Highway subgrade rolling stone early warning method and system

Abstract

The invention relates to the technical field of highway safety, and particularly discloses a highway subgrade rolling stone early warning method and system. The method comprises the following steps: collecting vibration data along a highway subgrade by using a distributed optical fiber sonic sensor; constructing an original event data set, simulating a rolling stone event by throwing stones and collecting blocking event vibration data in a normal traffic flow, generating a vibration energy distribution diagram, and identifying and marking energy high-value point locations to obtain data fragments; carrying out sample expansion processing on the rolling stone event data fragments, and carrying out stratified sampling and clustering analysis on the blocking event data fragments to form a balanced final event data set; converting the vibration data into a two-dimensional spectrogram; training the spectrogram by using a multi-scale convolutional neural network model; and classifying real-time vibration data by using the trained model, and accurately identifying a rolling stone event and a blocking event. The system comprises a data acquisition module, a data set preprocessing module, a feature conversion module and an intelligent identification module. Through efficient data processing and a deep learning model, rolling stone and interference events can be effectively distinguished, and the early warning precision and reliability are improved.

Description

Technical Field

[0001] This application relates to the field of highway safety technology, and in particular to a method and system for early warning of falling rocks on highway subgrade. Background Technology

[0002] The highway network is the artery of the national economy, and its safe and smooth operation is crucial. However, highways traversing mountainous and hilly areas face severe challenges to the stability of their roadbed slopes due to long-term exposure to the natural environment and engineering activities. Events such as rockfalls and small landslides are highly sudden, random, and extremely destructive. A single rock rolling down a slope can potentially trigger a serious traffic accident, causing casualties, vehicle damage, and traffic disruption, resulting in incalculable losses.

[0003] Chinese Patent, Publication No. CN108120428A, Publication Date: June 5, 2018, discloses a model test system and method for reconstructing and monitoring the trajectory of a rolling stone. The system includes a base with a simulated mountain on it, a rolling stone release device on the simulated mountain, a protective fence surrounding the simulated mountain, a camera and a laser scanner next to the simulated mountain, and a force acquisition unit on the protective fence to collect video images of the rolling stone's descent, changes in laser scanning point cloud data, and force monitoring data on the impact of the rolling stone on the protective fence. This invention uploads the video images of the rolling stone's descent, the changes in laser scanning point cloud data, and the force monitoring data on the impact of the rolling stone on the protective fence to a computer, comprehensively processes and reconstructs the rolling stone's trajectory, and monitors and warns of rolling stone intrusion.

[0004] The shortcomings of the above-mentioned technical solutions are as follows: they rely on physical model testing, requiring the construction of simulated mountains and the installation of multiple large devices (such as cameras, laser scanners, and protective fences), which leads to long deployment cycles, high costs, and inability to be directly applied to real highway subgrades. The rockfall release device and mountain filling need to be customized according to specific terrain, making rapid expansion difficult; multi-sensor data fusion processing (such as video overlay and point cloud analysis) is computationally complex and has poor real-time performance, which may not meet the needs of real-time early warning in highway operations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a method and system for early warning of roadbed rockfall, which can overcome the problems of complex deployment and poor real-time performance of traditional physical models, and achieve real-time and accurate event identification based on vibration data.

[0006] To achieve the above objectives, this application adopts the following technical solution: This application provides a method for early warning of falling rocks on highway subgrade, the method comprising: S1, uses distributed fiber optic acoustic sensors to collect vibration data along the roadbed; S2, construct the original event dataset, including: simulating a boulder rolling event by throwing stones of different weights to obtain vibration data of the boulder rolling event; collecting vibration data of high-energy blocking events in normal traffic flow; generating a time- and space-based vibration energy distribution map centered on the event occurrence point, identifying and marking the time and space locations of high-energy points to obtain marked points; taking a predetermined number of points in space centered on the marked points, and extracting a predetermined duration in time to obtain data segments; obtaining the original event dataset based on the data segments; S3: Expand the data segments related to the Rolling Stones incident in the original event dataset to form an expanded dataset; perform stratified sampling on the data segments related to the blocking incident in the original event dataset to obtain spectral features; based on the spectral features, use a clustering algorithm to aggregate the data segments related to the blocking incident into several categories according to feature similarity; randomly select samples from each category according to a preset ratio to form a negative sample training set; based on the expanded dataset and the negative sample training set, obtain the final event dataset. S4 converts the vibration data in the final event dataset into a two-dimensional spectrogram; S5 uses a multi-scale convolutional neural network model to train the spectrogram; S6 uses a trained convolutional neural network model to classify real-time vibration data and identify stone rolling events and blocking events.

