A method for detecting the compaction quality of a roadbed

Abstract

The application discloses a kind of detection methods of subgrade compaction quality, more specifically related to subgrade quality detection technical field, database based on image-vibration signal is constructed;The data of database includes the vibration signal and subgrade image signal collected;With the input of adjusted subgrade image signal and vibration signal, a feature preprocessing model is established, the adjusted subgrade surface image signal is extracted using a deep learning algorithm, and the vibration signal is extracted using a voice detection model;And build a multi-modal data compaction degree detection method;The compaction degree of subgrade is detected using the multi-modal data compaction degree detection method.

Description

Technical Field

[0001] This invention relates to the field of roadbed quality testing technology, and specifically to a method for testing the compaction quality of roadbed. Background Technology

[0002] As a major component of highway structures, the roadbed bears both static and dynamic loads. The compaction degree of the roadbed is a crucial parameter for evaluating its quality, directly affecting its performance and service life. Both the "Highway Roadbed Design Specification" (JTG D30-2015) and the "Highway Roadbed Construction Technical Specification" (JTGT3610-2019) consider compaction degree a key standard for roadbed construction control. Insufficient compaction during roadbed construction can lead to problems such as pavement collapse, rutting, water damage, and settlement. Therefore, controlling and monitoring roadbed compaction degree is of great significance.

[0003] Currently, commonly used compaction testing methods in my country include the ring cutter method, sand cone method, and water cone method. These methods all employ the simple equivalence principle, converting soil characteristic parameters with standard materials to obtain intuitive measurement data, thereby calculating the basic compaction data. However, these traditional methods have several drawbacks: First, sampling and testing after compaction cannot provide real-time measurement and assessment of the compaction state, exhibiting lag and difficulty in remediation; second, the representativeness of sampling tests may lead to errors, making it difficult to accurately reflect the compaction status of the entire road; furthermore, sampling and testing are arduous, time-consuming, and expensive; most importantly, during actual construction, unreasonable material gradation or excessively high or low moisture content may lead to the formation of "weak points," which traditional methods often miss, creating potential safety hazards for future road use.

[0004] To address this issue, domestic and international experts have proposed several intelligent compaction testing methods through extensive engineering verification, leading to the emergence of various compaction quality evaluation indicators. Intelligent compaction testing methods are characterized by high efficiency, large sample size, and good representativeness. Through mathematical statistical analysis, they play a crucial role in improving and ensuring highway construction quality and enhancing project durability. However, the application of intelligent compaction indicators is still not mature enough, lacking theoretical basis and practical experience, and also lacking a complete compaction quality evaluation system. Summary of the Invention

[0005] To address this issue, the present invention proposes a method for detecting the compaction quality of roadbeds. This method fully integrates images of compacted soil and vibration signals from the compaction process of a road roller, and comprehensively analyzes these two types of information to reasonably assess the compaction quality of the roadbed. This method provides an important reference for the digital construction of roadbed engineering and will play a positive role in the field of highway construction, solving the problems raised in the background art.

[0006] The technical solution provided by this invention is as follows: A method for detecting the quality of roadbed compaction, the method comprising:

[0007] Construct a database based on image-vibration signals; the database includes collected vibration signals and roadbed image signals.

[0008] Using adjusted roadbed image signals and vibration signals as input, a feature preprocessing model is established. Deep learning algorithms are used to extract features from the adjusted roadbed surface image signals, and a speech detection model is used to extract features from the vibration signals.

[0009] Construct a multimodal data compaction degree detection method;

[0010] The compaction degree of the roadbed was detected using a multimodal data compaction degree detection method.

[0011] Preferably, the acquired image signal data is matched one-to-one with the vibration signal data, a compaction degree classification label is added, and all image signal data are adjusted to the same resolution. Then, the adjusted image signal data is saved to form a database.

