature vector for locating broken wires in steel strands is constructed; the multidimensional feature vector is input into a machine learning model to train the model to learn the acoustic emission signal feature distribution of broken wires in steel strands corresponding to specific prestressed ducts. A pre-trained machine learning model is used to predict the location of broken wires in the prestressed ducts of the steel strands.

5. The method for locating broken wires in prestressed steel strands according to claim 4, characterized in that, Also includes: Using the characteristics of the aforementioned acoustic emission signals from the same source as samples, simulation features consistent with the statistical characteristics of real steel strand broken wire acoustic emission signals are generated based on the TransGAN generative adversarial network. Based on the simulation features, a simulated multidimensional feature vector for locating broken wires in steel strands is constructed; the simulated multidimensional feature vector and the multidimensional feature vector are input into a machine learning model to train the model to learn the acoustic emission signal feature distribution of broken wires in steel strands corresponding to specific prestressed ducts.

6. The method for locating broken wires in prestressed steel strands according to claim 5, characterized in that: The machine learning model employs an ensemble learning model, including the XGBoost algorithm, LightGBM algorithm, CATBoost algorithm, or RF algorithm.

7. A prestressed steel strand broken wire positioning system, characterized in that, include: The signal processing module performs noise reduction preprocessing on the collected acoustic emission signal of broken steel strands; The first wave arrival time identification module searches for abrupt change locations in the preprocessed acoustic emission signal based on the Hinkley criterion. An analysis window is constructed based on the mutation location, and the location of the first arrival point is determined based on the AIC criterion. The corresponding sampling time is the arrival time of the first wave. In the AIC criterion, a boundary penalty term is introduced to suppress the interference of small sample data fluctuations at the boundary of the analysis window on the first wave arrival time identification result. Formula (1) is its expression: （1） In the formula, For penalty weighting; This is a penalty term used to apply a smoothing penalty to segmentation points that are close to the edge of the analysis window or have insufficient effective sample size; The positioning module calculates the time difference between the arrival times of the first wave and the arrival times of the same acoustic emission signals at different acoustic emission sensors to determine the specific location of the broken wire in the steel strand.

8. The prestressed steel strand broken wire positioning system according to claim 7, characterized in that, Also includes: The signal acquisition module, which is communicatively connected to the signal processing module, includes a waveguide rod and an acoustic emission sensor; The waveguide rod has a preset length and is used to connect prestressed steel strands located at different channel positions; The two acoustic emission sensors are fixedly installed at the beginning and end of the waveguide rod, respectively, to collect acoustic emission signals generated by broken steel strands in different prestressed ducts.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor to enable the processor to perform the prestressed steel strand broken wire positioning method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the prestressed steel strand broken wire positioning method according to any one of claims 1 to 6.

Images