[0007] As a preferred technical solution, in S3, the data segments about the rolling stone event in the original event dataset are expanded to form an expanded dataset, which includes: taking the time point of the marked point as the center, performing a translation operation on the time axis to generate new samples at predetermined time intervals; linearly scaling the amplitude of the vibration data to simulate rolling stone events of different intensities; and superimposing the vibration data with the background noise data to simulate real environmental interference.

[0008] As a preferred technical solution, in S3, the spectral characteristics include: the spectral entropy and subband variance of the vibration data; the formula for calculating the spectral entropy of the vibration data is: ,in , The frequency spectrum of the vibration data is given. The sub-band variance of the vibration data is calculated as follows: the frequency spectrum of the vibration data is divided into M sub-bands, each sub-band containing the following frequency points. Then the variance of the m-th sub-band is: ,in Let be the spectral mean of the m-th sub-band.

[0009] As a preferred technical solution, in S4, converting the vibration data in the final event dataset into a two-dimensional spectrogram includes: treating the data segments in the final event dataset as one-dimensional time series of several channels, performing a short-time Fourier transform on the vibration data of each channel to obtain the two-dimensional spectrogram of each channel; ensuring that the key features of the vibration data are not destroyed, adjusting the parameters to make the generated spectrogram close to 1:1 in length and width; and merging the two-dimensional spectrograms of each channel into a multi-channel two-dimensional spectrogram.

[0010] As a preferred technical solution, S5 uses a multi-scale convolutional neural network model to train the spectrogram, including: constructing a convolutional neural network model based on the Inception v1 architecture, which uses convolutional kernels of various sizes in parallel to capture features at different scales at the same level; using the structure type information of the vibration data acquisition points as additional input features and fusing them before the fully connected layer of the convolutional neural network model, enabling the model to learn the key knowledge that "the same rolling stone event may exhibit different vibration characteristics on different structures"; optimizing hyperparameters such as learning rate and model depth before training, using a narrow search range and K-fold cross-validation to determine the optimal hyperparameter combination; and monitoring the performance of the convolutional neural network model using a validation set during training and employing strategies such as early stopping to prevent overfitting.

[0011] As a preferred technical solution, in S1, the collection of vibration data along the roadbed using a distributed optical fiber acoustic sensor includes: deploying the distributed optical fiber acoustic sensor in the equipment room of the highway, and continuously collecting vibration data using a communication optical cable deployed under the median strip of the highway; the distributed optical fiber acoustic sensor operates at a preset sampling rate and has a preset spatial resolution, and continuously monitors the vibration data along the roadbed.

[0012] This application also provides a roadbed rockfall early warning system, the system comprising: The data acquisition module is used to collect vibration data along the roadbed using distributed fiber optic acoustic sensors. The dataset preprocessing module is used to construct the original event dataset and perform expansion and stratified sampling on the original event dataset to obtain the final event dataset. The original event dataset includes vibration data obtained by simulating a rockfall event and collecting blocking events in normal traffic flow, and high-energy points are identified and marked based on the vibration energy distribution map to generate data segments. The feature conversion module is used to convert the vibration data in the final event dataset into a two-dimensional spectrogram. The intelligent recognition module is used to train a two-dimensional spectrogram using a multi-scale convolutional neural network model, and to classify real-time vibration data using the trained model to identify stone rolling events and blockage events.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: This application employs distributed fiber optic acoustic sensors to continuously collect vibration data along the highway subgrade, avoiding complex physical equipment deployment. A high-quality training set is generated by constructing a raw event dataset, including simulated rockfall events and normal traffic flow data, and utilizing amplification and stratified sampling. The vibration data is then converted into two-dimensional spectrograms and trained and classified using a multi-scale convolutional neural network (CNN). This data-driven approach achieves end-to-end automated processing without manual intervention. The application directly classifies the spectrograms in real time using a CNN model, significantly improving the speed of early warning and making it suitable for highway operation scenarios.