[0012] Preferably, the feature preprocessing model uses EfficientNetV2 network to extract image features and VGGish network to extract vibration features as preprocessing data; in EfficientNetV2 network, SimAM attention is used instead of the traditional SE attention mechanism to improve the efficiency and accuracy of feature extraction, while VGGish network focuses on the feature extraction of vibration signals.

[0013] Preferably, a compaction quality detection method based on multimodal data is constructed by using preprocessed data as input to the AVCA multimodal fusion network;

[0014] The AVCA network, based on MLP and Transformer, processes data by removing the original text feature encoding. Image features processed by the EfficientNetV2 model and vibration signal features processed by the VGGish model are used as input to the AVCA network and fed into its encoding layer for feature encoding. Subsequently, these features undergo deep feature extraction via a cross-attention module to discover and fuse the correlation between image and audio features. Next, projection and reconstruction layers are used to explore feature similarities between different modalities. Finally, loss calculation and output compaction degree prediction classification are used to achieve effective data processing and analysis.

[0015] Preferably, the coding layer uses a pre-trained feature extractor to extract features of vibration and visual modalities, including a vibration signal feature extractor and an image effect extractor. These features are fed into a cross-attention module for further processing after passing through the encoder in parallel.

[0016] The encoder contains a hidden layer whose structure varies depending on the output of the image and vibration feature extraction network; this hidden layer contains a linear layer, BN normalization, activation function, and dropout layer; if no hidden layer is needed, the encoder contains only a linear layer, BN normalization, activation function, and dropout layer.

[0017] To meet the input requirements, the special image needs to be adjusted. If the special image has a large number of channels, but the required number of channels for the input is small, and different Dropout methods have a significant impact on the model, a progressive descent strategy is needed to buffer this. The hidden layers here exist to address this situation.

[0018] Preferably, in the cross-attention module, features between vibration signals and image signals are shared; after feature input from the coding layer, image features and vibration features are concatenated and then fused continuously through a self-attention module, an LN normalization layer, a feedforward layer, and another LN normalization layer; the feedforward layer includes a linear layer, a GELU activation function, a Dropout layer, another linear layer, and a Dropout layer; subsequently, the final attention map is obtained by short-circuiting and fusing these processed features with feature maps from the image coding layer and the vibration signal coding layer, respectively.

[0019] Preferably, the projection layer and reconstruction layer contain the same content as the encoding layer, including a linear layer, BN normalization, activation function, and Dropout layer.

[0020] Preferably, the loss function Loss is a regularized loss function that retains information from each input mode and is measured using mean squared error.

[0021]

[0022] These are the feature data after the image data and vibration data have been reconstructed; φ v φ a It is preprocessed feature data of images and vibration signals; θ v θ a It is the output of image and vibration data features after passing through the projection layer, where n represents the number of training data, b i For each training data's true value, c iThis represents the model's predicted value for the data.

[0023] Preferably, a trained multimodal data compaction detection method is used in conjunction with an embedded device to detect the compaction of the roadbed on site.

[0024] The present invention has the following advantages:

[0025] This invention, through an improved multi-source data model, collects fewer types of data compared to existing technologies, is easier to deploy, and can be met by existing equipment. The lightweight model can also complete the training and testing of multimodal data more quickly.

[0026] By extracting image and signal data, this invention significantly improves the prediction accuracy of roadbed compaction compared to existing technologies, which helps to enhance dynamic management and improve construction efficiency. Attached Figure Description

[0027] Figure 1 Attention map provided for this invention;

[0028] Figure 2 shows the results of the collected data, trained model, and actual measurements of this invention. Detailed Implementation

[0029] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention provides a method for detecting the compaction quality of roadbed, the method comprising:

[0031] Construct a database based on image-vibration signals; the database includes collected vibration signals and roadbed image signals.

[0032] Specifically, the collected image signal data and vibration signal data are matched one-to-one, classification labels are added, and all image signal data are adjusted to the same resolution. The adjusted image signal data is then saved to form a database.

