Bridge prestressed hole grouting density rapid quantitative detection method

Abstract

The application discloses a kind of bridge prestress hole grouting density rapid quantitative detection method, the method uses two-stage detection strategy, qualitative identification whether there is grouting defect in the first stage based on the multi-domain feature of impact echo signal is combined LightGBM model;In the second stage, for the measuring point identified as grouting defect in the first stage, the quantification of grouting insufficiency is realized in the second stage based on the impact echo signal spectrum and the integrated learning classifier model.The method can realize the rapid quantitative detection of bridge prestress hole grouting density while guaranteeing the objectivity of detection result, and weaken the dependence of detection result on the working experience of detection personnel.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing of internal defects in prestressed bridges, specifically to a rapid quantitative testing method for the compaction of grouting in prestressed ducts of bridges. Background Technology

[0002] Grouting of bridge prestressed ducts is a crucial step in post-tensioned construction. This process ensures that the grout fully coats the prestressed steel bars and strands, effectively transferring prestress while preventing corrosion of the prestressed tendons due to water and air, which could reduce bridge durability. Defects in the grouting of bridge prestressed ducts are often difficult to detect due to their concealed location.

[0003] The impact-echo method has become one of the more widely used methods due to its simplicity and requirement of only a single test surface. This method determines the presence of defects by analyzing the dominant frequency drift in the signal spectrum. However, the interpretation of results using this method has several problems:

[0004] (1) The number of measurement points for the impact echo method is large, and the efficiency is low when judged manually;

[0005] (2) Manual judgment is subjective and relies on the engineering experience of the testing personnel;

[0006] (3) Current technical methods can only determine whether there are defects, that is, qualitatively identify the grouting quality, but cannot specifically estimate and quantify the degree of grouting fullness when grouting defects exist. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a rapid quantitative detection method for the grout density of prestressed ducts in bridges. This method improves detection efficiency and objectivity while reducing reliance on the experience of testing personnel.

[0008] To achieve the above-mentioned technical objectives, the present invention employs the following technical means:

[0009] A rapid quantitative detection method for the grouting density of prestressed ducts in bridges employs a two-stage detection strategy that integrates rapid qualitative identification of grouting quality in the first stage with quantitative detection of grouting quality in the second stage. The method includes the following steps:

[0010] S1. The impact echo method is used to obtain the impact echo signal sample sets corresponding to full grout, empty grout, and different grout fullness indices, where the grout fullness index is defined as follows:

[0011] FI = h / d × 100%, where:

[0012] FI is the grout fullness index; h is the vertical height of the grout; d is the diameter of the corrugated pipe.

[0013] S2. In the first stage, the sample set corresponding to the full grout measurement points is labeled as full, and the samples corresponding to empty grout and different grout fullness indices are merged and labeled as unfull. The sample sets belonging to the full and unfull labels are combined into a total sample set. 80% of the samples in each category are evenly selected as the training set samples under the total sample set, and the remaining 20% ​​of the samples are used as the test set samples under the total sample set. The training set samples are used to construct a rapid qualitative identification model for grouting defects, and the test set samples are used to test the prediction effect of the rapid qualitative identification model for grouting defects.

[0014] S3. In the second stage, for the sample set labeled as unfull, the samples containing empty grout and samples corresponding to different grout fullness indices in this sample set are relabeled as 0% and the corresponding grout fullness index value, respectively. Then, the sample set to which the relabeled multi-class labels belong is combined into a new total sample set. 80% of the samples in each category of the new total sample set are evenly selected as the training set samples under the new total sample set, and the remaining 20% ​​of the samples are used as the test set samples under the new total sample set. The training set samples are used to construct a quantitative detection model for grout fullness, and the test set samples are used to test the prediction effect of the quantitative detection model for grout fullness.