[0014] In this application, the data fragments of the rockfall event are expanded, and the data of the blocking event are clustered and sampled to ensure the balance and representativeness of the dataset and reduce false alarms. The CNN model can learn the deep features of the vibration signal, effectively distinguishing between rockfalls and traffic interference, and overcoming the problem of optical equipment being susceptible to environmental interference in the prior art. In addition, the distributed fiber optic sensors can cover long-distance roadbeds, are insensitive to environmental changes, and improve the robustness and deployment flexibility of the system. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a method for early warning of falling rocks on highway subgrade; Figure 2 Flowchart of the steps to construct the original event dataset; Figure 3 A two-dimensional spectrogram as input to a convolutional neural network model. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in specific embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0017] like Figure 1 As shown, this application provides a method for early warning of roadbed rockfall, the method comprising: S1 uses distributed fiber optic acoustic sensors to collect vibration data along the roadbed.

[0018] S2, as Figure 2As shown, the original event dataset is constructed, including: simulating a boulder rolling event by throwing stones of different weights to obtain vibration data; collecting vibration data of high-energy blocking events in normal traffic flow; generating a time- and space-based vibration energy distribution map centered on the event occurrence point, identifying and labeling the time and spatial locations of high-energy points, obtaining labeled points; selecting a predetermined number of points spatially centered on the labeled points, and extracting data segments of predetermined duration in time; and obtaining the original event dataset based on the data segments.

[0019] S3 involves expanding the data segments related to the Rolling Stones incident in the original event dataset to create an expanded dataset. Then, stratified sampling is performed on the data segments related to the blocking incident in the original event dataset to obtain spectral features. Based on these spectral features, a clustering algorithm is used to aggregate the data segments related to the blocking incident into several categories according to feature similarity. Samples are randomly selected from each category according to a predetermined ratio to form a negative sample training set. The final event dataset is obtained based on the expanded dataset and the negative sample training set.

[0020] S4 converts the vibration data in the final event dataset into a two-dimensional spectrogram.

[0021] S5 uses a multi-scale convolutional neural network model to train the spectrogram.

[0022] S6 uses a trained convolutional neural network model to classify real-time vibration data and identify stone rolling events and blocking events.

[0023] As a preferred technical solution, in step S1, collecting vibration data along the roadbed using a distributed fiber optic acoustic sensor includes: deploying the distributed fiber optic acoustic sensor in a control room within the highway and continuously collecting vibration data using a communication optical cable deployed beneath the central median strip. The distributed fiber optic acoustic sensor operates at a preset sampling rate and has a preset spatial resolution, continuously monitoring vibration data along the roadbed. In this application, the distributed fiber optic acoustic sensor used has a sampling rate of 2000 Hz and a spatial resolution of 10 meters.

[0024] In S2, rocks of varying weights are thrown to simulate rockfall events and obtain vibration data. Controlled field experiments are conducted to collect data on potential roadbed obstruction events such as falling rocks, spilled materials, and rockfalls from slopes. Specifically, data is collected using an aerial work platform by throwing rocks from 1 to 7 meters at 1-meter intervals, with rock weights ranging from 5 kg to 25 kg.

[0025] Specifically, based on the on-site experimental recording points and recording time, a vibration energy distribution map of a nearby 21-second × 21-point area is generated. The rolling stone event usually presents as a "bright spot" with significantly enhanced energy at a specific time and location. The specific time and location where obvious high energy values ​​appear are manually identified and marked.