[0033] Using adjusted roadbed image signals and vibration signals as input, a feature preprocessing model is established. Deep learning algorithms are used to extract features from the adjusted road surface image signals, and a speech detection model is used to extract features from the vibration signals.

[0034] Construct a multimodal data compaction degree detection method;

[0035] The compaction degree of the roadbed was detected using a multimodal data compaction degree detection method.

[0036] Preferably, the feature preprocessing model uses EfficientNetV2 network to extract image features and VGGish network to extract vibration features as preprocessing data; in EfficientNetV2 network, SimAM attention is used instead of the traditional SE attention mechanism to improve the efficiency and accuracy of feature extraction, while VGGish network focuses on the feature extraction of vibration signals.

[0037] Preferably, a compaction quality detection method based on multimodal data is constructed by using preprocessed data as input to the AVCA multimodal fusion network;

[0038] The AVCA network, based on MLP and Transformer, processes data by removing the original text feature encoding. Image features processed by the EfficientNetV2 model and vibration signal features processed by the VGGish model are used as input to the AVCA network and fed into its encoding layer for feature encoding. Subsequently, these features undergo deep feature extraction via a cross-attention module to discover and fuse the correlation between image and audio features. Next, projection and reconstruction layers are used to explore feature similarities between different modalities. Finally, loss calculation and output compaction degree prediction classification achieve effective data processing and analysis. The classification labels are <90, 90, ..., 95, ..., 98, >99, where 90 represents a compaction degree of 90-91%. For specific projects, the number of labels can be reduced to improve efficiency, such as <90, 90-93, 93-95, 95-98, >98.

[0039] Preferably, the encoding layer utilizes a pre-trained feature extractor to extract features of vibration and visual modalities, including a vibration signal feature extractor and an image effect extractor. These features are fed into a cross-attention module for further processing after passing through the encoder in parallel. The encoder contains a hidden layer whose structure varies depending on the output of the image and vibration feature extraction network. This hidden layer contains a linear layer, BN normalization, activation function, and Dropout layer. If a hidden layer is not required, the encoder contains only a linear layer, BN normalization, activation function, and Dropout layer.

[0040] Preferably, in the cross-attention module, features between the vibration signal and the image signal are shared; after feature input from the encoding layer, image features and vibration features are concatenated and then fused sequentially through a Self-Attention module, an LN normalization layer, a feedforward layer, and another LN normalization layer; the feedforward layer includes a linear layer, a GELU activation function, a Dropout layer, another linear layer, and a Dropout layer; subsequently, by short-circuiting these processed features with feature maps from the image encoding layer and the vibration signal encoding layer respectively, the final attention map is obtained, as shown below. Figure 1 As shown.

[0041] Preferably, the projection layer and reconstruction layer contain the same content as the encoding layer, including a linear layer, BN normalization, activation function, and Dropout layer.

[0042] Preferably, the loss function Loss is a regularized loss function that retains information from each input mode and is measured using mean squared error.

[0043]

[0044] These are the feature data after the image data and vibration data have been reconstructed; φ v φ a It is preprocessed feature data of images and vibration signals; θ v θ a It is the output of image and vibration data features after passing through the projection layer, where n represents the number of training data, b i For each training data's true value, c i This represents the model's predicted value for the data.

[0045] Preferably, a trained multimodal data compaction detection method is used in conjunction with an embedded device to detect the compaction of the roadbed on site.

[0046] Based on the above, the subgrade compaction quality is tested as follows:

[0047] Before construction, an industrial camera is installed on the road roller to collect image data, and an accelerometer is used to collect vibration signals, which are then transmitted to the terminal.

[0048] During the data acquisition phase (first stage of construction), images and vibration signals are acquired in real time, and the coordinate information of the location is read from the GNSS system. The compaction degree is then determined using the sand cone method at the corresponding location. A basic dataset containing images, vibration data, and output compaction degrees is completed.

[0049] Based on the dataset, the model proposed in this method is trained, and the loss and accuracy are monitored until the accuracy reaches 0.9, and the model weight file is saved.

