Filling compaction quality classification evaluation method based on elastic wave frequency dispersion image

Abstract

The invention relates to the technical field of earth-rock filling construction, in particular to a filling compaction quality classification evaluation method based on elastic wave frequency dispersion images, which comprises the following steps: carrying out fluctuation test on earth-rock filling materials with different compaction qualities, and collecting elastic wave signals; the compaction quality of the earth-rock filling material is determined by the dry density and porosity of a filling body; converting the elastic wave signal into a frequency dispersion image based on multiple wavelet transformations; constructing a training data set for model training; the training data set comprises a frequency dispersion image with a preset compaction quality grade label; and constructing a convolutional neural network, training based on the training data set to obtain a compaction quality evaluation model, and evaluating the compaction quality of the to-be-tested earth-rock filling material by taking the frequency dispersion image of the to-be-tested earth-rock filling material as input. A frequency dispersion image of an elastic wave signal is used as a sensitive representation of the compaction quality of the earth-rock filling material, and feature learning and classification are performed on the frequency dispersion image through a convolutional neural network model, so that lossless intelligent identification of different compaction grades is realized.

Description

Technical Field

[0001] This application relates to the field of earthwork filling construction technology, and more specifically, to a method for classifying and evaluating the compaction quality of filling based on elastic wave dispersion images. Background Technology

[0002] Earth-rock embankment projects are widely used in modern water conservancy projects and high dam construction due to their safety, reliability, strong construction adaptability, and excellent seismic performance. The main load-bearing material in earth-rock embankment projects is the earth and rock fill material, and its compaction quality directly determines the deformation characteristics, seepage resistance, and overall stability of the embankment project. Poorly compacted areas often become potential locations for stress concentration and leakage in earth-rock embankment projects. Therefore, accurately evaluating the compaction quality of earth-rock embankment projects is crucial to ensuring the long-term safe operation of dams.

[0003] Traditional methods for testing compaction quality mainly include the water-filling method, the ring cutter method, and the nuclear density meter method. These methods estimate the degree of compaction by directly measuring the dry density and moisture content of the soil. While the results are intuitive, they are significantly destructive, and each test only reflects the condition of a localized point, failing to reflect the compaction uniformity over a large area. Furthermore, these methods largely rely on on-site sampling and laboratory analysis, which is not only complex and time-consuming but also significantly destructive. For example, the water-filling method requires excavating sampling points, damaging the compaction surface structure of earth-fill projects; the ring cutter method can only measure point-like compaction information and cannot reflect overall uniformity; while the nuclear density meter offers some convenience, it is restricted by radiation source management and has limited detection depth. These problems make traditional testing methods insufficient to meet the needs of modern high-fill, rapid-construction large-scale earth-fill projects.

[0004] In recent years, non-destructive testing (NDT) technology has gradually become a research hotspot for health monitoring in earthwork and rockfill engineering. Among them, the elastic wave method is considered a highly promising means of compaction quality testing because it can reflect the elastic and structural characteristics of the medium without disturbing the structure. When elastic waves propagate in earthwork and rockfill, their wave velocity, dispersion pattern, and attenuation law are all affected by porosity, particle contact stiffness, and density. Current research has used empirical relationships between P-wave or S-wave velocity and density and porosity to estimate compaction quality. For example, rapid estimation can be achieved by establishing an empirical fitting curve of wave velocity-compaction degree. However, this method ignores dispersion and attenuation characteristics and relies only on a single wave velocity index, making it difficult to capture the nonlinear characteristics of energy propagation in complex fill materials, resulting in insufficient identification accuracy. Therefore, there is an urgent need for a new evaluation method that can automatically extract the dispersion energy characteristics of elastic waves and achieve intelligent identification of compaction quality. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method for classifying and evaluating the compaction quality of earthwork based on elastic wave dispersion images. This method utilizes the dispersion images of elastic wave signals as a sensitive representation of the compaction quality of earthwork projects. By employing a convolutional neural network model to learn and classify the features of the dispersion images, it achieves non-destructive intelligent identification of different compaction levels.