[0015] S4. Detection process: For the impact echo signal of the test point, the first stage of rapid qualitative identification of grouting quality is performed. In step S2, the rapid qualitative identification model of grouting defects adopts an integrated machine learning model. Taking the multi-domain features of the signal as input, it directly outputs whether the test point is full or unfull. If the predicted output is full, it indicates that there is no grouting defect at the test point. The grouting fullness index is directly output as 100%, and the detection of the test point is skipped to the next test point. If the test point is predicted to output as unfull, it indicates that the test point is not grouting full. Then, the second stage of quantitative detection of grouting quality is entered. In step S3, the quantification detection of the degree of grouting incompleteness takes the spectrum of the signal as input and uses a deep learning model to predict and output the fullness index under the condition of grouting incompleteness.

[0016] In step S4, the integrated machine learning model adopts the LightGBM classification model, and the optimal hyperparameter combination is obtained through cross-validation and grid search methods.

[0017] The deep learning model adopts an ensemble voting classifier model, which uses four convolutional neural networks—AlexNet, MobileNet, SqueezeNet, and ShuffleNet—as base models. It uses an average soft voting method to predict the final output result based on the output class of each model and the probability corresponding to the output class.

[0018] The LightGBM classification model uses the multi-domain features of the impulse echo signal as input, including time-domain features, frequency-domain features, and time-frequency-domain features.

[0019] The integrated voting classifier model uses a two-dimensional spectrogram obtained by processing the impulse echo signal through continuous wavelet transform as input.

[0020] The time-domain features include the waveform's peak factor, waveform factor, impulse factor, margin factor, and kurtosis factor.

[0021] The frequency domain features include: average frequency domain amplitude, root mean square frequency value, standard deviation of frequency value, and center frequency value;

[0022] The time-frequency domain feature is an energy sequence vector [E] composed of the relative energy values ​​of signals within the frequency range of 2-20kHz, at 2kHz intervals, i.e., within the sub-frequency bands of 2-4kHz, 4-6kHz, 6-8kHz, ..., 18-20kHz. 2-4kHz E 4-6kHz ,...,E 18-20kHz ].

[0023] Beneficial effects:

[0024] The present invention provides a rapid quantitative detection method for the compaction of grouting in prestressed ducts of bridges, which has the following advantages compared with existing grouting defect detection methods:

[0025] 1. This invention can quickly identify the presence or absence of grouting defects, with high detection efficiency, and is suitable for rapid screening and location of grouting defects in prestressed ducts;

[0026] 2. The test results of this invention are highly objective, which reduces the subjectivity of the testers in judging the results and the dependence of the testers' engineering experience on the interpretation of the test results;

[0027] 3. This invention can not only quickly diagnose and locate grouting defects, but also quantitatively estimate the degree of fullness or fullness index of the measuring points where grouting defects exist. Attached Figure Description

[0028] Figure 1 The two-stage strategy detection flowchart provided by this invention;

[0029] Figure 2 Simulation diagram of the web of a prestressed box girder for bridges with different grouting fullness levels;

[0030] Wherein, 1 represents the reinforcing bar; 2 represents the corrugated pipe; and 3 represents the web.

[0031] Figure 3 Signal waveforms and spectrum diagrams for different degrees of grout fullness;

[0032] Figure 4 These are the model test results from the first stage of rapid identification of grouting defects;

[0033] Figure 5 Test the accuracy of the model under different domain feature selections;

[0034] Figure 6 The graph shows the quantitative test results of the degree of incomplete grouting under the condition of incomplete grouting in the second stage.

[0035] Figure 7 To quantify the prediction accuracy of the second-stage grouting fullness. Detailed Implementation

[0036] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0037] A rapid quantitative detection method for the grouting density of prestressed ducts in bridges is proposed. This method employs a two-stage detection strategy and includes the following steps:

[0038] First, impact echo signals corresponding to full grout, empty grout, and different grout fullness levels were obtained using the impact echo method to form a total sample set. Based on the grout fullness label, 80% of the sample sets under each fullness category were selected as the training set, and the remaining 20% ​​were used as the test set. The grout fullness level can be represented by the grout fullness index FI = h / d, where FI is the grout fullness index; h is the vertical height of the grout; and d is the bellows diameter.