[0050] During the roadbed compaction construction, image data continues to be collected and input into the trained model to obtain the output compaction value, thereby achieving rapid real-time detection of compaction quality.

[0051] Figure 2 shows the results of data collection, model training, and field measurements using this method. The results are compared with those obtained using traditional methods, demonstrating accuracy and meeting precision requirements, thus enabling efficient detection of compaction degree.

[0052] Figures 2(a) and 2(b) show the vibration signal and image data of test point 1. The model prediction obtained by this method is 90%, that is, the compaction degree is 90-91%, while the measured compaction degree is 90.3%.

[0053] Figures 2(c) and 2(d) show the vibration signal and image data of a certain test point 2. The model prediction obtained by this method is 96%, that is, the compaction degree is 96-97%, while the measured compaction degree is 96.2%.

[0054] The accuracy of the above embodiments meets industry requirements, and the completed testing is beneficial to improving project quality and construction efficiency.

[0055] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for detecting the compaction quality of roadbed, characterized in that: The detection methods include: Construct a database based on image-vibration signals; the database includes collected vibration signals and roadbed image signals. Using adjusted roadbed image signals and vibration signals as inputs, a feature preprocessing model is established. Deep learning algorithms are used to extract features from the adjusted road surface image signals, and a speech detection model is used to extract features from the vibration signals. A multimodal data compaction detection method is constructed, and the roadbed compaction is detected using the multimodal data compaction detection method. The feature preprocessing model uses EfficientNetV2 network to extract image features and VGGish network to extract vibration features as preprocessing data. In EfficientNetV2 network, SimAM attention is used instead of the traditional SE attention mechanism, while VGGish network focuses on feature extraction of vibration signal. Using preprocessed data as input to the AVCA multimodal fusion network, a compaction quality detection method based on multimodal data is constructed, as follows: The AVCA network, based on MLP and Transformer, processes data by removing the original text feature encoding. Image features processed by the EfficientNetV2 model and vibration signal features processed by the VGGish model are used as inputs to the AVCA network and fed into its encoding layer for feature encoding. This encoding layer utilizes pre-trained feature extractors to extract features from vibration and visual modalities, including vibration signal feature extractors and image effect extractors. These features, after passing through the encoder in parallel, are fed into a cross-attention module for further processing. Subsequently, these features undergo deep feature extraction through the cross-attention module to discover and fuse the correlation between image features and vibration signal features. Next, projection and reconstruction layers are used to explore feature similarities between different modalities. Finally, the loss is calculated using a loss function, and a predicted classification of compaction degree is output. In the cross-attention module, features between vibration signals and image signals are shared. After feature input from the coding layer, image features and vibration features are concatenated and then fused through a self-attention module, an LN normalization layer, a feedforward layer, and another LN normalization layer. Subsequently, the final attention map is obtained by short-circuiting and fusing these processed features with feature maps from the image coding layer and the vibration signal coding layer, respectively. The projection layer and reconstruction layer contain the same content as the encoding layer, including a linear layer, BN normalization, activation function, and Dropout layer.

2. The method for detecting the compaction quality of roadbed according to claim 1, characterized in that: The acquired image signal data is matched one-to-one with the vibration signal data, classification labels are added, and all image signal data are adjusted to the same resolution. The adjusted image signal data is then saved to form a database.

3. The method for detecting the compaction quality of roadbed according to claim 1, characterized in that: The loss function uses a regularized loss function to preserve information from each input mode and is measured using mean squared error. , ; , These are the feature data after the image data and vibration data have been reconstructed, respectively. , It consists of preprocessed feature data of images and vibration signals; , It is the output of image and vibration data features after passing through the projection layer, where n represents the number of training data. For the true value of each training data point, This represents the model's predicted value for the data.

4. The method for detecting the compaction quality of roadbed according to claim 1, characterized in that: A trained multimodal data compaction detection method, combined with embedded equipment, is used to detect the compaction of the roadbed on site.

Images