[0006] The technical solution provided in this application is as follows:

[0007] A method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images includes:

[0008] Wave tests were conducted on earth and rock fill materials with different compaction qualities to collect elastic wave signals; the compaction quality of earth and rock fill projects is determined by the dry density and porosity of the earth and rock fill materials.

[0009] The elastic wave signal is converted into a dispersion image based on multiple wavelet transforms;

[0010] Construct a training dataset for model training; the training dataset includes dispersion images labeled with preset compaction quality levels;

[0011] A convolutional neural network is constructed and a compaction quality assessment model is trained based on the training dataset. The compaction quality assessment model uses the dispersion image of the soil and rock fill material to be tested as input to evaluate the compaction quality of the soil and rock fill project to be tested.

[0012] One possible implementation involves conducting wave tests on earth-rock fill materials with different compaction masses and collecting elastic wave signals, including:

[0013] The earth and rock fill material is compacted by a compaction device, and fluctuation tests are performed on several areas of the compacted surface.

[0014] During testing, an external triggering method is used for excitation. A high-frequency signal is generated by hammering a steel plate, and the elastic wave signal is received by multiple detectors arranged at equal intervals.

[0015] One possible implementation involves converting the elastic wave signal into a dispersion image based on multiple wavelet transforms, including:

[0016] In the time domain, a wavelet transform is performed on an elastic wave signal to obtain the amplitude and phase information of the wave components at different frequencies in time, and the elastic wave signal is converted to the time-frequency f domain.

[0017] Perform wavelet transform on the obtained time-frequency domain signal along the x-axis to transform the time-frequency domain signal to the frequency f-wavenumber domain;

[0018] Based on the mapping relationship between frequency and wavenumber, the frequency-wavenumber spectrum in the frequency-wavenumber domain is transformed into the dispersion energy spectrum in the fV domain.

[0019] One possible implementation involves constructing a training dataset for model training, including:

[0020] The porosity of earth and rock fill material is used as the standard for classifying compaction quality grades to classify the compaction quality of earth and rock fill projects.

[0021] Labels are set for each dispersion image based on the compaction quality grade classification results to obtain dispersion images with preset compaction quality grade labels, and the training dataset is constructed.

[0022] One possible implementation involves training a compaction quality assessment model based on the training dataset, including:

[0023] According to a preset ratio, the training dataset is divided into a training set for model training, a validation set for adjusting hyperparameters, and a validation set for testing the performance of the compaction quality assessment model.

[0024] In one possible implementation, after generating the compaction quality assessment model, the method further includes:

[0025] The loss value of the compaction quality assessment model is analyzed using the cross-entropy loss function in the multi-class case, and is expressed as:

[0026]

[0027] In the formula, The number of categories; For sign functions, the values ​​include 0 and 1. When the first... Each sample belongs to When the class is selected, the value is 1; otherwise, it is 0. Indicates the prediction result, the first The predicted probability that a sample belongs to this class; Indicates the first The loss value for each sample;

[0028] The performance of the compaction quality assessment model is evaluated based on the training loss curve and the validation loss curve.

[0029] In one possible implementation, after generating the compaction quality assessment model, the method further includes assessing the accuracy of the compaction quality assessment model. Accuracy Recall rate and Fraction The reliability of the compaction quality assessment model is measured.

[0030] One possible implementation involves evaluating the compaction quality of the earth-rock fill project under test based on a compaction quality assessment model, including:

[0031] Based on the compaction quality assessment model, the dispersion image features of the earth-rock fill project to be tested are extracted, and the compaction quality evaluation category label and probability value of the earth-rock fill project to be tested are output.

[0032] Compare the compaction quality evaluation category label with the category of actual compaction quality;

[0033] The compaction quality evaluation category is determined based on the probability value, and the compaction quality evaluation of the earth-rock filling project under test is completed.