[0039] Furthermore, samples corresponding to full grouting points are labeled as "full," while samples corresponding to empty grouting points and different grouting fullness are merged and labeled as "unfull." Multi-domain features of the impact echo signal are calculated, including time-domain, frequency-domain, and time-frequency-domain features. Time-domain features include waveform peak factor, shape factor, impulse factor, margin factor, and kurtosis factor. Frequency-domain features include the average frequency amplitude, root mean square frequency value, standard deviation frequency value, and center frequency value. Time-frequency-domain features are the relative energy values ​​of the signal within a specific frequency range. Using the multi-domain feature parameter vector as input and the labels "full" and "unfull" as outputs, the LightGBM model for rapid qualitative identification of grouting defects in the first stage is constructed and trained using training set samples. The optimal hyperparameter combination is obtained through cross-validation and grid search methods.

[0040] Furthermore, the sample signals corresponding to empty grout and different grout fullness are labeled as 0% and the corresponding grout fullness index, respectively. The two-dimensional spectrum of the signal obtained by continuous wavelet transform is used as input, and the grout fullness index is used as output. Four convolutional neural network models, namely AlexNet, MobileNet, SqueezeNet and ShuffleNet, are used as base models to construct and train an integrated voting classifier model.

[0041] Furthermore, such as Figure 1 As shown, for a sample signal to be tested, the multi-domain feature value of the signal is first calculated as the input of the LightGBM model, and the output prediction label is full or unfull. If the output label is full, it indicates that there is no grouting defect at the measuring point, and the grouting fullness index is directly output as 100% and the next measuring point is detected. If the output prediction label is unfull, it indicates that the measuring point is not grouting, and the second stage of incompleteness quantification detection is entered. The spectrum of the sample signal to be tested is used as input, and the ensemble voting classifier model is used to output the fullness index under the condition of incomplete grouting. Example

[0042] The present invention will be further illustrated by the following embodiments, but these should not be construed as limiting the scope of protection of the present invention.

[0043] like Figure 2 As shown, concrete slabs with different degrees of voids are preset to simulate the web of a prestressed box girder of a bridge under different degrees of grouting fullness. The specific grouting fullness indices include: 0%, 20%, 40%, 60%, 80%, and 100%.

[0044] like Figure 3 As shown, 500 sample signals for each working condition were obtained using the impulse echo method, totaling 3000 signal samples. The waveforms and spectrograms of these 6 types of sample signals are representative.

[0045] For the first stage of detection, 500 samples with a fullness index of 100% were labeled as "full," and 2500 samples with fullness indices of 0%, 20%, 40%, 60%, and 80% were labeled as "unfull." 80% of the samples from both labels were randomly selected as the training set, and the remaining 20% ​​as the test set. The multi-domain feature parameters of the sample signals were used as input, and the training and testing structures were as follows: Figure 4 The confusion matrix shown indicates that the accuracy of the first-stage rapid qualitative identification reached 98%.

[0046] Figure 5As shown, the test accuracy of models using different domain features as inputs reveals that the test accuracy when using multi-domain features as inputs is significantly higher than the test accuracy when using only the time domain, frequency domain, or time-frequency domain.

[0047] Five sample signals, representing 0%, 20%, 40%, 60%, and 80% respectively, were labeled as corresponding grout fullness indices. Using the two-dimensional spectrogram of the signals as input, a soft-voting classifier model was integrated using four base models: AlexNet, MobileNet, SqueezeNet, and ShuffleNet. Similarly, 80% (400 samples) of each class were selected as the training set, and 20% (100 samples) of each class were selected as the test set. The integrated soft-voting classifier model was trained and tested, and the prediction results on the test set are shown below. Figure 6 The confusion matrix shown.

[0048] like Figure 7 As shown, the prediction accuracy of the statistical model shows that the prediction accuracy of the second-stage ensemble soft-voting classifier model is significantly higher than that of the individual models.