[0026] Assume the vibration signal at the sampling point is The corresponding vibrational energy is: ; Subsequently, taking two points to the left and two points to the right of the marked point as the center, a total of 5 points corresponding to 50 meters, and taking the time of the event as the end point, and then taking 3 seconds forward, we finally obtained a raw event data segment with a size of 3 seconds × 5 points.

[0027] Using the same sampling procedure, data on high-energy, suspected disruption events within normal traffic flow on highways were collected and categorized as "Other". This resulted in a standard-sized raw event dataset with precise time-location labels.

[0028] As a preferred technical solution, in S3, the data segments related to the rolling stone event in the original event dataset are expanded to form the expanded dataset. This expansion includes: translating the data along the time axis with the time points of the marked locations as the center, generating new samples at predetermined time intervals; linearly scaling the amplitude of the vibration data to simulate rolling stone events of different intensities; and superimposing the vibration data with background noise data to simulate real-world environmental disturbances.

[0029] Specifically, to address the imbalance between the limited amount of blocking event data and the massive amount of "other" category data, a strategy combining "event data expansion" and "stratified sampling of non-Rolling Stone event data" is adopted. The "event data expansion" involves three types of expansion for the labeled and clipped event data: Time Shift: In real-world scenarios, the rolling stone can occur at any time, not just at the precise time point we labeled. Therefore, we shift the data matrix along the time axis. At 100-millisecond intervals, we shift the data no more than 1 second before and after the original event time point, generating new data samples. This operation ensures that the model learns not features at a fixed point in time, but rather the complete waveform feature pattern before and after the event, enhancing the model's temporal robustness.

[0030] Energy Adjustment: The energy (signal amplitude) of the rolling stone varies depending on the size of the stone, the drop, and the distance of the landing point from the optical fiber. Field experiments show that the detection rate decreases when the energy is too low. Therefore, we linearly scaled the amplitude values ​​of the data matrix, reducing them from the original energy value to at least 0.7 times in 0.05 increments. This simulates rolling stone events of varying intensities, forcing the model to learn the essential characteristic of constant amplitude, rather than relying solely on amplitude magnitude for judgment.

[0031] Noise superposition: Real road environments are filled with various background vibrations, such as vehicle traffic. To adapt the model to complex environments, we superimpose the processed event data with background noise data (non-rockfall event data) randomly collected online. This simulates a rockfall scenario against a traffic background, significantly improving the realism of the generated data and the model's robustness to interference.

[0032] As a preferred technical solution, in S3, the spectral characteristics include: the spectral entropy and subband variance of the vibration data. The formula for calculating the spectral entropy of the vibration data is: ,in , The frequency spectrum of the vibration data is shown. The sub-band variance of the vibration data is calculated as follows: the frequency spectrum of the vibration data is divided into M sub-bands, each sub-band containing the following frequency points. Then the variance of the m-th sub-band is: ,in Let be the spectral mean of the m-th sub-band.

[0033] Specifically, due to the massive amount of data in the "Other" category, directly using all the data is impractical and would lead to class imbalance. We employ stratified sampling to ensure representativeness: First, all "Other" data are similarly segmented into 3-second × 5-point segments, and their spectral characteristics are calculated, including the signal's spectral entropy, subband variance, and other features. Assuming the power spectrum of the signal collected within 3 seconds at the sampling points is... The corresponding spectral entropy: ; in: ; Suppose the signal spectrum is divided into M sub-bands, and each sub-band contains the following frequency points: Then the variance of the m-th sub-band is: ; in Let be the power spectral mean of the m-th sub-band.

[0034] Then, a clustering algorithm is used to group these data into several categories based on feature similarity. Finally, samples are randomly drawn from each cluster according to a preset ratio to form a negative sample training set. This ensures that the drawn negative samples can cover various types of background vibrations.

[0035] Ultimately, the constructed dataset contains two main categories of data: one is a large number of diverse positive samples of "Rolling Stone events" generated by expanding a small amount of original event data; the other is a representative negative sample of "non-Rolling Stone events" obtained through stratified sampling.