[0034] Compared with the prior art, the technical solution provided in this application has the following beneficial effects:

[0035] (1) Non-destructive testing and high efficiency: This method only requires the acquisition of elastic wave signals, without sampling or damaging the soil and rock filling structure, and the entire testing process is non-destructive; by analyzing the physical correlation between the elastic wave dispersion image and the porosity of the soil and rock filling material, a physical characterization mechanism for compaction quality is established; through the automatic recognition of the convolutional neural network model, the compaction quality evaluation can be completed in a few minutes, and the testing efficiency is tens of times higher than that of traditional methods.

[0036] (2) Clear physical correlation: By establishing the correspondence between elastic wave dispersion energy and porosity of soil and rock fill material, the mechanistic explanation of compaction quality is realized, overcoming the "black box" characteristics of traditional empirical models.

[0037] (3) Strong robustness and applicability: The convolutional neural network algorithm is used to automatically learn the energy distribution law of the dispersion image and realize the automatic identification of different compaction levels. Through comparative research on various network structures, the VGG16 model performs stably in different classification tasks. The model has a small number of parameters and low computational cost, making it suitable for deployment on mobile terminals or embedded devices at the engineering site to achieve rapid identification of compaction quality.

[0038] (4) Significant economic and safety benefits: This method can reduce the cost of laboratory testing and manual sampling, while avoiding damage to the compaction surface, and realize non-destructive, rapid and automated compaction quality testing. It overcomes the defects of traditional methods, such as strong destructiveness, small coverage and low efficiency, and improves construction safety and testing economy, providing technical support for intelligent construction and monitoring of earth and rock filling projects. Attached Figure Description

[0039] Figure 1 This is a flowchart of a method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images, provided in Embodiment 1 of this application.

[0040] Figure 2 This is a schematic diagram illustrating the recognition effect of the VGG16 model for compaction quality classification provided in Embodiment 1 of this application.

[0041] Figure 3 This is a schematic diagram illustrating the recognition effect of the 5-category VGG16 model for compaction quality provided in Embodiment 1 of this application.

[0042] Figure 4 This is a flowchart of the test procedure for conducting fluctuation tests on earth and rock fill projects, provided in Embodiment 2 of this application.

[0043] Figure 5 This is a schematic diagram of the compaction chamber surface and its location arrangement provided in Embodiment 1 of this application.

[0044] Figure 6 This is an example diagram of dispersion image conversion provided in Embodiment 2 of this application.

[0045] Figure 7 This is a Grad-CAM visualization heatmap provided in Embodiment 3 of this application. Detailed Implementation

[0046] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0047] Example 1

[0048] See Figure 1 This is a flowchart of a method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images, provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include:

[0049] Step 101: Conduct wave tests on soil and rock fill materials with different compaction masses and collect elastic wave signals.

[0050] The compaction quality of earth-rock fill projects is determined by the dry density and porosity of the fill material. The aforementioned fluctuation test of the earth-rock fill material involves compacting the fill material with rollers to ensure that every area of ​​the compacted surface is compacted, and then performing fluctuation tests on nine areas of the compacted surface. The test uses an external triggering method, where a high-frequency signal is generated by striking a steel plate with a hammer, and the elastic wave signal is received by multiple equally spaced detectors.

[0051] Step 102: Convert the above-mentioned elastic wave signal obtained from the test into a dispersion image based on multiple wavelet transforms.

[0052] Specifically, the elastic wave signal obtained from the aforementioned wave test is the time-domain information of the elastic wave. Directly analyzing the time-domain information of the elastic wave makes it difficult to obtain the corresponding relationship between the elastic wave and the compaction quality of the soil and rock fill material. Therefore, this application performs multiple wavelet transforms on the time-domain information of the elastic wave, ultimately converting it into the dispersion energy spectrum of the elastic wave, i.e., the dispersion image, and then analyzes the corresponding relationship between the dispersion energy spectrum and the compaction quality of the soil and rock fill material.

[0053] Step 103: Construct a training dataset for model training. This training dataset includes labeled dispersive images. The labels on the dispersive images represent the compaction quality level of the corresponding earthwork filling project.

[0054] Specifically, the model is trained using supervised learning, learning from labeled training datasets to deepen its understanding of the characteristics of various compaction quality levels in earth-rock fill projects. Finally, the model uses the prediction dataset to classify and identify the compaction quality of the earth-rock fill material. Therefore, before training the model, it is necessary to segment the dispersion images of different compaction qualities of the earth-rock fill material to classify them into different compaction quality levels.