Claims

1. A rapid quantitative detection method for the grouting density of prestressed ducts in bridges, characterized in that, A two-stage detection strategy is adopted, which integrates rapid qualitative identification of grouting quality in the first stage and quantitative detection of grouting quality in the second stage. The strategy includes the following steps: S1. The impact echo method is used to obtain the impact echo signal sample sets corresponding to full grout, empty grout, and different grout fullness indices, where the grout fullness index is defined as follows: FI = h / d × 100%, where: FI is the grout fullness index; h is the vertical height of the grout; d is the diameter of the corrugated pipe. S2. In the first stage, the sample set corresponding to the full grout measurement points is labeled as full, and the samples corresponding to empty grout and different grout fullness indices are merged and labeled as unfull. The sample sets belonging to the full and unfull labels are combined into a total sample set. 80% of the samples in each category are evenly selected as the training set samples under the total sample set, and the remaining 20% ​​of the samples are used as the test set samples under the total sample set. The training set samples are used to construct a rapid qualitative identification model for grouting defects, and the test set samples are used to test the prediction effect of the rapid qualitative identification model for grouting defects. S3. In the second stage, for the sample set labeled as unfull, the samples containing empty grout and samples corresponding to different grout fullness indices in this sample set are relabeled as 0% and the corresponding grout fullness index value, respectively. Then, the sample set to which the relabeled multi-class labels belong is combined into a new total sample set. 80% of the samples in each category of the new total sample set are evenly selected as the training set samples under the new total sample set, and the remaining 20% ​​of the samples are used as the test set samples under the new total sample set. The training set samples are used to construct a quantitative detection model for grout fullness, and the test set samples are used to test the prediction effect of the quantitative detection model for grout fullness. S4. Detection process: For the impact echo signal of the test point, the first stage of rapid qualitative identification of grouting quality is performed. In step S2, the rapid qualitative identification model of grouting defects adopts an integrated machine learning model. Taking the multi-domain features of the signal as input, it directly outputs whether the test point is full or unfull. If the predicted output is full, it indicates that there is no grouting defect at the test point. The grouting fullness index is directly output as 100%, and the detection of the test point is skipped to the next test point. If the test point is predicted to output as unfull, it indicates that the test point is not grouting full. Then, the second stage of quantitative detection of grouting quality is entered. In step S3, the quantification detection of the degree of grouting incompleteness takes the spectrum of the signal as input and uses a deep learning model to predict and output the fullness index under the condition of grouting incompleteness.

2. The rapid quantitative detection method for grout density of prestressed ducts in bridges according to claim 1, characterized in that, In step S4, the integrated machine learning model adopts the LightGBM classification model, and the optimal hyperparameter combination is obtained through cross-validation and grid search methods.

3. The rapid quantitative detection method for grout density of prestressed ducts in bridges according to claim 1, characterized in that, The deep learning model employs an ensemble voting classifier model, which uses four convolutional neural networks—AlexNet, MobileNet, SqueezeNet, and ShuffleNet—as base models. It predicts the final output result using an average soft voting method based on the output class of each model and the probability corresponding to that output class.

4. The rapid quantitative detection method for grout density of prestressed ducts in bridges according to claim 2, characterized in that, The LightGBM classification model uses the multi-domain features of the impulse echo signal as input, including time-domain features, frequency-domain features, and time-frequency-domain features. The integrated voting classifier model uses a two-dimensional spectrogram obtained by processing the impulse echo signal through continuous wavelet transform as input.

5. The rapid quantitative detection method for grout density of prestressed ducts in bridges according to claim 4, characterized in that, The time-domain features include the waveform's peak factor, waveform factor, impulse factor, margin factor, and kurtosis factor. The frequency domain features include: average frequency domain amplitude, root mean square frequency value, standard deviation of frequency value, and center frequency value; The time-frequency domain feature is an energy sequence vector [E] composed of the relative energy values ​​of signals within the frequency range of 2-20kHz, at 2kHz intervals, i.e., within the sub-frequency bands of 2-4kHz, 4-6kHz, 6-8kHz, ..., 18-20kHz. 2-4kHz E 4-6kHz ,...,E 18-20kHz ].

Images