[0036] As a preferred technical solution, in step S4, converting the vibration data in the final event dataset into a two-dimensional spectrogram includes: treating the data segments in the final event dataset as one-dimensional time series of several channels; performing a short-time Fourier transform on the vibration data of each channel to obtain the two-dimensional spectrogram of each channel; and adjusting parameters to make the generated spectrogram approximately 1:1 in length and width, while ensuring that the key features of the vibration data are not destroyed. Finally, the two-dimensional spectrograms of each channel are merged into a multi-channel two-dimensional spectrogram.

[0037] Specifically, raw one-dimensional vibration signals are difficult to extract effective features directly at low sampling rates. This application combines signal processing with deep learning, converting the time-domain signal into a spectrogram using Fast Fourier Transform, transforming the feature extraction problem into an image recognition problem, thereby leveraging the powerful image processing capabilities of CNNs. The model itself employs a multi-scale Inception structure to capture features at different scales and incorporates prior knowledge of the structure to improve classification accuracy.

[0038] First, each 3-second × 5-point data point obtained above is considered as a one-dimensional time series with 5 channels. A short-time Fourier transform is then performed on the data from each channel to convert it into a two-dimensional spectrogram.

[0039] The key to this step is optimizing the hyperparameters (such as window length and overlap ratio). The principle is to adjust the parameters to make the generated spectrogram approximately 1:1 in length and width, without compromising key signal features, as convolutional neural networks typically perform better when processing square images. Ultimately, the spectrograms from the five channels can be merged into a single multi-channel two-dimensional image, which serves as input to the CNN model, such as... Figure 3 As shown.

[0040] As a preferred technical solution, S5 uses a multi-scale convolutional neural network (CNN) model to train the spectrograms. This includes: constructing a CNN model based on the Inception v1 architecture; the CNN model uses convolutional kernels of various sizes in parallel, enabling it to capture features at different scales within the same layer. Structural type information (such as roadbed and slope) from the vibration data acquisition points is used as additional input features and fused before the fully connected layers of the CNN model. This allows the CNN model to learn the key knowledge that "the same rockfall event may exhibit different vibration characteristics on different structures." Before training, hyperparameters such as the learning rate and model depth are optimized using a narrow search range and K-fold cross-validation to determine the optimal hyperparameter combination. During training, the performance of the CNN model is monitored using a validation set, and strategies such as early stopping are employed to prevent overfitting.

[0041] This application also provides a roadbed rockfall warning system, the system comprising: The data acquisition module is used to collect vibration data along the roadbed using distributed fiber optic acoustic sensors.

[0042] The dataset preprocessing module is used to construct the original event dataset and perform expansion and stratified sampling on the original event dataset to obtain the final event dataset. The original event dataset includes vibration data obtained by simulating stone rolling events and collecting data on traffic disruption events in normal traffic flow. Based on the vibration energy distribution map, high-energy points are identified and labeled to generate data segments.

[0043] The feature transformation module is used to convert the vibration data in the final event dataset into a two-dimensional spectrogram.

[0044] The intelligent recognition module is used to train a two-dimensional spectrogram using a multi-scale convolutional neural network model, and to classify real-time vibration data using the trained model to identify stone rolling events and blockage events.

[0045] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "a" or "one," and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. Unless otherwise stated, terms such as "front," "back," "left," "right," "lower," and / or "upper" are for illustrative purposes only and are not limited to a location or spatial orientation. Terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0046] The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0047] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for early warning of rockfall on highway subgrade, characterized in that, The method includes: S1, uses distributed fiber optic acoustic sensors to collect vibration data along the roadbed; S2, construct the original event dataset, including: simulating a boulder rolling event by throwing stones of different weights to obtain vibration data of the boulder rolling event; collecting vibration data of high-energy blocking events in normal traffic flow; generating a time- and space-based vibration energy distribution map centered on the event occurrence point, identifying and marking the time and space locations of high-energy points to obtain marked points; taking a predetermined number of points in space centered on the marked points, and extracting a predetermined duration in time to obtain data segments; obtaining the original event dataset based on the data segments; S3: Expand the data segments related to the Rolling Stones incident in the original event dataset to form an expanded dataset; perform stratified sampling on the data segments related to the blocking incident in the original event dataset to obtain spectral features; based on the spectral features, use a clustering algorithm to aggregate the data segments related to the blocking incident into several categories according to feature similarity; randomly select samples from each category according to a preset ratio to form a negative sample training set; based on the expanded dataset and the negative sample training set, obtain the final event dataset. S4 converts the vibration data in the final event dataset into a two-dimensional spectrogram; S5 uses a multi-scale convolutional neural network model to train the spectrogram; S6 uses a trained convolutional neural network model to classify real-time vibration data and identify stone rolling events and blocking events.