[0055] In some embodiments, the porosity of the soil-rock fill material obtained by the water-filling method is used as the basis for evaluating the compaction quality of the soil-rock fill project. The compaction quality of the soil-rock fill material is pre-classified into different grades according to project needs. The different compaction quality grades of the soil-rock fill material are used as labels for dispersion images. Labels are set for the dispersion images obtained from the fluctuation test, thereby constructing a large dataset. Optionally, the above-mentioned compaction quality grades can be set according to actual conditions, and this application embodiment does not impose specific limitations.

[0056] Step 104: Construct a convolutional neural network and train it based on the above training dataset to obtain a compaction quality assessment model.

[0057] Among them, VGG16 maintains high model accuracy with extremely low parameter count through depthwise separable convolution and inverted residual structure, and its lightweight design meets the needs of actual engineering sites.

[0058] Step 105: After obtaining the dispersion image of the soil and rock fill project to be tested according to the methods in steps 101 to 102, input the dispersion image into the compaction quality assessment model to determine the compaction quality evaluation category of the soil and rock fill material to be tested.

[0059] Specifically, following steps 101-102 as described above, the acquired one-dimensional elastic wave data is converted into a dispersion image. This dispersion image is then input into the compaction quality assessment model to identify the compaction quality evaluation category label. The compaction quality evaluation category label is compared with the category of the actual compacted mass to verify its identification effect and accuracy. Then, based on these probability values, the most probable compaction quality evaluation category is determined, which is the compaction quality evaluation category corresponding to the analyzed dispersion image, and its name and corresponding probability value are returned. However, it should be noted that the compaction quality evaluation category to be identified must be within the range of specific parameters (such as porosity, gradation, etc.) contained in the trained dataset.

[0060] To verify the model's recognition and classification performance under different classification conditions for compaction quality of earth and rock fill materials, two classification scenarios were set up. One scenario categorizes the compaction quality of earth and rock fill materials into two levels: qualified and unqualified, based on the porosity required by the earth and rock fill engineering design. The other scenario categorizes the compaction quality of earth and rock fill materials into five more refined compaction quality levels. The classification settings for compaction quality levels can be adjusted according to engineering requirements. The size and label settings of the dispersive image dataset are shown in Tables 1 and 2.

[0061] Table 1: Size and label settings of the compaction quality two-class dispersion image dataset

[0062]

[0063] Table 2: Size and label settings of the 5-category dispersion image dataset for compaction quality

[0064]

[0065] The compaction quality assessment model based on VGG16 achieved partial identification results on the compaction quality evaluation category test set, as shown below. Figure 2 , Figure 3 As shown. To better demonstrate the model training effect, this application loads the model parameters corresponding to the highest accuracy into the original model. The titles of the visualized recognition result images are the model's predicted labels and probability values, respectively. Test results show that the VGG16-based compaction quality assessment model provided in this application has the same recognition results as the actual situation, and the probability values ​​are high. It can effectively extract dispersive image features and realize the evaluation of the compaction quality of earth and rock fill materials.

[0066] Example 2

[0067] This application utilizes the dispersion image of elastic wave signals as a sensitive representation of the compaction quality in earth-rock filling engineering. A convolutional neural network model is used to learn and classify the features of the dispersion image, achieving non-destructive intelligent identification of different compaction levels. Embodiment Two of this application, based on Embodiment One, supplements and refines the content of each part of the technical solution to ensure full disclosure of the technical solution.

[0068] This application first conducts wave tests on soil and rock fill materials with different compaction masses, and uses wavelet transform to convert the obtained elastic wave signals into dispersion images.

[0069] To accomplish the above steps, a test site must be constructed in advance. For example... Figure 4 As shown, the test site was constructed using stone bricks, topped with a rain shelter, and the bottom was plastered with concrete to ensure a flat surface. After the test site was constructed, soil and rock fill materials of various grades were prepared according to the test gradations. The soil and rock fill materials were evenly laid on the test platform to a thickness of 20 cm. After the soil and rock fill materials were laid, the compaction equipment was started, and the soil and rock fill materials were compacted using a forward and backward method to ensure that every area of ​​the compacted surface was compacted. Figure 5 As shown in the illustration, the test compaction surface is uniformly divided into 9 regions in this embodiment. After every two passes of compaction, a fluctuation test is performed on each of the 9 regions of the compaction surface. By detecting the porosity of different regions of the compaction surface, more data is obtained to meet the data requirements of the convolutional neural network. The excitation method uses an external triggering method, where a high-frequency signal is generated by striking the steel plate with a hammer. Each strike completes one fluctuation test.

[0070] To collect elastic wave signals during the test, 12 geophones were arranged at equal intervals of 10 cm along the length. The seismic source and geophones were positioned on the center line of the area, with the seismic source 20 cm from the first geophone. A steel plate was placed at the seismic source. To meet the dataset size required for subsequent convolutional neural network training, this testing method was used to conduct multiple wave tests on nine areas of the compacted earth-rock fill surface.

[0071] After the fluctuation test is completed, the compaction quality, namely dry density and porosity, of the soil and rock fill material in each area is tested using the test pit water filling method. The test pit diameter is 20cm. A fluctuation test and test pit water filling test are performed after every two passes of compaction, for a total of 10 passes of compaction, requiring 5 fluctuation tests and test pit water filling tests.

[0072] Furthermore, in order to determine the relationship between elastic waves and the compaction quality of earth and rock fill, this application performs multiple wavelet transforms on the time-domain information of elastic waves, and finally converts it into the dispersion energy spectrum of elastic waves, and then analyzes the relationship between the dispersion energy spectrum and the compaction quality of earth and rock fill.

[0073] In the time domain, performing a wavelet transform on a single elastic wave signal yields the amplitude and phase information of wave components at different frequencies. Similarly, in the spatial domain, performing a similar wavelet transform on multiple elastic wave signals provides the amplitude and phase information of wave components at different wavenumbers. Therefore, by performing multiple wavelet transforms on multiple elastic wave signals, the multi-channel seismic signals can be converted into a dispersive energy spectrum, allowing for the analysis of the energy distribution of a shot gather recording in the frequency-wavenumber domain.

[0074] As a feasible approach, the first step is to record... Perform wavelet transform along the time axis, Domain signal transformation to The (time-frequency) domain is represented as:

[0075]

[0076] Then get signal edge Perform wavelet transform on the axis, and Domain signal transformation to The (frequency-wavenumber) domain is represented as:

[0077]

[0078] In the formula, Indicates wave number, This represents the frequency, and after two wavelet transforms, The signal in the domain is transformed to The domain, and because the wave number and frequency have the following relationship:

[0079]

[0080] In the formula, This represents the wave number. Therefore, it can be calculated through mapping. Frequency wavenumber spectrum transformation of the domain The dispersion energy spectrum of the frequency-wavenumber domain is used to analyze the energy distribution of a shot gather record in the frequency-wavenumber domain. Figure 6 It can be seen that the dispersion energy spectrum of the compacted material of earth and rock fill engineering obtained by this method has good dispersion energy imaging quality, good dispersion energy continuity and high phase velocity resolution, which can meet the requirements of analyzing the correlation between dispersion image and compaction quality of earth and rock fill material.

[0081] After obtaining the above dispersion images, this application classifies the compaction quality of earth and rock fill into different levels according to the requirements, sets labels for the dispersion images under different compaction quality levels, and constructs a large dataset.

[0082] In some embodiments, the total number of dispersive images obtained after multiple wavelet transforms is 372. 70% of the dispersive images are randomly selected for training; using a higher proportion of the training set helps to obtain a more robust classification model. Specifically, a convolutional neural network model is used to train the aforementioned 70% dataset constructed from the dispersive images to extract the energy distribution features of the dispersive images, thereby establishing the relationship between the dispersive images and the compaction quality of the earth-rock fill material.