2. The method for early warning of roadbed rockfall as described in claim 1, characterized in that, In S3, the data segments about the rolling stone event in the original event dataset are expanded to form the expanded dataset. This includes: translating the data along the time axis with the time point of the marked point as the center to generate new samples at predetermined time intervals; linearly scaling the amplitude of the vibration data to simulate rolling stone events of different intensities; and superimposing the vibration data with the background noise data to simulate real environmental interference.

3. The method for early warning of roadbed rockfall as described in claim 1, characterized in that, In S3, the spectral features include: the spectral entropy and subband variance of the vibration data; the formula for calculating the spectral entropy of the vibration data is: ,in , The frequency spectrum of the vibration data is given. The sub-band variance of the vibration data is calculated as follows: the frequency spectrum of the vibration data is divided into M sub-bands, each sub-band containing the following frequency points. Then the variance of the m-th sub-band is: ,in Let be the spectral mean of the m-th sub-band.

4. The method for early warning of roadbed rockfall as described in claim 1, characterized in that, In S4, converting the vibration data in the final event dataset into a two-dimensional spectrogram includes: treating the data segments in the final event dataset as one-dimensional time series of several channels, performing a short-time Fourier transform on the vibration data of each channel to obtain the two-dimensional spectrogram of each channel; adjusting the parameters to make the generated spectrogram approximately 1:1 in length and width without destroying the key features of the vibration data; and merging the two-dimensional spectrograms of each channel into a multi-channel two-dimensional spectrogram.

5. A method for early warning of roadbed rockfall according to claim 1, characterized in that, S5, training the spectrograms using a multi-scale convolutional neural network model includes: constructing a convolutional neural network model based on the Inception v1 architecture, which uses convolutional kernels of various sizes in parallel to capture features at different scales at the same level; using the structure type information of the vibration data acquisition points as additional input features and fusing them before the fully connected layer of the convolutional neural network model, enabling the model to learn the key knowledge that "the same rolling stone event may exhibit different vibration characteristics on different structures"; optimizing hyperparameters such as learning rate and model depth before training, using a narrow search range and K-fold cross-validation to determine the optimal hyperparameter combination; and monitoring the performance of the convolutional neural network model using a validation set during training and employing strategies such as early stopping to prevent overfitting.

6. The method for early warning of roadbed rockfall as described in claim 1, characterized in that, In S1, the collection of vibration data along the roadbed using distributed fiber optic acoustic sensors includes: deploying distributed fiber optic acoustic sensors in the equipment room of the highway and continuously collecting vibration data using communication optical cables deployed under the central median strip of the highway; the distributed fiber optic acoustic sensors operate at a preset sampling rate and have a preset spatial resolution, and continuously monitor the vibration data along the roadbed.

7. A roadbed rockfall early warning system, characterized in that, include: The data acquisition module is used to collect vibration data along the roadbed using distributed fiber optic acoustic sensors. The dataset preprocessing module is used to construct the original event dataset and perform expansion and stratified sampling on the original event dataset to obtain the final event dataset. The original event dataset includes vibration data obtained by simulating a rockfall event and collecting blocking events in normal traffic flow, and high-energy points are identified and marked based on the vibration energy distribution map to generate data segments. The feature conversion module is used to convert the vibration data in the final event dataset into a two-dimensional spectrogram. The intelligent recognition module is used to train a two-dimensional spectrogram using a multi-scale convolutional neural network model, and to classify real-time vibration data using the trained model to identify stone rolling events and blockage events.

Images