[0083] Furthermore, 20% of the dispersive images were selected as a validation set for adjusting hyperparameters; a smaller validation set can speed up the tuning process. The remaining 10% of the dispersive images were used to construct a test set, which was then input into the trained model to test the classification and recognition performance of the selected convolutional neural network model for the compaction quality grade of earth and rock fill.

[0084] Example 3

[0085] This application selects the VGG16 convolutional neural network model as the compaction quality assessment model. In Embodiment 3 of this application, these three representative convolutional neural network models are used to assess the compaction quality of earth-rock filling projects, and their recognition effects are compared to verify that the VGG16 convolutional neural network model selected in this application is the best-performing convolutional neural network model.

[0086] Specifically, the global modeling advantage of the Swin Transformer helps reveal the correlation between local anomalies and the overall distribution of data features, breaking through the limitations of the local receptive field of traditional convolutional neural networks. ResNet18, as a lightweight representative of classic residual networks, solves the gradient degradation problem of deep networks with its residual skip connection structure, providing a robust benchmark for the performance of traditional convolutional neural networks. VGG16 maintains high model accuracy with extremely low parameter count through depthwise separable convolutions and inverted residual structures; its lightweight design meets the needs of practical engineering applications. These three models respectively demonstrate the global perception capability of cutting-edge Transformer architectures, the robust feature extraction characteristics of traditional CNNs, and the engineering applicability of lightweight networks.

[0087] While deep neural networks, represented by CNNs, exhibit significant advantages in feature extraction, their internal decision-making mechanisms are often considered opaque "black boxes," a characteristic that limits model reliability. To overcome this limitation, this application introduces Gradient Weighted Class Activation Mapping (Grad-CAM) technology to visualize and analyze key feature regions in the elastic wave dispersion image recognition task. Grad-CAM generates high-resolution heatmaps by calculating the global average weight of the target class gradient relative to the final convolutional layer feature map to quantify the contribution of different image regions to the model's prediction. For the three models used in Embodiment 3 of this application—Swin Transformer, ResNet18, and VGG16—the spatial distribution characteristics of their regions of interest are analyzed using heatmaps.

[0088] See Figure 7 The image shows Grad-CAM visualization heatmaps for the three models provided in Embodiment 3 of this application. The red areas in the heatmap represent the level of attention the model pays to the dispersion map at that location; the darker the color, the higher the level of attention.

[0089] Furthermore, for model training and performance evaluation, the loss value used in the network structure is analyzed using the cross-entropy loss function in the multi-class classification case, expressed as:

[0090]

[0091] In the formula, The number of categories. For sign functions, the values ​​include 0 and 1. When the first... Each sample belongs to The value is 1 if the class is selected, otherwise it is 0. Indicates the prediction result, the first The predicted probability that a sample belongs to this class. Indicates the first The loss value for each sample.

[0092] The performance of the models is evaluated using two metrics: training loss curves and validation loss curves obtained from training three different network models. The training loss curve is used to understand the performance during the model training process, while the validation loss curve is used to evaluate the model's generalization ability.

[0093] To comprehensively evaluate and compare the performance of the four network models, corresponding reliability metrics were selected to measure the model performance. Based on this, the various reliability metrics of the models were analyzed, and the performance of each model was evaluated.

[0094] Reliability indicators are selected based on accuracy. Accuracy Recall rate and Fraction The measurement is performed. The specific calculation formula is as follows:

[0095]

[0096]

[0097]

[0098]

[0099] In the formula, The accuracy of the model is the proportion of samples correctly classified by the model out of the total number of samples, reflecting the overall accuracy of the model. , This refers to the situation where a positive example is predicted and actually is a positive example. , This refers to a situation where a positive example is predicted, but a negative example actually occurs. , This refers to a situation where a negative example is predicted but a positive example actually occurs. , This refers to a situation where a negative example is predicted and a negative example actually occurs. and The higher, and The lower the value, the higher the accuracy of the model; Recall represents the model's performance; a higher recall indicates better performance in recognizing high-quality dispersive images. Precision represents the proportion of truly positive samples among all predicted positive samples, reflecting the model's ability to distinguish low-quality dispersive images; higher precision indicates more accurate model discrimination. The score is the harmonic mean of precision and recall, comprehensively reflecting the robustness of the model; a higher score indicates better performance. The scores indicate that the model demonstrates greater robustness while maintaining a balance between accuracy and recall.

[0100] In summary, through comparative studies of various network structures, the VGG16 model demonstrates stable performance across different classification tasks. It features a small number of model parameters and low computational cost, making it suitable for deployment on mobile or embedded devices in engineering sites to achieve rapid identification of compaction mass.

[0101] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images, characterized in that, include: Wave tests were conducted on earth-rock fill materials with different compaction masses to collect elastic wave signals; The compaction quality of earth-rock fill material is determined by the dry density and porosity of the fill body; The elastic wave signal is converted into a dispersion image based on multiple wavelet transforms; Construct a training dataset for model training; the training dataset includes dispersion images labeled with preset compaction quality levels; A convolutional neural network is constructed and a compaction quality assessment model is trained based on the training dataset. The compaction quality assessment model uses the dispersion image of the soil and rock fill material to be tested as input to evaluate the compaction quality of the soil and rock fill material to be tested.

2. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, Wave tests were conducted on earth-rock fill materials with different compaction masses to collect elastic wave signals, including: The earth and rock fill material is compacted by a compaction device, and fluctuation tests are performed on several areas of the compacted surface. During testing, an external triggering method is used for excitation. A high-frequency signal is generated by hammering a steel plate, and the elastic wave signal is received by multiple detectors arranged at equal intervals.

3. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, The elastic wave signal is converted into a dispersion image based on multiple wavelet transforms, including: In the time domain, a wavelet transform is performed on an elastic wave signal to obtain the amplitude and phase information of the wave components at different frequencies in time, thus converting the elastic wave signal into a time-frequency domain. domain; The obtained time-frequency domain signal along Perform wavelet transform on the axis to transform the time-frequency domain signal to frequency domain. wave number domain; Based on the mapping relationship between frequency and wavenumber, the frequency-wavenumber spectrum in the frequency-wavenumber domain is transformed into... The dispersion energy spectrum of the domain.

4. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, Construct a training dataset for model training, including: The porosity of earth and rock fill material is used as the standard for classifying compaction quality grades, and the compaction quality of earth and rock fill material is classified into grades. Labels are set for each dispersion image based on the compaction quality grade classification results to obtain dispersion images with preset compaction quality grade labels, and the training dataset is constructed.

5. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, A compaction quality assessment model is trained based on the aforementioned training dataset, including: According to a preset ratio, the training dataset is divided into a training set for model training, a validation set for adjusting hyperparameters, and a validation set for testing the performance of the compaction quality assessment model.

6. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, After generating the compaction quality assessment model, the method further includes: The loss value of the compaction quality assessment model is analyzed using the cross-entropy loss function in the multi-class case, and is expressed as: In the formula, The number of categories; For sign functions, the values ​​include 0 and 1. When the first... Each sample belongs to When the class is selected, the value is 1; otherwise, it is 0. Indicates the prediction result, the first The predicted probability that a sample belongs to this class; Indicates the first The loss value for each sample; The performance of the compaction quality assessment model is evaluated based on the training loss curve and the validation loss curve.

7. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, After generating the compaction quality assessment model, the method also includes assessing the accuracy of the compaction quality assessment model. Accuracy Recall rate and Fraction The reliability of the compaction quality assessment model is measured.

8. The method for classifying and evaluating the compaction quality of fill based on elastic wave dispersion images according to claim 1, characterized in that, The compaction quality of the tested earth-rock fill material is evaluated based on the compaction quality assessment model, including: Based on the compaction quality assessment model, the dispersion image features of the soil and rock fill material to be tested are extracted, and the compaction quality evaluation category label and probability value of the soil and rock fill material to be tested are output. Compare the compaction quality evaluation category label with the category of actual compaction quality; The compaction quality evaluation category is determined based on the probability value, and the compaction quality evaluation of the soil and rock fill material to be tested is completed.

Images